{"citing_id": "2305.01064v1", "cited_id": "1910.11333", "section_title": "Summary", "citation": "This finding weakens the claim made in #REFR of achieving a successful sampling task, namely achieving approximate sampling for some specific distribution, and may shed doubt on the potential of NISQ systems to achieve sampling tasks.", "text_before_citation": ["In this paper we studied three main issues, mostly from a statistical point of view.", "The first issue, already demonstrated in #OTHEREFR is the large gap between the Google samples and the Google noise model or any specific noise model we studied."], "text_after_citation": ["The second issue concerns Formula (77) of the Google paper (4) with its simple independence-like form.", "This, and the noise and calibration involved in the experimental process, make it surprising that Google's fidelity estimates appear to be rather close to Formula (77) for hundreds of different experiments.", "Surprises do occur in science, and may lead to important scientific discoveries, but from a statistical point of view, the subsumed independence between components of systems such as quantum computers, which are known to be sensitive to noise and errors caused by interactions with their environment, is striking.", "Related questions arise from the systematic deviation of Formula (77) for patch circuits, and the large difference between the F XEB fidelity of the two patches. The remarkable verification of Kalachev et al. #OTHEREFR and Liu et al.", "#OTHEREFR weakened these concerns by showing that Formula (77) provides a good approximation to hundreds of circuits for which the Google team stated that they had not computed the amplitudes."], "citing_paper_content": {"title": "Questions And Concerns About Google'S Quantum Supremacy Claim", "abstract": "In October 2019, Nature published a paper [6] describing an experimental work that was performed at Google. The paper claims to demonstrate quantum (computational) supremacy on a 53-qubit quantum computer. Since then we have been involved in a long-term project to study various statistical aspects of the Google experiment. In [30] we studied Google's statistical framework that we found to be very sound and offered some technical improvements. This document describes three main concerns (based on statistical analysis) about the Google 2019 experiment. The first concern is that the data do not agree with Google's noise model (or any other specific model). The second concern is that a crucial simple formula for a priori estimation of the fidelity seems to involve an unexpected independence assumption, and yet it gives very accurate predictions. The third concern is about statistical properties of the calibration process. 1"}, "cited_paper_content": {"title": "Supplementary Information For\"Quantum Supremacy Using A Programmable Superconducting Processor\"", "abstract": "The promise of quantum computers is that certain computational tasks might be executed exponentially faster on a quantum processor than on a classical processor1. A fundamental challenge is to build a high-fidelity processor capable of running quantum algorithms in an exponentially large computational space. Here we report the use of a processor with programmable superconducting qubits2\u20137 to create quantum states on 53 qubits, corresponding to a computational state-space of dimension 253 (about 1016). Measurements from repeated experiments sample the resulting probability distribution, which we verify using classical simulations. Our Sycamore processor takes about 200 seconds to sample one instance of a quantum circuit a million times\u2014our benchmarks currently indicate that the equivalent task for a state-of-the-art classical supercomputer would take approximately 10,000 years. This dramatic increase in speed compared to all known classical algorithms is an experimental realization of quantum supremacy8\u201314 for this specific computational task, heralding a much-anticipated computing paradigm. Quantum supremacy is demonstrated using a programmable superconducting processor known as Sycamore, taking approximately 200 seconds to sample one instance of a quantum circuit a million times, which would take a state-of-the-art supercomputer around ten thousand years to compute."}, "keywords": ["NISQ systems"], "citation_intent": "background"}
{"citing_id": "2303.05283v1", "cited_id": "1909.05787", "section_title": "B. Haptic Communication", "citation": "Alternatively, the transmitter can predict its future haptic data and transmit the predicted data to compensate for the transmission delay #REFR .", "text_before_citation": ["In the tele-operated needle insertion, the force/torque feedback from the patient is predicted by inputting the force/torque commands of the surgeon to the hidden Markov model (HMM) #OTHEREFR .", "Audiovisual data collected in the interaction with a surface material are input to a neural network-based semantic learning algorithm to predict the texture of the surface material #OTHEREFR .", "Haptic data can be predicted either at the receiver side or at the transmitter side to compensate for an excessive delay or packet loss.", "The receiver can predict the haptic data from the transmitter when an excessive delay occurs #OTHEREFR .", "For example, digital twin-based prediction can be used by the receiver for low-latency interactions #OTHEREFR ."], "text_after_citation": ["In this case, the prediction of whether haptic interaction is about to occur can assist to determine whether the haptic data prediction and the subsequent transmission are necessary #OTHEREFR .", "Haptic data prediction algorithms, such as AI-based ones, can be computing-intensive.", "To this end, they can be implemented using computing resources in the network to satisfy the stringent delay requirements #OTHEREFR , #OTHEREFR .", "In a teleoperation scenario, each of the two interacting haptic devices is associated with one edge server which caches the haptic interaction data, trains and implements the LSTM networkbased prediction algorithm, and delivers the predicted haptic data to its associated haptic device #OTHEREFR .", "Furthermore, with close proximity, auxiliary robots can be deployed around haptic devices to implement haptic data prediction and deliver the results to the devices using device-to-device (D2D) communications #OTHEREFR ."], "citing_paper_content": {"title": "Toward Immersive Communications In 6G", "abstract": "The sixth generation (6G) networks are expected to enable immersive communications and bridge the physical and the virtual worlds. Integrating extended reality, holography, and haptics, immersive communications will revolutionize how people work, entertain, and communicate by enabling lifelike interactions. However, the unprecedented demand for data transmission rate and the stringent requirements on latency and reliability create challenges for 6G networks to support immersive communications. In this survey article, we present the prospect of immersive communications and investigate emerging solutions to the corresponding challenges for 6G. First, we introduce use cases of immersive communications, in the fields of entertainment, education, and healthcare. Second, we present the concepts of immersive communications, including extended reality, haptic communication, and holographic communication, their basic implementation procedures, and their requirements on networks in terms of transmission rate, latency, and reliability. Third, we summarize the potential solutions to addressing the challenges from the aspects of communication, computing, and networking. Finally, we discuss future research directions and conclude this study."}, "cited_paper_content": {"title": "Prediction And Communication Co-Design For Ultra-Reliable And Low-Latency Communications", "abstract": "Ultra-reliable and low-latency communications (URLLC) are considered as one of three new application scenarios in the fifth generation cellular networks. In this work, we aim to reduce the user experienced delay through prediction and communication co-design, where each mobile device predicts its future states and sends them to a data center in advance. Since predictions are not error-free, we consider prediction errors and packet losses in communications when evaluating the reliability of the system. Then, we formulate an optimization problem that maximizes the number of URLLC services supported by the system by optimizing time and frequency resources and the prediction horizon. Simulation results verify the effectiveness of the proposed method, and show that the tradeoff between user experienced delay and reliability can be improved significantly via prediction and communication co-design. Furthermore, we carried out an experiment on the remote control in a virtual factory, and validated our concept on prediction and communication co-design with the practical mobility data generated by a real tactile device."}, "keywords": ["future haptic data", "transmission delay"], "citation_intent": "background"}
{"citing_id": "2303.08610v1", "cited_id": "2002.04745", "section_title": "Blind Estimation System", "citation": "For the graph decoder, we use the 6-layer transformer encoder with the pre-layer normalization #REFR and 16 heads.", "text_before_citation": ["We reuse the prototype decoder for the parameter estimation; we add another projection head, append a task token xT to differentiate the two tasks, and remove the causal attention mask.", "Since each parameter has a different range and scale, we translate and rescale the ground-truth value to fit into [0, 1] range. Architecture Details.", "The FFT size, hop length, and the number of Mel filter banks of the reference Mel spectrogram are 1536, 384, and 256, respectively.", "The convolutional backbone is a VGGish model #OTHEREFR with the following modifications: (i) depthwise separable convolutions #OTHEREFR , (ii) channels divided into four groups with dilations of #OTHEREFR , (iii) and the use of layer normalization #OTHEREFR .", "We used a transformer decoder layer with 1 (singing) and 6 (drum) queries for the pooling."], "text_after_citation": ["While the original paper used eigenvectors of the normalized Laplacian matrix as node id encoding, we use the sinusoidal embeddings since the eigenvectors are intractable during the decoding. Training.", "We train the prototype decoding task in a teacher-forcing manner using the cross-entropy losses with label-smoothing of 0.1.", "At the same time, we train the parameter estimation task by feeding an oracle prototype G0 as input and using the l1 distance as an objective.", "We use AdamW #OTHEREFR optimizer, a linear learning rate scheduler with 5e-4 peak learning rate, 50k warmup steps, 200k total training steps, and batch size of 32."], "citing_paper_content": {"title": "Blind Estimation Of Audio Processing Graph", "abstract": "Musicians and audio engineers sculpt and transform their sounds by connecting multiple processors, forming an audio processing graph. However, most deep-learning methods overlook this real-world practice and assume fixed graph settings. To bridge this gap, we develop a system that reconstructs the entire graph from a given reference audio. We first generate a realistic graph-reference pair dataset and train a simple blind estimation system composed of a convolutional reference encoder and a transformer-based graph decoder. We apply our model to singing voice effects and drum mixing estimation tasks. Evaluation results show that our method can reconstruct complex signal routings, including multi-band processing and sidechaining."}, "cited_paper_content": {"title": "On Layer Normalization In The Transformer Architecture", "abstract": "The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normalization between the residual blocks, the expected gradients of the parameters near the output layer are large. Therefore, using a large learning rate on those gradients makes the training unstable. The warm-up stage is practically helpful for avoiding this problem. On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization. This motivates us to remove the warm-up stage for the training of Pre-LN Transformers. We show in our experiments that Pre-LN Transformers without the warm-up stage can reach comparable results with baselines while requiring significantly less training time and hyper-parameter tuning on a wide range of applications."}, "keywords": ["pre-layer normalization"], "citation_intent": "method"}
{"citing_id": "2305.00604v1", "cited_id": "1912.10985", "section_title": "Introduction", "citation": "Inspired by K-FAC, there have been other works discussing approximations of G \u03b8 and its inverse #REFR .", "text_before_citation": ["where \u03bb > 0 is the so-called Tikhonov regularization parameter.", "It is well-known #OTHEREFR , #OTHEREFR , that under the assumption of approximating the model f with its first-order Taylor expansion, the Hessian corresponds with the so-called generalized Gauss-Newton (GGN) matrix G \u03b8 , and hence (4) can be expressed as", "EQUATION", "A major practical limitation of (5) is the computation of the inverse term.", "A method that alleviates this difficulty is known as Kronecker-Factored Approximate Curvature (K-FAC) #OTHEREFR which approximates the block-diagonal (i.e., layer-wise) empirical Hessian or GGN matrix."], "text_after_citation": ["In the following, we discuss a popular approach that allows for (moderately) efficient computation.", "The generalized Gauss-Newton matrix G \u03b8 is defined as", "EQUATION", "where J and \u2207 2 denote the Jacobian and Hessian matrices, respectively.", "Correspondingly, the diagonal block of G \u03b8 corresponding to the weights of the ith layer"], "citing_paper_content": {"title": "Isaac Newton: Input-Based Approximate Curvature For Newton'S Method", "abstract": "We present ISAAC (Input-baSed ApproximAte Curvature), a novel method that conditions the gradient using selected second-order information and has an asymptotically vanishing computational overhead, assuming a batch size smaller than the number of neurons. We show that it is possible to compute a good conditioner based on only the input to a respective layer without a substantial computational overhead. The proposed method allows effective training even in small-batch stochastic regimes, which makes it competitive to first-order as well as second-order methods."}, "cited_paper_content": {"title": "Backpack: Packing More Into Backprop", "abstract": "Automatic differentiation frameworks are optimized for exactly one thing: computing the average mini-batch gradient. Yet, other quantities such as the variance of the mini-batch gradients or many approximations to the Hessian can, in theory, be computed efficiently, and at the same time as the gradient. While these quantities are of great interest to researchers and practitioners, current deep learning software does not support their automatic calculation. Manually implementing them is burdensome, inefficient if done naively, and the resulting code is rarely shared. This hampers progress in deep learning, and unnecessarily narrows research to focus on gradient descent and its variants; it also complicates replication studies and comparisons between newly developed methods that require those quantities, to the point of impossibility. To address this problem, we introduce BackPACK, an efficient framework built on top of PyTorch, that extends the backpropagation algorithm to extract additional information from first-and second-order derivatives. Its capabilities are illustrated by benchmark reports for computing additional quantities on deep neural networks, and an example application by testing several recent curvature approximations for optimization."}, "keywords": ["approximations"], "citation_intent": "background"}
{"citing_id": "2304.01552v1", "cited_id": "1707.09835", "section_title": "Discussion", "citation": "Meta-SGD #REFR is a well-known constant diagonal preconditioner (i.e., diag(a 1 , \u2022 \u2022 \u2022 , a n )) that does not need to be positive definite and we also investigate its modification with a constraint on positive definiteness. The results are shown in Table 8 .", "text_before_citation": ["simple constant preconditioners: Approximate GAP is a low-complexity method where SVD Meta-SGD #OTHEREFR 2.4218 \u00d7 10 5 100.0% MC #OTHEREFR 2.7106 \u00d7 10 6 2140.4% PAMELA #OTHEREFR 1.6239 \u00d7 10 5 34.1% MH #OTHEREFR 7.2196 \u00d7 10 7 59586.7% Sparse-MAML #OTHEREFR 2.4218 \u00d7 10 5 100.0% GAP 1.2131 \u00d7 10 5 0.2% Table 9 . Ablation study of PGAP on mini-ImageNet.", "Performance of the GAP-trained model is significantly affected by not applying PGAP.", "operation is avoided by approximating GAP with a constant diagonal preconditioner.", "A natural question to ask is how does Approximate GAP compare with other constant diagonal preconditioners.", "To answer this question, we have compared Approximate GAP with Meta-SGD and a modified Meta-SGD."], "text_after_citation": ["It can be observed that enforcing positive definiteness can improve Meta-SGD.", "Furthermore, an additional improvement can be achieved by Approximate GAP.", "While both modified Meta-SGD and Approximate GAP are positive definite, Approximate GAP is different because it inherits an additional constraint from GAP -a block diagonal structure where a constant diagonal matrix M is repeated (i.e., blkdiag(M, \u2022 \u2022 \u2022 , M )).", "The inherited constraint provides a gain over the modified Meta-SGD.", "Does GAP learn a useful preconditioner: While a Riemannian metric can be helpful, it does not mean any Riemannian metric will result in an improvement."], "citing_paper_content": {"title": "Meta-Learning With A Geometry-Adaptive Preconditioner", "abstract": "Model-agnostic meta-learning (MAML) is one of the most successful meta-learning algorithms. It has a bi-level optimization structure where the outer-loop process learns a shared initialization and the inner-loop process optimizes task-specific weights. Although MAML relies on the standard gradient descent in the inner-loop, recent studies have shown that controlling the inner-loop's gradient descent with a meta-learned preconditioner can be beneficial. Existing preconditioners, however, cannot simultaneously adapt in a task-specific and path-dependent way. Additionally, they do not satisfy the Riemannian metric condition, which can enable the steepest descent learning with preconditioned gradient. In this study, we propose Geometry-Adaptive Preconditioned gradient descent (GAP) that can overcome the limitations in MAML; GAP can efficiently meta-learn a preconditioner that is dependent on task-specific parameters, and its preconditioner can be shown to be a Riemannian metric. Thanks to the two properties, the geometryadaptive preconditioner is effective for improving the innerloop optimization. Experiment results show that GAP outperforms the state-of-the-art MAML family and preconditioned gradient descent-MAML (PGD-MAML) family in a variety of few-shot learning tasks. Code is available at: https://github.com/Suhyun777/CVPR23-GAP."}, "cited_paper_content": {"title": "Meta-Sgd: Learning To Learn Quickly For Few-Shot Learning", "abstract": "Few-shot learning is challenging for learning algorithms that learn each task in isolation and from scratch. In contrast, meta-learning learns from many related tasks a meta-learner that can learn a new task more accurately and faster with fewer examples, where the choice of meta-learners is crucial. In this paper, we develop Meta-SGD, an SGD-like, easily trainable meta-learner that can initialize and adapt any differentiable learner in just one step, on both supervised learning and reinforcement learning. Compared to the popular meta-learner LSTM, Meta-SGD is conceptually simpler, easier to implement, and can be learned more efficiently. Compared to the latest meta-learner MAML, Meta-SGD has a much higher capacity by learning to learn not just the learner initialization, but also the learner update direction and learning rate, all in a single meta-learning process. Meta-SGD shows highly competitive performance for few-shot learning on regression, classification, and reinforcement learning."}, "keywords": ["Meta-SGD"], "citation_intent": "result"}
{"citing_id": "2304.14106v1", "cited_id": "1907.09190", "section_title": "Opinion Analysis", "citation": "Nonetheless, this study is limited in its scope of long-form QA task on ELI5 #REFR dataset, and further research is required to explore more elaborate knowledge analysis for more tasks and to evaluate the generalizability of our findings.", "text_before_citation": ["This suggests that ChatGPT is becoming better at generating responses that contain semantic frames or frames used to represent entities and events.", "Similarly, we can observe an decreasing trend in the frequency of argument role features such as Agent, Patient, and Theme in ChatGPT's responses over time.", "This indicates an improvement in ChatGPT's ability to recognize the argument roles of entities and events in generating responses that convey the meaning correctly.", "Overall, the results presented in Figure 6 , 7 and 8 demonstrate that ChatGPT's patterns on knowledge are showing improvement over time.", "The decreasing frequency values of named entity, relation, opinion and frame features suggest that ChatGPT has become more proficient in generating responses that contain less complex knowledge to fit the instruction \"Explain like I am five\"."], "text_after_citation": ["P e r s o n -0 1 -8 O r g a n iz a t io n -0 1 -1 8 L o c a t io n -0 1 -1 8 W o r k -0 1 -1 8 T im e -0 1 -1 R e la t io n -0 1 -1 8 P e r s o n -0 3 -0 5 O r g a n iz a t io n -0 3 -0 5 L o c a t io n -0 3 -0 5 W o r k -0 3 -0 5 T im e -0 3 -0 5 R e la t io n -0 3 -0 5 P e r s o n -0 4 -0 9 O r g a n iz a t io n -0 4 -0 9 L o c a t io n -0 4 -0 9 W o r k -0 4 -0 9 T im e -0 4 -0 9 R e la t io n -0 4 -0 9 O p in io n -0 1 -1 8 S e n t im e n t -0 1 -1 8 O p in io n -0 3 -0 5 S e n t im e n t -0 3 -0 5 O p in io n -0 4 -0 9 S e n t im e n t -0 4 -0 9"], "citing_paper_content": {"title": "Chatlog: Recording And Analyzing Chatgpt Across Time", "abstract": "While there are abundant researches about evaluating ChatGPT on natural language understanding and generation tasks, few studies have investigated how ChatGPT's behavior changes over time. In this paper, we collect a coarse-to-fine temporal dataset called ChatLog, consisting of two parts that update monthly and daily: ChatLog-Monthly is a dataset of 38,730 question-answer pairs collected every month including questions from both the reasoning and classification tasks. ChatLog-Daily, on the other hand, consists of ChatGPT's responses to 1000 identical questions for long-form generation every day. We conduct comprehensive automatic and human evaluation to provide the evidence for the existence of ChatGPT evolving patterns. We further analyze the unchanged characteristics of ChatGPT over time by extracting its knowledge and linguistic features. We find some stable features to improve the robustness of a RoBERTa-based detector on new versions of ChatGPT. We will continuously maintain our project at https://github.com/ THU-KEG/ChatLog."}, "cited_paper_content": {"title": "Eli5: Long Form Question Answering", "abstract": "We introduce the first large-scale corpus for long-form question answering, a task requiring elaborate and in-depth answers to open-ended questions. The dataset comprises 270K threads from the Reddit forum ``Explain Like I'm Five'' (ELI5) where an online community provides answers to questions which are comprehensible by five year olds. Compared to existing datasets, ELI5 comprises diverse questions requiring multi-sentence answers. We provide a large set of web documents to help answer the question. Automatic and human evaluations show that an abstractive model trained with a multi-task objective outperforms conventional Seq2Seq, language modeling, as well as a strong extractive baseline. However, our best model is still far from human performance since raters prefer gold responses in over 86% of cases, leaving ample opportunity for future improvement."}, "keywords": ["long-form QA task"], "citation_intent": "result"}
{"citing_id": "2303.07926v1", "cited_id": "1712.01980", "section_title": "Introduction", "citation": "Following this framework, semiring semantics for full first-order logic (FO) were developed in #REFR .", "text_before_citation": ["Prior to this work, these adaptations of team semantics have been studied in isolation from one another.", "Data provenance provides means to describe the origins of data, allowing to give information about the witnesses to a query, or determining how a certain output is derived.", "Provenance semirings were introduced in #OTHEREFR to devise a general framework that allows to uniformly treat extensions of positive relational algebra, where the tuples have annotations that reflect very diverse information.", "Some motivating examples of said relations come from incomplete and probabilistic databases, and bag semantics.", "This semiring framework captures a notion of data provenance called howprovenance, where the semiring operations essentially capture how each output is produced from the source."], "text_after_citation": ["The semiring semantics for FO refines, in particular, the classical Boolean semantics by allowing formulae to be evaluated as values from a semiring.", "This allows for example counting proof trees, or winning strategies in the model checking game for A and \u03c6.", "In databases, dependencies are applied as integrity constraints (ICs) that specify sets of rules that the database needs to satisfy.", "Formal analysis of the rules is facilitated by viewing them as FO sentences that usually follow certain syntactic patterns.", "This approach is sometimes inadequate because query languages such as SQL operate with multisets (i.e., bags) of tuples instead of sets. Recently, (Chu et al."], "citing_paper_content": {"title": "Unified Foundations Of Team Semantics Via Semirings", "abstract": "Semiring semantics for first-order logic provides a way to trace how facts represented by a model are used to deduce satisfaction of a formula. Team semantics is a framework for studying logics of dependence and independence in diverse contexts such as databases, quantum mechanics, and statistics by extending first-order logic with atoms that describe dependencies between variables. Combining these two, we propose a unifying approach for analysing the concepts of dependence and independence via a novel semiring team semantics, which subsumes all the previously considered variants for first-order team semantics. In particular, we study the preservation of satisfaction of dependencies and formulae between different semirings. In addition we create links to reasoning tasks such as provenance, counting, and repairs."}, "cited_paper_content": {"title": "Semiring Provenance For First-Order Model Checking", "abstract": "Given a first-order sentence, a model-checking computation tests whether the sentence holds true in a given finite structure. Data provenance extracts from this computation an abstraction of the manner in which its result depends on the data items that describe the model. Previous work on provenance was, to a large extent, restricted to the negation-free fragment of first-order logic and showed how provenance abstractions can be usefully described as elements of commutative semirings --- most generally as multivariate polynomials with positive integer coefficients. ::: In this paper we introduce a novel approach to dealing with negation and a corresponding commutative semiring of polynomials with dual indeterminates. These polynomials are used to perform reverse provenance analysis, i.e., finding models that satisfy various properties under given provenance tracking assumptions."}, "keywords": ["full first-order logic"], "citation_intent": "method"}
{"citing_id": "2303.07641v1", "cited_id": "1911.10683", "section_title": "Table Structure Recognition", "citation": "These results show that the proposed model works well on both simple and complex tables and outperforms EDD #REFR ) by about 8.1% and the best model in (Nassar et al., 2022) by about 2%.", "text_before_citation": ["In this experiment, we evaluate the effectiveness of WSTabNet for recognizing the structure of the table images on PubTabNet and FinTabNet datasets. Table  5 shows the table structure recognition performance (TEDS-struc.", "scores) of the proposed model and the previous table structure recognition methods on the validation set of the PubTabNet and the test set of the FinTabNet.", "On the test set of the FinTabNet dataset, the proposed model achieved TEDS-struc.", "of 99.06%, 98.33%, and 98.72% on the simple, complex, and all table images, respectively."], "text_after_citation": ["On the validation set of the PubTabNet dataset, the proposed model achieved TEDS-struc.", "of 97.74% on all table images which again improves EDD by about 7.8% and the best model in (Nassar et al., 2022) by about 1%.", "Note that all other methods except EDD are fully supervised approaches that require both HTML and cell bounding boxes annotations in the training step. (FT) Model was trained on PubTabNet and then finetuned."], "citing_paper_content": {"title": "Rethinking Image-Based Table Recognition Using Weakly Supervised Methods", "abstract": "Most of the previous methods for table recognition rely on training datasets containing many richly annotated table images. Detailed table image annotation, e.g., cell or text bounding box annotation, however, is costly and often subjective. In this paper, we propose a weakly supervised model named WSTabNet for table recognition that relies only on HTML (or LaTeX) code-level annotations of table images. The proposed model consists of three main parts: an encoder for feature extraction, a structure decoder for generating table structure, and a cell decoder for predicting the content of each cell in the table. Our system is trained end-to-end by stochastic gradient descent algorithms, requiring only table images and their ground-truth HTML (or LaTeX) representations. To facilitate table recognition with deep learning, we create and release WikiTableSet, the largest publicly available image-based table recognition dataset built from Wikipedia. WikiTableSet contains nearly 4 million English table images, 590K Japanese table images, and 640k French table images with corresponding HTML representation and cell bounding boxes. The extensive experiments on WikiTableSet and two large-scale datasets: FinTabNet and PubTabNet demonstrate that the proposed weakly supervised model achieves better, or similar accuracies compared to the state-of-the-art models on all benchmark datasets."}, "cited_paper_content": {"title": "Image-Based Table Recognition: Data, Model, And Evaluation", "abstract": "Important information that relates to a specific topic in a document is often organized in tabular format to assist readers with information retrieval and comparison, which may be difficult to provide in natural language. However, tabular data in unstructured digital documents, e.g., Portable Document Format (PDF) and images, are difficult to parse into structured machine-readable format, due to complexity and diversity in their structure and style. To facilitate image-based table recognition with deep learning, we develop the largest publicly available table recognition dataset PubTabNet (this https URL), containing 568k table images with corresponding structured HTML representation. PubTabNet is automatically generated by matching the XML and PDF representations of the scientific articles in PubMed Central Open Access Subset (PMCOA). We also propose a novel attention-based encoder-dual-decoder (EDD) architecture that converts images of tables into HTML code. The model has a structure decoder which reconstructs the table structure and helps the cell decoder to recognize cell content. In addition, we propose a new Tree-Edit-Distance-based Similarity (TEDS) metric for table recognition. The experiments demonstrate that the EDD model can accurately recognize complex tables solely relying on the image representation, outperforming the state-of-the-art by 7.7% absolute TEDS score."}, "keywords": ["complex tables"], "citation_intent": "result"}
{"citing_id": "2303.09949v1", "cited_id": "1909.08423", "section_title": "Properties Of Our Ansatz", "citation": "This is in contrast to previous approaches which were based on incorporating ab-initio orbitals #REFR , which could not reach chemical accuracy even for small molecules.", "text_before_citation": ["Therefore during supervised pre-training, the undetermined sign of the reference orbitals becomes irrelevant, leading to faster convergence as demonstrated in Sec. 2.4.", "\u2022 Locality: When using localized HF-orbitals as input, the resulting TAOs are also localized.", "Localized HF-orbitals are orbitals which have non-zero orbital features c Ik only on some subset of atoms.", "Since we enforce the backflow f a \u03b8 to be antisymmetric (and thus f a (0) = 0), the resulting TAOs have zero contribution from atoms I with c Ik = 0.", "\u2022 High expressivity: We empirically find that our ansatz is sufficiently expressive to model ground-state wavefunctions to high accuracy."], "text_after_citation": [], "citing_paper_content": {"title": "Towards A Foundation Model For Neural Network Wavefunctions", "abstract": "Deep neural networks have become a highly accurate and powerful wavefunction ansatz in combination with variational Monte Carlo methods for solving the electronic Schr\u00f6dinger equation. However, despite their success and favorable scaling, these methods are still computationally too costly for wide adoption. A significant obstacle is the requirement to optimize the wavefunction from scratch for each new system, thus requiring long optimization. In this work, we propose a novel neural network ansatz, which effectively maps uncorrelated, computationally cheap Hartree-Fock orbitals, to correlated, high-accuracy neural network orbitals. This ansatz is inherently capable of learning a single wavefunction across multiple compounds and geometries, as we demonstrate by successfully transferring a wavefunction model pre-trained on smaller fragments to larger compounds. Furthermore, we provide ample experimental evidence to support the idea that extensive pre-training of a such a generalized wavefunction model across different compounds and geometries could lead to a foundation wavefunction model. Such a model could yield high-accuracy ab-initio energies using only minimal computational effort for fine-tuning and evaluation of observables."}, "cited_paper_content": {"title": "Deep Neural Network Solution Of The Electronic Schr\u00f6dinger Equation.", "abstract": "The electronic Schrodinger equation describes fundamental properties of molecules and materials, but can only be solved analytically for the hydrogen atom. The numerically exact full configuration-interaction method is exponentially expensive in the number of electrons. Quantum Monte Carlo is a possible way out: it scales well to large molecules, can be parallelized, and its accuracy has, as yet, only been limited by the flexibility of the used wave function ansatz. Here we propose PauliNet, a deep-learning wave function ansatz that achieves nearly exact solutions of the electronic Schrodinger equation. PauliNet has a multireference Hartree-Fock solution built in as a baseline, incorporates the physics of valid wave functions, and is trained using variational quantum Monte Carlo (VMC). PauliNet outperforms comparable state-of-the-art VMC ansatzes for atoms, diatomic molecules and a strongly-correlated hydrogen chain by a margin and is yet computationally efficient. We anticipate that thanks to the favourable scaling with system size, this method may become a new leading method for highly accurate electronic-strucutre calculations on medium-sized molecular systems."}, "keywords": ["ab-initio orbitals"], "citation_intent": "result"}
{"citing_id": "2303.02918v2", "cited_id": "1912.01703", "section_title": "A.1. Experimental Settings", "citation": "Our code is implemented using PyTorch #REFR and PyTorch-Geometric , and all our experiments are ran on Nvidia RTX3090 GPUs with 24GB of memory.", "text_before_citation": [], "text_after_citation": ["Natural baselines.", "RFP is compared with a number of popular and recent methods.", "In particular, we focus on the comparison of RNF (Abboud et al., 2020) , Laplacian PE #OTHEREFR and PowerEmbed #OTHEREFR which is arguably the closest approach to RFP.", "Additionally, we introduce further baselines based on our RFP method, such as including both the RFP and eigenvectors of the propagation operators without a complete RFP trajectory in order to demonstrate the importance of considering the complete trajectory.", "For the baseline case of Laplacian eigenvectors only (denoted byL EIGVECS \u2020 ) we use the eigenvectors corresponding to the smallest eigenvalues, in order to be consistent with the literature #OTHEREFR ."], "citing_paper_content": {"title": "Graph Positional Encoding Via Random Feature Propagation", "abstract": "Two main families of node feature augmentation schemes have been explored for enhancing GNNs: random features and spectral positional encoding. Surprisingly, however, there is still no clear understanding of the relation between these two augmentation schemes. Here we propose a novel family of positional encoding schemes which draws a link between the above two approaches and improves over both. The new approach, named Random Feature Propagation (RFP), is inspired by the power iteration method and its generalizations. It concatenates several intermediate steps of an iterative algorithm for computing the dominant eigenvectors of a propagation matrix, starting from random node features. Notably, these propagation steps are based on graph-dependent propagation operators that can be either predefined or learned. We explore the theoretical and empirical benefits of RFP. First, we provide theoretical justifications for using random features, for incorporating early propagation steps, and for using multiple random initializations. Then, we empirically demonstrate that RFP significantly outperforms both spectral PE and random features in multiple node classification and graph classification benchmarks."}, "cited_paper_content": {"title": "Pytorch: An Imperative Style, High-Performance Deep Learning Library", "abstract": "Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. ::: In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. ::: We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks."}, "keywords": ["PyTorch"], "citation_intent": "method"}
{"citing_id": "2303.02691v1", "cited_id": "2002.07530", "section_title": "Algorithm And Regret Guarantee", "citation": "At the same time, in near-stationary environments (P T is small enough), our result can recover to the performance of LogUCB1 algorithm #REFR .", "text_before_citation": ["Theorem 3.", "For all \u03b3 \u2208 (1/T, 1), \u03bb = d log(T )/c \u00b5 , the dynamic regret of SCB-WeightUCB is bounded with probability at least 1 \u2212 1/T , by", "R T \u2264 O k 2 \u00b5 \u221a c \u00b5 1 (1 \u2212 \u03b3) 3 /2 P T + k \u00b5 \u221a c \u00b5 d(1 \u2212 \u03b3) 1 /2 T .", "By setting \u03b3 = 1 \u2212 max{1/T, k \u00b5 P T /(dT )}, we achieve R T \u2264 \uf8f1 \uf8f4 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f4 \uf8f3 O k 5/4 \u00b5 c 1/2 \u00b5 d 3 /4 P 1 /4 T T 3 /4 when P T \u2265 d k\u00b5T , O k\u00b5 c 1/2 \u00b5 d \u221a T when 0 \u2264 P T < d k\u00b5T .", "Compared to GLB, we improve the order of c \u00b5 from c \u22121 \u00b5 to c \u2212 1 /2 \u00b5 by exploiting the self-concordant properties."], "text_after_citation": ["The proof of Theorem 3 is presented in Appendix C.2.", "In addition, for the piecewise-stationary SCB, we propose SCB-PW-WeightUCB algorithm that gets rid of influence of c \u00b5 and thus directly improves upon .", "Theorem 4.", "For all \u03b3 \u2208 (1/2, 1), D = log(T )/ log(1/\u03b3) and \u03bb = d log(T )/c \u00b5 , the regret of SCB-PW-WeightUCB is bounded with probability at least 1 \u2212 1/T , by", "R T \u2264 O 1 1 \u2212 \u03b3 \u0393 T + 1 \u221a 1 \u2212 \u03b3 + d (1 \u2212 \u03b3)T ."], "citing_paper_content": {"title": "Revisiting Weighted Strategy For Non-Stationary Parametric Bandits", "abstract": "Non-stationary parametric bandits have attracted much attention recently. There are three principled ways to deal with non-stationarity, including sliding-window, weighted, and restart strategies. As many non-stationary environments exhibit gradual drifting patterns, the weighted strategy is commonly adopted in real-world applications. However, previous theoretical studies show that its analysis is more involved and the algorithms are either computationally less efficient or statistically suboptimal. This paper revisits the weighted strategy for non-stationary parametric bandits. In linear bandits (LB), we discover that this undesirable feature is due to an inadequate regret analysis, which results in an overly complex algorithm design. We propose a refined analysis framework, which simplifies the derivation and importantly produces a simpler weight-based algorithm that is as efficient as window/restart-based algorithms while retaining the same regret as previous studies. Furthermore, our new framework can be used to improve regret bounds of other parametric bandits, including Generalized Linear Bandits (GLB) and Self-Concordant Bandits (SCB). For example, we develop a simple weighted GLB algorithm with an O(k 5 /4 \u00b5 c \u2212 3 /4 \u00b5 d 3 /4 P 1 /4 T T 3 /4) regret, improving the O(k 2 \u00b5 c \u22121 \u00b5 d 9 /10 P 1 /5 T T 4 /5) bound in prior work, where k \u00b5 and c \u00b5 characterize the reward model's nonlinearity, P T measures the non-stationarity, d and T denote the dimension and time horizon."}, "cited_paper_content": {"title": "Improved Optimistic Algorithms For Logistic Bandits", "abstract": "The generalized linear bandit framework has attracted a lot of attention in recent years by extending the well-understood linear setting and allowing to model richer reward structures. It notably covers the logistic model, widely used when rewards are binary. For logistic bandits, the frequentist regret guarantees of existing algorithms are $\\tilde{\\mathcal{O}}(\\kappa \\sqrt{T})$, where $\\kappa$ is a problem-dependent constant. Unfortunately, $\\kappa$ can be arbitrarily large as it scales exponentially with the size of the decision set. This may lead to significantly loose regret bounds and poor empirical performance. In this work, we study the logistic bandit with a focus on the prohibitive dependencies introduced by $\\kappa$. We propose a new optimistic algorithm based on a finer examination of the non-linearities of the reward function. We show that it enjoys a $\\tilde{\\mathcal{O}}(\\sqrt{T})$ regret with no dependency in $\\kappa$, but for a second order term. Our analysis is based on a new tail-inequality for self-normalized martingales, of independent interest."}, "keywords": ["LogUCB1 algorithm"], "citation_intent": "result"}
{"citing_id": "2303.17835v1", "cited_id": "1201.0490", "section_title": "Support Vector Classifier", "citation": "The linear kernel SVC implementation LinearSVC from Scikit-learn library #REFR was used to conduct the experiment.", "text_before_citation": ["The experiments were repeated with the SVC model to the same two datasets."], "text_after_citation": ["Linear kernel SVC was chosen due to large dataset size.", "Other kernel types were tested, however they did not scale to the large number of samples.", "The samples were normalized using the Scikit-learn StandardScaler to ease the model convergence. Table 2 displays the results from the experiments.", "The proposed method is clearly superior to the conventional method in both experiments.", "The performance in the statistical change dataset is considerably worse when compared to the shift change dataset."], "citing_paper_content": {"title": "Improved Difference Images For Change Detection Classifiers In Sar Imagery Using Deep Learning", "abstract": "Satellite-based Synthetic Aperture Radar (SAR) images can be used as a source of remote sensed imagery regardless of cloud cover and day-night cycle. However, the speckle noise and varying image acquisition conditions pose a challenge for change detection classifiers. This paper proposes a new method of improving SAR image processing to produce higher quality difference images for the classification algorithms. The method is built on a neural network-based mapping transformation function that produces artificial SAR images from a location in the requested acquisition conditions. The inputs for the model are: previous SAR images from the location, imaging angle information from the SAR images, digital elevation model, and weather conditions. The method was tested with data from a location in NorthEast Finland by using Sentinel-1 SAR images from European Space Agency, weather data from Finnish Meteorological Institute, and a digital elevation model from National Land Survey of Finland. In order to verify the method, changes to the SAR images were simulated, and the performance of the proposed method was measured using experimentation where it gave substantial improvements to performance when compared to a more conventional method of creating difference images."}, "cited_paper_content": {"title": "Scikit-Learn: Machine Learning In Python", "abstract": "Scikit-learn is a Python module integrating a wide range of state-of-the-art machine learning algorithms for medium-scale supervised and unsupervised problems. This package focuses on bringing machine learning to non-specialists using a general-purpose high-level language. Emphasis is put on ease of use, performance, documentation, and API consistency. It has minimal dependencies and is distributed under the simplified BSD license, encouraging its use in both academic and commercial settings. Source code, binaries, and documentation can be downloaded from http://scikit-learn.sourceforge.net."}, "keywords": ["Scikit-learn library"], "citation_intent": "method"}
{"citing_id": "2303.00298v1", "cited_id": "1812.01601", "section_title": "Related Work", "citation": "HMMR #REFR learns the human dynamics to predict pose and shape for past and future frames.", "text_before_citation": ["Regression-based scheme has recently received extensive research #OTHEREFR Lin et al., 2021b,a) , due to its directness and effectiveness.", "HMR #OTHEREFR , is the representative regression-based methods, using an image encoder and regressor to predict the pose, shape and camera parameters.", "To train the model well and make sure the realistic of the pose and shape, the reprojection loss and adversarial loss are introduced to leverage unpaired 2D-to-3D supervision.", "In addition, several non-parameteric mesh regression methods are proposed to directly regress the mesh vertices coordinates, including Pose2Mesh #OTHEREFR , Convolution Mesh Regression #OTHEREFR , I2L-MeshNet , and the Transformer-based METRO #OTHEREFR and Mesh Graphormer #OTHEREFR .", "Beyond estimating pose and shape from a singe image, video-based methods consider to fully dig the temporal motion information hidden in video data to improve the accuracy and robustness."], "text_after_citation": ["VIBE #OTHEREFR encodes temporal feature using a GRU and adopts an adversarial learning framework to learn kinematically plausible motion from a large-scale motion capture dataset.", "TCMR #OTHEREFR introduces PoseForecast to forecast additional temporal features from past and future frames without a current frame.", "MAED #OTHEREFR proposes to use a spatial-temporal encoder to learn temporally enhanced image features and regress the joint rotations following a defined kinematic topology.", "In contrast to these methods, our goal is to encode joint-level features, shape and camera information separately, rather than encoding all the information into a unified image feature vector.", "Since we use independent tokens to encode the rotational information of each joint, we can model the inner temporal patterns when each joint rotates over time."], "citing_paper_content": {"title": "Capturing The Motion Of Every Joint: 3D Hu-Man Pose And Shape Estimation With Indepen-Dent Tokens", "abstract": "In this paper we present a novel method to estimate 3D human pose and shape from monocular videos. This task requires directly recovering pixel-alignment 3D human pose and body shape from monocular images or videos, which is challenging due to its inherent ambiguity. To improve precision, existing methods highly rely on the initialized mean pose and shape as prior estimates and parameter regression with an iterative error feedback manner. In addition, video-based approaches model the overall change over the image-level features to temporally enhance the single-frame feature, but fail to capture the rotational motion at the joint level, and cannot guarantee local temporal consistency. To address these issues, we propose a novel Transformer-based model with a design of independent tokens. First, we introduce three types of tokens independent of the image feature: joint rotation tokens, shape token, and camera token. By progressively interacting with image features through Transformer layers, these tokens learn to encode the prior knowledge of human 3D joint rotations, body shape, and position information from large-scale data, and are updated to estimate SMPL parameters conditioned on a given image. Second, benefiting from the proposed token-based representation, we further use a temporal model to focus on capturing the rotational temporal information of each joint, which is empirically conducive to preventing large jitters in local parts. Despite being conceptually simple, the proposed method attains superior performances on the 3DPW and Human3.6M datasets. Using ResNet-50 and Transformer architectures, it obtains 42.0 mm error on the PA-MPJPE metric of the challenging 3DPW, outperforming state-of-the-art counterparts by a large margin. Code will be publicly available 1 ."}, "cited_paper_content": {"title": "Learning 3D Human Dynamics From Video", "abstract": "From an image of a person in action, we can easily guess the 3D motion of the person in the immediate past and future. This is because we have a mental model of 3D human dynamics that we have acquired from observing visual sequences of humans in motion. We present a framework that can similarly learn a representation of 3D dynamics of humans from video via a simple but effective temporal encoding of image features. At test time, from video, the learned temporal representation give rise to smooth 3D mesh predictions. From a single image, our model can recover the current 3D mesh as well as its 3D past and future motion. Our approach is designed so it can learn from videos with 2D pose annotations in a semi-supervised manner. Though annotated data is always limited, there are millions of videos uploaded daily on the Internet. In this work, we harvest this Internet-scale source of unlabeled data by training our model on unlabeled video with pseudo-ground truth 2D pose obtained from an off-the-shelf 2D pose detector. Our experiments show that adding more videos with pseudo-ground truth 2D pose monotonically improves 3D prediction performance. We evaluate our model on the recent challenging dataset of 3D Poses in the Wild and obtain state-of-the-art performance on the 3D prediction task without any fine-tuning. The project website with video can be found at https://akanazawa.github.io/human_dynamics/."}, "keywords": ["pose", "human dynamics"], "citation_intent": "background"}
{"citing_id": "2303.07228v2", "cited_id": "1701.03081", "section_title": "D. Extending The Method To The Two-Way Distillable Entanglement", "citation": "However,\u00ca npt rev (\u2022) is efficiently computable by SDP and tightens the upper bound of the example states illustrated in #REFR , whose details can be found in Appendix C.", "text_before_citation": ["It also follows an efficiently computable relaxation as\u00ca npt", "rev (\u03c1 AB ) = [1 \u2212 2 \u2212R max,PPT (\u03c1 AB ) ] i \u03bb i S(B) \u03c8 i , where \u03c9 AB = i \u03bb i |\u03c8 i \u03c8 i |", "is the spectral decomposition of the PPT-squeezed state \u03c9 AB of \u03c1 AB .", "In fact,\u00ca npt rev (\u2022) can be interpreted as an easily computable version of the bound E MP (\u2022) in #OTHEREFR , utilizing the convexity of E D,\u2194 (\u03c1 AB ) on the convex decomposition of \u03c1 AB into MC states and PPT states.", "Since the set of all MC states is not convex, tracking all possible decompositions to compute E MP (\u2022) is generally hard."], "text_after_citation": ["We note that R(\u03c1 AB ) \u2264 E MP (\u03c1 AB ) \u2264 E npt rev (\u03c1 AB ) where R(\u2022) is the Rains bound for the two-way distillable entanglement.", "Nevertheless,\u00ca npt rev (\u2022) connects the reverse max-relative entropy of NPT entanglement with the entanglement of formation, and we believe such connection would shed light on the study of other quantum resource theories as well."], "citing_paper_content": {"title": "Estimate Distillable Entanglement And Quantum Capacity By Squeezing Useless Entanglement", "abstract": "Entanglement distillation is crucial in quantum information processing. But it remains challenging to estimate the distillable entanglement and its closely related essential quantity, the quantum capacity of a noisy quantum channel. In this work, we propose methods for evaluating both quantities by squeezing out useless entanglement within a state or a quantum channel, whose contributions are expected to be ignored for the distillable entanglement or the quantum capacity, respectively. We first consider a general resource measure called the reverse divergence of resources to quantify the minimum divergence between a target state and the set of free states. We then introduce the reverse max-relative entropy of entanglement and apply it to establish efficiently computable upper bounds on the distillable entanglement. We also extend the reverse divergence of resources to quantum channels and derive upper bounds on the quantum capacity. We further apply our method to investigate purifying the maximally entangled states under practical noises, such as depolarizing and amplitude damping noises, and notably establish improvements in estimating the one-way distillable entanglement. Our bounds also offer useful benchmarks for evaluating the quantum capacities of qubit quantum channels of interest, including the Pauli channels and the random mixed unitary channels. CONTENTS I. Introduction A. Background B. Main contributions II. Reverse divergence of resources A. Preliminaries B. Reverse divergence of resources III. Applications on distillable entanglement A. Upper bound on the one-way distillable entanglement B. Continuity bounds of the one-way distillable entanglement C. Examples of less-entangled states D. Extending the method to the two-way distillable entanglement IV. Applications on quantum channel capacity A. Quantum capacity of qubit channels V. Concluding remarks Acknowledgements."}, "cited_paper_content": {"title": "Useful States And Entanglement Distillation", "abstract": "We derive general upper bounds on the distillable entanglement of a mixed state under one-way and two-way local operations and classical communication (LOCC). In both cases, the upper bound is based on a convex decomposition of the state into \u201cuseful\u201d and \u201cuseless\u201d quantum states. By \u201cuseful,\u201d we mean a state whose distillable entanglement is non-negative and equal to its coherent information (and thus given by a single-letter, tractable formula). On the other hand, \u201cuseless\u201d states are undistillable, i.e., their distillable entanglement is zero. We prove that in both settings, the distillable entanglement is convex on such decompositions. Hence, an upper bound on the distillable entanglement is obtained from the contributions of the useful states alone, being equal to the convex combination of their coherent informations. Optimizing over all such decompositions of the input state yields our upper bound. The useful and useless states are given by degradable and antidegradable states in the one-way LOCC setting, and by maximally correlated and positive partial transpose (PPT) states in the two-way LOCC setting, respectively. We also illustrate how our method can be extended to quantum channels. Interpreting our upper bound as a convex roof extension, we show that it reduces to a particularly simple, non-convex optimization problem for the classes of isotropic states and Werner states. In the one-way LOCC setting, this non-convex optimization yields an upper bound on the quantum capacity of the qubit depolarizing channel that is strictly tighter than previously known bounds for large values of the depolarizing parameter. In the two-way LOCC setting, the non-convex optimization achieves the PPT-relative entropy of entanglement for both isotropic and Werner states."}, "keywords": ["example states"], "citation_intent": "background"}
{"citing_id": "2304.13633v1", "cited_id": "1807.03748", "section_title": "Introduction", "citation": "InfoNCE #REFR (Information Noise-Contrastive Estimation) is based on contrastive learning, which uses noise-contrastive estimation (Noise-Contrastive Estimation, NCE) as a bound to approximate mutual information, and uses neural networks to parameterize critic in NCE, so that it can be used more flexibly.", "text_before_citation": ["EQUATION", "Where is a multidimensional variable with ( > 2) dimensions, and ( ) obtains the total correlation among all dimensions of .", "Since the exact computation of mutual information and total correlation is only tractable for discrete variables, or for a limited family of problems where the probability distributions are known #OTHEREFR .", "Therefore, many methods have been proposed to estimate MI based on theoretical upper or lower bounds.", "For instance, MINE #OTHEREFR maximizes an implicit function lower bound obtained by Jensen's inequality to estimate the mutual information between two random variables."], "text_after_citation": ["The CLUB #OTHEREFR algorithm employs unsupervised learning methods of logarithmic comparison and log-linear models to avoid overfitting and improve the accuracy of estimates.", "CLUB uses a technique called Contrastive Log-ratio Upper Bound to calculate the confidence interval and upper bound of the estimated TC value.", "This technique is based on the distribution of log comparison values and some statistical assumptions, which can effectively control the error and confidence of the estimate.", "The same issue of an intractable computation arises when estimating the TC value of multidimensional variables without knowledge of the ( ) distribution.", "In representation learning, a straightforward prior distribution assumption of ( ) is frequently made to handle this issue #OTHEREFR ."], "citing_paper_content": {"title": "Understanding The Limitation Of Total Correlation Estimation Based On Mutual Information Bounds", "abstract": "The total correlation(TC) is a crucial index to measure the correlation between marginal distribution in multidimensional random variables, and it is frequently applied as an inductive bias in representation learning. Previous research has shown that the TC value can be estimated using mutual information boundaries through decomposition. However, we found through theoretical derivation and qualitative experiments that due to the use of importance sampling in the decomposition process, the bias of TC value estimated based on MI bounds will be amplified when the proposal distribution in the sampling differs significantly from the target distribution. To reduce estimation bias issues, we propose a TC estimation correction model based on supervised learning, which uses the training iteration loss sequence of the TC estimator based on MI bounds as input features to output the true TC value. Experiments show that our proposed method can improve the accuracy of TC estimation and eliminate the variance generated by the TC estimation process."}, "cited_paper_content": {"title": "Representation Learning With Contrastive Predictive Coding", "abstract": "While supervised learning has enabled great progress in many applications, unsupervised learning has not seen such widespread adoption, and remains an important and challenging endeavor for artificial intelligence. In this work, we propose a universal unsupervised learning approach to extract useful representations from high-dimensional data, which we call Contrastive Predictive Coding. The key insight of our model is to learn such representations by predicting the future in latent space by using powerful autoregressive models. We use a probabilistic contrastive loss which induces the latent space to capture information that is maximally useful to predict future samples. It also makes the model tractable by using negative sampling. While most prior work has focused on evaluating representations for a particular modality, we demonstrate that our approach is able to learn useful representations achieving strong performance on four distinct domains: speech, images, text and reinforcement learning in 3D environments."}, "keywords": ["mutual information", "contrastive learning"], "citation_intent": "method"}
{"citing_id": "2303.00091v1", "cited_id": "1608.03983", "section_title": "Details Of Model Training", "citation": "The VLP model was pre-trained using the Adam optimizer with decoupled weight decay (AdamW), along with a warmup cosine annealing scheduler #REFR ).", "text_before_citation": [], "text_after_citation": ["The initial learning rate for the warm-up phase was set to 1e\u22125, which was gradually increased to a maximum learning rate of 1e \u2212 4 before being decayed using the cosine annealing scheduler.", "The visual encoder was initialized with self-supervised weights from ImageNet, while the text encoder was trained from scratch.", "The model was trained for 15 epochs with a warm-up epoch of 5, and a batch size of 12.", "All VLP model training was performed on a GeForce RTX 3090.", "For fine-tuning, we used the same optimizer and scheduler, with a learning rate of 1e\u22124 and a warm-up learning rate of 1e\u2212 5."], "citing_paper_content": {"title": "Improving Medical Speech-To-Text Accuracy With Vision-Language Pre-Training Model", "abstract": "Automatic Speech Recognition (ASR) is a technology that converts spoken words into text, facilitating interaction between humans and machines. One of the most common applications of ASR is Speech-To-Text (STT) technology, which simplifies user workflows by transcribing spoken words into text. In the medical field, STT has the potential to significantly reduce the workload of clinicians who rely on typists to transcribe their voice recordings. However, developing an STT model for the medical domain is challenging due to the lack of sufficient speech and text datasets. To address this issue, we propose a medical-domain text correction method that modifies the output text of a general STT system using the Vision Language Pre-training (VLP) method. VLP combines textual and visual information to correct text based on image knowledge. Our extensive experiments demonstrate that the proposed method offers quantitatively and clinically significant improvements in STT performance in the medical field. We further show that multi-modal understanding of image and text information outperforms single-modal understanding using only text information."}, "cited_paper_content": {"title": "Sgdr: Stochastic Gradient Descent With Warm Restarts", "abstract": "Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets, where we demonstrate new state-of-the-art results at 3.14% and 16.21%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at https://github.com/loshchil/SGDR"}, "keywords": ["VLP model", "Adam optimizer"], "citation_intent": "method"}
{"citing_id": "2303.11866v1", "cited_id": "1810.04805", "section_title": "Conclusion & Future Work", "citation": "For the text encoder, we choose between bert-base-uncased #REFR and SimCSE Gao et al. (2021) .", "text_before_citation": ["Our approach presents one simple but effective way to use that knowledge.", "We follow the standard training procedure ( \u00a72.4) and train a CLIP-base model where both of the encoders are initialized randomly, instead of using weights initialized from unimodally pretrained models (DeIT Touvron et al. (2021) and SimCSE Gao et al. (2021) ). We train three models, one for each dataset size. The results can be seen in Fig 6.", "Compared to the randomly initialized model, the pretrained model is substantially better across all three datasets and all 3 model sizes.", "However, it is likely that the benefit of unimodal pretraining will be diminished as the number of training pairs available for multimodal vision-language pretraining increases, although we do not explore this.", "We train LilT-base models with encoders initialized from different kinds of pretraining methods."], "text_after_citation": ["For the image encoder, we choose between DeiTTouvron et al. 2021"], "citing_paper_content": {"title": "Contrastive Alignment Of Vision To Language Through Parameter-Efficient Transfer Learn-Ing", "abstract": "Contrastive vision-language models (e.g. CLIP) are typically created by updating all the parameters of a vision model and language model through contrastive training. Can such models be created by a small number of parameter updates to an already-trained language model and vision model? The literature describes techniques that can create vision-language models by updating a small number of parameters in a language model, but these require already aligned visual representations and are non-contrastive, hence unusable for latency-sensitive applications such as neural search. We explore the feasibility and benefits of parameter-efficient contrastive vision-language alignment through transfer learning: creating a model such as CLIP by minimally updating an already-trained vision and language model. We find that a minimal set of parameter updates (<7%) can achieve the same performance as full-model training, and updating specific components (<1% of parameters) can match 75% of full-model training. We describe a series of experiments: we show that existing knowledge is conserved more strongly in parameter-efficient training and that parameter-efficient scaling scales with model and dataset size. Where paired-image text data is scarce but strong multilingual language models exist (e.g. low resource languages), parameter-efficient training is even preferable to full-model training. Given a fixed compute budget, parameter-efficient training allows training larger models on the same hardware, achieving equivalent performance in less time. Parameter-efficient training hence constitutes an energyefficient and effective training strategy for contrastive vision-language models that may be preferable to the full-model training paradigm for common use cases. Code and weights at https://github.com/codezakh/LilT."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["text encoder"], "citation_intent": "method"}
{"citing_id": "2303.08694v1", "cited_id": "1204.3476", "section_title": "Introduction", "citation": "In the last decades multilevel Monte Carlo (MLMC) methods have been applied to a plethora of problems in stochastic modelling and uncertainty quantification (see e.g. [5, 11, 14, 19, #REFR ).", "text_before_citation": [], "text_after_citation": ["The method relies on a hierarchy of approximations arranged as a telescoping sum, resulting in a variance reduction.", "In fact, in certain situations the multilevel estimator has asymptotically the same computational complexity as one solve of the deterministic problem on the finest discretization of the hierarchy.", "Having a preset finest discretization level the multilevel estimator is a biased estimator.", "The continuous level Monte Carlo (CLMC) estimator circumvents this issue by defining the estimator as a stochastic process (see [12] ).", "The level of refinement is here given by an exponentially distributed random variable, which in turn means that the computational complexity of a CLMC estimator relies heavily on the concrete samples of this random variable in the simulation."], "citing_paper_content": {"title": "Quasi Continuous Level Monte Carlo For Random Elliptic Pdes", "abstract": "This paper provides a framework in which multilevel Monte Carlo and continuous level Monte Carlo can be compared. In continuous level Monte Carlo the level of refinement is determined by an exponentially distributed random variable, which therefore heavily influences the computational complexity. We propose in this paper a variant of the algorithm, where the exponentially distributed random variable is generated by a quasi Monte Carlo sequence, resulting in a significant variance reduction. In the examples presented the quasi continuous level Monte Carlo algorithm outperforms multilevel and continuous level Monte Carlo by a clear margin."}, "cited_paper_content": {"title": "Further Analysis Of Multilevel Monte Carlo Methods For Elliptic Pdes With Random Coefficients", "abstract": "We consider the application of multilevel Monte Carlo methods to elliptic PDEs with random coefficients. We focus on models of the random coefficient that lack uniform ellipticity and boundedness with respect to the random parameter, and that only have limited spatial regularity. We extend the finite element error analysis for this type of equation, carried out recently by Charrier, Scheichl and Teckentrup, to more difficult problems, posed on non--smooth domains and with discontinuities in the coefficient. For this wider class of model problem, we prove convergence of the multilevel Monte Carlo algorithm for estimating any bounded, linear functional and any continuously Fr\\'echet differentiable non--linear functional of the solution. We further improve the performance of the multilevel estimator by introducing level dependent truncations of the Karhunen--Lo\\`eve expansion of the random coefficient. Numerical results complete the paper."}, "keywords": ["Monte Carlo (MLMC)"], "citation_intent": "method"}
{"citing_id": "2303.16761v1", "cited_id": "1806.00525", "section_title": "Arxiv:2303.16761V1 [Cs.Ir] 23 Mar 2023", "citation": "To validate the effectiveness of our approach, we conduct dialogue-to-video experiments on a benchmark dataset AVSD #REFR .", "text_before_citation": ["Such discussion contains subtle information about the video of interest and thus cannot be treated as a plain-text query.", "Therefore, to incorporate the conversational information from dialogues, we propose a novel dialogue-to-video retrieval approach.", "In our proposed model, we sequentially encode each turn of the dialogue to obtain a dialogue-aware query representation with the purpose of retaining the dialogue information.", "Then we calculate the similarity between this dialogue-aware query representation and individual frames in the video in order to obtain a weighted video representation.", "Finally, we use the video representation to compute an overall similarity score with the dialogue-aware query."], "text_after_citation": ["Experimental results show that our approach achieves significant improvements over previous state-of-the-art models including FiT and ViReD #OTHEREFR .", "In this section, we describe how our dialogue-to-video retrieval system works.", "Our retrieval system consists of two major components: 1) a temporal-aware video encoder responsible for encoding the image frames in video with temporal information.", "2) a dialogue-query encoder responsible for encoding the dialogue query with conversational information.", "As shown in Figure 1 , our model receives video-query pairs and produces similarity scores."], "citing_paper_content": {"title": "Dialogue-To-Video Retrieval", "abstract": "Recent years have witnessed an increasing amount of dialogue/conversation on the web especially on social media. That inspires the development of dialogue-based retrieval, in which retrieving videos based on dialogue is of increasing interest for recommendation systems. Different from other video retrieval tasks, dialogue-to-video retrieval uses structured queries in the form of user-generated dialogue as the search descriptor. We present a novel dialogue-to-video retrieval system, incorporating structured conversational information. Experiments conducted on the AVSD dataset show that our proposed approach using plain-text queries improves over the previous counterpart model by 15.8% on R@1. Furthermore, our approach using dialogue as a query, improves retrieval performance by 4."}, "cited_paper_content": {"title": "Audio Visual Scene-Aware Dialog (Avsd) Challenge At Dstc7", "abstract": "Scene-aware dialog systems will be able to have conversations with users about the objects and events around them. Progress on such systems can be made by integrating state-of-the-art technologies from multiple research areas including end-to-end dialog systems visual dialog, and video description. We introduce the Audio Visual Scene Aware Dialog (AVSD) challenge and dataset. In this challenge, which is one track of the 7th Dialog System Technology Challenges (DSTC7) workshop1, the task is to build a system that generates responses in a dialog about an input video"}, "keywords": ["video", "AVSD"], "citation_intent": "method"}
{"citing_id": "2304.13085v2", "cited_id": "1412.6980", "section_title": "Implementation Details", "citation": "The Adaptive Moment Estimation (Adam) #REFR is used as our optimizer with a learning rate of 0.0001 and a batch size of 32.", "text_before_citation": ["We utilize the RawNet2 [4] model as the backbone for feature learning."], "text_after_citation": ["The loss weight \u03bb is set as 0.5 in the experiment.", "To report the detection performance, we calculate the Equal Error Rate (EER) following previous studies #OTHEREFR 29, #OTHEREFR . RawNet2 #OTHEREFR .", "The RawNet2 model is based on DNN speaker embedding extraction with the raw waveform as inputs.", "This powerful model uses a technique named feature map scaling which scales feature maps similar to squeezeexcitation.", "It performed the best in the ASVspoof 2021 Speech Deepfake track."], "citing_paper_content": {"title": "Ai-Synthesized Voice Detection Using Neural Vocoder Artifacts", "abstract": "Advancements in AI-synthesized human voices have created a growing threat of impersonation and disinformation, making it crucial to develop methods to detect synthetic human voices. This study proposes a new approach to identifying synthetic human voices by detecting artifacts of vocoders in audio signals. Most DeepFake audio synthesis models use a neural vocoder, a neural network that generates waveforms from temporal-frequency representations like mel-spectrograms. By identifying neural vocoder processing in audio, we can determine if a sample is synthesized. To detect synthetic human voices, we introduce a multi-task learning framework for a binaryclass RawNet2 model that shares the feature extractor with a vocoder identification module. By treating vocoder identification as a pretext task, we constrain the feature extractor to focus on vocoder artifacts and provide discriminative features for the final binary classifier. Our experiments show that the improved RawNet2 model based on vocoder identification achieves high classification performance on the binary task overall. Codes and data can be found at https://github.com/csun22/Synthetic-Voice-Detection-Vocoder-Artifacts."}, "cited_paper_content": {"title": "Adam: A Method For Stochastic Optimization", "abstract": "We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm."}, "keywords": ["learning rate", "Adaptive Moment Estimation"], "citation_intent": "method"}
{"citing_id": "2304.07971v1", "cited_id": "1808.10523", "section_title": "Related Works 6.1 Collaborative Filtering", "citation": "In recent years, this type of global relationships has gradually received more attention and has been incorporated in CF models in the form of interaction graph #REFR .", "text_before_citation": ["Since CF can be considered as a task to complete entries in the user-item interaction matrix, Matrix Factorization (MF) #OTHEREFR , as a strategy for matrix completion, naturally becomes the foundation of the mainstream approach in CF.", "MF assumes that the user-item interaction matrix is low-rank and can be recovered by learning the embedding vectors of users and items.", "Most MF models generate predicted entries by the dot product of user and item embedding vectors, while they can be optimized by minimizing the error of individual entries #OTHEREFR or maximizing the difference between positive and negative samples #OTHEREFR .", "They are both widely adopted in the subsequent proposed methods which introduce refined structures like neural networks #OTHEREFR .", "These approaches enable a light design by focusing on the modeling of single user-item entry, while neglecting the synergy between different interactions."], "text_after_citation": ["With the emerging of Graph Convolutional Networks (GCN), GCN models #OTHEREFR quickly become the state-of-the-art in MF-based models and are continuously improved to achieve advances in efficiency and accuracy #OTHEREFR .", "Unlike MF-based models, another class of method implements CF by treating the user's historical interactions as features, and modeling the relationships between items #OTHEREFR .", "A classical approach is the linear autoencoder #OTHEREFR , which models an item-item relationship matrix to encode user features.", "This idea is then extended by subsequent studies and applied to nonlinear denoising autoencoders #OTHEREFR and variational autoencoders #OTHEREFR .", "A recent work #OTHEREFR considers CF in terms of graph signal processing and proposes a framework for signal-based models, which can incorporate the linear autoencoders and the ideal case of MF and GCN models."], "citing_paper_content": {"title": "Collaborative Residual Metric Learning", "abstract": "In collaborative filtering, distance metric learning has been applied to matrix factorization techniques with promising results. However, matrix factorization lacks the ability of capturing collaborative information, which has been remarked by recent works and improved by interpreting user interactions as signals. This paper aims to find out how metric learning connect to these signal-based models. By adopting a generalized distance metric, we discovered that in signal-based models, it is easier to estimate the residual of distances, which refers to the difference between the distances from a user to a target item and another item, rather than estimating the distances themselves. Further analysis also uncovers a link between the normalization strength of interaction signals and the novelty of recommendation, which has been overlooked by existing studies. Based on the above findings, we propose a novel model to learn a generalized distance user-item distance metric to capture user preference in interaction signals by modeling the residuals of distance. The proposed CoRML model is then further improved in training efficiency by a newly introduced approximated ranking weight. Extensive experiments conducted on 4 public datasets demonstrate the superior performance of CoRML compared to the state-of-the-art baselines in collaborative filtering, along with high efficiency and the ability of providing novelty-promoted recommendations, shedding new light on the study of metric learning-based recommender systems. CCS CONCEPTS \u2022 Information systems \u2192 Recommender systems; Collaborative filtering."}, "cited_paper_content": {"title": "Spectral Collaborative Filtering", "abstract": "Despite the popularity of Collaborative Filtering (CF), CF-based methods are haunted by the cold-start problem, which has a significantly negative impact on users' experiences with Recommender Systems (RS). In this paper, to overcome the aforementioned drawback, we first formulate the relationships between users and items as a bipartite graph. Then, we propose a new spectral convolution operation directly performing in the spectral domain, where not only the proximity information of a graph but also the connectivity information hidden in the graph are revealed. With the proposed spectral convolution operation, we build a deep recommendation model called Spectral Collaborative Filtering (SpectralCF). Benefiting from the rich information of connectivity existing in the spectral domain, SpectralCF is capable of discovering deep connections between users and items and therefore, alleviates the cold-start problem for CF. To the best of our knowledge, SpectralCF is the first CF-based method directly learning from the spectral domains of user-item bipartite graphs. We apply our method on several standard datasets. It is shown that SpectralCF significantly out-performs state-of-the-art models. Code and data are available at https://github.com/lzheng21/SpectralCF."}, "keywords": ["interaction graph"], "citation_intent": "background"}
{"citing_id": "2303.02995v1", "cited_id": "1906.02890", "section_title": "Introduction", "citation": "To be specific, for modeling hierarchies in natural language, we share similar intuitions with the previous studies on unsupervised grammar induction, which aim at unsupervised hierarchical mining #REFR Drozdov et al., 2019) .", "text_before_citation": ["To this end, we introduce hierarchy-aware attention into CLIP, denoted as HiCLIP.", "Hierarchy-aware attention applies an attention mask to the conventional attention mechanism to indicate the tendency to merge certain vision patches and language tokens into groups because they are spatially and semantically or visually similar.", "We generalize hierarchy-aware attention to both images and texts, where its mask is obtained by first calculating the neighbouring affinity score among adjacent patches or tokens, and then propagating the scores across any given patch or token pairs.", "In addition, we formulate the affinity score with an increasing trend as the layer gets deeper to ensure the merged groups remains the same.", "In this way, we progressively aggregate hierarchies in a layer-by-layer manner for both images and texts."], "text_after_citation": ["Tree Transformer proposes a similar modified attention mechanism which is essentially a special case of hierarchy-aware attention, where the attention mask is instantiated as constituent prior to encourage the merging of semantically similar tokens.", "Capturing hierarchies in visual contents is more challenging, because spatial correlation should also be considered in addition to visual similarities.", "Therefore, we extend the hierarchy-aware attention to Vision Transformers #OTHEREFR by creating a Group Transformer to progressively aggregate image patches into semantic groups until all patches are merged in one common group which is the original image.", "Different from the 1D scenario in Tree Transformer, the neighboring affinity score is computed among the four adjacent neighbors of each image patch ( Fig. 1 (a) ).", "Afterwards, we propagate neighboring affinity scores by comparing two special paths connecting image patches on the 2D grid graph."], "citing_paper_content": {"title": "Hiclip: Contrastive Language-Image Pre-Training With Hierarchy-Aware Attention", "abstract": "The success of large-scale contrastive vision-language pretraining (CLIP) has benefited both visual recognition and multimodal content understanding. The concise design brings CLIP the advantage in inference efficiency against other visionlanguage models with heavier cross-attention fusion layers, making it a popular choice for a wide spectrum of downstream tasks. However, CLIP does not explicitly capture the hierarchical nature of high-level and fine-grained semantics conveyed in images and texts, which is arguably critical to vision-language understanding and reasoning. To this end, we equip both the visual and language branches in CLIP with hierarchy-aware attentions, namely Hierarchy-aware CLIP (HiCLIP), to progressively discover semantic hierarchies layer-by-layer from both images and texts in an unsupervised manner. As a result, such hierarchical aggregation significantly improves the cross-modal alignment. To demonstrate the advantages of HiCLIP, we conduct qualitative analysis on its unsupervised hierarchy induction during inference, as well as extensive quantitative experiments on both visual recognition and vision-language downstream tasks. 1 * This work was conducted while interning at ByteDance."}, "cited_paper_content": {"title": "Visually Grounded Neural Syntax Acquisition", "abstract": "We present the Visually Grounded Neural Syntax Learner (VG-NSL), an approach for learning syntactic representations and structures without any explicit supervision. The model learns by looking at natural images and reading paired captions. VG-NSL generates constituency parse trees of texts, recursively composes representations for constituents, and matches them with images. We define concreteness of constituents by their matching scores with images, and use it to guide the parsing of text. Experiments on the MSCOCO data set show that VG-NSL outperforms various unsupervised parsing approaches that do not use visual grounding, in terms of F1 scores against gold parse trees. We find that VGNSL is much more stable with respect to the choice of random initialization and the amount of training data. We also find that the concreteness acquired by VG-NSL correlates well with a similar measure defined by linguists. Finally, we also apply VG-NSL to multiple languages in the Multi30K data set, showing that our model consistently outperforms prior unsupervised approaches."}, "keywords": ["unsupervised grammar induction"], "citation_intent": "result"}
{"citing_id": "2304.11928v1", "cited_id": "1712.02560", "section_title": "1) Distribution Divergence", "citation": "The authors propose the sliced Wasserstein discrepancy, which considers the properties of the underlying geometry of probability space and thus im-proves upon the L1 norm used in #REFR . Further follow-up work is presented in Li et al.", "text_before_citation": ["In the third step, the feature extractor is then optimized to minimize the MCD for the target domain.", "This causes those samples from the target domain far away from the source domain distribution to move closer to the source domain distribution (here, the support of the source domain is given, and the two classifiers agree).", "The three steps are iterated, which results in an adversarial optimization.", "Lee et al.", "#OTHEREFR advance this work by introducing an improved way of computing the discrepancy between the classifier probability outputs."], "text_after_citation": ["#OTHEREFR where two classifiers are trained on the source domain while also updating the feature generator.", "Then they maximize the classifier discrepancy on the target domain while ensuring the source domain classification stays the same.", "In the final stage, they train the feature generator to minimize the classifier discrepancy, pushing the target domain data to the statistical support of the source domain.", "Other methods for distribution divergence minimization: MMD #OTHEREFR uses the soft paste algorithm, combining two images by a weighted overlay (see Section III-B).", "A reference source domain image is pasted into a target and source domain image."], "citing_paper_content": {"title": "Survey On Unsupervised Domain Adaptation For Semantic Segmentation For Visual Perception In Automated Driving", "abstract": "The research leading to these results is funded by the German Federal Ministry for Economic Affairs and Energy within the project \"KI Delta Learning\" (F\u00f6rderkennzeichen 19A19013C and 19A19013K). The authors would like to thank the consortium for the successful cooperation."}, "cited_paper_content": {"title": "Maximum Classifier Discrepancy For Unsupervised Domain Adaptation", "abstract": "In this work, we present a method for unsupervised domain adaptation (UDA), where we aim to transfer knowledge from a label-rich domain (i.e., a source domain) to an unlabeled domain (i.e., a target domain). Many adversarial learning methods have been proposed for this task. These methods train domain classifier networks (i.e., a discriminator) to discriminate distinguish the features as either a source or target and train a feature generator network to mimic the discriminator.However, the domain classifier only tries to distinguish the features as a source or target and thus does not consider task-specific decision boundaries between classes. Therefore, a trained generator can generate ambiguous features near class boundaries. ::: To solve the problem, we propose a new approach that attempts to align distributions of source and target by utilizing the task-specific decision boundaries. We propose to utilize task-specific classifiers as discriminators that try to detect target samples that are far from the support of the source. A feature generator learns to generate target features inside the support to fool the classifiers. Since the generator uses feedback from task-specific classifiers, it avoids generating target features near class boundaries. Our method outperforms other methods on several datasets of image classification and semantic segmentation."}, "keywords": ["sliced Wasserstein discrepancy"], "citation_intent": "background"}
{"citing_id": "2304.12217v1", "cited_id": "physics/0508025", "section_title": "Introduction", "citation": "First, the standard indicator and list view exemplified by Google Scholar, as shown in Figure 1 (a), provides key bibliographic indicators (# of citations, h-index #REFR , etc.) as an overview of the scholar.", "text_before_citation": ["Commonly, scholar profiling is composed of two steps: 1) accurate extraction of 360-degree academic demographics of a scholar from the web, which is normally considered an information retrieval task; and 2) appropriate aggregation, analysis, and representation of the extracted academic data, which is largely a data mining task.", "This work focuses on the latter step and formulates the impactoriented scholar profiling problem -how to arrange a scholar's academic data to best represent his/her scientific impact.", "Here the impact is defined as the breadth and depth of one's scientific contribution, as well as their community recognition.", "The problem is closely related to important applications of scholar profiles, such as serving possible references for academic award selection and tenure evaluation #OTHEREFR , and predicting one's future academic output [20] .", "Existing scholar profiling websites generally design two types of views."], "text_after_citation": ["The view is complemented with multiple lists as the context (publications, co-authors, etc.).", "The indicators are believed to be helpful for scholar ranking, though the supplemented paper/coauthor lists are normally unprocessed.", "There is a missed opportunity to exploit the structure within these lists for impact-oriented contextual tasks beyond ranking, such as reasoning and understanding of scholar profiles.", "Second, scholarly websites from academics often exhibit various bibliometric networks.", "These networks somehow mitigate the drawback of indicator and list by illustrating the relational context of a scholar."], "citing_paper_content": {"title": "Impact-Oriented Contextual Scholar Profiling Using Self-Citation Graphs", "abstract": "Quantitatively profiling a scholar's scientific impact is important to modern research society. Current practices with bibliometric indicators (e.g., h-index), lists, and networks perform well at scholar ranking, but do not provide structured context for scholar-centric, analytical tasks such as profile reasoning and understanding. This work presents GeneticFlow (GF), a suite of novel graph-based scholar profiles that fulfill three essential requirements: structured-context, scholar-centric, and evolution-rich. We propose a framework to compute GF over large-scale academic data sources with millions of scholars. The framework encompasses a new unsupervised advisoradvisee detection algorithm, a well-engineered citation type classifier using interpretable features, and a fine-tuned graph neural network (GNN) model. Evaluations are conducted on the real-world task of scientific award inference. Experiment outcomes show that the F1 score of best GF profile significantly outperforms alternative methods of impact indicators and bibliometric networks in all the 6 computer science fields considered. Moreover, the core GF profiles, with 63.6%\u223c66.5% nodes and 12.5%\u223c29.9% edges of the full profile, still significantly outrun existing methods in 5 out of 6 fields studied. Visualization of GF profiling result also reveals human explainable patterns for high-impact scholars."}, "cited_paper_content": {"title": "An Index To Quantify An Individual'S Scientific Research Output", "abstract": "I propose the index $h$, defined as the number of papers with citation number higher or equal to $h$, as a useful index to characterize the scientific output of a researcher."}, "keywords": ["Google Scholar", "h-index"], "citation_intent": "method"}
{"citing_id": "2304.07769v1", "cited_id": "2001.06591", "section_title": "Related Work", "citation": "Usually, generative models attempt to learn the reconstruction of the data and use this reconstruction to identify anomalous samples #REFR .", "text_before_citation": ["Deep support vector data description (DSVDD) finds a hypersphere to enclose the representation of normal samples #OTHEREFR .", "Liu and Gryllias constructed frequency domain features using cyclic spectral analysis and applied them in the support vector data description (SVDD) framework.", "This method has been proven robust against outliers and can achieve a high detection rate for detecting anomalies #OTHEREFR .", "In #OTHEREFR , researchers presented a new approach to identify imagery anomalies by training the model on normal images altered by geometric transformation.", "In this model, the classifier calculates the anomaly score using softmax statistics."], "text_after_citation": ["For instance, auto-encoders model the normal data distribution, and the reconstruction error is used as the anomaly score #OTHEREFR .", "Deep structured energy-based models (DSEBMs) learn an energy-based model and map each sample to an energy score #OTHEREFR .", "Deep autoencoding Gaussian mixture model (DAGMM) estimates a mixed Gaussian distribution by using an encoder for normal samples #OTHEREFR .", "A recent line of work on anomaly detection has focused on adversarial neural networks.", "For example, this structure has been used to identify anomalies in medical images #OTHEREFR ."], "citing_paper_content": {"title": "Regularized Complete Cycle Consistent Gan For Anomaly Detection", "abstract": "This study presents an adversarial method for anomaly detection in real-world applications, leveraging the power of generative adversarial neural networks (GANs) through cycle consistency in reconstruction error. Previous methods suffer from the high variance between class-wise accuracy which leads to not being applicable for all types of anomalies. The proposed method named RCALAD tries to solve this problem by introducing a novel discriminator to the structure, which results in a more efficient training process. Additionally, RCALAD employs a supplementary distribution in the input space to steer reconstructions toward the normal data distribution, effectively separating anomalous samples from their reconstructions and facilitating more accurate anomaly detection. To further enhance the performance of the model, two novel anomaly scores are introduced. The proposed model has been thoroughly evaluated through extensive experiments on six various datasets, yielding results that demonstrate its superiority over existing state-of-the-art models. The code is readily available to the research community at https://github.com/zahraDehghanian97/RCALAD."}, "cited_paper_content": {"title": "Regularized Cycle Consistent Generative Adversarial Network For Anomaly Detection", "abstract": "In this paper, we investigate algorithms for anomaly detection. Previous anomaly detection methods focus on modeling the distribution of non-anomalous data provided during training. However, this does not necessarily ensure the correct detection of anomalous data. We propose a new Regularized Cycle Consistent Generative Adversarial Network (RCGAN) in which deep neural networks are adversarially trained to better recognize anomalous samples. This approach is based on leveraging a penalty distribution with a new definition of the loss function and novel use of discriminator networks. It is based on a solid mathematical foundation, and proofs show that our approach has stronger guarantees for detecting anomalous examples compared to the current state-of-the-art. Experimental results on both real-world and synthetic data show that our model leads to significant and consistent improvements on previous anomaly detection benchmarks. Notably, RCGAN improves on the state-of-the-art on the KDDCUP, Arrhythmia, Thyroid, Musk and CIFAR10 datasets."}, "keywords": ["anomalous samples"], "citation_intent": "background"}
{"citing_id": "2303.05071v1", "cited_id": "1903.11027", "section_title": "Results", "citation": "NuScenes #REFR presents a greater challenge for 3D SOT task than KITTI due to its larger data volumes and lower frequency for annotated frames (2Hz for NuScenes v.s. 10Hz for KITTI and WOD).", "text_before_citation": ["However, only MBP-Track can accurately track the target after the occlusion disappears, owing to the sufficient use of temporal information.", "To demonstrate the generalization ability of our proposed MBPTrack, we evaluate the KITTI pretrained models on WOD #OTHEREFR , following previous work #OTHEREFR .", "The corresponding categories between KITTI and WOD datasets are Car\u2192Vehicle and Pedestrian\u2192Pedestrian. The experimental results, as presented in Tab.", "2, indicate that MBP-Track yields competitive or better tracking results than other methods under different levels of sparsity.", "In conclusion, our proposed method not only precisely tracks targets of all sizes but also generalizes well to unseen scenarios."], "text_after_citation": ["We conduct a comparison of our approach with previous methods on the NuScenes dataset following M2-Track #OTHEREFR . As shown in Tab.", "3, our method achieves a consistent and large performance gain compared with the previous state-of-the-art method, M2-Track.", "Leveraging the rich temporal and spatial information contained in the historical frames, MBPTrack exhibits superior performance over methods that only consider two frames when large appearance variation occurs between them. Fig.", "4 shows the model complexity and average inference time of different components in the Car category on KITTI.", "Our experiments are conducted on a single NVIDIA RTX 3090."], "citing_paper_content": {"title": "Mbptrack: Improving 3D Point Cloud Tracking With Memory Networks And Box Priors", "abstract": "3D single object tracking has been a crucial problem for decades with numerous applications such as autonomous driving. Despite its wide-ranging use, this task remains challenging due to the significant appearance variation caused by occlusion and size differences among tracked targets. To address these issues, we present MBPTrack, which adopts a Memory mechanism to utilize past information and formulates localization in a coarse-to-fine scheme using Box Priors given in the first frame. Specifically, past frames with targetness masks serve as an external memory, and a transformer-based module propagates tracked target cues from the memory to the current frame. To precisely localize objects of all sizes, MBPTrack first predicts the target center via Hough voting. By leveraging box priors given in the first frame, we adaptively sample reference points around the target center that roughly cover the target of different sizes. Then, we obtain dense feature maps by aggregating point features into the reference points, where localization can be performed more effectively. Extensive experiments demonstrate that MBPTrack achieves state-of-the-art performance on KITTI, nuScenes and Waymo Open Dataset, while running at 50 FPS on a single RTX3090 GPU."}, "cited_paper_content": {"title": "Nuscenes: A Multimodal Dataset For Autonomous Driving", "abstract": "Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology. Image-based benchmark datasets have driven development in computer vision tasks such as object detection, tracking and segmentation of agents in the environment. Most autonomous vehicles, however, carry a combination of cameras and range sensors such as lidar and radar. As machine learning based methods for detection and tracking become more prevalent, there is a need to train and evaluate such methods on datasets containing range sensor data along with images. In this work we present nuTonomy scenes (nuScenes), the first published dataset to carry the full autonomous vehicle sensor suite: 6 cameras, 5 radars and 1 lidar, all with full 360 degree field of view. nuScenes comprises 1000 scenes, each 20s long and fully annotated with 3D bounding boxes for 23 classes and 8 attributes. It has 7x as many annotations and 100x as many images as the pioneering KITTI dataset. We define novel 3D detection and tracking metrics. We also provide careful dataset analysis as well as baselines for lidar and image based detection and tracking. Data, development kit and more information are available online at this http URL."}, "keywords": ["annotated frames", "3D SOT task"], "citation_intent": "background"}
{"citing_id": "2303.13009v1", "cited_id": "1804.02516", "section_title": "B. Dataset Details", "citation": "On the other hand, the 1kB split uses the identical 1,000 test split of 1kA for the test, whereas the train set is a subset of 1kA's containing 6,656 samples #REFR .", "text_before_citation": ["The train dataset contains 1,261 samples, and the test set contains 439 samples, respectively.", "MSRVTT.", "The original MSRVTT-full #OTHEREFR dataset, used on video captioning task, contains 6,513 train, 497 validation, and 2,990 test samples.", "However, we have observed a wide range of dataset split variations throughout research on text-to-video retrieval.", "One split variant randomly samples 1,000 clip-text pairs from the test set for evaluation and uses the rest of the 9,000 samples as train data #OTHEREFR , which is commonly denoted as the 1kA split."], "text_after_citation": ["Another commonly used data split also uses the identical 1,000 test set, while adopting both the train and validation set from the standard MSRVTT for training.", "We evaluated our method on two split protocols most prominently observed in the literature, 1kA, and 7k.", "For convenience, we denote the former as MSRVTT-9k and the latter as MSRVTT-7k. MSVD-QA.", "MSVD-QA #OTHEREFR contains 47k open-ended questions on 2k videos, derived from the original MSVD dataset #OTHEREFR .", "We construct the answer set with 1,000 most frequently appeared answers."], "citing_paper_content": {"title": "Meltr: Meta Loss Transformer For Learning To Fine-Tune Video Foundation Models", "abstract": "Foundation models have shown outstanding performance and generalization capabilities across domains. Since most studies on foundation models mainly focus on the pretraining phase, a naive strategy to minimize a single task-specific loss is adopted for fine-tuning. However, such fine-tuning methods do not fully leverage other losses that are potentially beneficial for the target task. Therefore, we propose MEta Loss TRansformer (MELTR), a plug-in module that automatically and non-linearly combines various loss functions to aid learning the target task via auxiliary learning. We formulate the auxiliary learning as a bi-level optimization problem and present an efficient optimization algorithm based on Approximate Implicit Differentiation (AID). For evaluation, we apply our framework to various video foundation models (UniVL, Violet and All-in-one), and show significant performance gain on all four downstream tasks: text-to-video retrieval, video question answering, video captioning, and multimodal sentiment analysis. Our qualitative analyses demonstrate that MELTR adequately 'transforms' individual loss functions and 'melts' them into an effective unified loss."}, "cited_paper_content": {"title": "Learning A Text-Video Embedding From Incomplete And Heterogeneous Data", "abstract": "Joint understanding of video and language is an active research area with many applications. Prior work in this domain typically relies on learning text-video embeddings. One difficulty with this approach, however, is the lack of large-scale annotated video-caption datasets for training. To address this issue, we aim at learning text-video embeddings from heterogeneous data sources. To this end, we propose a Mixture-of-Embedding-Experts (MEE) model with ability to handle missing input modalities during training. As a result, our framework can learn improved text-video embeddings simultaneously from image and video datasets. We also show the generalization of MEE to other input modalities such as face descriptors. We evaluate our method on the task of video retrieval and report results for the MPII Movie Description and MSR-VTT datasets. The proposed MEE model demonstrates significant improvements and outperforms previously reported methods on both text-to-video and video-to-text retrieval tasks. Code is available at: https://github.com/antoine77340/Mixture-of-Embedding-Experts"}, "keywords": ["6,656 samples", "train set"], "citation_intent": "method"}
{"citing_id": "2303.14012v1", "cited_id": "1612.00593", "section_title": "I. Introduction", "citation": "We then use PointNet #REFR architecture to form a mapping from point-cloud observation of the target object, and pose of the deformable tool to 3D representation of the contact points between the two bodies.", "text_before_citation": ["However, this contact reasoning concept is more suitable for the control aspect than for the planning aspect due to its ability to track the interaction in real time.", "Thus, the question of how to predict the interaction between deformable and rigid objects and exploit such interactions for planning remains open.", "To address the aforementioned open issues, we propose Sequence Planning with deformable-ON-rigid contact prediction from GEometric features (SPONGE), a sequence planning pipeline powered by a contact prediction model that predicts contact between deformable and rigid bodies, with the aim of providing robots with the aforementioned human-like planning skill in order to efficiently automate downstream deformable object manipulation tasks such as cleaning dishes (Fig. 1 ).", "Instead of contact reasoning, in this paper we tackle the concept of contact prediction of a 3D deformable tool acting on rigid objects, which is better suited for planning purposes.", "We take a data-driven approach with physics-based simulation to model the interactions between 3D deformable objects and rigid objects."], "text_after_citation": ["The trained contact prediction model is then used as the driving force behind the planning of a subsequent task.", "Finally, we deploy SPONGE in the real world to demonstrate that the contact prediction model trained only with synthetic data from physics-based simulation can help to produce an efficient plan for a manipulation task to be successfully executed in the real world.", "In summary, the main contributions of this paper are as follows:", "\u2022 A deep learning-based contact prediction model that predicts the contacts between 3D deformable and rigid objects under interactions.", "\u2022 A planning pipeline powered by the proposed contact prediction model to efficiently automate deformable object manipulation tasks."], "citing_paper_content": {"title": "Sponge: Sequence Planning With Deformable-On-Rigid Contact Prediction From Geometric Features", "abstract": "Planning robotic manipulation tasks, especially those that involve interaction between deformable and rigid objects, is challenging due to the complexity in predicting such interactions. We introduce SPONGE, a sequence planning pipeline powered by a deep learning-based contact prediction model for contacts between deformable and rigid bodies under interactions. The contact prediction model is trained on synthetic data generated by a developed simulation environment to learn the mapping from point-cloud observation of a rigid target object and the pose of a deformable tool, to 3D representation of the contact points between the two bodies. We experimentally evaluated the proposed approach for a dish cleaning task both in simulation and on a real Franka Emika Panda with real-world objects. The experimental results demonstrate that in both scenarios the proposed planning pipeline is capable of generating high-quality trajectories that can accomplish the task by achieving more than 90% area coverage on different objects of varying sizes and curvatures while minimizing travel distance. Code and video are available at: https://irobotics.aalto.fi/sponge/."}, "cited_paper_content": {"title": "Pointnet: Deep Learning On Point Sets For 3D Classification And Segmentation", "abstract": "Point cloud is an important type of geometric data structure. Due to its irregular format, most researchers transform such data to regular 3D voxel grids or collections of images. This, however, renders data unnecessarily voluminous and causes issues. In this paper, we design a novel type of neural network that directly consumes point clouds, which well respects the permutation invariance of points in the input. Our network, named PointNet, provides a unified architecture for applications ranging from object classification, part segmentation, to scene semantic parsing. Though simple, PointNet is highly efficient and effective. Empirically, it shows strong performance on par or even better than state of the art. Theoretically, we provide analysis towards understanding of what the network has learnt and why the network is robust with respect to input perturbation and corruption."}, "keywords": ["deformable tool", "PointNet architecture"], "citation_intent": "method"}
{"citing_id": "2304.06632v1", "cited_id": "1406.2661", "section_title": "B. Necessary Conditions Of Aigc", "citation": "Goodfellow proposed the first generative model, Generative Adversarial Network (GAN), in 2014 #REFR . Table II shows the evolution timeline of generative algorithms.", "text_before_citation": ["The algorithms simulate the simple human brain, which improves the learning model through experience accumulation #OTHEREFR .", "Neural network models further emulate the signal processing and thinking mechanisms of human brain nerves #OTHEREFR , #OTHEREFR .", "Furthermore, generative algorithms, such as Google's Transformer architecture #OTHEREFR , draw on human attention mechanisms to enable the completion of multiple tasks by an algorithm.", "They can model the probability distribution of the input data and then generate new data.", "They are trained by processing large amounts of text data, then fine-tuned on specific tasks with labeled data, enabling them to interact according to contextual content and chat in a manner similar to human beings."], "text_after_citation": ["In most cases, the significance of GAN is a source of inspiration for many popular variations and architectures.", "The transformer model has a wide range of applications in various domains (including NLP and CV).", "In addition, several pre-training models, such as BERT, GPT-3, and LaMDA, have been developed based on the Transformer model.", "The diffusion model is currently the most advanced image generation model because of its optimized performance.", "With the development of generative models, language models have also made great progress. For example, Devlin et al."], "citing_paper_content": {"title": "Ai-Generated Content (Aigc): A Survey", "abstract": "To address the challenges of digital intelligence in the digital economy, artificial intelligence-generated content (AIGC) has emerged. AIGC uses artificial intelligence to assist or replace manual content generation by generating content based on user-inputted keywords or requirements. The development of large model algorithms has significantly strengthened the capabilities of AIGC, which makes AIGC products a promising generative tool and adds convenience to our lives. As an upstream technology, AIGC has unlimited potential to support different downstream applications. It is important to analyze AIGC's current capabilities and shortcomings to understand how it can be best utilized in future applications. Therefore, this paper provides an extensive overview of AIGC, covering its definition, essential conditions, cutting-edge capabilities, and advanced features. Moreover, it discusses the benefits of large-scale pretrained models and the industrial chain of AIGC. Furthermore, the article explores the distinctions between auxiliary generation and automatic generation within AIGC, providing examples of text generation. The paper also examines the potential integration of AIGC with the Metaverse. Lastly, the article highlights existing issues and suggests some future directions for application. Impact Statement-It is necessary for academia and industry to take an overview of what AIGC is, how AIGC works, how AIGC changes our lifestyles, and what AIGC will be in the future. This article proposes a survey of AIGC from its definition, pros, cons, applications, current challenges, and future directions to answer these urgent questions. We summarize the existing major literature, which helps relevant researchers become familiar with and understand the existing works and unsolved problems. Based on the review of literature and the commercialization of scientific and research findings, we conduct some cutting-edge AIGC research. In particular, the challenges and future directions of AIGC can be helpful for developing AI. Relevant technologies of AIGC will boost the development of artificial intelligence, better serve human society, and achieve sustainable development."}, "cited_paper_content": {"title": "Generative Adversarial Networks", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["generative algorithms", "Generative Adversarial Network"], "citation_intent": "background"}
{"citing_id": "2304.02714v1", "cited_id": "1505.04597", "section_title": "B. Networks And Algorithms 1) Whistle Extraction Network:", "citation": "The generators follow the U-Net #REFR architecture, which has 6 U-Net blocks with a basic width of 64. InstanceNorm layers are used in the U-Net blocks.", "text_before_citation": ["It contains 4 convolutional layers with a stride of 2 and a fully connected layer.", "The networks are optimized by Adam optimizers (initial learning rate = 1 \u00d7 10 \u22124 , betas = [0.5, 0.9], batch size = 64) for 3 \u00d7 10 4 and 5 \u00d7 10 4 iterations on the reduced and full datasets, respectively.", "In each WGAN training iteration, the discriminator is optimized for 5 steps while the generator is optimized for 1 step, where the network parameters are updated by applying the optimizer to one minibatch of data in each step.", "For sample selection, we used T e =70, T c =0.5, T p =64.", "3) CycleGAN: The GAN model that we used to add whistles on synthetic noise employs the CycleGAN architecture of #OTHEREFR ."], "text_after_citation": ["The discriminator is a fully convolutional network with 3 convolutional layers.", "We trained the generators and discriminators with Adam optimizers (learning rate = 2 \u00d7 10 \u22124 , betas = [0.5, 0.999], batch size = 64) for 25,120 iterations (160 epochs for 10,000 real positive samples) for the reduced dataset and 50 epochs for the full dataset.", "We set \u03bb 0 =10, \u03bb 1 =0.5, and \u03bb 2 =0.5 for Eq. 11.", "We apply a random \u03b3 following a unified distribution between (0.5, 1.5) in Eq. 3."], "citing_paper_content": {"title": "Learning Stage-Wise Gans For Whistle Extraction In Time-Frequency Spectrograms", "abstract": "Whistle contour extraction aims to derive animal whistles from time-frequency spectrograms as polylines. For toothed whales, whistle extraction results can serve as the basis for analyzing animal abundance, species identity, and social activities. During the last few decades, as long-term recording systems have become affordable, automated whistle extraction algorithms were proposed to process large volumes of recording data. Recently, a deep learning-based method demonstrated superior performance in extracting whistles under varying noise conditions. However, training such networks requires a large amount of labor-intensive annotation, which is not available for many species. To overcome this limitation, we present a framework of stage-wise generative adversarial networks (GANs), which compile new whistle data suitable for deep model training via three stages: generation of background noise in the spectrogram, generation of whistle contours, and generation of whistle signals. By separating the generation of different components in the samples, our framework composes visually promising whistle data and labels even when few expert annotated data are available. Regardless of the amount of human-annotated data, the proposed data augmentation framework leads to a consistent improvement in performance of the whistle extraction model, with a maximum increase of 1.69 in the whistle extraction mean F1-score. Our stage-wise GAN also surpasses one single GAN in improving whistle extraction models with augmented data."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["InstanceNorm layers"], "citation_intent": "method"}
{"citing_id": "2303.11041v1", "cited_id": "1807.08555", "section_title": "Results", "citation": "We compare our loss (editing loss) with the following baselines: #REFR No Editing: the initial segmentation y init is used as the final segmentation\u0177, and the overall error in this case is the distance from the CAS contours to y init . This should serve as an upper bound for error.", "text_before_citation": ["We use the editing evaluation metric D and report the 95 th percentile of the overall editing error, the error near the user input, and far from the user input (mm).", "The near and far regions are defined by thresholding A at 0.5.", "For the CV results, we report the mean and standard deviation over the five folds.", "The statistical significance is computed for the difference with InterCNN.", "\u2020 : p-value < 0.01, \u2021 : p-value < 0.001."], "text_after_citation": ["(2) CE Loss: an editing model trained using the standard CE segmentation loss.", "(3) Dice Loss #OTHEREFR : an editing model trained using Dice segmentation loss.", "4InterCNN #OTHEREFR : for every training sample, simulated user edits based on the prediction are accumulated with any previous edits and re-input to the model for 10 iterations, trained using CE loss.", "We report the results after a single edit (the furthest CAS contour from\u0177) in Table 1 .", "A single training epoch takes \u2248 3 minutes for all the models except for InterCNN, which takes \u2248 14 minutes, on a single NVIDIA Tesla V100 GPU."], "citing_paper_content": {"title": "From Sparse To Precise: A Practical Editing Approach For Intracardiac Echocardiography Segmentation", "abstract": "Accurate and safe catheter ablation procedures for patients with atrial fibrillation require precise segmentation of cardiac structures in Intracardiac Echocardiography (ICE) imaging. Prior studies have suggested methods that employ 3D geometry information from the ICE transducer to create a sparse ICE volume by placing 2D frames in a 3D grid, enabling training of 3D segmentation models. However, the resulting 3D masks from these models can be inaccurate and may lead to serious clinical complications due to the sparse sampling in ICE data, frames misalignment, and cardiac motion. To address this issue, we propose an interactive editing framework that allows users to edit segmentation output by drawing scribbles on a 2D frame. The user interaction is mapped to the 3D grid and utilized to execute an editing step that modifies the segmentation in the vicinity of the interaction while preserving the previous segmentation away from the interaction. Furthermore, our framework accommodates multiple edits to the segmentation output in a sequential manner without compromising previous edits. This paper presents a novel loss function and a novel evaluation metric specifically designed for editing. Results from cross-validation and testing indicate that our proposed loss function outperforms standard losses and training strategies in terms of segmentation quality and following user input. Additionally, we show quantitatively and qualitatively that subsequent edits do not compromise previous edits when using our method, as opposed to standard segmentation losses. Overall, our approach enhances the accuracy of the segmentation while avoiding undesired changes away from user interactions and without compromising the quality of previously edited regions, leading to better patient outcomes."}, "cited_paper_content": {"title": "Iterative Interaction Training For Segmentation Editing Networks", "abstract": "Automatic segmentation has great potential to facilitate morphological measurements while simultaneously increasing efficiency. Nevertheless often users want to edit the segmentation to their own needs and will need different tools for this. There has been methods developed to edit segmentations of automatic methods based on the user input, primarily for binary segmentations. Here however, we present an unique training strategy for convolutional neural networks (CNNs) trained on top of an automatic method to enable interactive segmentation editing that is not limited to binary segmentation. By utilizing a robot-user during training, we closely mimic realistic use cases to achieve optimal editing performance. In addition, we show that an increase of the iterative interactions during the training process up to ten improves the segmentation editing performance substantially. Furthermore, we compare our segmentation editing CNN (interCNN) to state-of-the-art interactive segmentation algorithms and show a superior or on par performance."}, "keywords": ["initial segmentation", "final segmentation\u0177"], "citation_intent": "method"}
{"citing_id": "2304.03510v1", "cited_id": "2001.01202", "section_title": "Results And Discussion", "citation": "The detection error with the deep features method #REFR on WL images indicates a performance similar to that of both the VIS and NIR spectral bands.", "text_before_citation": ["The deep features method #OTHEREFR , which is based on facial features, indicates less variation in degraded performance, especially in the NIR bands.", "However, the Hierarchical Deep Residual SLERP #OTHEREFR , which is based on texture features that indicate higher variation in the detection performance, especially in NIR spectral bands, compared to the deep features method #OTHEREFR .", "The possible degradation of the Hierarchical Deep Residual SLERP #OTHEREFR can be attributed to the lack of texture information in the NIR spectral bands.", "Thus, the detection performance of D-MAD in the NIR spectrum depends on the type of feature used for morphing de-tection.", "\u2022 The use of wholeLight (WL) that is captured without any spectral filtering indicates the varied detection performance with respect to D-MAD techniques."], "text_after_citation": ["Because the deep features method #OTHEREFR is based on facial features, it is less sensitive to different spectral bands, which may be due to the backbone model that is trained only using VIS images.", "However, the performance of Hierarchical Deep Residual SLERP #OTHEREFR on WL images shows degraded performance compared to the VIS spectral bands.", "Table 3 lists the performance of the D-MAD algorithms on the proposed multispectral framework and visible images.", "Both multispectral and visible data are collected using the same data subjects, and this can provide insights into the utility of multispectral images for reliable morphing attack detection.", "Furthermore, the performance of the multispectral framework was presented by fusing the results of all individual spectral bands."], "citing_paper_content": {"title": "Multispectral Imaging For Differential Face Morphing Attack Detection: A Preliminary Study", "abstract": "Face morphing attack detection is emerging as an increasingly challenging problem owing to advancements in high-quality and realistic morphing attack generation. Reliable detection of morphing attacks is essential because these attacks are targeted for border control applications. This paper presents a multispectral framework for differential morphing-attack detection (D-MAD). The D-MAD methods are based on using two facial images that are captured from the ePassport (also called the reference image) and the trusted device (for example, Automatic Border Control (ABC) gates) to detect whether the face image presented in ePassport is morphed. The proposed multispectral D-MAD framework introduce a multispectral image captured as a trusted capture to capture seven different spectral bands to detect morphing attacks. Extensive experiments were conducted on the newly created datasets with 143 unique data subjects that were captured using both visible and multispectral cameras in multiple sessions. The results indicate the superior performance of the proposed multispectral framework compared to visible images."}, "cited_paper_content": {"title": "Deep Face Representations For Differential Morphing Attack Detection", "abstract": "The vulnerability of facial recognition systems to face morphing attacks is well known. Many different approaches for morphing attack detection have been proposed in the scientific literature. However, the morphing attack detection algorithms proposed so far have only been trained and tested on datasets whose distributions of image characteristics are either very limited (e.g. only created with a single morphing tool) or rather unrealistic (e.g. no print-scan transformation). As a consequence, these methods easily overfit on certain image types and the results presented cannot be expected to apply to real-world scenarios. For example, the results of the latest NIST Face Recognition Vendor Test MORPH show that the submitted MAD algorithms lack robustness and performance when considering unseen and challenging datasets. In this work, subsets of the FERET and FRGCv2 face databases are used to create a large realistic database for training and testing of morphing attack detection algorithms, containing a large number of ICAO-compliant bona fide facial images, corresponding unconstrained probe images, and morphed images created with four different tools. Furthermore, multiple post-processings are applied on the reference images, e.g. print-scan and JPEG2000 compression. On this database, previously proposed differential morphing algorithms are evaluated and compared. In addition, the application of deep face representations for differential morphing attack detection algorithms is investigated. It is shown that algorithms based on deep face representations can achieve very high detection performance (less than 3\\%~\\mbox{D-EER}) and robustness with respect to various post-processings. Finally, the limitations of the developed methods are analyzed."}, "keywords": ["deep features method"], "citation_intent": "result"}
{"citing_id": "2303.05796v2", "cited_id": "1906.04032", "section_title": "A.4 Model Details", "citation": "As seen in Table 5 , we found that a more expressive normalizing flow with resampled base #REFR Stimper et al., 2022) improves significantly the results over a simpler radial normalizing flow (Rezende & Mohamed, 2015) across all the metrics.", "text_before_citation": ["The list of core architectures used across the experiments are: ResNet18 / ResNet50 / EfficientNet / Swin #OTHEREFR Tan & Le, 2021; from the torchvision repository 2 and Wide-ResNet-28-10 (Zagoruyko & Komodakis, 2016) from the original implementation of DUE.", "Except for the experiment on architecture type and size where ResNet18 has output channels for the residual blocks with size [64, 128, 256, 512] , ResNet18 has output channels for the residual blocks with size [32, 64, 128, 256] which causes small differences in final accuracy.", "Uncertainty head.", "For DUE we use the original implementation 3 with by default we use the RBF kernel function.", "For NatPN we use the original implementation 4 but change the uncertainty head with a more expressive density estimator."], "text_after_citation": ["For all the experiments (except toys where we use radial flow) we use NSF-R with 16 layers."], "citing_paper_content": {"title": "Workshop On The Pitfalls Of Limited Data And Computation For Trustworthy Ml, Iclr 2023 Training, Architecture, And Prior For Deterministic Uncertainty Methods", "abstract": "Accurate and efficient uncertainty estimation is crucial to build reliable Machine Learning (ML) models capable to provide calibrated uncertainty estimates, generalize and detect Out-Of-Distribution (OOD) datasets. To this end, Deterministic Uncertainty Methods (DUMs) is a promising model family capable to perform uncertainty estimation in a single forward pass. This work investigates important design choices in DUMs: (1) we show that training schemes decoupling the core architecture and the uncertainty head schemes can significantly improve uncertainty performances. (2) we demonstrate that the core architecture expressiveness is crucial for uncertainty performance and that additional architecture constraints to avoid feature collapse can deteriorate the trade-off between OOD generalization and detection. (3) Contrary to other Bayesian models, we show that the prior defined by DUMs do not have a strong effect on the final performances."}, "cited_paper_content": {"title": "Neural Spline Flows", "abstract": "A normalizing flow models a complex probability density as an invertible transformation of a simple base density. Flows based on either coupling or autoregressive transforms both offer exact density evaluation and sampling, but rely on the parameterization of an easily invertible elementwise transformation, whose choice determines the flexibility of these models. Building upon recent work, we propose a fully-differentiable module based on monotonic rational-quadratic splines, which enhances the flexibility of both coupling and autoregressive transforms while retaining analytic invertibility. We demonstrate that neural spline flows improve density estimation, variational inference, and generative modeling of images."}, "keywords": ["expressive normalizing flow"], "citation_intent": "result"}
{"citing_id": "2303.03319v1", "cited_id": "quant-ph/0401091", "section_title": "Path Finding In Arbitrary Graphs", "citation": "However, we provide an alternative approach that, while not as fast as Algorithm 3, still provides an improvement over the algorithm of #REFR for graphs in which all (self-avoiding) paths from s to t are short.", "text_before_citation": ["For the more general case of st-path G (x) when G(x) is not known to only have one st-path, while it is possible that an algorithm similar to Algorithm 3 would work, we have not been able to bound the running time effectively.", "This is because in the case of a single path, once you find an intermediate edge on the path, the longest paths from s and t to that edge must be shorter than the length of the longest path from s to t.", "This ensures that subproblems take shorter time than the original problem. With multiple paths, we no longer have that guarantee."], "text_after_citation": ["Our approach does not make use of our path-edge sampling algorithm as a subroutine, and instead uses the path detection algorithm of Lemma 8 to decide whether there are paths through various subgraphs, and then uses that information to find each edge in a path in order from s to t.", "In this way, we avoid the problem of subproblems being larger than the original problem, since if the longest path from s to t has length L, and the first edge we find on the path is (s, u), then longest path from u to t that doesn't go through s must have length at most L \u2212 1. However, we lose the advantage of a divide-and-conquer approach.", "To find the first edge on a path, we use a group testing approach.", "We divide the neighbors of s in G into two sets, S 1 and S 2 and run path detection algorithms in parallel on two subgraphs of G(x), one with edges from s removed, except those to vertices in S 1 (that is, G \u2212 {{s,u}\u2208E:u\u2208S 1 } ), and one with edges from s removed, except those to vertices in S 2 .", "We will detect which of these subgraphs contains a path, and we will know there is a path whose first edge goes from s to a vertex in the corresponding set (S 1 or S 2 )."], "citing_paper_content": {"title": "Quantum Algorithm For Path-Edge Sampling", "abstract": "We present a quantum algorithm for sampling an edge on a path between two nodes s and t in an undirected graph given as an adjacency matrix, and show that this can be done in query complexity that is asymptotically the same, up to log factors, as the query complexity of detecting a path between s and t. We use this path sampling algorithm as a subroutine for st-path finding and st-cut-set finding algorithms in some specific cases. Our main technical contribution is an algorithm for generating a quantum state that is proportional to the positive witness vector of a span program."}, "cited_paper_content": {"title": "Quantum Query Complexity Of Some Graph Problems", "abstract": "Quantum algorithms for graph problems are considered, both in the adjacency matrix model and in an adjacency list-like array model. We give almost tight lower and upper bounds for the bounded error quantum query complexity of Connectivity, Strong Connectivity, Minimum Spanning Tree, and Single Source Shortest Paths. For example we show that the query complexity of Minimum Spanning Tree is in \u0398(n 3/2) in the matrix model and in \\(\\Theta(\\sqrt{nm})\\) in the array model, while the complexity of Connectivity is also in \u0398(n 3/2) in the matrix model, but in \u0398(n) in the array model. The upper bounds utilize search procedures for finding minima of functions under various conditions."}, "keywords": ["algorithm"], "citation_intent": "method"}
{"citing_id": "2303.06544v1", "cited_id": "1609.04836", "section_title": "Other Hyperparameters", "citation": "In our preliminary experiments, small batches provided models with worse generalization, and sets of 256 and 512 objects produced similar results; hence, to prioritize a flat minimum #REFR , we chose 256 objects in each batch.", "text_before_citation": ["Each of those hyperparameters is relevant and has an impact on the final results, but focusing on generalization, it is essential to highlight the early stopping criteria.", "The patience hyperparameter allows managing the overfitting; when a higher patience value is used, the overfitting is increased, and obviously, the performance in the testing set is decreased.", "Another relevant pattern was observed in the batch size; the impact of the size batch on optimization results has been studied.", "The evidence shows that a small batch size generates flat minima (better generalization), and on the other hand, bigger batches tend to lead to sharp minima #OTHEREFR .", "However, this evidence is based on a representative training set, so each batch is a good sample; in a case considering data shift, that is not true, and because of that, small batches can provide very wrong directions for the optimizer."], "text_after_citation": [], "citing_paper_content": {"title": "Informative Regularization For A Multi-Layer Perceptron Rr Lyrae Classifier Under Data Shift", "abstract": "In recent decades, machine learning has provided valuable models and algorithms for processing and extracting knowledge from time-series surveys. Different classifiers have been proposed and performed to an excellent standard. Nevertheless, few papers have tackled the data shift problem in labeled training sets, which occurs when there is a mismatch between the data distribution in the training set and the testing set. This drawback can damage the prediction performance in unseen data. Consequently, we propose a scalable and easily adaptable approach based on an informative regularization and an ad-hoc training procedure to mitigate the shift problem during the training of a multi-layer perceptron for RR Lyrae classification. We collect ranges for characteristic features to construct a symbolic representation of prior knowledge, which was used to model the informative regularizer component. Simultaneously, we design a two-step back-propagation algorithm to integrate this knowledge into the neural network, whereby one step is applied in each epoch to minimize classification error, while another is applied to ensure regularization. Our algorithm defines a subset of parameters (a mask) for each loss function. This approach handles the forgetting effect, which stems from a trade-off between these loss functions (learning from data versus learning expert knowledge) during training. Experiments were conducted using recently proposed shifted benchmark sets for RR Lyrae stars, outperforming baseline models by up to 3% through a more reliable classifier. Our method provides a new path to incorporate knowledge from characteristic features into artificial neural networks to manage the underlying data shift problem."}, "cited_paper_content": {"title": "On Large-Batch Training For Deep Learning: Generalization Gap And Sharp Minima", "abstract": "The stochastic gradient descent (SGD) method and its variants are algorithms of choice for many Deep Learning tasks. These methods operate in a small-batch regime wherein a fraction of the training data, say $32$-$512$ data points, is sampled to compute an approximation to the gradient. It has been observed in practice that when using a larger batch there is a degradation in the quality of the model, as measured by its ability to generalize. We investigate the cause for this generalization drop in the large-batch regime and present numerical evidence that supports the view that large-batch methods tend to converge to sharp minimizers of the training and testing functions - and as is well known, sharp minima lead to poorer generalization. In contrast, small-batch methods consistently converge to flat minimizers, and our experiments support a commonly held view that this is due to the inherent noise in the gradient estimation. We discuss several strategies to attempt to help large-batch methods eliminate this generalization gap."}, "keywords": ["batch", "small batches"], "citation_intent": "result"}
{"citing_id": "2304.11267v1", "cited_id": "1406.2661", "section_title": "Related Work", "citation": "Image generation has garnered significant research interest, particularly in recent years, with the advent of GANs #REFR .", "text_before_citation": [], "text_after_citation": ["GANs consist of two neural networks, a generator and a discriminator, that compete with each other to create realistic images #OTHEREFR .", "While GANs demonstrate the capability to generate high resolution images with good perceptual quality #OTHEREFR , it suffers from the difficulty to train effectively #OTHEREFR .", "VAEs #OTHEREFR emerged as a popular generative model, utilizing probabilistic graphical models to generate images through latent space representation, enabling efficient synthesis but with lower sample quality than GANs.", "Denoising Diffusion Probabilistic Model (DDPM) #OTHEREFR marked a significant milestone in the development of diffusion-based generative models.", "DDPM demonstrated the potential of these models to generate high-quality images through a series of iterative noise-removal steps."], "citing_paper_content": {"title": "Speed Is All You Need: On-Device Acceleration Of Large Diffusion Models Via Gpu-Aware Optimizations", "abstract": "The rapid development and application of foundation models have revolutionized the field of artificial intelligence. Large diffusion models have gained significant attention for their ability to generate photorealistic images and support various tasks. On-device deployment of these models provides benefits such as lower server costs, offline functionality, and improved user privacy. However, common large diffusion models have over 1 billion parameters and pose challenges due to restricted computational and memory resources on devices. We present a series of implementation optimizations for large diffusion models that achieve the fastest reported inference latency to-date(under 12 seconds for Stable Diffusion 1.4 without INT8 quantization for a 512 \u00d7 512 image with 20 iterations) on GPUequipped mobile devices. These enhancements broaden the applicability of generative AI and improve the overall user experience across a wide range of devices."}, "cited_paper_content": {"title": "Generative Adversarial Networks", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["Image generation", "GANs"], "citation_intent": "background"}
{"citing_id": "2303.08977v1", "cited_id": "1811.10180", "section_title": "Motion Deblurring", "citation": "In particular, we point out that our results are sharper than the EDI #REFR reconstruction, which assumes all events correspond to the same amount of intensity change.", "text_before_citation": ["We use three image quality metrics to compare DeblurSR with different baseline approaches: the Mean Squared Error (MSE, lower is better), the Peak Signal-to-Noise Ratio (PSNR, higher is better), and the Structural Similarity Index Measure (SSIM, higher is better).", "Our method demonstrates an impressive ability in motion deblurring.", "Specifically, on the REDS dataset, De-blurSR improves the current best-performing method by 12.3% in MSE, 4.7% in PSNR, and 4.9% in SSIM.", "On HQF, DeblurSR outperforms the state-of-the-art approach by 22.2% in MSE, 10.1% in PSNR, and 14.0% in SSIM.", "Qualitatively, as shown in Figure 4 and Figure 5 , De-blurSR generates smooth and sharp frames."], "text_after_citation": ["Meanwhile, our recon-structed frames are significantly less noisy than the eSL-Net #OTHEREFR reconstruction, which overly emphasizes texture details.", "Compared with E-CIR #OTHEREFR , DeblurSR offers more realistic details around the thin edges.", "The difference between E-CIR and DeblurSR is sometimes subtle and hard to notice from static images.", "We encourage readers to watch the animated visualizations provided in the supplementary material, which additionally demonstrate the temporal smoothness of our results in the video format.", "DeblurSR is computationally efficient."], "citing_paper_content": {"title": "Deblursr: Event-Based Motion Deblurring Under The Spiking Representation", "abstract": "We present DeblurSR, a novel motion deblurring approach that converts a blurry image into a sharp video. DeblurSR utilizes event data to compensate for motion ambiguities and exploits the spiking representation to parameterize the sharp output video as a mapping from time to intensity. Our key contribution, the Spiking Representation (SR), is inspired by the neuromorphic principles determining how biological neurons communicate with each other in living organisms. We discuss why the spikes can represent sharp edges and how the spiking parameters are interpreted from the neuromorphic perspective. DeblurSR has higher output quality and requires fewer computing resources than state-of-the-art event-based motion deblurring methods. We additionally show that our approach easily extends to video super-resolution when combined with recent advances in implicit neural representation. The implementation and animated visualization of DeblurSR are available at https://github.com/chensong1995/DeblurSR."}, "cited_paper_content": {"title": "Bringing A Blurry Frame Alive At High Frame-Rate With An Event Camera", "abstract": "Event-based cameras can measure intensity changes (called \u2018events\u2019) with microsecond accuracy under high-speed motion and challenging lighting conditions. With the active pixel sensor (APS), the event camera allows simultaneous output of the intensity frames. However, the output images are captured at a relatively low frame-rate and often suffer from motion blur. A blurry image can be regarded as the integral of a sequence of latent images, while the events indicate the changes between the latent images. Therefore, we are able to model the blur-generation process by associating event data to a latent image. In this paper, we propose a simple and effective approach, the Event-based Double Integral (EDI) model, to reconstruct a high frame-rate, sharp video from a single blurry frame and its event data. The video generation is based on solving a simple non-convex optimization problem in a single scalar variable. Experimental results on both synthetic and real images demonstrate the superiority of our EDI model and optimization method in comparison to the state-of-the-art."}, "keywords": ["intensity change", "events"], "citation_intent": "result"}
{"citing_id": "2303.01155v2", "cited_id": "1606.05830", "section_title": "I. Introduction", "citation": "Visual SLAM (VSLAM) systems can employ a wide range of vision sensors, such as monocular, stereo, Red Green Blue-Depth (RGB-D), omnidirectional, and eventbased cameras to estimate the environmental map while localizing the camera #REFR .", "text_before_citation": [], "text_after_citation": ["The primary advantage of vision sensors is that they need low-cost hardware to supply rich visual and semantic information from surroundings for various tasks #OTHEREFR .", "Semantic data, which refers to high-level information acquired from the environment, make VSLAM tasks more robust and expand the range of applications that can employ the reconstructed maps #OTHEREFR , #OTHEREFR .", "For instance, robots may need to identify objects in the scene Fig.", "1 : The final reconstructed map of the environment using the proposed method in a hierarchical representation.", "Accordingly, the primary map entities are detected 3D points extracted from the environment and visited ArUco markers."], "citing_paper_content": {"title": "Marker-Based Visual Slam Leveraging Hierarchical Representations", "abstract": "Fiducial markers can encode rich information about the environment and can aid Visual SLAM (VSLAM) approaches in reconstructing maps with practical semantic information. Current marker-based VSLAM approaches mainly utilize markers for improving feature detections in low-feature environments and/or for incorporating loop closure constraints, generating only low-level geometric maps of the environment prone to inaccuracies in complex environments. To bridge this gap, this paper presents a VSLAM approach utilizing a monocular camera along with fiducial markers to generate hierarchical representations of the environment while improving the camera pose estimate. The proposed approach detects semantic entities from the surroundings, including walls, corridors, and rooms encoded within markers, and appropriately adds topological constraints among them. Experimental results on a real-world dataset collected with a robot demonstrate that the proposed approach outperforms a traditional marker-based VSLAM baseline in terms of accuracy, given the addition of new constraints while creating enhanced map representations. Furthermore, it shows satisfactory results when comparing the reconstructed map quality to the one reconstructed using a LiDAR SLAM approach."}, "cited_paper_content": {"title": "Past, Present, And Future Of Simultaneous Localization And Mapping: Toward The Robust-Perception Age", "abstract": "Simultaneous localization and mapping (SLAM) consists in the concurrent construction of a model of the environment (the map ), and the estimation of the state of the robot moving within it. The SLAM community has made astonishing progress over the last 30 years, enabling large-scale real-world applications and witnessing a steady transition of this technology to industry. We survey the current state of SLAM and consider future directions. We start by presenting what is now the de-facto standard formulation for SLAM. We then review related work, covering a broad set of topics including robustness and scalability in long-term mapping, metric and semantic representations for mapping, theoretical performance guarantees, active SLAM and exploration, and other new frontiers. This paper simultaneously serves as a position paper and tutorial to those who are users of SLAM. By looking at the published research with a critical eye, we delineate open challenges and new research issues, that still deserve careful scientific investigation. The paper also contains the authors\u2019 take on two questions that often animate discussions during robotics conferences: Do robots need SLAM? and Is SLAM solved?"}, "keywords": ["Visual SLAM (VSLAM)"], "citation_intent": "background"}
{"citing_id": "2303.08595v1", "cited_id": "2003.02389", "section_title": "Background And Related Works", "citation": "Its counterpart in structured pruning-Iterative L1-norm-based pruning (ILP) #REFR , which removes filters based on their L1-norm values, cannot effectively prune a model while maintaining its accuracy.", "text_before_citation": ["But these methods have to explore a large search space of all available layer-wise sparsity, which is time consuming when neural networks are large and datasets are complex.", "Therefore, there is a great need for an automatic, iterative, structured pruning solution that can automatically and efficiently generate small, accurate, and hardware-efficient models. The challenges are three-fold.", "First, how to effectively identify the insignificant parameters in a model to prune? Existing works have explored different mechanisms, e.g., L2-Norm in Soft Filter Pruning (SFP) #OTHEREFR , Soft Channel Pruning (SCP) [Kang and Han, 2020] and EagleEye #OTHEREFR , geometric median in FPGM #OTHEREFR , Hessian in EigenDamage , Empirical Sensitivity in Provable Filter Pruning (PFP) #OTHEREFR , adversarial knockoff features in SCOP #OTHEREFR , polarization regularizer in Neuron-level Structured Pruning (NSP) #OTHEREFR , LASSO regression in Channel Pruning (CP) #OTHEREFR , and other information considering the relationship between neighboring layers (Gate Batch Normalization (GBN) #OTHEREFR , Sparse Structure Selection (SSS) #OTHEREFR , Hinge #OTHEREFR , Pruning From Scratch (PFS) and Stripe-Wise Pruning (SWP) #OTHEREFR ).", "In comparison, activation-based attention, proposed in this paper, can more effectively capture the importance of filters, and pruning based on attention values can produce much better models, as quantitatively shown in our evaluation (Section 4).", "Second, how to design an effective iterative pruning process to recover the accuracy loss caused by structured pruning? LTH-based iterative pruning is a promising approach, but it has only been shown to work with unstructured pruning such as IMP."], "text_after_citation": ["For example, ILP can prune ResNet-50 by at most 11.5% of parameters when the maximum accuracy loss is limited to 1% on ImageNet.", "So directly applying iterative pruning with existing weightmagnitude-based structured pruning methods does not produce accurate pruned models.", "This paper proposes a novel LTH-based iterative, structured pruning solution using attentions, and it significantly outperforms ILP and other related structured pruning works that involve an iterative process (GDP #OTHEREFR , ACTD , Quantization and Pruning (QP) #OTHEREFR , IMP-Refill and IMP-Regroup #OTHEREFR ).", "The third challenge is how to automate the pruning process so it does not require any human intervention? The existing structured pruning works all require difficult hand-tuning of many hyper-parameters, e.g., DCP #OTHEREFR and MDP need multiple hyper-parameters to balance the original taskspecific loss and the additional pruning loss; VCNNP #OTHEREFR requires careful settings of \u03c4 and \u03b8 to decide which filters to prune; DMC #OTHEREFR and DeepHoyer ] require parameters to decide the regularization strength with different settings for different datasets and models.", "To address this challenge, this paper proposes a fully automated pruning solution that can automatically generate pruned models that meet users' diverse model accuracy, size, and speed requirements."], "citing_paper_content": {"title": "Automatic Attention Pruning: Improving And Automating Model Pruning Using Attentions", "abstract": "Pruning is a promising approach to compress deep learning models in order to deploy them on resource-constrained edge devices. However, many existing pruning solutions are based on unstructured pruning, which yields models that cannot efficiently run on commodity hardware; and they often require users to manually explore and tune the pruning process, which is timeconsuming and often leads to sub-optimal results. To address these limitations, this paper presents Automatic Attention Pruning (AAP), an adaptive, attention-based, structured pruning approach to automatically generate small, accurate, and hardware-efficient models that meet user objectives. First, it proposes iterative structured pruning using activation-based attention maps to effectively identify and prune unimportant filters. Then, it proposes adaptive pruning policies for automatically meeting the pruning objectives of accuracy-critical, memory-constrained, and latency-sensitive tasks. A comprehensive evaluation shows that AAP substantially outperforms the state-of-the-art structured pruning works for a variety of model architectures. Our code is at: https://github.com/kaiqi123/ Automatic-Attention-Pruning.git."}, "cited_paper_content": {"title": "Comparing Rewinding And Fine-Tuning In Neural Network Pruning", "abstract": "Many neural network pruning algorithms proceed in three steps: train the network to completion, remove unwanted structure to compress the network, and retrain the remaining structure to recover lost accuracy. The standard retraining technique, fine-tuning, trains the unpruned weights from their final trained values using a small fixed learning rate. In this paper, we compare fine-tuning to alternative retraining techniques. Weight rewinding (as proposed by Frankle et al., (2019)), rewinds unpruned weights to their values from earlier in training and retrains them from there using the original training schedule. Learning rate rewinding (which we propose) trains the unpruned weights from their final values using the same learning rate schedule as weight rewinding. Both rewinding techniques outperform fine-tuning, forming the basis of a network-agnostic pruning algorithm that matches the accuracy and compression ratios of several more network-specific state-of-the-art techniques."}, "keywords": ["structured pruning-Iterative L1-norm-based", "pruning"], "citation_intent": "background"}
{"citing_id": "2303.16053v1", "cited_id": "1902.07891", "section_title": "Benchmark Results On Mpeblink Dataset", "citation": "A similar conclusion has already been made in #REFR but the properties of multi-person and long video in MPEblink make it worse for landmark-based eyeblink detection methods.", "text_before_citation": ["\u2022 For Blink-AP, InstBlink significantly outperforms the others by large margins (i.e., 21% at least of Blink-AP 50 and 6% at least of Blink-AP 75 ), which verifies the superiority of the proposed framework.", "We argue that one essential reason is that our framework can model a better long-term temporal eyeblink representation than the frame-based method #OTHEREFR and the sliding window based methods #OTHEREFR .", "Moreover, under the proposed framework, eyeblink features can be facilitated via face's global context (e.g., head pose and illumination condition) with joint optimization and interaction, while previous works that utilize a sequential manner cannot.", "From Table 2 , it can also be summarized that the landmark-based methods #OTHEREFR perform poorer than the region-based counterparts.", "We think that one essential reason is that landmark detection is unreliable under unconstrained conditions."], "text_after_citation": ["Our method is region-based, but localizes eye region in an implicit way where global face context is included and thus becomes more robust.", "\u2022 InstBlink also outperforms others on Inst-AP.", "We think that is because our framework can better model the longterm spatio-temporal instance representations, while the counterparts achieve tracking under a tracking-by-detection framework, which contains limited spatio-temporal modeling and may suffer from heavy occlusion. Inference speed analysis.", "The result is listed in Table 3 , assuming the 4 compared methods use InsightFace for face & landmark detection and InstBlink inferences within a clip length of 36.", "It can be seen that InstBlink is also of high inference speed (i.e., 112 FPS for network forwarding) while the real-time capacity of other methods is not superior."], "citing_paper_content": {"title": "Real-Time Multi-Person Eyeblink Detection In The Wild For Untrimmed Video", "abstract": "Real-time eyeblink detection in the wild can widely serve for fatigue detection, face anti-spoofing, emotion analysis, etc. The existing research efforts generally focus on singleperson cases towards trimmed video. However, multiperson scenario within untrimmed videos is also important for practical applications, which has not been well concerned yet. To address this, we shed light on this research field for the first time with essential contributions on dataset, theory, and practices. In particular, a large-scale dataset termed MPEblink that involves 686 untrimmed videos with 8748 eyeblink events is proposed under multiperson conditions. The samples are captured from unconstrained films to reveal \"in the wild\" characteristics. Meanwhile, a real-time multi-person eyeblink detection method is also proposed. Being different from the existing counterparts, our proposition runs in a one-stage spatio-temporal way with end-to-end learning capacity. Specifically, it simultaneously addresses the sub-tasks of face detection, face tracking, and human instance-level eyeblink detection. This paradigm holds 2 main advantages: (1) eyeblink features can be facilitated via the face's global context (e.g., head pose and illumination condition) with joint optimization and interaction, and (2) addressing these sub-tasks in parallel instead of sequential manner can save time remarkably to meet the real-time running requirement. Experiments on MPEblink verify the essential challenges of real-time multiperson eyeblink detection in the wild for untrimmed video. Our method also outperforms existing approaches by large margins and with a high inference speed."}, "cited_paper_content": {"title": "Towards Real-Time Eyeblink Detection In The Wild:Dataset,Theory And Practices", "abstract": "Effective and real-time eyeblink detection is of wide-range applications, such as deception detection, drive fatigue detection, face anti-spoofing, etc. Although numerous of efforts have already been paid, most of them focus on addressing the eyeblink detection problem under the constrained indoor conditions with the relative consistent subject and environment setup. Nevertheless, towards the practical applications eyeblink detection in the wild is more required, and of greater challenges. However, to our knowledge this has not been well studied before. In this paper, we shed the light to this research topic. A labelled eyeblink in the wild dataset (i.e., HUST-LEBW) of 673 eyeblink video samples (i.e., 381 positives, and 292 negatives) is first established by us. These samples are captured from the unconstrained movies, with the dramatic variation on human attribute, human pose, illumination condition, imaging configuration, etc. Then, we formulate eyeblink detection task as a spatial-temporal pattern recognition problem. After locating and tracking human eye using SeetaFace engine and KCF tracker respectively, a modified LSTM model able to capture the multi-scale temporal information is proposed to execute eyeblink verification. A feature extraction approach that reveals appearance and motion characteristics simultaneously is also proposed. The experiments on HUST-LEBW reveal the superiority and efficiency of our approach. It also verifies that, the existing eyeblink detection methods cannot achieve satisfactory performance in the wild."}, "keywords": ["landmark-based eyeblink detection"], "citation_intent": "result"}
{"citing_id": "2304.14570v1", "cited_id": "1803.00188", "section_title": "Design Principles In Kaiaulu", "citation": "We chose to use the R language #REFR , due to the familiarity of the authors with the language and a preference for its package architecture.", "text_before_citation": ["In this section, we discuss how our design principles are translated into Kaiaulu's specific design decisions.", "In the following section, we fully flesh out Kaiaulu's modules and features.", "Batch Mode, Interactive Mode, and Literate Programming in Kaiaulu."], "text_after_citation": ["Minimally, the structure of an R package consists of the package metadata and its API.", "In addition, the R ecosystem encourages and promotes best practices to include documentation packages called vignettes, which leads R users to expect an API and R Notebooks when installing packages from CRAN (The Comprehensive R Archive Network).", "#OTHEREFR CRAN treats R Notebooks as first class citizens in an R package #OTHEREFR showing on each package's website any R Notebooks available.", "Because of R package structure, complying with familiar software abstractions (see Section 2.5) automatically brings the benefits of literate programming (see Section 2.3).", "Abstraction Debt in Kaiaulu."], "citing_paper_content": {"title": "Building The Msr Tool Kaiaulu: Design Principles And Experiences", "abstract": "Background: Since Alitheia Core was proposed and subsequently retired, tools that support empirical studies of software projects continue to be proposed, such as Codeface, Codeface4Smells, Grimoire-Lab and SmartSHARK, but they all make different design choices and provide overlapping functionality. Aims: We seek to understand the design decisions adopted by these tools-the good and the bad-along with their consequences, to understand why their authors reinvented functionality already present in other tools, and to help inform the design of future tools. Method: We used action research to evaluate the tools, and to determine a set of principles and anti-patterns to motivate a new tool design. Results: We identified 7 major design choices among the tools: 1) Abstraction Debt, 2) the use of Project Configuration Files, 3) the choice of Batch or Interactive Mode, 4) Minimal Paths to Data, 5) Familiar Software Abstractions, 6) Licensing and 7) the Perils of Code Reuse. Building on the observed good and bad design decisions, we created our own tool architecture and implemented it as an R package. Conclusions: Tools should not require onerous setup for users to obtain data. Authors should consider the conventions and abstractions used by their chosen language and build upon these instead of redefining them. Tools should encourage best practices in experiment reproducibility by leveraging self-contained and readable schemas that are used for tool automation, and reuse must be done with care to avoid depending on dead code."}, "cited_paper_content": {"title": "Xnmt: The Extensible Neural Machine Translation Toolkit", "abstract": "This paper describes XNMT, the eXtensible Neural Machine Translation toolkit. XNMT distin- guishes itself from other open-source NMT toolkits by its focus on modular code design, with the purpose of enabling fast iteration in research and replicable, reliable results. In this paper we describe the design of XNMT and its experiment configuration system, and demonstrate its utility on the tasks of machine translation, speech recognition, and multi-tasked machine translation/parsing. XNMT is available open-source at this https URL"}, "keywords": ["R language"], "citation_intent": "method"}
{"citing_id": "2304.03481v1", "cited_id": "1902.10811", "section_title": "-Number Of Branches In The Ladder Self-Attention Block.", "citation": "The top-1 accuracy of our PSLT and DeiT-S on both the ImageNet and ImageNet-V2 validation set conforms to the linear fit in #REFR .", "text_before_citation": ["-Stacking the proposed modules.", "As described in Section 3.1, our proposed PSLT is composed of the proposed ladder self-attention block, Light FFN and the pixel-adaptive fusion module.", "Table 15 shows the performance of models by stacking the proposed modules one by one, which demonstrates the effectiveness of the modules.", "Compared to the ViT, although applying our ladder self-attetion (LSA) and Light FFN (LFFN) does not improve the performance, the number of parameters and FLOPs are largely reduced, where the parameters have been reduced by 9 times and 15 times respectively, and the FLOPs have been reduced by 27 times and 37 times respectively.", "The pixel-adaptive fusion module (PAFM) improves the model performance by integrating information of all branches in the block."], "text_after_citation": ["Our PSLT (including DeiT-S) achieves smaller improvement over ViT-B on ImageNet-V2 validation (2% and 1% top-1 accuracy improvement on ImageNet and ImageNet-V2 validation set, respectively).", "Such a phenomenon has been similarly observed in #OTHEREFR , #OTHEREFR .", "As described in #OTHEREFR , this shows the model generalization ability is challenged on different validation set."], "citing_paper_content": {"title": "Pslt: A Light-Weight Vision Transformer With Ladder Self-Attention And Progressive Shift", "abstract": "Vision Transformer (ViT) has shown great potential for various visual tasks due to its ability to model long-range dependency. However, ViT requires a large amount of computing resource to compute the global self-attention. In this work, we propose a ladder self-attention block with multiple branches and a progressive shift mechanism to develop a lightweight transformer backbone that requires less computing resources (e.g. a relatively small number of parameters and FLOPs), termed Progressive Shift Ladder Transformer (PSLT). First, the ladder self-attention block reduces the computational cost by modelling local self-attention in each branch. In the meanwhile, the progressive shift mechanism is proposed to enlarge the receptive field in the ladder self-attention block by modelling diverse local self-attention for each branch and interacting among these branches. Second, the input feature of the ladder self-attention block is split equally along the channel dimension for each branch, which considerably reduces the computational cost in the ladder self-attention block (with nearly 1 3 the amount of parameters and FLOPs), and the outputs of these branches are then collaborated by a pixel-adaptive fusion. Therefore, the ladder self-attention block with a relatively small number of parameters and FLOPs is capable of modelling long-range interactions. Based on the ladder self-attention block, PSLT performs well on several vision tasks, including image classification, objection detection and person re-identification. On the ImageNet-1k dataset, PSLT achieves a top-1 accuracy of 79.9% with 9.2M parameters and 1.9G FLOPs, which is comparable to several existing models with more than 20M parameters and 4G FLOPs. Code is available at https://isee-ai.cn/wugaojie/PSLT.html."}, "cited_paper_content": {"title": "Do Imagenet Classifiers Generalize To Imagenet?", "abstract": "We build new test sets for the CIFAR-10 and ImageNet datasets. Both benchmarks have been the focus of intense research for almost a decade, raising the danger of overfitting to excessively re-used test sets. By closely following the original dataset creation processes, we test to what extent current classification models generalize to new data. We evaluate a broad range of models and find accuracy drops of 3% - 15% on CIFAR-10 and 11% - 14% on ImageNet. However, accuracy gains on the original test sets translate to larger gains on the new test sets. Our results suggest that the accuracy drops are not caused by adaptivity, but by the models' inability to generalize to slightly \"harder\" images than those found in the original test sets."}, "keywords": ["top-1 accuracy"], "citation_intent": "result"}
{"citing_id": "2303.17857v1", "cited_id": "2002.02445", "section_title": "I. Introduction", "citation": "A framework used for dataset generation was also proposed for cooperatively exploiting both visual and wireless data #REFR .", "text_before_citation": ["The traditional vision-based object tracking algorithms, such as the sparse representation and correlation filtering, have difficulty in accurately locating high-mobility UEs in real-time #OTHEREFR .", "Furthermore, in complex environments, tracking multiple UEs in the face of blockage and uneven light, the localization accuracy of UEs tends to degrade significantly.", "As a result, the traditional CV-related ML algorithms cannot satisfy the high location accuracy required by mmWave communications.", "Finally, gathering abundant labelled data from real-world environments including both visual data and wireless signals to train ML models is still challenging.", "Nevertheless, vision-assisted beam management methods that predict the optimal mmWave beams have been investigated in the last few years."], "text_after_citation": ["However, these methods may still be plagued by a number of issues:", "\u2022 The robustness of the existing ML models for visionassisted beam management has to be improved #OTHEREFR .", "When the classical image classification models designed for prediction are used for mmWave link blockage prediction and beam prediction, the target accuracy cannot be always satisfied, for example due to the over-fitting issues.", "In this context, a preliminary study was conducted in #OTHEREFR by relying on a simple dataset, for investigating visionassisted beam management in multi-user scenarios.", "\u2022 The scalability of the methods is not guaranteed in practical scenarios."], "citing_paper_content": {"title": "Vision-Assisted Mmwave Beam Management For Next-Generation Wireless Systems: Concepts, Solutions And Open Challenges", "abstract": "Beamforming techniques have been widely used in the millimeter wave (mmWave) bands to mitigate the path loss of mmWave radio links as the narrow straight beams by directionally concentrating the signal energy. However, traditional mmWave beam management algorithms usually require excessive channel state information overhead, leading to extremely high computational and communication costs. This hinders the widespread deployment of mmWave communications. By contrast, the revolutionary vision-assisted beam management system concept employed at base stations (BSs) can select the optimal beam for the target user equipment (UE) based on its location information determined by machine learning (ML) algorithms applied to visual data, without requiring channel information. In this paper, we present a comprehensive framework for a vision-assisted mmWave beam management system, its typical deployment scenarios as well as the specifics of the framework. Then, some of the challenges faced by this system and their efficient solutions are discussed from the perspective of ML. Next, a new simulation platform is conceived to provide both visual and wireless data for model validation and performance evaluation. Our simulation results indicate that the vision-assisted beam management is indeed attractive for next-generation wireless systems."}, "cited_paper_content": {"title": "Viwi Vision-Aided Mmwave Beam Tracking: Dataset, Task, And Baseline Solutions", "abstract": "Vision-aided wireless communication is motivated by the recent advances in deep learning and computer vision as well as the increasing dependence on line-of-sight links in millimeter wave (mmWave) and terahertz systems. By leveraging vision, this new research direction enables an interesting set of new capabilities such as vision-aided mmWave beam and blockage prediction, proactive hand-off, and resource allocation among others. These capabilities have the potential of reliably supporting highly-mobile applications such as vehicular/drone communications and wireless virtual/augmented reality in mmWave and terahertz systems. Investigating these interesting applications, however, requires the development of special dataset and machine learning tasks. Based on the Vision-Wireless (ViWi) dataset generation framework [1], this paper develops an advanced and realistic scenario/dataset that features multiple base stations, mobile users, and rich dynamics. Enabled by this dataset, the paper defines the vision-wireless mmWave beam tracking task (ViWi-BT) and proposes a baseline solution that can provide an initial benchmark for the future ViWi-BT algorithms."}, "keywords": ["wireless data"], "citation_intent": "method"}
{"citing_id": "2303.07200v1", "cited_id": "1907.12207", "section_title": "Objective Function:", "citation": "Additionally, it is important that function f that can learn a fruitful representation and complex data dependencies #REFR .", "text_before_citation": ["EQUATION", "where F * s is the final selected feature set, J is a desired loss function, and f (x", "(i)", "Fs ; \u03b8) is a classification function parameterized by \u03b8 aiming at estimating the target for the i-th sample using a subset of features x (i) Fs . Solving this optimization problem can be a challenging task.", "As the choice of feature subset F s grows exponentially with increasing number of features d, solving Equation 1 is a NP-hard problem."], "text_after_citation": ["We choose artificial neural networks due to their high expressive power; a simple one-hidden layer feed-forward neural network is known to be a universal approximator #OTHEREFR .", "Finally, as we aim to select features in a computationally efficient manner, in this paper, we choose sparse neural networks to represent the data and perform feature selection."], "citing_paper_content": {"title": "Supervised Feature Selection With Neuron Evolution In Sparse Neural Networks", "abstract": "Feature selection that selects an informative subset of variables from data not only enhances the model interpretability and performance but also alleviates the resource demands. Recently, there has been growing attention on feature selection using neural networks. However, existing methods usually suffer from high computational costs when applied to high-dimensional datasets. In this paper, inspired by evolution processes, we propose a novel resource-efficient supervised feature selection method using sparse neural networks, named \"NeuroFS\". By gradually pruning the uninformative features from the input layer of a sparse neural network trained from scratch, NeuroFS derives an informative subset of features efficiently. By performing several experiments on 11 low and high-dimensional real-world benchmarks of different types, we demonstrate that NeuroFS achieves the highest ranking-based score among the considered state-of-the-art supervised feature selection models. The code is available on GitHub 1 ."}, "cited_paper_content": {"title": "Lassonet: Neural Networks With Feature Sparsity", "abstract": "We propose a neural network model, with a separate linear (residual) term, that explicitly bounds the input layer weights for a feature by the linear weight for that feature. The model can be seen as a modification of so-called residual neural networks to produce a path of models that are feature-sparse, that is, use only a subset of the features. This is analogous to the solution path from the usual Lasso ($\\ell_1$-regularized) linear regression. We call the proposed procedure {\\tt LassoNet} and develop a projected proximal gradient algorithm for its optimization. This approach can sometimes give as low or lower test error than a standard neural network, and its feature selection provides more interpretable solutions. We illustrate the method using both simulated and real data examples, and show that it is often able to achieve competitive performance with a much smaller number of input feature"}, "keywords": ["fruitful representation"], "citation_intent": "background"}
{"citing_id": "2304.04911v1", "cited_id": "1904.12901", "section_title": "D. Problem Formulation", "citation": "The assumption of the MDP does not hold on most realworld systems due to delays and hence those maybe only partially observable #REFR .", "text_before_citation": ["In this force trajectory tracking problem, the trajectories are restricted to sinusoidal waveforms between 0.05 to 0.35 Hz.", "This is because (as discussed in Section III-B) the input motor current is saturated to constrain the pendulum within safe position bounds and avoid any damage to the ball screw during learning.", "In addition, the initial random policy with abruptly switching high current values can induce vibrationbased damage to the system."], "text_after_citation": ["This problem is often countered by adding more information in the observation space of the agent.", "The observation vector includes the following: the absolute encoder-based angle, q [rad], angular velocity,q", "EQUATION", "where the divisor, D = 10 6 , maintains an episodic reward between [-30, 0] .", "In absence of this, the summation of reward per time step was observed to undergo catastrophic forgetting #OTHEREFR resulting in unlearning and poor performance."], "citing_paper_content": {"title": "Real-Time Model-Free Deep Reinforcement Learning For Force Control Of A Series Elastic Actuator", "abstract": "Many state-of-the art robotic applications utilize series elastic actuators (SEAs) with closed-loop force control to achieve complex tasks such as walking, lifting, and manipulation. Model-free PID control methods are more prone to instability due to nonlinearities in the SEA where cascaded model-based robust controllers can remove these effects to achieve stable force control. However, these model-based methods require detailed investigations to characterize the system accurately. Deep reinforcement learning (DRL) has proved to be an effective model-free method for continuous control tasks, where few works deal with hardware learning. This paper describes the training process of a DRL policy on hardware of an SEA pendulum system for tracking force control trajectories from 0.05-0.35 Hz at 50 N amplitude using the Proximal Policy Optimization (PPO) algorithm. Safety mechanisms are developed and utilized for training the policy for 12 hours (overnight) without an operator present within the full 21 hours training period. The tracking performance is evaluated showing improvements of 25 N in mean absolute error when comparing the first 18 min. of training to the full 21 hours for a 50 N amplitude, 0.1 Hz sinusoid desired force trajectory. Finally, the DRL policy exhibits better tracking and stability margins when compared to a model-free PID controller for a 50 N chirp force trajectory."}, "cited_paper_content": {"title": "Challenges Of Real-World Reinforcement Learning", "abstract": "Reinforcement learning (RL) has proven its worth in a series of artificial domains, and is beginning to show some successes in real-world scenarios. However, much of the research advances in RL are often hard to leverage in real-world systems due to a series of assumptions that are rarely satisfied in practice. We present a set of nine unique challenges that must be addressed to productionize RL to real world problems. For each of these challenges, we specify the exact meaning of the challenge, present some approaches from the literature, and specify some metrics for evaluating that challenge. An approach that addresses all nine challenges would be applicable to a large number of real world problems. We also present an example domain that has been modified to present these challenges as a testbed for practical RL research."}, "keywords": ["realworld systems"], "citation_intent": "background"}
{"citing_id": "2304.05405v2", "cited_id": "2001.00326", "section_title": "Sdarts.", "citation": "E 2 NAS successfully overperformed previous works on the three datasets available in NAS-Bench-201 #REFR (e.g., a +29.45 % top 1 accuracy improvement on ImageNet-16-120).", "text_before_citation": ["EQUATION", "where U is a function updating the hidden state \u210e obtained by aggregating all its predecessors with function G.", "Since", "EQUATION", "where L is the Cross-Entropy loss, R is a 2 regularization term, is a value that balances between optimizing the current architecture (exploitation) or preventing other alternatives from vanishing (exploration)."], "text_after_citation": ["Zhang et al.", "#OTHEREFR proposed iDARTS, a solution that reformulates the optimization process of DARTS with a Neumann-approximation of the Implicit Theorem Function (IFT) #OTHEREFR . Concretely, the architectural parameter gradients \u0394 L (see Eq. 3) are calculated as follows:", "EQUATION", "where L is the validation loss, L is the training loss, and are the network weights.", "However, it is computationally intensive to compute the inverse of the Hessian matrix"], "citing_paper_content": {"title": "Efficient Automation Of Neural Network Design: A Survey On Differentiable Neural Architecture Search", "abstract": "In the past few years, Differentiable Neural Architecture Search (DNAS) rapidly imposed itself as the trending approach to automate the discovery of deep neural network architectures. This rise is mainly due to the popularity of DARTS, one of the first major DNAS methods. In contrast with previous works based on Reinforcement Learning or Evolutionary Algorithms, DNAS is faster by several orders of magnitude and uses fewer computational resources. In this comprehensive survey, we focus specifically on DNAS and review recent approaches in this field. Furthermore, we propose a novel challenge-based taxonomy to classify DNAS methods. We also discuss the contributions brought to DNAS in the past few years and its impact on the global NAS field. Finally, we conclude by giving some insights into future research directions for the DNAS field. CCS Concepts: \u2022 General and reference \u2192 Surveys and overviews; \u2022 Computing methodologies \u2192 Search methodologies; Computer vision."}, "cited_paper_content": {"title": "Nas-Bench-201: Extending The Scope Of Reproducible Neural Architecture Search", "abstract": "Neural architecture search (NAS) has achieved breakthrough success in a great number of applications in the past few years. It could be time to take a step back and analyze the good and bad aspects in the field of NAS. A variety of algorithms search architectures under different search space. These searched architectures are trained using different setups, e.g., hyper-parameters, data augmentation, regularization. This raises a comparability problem when comparing the performance of various NAS algorithms. NAS-Bench-101 has shown success to alleviate this problem. In this work, we propose an extension to NAS-Bench-101: NAS-Bench-102 with a different search space, results on multiple datasets, and more diagnostic information. NAS-Bench-102 has a fixed search space and provides a unified benchmark for almost any up-to-date NAS algorithms. The design of our search space is inspired by the one used in the most popular cell-based searching algorithms, where a cell is represented as a directed acyclic graph. Each edge here is associated with an operation selected from a predefined operation set. For it to be applicable for all NAS algorithms, the search space defined in NAS-Bench-102 includes all possible architectures generated by 4 nodes and 5 associated operation options, which results in 15,625 neural cell candidates in total. The training log using the same setup and the performance for each architecture candidate are provided for three datasets. This allows researchers to avoid unnecessary repetitive training for selected architecture and focus solely on the search algorithm itself. The training time saved for every architecture also largely improves the efficiency of most NAS algorithms and presents a more computational cost friendly NAS community for a broader range of researchers. We provide additional diagnostic information such as fine-grained loss and accuracy, which can give inspirations to new designs of NAS algorithms. In further support of the proposed NAS-Bench-102, we have analyzed it from many aspects and benchmarked 10 recent NAS algorithms, which verify its applicability."}, "keywords": ["NAS", "E 2 NAS"], "citation_intent": "result"}
{"citing_id": "2304.08756v2", "cited_id": "1801.08297", "section_title": "Related Work", "citation": "NDDR-CNN #REFR proposes a novel CNN structure to learn a discriminative feature embedding for each task.", "text_before_citation": ["As we will explain in the next section, directly extending the one-shot weight-sharing strategy from single-task NAS to multi-task NAS incurs interferences among tasks, leading to significant performance degradation.", "AutoTaskFormer tackles this problem using a novel NAS pipeline that involves two search stages instead of one, specifically designed for multi-task learning.", "Multi-task Learning (MTL). MTL aims to learn parameter sharing across multiple tasks.", "For example, Cross-Stitch #OTHEREFR manually employs an additional group of shared units to merge multiple task backbones.", "Sluice #OTHEREFR jointly learns a latent multi-task architecture and task-specific models."], "text_after_citation": ["There are several works that design task-sharing architectures through NAS.", "In particular, MTL-NAS #OTHEREFR starts with two taskspecific networks and seeks optimal edges between the intertask branches.", "AdaShare #OTHEREFR learns the sharing weights through a task-specific policy that selectively chooses which layers to execute for a given task.", "AutoMTL #OTHEREFR proposes a source-to-source compiler that transforms the backbone CNN into a supermodel.", "However, due to the bottom-up approach of combining task-specific networks, the sharing between tasks and networks needs to be meticulously designed and hence can only support a small number of tasks."], "citing_paper_content": {"title": "Autotaskformer: Searching Vision Transformers For Multi-Task Learning", "abstract": "Vision Transformers have shown great performance in single tasks such as classification and segmentation. However, real-world problems are not isolated, which calls for vision transformers that can perform multiple tasks concurrently. Existing multi-task vision transformers are handcrafted and heavily rely on human expertise. In this work, we propose a novel one-shot neural architecture search framework, dubbed AutoTaskFormer (Automated Multi-Task Vision TransFormer), to automate this process. AutoTask-Former not only identifies the weights to share across multiple tasks automatically, but also provides thousands of well-trained vision transformers with a wide range of parameters (e.g., number of heads and network depth) for deployment under various resource constraints. Experiments on both small-scale (2-task Cityscapes and 3-task NYUv2) and large-scale (16-task Taskonomy) datasets show that Auto-TaskFormer outperforms state-of-the-art handcrafted vision transformers in multi-task learning. The entire code and models will be open-sourced."}, "cited_paper_content": {"title": "Nddr-Cnn: Layerwise Feature Fusing In Multi-Task Cnns By Neural Discriminative Dimensionality Reduction", "abstract": "In this paper, we propose a novel Convolutional Neural Network (CNN) structure for general-purpose multi-task learning (MTL), which enables automatic feature fusing at every layer from different tasks. This is in contrast with the most widely used MTL CNN structures which empirically or heuristically share features on some specific layers (e.g., share all the features except the last convolutional layer). The proposed layerwise feature fusing scheme is formulated by combining existing CNN components in a novel way, with clear mathematical interpretability as discriminative dimensionality reduction, which is referred to as Neural Discriminative Dimensionality Reduction (NDDR). Specifically, we first concatenate features with the same spatial resolution from different tasks according to their channel dimension. Then, we show that the discriminative dimensionality reduction can be fulfilled by 1x1 Convolution, Batch Normalization, and Weight Decay in one CNN. The use of existing CNN components ensures the end-to-end training and the extensibility of the proposed NDDR layer to various state-of-the-art CNN architectures in a \"plug-and-play\" manner. The detailed ablation analysis shows that the proposed NDDR layer is easy to train and also robust to different hyperparameters. Experiments on different task sets with various base network architectures demonstrate the promising performance and desirable generalizability of our proposed method. The code of our paper is available at https://github.com/ethanygao/NDDR-CNN."}, "keywords": ["novel CNN structure"], "citation_intent": "background"}
{"citing_id": "2305.00455v1", "cited_id": "1706.03762", "section_title": "Causal Semantics Extractor", "citation": "In #REFR , the vanilla attention matrix is based on the calculation of all the query-key pairs.", "text_before_citation": ["According to #OTHEREFR , the computation of the vanilla attention matrix A \u2208 R n\u00d7n is based on the dot-product.", "It is defined as A = softmax QK \u221a d ; Q = TW q , K = TW k , where the query matrix Q \u2208 R n\u00d7d and key matrix K \u2208 R n\u00d7d are generated by the linear projection of the input token matrix T \u2208 R n\u00d7dm based on the learnable weights matrices W q \u2208 R dm\u00d7d and W k \u2208 R dm\u00d7d . n indicates the total number of input tokens.", "d represents the embedding dimension and d m denotes the dimension of an input token.", "The new value matrix V new \u2208 R n\u00d7d can be obtained via", "V new = A V; V = TW v , where the value matrix V \u2208 R n\u00d7d and W v \u2208 R dm\u00d7d ."], "text_after_citation": ["However, in the proposed Causal Semantics Extractor, only the top \u03ba most similar keys and values for each query are used to compute the causal attention matrix.", "Similar to #OTHEREFR , all the queries and keys are calculated by the dot-product.", "Then, the row-wise top \u03ba elements are used for the softmax calculation.", "In the proposed Causal Semantics Extractor, the value matrix #OTHEREFR , for the input video, and VisualAtten(\u2022) indicates a visual attention mechanism based on the element-wise multiplication.", "X mul = FC(Z ta Z va ), where denotes the operation of feature concatenation and FC(\u2022) indicates a fully connected layer."], "citing_paper_content": {"title": "Causalainer: Causal Explainer For Automatic Video Summarization", "abstract": "The goal of video summarization is to automatically shorten videos such that it conveys the overall story without losing relevant information. In many application scenarios, improper video summarization can have a large impact. For example in forensics, the quality of the generated video summary will affect an investigator's judgment while in journalism it might yield undesired bias. Because of this, modeling explainability is a key concern. One of the best ways to address the explainability challenge is to uncover the causal relations that steer the process and lead to the result. Current machine learning-based video summarization algorithms learn optimal parameters but do not uncover causal relationships. Hence, they suffer from a relative lack of explainability. In this work, a Causal Explainer, dubbed Causalainer, is proposed to address this issue. Multiple meaningful random variables and their joint distributions are introduced to characterize the behaviors of key components in the problem of video summarization. In addition, helper distributions are introduced to enhance the effectiveness of model training. In visual-textual input scenarios, the extra input can decrease the model performance. A causal semantics extractor is designed to tackle this issue by effectively distilling the mutual information from the visual and textual inputs. Experimental results on commonly used benchmarks demonstrate that the proposed method achieves state-of-the-art performance while being more explainable."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["vanilla attention matrix"], "citation_intent": "method"}
{"citing_id": "2304.08597v1", "cited_id": "1902.00532", "section_title": "Key Limitations", "citation": "Time Constrained approaches have been proposed that limited the amount of time available for the AutoML system to search through the space, however, this came with a trade-off -balancing between time and pipeline accuracy #REFR . Thus the resulting pipeline was mostly sub-optimal.", "text_before_citation": ["While the iterative search (such as grid search or random search) has shown success, however, the search space explodes when pipelines contain many steps with an even greater number of choices.", "This explosion in the number of candidate pipelines is prohibitive of compute time and cost of resources needed to support searching through large hyper-parameter search space.", "The significant time cost of AutoML systems comes from evaluating the end-to-end pipeline.", "As the quality of the pipeline is only measured after the completion of the last step."], "text_after_citation": [], "citing_paper_content": {"title": "Etop: Early Termination Of Pipelines For Faster Training Of Automl System", "abstract": "Recent advancements in software and hardware technologies have enabled the use of AI/ML models in everyday applications has significantly improved the quality of service rendered. However, for a given application, finding the right AI/ML model is a complex and costly process, that involves the generation, training, and evaluation of multiple interlinked steps (called pipelines), such as data pre-processing, feature engineering, selection, and model tuning. These pipelines are complex (in structure) and costly (both in compute resource and time) to execute end-to-end, with a hyper-parameter associated with each step. Au-toML systems automate the search of these hyper-parameters but are slow, as they rely on optimizing the pipelines' end output. We propose \"the eTOP Framework\" which works on top of any AutoML system and decides whether or not to execute the pipeline to the end or terminate at an intermediate step. Experimental evaluation on 26 benchmark datasets 3 and integration of eTOPwith MLBox 4 reduces the training time of the AutoML system upto 40x than baseline MLBox."}, "cited_paper_content": {"title": "Hyper-Parameter Tuning Under A Budget Constraint", "abstract": "We study a budgeted hyper-parameter tuning problem, where we optimize the tuning result under a hard resource constraint. We propose to solve it as a sequential decision making problem, such that we can use the partial training progress of configurations to dynamically allocate the remaining budget. Our algorithm combines a Bayesian belief model which estimates the future performance of configurations, with an action-value function which balances exploration-exploitation tradeoff, to optimize the final output. It automatically adapts the tuning behaviors to different constraints, which is useful in practice. Experiment results demonstrate superior performance over existing algorithms, including the-state-of-the-art one, on real-world tuning tasks across a range of different budgets."}, "keywords": ["AutoML system", "trade-off -balancing"], "citation_intent": "background"}
{"citing_id": "2303.10135v2", "cited_id": "1706.03762", "section_title": "B. Assembly Graphs", "citation": "Both the part-id and surface-id fields are encoded with a d-dimensional Sinusoidal Positional Encoding #REFR .", "text_before_citation": ["N, part-id \u2208 R d ].", "There are three atomic part types: long profile, short profile and angle bracket.", "2) Surface Nodes: Different to the one in #OTHEREFR , we associate each surface node", "v s i \u2208 V s with the features \u03c6 (v s i ) = [surface-type \u2208 N, surface-id \u2208 R d ].", "There are two surface types (long and short) for profiles and one (lateral) for brackets."], "text_after_citation": ["3) Surface-to-Surface Edges: We design a fullyconnected graph for all surface nodes V s to capture the relation between untouched surfaces, which is more fine-grained than those in #OTHEREFR with only connects between touched surfaces.", "These edges are assigned with a feature \u03c6 (e i ) \u2208 R, indicating the relation between the two surfaces: \u03c6 (e i ) = relative distance (parallel); 1 (belong to the same part); \u22121 (orthogonal); 0 (same-surface loop).", "4) Surface-to-Part Edges: These connect each surface and part node pair", "(v s i , v p j ) \u2208 V s \u00d7 V p , where surface v s i belongs to the part v p", "j ."], "citing_paper_content": {"title": "Efficient And Feasible Robotic Assembly Sequence Planning Via Graph Representation Learning", "abstract": "Automatic Robotic Assembly Sequence Planning (RASP) can significantly improve productivity and resilience in modern manufacturing along with the growing need for greater product customization. One of the main challenges in realizing such automation resides in efficiently finding solutions from a growing number of potential sequences for increasingly complex assemblies. Besides, costly feasibility checks are always required for the robotic system. To address this, we propose a holistic graphical approach including a graph representation called Assembly Graph for product assemblies and a policy architecture, Graph Assembly Processing Network, dubbed GRACE for assembly sequence generation. Secondly, we use GRACE to extract meaningful information from the graph input and predict assembly sequences in a step-by-step manner. In experiments, we show that our approach can predict feasible assembly sequences across product variants of aluminum profiles based on data collected in simulation of a dual-armed robotic system. We further demonstrate that our method is capable of detecting infeasible assemblies, substantially alleviating the undesirable impacts from false predictions, and hence facilitating real-world deployment soon. Code and training data will be open-sourced. Equal Contribution."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["part-id", "Encoding"], "citation_intent": "method"}
{"citing_id": "2303.05183v1", "cited_id": "1912.01703", "section_title": "Implementation Details", "citation": "All models are trained using Python 3.10.4, Pytorch 1.11.0 #REFR , and an Nvidia Tesla V100 GPU.", "text_before_citation": ["The learning rate for the noise estimator decreases by half every 10 epochs with 50 epochs trained, while the learning rate for the denoising network is halved every 20 epochs with 100 epochs trained.", "As for the hyper-parameter in the adaptive re-visible loss, we set \u03bb = 3 as the initial value and progressively increase it to 11.", "The denoising network and noise estimation are jointly optimized during training.", "The patches of size 128 \u00d7 128 are randomly cropped for training.", "Note that the details of the global masker and global mask mapper remain the same as Blind2Unblind #OTHEREFR ."], "text_after_citation": ["Datasets We consider five types of Poisson-Gaussian noise in synthetic noise estimation: (1) P G1 : \u03b1 = 0.1, \u03c3 = 0.02, (2) P G2 : \u03b1 = 0.1, \u03c3 = 0.0002, (3) P G3 : \u03b1 = 0.05, \u03c3 = 0.02, (4) P G4 : \u03b1 = 0.05, \u03c3 = 0.0002, (5) P G5 : \u03b1 = 0.01, \u03c3 = 0.02.", "For grayscale images, we use BSD400 #OTHEREFR for the training set, and for sRGB images, we use CBSD432 #OTHEREFR .", "The noise levels are estimated on standard BSD68 #OTHEREFR and CBSD68 #OTHEREFR for grayscale and sRGB images, respectively.", "For synthetic denoising, we use ILSVRC2012 #OTHEREFR validation set for sRGB image denoising and BSD400 #OTHEREFR for grayscale image denoising.", "Specifically, following the setting in #OTHEREFR , we select 44328 images with sizes between 256 \u00d7 256 and 512\u00d7512 pixels from ILSVRC2012 validation set for training."], "citing_paper_content": {"title": "Blind2Sound: Self-Supervised Image Denoising Without Residual Noise", "abstract": "Self-supervised blind denoising for Poisson-Gaussian noise remains a challenging task. Pseudo-supervised pairs constructed from single noisy images re-corrupt the signal and degrade the performance. The visible blindspots solve the information loss in masked inputs. However, without explicitly noise sensing, mean square error as an objective function cannot adjust denoising intensities for dynamic noise levels, leading to noticeable residual noise. In this paper, we propose Blind2Sound, a simple yet effective approach to overcome residual noise in denoised images. The proposed adaptive re-visible loss senses noise levels and performs personalized denoising without noise residues while retaining the signal lossless. The theoretical analysis of intermediate medium gradients guarantees stable training, while the Cramer Gaussian loss acts as a regularization to facilitate the accurate perception of noise levels and improve the performance of the denoiser. Experiments on synthetic and real-world datasets show the superior performance of our method, especially for single-channel images. The code is available in supplementary materials."}, "cited_paper_content": {"title": "Pytorch: An Imperative Style, High-Performance Deep Learning Library", "abstract": "Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. ::: In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. ::: We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks."}, "keywords": ["Pytorch"], "citation_intent": "method"}
{"citing_id": "2303.04654v1", "cited_id": "1608.03983", "section_title": "Implementation Details", "citation": "We utilize the AdamW optimizer [20] and the CosineAnnealing learning rate scheduler #REFR with their default parameters.", "text_before_citation": ["The FlyingThings3D dataset #OTHEREFR is used as the training dataset, which contains 800 pairs of synthetic RGBD images.", "The Middle-bury2014 dataset #OTHEREFR is used as the testing dataset, which contains 23 real-world RGBD images.", "There is a significant domain gap between the synthetic and real-world images, making it suitable for evaluating the generalizability of the pre-trained DfF models.", "We simulate focal stacks with the RGBD images for training and testing, using a stack size of #OTHEREFR The lens data comes from lensnet.com #OTHEREFR 10.", "The focus distances are chosen linearly from the minimum (20 cm) and maximum (20 m) depth range of each image, randomly perturbed."], "text_after_citation": ["The training batch size is 16, and the initial learning rate is 1e \u22124 .", "Each DfF model is trained for 400 epochs on a single A100 GPU.", "Following training, we evaluate each model directly on the testing focal stacks without any fine-tuning.", "In real-world experiments, we use the Canon EOS R camera and the RF 50mm F/1.8 lens discussed in Section 4.", "The training procedure keeps the same as the previous experiment."], "citing_paper_content": {"title": "Aberration-Aware Depth-From-Focus", "abstract": "Computer vision methods for depth estimation usually use simple camera models with idealized optics. For modern machine learning approaches, this creates an issue when attempting to train deep networks with simulated data, especially for focus-sensitive tasks like Depth-from-Focus. In this work, we investigate the domain gap caused by off-axis aberrations that will affect the decision of the best-focused frame in a focal stack. We then explore bridging this domain gap through aberration-aware training (AAT). Our approach involves a lightweight network that models lens aberrations at different positions and focus distances, which is then integrated into the conventional network training pipeline. We evaluate the generality of pretrained models on both synthetic and real-world data. Our experimental results demonstrate that the proposed AAT scheme can improve depth estimation accuracy without fine-tuning the model or modifying the network architecture. Our code and models will be made publicly available."}, "cited_paper_content": {"title": "Sgdr: Stochastic Gradient Descent With Warm Restarts", "abstract": "Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets, where we demonstrate new state-of-the-art results at 3.14% and 16.21%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at https://github.com/loshchil/SGDR"}, "keywords": ["CosineAnnealing learning rate"], "citation_intent": "method"}
{"citing_id": "2303.08431v2", "cited_id": "1801.05039", "section_title": "Introduction", "citation": "In the seminal work of #REFR , the authors studied an LQR problem with deterministic dynamics over an infinite horizon.", "text_before_citation": ["In order to maximize the accumulated reward over time, the agent learns to select actions based on past experiences (exploitation) and by making new choices (exploration).", "In recent years, we have witnessed successful development of RL systems in various applications, including robotics control #OTHEREFR , AlphaGo and Atari games #OTHEREFR , autonomous driving #OTHEREFR , and stock trading #OTHEREFR .", "Despite its practical success, theoretical understanding of RL is still limited and at its primitive stage.", "To establish a better foundation of RL, there has been a surge of theoretical works in recent years on the Linear Quadratic Regulator (LQR) problem.", "This problem is a special class of control problems with linear dynamics and quadratic cost functions #OTHEREFR ."], "text_after_citation": ["They proved that the simple policy gradient method converges to the globally optimal solution with a linear rate (despite nonconvexity of the objective).", "Their key idea is to utilize the Riccati equation (an algebraic-equation characterization that only works for LQR problems) and show that the cost function enjoys a \"gradient dominance\" property.", "This result has been extended to other settings such as linear dynamics with additive or multiplicative Gaussian noise, finite-time horizon, and modifications of the vanilla policy-gradient method in follow-up works #OTHEREFR .", "Other aspects in the learning of LQR, such as the trade-off between exploration and exploitation, have also been studied recently #OTHEREFR .", "Despite the desirable theoretical properties of LQR, this setting is limited in practice due to the nonlinear nature of many real-world dynamic systems."], "citing_paper_content": {"title": "Policy Gradient Converges To The Globally Optimal Policy For Nearly Linear-Quadratic Regulators", "abstract": "Nonlinear control systems with partial information to the decision maker are prevalent in a variety of applications. As a step toward studying such nonlinear systems, this work explores reinforcement learning methods for finding the optimal policy in the nearly linear-quadratic regulator systems. In particular, we consider a dynamic system that combines linear and nonlinear components, and is governed by a policy with the same structure. Assuming that the nonlinear component comprises kernels with small Lipschitz coefficients, we characterize the optimization landscape of the cost function. Although the cost function is nonconvex in general, we establish the local strong convexity and smoothness in the vicinity of the global optimizer. Additionally, we propose an initialization mechanism to leverage these properties. Building on the developments, we design a policy gradient algorithm that is guaranteed to converge to the globally optimal policy with a linear rate."}, "cited_paper_content": {"title": "Global Convergence Of Policy Gradient Methods For The Linear Quadratic Regulator", "abstract": "Direct policy gradient methods for reinforcement learning and continuous control problems are a popular approach for a variety of reasons: 1) they are easy to implement without explicit knowledge of the underlying model 2) they are an \"end-to-end\" approach, directly optimizing the performance metric of interest 3) they inherently allow for richly parameterized policies. A notable drawback is that even in the most basic continuous control problem (that of linear quadratic regulators), these methods must solve a non-convex optimization problem, where little is understood about their efficiency from both computational and statistical perspectives. In contrast, system identification and model based planning in optimal control theory have a much more solid theoretical footing, where much is known with regards to their computational and statistical properties. This work bridges this gap showing that (model free) policy gradient methods globally converge to the optimal solution and are efficient (polynomially so in relevant problem dependent quantities) with regards to their sample and computational complexities."}, "keywords": ["LQR problem"], "citation_intent": "background"}
{"citing_id": "2304.05561v1", "cited_id": "1506.02753", "section_title": "Comparison With Non-Adversarial Embedding Inversion", "citation": "We have compared our results with the ones obtained using the non-adversarial embedding inversion approach by Dosovitskiy et al. in #REFR .", "text_before_citation": [], "text_after_citation": ["With access to the original data and to all layers of the original model (both missing in the adversarial settings), the approach in #OTHEREFR trains an inversion network minimizing the imagespace L2 loss.", "In #OTHEREFR , the lowest reconstruction errors are obtained when the inversion network is trained together with the original model in the auto-encoder setup.", "We adopt this case as our baseline, which provides an upper-bound to the reconstruction quality.", "To enable the comparison with our approach, we implemented the baseline on ResNet50 face recognition model.", "As shown in Figure 8 , despite the adversarial conditions, our attack can craft reconstructions that are close to the baseline ones."], "citing_paper_content": {"title": "On The Adversarial Inversion Of Deep Biometric Representations", "abstract": "Biometric authentication service providers often claim that it is not possible to reverse-engineer a user's raw biometric sample, such as a fingerprint or a face image, from its mathematical (feature-space) representation. This is presented as a security feature of the system against an attacker who may be able to retrieve the template in the feature space. In this paper, we investigate this claim on the specific example of deep neural network (DNN) embeddings. Inversion of DNN embeddings has been investigated for explaining deep image representations or synthesizing normalized images. When setting the inversion, existing studies leverage full access to all layers of the original model, as well as all possible information on the original dataset-including, in some cases, the original image to be reconstructed. For the biometric authentication use case, we need to investigate this under adversarial settings where an attacker has access to a feature-space representation but no direct access to the exact original dataset nor the original learned model. Instead, we assume varying degree of attacker's background knowledge about the distribution of the dataset as well as the original learned model (architecture and training process). In the worst case, we assume attacker has no knowledge of the original data distribution nor the original model. In these cases, we show that the attacker can exploit off-the-shelf DNN models and public datasets, to mimic the behaviour of the original learned model to varying degrees of success, based only on the obtained representation and attacker's prior knowledge. We propose a two-pronged attack that first infers the original DNN by exploiting the model footprint on the embedding, and then reconstructs the raw data by using the inferred model. We show the practicality of the attack on popular DNNs trained for two prominent biometric modalities, face and fingerprint recognition. The attack can effectively infer the original recognition model (mean accuracy 83% for faces, 86% for fingerprints), and can craft effective biometric reconstructions that are successfully authenticated with 1-vs-1 authentication accuracy of up to 92% for some models."}, "cited_paper_content": {"title": "Inverting Visual Representations With Convolutional Networks", "abstract": "Feature representations, both hand-designed and learned ones, are often hard to analyze and interpret, even when they are extracted from visual data. We propose a new approach to study image representations by inverting them with an up-convolutional neural network. We apply the method to shallow representations (HOG, SIFT, LBP), as well as to deep networks. For shallow representations our approach provides significantly better reconstructions than existing methods, revealing that there is surprisingly rich information contained in these features. Inverting a deep network trained on ImageNet provides several insights into the properties of the feature representation learned by the network. Most strikingly, the colors and the rough contours of an image can be reconstructed from activations in higher network layers and even from the predicted class probabilities."}, "keywords": ["non-adversarial embedding inversion"], "citation_intent": "result"}
{"citing_id": "2303.12187v1", "cited_id": "1510.08484", "section_title": "Pre-Training", "citation": "The noise audio clips in the categories of \"natural\", \"music\" and \"babble\" are sampled from the MUSAN dataset #REFR .", "text_before_citation": ["The implication of baseline and all is the same as Table 2 . CMLR afterward.", "Our model is trained on 4 A100 GPUs with max tokens 2500, one batch size of data containing at most 100 seconds of audio per GPU.", "It takes around 32 hours to finish a phase, where 400k steps are.", "We do not use accumulated gradient, update freq is set to 1.", "One thing to note is that, among all pre-training and fine-tuning stages, there is no extra noise added. Random noise is included in the decoding stage."], "text_after_citation": [], "citing_paper_content": {"title": "Practice Of The Conformer Enhanced Audio-Visual Hubert On Mandarin And English", "abstract": "Considering the bimodal nature of human speech perception, lips, and teeth movement has a pivotal role in automatic speech recognition. Benefiting from the correlated and noise-invariant visual information, audiovisual recognition systems enhance robustness in multiple scenarios. In previous work, audiovisual HuBERT appears to be the finest practice incorporating modality knowledge. This paper outlines a mixed methodology, named conformer enhanced AV-HuBERT, boosting the AV-HuBERT system's performance a step further. Compared with baseline AV-HuBERT, our method in the one-phase evaluation of clean and noisy conditions achieves 7% and 16% relative WER reduction on the English AVSR benchmark dataset LRS3. Furthermore, we establish a novel 1000h Mandarin AVSR dataset CSTS. On top of the baseline AV-HuBERT, we exceed the WeNet ASR system by 14% and 18% relatively on MISP and CMLR by pre-training with this dataset. The conformerenhanced AV-HuBERT we proposed brings 7% on MISP and 6% CER reduction on CMLR, compared with the baseline AV-HuBERT system."}, "cited_paper_content": {"title": "Musan: A Music, Speech, And Noise Corpus", "abstract": "This report introduces a new corpus of music, speech, and noise. This dataset is suitable for training models for voice activity detection (VAD) and music/speech discrimination. Our corpus is released under a flexible Creative Commons license. The dataset consists of music from several genres, speech from twelve languages, and a wide assortment of technical and non-technical noises. We demonstrate use of this corpus for music/speech discrimination on Broadcast news and VAD for speaker identification."}, "keywords": ["MUSAN dataset"], "citation_intent": "method"}
{"citing_id": "2303.10460v1", "cited_id": "1204.0867", "section_title": "Iv. Choosing A Spanning Tree Which Minimizes The Total Number Of Transmissions Used", "citation": "For a single uniprior ICP represented by the informationflow graph G, since the bandwidth-optimal index code in #REFR gives a separate code for each strongly connected component of G, in the rest of this paper, we consider all informationflow graphs to be strongly connected.", "text_before_citation": [], "text_after_citation": ["Further, for an index code C T obtained from a spanning tree T , both the notations l max (C T ) as well as l max (T ) are used interchangeably to mean the maximum number of transmissions used to decode a single requested message at any receiver while using the index code C T . Example 1.", "Consider the single uniprior ICP represented by the information-flow graph, G shown in Fig. 1 .", "For this graph, Algorithm 2 in #OTHEREFR forms the connected graph on 4 labeled nodes and finds a spanning tree of diameter two.", "There are four possible spanning trees of diameter two for a labeled K 4 , which are shown in Fig. 2 .", "Corresponding to each of these spanning trees, the bandwidth-optimal index codes satisfying the min-max probability of error criterion are given in Table I ."], "citing_paper_content": {"title": "Average Probability Of Error For Single Uniprior Index Coding Over Rayleigh Fading Channel", "abstract": "Ong and Ho developed optimal linear index codes for single uniprior index coding problems (ICPs) by finding a spanning tree for each of the strongly connected components of the corresponding information-flow graphs, following which Thomas et al. considered the same class of ICPs over Rayleigh fading channel. They developed the min-max probability of error criterion for choosing an index code which minimized the probability of error at the receivers and showed that there always exist optimal linear index codes for which any receiver takes at most two transmissions to decode a requested message. Motivated by the above works, this paper considers single uniprior ICPs over Rayleigh fading channels for which minimizing average probability of error is shown to be a criterion for further selection of index codes. The optimal index code w.r.t this criterion is shown to be one that minimizes the total number of transmissions used for decoding the message requests at all the receivers. An algorithm that generates a spanning tree which has a lower value of this metric as compared to the optimal star graph is also presented. For a given set of parameters of single uniprior ICPs, a lower bound for the total number of transmissions used by any optimal index code is derived, and a class of ICPs for which this bound is tight is identified."}, "cited_paper_content": {"title": "Optimal Index Codes For A Class Of Multicast Networks With Receiver Side Information", "abstract": "This paper studies a special class of multicast index coding problems where a sender transmits messages to multiple receivers, each with some side information. Here, each receiver knows a unique message a priori, and there is no restriction on how many messages each receiver requests from the sender. For this class of multicast index coding problems, we obtain the optimal index code, which has the shortest codelength for which the sender needs to send in order for all receivers to obtain their (respective) requested messages. This is the first class of index coding problems where the optimal index codes are found. In addition, linear index codes are shown to be optimal for this class of index coding problems."}, "keywords": ["bandwidth-optimal index code"], "citation_intent": "background"}
{"citing_id": "2304.04875v1", "cited_id": "1812.01601", "section_title": "Ablation Study", "citation": "Table 8 shows how 3DPW changes the 3D pseudo-GTs compared to other ITW datasets, such as MPII, LSPET, and InstaVariety #REFR .", "text_before_citation": ["This shows that our recipes are specially designed for the annotation networks f to obtain beneficial 3D pseudo-GTs.", "The reason for the small effect when the recipes are applied to the estimation networks g is that the estimation networks g are trained with 3D pseudo-GTs, while annotation networks f are trained with 2D GTs of ITW datasets without 3D evidence.", "The absence of 3D evidence when training annotation networks f results in severe ambiguities, which can be cured by our recipes.", "On the other hand, as estimation networks g are fully supervised with 3D pseudo-GTs, they suffer less from ambiguities.", "For each recipe, None represents both annotation and estimation networks are trained with the remaining other two recipes. Effect of training annotation network f on 3DPW."], "text_after_citation": ["As the table shows, adding other ITW datasets does not obtain the performance gain of g compared to H36M+MI+COCO.", "This is because adding ITW datasets does not contribute to relieving the depth ambiguity as they provide only 2D GTs.", "On the other hand, 3DPW provides 3D GTs, largely helpful to alleviate the depth ambiguity.", "Importantly, the 3D errors of g on MuPoTS decrease as well, which implies that using 3DPW as an additional training set is beneficial for multiple 3D ITW benchmarks.", "It is noticeable that InstaVariety has 95 times more images than 3DPW, while much less helpful for the beneficial 3D pseudo-GTs."], "citing_paper_content": {"title": "Three Recipes For Better 3D Pseudo-Gts Of 3D Human Mesh Estimation In The Wild", "abstract": "Recovering 3D human mesh in the wild is greatly challenging as in-the-wild (ITW) datasets provide only 2D pose ground truths (GTs). Recently, 3D pseudo-GTs have been widely used to train 3D human mesh estimation networks as the 3D pseudo-GTs enable 3D mesh supervision when training the networks on ITW datasets. However, despite the great potential of the 3D pseudo-GTs, there has been no extensive analysis that investigates which factors are important to make more beneficial 3D pseudo-GTs. In this paper, we provide three recipes to obtain highly beneficial 3D pseudo-GTs of ITW datasets. The main challenge is that only 2D-based weak supervision is allowed when obtaining the 3D pseudo-GTs. Each of our three recipes addresses the challenge in each aspect: depth ambiguity, sub-optimality of weak supervision, and implausible articulation. Experimental results show that simply retraining state-of-the-art networks with our new 3D pseudo-GTs elevates their performance to the next level without bells and whistles. The 3D pseudo-GT is publicly available 1 ."}, "cited_paper_content": {"title": "Learning 3D Human Dynamics From Video", "abstract": "From an image of a person in action, we can easily guess the 3D motion of the person in the immediate past and future. This is because we have a mental model of 3D human dynamics that we have acquired from observing visual sequences of humans in motion. We present a framework that can similarly learn a representation of 3D dynamics of humans from video via a simple but effective temporal encoding of image features. At test time, from video, the learned temporal representation give rise to smooth 3D mesh predictions. From a single image, our model can recover the current 3D mesh as well as its 3D past and future motion. Our approach is designed so it can learn from videos with 2D pose annotations in a semi-supervised manner. Though annotated data is always limited, there are millions of videos uploaded daily on the Internet. In this work, we harvest this Internet-scale source of unlabeled data by training our model on unlabeled video with pseudo-ground truth 2D pose obtained from an off-the-shelf 2D pose detector. Our experiments show that adding more videos with pseudo-ground truth 2D pose monotonically improves 3D prediction performance. We evaluate our model on the recent challenging dataset of 3D Poses in the Wild and obtain state-of-the-art performance on the 3D prediction task without any fine-tuning. The project website with video can be found at https://akanazawa.github.io/human_dynamics/."}, "keywords": ["3D pseudo"], "citation_intent": "result"}
{"citing_id": "2303.12942v1", "cited_id": "1907.07374", "section_title": "Iii. Research Methodology", "citation": "Researchers in #REFR provide a survey of interpretability and explainability related to ML algorithms, categorizing different interpretations proposed by different research works. Their categorization also pertains to medical applications.", "text_before_citation": ["In #OTHEREFR , the authors present a survey that explores the current trends and challenges of using visual analytics to interpret deep learning models based on XAI methods, as well as future directions of research in this area.", "Two perspectives have been taken into consideration, model usage and visual approaches.", "Model usage focuses on the performance of the AI system and how it behaves in different scenarios, while visual approaches examine the behavior of the AI system by looking at its outputs, such as predictions or classifications, over time.", "The researchers identified several research questions in light of their findings, then discussed the research directions that could be pursued in the future.", "By using XAI methods in the field of visual analytics, this survey provides guidance to better interpret neural networks."], "text_after_citation": ["According to the researchers, interpretability research can broadly be classified in many ways.", "It can range from methods that provide clearly interpreted information to analysis of complex patterns.", "The authors present attempts to mathematically formalize interpretability, as well as efforts to visualize it and evaluate its impact on task performance.", "These different approaches aim to better understand and improve the interpretability of algorithms.", "The investigators in #OTHEREFR provide a literature review of the use of XAI in the field of cybersecurity."], "citing_paper_content": {"title": "A Survey On Explainable Artificial Intelligence For Network Cybersecurity", "abstract": "The \"black-box\" nature of artificial intelligence (AI) models has been the source of many concerns in their use for critical applications. Explainable Artificial Intelligence (XAI) is a rapidly growing research field that aims to create machine learning models that can provide clear and interpretable explanations for their decisions and actions. In the field of network cybersecurity, XAI has the potential to revolutionize the way we approach network security by enabling us to better understand the behavior of cyber threats and to design more effective defenses. In this survey, we review the state of the art in XAI for cybersecurity in network systems and explore the various approaches that have been proposed to address this important problem. The review follows a systematic classification of network-driven cybersecurity threats and issues. We discuss the challenges and limitations of current XAI methods in the context of cybersecurity and outline promising directions for future research."}, "cited_paper_content": {"title": "A Survey On Explainable Artificial Intelligence (Xai): Towards Medical Xai", "abstract": "Recently, artificial intelligence, especially machine learning has demonstrated remarkable performances in many tasks, from image processing to natural language processing, especially with the advent of deep learning. Along with research progress, machine learning has encroached into many different fields and disciplines. Some of them, such as the medical field, require high level of accountability, and thus transparency, which means we need to be able to explain machine decisions, predictions and justify their reliability. This requires greater interpretability, which often means we need to understand the mechanism underlying the algorithms. Unfortunately, the black-box nature of the deep learning is still unresolved, and many machine decisions are still poorly understood. We provide a review on interpretabilities suggested by different research works and categorize them, with the intention of providing alternative perspective that is hopefully more tractable for future adoption of interpretability standard. We explore further into interpretability in the medical field, illustrating the complexity of interpretability issue."}, "keywords": ["explainability", "interpretability"], "citation_intent": "background"}
{"citing_id": "2304.09527v1", "cited_id": "1604.03650", "section_title": "Stereo Synthesis Comparison", "citation": "In Figure 5 , the poles generated by Xie et al. #REFR and Godard et al.", "text_before_citation": ["Although they perform well on visible regions, they are inefficient on handling the disoccluded regions.", "Thanks to the two rectification strategies, our method is able to identify the error regions caused by the first stage and inpaint it with reasonable contents.", "Specifically, the pruning-based rectification detects the intractable area of warping while the bidirectional matching rectification finds out the inconsistent pixels between the stereo views.", "On the other hands, the exemplar synthesis results are shown in Figure 5 , and we can see that there are some artifacts in their results.", "In particular, their synthesized stereo either produces blurring artifacts, or contains severe distortion."], "text_after_citation": ["#OTHEREFR (highlighted in the red rectangles) are distorted, while the ones generated by our method are very close to the ground-truth image.", "Although the pole in the synthesized results produced by Luo et al. #OTHEREFR and Gonzalez et al.", "#OTHEREFR are straight, their image details are not well preserved.", "Also, the shape of traffic lights in Gonzalez et al. #OTHEREFR is blurry.", "Moreover, the contour of the van in the green rectangles generated by previous methods is blurry or distorted, while our method can produce faithful result."], "citing_paper_content": {"title": "Single-View View Synthesis With Self-Rectified Pseudo-Stereo", "abstract": "Synthesizing novel views from a single view image is a highly ill-posed problem. We discover an effective solution to reduce the learning ambiguity by expanding the single-view view synthesis problem to a multi-view setting. Specifically, we leverage the reliable and explicit stereo prior to generate a pseudostereo viewpoint, which serves as an auxiliary input to construct the 3D space. In this way, the challenging novel view synthesis process is decoupled into two simpler problems of stereo synthesis and 3D reconstruction. In order to synthesize a structurally correct and detail-preserved stereo image, we propose a selfrectified stereo synthesis to amend erroneous regions in an identify-rectify manner. Hard-to-train and incorrect warping samples are first discovered by two strategies, 1) pruning the network to reveal low-confident predictions; and 2) bidirectionally matching between stereo images to allow the discovery of improper mapping. These regions are then inpainted to form the final pseudo-stereo. With the aid of this extra input, a preferable 3D reconstruction can be easily obtained, and our"}, "cited_paper_content": {"title": "Deep3D: Fully Automatic 2D-To-3D Video Conversion With Deep Convolutional Neural Networks", "abstract": "As 3D movie viewing becomes mainstream and the Virtual Reality (VR) market emerges, the demand for 3D contents is growing rapidly. Producing 3D videos, however, remains challenging. In this paper we propose to use deep neural networks to automatically convert 2D videos and images to a stereoscopic 3D format. In contrast to previous automatic 2D-to-3D conversion algorithms, which have separate stages and need ground truth depth map as supervision, our approach is trained end-to-end directly on stereo pairs extracted from existing 3D movies. This novel training scheme makes it possible to exploit orders of magnitude more data and significantly increases performance. Indeed, Deep3D outperforms baselines in both quantitative and human subject evaluations."}, "keywords": ["Figure", "Godard"], "citation_intent": "background"}
{"citing_id": "2304.07442v1", "cited_id": "1312.5602", "section_title": "Iv. Our Algorithm", "citation": "To overcome this issue, we develop a technique inspired by work in reinforcement learning #REFR .", "text_before_citation": ["These updates do not cause a significant overall decrease in optimizee's cost function.", "This effect is more pronounced when we operate in parameter space as opposed to gradient space since the LSTM network does not get any information about the curvature of the optimization surface.", "Running optimization in a finite horizon window in parameter space can then cause LSTM to suggest incorrect updates to the optimizee network.", "In the specific case of QNNs, the parameters correspond to a rotation of given input state about a particular axis.", "An incorrect update can very easily cause the quantum state to be rotated incorrectly and therefore lead to an increase in the value of the objective function we're interested in minimizing."], "text_after_citation": ["Throughout training, we keep track of past history of parameters and their corresponding cost function values in a \"replay buffer\".", "At the start of training we instantiate a double ended queue dubbed as a replay-buffer B of a finite capacity R. For metaiteration t = 1 . . .", "T we observe a history of parameters \u03b8 t and the corresponding cost C(\u03b8 t ).", "If C(\u03b8 t+1 ) < C(\u03b8 t ) then we add the state s = [\u03b8 t+1 , C(\u03b8 t+1 ), \u2206C(\u03b8), h t+1 ] to the replay buffer.", "Once the meta-iteration ends and L(\u03a6) is computed, if the QNN cost function is diverging, we seed the parameters for the next meta-iteration by performing the following update:"], "citing_paper_content": {"title": "Learning To Optimize Quantum Neural Network Without Gradients", "abstract": "Quantum Machine Learning is an emerging subfield in machine learning where one of the goals is to perform pattern recognition tasks by encoding data into quantum states. This extension from classical to quantum domain has been made possible due to the development of hybrid quantum-classical algorithms that allow a parameterized quantum circuit to be optimized using gradient based algorithms that run on a classical computer. The similarities in training of these hybrid algorithms and classical neural networks has further led to the development of Quantum Neural Networks (QNNs). However, in the current training regime for QNNs, the gradients w.r.t objective function have to be computed on the quantum device. This computation is highly non-scalable and is affected by hardware and sampling noise present in the current generation of quantum hardware. In this paper, we propose a training algorithm that does not rely on gradient information. Specifically, we introduce a novel metaoptimization algorithm that trains a meta-optimizer network to output parameters for the quantum circuit such that the objective function is minimized. We empirically and theoretically show that we achieve a better quality minima in fewer circuit evaluations than existing gradient based algorithms on different datasets."}, "cited_paper_content": {"title": "Playing Atari With Deep Reinforcement Learning", "abstract": "We present the first deep learning model to successfully learn control policies directly from high-dimensional sensory input using reinforcement learning. The model is a convolutional neural network, trained with a variant of Q-learning, whose input is raw pixels and whose output is a value function estimating future rewards. We apply our method to seven Atari 2600 games from the Arcade Learning Environment, with no adjustment of the architecture or learning algorithm. We find that it outperforms all previous approaches on six of the games and surpasses a human expert on three of them."}, "keywords": ["reinforcement learning"], "citation_intent": "method"}
{"citing_id": "2304.03006v1", "cited_id": "1602.05629", "section_title": "A. Federated Learning", "citation": "However, for the federated model, which has seen no training data (insuring the data privacy), these local updates are then collated, in our case via a block to be added to the blockchain, and averaged based on the FedAvg algorithm #REFR .", "text_before_citation": ["The goal for every model in the system, both the local models and the federated (global) model is to minimize the loss with respect to the model parameter: min", "EQUATION", "Where is a chosen loss function, consistent across all participating models, x, y are the training (input and desired output) vectors and \u03c9 is the model's parameters."], "text_after_citation": ["Hence, for each local model index m = 1, 2, ...., M of a participating IoT device, which has performed local training (1) either on the device itself or via a processing node, over its training set T m :", "EQUATION", "Where", "T = M m=1", "|T m |."], "citing_paper_content": {"title": "Iot Federated Blockchain Learning At The Edge", "abstract": "IoT devices are sorely underutilised in the medical field, especially within machine learning for medicine, yet they offer unrivalled benefits. IoT devices are low cost, energy efficient, small and intelligent devices [1]. In this paper, we propose a distributed federated learning framework for IoT devices, more specifically for IoMT (Internet of Medical Things), using blockchain to allow for a decentralised scheme improving privacy and efficiency over a centralised system; this allows us to move from the cloud based architectures, that are prevalent, to the edge. The system is designed for three paradigms: 1) Training neural networks on IoT devices to allow for collaborative training of a shared model whilst decoupling the learning from the dataset [2] to ensure privacy [3]. Training is performed in an online manner simultaneously amongst all participants, allowing for training of actual data that may not have been present in a dataset collected in the traditional way and dynamically adapt the system whilst it is being trained. 2) Training of an IoMT system in a fully private manner such as to mitigate the issue with confidentiality of medical data and to build robust, and potentially bespoke [4], models where not much, if any, data exists. 3) Distribution of the actual network training, something federated learning itself does not do, to allow hospitals, for example, to utilize their spare computing resources to train network models."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["federated model"], "citation_intent": "method"}
{"citing_id": "2305.02468v1", "cited_id": "1909.02027", "section_title": "Datasets", "citation": "CLINC150 #REFR This dataset is multi-domain dataset which contains 23,700 utterances that cover 150 intent classes over 10 domains.", "text_before_citation": ["The dialogues cover a wide range of domains and topics.", "MultiWOZ 2.2 #OTHEREFR is the improved version of MultiWOZ 2.1 #OTHEREFR ) that has corrected annotation errors, inconsistencies, and ontology issues, also added span annotations to standardize.", "MultiWOZ -2.1, 2.2 The MultiWOZ", "Banking 77 #OTHEREFR This dataset is a collection of 77 real-life customer banking service queries. It consists of 13,083 utterances.", "Each query is labeled with a single intent, however, it is hard to differentiate because they correspond to very similar tasks."], "text_after_citation": ["HWU64"], "citing_paper_content": {"title": "Task-Optimized Adapters For An End-To-End Task-Oriented Dialogue System", "abstract": "Task-Oriented Dialogue (TOD) systems are designed to carry out specific tasks by tracking dialogue states and generating appropriate responses to help users achieve defined goals. Recently, end-to-end dialogue models pre-trained based on large datasets have shown promising performance in the conversational system. However, they share the same parameters to train tasks of the dialogue system (NLU, DST, NLG), so debugging each task is challenging. Also, they require a lot of effort to fine-tune large parameters to create a taskoriented chatbot, making it difficult for nonexperts to handle. Therefore, we intend to train relatively lightweight and fast models compared to PLM. In this paper, we propose an End-to-end TOD system with Task-Optimized Adapters which learn independently per task, adding only small number of parameters after fixed layers of pre-trained network. We also enhance the performance of the DST and NLG modules through reinforcement learning, overcoming the learning curve that has lacked at the adapter learning and enabling the natural and consistent response generation that is appropriate for the goal. Our method is a model-agnostic approach and does not require prompt-tuning as only input data without a prompt. As results of the experiment, our method shows competitive performance on the MultiWOZ benchmark compared to the existing end-to-end models. In particular, we attain state-of-the-art performance on the DST task of 2.2 dataset."}, "cited_paper_content": {"title": "An Evaluation Dataset For Intent Classification And Out-Of-Scope Prediction", "abstract": "Task-oriented dialog systems need to know when a query falls outside their range of supported intents, but current text classification corpora only define label sets that cover every example. We introduce a new dataset that includes queries that are out-of-scope---i.e., queries that do not fall into any of the system's supported intents. This poses a new challenge because models cannot assume that every query at inference time belongs to a system-supported intent class. Our dataset also covers 150 intent classes over 10 domains, capturing the breadth that a production task-oriented agent must handle. We evaluate a range of benchmark classifiers on our dataset along with several different out-of-scope identification schemes. We find that while the classifiers perform well on in-scope intent classification, they struggle to identify out-of-scope queries. Our dataset and evaluation fill an important gap in the field, offering a way of more rigorously and realistically benchmarking text classification in task-driven dialog systems."}, "keywords": ["multi-domain dataset", "150 intent classes"], "citation_intent": "background"}
{"citing_id": "2303.11546v1", "cited_id": "1706.02677", "section_title": "Implementation Details", "citation": "A weight decay is set to 0.01, with a linear warmup #REFR over t warm =1k iterations, followed by a linear decay.", "text_before_citation": ["Mapillary #OTHEREFR involves 18,000 training images and 2,000 validation images with diverse resolutions.", "For brevity, we denote GTA, SYNTHIA, Cityscapes, BDD, and Mapillary as G, S, C, B, and M, respectively. Network architecture.", "We conduct experiments using ResNet #OTHEREFR as an encoder architecture and DeepLabV3+ #OTHEREFR as a semantic segmentation decoder architecture.", "In all experiments, encoders are initialized with an ImageNet #OTHEREFR pre-trained model. Training. We adopt an AdamW [38] optimizer.", "An initial learning rate is set to 3\u00d710 \u22125 for the encoder and 3 \u00d7 10 \u22124 for the decoder, 40k training iterations, a batch size of 4."], "text_after_citation": ["We use random scaling in the range [0.5, 2.0] and random cropping with a size of 768\u00d7768.", "We apply additional data augmentation techniques, including random flipping and color jittering.", "We set the texture regularization parameters as u l =5 \u00d7 10 \u2212l\u22122 , and the texture generalization parameters as v l =5 \u00d7 10 \u2212l\u22122 .", "The original task loss and the stylized task loss weights are set to \u03b1 orig =0.5 and \u03b1 styl =0.5, respectively. We set the RSM threshold \u03c4 =0.1."], "citing_paper_content": {"title": "Texture Learning Domain Randomization For Domain Generalized Segmentation", "abstract": "Deep Neural Networks (DNNs)-based semantic segmentation models trained on a source domain often struggle to generalize to unseen target domains, i.e., a domain gap problem. Texture often contributes to the domain gap, making DNNs vulnerable to domain shift because they are prone to be texture-biased. Existing Domain Generalized Semantic Segmentation (DGSS) methods have alleviated the domain gap problem by guiding models to prioritize shape over texture. On the other hand, shape and texture are two prominent and complementary cues in semantic segmentation. This paper argues that leveraging texture is crucial for improving performance in DGSS. Specifically, we propose a novel framework, coined Texture Learning Domain Randomization (TLDR). TLDR includes two novel losses to effectively enhance texture learning in DGSS: (1) a texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pretrained model and (2) a texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner. Extensive experimental results demonstrate the superiority of the proposed TLDR; e.g., TLDR achieves 46.5 mIoU on GTA\u2192Cityscapes using ResNet-50, which improves the prior state-of-the-art method by 1.9 mIoU."}, "cited_paper_content": {"title": "Accurate, Large Minibatch Sgd: Training Imagenet In 1 Hour", "abstract": "Deep learning thrives with large neural networks and large datasets. However, larger networks and larger datasets result in longer training times that impede research and development progress. Distributed synchronous SGD offers a potential solution to this problem by dividing SGD minibatches over a pool of parallel workers. Yet to make this scheme efficient, the per-worker workload must be large, which implies nontrivial growth in the SGD minibatch size. In this paper, we empirically show that on the ImageNet dataset large minibatches cause optimization difficulties, but when these are addressed the trained networks exhibit good generalization. Specifically, we show no loss of accuracy when training with large minibatch sizes up to 8192 images. To achieve this result, we adopt a hyper-parameter-free linear scaling rule for adjusting learning rates as a function of minibatch size and develop a new warmup scheme that overcomes optimization challenges early in training. With these simple techniques, our Caffe2-based system trains ResNet-50 with a minibatch size of 8192 on 256 GPUs in one hour, while matching small minibatch accuracy. Using commodity hardware, our implementation achieves ~90% scaling efficiency when moving from 8 to 256 GPUs. Our findings enable training visual recognition models on internet-scale data with high efficiency."}, "keywords": ["weight decay", "linear warmup"], "citation_intent": "method"}
{"citing_id": "2304.09715v1", "cited_id": "1707.03167", "section_title": "Iv. Experiments", "citation": "Proposed by RegNet #REFR , it is the first Kitti split introduced for this task, and as shown by our results in Table II it remains the most challenging.", "text_before_citation": ["It provides enough samples and an accurate calibration ground truth.", "Using the same dataset as previous works will help us compare our results.", "However, those previous publications used different splits of the Kitti dataset.", "We present those splits in Table I and will refer to them as \u03b1, \u03b2, \u03b3, and \u03b4.", "From the splits presented in Table I , we choose \u03b1 as our reference."], "text_after_citation": ["This is because it uses samples from separate days for training/validation and testing, with different camera intrinsics.", "In comparison, splits \u03b2, \u03b3, and \u03b4 used in other publications include in their testing sets samples recorded the same day and with the same camera intrinsics as their training sets.", "\u03b4 even uses spatially redundant samples (some scenes are captured in the same location in training and testing sets) as acknowledged in #OTHEREFR .", "The split and dataset choice are not trivial as results in Table II show that the same network (here UniCal) trained and tested on different splits will perform differently.", "The splits with more similarities between the training and testing sets will get seemingly better results, whether those similarities are mostly in the camera intrinsics as in \u03b2, or even in the location where the samples were collected, as in \u03b4."], "citing_paper_content": {"title": "Unical: A Single-Branch Transformer-Based Model For Camera-To-Lidar Calibration And Validation", "abstract": "We introduce a novel architecture, UniCal, for Camera-to-LiDAR (C2L) extrinsic calibration which leverages self-attention mechanisms through a Transformer-based backbone network to infer the 6-degree of freedom (DoF) relative transformation between the sensors. Unlike previous methods, UniCal performs an early fusion of the input camera and LiDAR data by aggregating camera image channels and LiDAR mappings into a multi-channel unified representation before extracting their features jointly with a single-branch architecture. This single-branch architecture makes UniCal lightweight, which is desirable in applications with restrained resources such as autonomous driving. Through experiments, we show that UniCal achieves state-of-the-art results compared to existing methods. We also show that through transfer learning, weights learned on the calibration task can be applied to a calibration validation task without retraining the backbone."}, "cited_paper_content": {"title": "Regnet: Multimodal Sensor Registration Using Deep Neural Networks", "abstract": "In this paper, we present RegNet, the first deep convolutional neural network (CNN) to infer a 6 degrees of freedom (DOF) extrinsic calibration between multimodal sensors, exemplified using a scanning LiDAR and a monocular camera. Compared to existing approaches, RegNet casts all three conventional calibration steps (feature extraction, feature matching and global regression) into a single real-time capable CNN. Our method does not require any human interaction and bridges the gap between classical offline and target-less online calibration approaches as it provides both a stable initial estimation as well as a continuous online correction of the extrinsic parameters. During training we randomly decalibrate our system in order to train RegNet to infer the correspondence between projected depth measurements and RGB image and finally regress the extrinsic calibration. Additionally, with an iterative execution of multiple CNNs, that are trained on different magnitudes of decalibration, our approach compares favorably to state-of-the-art methods in terms of a mean calibration error of 0.28 degrees for the rotational and 6 cm for the translation components even for large decalibrations up to 1.5 m and 20 degrees."}, "keywords": ["RegNet"], "citation_intent": "background"}
{"citing_id": "2303.00844v1", "cited_id": "1909.09564", "section_title": "Lad-Lasso-Based Omp", "citation": "Finally, we consider the LAD-LASSO loss function G LAD \u2113 1 w defined in #REFR .", "text_before_citation": [], "text_after_citation": ["In this case, we restrict ourselves to the real-valued case for the sake of simplicity.", "In order to formulate the corresponding greedy selection rule, we need to introduce some auxiliary notation.", "First, we define an augmented version A \u2208 R (m+1)\u00d7N of the matrix A \u2208 R m\u00d7N as A := A \u03bbw * or, equivalently, A ij :=", "EQUATION", "In addition, given x \u2208 R N , we consider N augmentations of the residual vector r = Ax \u2212 y \u2208 R N as the vectors r j \u2208 R m+1 , defined by"], "citing_paper_content": {"title": "The Greedy Side Of The Lasso: New Algorithms For Weighted Sparse Recovery Via Loss Function-Based Orthogonal Matching Pursuit", "abstract": "We propose a class of greedy algorithms for weighted sparse recovery by considering new loss function-based generalizations of Orthogonal Matching Pursuit (OMP). Given a (regularized) loss function, the proposed algorithms alternate the iterative construction of the signal support via greedy index selection and a signal update based on solving a local data-fitting problem restricted to the current support. We show that greedy selection rules associated with popular weighted sparsity-promoting loss functions admit explicitly computable and simple formulas. Specifically, we consider \u2113 0-and \u2113 1-based versions of the weighted LASSO (Least Absolute Shrinkage and Selection Operator), the Square-Root LASSO (SR-LASSO) and the Least Absolute Deviations LASSO (LAD-LASSO). Through numerical experiments on Gaussian compressive sensing and high-dimensional function approximation, we demonstrate the effectiveness of the proposed algorithms and empirically show that they inherit desirable characteristics from the corresponding loss functions, such as SR-LASSO's noise-blind optimal parameter tuning and LAD-LASSO's fault tolerance. In doing so, our study sheds new light on the connection between greedy sparse recovery and convex relaxation."}, "cited_paper_content": {"title": "Sparse Harmonic Transforms Ii: Best $S$-Term Approximation Guarantees For Bounded Orthonormal Product Bases In Sublinear-Time", "abstract": "In this paper, we develop a sublinear-time compressive sensing algorithm for approximating functions of many variables which are compressible in a given Bounded Orthonormal Product Basis (BOPB). The resulting algorithm is shown to both have an associated best $s$-term recovery guarantee in the given BOPB, and also to work well numerically for solving sparse approximation problems involving functions contained in the span of fairly general sets of as many as $\\sim10^{230}$ orthonormal basis functions. All code is made publicly available. ::: As part of the proof of the main recovery guarantee new variants of the well known CoSaMP algorithm are proposed which can utilize any sufficiently accurate support identification procedure satisfying a {Support Identification Property (SIP)} in order to obtain strong sparse approximation guarantees. These new CoSaMP variants are then proven to have both runtime and recovery error behavior which are largely determined by the associated runtime and error behavior of the chosen support identification method. The main theoretical results of the paper are then shown by developing a sublinear-time support identification algorithm for general BOPB sets which is robust to arbitrary additive errors. Using this new support identification method to create a new CoSaMP variant then results in a new robust sublinear-time compressive sensing algorithm for BOPB-compressible functions of many variables."}, "keywords": ["LAD-LASSO loss function"], "citation_intent": "method"}
{"citing_id": "2303.09128v1", "cited_id": "1910.10683", "section_title": "Models", "citation": "We have experimented with two large language models for code: (1) CodeT5 , which is an encoder-decoder model based on T5 #REFR and (2) Codex , which is a decoder only model based on GPT-3 (Brown et al., 2020) .", "text_before_citation": [], "text_after_citation": ["Both T5 and GPT-3 have strong zero-shot learning #OTHEREFR and transfer learning capabilities, and their versions for programming languages exhibit strong performance across multiple benchmarks.", "The two models are of different sizesthe CodeT5 is using 700M parameters T5-large architecture, and the Codex model uses GPT-3 architecture with more than 100B parameters.", "We have provided a more detailed discussion of these models in the Appendix, Section 7.3.", "Next, we describe the methods we have employed to improve out-of-domain generalization for each model."], "citing_paper_content": {"title": "Exploring Distributional Shifts In Large Language Models For Code Analysis", "abstract": "We systematically study the capacity of two large language models for code-CodeT5 and Codex-to generalize to out-of-domain data. In this study, we consider two fundamental applications-code summarization, and code generation. We split data into domains following its natural boundaries-by an organization, by a project, and by a module within the software project. This makes recognition of in-domain vs out-of-domain data at the time of deployment trivial. We establish that samples from each new domain present both models with a significant challenge of distribution shift. We study how well different established methods can adapt models to better generalize to new domains. Our experiments show that while multitask learning alone is a reasonable baseline, combining it with few-shot finetuning on examples retrieved from training data can achieve very strong performance. In fact, according to our experiments, this solution can outperform direct finetuning for very low-data scenarios. Finally, we consider variations of this approach to create a more broadly applicable method to adapt to multiple domains at once. We find that in the case of code generation, a model adapted to multiple domains simultaneously performs on par with those adapted to each domain individually."}, "cited_paper_content": {"title": "Exploring The Limits Of Transfer Learning With A Unified Text-To-Text Transformer", "abstract": "Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new \"Colossal Clean Crawled Corpus\", we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code."}, "keywords": ["two large language", "decoder"], "citation_intent": "method"}
{"citing_id": "2303.16206v1", "cited_id": "2003.12039", "section_title": "Related Work", "citation": "Most similar to our approach is arguably RAFT #REFR , which poses optical flow estimation as an optimization problem and uses a recurrent neural network to find an optimal descent direction.", "text_before_citation": ["Learning to Optimize.", "Recent methods have investigated incorporating optimization problems into neural network architectures #OTHEREFR .", "These methods design custom neural network layers that mimic certain canonical optimization problems.", "As a result, parametrized optimization problems can be inserted into neural networks.", "Not all optimization algorithms can easily be written as a layer and another strategy is to train a network to learn iterative updates from data (Adler &\u00d6ktem, 2018; #OTHEREFR ."], "text_after_citation": ["In a similar vein, we aim to learn the steganography optimization problem with a convolutional neural network without specially designed optimization layers."], "citing_paper_content": {"title": "Learning Iterative Neural Optimizers For Im-Age Steganography", "abstract": "Image steganography is the process of concealing secret information in images through imperceptible changes. Recent work has formulated this task as a classic constrained optimization problem. In this paper, we argue that image steganography is inherently performed on the (elusive) manifold of natural images, and propose an iterative neural network trained to perform the optimization steps. In contrast to classical optimization methods like L-BFGS or projected gradient descent, we train the neural network to also stay close to the manifold of natural images throughout the optimization. We show that our learned neural optimization is faster and more reliable than classical optimization approaches. In comparison to previous state-of-the-art encoder-decoder based steganography methods, it reduces the recovery error rate by multiple orders of magnitude and achieves zero error up to 3 bits per pixel (bpp) without the need for error-correcting codes."}, "cited_paper_content": {"title": "Raft: Recurrent All-Pairs Field Transforms For Optical Flow", "abstract": "We introduce Recurrent All-Pairs Field Transforms (RAFT), a new deep network architecture for optical flow. RAFT extracts per-pixel features, builds multi-scale 4D correlation volumes for all pairs of pixels, and iteratively updates a flow field through a recurrent unit that performs lookups on the correlation volumes. RAFT achieves state-of-the-art performance, with strong cross-dataset generalization and high efficiency in inference time, training speed, and parameter count. Code is available \\url{https://github.com/princeton-vl/RAFT}."}, "keywords": ["recurrent neural network", "optical flow estimation"], "citation_intent": "method"}
{"citing_id": "2304.12012v1", "cited_id": "1602.05629", "section_title": "Introduction", "citation": "Under certain conditions #REFR , this procedure is guaranteed to converge to a final global model representing an optimal consensus among the hospitals participating in the experiment.", "text_before_citation": ["The need for large amounts of data to develop Artificial Intelligence (AI) in healthcare has motivated a number of national and international initiatives aimed at creating medical data lakes accessible to researchers, such as the French Health Data Hub #OTHEREFR , the UK BioBank #OTHEREFR , the US ADNI #OTHEREFR and TCGA #OTHEREFR , among the many #OTHEREFR .", "In spite of these initiatives, there are still major bottlenecks preventing the widespread availability of large centralized repositories of healthcare information #OTHEREFR .", "To overcome these limitations, Federated Learning (FL) has been proposed as a working paradigm to enable the training of ML models on large datasets from diverse sources while guaranteeing the respect of data privacy and governance.", "The basic paradigm of FL consists of iterating the following steps: i) model training is performed locally in the hospitals starting from a common initialization, ii) the resulting model parameters are subsequently shared (instead of the data) and aggregated, to define a global model iii) transmitted back to the hospitals to initiate a new local training step."], "text_after_citation": ["FL is particularly suited for applications in sensitive domains, such as healthcare and biomedical research #OTHEREFR .", "The current societal and economical interest in FL for healthcare is paramount [56, #OTHEREFR , as demonstrated by the several largescale medical research projects based on FL at the national and international level, focusing for example on rare hematological diseases 1 , drug development 2 , blood cancer #OTHEREFR , among the many #OTHEREFR .", "In spite of the current popularity, the real-world implementation of FL is complex and requires research and development actions at the crossroad between different domains spanning data science, software programming, networking, and security.", "Today, several FL software frameworks are currently being proposed to data scientists and users, based on different design spaces, goals, and with varying degrees of software maturity.", "Nevertheless, most of these frameworks are not designed to find seamless application in medical use-cases, due to the specific challenges and requirements of working with medical data and hospital infrastructures."], "citing_paper_content": {"title": "Fed-Biomed: Open, Transparent And Trusted Federated Learning For Real-World Healthcare Applications", "abstract": "The real-world implementation of federated learning is complex and requires research and development actions at the crossroad between different domains ranging from data science, to software programming, networking, and security. While today several FL libraries are proposed to data scientists and users, most of these frameworks are not designed to find seamless application in medical use-cases, due to the specific challenges and requirements of working with medical data and hospital infrastructures. Moreover, governance, design principles, and security assumptions of these frameworks are generally not clearly illustrated, thus preventing the adoption in sensitive applications. Motivated by the current technological landscape of FL in healthcare, in this document we present"}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["hospitals", "optimal consensus"], "citation_intent": "method"}
{"citing_id": "2303.00516v1", "cited_id": "1511.06581", "section_title": "Interaction Task", "citation": "Since M 0 now includes continuous features we adopt Dueling DQN #REFR , a Deep RL algorithm.", "text_before_citation": ["In addition, we consider a ground MDP M 0,d at which the robot movements are modelled using continuous features.", "The state space S 0 now contains continuous vectors (x, y, \u03b8, v), representing pose and velocity of agent's mobile base on the plane.", "The discrete set of actions A 0 allows to accelerate, decelerate, rotate, and a special action denotes the initiation of an interaction.", "There exists goal MDPs, M 2 , M 1 , M 0 that capture both the dynamics and the task defined above, which can be obtained through a suitable composition of each M i,d and D #OTHEREFR .", "Therefore, we can still apply our tecnique to the composed goal MDP."], "text_after_citation": ["The plot in Figure 6 shows a training comparison between the Dueling DQN agent alone (dot-dashed brown), and Dueling DQN receiving rewards from the grid abstraction (green).", "As we can see, our method allows to provide useful exploration bias even in case of extremely sparse goal states, as in this case."], "citing_paper_content": {"title": "Exploiting Multiple Abstractions In Episodic Rl Via Reward Shaping", "abstract": "One major limitation to the applicability of Reinforcement Learning (RL) to many practical domains is the large number of samples required to learn an optimal policy. To address this problem and improve learning efficiency, we consider a linear hierarchy of abstraction layers of the Markov Decision Process (MDP) underlying the target domain. Each layer is an MDP representing a coarser model of the one immediately below in the hierarchy. In this work, we propose a novel form of Reward Shaping where the solution obtained at the abstract level is used to offer rewards to the more concrete MDP, in such a way that the abstract solution guides the learning in the more complex domain. In contrast with other works in Hierarchical RL, our technique has few requirements in the design of the abstract models and it is also tolerant to modeling errors, thus making the proposed approach practical. We formally analyze the relationship between the abstract models and the exploration heuristic induced in the lower-level domain. Moreover, we prove that the method guarantees optimal convergence and we demonstrate its effectiveness experimentally."}, "cited_paper_content": {"title": "Dueling Network Architectures For Deep Reinforcement Learning", "abstract": "In recent years there have been many successes of using deep representations in reinforcement learning. Still, many of these applications use conventional architectures, such as convolutional networks, LSTMs, or auto-encoders. In this paper, we present a new neural network architecture for model-free reinforcement learning. Our dueling network represents two separate estimators: one for the state value function and one for the state-dependent action advantage function. The main benefit of this factoring is to generalize learning across actions without imposing any change to the underlying reinforcement learning algorithm. Our results show that this architecture leads to better policy evaluation in the presence of many similar-valued actions. Moreover, the dueling architecture enables our RL agent to outperform the state-of-the-art on the Atari 2600 domain."}, "keywords": ["Deep RL algorithm", "Dueling DQN"], "citation_intent": "method"}
{"citing_id": "2303.01068v1", "cited_id": "1312.6199", "section_title": "Introduction", "citation": "In spite of the impressive performance of Deep Neural Networks (DNNs) in different fields, from computer vision to Natural Language Processing, these models are shown to be vulnerable to adversarial attacks, i.e., small perturbations of the input data #REFR .", "text_before_citation": [], "text_after_citation": ["In recent years many works have been proposed to evaluate the robustness of DNN models, and design methods to make them more robust to the perturbations to their inputs #OTHEREFR .", "Neural Machine Translation (NMT) models, which take an input sentence and automatically generate its translation, have reached impressive performance by us-ing DNN models such as transformers #OTHEREFR .", "Due to their performance, NMT models are widely used in different applications.", "However, faulty outputs of such models may pose serious threats, especially in securityimportant applications.", "Adversarial attacks against NMT models have been studied in the recent literature."], "citing_paper_content": {"title": "Targeted Adversarial Attacks Against Neural Machine Translation", "abstract": "Neural Machine Translation (NMT) systems are used in various applications. However, it has been shown that they are vulnerable to very small perturbations of their inputs, known as adversarial attacks. In this paper, we propose a new targeted adversarial attack against NMT models. In particular, our goal is to insert a predefined target keyword into the translation of the adversarial sentence while maintaining similarity between the original sentence and the perturbed one in the source domain. To this aim, we propose an optimization problem, including an adversarial loss term and a similarity term. We use gradient projection in the embedding space to craft an adversarial sentence. Experimental results show that our attack outperforms Seq2Sick, the other targeted adversarial attack against NMT models, in terms of success rate and decrease in translation quality. Our attack succeeds in inserting a keyword into the translation for more than 75% of sentences while similarity with the original sentence stays preserved 1 ."}, "cited_paper_content": {"title": "Intriguing Properties Of Neural Networks", "abstract": "Deep neural networks are highly expressive models that have recently achieved state of the art performance on speech and visual recognition tasks. While their expressiveness is the reason they succeed, it also causes them to learn uninterpretable solutions that could have counter-intuitive properties. In this paper we report two such properties. ::: First, we find that there is no distinction between individual high level units and random linear combinations of high level units, according to various methods of unit analysis. It suggests that it is the space, rather than the individual units, that contains of the semantic information in the high layers of neural networks. ::: Second, we find that deep neural networks learn input-output mappings that are fairly discontinuous to a significant extend. We can cause the network to misclassify an image by applying a certain imperceptible perturbation, which is found by maximizing the network's prediction error. In addition, the specific nature of these perturbations is not a random artifact of learning: the same perturbation can cause a different network, that was trained on a different subset of the dataset, to misclassify the same input."}, "keywords": ["adversarial attacks", "Deep Neural Networks"], "citation_intent": "background"}
{"citing_id": "2303.10385v1", "cited_id": "1706.03762", "section_title": "C. Vector Encoder", "citation": "We use a transformer #REFR encoder composed of 6 layers to model the high-order interactions on these features.", "text_before_citation": ["1) Polyline Encoder: All the three types of polylines are encoded by a shared polyline encoder based on attention mechanism. As illustrated in Fig.", "3 , given a polyline After a MSA layer, the information of all other tokens is aggregated on the learnable token, which is treated as polyline feature h P .", "P = {v 1 , v 2 , \u2022 \u2022 \u2022 , v", "2) Interaction-Aware Transformer: Social interaction modeling is a standard method to capture the interactions between agents #OTHEREFR , #OTHEREFR .", "We go one step further by jointly modeling the interactions between agents, scene context, and occlusion for the inference task."], "text_after_citation": ["Each transformer layer consists of a MSA block followed by a position-wise fully connected feedforward network (FFN).", "A residual connection is added after each block, followed by a LN."], "citing_paper_content": {"title": "Social Occlusion Inference With Vectorized Representation For Autonomous Driving", "abstract": "Autonomous vehicles must be capable of handling the occlusion of the environment to ensure safe and efficient driving. In urban environment, occlusion often arises due to other vehicles obscuring the perception of the ego vehicle. Since the occlusion condition can impact the trajectories of vehicles, the behavior of other vehicles is helpful in making inferences about the occlusion as a remedy for perceptual deficiencies. This paper introduces a novel social occlusion inference approach that learns a mapping from agent trajectories and scene context to an occupancy grid map (OGM) representing the view of ego vehicle. Specially, vectorized features are encoded through the polyline encoder to aggregate features of vectors into features of polylines. A transformer module is then utilized to model the high-order interactions of polylines. Importantly, occlusion queries are proposed to fuse polyline features and generate the OGM without the input of visual modality. To verify the performance of vectorized representation, we design a baseline based on a fully transformer encoder-decoder architecture mapping the OGM with occlusion and historical trajectories information to the ground truth OGM. We evaluate our approach on an unsignalized intersection in the INTERACTION dataset, which outperforms the state-of-the-art results."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["transformer encoder"], "citation_intent": "method"}
{"citing_id": "2303.00191v1", "cited_id": "1504.04909", "section_title": "Definition.", "citation": "The vanilla MAP-Elites #REFR mutates randomly sampled solutions in the archive with a genetic operator; generated solutions are added to the archive if their objective value exceeds that of the solution currently occupying their corresponding archive cell.", "text_before_citation": ["The QD objective is to find, for each \u2208 , a solution \u2208 R such that ( ) = and ( ) is maximized:", "EQUATION", "However, since is continuous, this objective would require infinite memory to solve, so we relax the QD objective to finding an archive (i.e., a finite set) of representative solutions \u0398 \u2286 R .", "A special case of QD is differentiable quality diversity (DQD) #OTHEREFR , where the objective and measure functions are first-order differentiable with gradients \u2207 and \u2207 .", "Algorithms based on MAP-Elites #OTHEREFR tessellate the measure space into cells, and \u0398 is constrained such that each of its solutions falls into a different cell of the tessellation based on its measure values."], "text_after_citation": ["Since its inception, MAP-Elites extensions have included new genetic operators, such as the Iso+LineDD operator inspired by crossover #OTHEREFR , as well as new methods for tessellating the measure space to create the archive.", "For example, MAP-Elites with Sliding Boundaries (MESB) adapts the size of grid cells online to reflect the distribution of solutions in measure space #OTHEREFR , while CVT-MAP-Elites #OTHEREFR precomputes a centroidal Voronoi tessellation (CVT) #OTHEREFR of the measure space that defines the archive cells.", "Algorithms based on Novelty Search #OTHEREFR maintain an unstructured archive where each solution must be novel by being a certain distance away from its nearest neighbors in measure space.", "A genetic algorithm then optimizes a population of solutions to achieve further novelty.", "While Novelty Search itself is a purely diversity-driven approach, many of its successors are designed for QD; for instance, Novelty Search with Local Competition (NSLC) #OTHEREFR balances between optimizing for the objective and novelty via multi-objective evolutionary algorithms. QD algorithms have started to incorporate modern optimization algorithms."], "citing_paper_content": {"title": "Pyribs: A Bare-Bones Python Library For Quality Diversity Optimization", "abstract": "Recent years have seen a rise in the popularity of quality diversity (QD) optimization, a branch of optimization that seeks to find a collection of diverse, high-performing solutions to a given problem. To grow further, we believe the QD community faces two challenges: developing a framework to represent the field's growing array of algorithms, and implementing that framework in software that supports a range of researchers and practitioners. To address these challenges, we have developed pyribs, a library built on a highly modular conceptual QD framework. By replacing components in the conceptual framework, and hence in pyribs, users can compose algorithms from across the QD literature; equally important, they can identify unexplored algorithm variations. Furthermore, pyribs makes this framework simple, flexible, and accessible, with a userfriendly API supported by extensive documentation and tutorials. This paper overviews the creation of pyribs, focusing on the conceptual framework that it implements and the design principles that have guided the library's development. Pyribs is available at https://pyribs.org CCS CONCEPTS \u2022 Computing methodologies \u2192 Search methodologies; \u2022 Software and its engineering \u2192 Software libraries and repositories."}, "cited_paper_content": {"title": "Illuminating Search Spaces By Mapping Elites", "abstract": "Many fields use search algorithms, which automatically explore a search space to find high-performing solutions: chemists search through the space of molecules to discover new drugs; engineers search for stronger, cheaper, safer designs, scientists search for models that best explain data, etc. The goal of search algorithms has traditionally been to return the single highest-performing solution in a search space. Here we describe a new, fundamentally different type of algorithm that is more useful because it provides a holistic view of how high-performing solutions are distributed throughout a search space. It creates a map of high-performing solutions at each point in a space defined by dimensions of variation that a user gets to choose. This Multi-dimensional Archive of Phenotypic Elites (MAP-Elites) algorithm illuminates search spaces, allowing researchers to understand how interesting attributes of solutions combine to affect performance, either positively or, equally of interest, negatively. For example, a drug company may wish to understand how performance changes as the size of molecules and their cost-to-produce vary. MAP-Elites produces a large diversity of high-performing, yet qualitatively different solutions, which can be more helpful than a single, high-performing solution. Interestingly, because MAP-Elites explores more of the search space, it also tends to find a better overall solution than state-of-the-art search algorithms. We demonstrate the benefits of this new algorithm in three different problem domains ranging from producing modular neural networks to designing simulated and real soft robots. Because MAP- Elites (1) illuminates the relationship between performance and dimensions of interest in solutions, (2) returns a set of high-performing, yet diverse solutions, and (3) improves finding a single, best solution, it will advance science and engineering."}, "keywords": ["objective value", "genetic operator"], "citation_intent": "method"}
{"citing_id": "2303.00897v1", "cited_id": "1806.00582", "section_title": "Preliminaries", "citation": "Based on the findings with Non-IID data #REFR , we expect datasets with similar data distributions to provide similar \u03a8(\u2022) values.", "text_before_citation": ["The federated client clustering results, i.e., the subjective term in Equation 1, determines the performance of CFL to some extent.", "Therefore, Equation 1motivates that the federated client clustering is a vital component in CFL.", "We design a distribution extractor function \u03a8(D) = Normalize( \u2202 (\u03c8;D) \u2202\u03c8 ), which indicates the updated direction toward the local minimum corresponding for the input dataset D, anchor model \u03c8 and loss function .", "We do not optimize the anchor model \u03c8 and maintain the loss function constant across all datasets in our implementations.", "As a result, the \u03a8(\u2022) output can be viewed as a representation of data distribution corresponding to the input dataset."], "text_after_citation": ["Then, we use cosine similarity to evaluate the distribution similarity of the two decentralized datasets, i.e., given any two unknown datasets D i , D j , the similarity is determined as:", "cos(\u03a8(D i ), \u03a8(D j )) = \u03a8(D i ) \u2022 \u03a8(D j ) \u03a8(D i ) \u03a8(D j ) .", "To better support our assumptions, we implement observation experiments on cosine similarity, as shown in Figure 2 , in which we augment MNIST/Fashion-MNIST dataset and partition them with varying levels of augmentation.", "The results reveal a significant difference in cosine similarity values.", "As a result, we conclude that \u03a8(\u2022) could represent a local data distribution, and clients with similar local data distributions (at both label and feature levels) have higher cosine similarity."], "citing_paper_content": {"title": "Stochastic Clustered Federated Learning", "abstract": "Federated learning is a distributed learning framework that takes full advantage of private data samples kept on edge devices. In real-world federated learning systems, these data samples are often decentralized and Non-Independently Identically Distributed (Non-IID), causing divergence and performance degradation in the federated learning process. As a new solution, clustered federated learning groups federated clients with similar data distributions to impair the Non-IID effects and train a better model for every cluster. This paper proposes StoCFL, a novel clustered federated learning approach for generic Non-IID issues. In detail, StoCFL implements a flexible CFL framework that supports an arbitrary proportion of client participation and newly joined clients for a varying FL system, while maintaining a great improvement in model performance. The intensive experiments are conducted by using four basic Non-IID settings and a real-world dataset. The results show that StoCFL could obtain promising cluster results even when the number of clusters is unknown. Based on the client clustering results, models trained with StoCFL outperform baseline approaches in a variety of contexts."}, "cited_paper_content": {"title": "Federated Learning With Non-Iid Data", "abstract": "Federated learning enables resource-constrained edge compute devices, such as mobile phones and IoT devices, to learn a shared model for prediction, while keeping the training data local. This decentralized approach to train models provides privacy, security, regulatory and economic benefits. In this work, we focus on the statistical challenge of federated learning when local data is non-IID. We first show that the accuracy of federated learning reduces significantly, by up to 55% for neural networks trained for highly skewed non-IID data, where each client device trains only on a single class of data. We further show that this accuracy reduction can be explained by the weight divergence, which can be quantified by the earth mover's distance (EMD) between the distribution over classes on each device and the population distribution. As a solution, we propose a strategy to improve training on non-IID data by creating a small subset of data which is globally shared between all the edge devices. Experiments show that accuracy can be increased by 30% for the CIFAR-10 dataset with only 5% globally shared data."}, "keywords": ["datasets"], "citation_intent": "result"}
{"citing_id": "2303.05245v1", "cited_id": "1907.11346", "section_title": "Prior Work", "citation": "For example the errors of single view methods which estimate the position of people relative to the camera is on the order of 120mm where 100mm is in the depth direction #REFR . This is consistent with the presence of scale/depth ambiguity.", "text_before_citation": ["When estimating 3d position from camera data there is an inherent ambiguity due to scale/depth ambiguity.", "Many types of interesting types of objects have a scale ambiguity, such as people, cars and animals.", "The existence of scale ambiguity is reflected in the performance on competitive datasets."], "text_after_citation": ["In the multi view setting where the depth ambiguity can be eliminated the state of the art is on the order of 17mm #OTHEREFR .", "There are many ways to resolve the scale/depth ambiguity.", "One way to solve the problem is to add a depth sensor such as a lidar #OTHEREFR or structured light #OTHEREFR .", "However adding additional sensors come with several drawbacks such as price, complexity, range among others.", "For these reasons we will focus on methods which only use camera data."], "citing_paper_content": {"title": "Probabilistic 3D Regression With Projected Huber Distribution", "abstract": "Estimating probability distributions which describe where an object is likely to be from camera data is a task with many applications. In this work we describe properties which we argue such methods should conform to. We also design a method which conform to these properties. In our experiments we show that our method produces uncertainties which correlate well with empirical errors. We also show that the mode of the predicted distribution outperform our regression baselines. The code for our implementation is available online Focal Length Model P(x,y,z;) P(x/z,y/z;) P(x, z;) Figure 1: Overview of out method. Our method takes an image as input and the focal length which was used to capture the scene and produces a log concave probability distribution in world coordinates in a way which can model the ambiguities which are inherent for camera sensors. On the left we have visualized the level curves in projected coordinates with the projected ground truth center of the cylinder. On the right we show the level curves of the distribution from a birds eye view. The estimated mode is shown as a cross and the ground truth position is a dot."}, "cited_paper_content": {"title": "Camera Distance-Aware Top-Down Approach For 3D Multi-Person Pose Estimation From A Single Rgb Image", "abstract": "Although significant improvement has been achieved recently in 3D human pose estimation, most of the previous methods only treat a single-person case. In this work, we firstly propose a fully learning-based, camera distance-aware top-down approach for 3D multi-person pose estimation from a single RGB image. The pipeline of the proposed system consists of human detection, absolute 3D human root localization, and root-relative 3D single-person pose estimation modules. Our system achieves comparable results with the state-of-the-art 3D single-person pose estimation models without any groundtruth information and significantly outperforms previous 3D multi-person pose estimation methods on publicly available datasets. ::: The code is available in this https URL , this https URL."}, "keywords": ["scale/depth ambiguity", "position"], "citation_intent": "result"}
{"citing_id": "2303.12671v1", "cited_id": "1405.0312", "section_title": "Existing Datasets And Methods For Visual Question Answering", "citation": "The Microsoft COCO dataset #REFR is one of the large-scale datasets that impact many studies in computer vision tasks, including object detection, image classification, image captioning, and visual question answering.", "text_before_citation": ["In computer vision, the research purpose for VQA is to make computers understand the semantic context of images."], "text_after_citation": ["Several VQA datasets are built on the MS-COCO in different languages, such as the VQA #OTHEREFR , VQAv2 #OTHEREFR in English, FM-IQA #OTHEREFR for Chinese, the Japanese VQA #OTHEREFR for Japanese, and the ViVQA #OTHEREFR for Vietnamese.", "There are also other two benchmark datasets for training and fine-tuning VQA methods, including Visual Genome (VG-QA) [15] and GQA #OTHEREFR . VG-QA is a VQA dataset that contains real-world photographs.", "It is designed and constructed to emphasize the interactions and relationship between natural questions and particular regions on the images.", "The creation of VG-QA lays the groundwork for building GQA, another large VQA collection that make use of Visual Genome scene graph structures to feature compositional question answering and real world reasoning.", "Besides, in the natural language processing field, the SQuAD dataset #OTHEREFR has boosted many studies in question-answering and natural language understanding."], "citing_paper_content": {"title": "Integrating Image Features With Convolutional Sequence-To-Sequence Network For Multilingual Visual Question Answering", "abstract": "Visual Question Answering (VQA) is a task that requires computers to give correct answers for the input questions based on the images. This task can be solved by humans with ease but is a challenge for computers. The VLSP2022-EVJVQA shared task carries the Visual Question Answering task in the multilingual domain on a newly released dataset: UIT-EVJVQA, in which the questions and answers are written in three different languages: English, Vietnamese and Japanese. We approached the challenge as a sequenceto-sequence learning task, in which we integrated hints from pre-trained state-of-the-art VQA models and image features with Convolutional Sequence-to-Sequence network to generate the desired answers. Our results obtained up to 0.3442 by F1 score on the public test set, 0.4210 on the private test set, and placed 3 rd in the competition."}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["Microsoft COCO dataset"], "citation_intent": "background"}
{"citing_id": "2303.12304v1", "cited_id": "1812.11703", "section_title": "I. Introduction", "citation": "DW-Xcorr #REFR is a common similarity calculation method in the currently popular siamese network based trackers.", "text_before_citation": ["Specifically, in the process of similarity calculation, the generated similarity response map is difficult to focus on the target region, which will directly affect the effectiveness of feature decoding in the subsequent tracking head.", "Secondly, the classification branch and the regression branch in the tracking head are separate in processing the task.", "Specifically, the classification branch is responsible for distinguishing the target from the background, while the regression branch is responsible for locating the bounding box of all positive samples and does not consider the classification information.", "It results in the accuracy misalignment between the output feature maps.", "On the one hand, most of the methods #OTHEREFR [13] #OTHEREFR treat each channel in the feature map equally in the process of channel downscaling, making it difficult to focus the similarity response maps on the target region."], "text_after_citation": ["It convolves the two feature maps extracted by the siamese network channel by channel and outputs the final similarity response map.", "The similarity response map has a feature that objects of the same category have a higher response on the same channel, while the response of other channels is suppressed.", "However, existing siamese network based trackers usually use a modified ResNet-50 as the feature extraction network.", "As a result, the number of channels of the final output feature map is too large, which leads to an elevated computational effort and makes it difficult to meet the real-time requirements of object tracking.", "In order to reduce the computational effort, they #OTHEREFR [13] #OTHEREFR use 1\u00d71 convolution to decrease the feature map's channels."], "citing_paper_content": {"title": "Siamthn: Siamese Target Highlight Network For Visual Tracking", "abstract": "Siamese network based trackers develop rapidly in the field of visual object tracking in recent years. The majority of siamese network based trackers now in use treat each channel in the feature maps generated by the backbone network equally, making the similarity response map sensitive to background influence and hence challenging to focus on the target region. Additionally, there are no structural links between the classification and regression branches in these trackers, and the two branches are optimized separately during training. Therefore, there is a misalignment between the classification and regression branches, which results in less accurate tracking results. In this paper, a Target Highlight Module is proposed to help the generated similarity response maps to be more focused on the target region. To reduce the misalignment and produce more precise tracking results, we propose a corrective loss to train the model. The two branches of the model are jointly tuned with the use of corrective loss to produce more reliable prediction results. Experiments on 5 challenging benchmark datasets reveal that the method outperforms current models in terms of performance, and runs at 38 fps, proving its effectiveness and efficiency."}, "cited_paper_content": {"title": "Siamrpn++: Evolution Of Siamese Visual Tracking With Very Deep Networks", "abstract": "Siamese network based trackers formulate tracking as convolutional feature cross-correlation between target template and searching region. However, Siamese trackers still have accuracy gap compared with state-of-the-art algorithms and they cannot take advantage of feature from deep networks, such as ResNet-50 or deeper. In this work we prove the core reason comes from the lack of strict translation invariance. By comprehensive theoretical analysis and experimental validations, we break this restriction through a simple yet effective spatial aware sampling strategy and successfully train a ResNet-driven Siamese tracker with significant performance gain. Moreover, we propose a new model architecture to perform depth-wise and layer-wise aggregations, which not only further improves the accuracy but also reduces the model size. We conduct extensive ablation studies to demonstrate the effectiveness of the proposed tracker, which obtains currently the best results on four large tracking benchmarks, including OTB2015, VOT2018, UAV123, and LaSOT. Our model will be released to facilitate further studies based on this problem."}, "keywords": ["network based trackers", "currently popular siamese"], "citation_intent": "method"}
{"citing_id": "2303.16406v1", "cited_id": "1906.03327", "section_title": "Dataset Analysis", "citation": "Our videos and text queries are collected from the HowTo100M #REFR dataset, and hence our category labels match theirs. As shown in Fig.", "text_before_citation": ["Task Category Distribution."], "text_after_citation": ["2 , the most frequently occurring categories (for all text-video pairs and just videos with step captions) are \"Hobbies and Crafts\", \"Food and Entertaining\", and \"Home and Garden\".", "While these are the most common categories (similar to HowTo100M's most common categories), other categories still have a presence in our dataset.", "Dataset Statistics.", "We collected a total of 3.4K text-video pairs, which are 287 seconds long on average, with a total duration of 270 hours.", "Out of 3.4K videos, 1.8K videos are clippable to a moment; i.e., only a short clip (<75% of the original video) is relevant to the text query."], "citing_paper_content": {"title": "Hierarchical Video-Moment Retrieval And Step-Captioning", "abstract": "There is growing interest in searching for information from large video corpora. Prior works have studied relevant tasks, such as text-based video retrieval, moment retrieval, video summarization, and video captioning in isolation, without an end-to-end setup that can jointly search from video corpora and generate summaries. Such an endto-end setup would allow for many interesting applications, e.g., a text-based search that finds a relevant video from a video corpus, extracts the most relevant moment from that video, and segments the moment into important steps with captions. To address this, we present the HIREST (HIerarchical REtrieval and STep-captioning) dataset and propose a new benchmark that covers hierarchical information retrieval and visual/textual stepwise summarization from an instructional video corpus. HIREST consists of 3.4K text-video pairs from an instructional video dataset, where 1.1K videos have annotations of moment spans relevant to text query and breakdown of each moment into key instruction steps with caption and timestamps (totaling 8.6K step captions). Our hierarchical benchmark consists of video retrieval, moment retrieval, and two novel moment segmentation and step captioning tasks. In moment segmentation, models break down a video moment into instruction steps and identify start-end boundaries. In step captioning, models generate a textual summary for each step. We also present starting point task-specific and end-to-end joint baseline models for our new benchmark. While the baseline models show some promising results, there still exists large room for future improvement by the community."}, "cited_paper_content": {"title": "Howto100M: Learning A Text-Video Embedding By Watching Hundred Million Narrated Video Clips", "abstract": "Learning text-video embeddings usually requires a dataset of video clips with manually provided captions. However, such datasets are expensive and time consuming to create and therefore difficult to obtain on a large scale. In this work, we propose instead to learn such embeddings from video data with readily available natural language annotations in the form of automatically transcribed narrations. The contributions of this work are three-fold. First, we introduce HowTo100M: a large-scale dataset of 136 million video clips sourced from 1.22M narrated instructional web videos depicting humans performing and describing over 23k different visual tasks. Our data collection procedure is fast, scalable and does not require any additional manual annotation. Second, we demonstrate that a text-video embedding trained on this data leads to state-of-the-art results for text-to-video retrieval and action localization on instructional video datasets such as YouCook2 or CrossTask. Finally, we show that this embedding transfers well to other domains: fine-tuning on generic Youtube videos (MSR-VTT dataset) and movies (LSMDC dataset) outperforms models trained on these datasets alone. Our dataset, code and models will be publicly available at: www.di.ens.fr/willow/research/howto100m/."}, "keywords": ["videos"], "citation_intent": "method"}
{"citing_id": "2303.16178v1", "cited_id": "1902.01954", "section_title": "C. Baselines", "citation": "For our implementation, we use GRUs instead of LSTMs because they are much faster while providing comparable performance #REFR .", "text_before_citation": ["We use six baselines in this paper.", "We chose each baseline because it represents a family of similar approaches or is a well-cited approach used as a baseline in many papers.", "attendgru This baseline is a simple unidirectional RNNbased attentional neural encoder-decoder architecture.", "It takes only source code tokens as encoder input and English comment as decoder input. It was first introduced by Iyer et al.", "#OTHEREFR as an off-the-shelf NMT/NLG approach to generate source code summaries."], "text_after_citation": ["transformer This baseline is another simple encoderdecoder architecture, but it replaces the recurrent layers with stacked muti-head attention layers #OTHEREFR .", "As mentioned in Section II-B, transformers introduce a position embedding layer that captures the sequential order of tokens which allows the multi-head attention layer to process the entire sequence at the same time.", "On the encoder side, the multi-head attention layer computes dot-product based self-attention on the source code tokens.", "On the decoder side, there are two multi-head attention layers: a masked multi-head attention layer that computes self-attention on the comment tokens followed by a regular multi-head attention layer that computes attention between the encoder and the masked attention layer.", "ast-attendgru This baseline is an enhancement over the attendgru model by including AST information on the encoder side along with source code tokens."], "citing_paper_content": {"title": "Label Smoothing Improves Neural Source Code Summarization", "abstract": "Label smoothing is a regularization technique for neural networks. Normally neural models are trained to an output distribution that is a vector with a single 1 for the correct prediction, and 0 for all other elements. Label smoothing converts the correct prediction location to something slightly less than 1, then distributes the remainder to the other elements such that they are slightly greater than 0. A conceptual explanation behind label smoothing is that it helps prevent a neural model from becoming \"overconfident\" by forcing it to consider alternatives, even if only slightly. Label smoothing has been shown to help several areas of language generation, yet typically requires considerable tuning and testing to achieve the optimal results. This tuning and testing has not been reported for neural source code summarization-a growing research area in software engineering that seeks to generate natural language descriptions of source code behavior. In this paper, we demonstrate the effect of label smoothing on several baselines in neural code summarization, and conduct an experiment to find good parameters for label smoothing and make recommendations for its use."}, "cited_paper_content": {"title": "A Neural Model For Generating Natural Language Summaries Of Program Subroutines", "abstract": "Source code summarization -- creating natural language descriptions of source code behavior -- is a rapidly-growing research topic with applications to automatic documentation generation, program comprehension, and software maintenance. Traditional techniques relied on heuristics and templates built manually by human experts. Recently, data-driven approaches based on neural machine translation have largely overtaken template-based systems. But nearly all of these techniques rely almost entirely on programs having good internal documentation; without clear identifier names, the models fail to create good summaries. In this paper, we present a neural model that combines words from code with code structure from an AST. Unlike previous approaches, our model processes each data source as a separate input, which allows the model to learn code structure independent of the text in code. This process helps our approach provide coherent summaries in many cases even when zero internal documentation is provided. We evaluate our technique with a dataset we created from 2.1m Java methods. We find improvement over two baseline techniques from SE literature and one from NLP literature."}, "keywords": ["LSTMs"], "citation_intent": "method"}
{"citing_id": "2304.03730v1", "cited_id": "1909.04054", "section_title": "Dialog Encoder", "citation": "Our method to encode dialog data is inspired by the Sequential Sentence Classification (SCC) model #REFR which also is based on BERT encoder and organizes the dialog data in the hierarchical manner.", "text_before_citation": [", and regard it as the representation of the ith utterance.", "The representation of head token, T [CLS] , would be used in the following Gated Mechanism part as a kind of context information which encodes all the utterances.", "With the utterance representations of T i[SEP ] , the role {r 1 , ..., r i } and intent {e i , ..., e i } information (one-hot vectors) are concatenated to their corresponding utterance representations.", "After a MLP operation, we can obtain the final representation of each utterance by", "T new i[SEP ] = M LP (r i \u2295 e i \u2295 T i[SEP ] )."], "text_after_citation": ["However, our model is different from SCC model in two aspects.", "Firstly, the representation of head token [CLS] is additionally utilized as context information in the following modules.", "Secondly, our encoder can integrate the extra role and intent information."], "citing_paper_content": {"title": "Gated Mechanism Enhanced Multi-Task Learning For Dialog Routing", "abstract": "Currently, human-bot symbiosis dialog systems, e.g., pre-and after-sales in E-commerce, are ubiquitous, and the dialog routing component is essential to improve the overall efficiency, reduce human resource cost, and enhance user experience. Although most existing methods can fulfil this requirement, they can only model single-source dialog data and cannot effectively capture the underlying knowledge of relations among data and subtasks. In this paper, we investigate this important problem by thoroughly mining both the data-totask and task-to-task knowledge among various kinds of dialog data. To achieve the above targets, we propose a Gated Mechanism enhanced Multi-task Model (G3M), specifically including a novel dialog encoder and two tailored gated mechanism modules. The proposed method can play the role of hierarchical information filtering and is non-invasive to existing dialog systems. Based on two datasets collected from real world applications, extensive experimental results demonstrate the effectiveness of our method, which achieves the state-of-the-art performance by improving 8.7%/11.8% on RMSE metric and 2.2%/4.4% on F1 metric."}, "cited_paper_content": {"title": "Pretrained Language Models For Sequential Sentence Classification", "abstract": "As a step toward better document-level understanding, we explore classification of a sequence of sentences into their corresponding categories, a task that requires understanding sentences in context of the document. Recent successful models for this task have used hierarchical models to contextualize sentence representations, and Conditional Random Fields (CRFs) to incorporate dependencies between subsequent labels. In this work, we show that pretrained language models, BERT (Devlin et al., 2018) in particular, can be used for this task to capture contextual dependencies without the need for hierarchical encoding nor a CRF. Specifically, we construct a joint sentence representation that allows BERT Transformer layers to directly utilize contextual information from all words in all sentences. Our approach achieves state-of-the-art results on four datasets, including a new dataset of structured scientific abstracts."}, "keywords": ["dialog data", "Sequential Sentence Classification"], "citation_intent": "method"}
{"citing_id": "2303.15472v1", "cited_id": "1805.09662", "section_title": "Computational Overhead And The Number Of Parameters", "citation": "When using our model with the deeper backbone denoted denoted by ours \u2020, the number of model parameters increases, but it does not increase significantly compared to other comparison groups, where is still similar to that of LF-Net #REFR .", "text_before_citation": ["The right table shows the number of parameters in millions, where the first group (top) are descriptor-only models and the second group (bottom) are joint detection and description models.", "Our model in the first row has a second smallest model size among those of descriptoronly models."], "text_after_citation": [], "citing_paper_content": {"title": "Learning Rotation-Equivariant Features For Visual Correspondence", "abstract": "Extracting discriminative local features that are invariant to imaging variations is an integral part of establishing correspondences between images. In this work, we introduce a self-supervised learning framework to extract discriminative rotation-invariant descriptors using groupequivariant CNNs. Thanks to employing group-equivariant CNNs, our method effectively learns to obtain rotationequivariant features and their orientations explicitly, without having to perform sophisticated data augmentations. The resultant features and their orientations are further processed by group aligning, a novel invariant mapping technique that shifts the group-equivariant features by their orientations along the group dimension. Our group aligning technique achieves rotation-invariance without any collapse of the group dimension and thus eschews loss of discriminability. The proposed method is trained end-to-end in a self-supervised manner, where we use an orientation alignment loss for the orientation estimation and a contrastive descriptor loss for robust local descriptors to geometric/photometric variations. Our method demonstrates state-of-the-art matching accuracy among existing rotationinvariant descriptors under varying rotation and also shows competitive results when transferred to the task of keypoint matching and camera pose estimation."}, "cited_paper_content": {"title": "Lf-Net: Learning Local Features From Images", "abstract": "We present a novel strategy to learn a local feature pipeline from collections of images with deep networks, without the need for human supervision. To do so we frame the learning problem with a two-branch network. We posit that training both branches with a standard Siamese architecture is not feasible, as solving the correspondence problem jointly with feature learning is too challenging, and does not converge well enough to train from scratch. Instead, we propose to break differentiability on one branch and use ground-truth geometry to remove the burden of solving the correspondence problem while training, while keeping the other fully-differentiable. In order to train this setup with gradient-based methods, we optimize for the differentiable branch while using the parameters from the previous training step for the other, and demonstrate that both converge to a single, optimal solution. Our method can be trained with only the relative camera pose and depth information---furthermore, we show that this ground truth does not need to be perfect and can be easily obtained with off-the-shelf Structure-from-Motion solutions. Our models outperform the state of the art on sparse feature matching on both indoor and outdoor datasets, while running at 60+ fps for QVGA images."}, "keywords": ["model", "LF-Net"], "citation_intent": "result"}
{"citing_id": "2304.00595v1", "cited_id": "1912.01244", "section_title": "Vi. Solving The Conditions Of Optimality Using A Modified Physics Informed Neural Network", "citation": "IV, computing the uncontrolled PDFs for the deterministic dynamics (i.e., \u03b4 = 0) requires inverting #REFR .", "text_before_citation": ["Additionally, to satisfy compute constraints, we uniformly randomly sampled 35,000 samples every 40,000 epochs.", "For computing the Sinkhorn losses at the endpoint boundary conditions, we use the entropic regularization parameter (see (11)) \u03b5 = 0.1. Fig.", "2 depicts fifty optimally controlled state sample paths for this simulation.", "These sample paths are obtained via closed-loop simulation with the optimal control policy u opt resulting from the training of the PINN. Fig.", "3 shows the snaphsots of the univariate marginal PDFs under optimal control and the same without control, for the aforesaid numerical simulation. Following Sec."], "text_after_citation": ["We used the methodof-characteristics #OTHEREFR to solve the corresponding unforced Liouville PDE, thereby obtaining the uncontrolled joint PDF snapshots.", "The marginals \u03c1 unc i , i \u2208 3 , in Fig. 3 were obtained by numerically integrating these uncontrolled joints."], "citing_paper_content": {"title": "Optimal Mass Transport Over The Euler Equation", "abstract": "We consider the finite horizon optimal steering of the joint state probability distribution subject to the angular velocity dynamics governed by the Euler equation. The problem and its solution amounts to controlling the spin of a rigid body via feedback, and is of practical importance, for example, in angular stabilization of a spacecraft with stochastic initial and terminal states. We clarify how this problem is an instance of the optimal mass transport (OMT) problem with bilinear prior drift. We deduce both static and dynamic versions of the Eulerian OMT, and provide analytical and numerical results for the synthesis of the optimal controller."}, "cited_paper_content": {"title": "Wasserstein Proximal Algorithms For The Schr\\\"{O}Dinger Bridge Problem: Density Control With Nonlinear Drift", "abstract": "We study the Schrodinger bridge problem (SBP) with nonlinear prior dynamics. In control-theoretic language, this is a problem of minimum effort steering of a given joint state probability density function (PDF) to another over a finite time horizon, subject to a controlled stochastic differential evolution of the state vector. For generic nonlinear drift, we reduce the SBP to solving a system of forward and backward Kolmogorov partial differential equations (PDEs) that are coupled through the boundary conditions, with unknowns being the \"Schrodinger factors\" -- so named since their product at any time yields the optimal controlled joint state PDF at that time. We show that if the drift is a gradient vector field, or is of mixed conservative-dissipative nature, then it is possible to transform these PDEs into a pair of initial value problems (IVPs) involving the same forward Kolmogorov operator. Combined with a recently proposed fixed point recursion that is contractive in the Hilbert metric, this opens up the possibility to numerically solve the SBPs in these cases by computing the Schrodinger factors via a single IVP solver for the corresponding (uncontrolled) forward Kolmogorov PDE. The flows generated by such forward Kolmogorov PDEs, for the two aforementioned types of drift, in turn, enjoy gradient descent structures on the manifold of joint PDFs with respect to suitable distance functionals. We employ a proximal algorithm developed in our prior work, that exploits this geometric viewpoint, to solve these IVPs and compute the Schrodinger factors via weighted scattered point cloud evolution in the state space. We provide the algorithmic details and illustrate the proposed framework of solving the SBPs with nonlinear prior dynamics by numerical examples."}, "keywords": ["deterministic dynamics"], "citation_intent": "method"}
{"citing_id": "2304.05600v1", "cited_id": "1807.09840", "section_title": "A. External Results", "citation": "Similarly, we show results for the TUT18 task #REFR , sampled from the HARES benchmark, in Table A2 .", "text_before_citation": ["Here, we review results for our primary downstream tasks from external, high-performing models.", "We provide these results to contextualize performance on these tasks further, and compare to our strongest result in each case.", "In Table A1 , we assemble a collection of recent top results on VGGSound #OTHEREFR .", "We restrict ourselves to those with which we share an evaluation metric (Top-1 Accuracy, which is common in recent VGGSound evaluations).", "These methods involve any combination of supervision during pretraining, highly optimized state-of-the-art architectures like multimodal transformers, and larger pretraining sets."], "text_after_citation": ["Here, we compare against all results reported in the original HARES paper #OTHEREFR .", "The difference between the LibriSpeechpretrained [79] Wav2Vec 2.0 #OTHEREFR performance (which is very low) and the other methods, which are trained on Au-dioSet #OTHEREFR , illustrates the importance of pretraining data to this particular task.", "Our performance is competitive, but several AudioSet approaches with tuned architectures outperform our result, suggesting that our fully self-supervised pretraining and the relatively modest amount of pretraining data that we use may play a role. Arch."], "citing_paper_content": {"title": "Looking Similar, Sounding Different: Leveraging Counterfactual Cross-Modal Pairs For Audiovisual Representation Learning", "abstract": "2 [chihweiw, iorife, mkalayeh]@netflix.com Figure 1. (Left) Audiovisual scenes can be perceptually similar even as the words spoken in them differ, which may be a challenge to self-supervised audiovisual representation learning. (Right) We propose to leverage movie dubs during training and show that it improves the quality of learned representations on a wide range of tasks."}, "cited_paper_content": {"title": "A Multi-Device Dataset For Urban Acoustic Scene Classification", "abstract": "This paper introduces the acoustic scene classification task of DCASE 2018 Challenge and the TUT Urban Acoustic Scenes 2018 dataset provided for the task, and evaluates the performance of a baseline system in the task. As in previous years of the challenge, the task is defined for classification of short audio samples into one of predefined acoustic scene classes, using a supervised, closed-set classification setup. The newly recorded TUT Urban Acoustic Scenes 2018 dataset consists of ten different acoustic scenes and was recorded in six large European cities, therefore it has a higher acoustic variability than the previous datasets used for this task, and in addition to high-quality binaural recordings, it also includes data recorded with mobile devices. We also present the baseline system consisting of a convolutional neural network and its performance in the subtasks using the recommended cross-validation setup."}, "keywords": ["TUT18 task"], "citation_intent": "result"}
{"citing_id": "2304.14676v1", "cited_id": "1609.08138", "section_title": "1) From The", "citation": "Now it is easily verified that [0 N I N ]P \u22121 \u03c0 in (24) is the same as the matrix in #REFR . V.", "text_before_citation": ["EQUATION", "Therefore, this N -sum box is feasible according to #OTHEREFR .", "For the specified G, H matrices, the transfer matrix M Q is,", "EQUATION", "Recall that P \u22121 \u03c0 = P \u03c0 \u22121 ."], "text_after_citation": ["CONCLUSION Using the N -sum box abstraction #OTHEREFR , this work translated existing CSA coding schemes for classical MAC settings into Quantum CSA coding schemes, for entanglementassisted QMACs.", "This leads to immediate applications to Quantum PIR with secure and MDS-coded storage (Q-MDS-XSTPIR), as well as Quantum SDBMM.", "In both cases, the rate achieved with the QCSA scheme can be expressed as R Q = min{1, 2R C }, where R C is the rate achieved by the CSA scheme in the corresponding classical setting.", "Recent results in QPIR, QTPIR, QMDSTPIR can be recovered as special cases.", "An important direction for future work is to explore settings where certain qudits are erased or lost, which would extend the scope of QCSA schemes to allow stragglers."], "citing_paper_content": {"title": "Quantum Cross Subspace Alignment Codes Via The N -Sum Box Abstraction", "abstract": "Cross-subspace alignment (CSA) codes are used in various private information retrieval (PIR) schemes (e.g., with secure storage) and in secure distributed batch matrix multiplication (SDBMM). Using a recently developed N-sum box abstraction of a quantum multiple-access channel (QMAC), we translate CSA schemes over classical multiple-access channels into efficient quantum CSA schemes over a QMAC, achieving maximal superdense coding gain. Because of the N-sum box abstraction, the underlying problem of coding to exploit quantum entanglements for CSA schemes, becomes conceptually equivalent to that of designing a channel matrix for a MIMO MAC subject to given structural constraints imposed by the N-sum box abstraction, such that the resulting MIMO MAC is able to implement the functionality of a CSA scheme (encoding/decoding) over-the-air. Applications include Quantum PIR with secure and MDS-coded storage, as well as Quantum SDBMM."}, "cited_paper_content": {"title": "The Capacity Of Private Information Retrieval From Coded Databases", "abstract": "We consider the problem of private information retrieval (PIR) over a distributed storage system. The storage system consists of $N$ non-colluding databases, each storing a coded version of $M$ messages. In the PIR problem, the user wishes to retrieve one of the available messages without revealing the message identity to any individual database. We derive the information-theoretic capacity of this problem, which is defined as the maximum number of bits of the desired message that can be privately retrieved per one bit of downloaded information. We show that the PIR capacity in this case is $C=\\left(1+\\frac{K}{N}+\\frac{K^2}{N^2}+\\cdots+\\frac{K^{M-1}}{N^{M-1}}\\right)^{-1}=(1+R_c+R_c^2+\\cdots+R_c^{M-1})^{-1}=\\frac{1-R_c}{1-R_c^M}$, where $R_c$ is the rate of the $(N,K)$ code used. The capacity is a function of the code rate and the number of messages only regardless of the explicit structure of the storage code. The result implies a fundamental tradeoff between the optimal retrieval cost and the storage cost. The result generalizes the achievability and converse results for the classical PIR with replicating databases to the case of coded databases."}, "keywords": ["matrix", "N ]P \u22121"], "citation_intent": "background"}
{"citing_id": "2303.07853v1", "cited_id": "1512.04150", "section_title": "Introduction", "citation": "To the best of our knowledge, the current state-of-the-art in image-level-based WSSS methods use class activation maps (CAMs) #REFR to generate the pixel-level masks of an object from its image-level label.", "text_before_citation": ["Therefore, to reduce the time and resources required for generating pixel-wise masks, a wide range of research works focus on developing approaches that focus on weaker kinds of supervision.", "This is where Weakly Supervised Semantic Segmentation (WSSS) can be highly beneficial.", "WSSS approaches focus on generating the required masks with minimum supervision, such as image-level labels #OTHEREFR , bounding boxes #OTHEREFR , point annotations #OTHEREFR , and scribbles #OTHEREFR .", "This work focuses on generating semantic segmentation masks for medical images using the most straightforward, and least-supervised, image-level labels.", "We limit the scope of our work to just using image-level labels since they are the most inexpensive form of annotation."], "text_after_citation": ["The central idea of CAMs is to use any model trained with classification loss to generate activation maps that highlight the image regions responsible for the prediction decision.", "This results mostly in a rough localization of the objects rather than precise pixel-wise masks.", "The most popular CAM approaches focus on adding regularization loss to improve the quality of the CAM prediction #OTHEREFR or utilizing refinement methods that aim to enhance the CAM afterward #OTHEREFR .", "For example, adversarial erasing #OTHEREFR erases the most discriminative part of the CAM to force the model to consider different parts of an object. #OTHEREFR", "(Chang et al., 2020) use clustering to automatically sub-divide every class into sub-classes, implicitly generating distinctive classes for less discriminative parts of the CAM."], "citing_paper_content": {"title": "Boundarycam: A Boundary-Based Refinement Framework For Weakly Supervised Semantic Segmentation Of Medical Images", "abstract": "Weakly Supervised Semantic Segmentation (WSSS) with only image-level supervision is a promising approach to deal with the need for Segmentation networks, especially for generating a large number of pixel-wise masks in a given dataset. However, most stateof-the-art image-level WSSS techniques lack an understanding of the geometric features embedded in the images since the network cannot derive any object boundary information from just image-level labels. We define a boundary here as the line separating an object and its background, or two different objects. To address this drawback, we propose our novel BoundaryCAM framework, which deploys state-of-the-art class activation maps combined with various post-processing techniques in order to achieve fine-grained higher-accuracy segmentation masks. To achieve this, we investigate a state-of-the-art unsupervised semantic segmentation network that can be used to construct a boundary map, which enables BoundaryCAM to predict object locations with sharper boundaries. By applying our method to WSSS predictions, we were able to achieve up to 10% improvements even to the benefit of the current state-of-the-art WSSS methods for medical imaging. The framework is open-source and accessible online 1 ."}, "cited_paper_content": {"title": "Learning Deep Features For Discriminative Localization", "abstract": "In this work, we revisit the global average pooling layer proposed in [13], and shed light on how it explicitly enables the convolutional neural network (CNN) to have remarkable localization ability despite being trained on imagelevel labels. While this technique was previously proposed as a means for regularizing training, we find that it actually builds a generic localizable deep representation that exposes the implicit attention of CNNs on an image. Despite the apparent simplicity of global average pooling, we are able to achieve 37.1% top-5 error for object localization on ILSVRC 2014 without training on any bounding box annotation. We demonstrate in a variety of experiments that our network is able to localize the discriminative image regions despite just being trained for solving classification task1."}, "keywords": ["image-level label", "class activation maps"], "citation_intent": "method"}
{"citing_id": "2304.03717v1", "cited_id": "1512.03385", "section_title": "Experimental Results", "citation": "For the image part, we use a pre-trained ResNet101 #REFR , followed by the same layers.", "text_before_citation": ["Besides the simulation results reported in Figure 1 , we also conduct experiments on the MSCOCO-2014 dataset #OTHEREFR using more practical models. See Figure 2 for the results.", "For the text part, we use a pre-trained RoBERTa model #OTHEREFR , followed by a 3-layer fully-connected network with batch norm between layers."], "text_after_citation": ["In both parts, the width of the fully-connected layers and the output dimension are 768.", "We freeze the pre-trained parts of the model and only train the fully-connected parts.", "We measure the quality of the learned representation using its zero-shot performance on the MSCOCO-2014 validation set.", "Unlike common image classification datasets, images in the MSCOCO dataset usually have multiple labels, each corresponding to an object that appears in the image, and there are 80 categories in total.", "We regard a prediction to be correct if it matches one label."], "citing_paper_content": {"title": "On The Importance Of Contrastive Loss In Multimodal Learning", "abstract": "Recently, contrastive learning approaches (e.g., CLIP (Radford et al., 2021)) have received huge success in multimodal learning, where the model tries to minimize the distance between the representations of different views (e.g., image and its caption) of the same data point while keeping the representations of different data points away from each other. However, from a theoretical perspective, it is unclear how contrastive learning can learn the representations from different views efficiently, especially when the data is not isotropic. In this work, we analyze the training dynamics of a simple multimodal contrastive learning model and show that contrastive pairs are important for the model to efficiently balance the learned representations. In particular, we show that the positive pairs will drive the model to align the representations at the cost of increasing the condition number, while the negative pairs will reduce the condition number, keeping the learned representations balanced."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["pre-trained ResNet101"], "citation_intent": "method"}
{"citing_id": "2303.15965v1", "cited_id": "2002.08546", "section_title": "Classification: Organamnist", "citation": "SHOT #REFR showed comparable performance to SFHarmony when trained with a batchsize of 500 (85.27%), but was highly dependent on the modified source training.", "text_before_citation": ["Thus, robustness to the batchsize is vital if a SFDA method is to be used for harmonisation.", "We use a single source model to allow fair comparison, trained with a batchsize of 50.", "We wish to maximise performance across all sites: as harmonisation is normally framed as a joint domain adaptation problem #OTHEREFR , the average performance across all sites is reported.", "The results can be seen in Table 1 , alongside the baseline methods.", "It can be seen that SFHarmony outperforms the existing SFDA methods, especially when a small batchsize was used for training (86.22% for batchsize 5)."], "text_after_citation": ["Interestingly, several of the SFDA approaches outperformed the adversarial approaches despite them having access to the source data, possibly due to the instability of such approaches.", "The proposed D GM M loss is clearly able to align the features across sites using only the GMM summary statistics. This is demonstrated by Fig.", "2 , which shows the source and target features for each site before and after DA.", "Clearly the features overlap much more after DA, which both leads to the clear improvement in performance, and shows that the approach is achieving the harmonisation goals of the model having a shared feature embedding across sites.", "We tried modelling the features with K \u2208 {1, 2, 3} GMM components: visual inspection of the features suggested that at least 2 components would be beneficial."], "citing_paper_content": {"title": "Sfharmony: Source Free Domain Adaptation For Distributed Neuroimaging Analysis", "abstract": "To represent the biological variability of clinical neuroimaging populations, it is vital to be able to combine data across scanners and studies. However, different MRI scanners produce images with different characteristics, resulting in a domain shift known as the 'harmonisation problem'. Additionally, neuroimaging data is inherently personal in nature, leading to data privacy concerns when sharing the data. To overcome these barriers, we propose an Unsupervised Source-Free Domain Adaptation (SFDA) method, SFHarmony. Through modelling the imaging features as a Gaussian Mixture Model and minimising an adapted Bhattacharyya distance between the source and target features, we can create a model that performs well for the target data whilst having a shared feature representation across the data domains, without needing access to the source data for adaptation or target labels. We demonstrate the performance of our method on simulated and real domain shifts, showing that the approach is applicable to classification, segmentation and regression tasks, requiring no changes to the algorithm. Our method outperforms existing SFDA approaches across a range of realistic data scenarios, demonstrating the potential utility of our approach for MRI harmonisation and general SFDA problems. Our code is available at https://github.com/ nkdinsdale/SFHarmony."}, "cited_paper_content": {"title": "Do We Really Need To Access The Source Data? Source Hypothesis Transfer For Unsupervised Domain Adaptation", "abstract": "Unsupervised domain adaptation (UDA) aims to leverage the knowledge learned from a labeled source dataset to solve similar tasks in a new unlabeled domain. Prior UDA methods typically require to access the source data when learning to adapt the model, making them risky and inefficient for decentralized private data. In this work we tackle a novel setting where only a trained source model is available and investigate how we can effectively utilize such a model without source data to solve UDA problems. To this end, we propose a simple yet generic representation learning framework, named \\emph{Source HypOthesis Transfer} (SHOT). Specifically, SHOT freezes the classifier module (hypothesis) of the source model and learns the target-specific feature extraction module by exploiting both information maximization and self-supervised pseudo-labeling to implicitly align representations from the target domains to the source hypothesis. In this way, the learned target model can directly predict the labels of target data. We further investigate several techniques to refine the network architecture to parameterize the source model for better transfer performance. To verify its versatility, we evaluate SHOT in a variety of adaptation cases including closed-set, partial-set, and open-set domain adaptation. Experiments indicate that SHOT yields state-of-the-art results among multiple domain adaptation benchmarks."}, "keywords": ["modified source training"], "citation_intent": "result"}
{"citing_id": "2304.02945v1", "cited_id": "1201.0490", "section_title": "Svm And Binary Relevance", "citation": "We used the support vector machines implementation in the \"scikit-learn\" library for Python #REFR . For binary relevance, we employed the \"one vs.", "text_before_citation": [], "text_after_citation": ["rest\" classification approach using support vector machines as the base learner with the usual threshold, 0.5.", "For both a linear and a radial kernel, we performed a grid search to determine a suitable values for the hyperparameters C and gamma (gamma applies only to the radial kernel) to minimize 0/1 loss. The x-variables are the TF-IDF variables.", "The models were trained on the training data and evaluated on the validation data.", "For binary relevance, the grid search determined a linear kernel with C = 100."], "citing_paper_content": {"title": "Multi-Label Classification Of Open-Ended Questions With Bert", "abstract": "Open-ended questions in surveys are valuable because they do not constrain the respondent's answer, thereby avoiding biases. However, answers to open-ended questions are text data which are harder to analyze. Traditionally, answers were manually classified as specified in the coding manual. In the last 10 years, researchers have tried to automate coding. Most of the effort has gone into the easier problem of single label prediction, where answers are classified into a single code. However, open-ends that require multi-label classification, i.e., that are assigned multiple codes, occur frequently. In social science surveys, such open-ends are also frequently mildly multi-label. In mildly multi-label classifications, the average number of labels per answer text is relatively low (e.g. < 1.5). For example, the data set we analyze asks \"What do you think is the most important political problem in Germany at the moment?\" Even though the question asks for a single problem, some answers contain multiple problems. Of course, the average number of problems (or labels) per answer is still low. This paper focuses on multi-label classification of text answers to open-ended survey questions in social science surveys. We evaluate the performance of the transformer-based architecture BERT for the German language in comparison to traditional multi-label algorithms (Binary Relevance, Label Powerset, ECC) in a German social science survey, the GLES Panel (N=17,584, 55 labels). Because our data set requires at least one label per text answer, we also propose a modification in case the classification methods fail to predict any labels. We evaluate the algorithms on 0/1 loss: zero loss occurs only when all labels are predicted correctly; a mistake on one label incurs the full loss (1). This loss corresponds to the reality of manual text classification: you code the whole text answer with all labels, even if only a single suspicious label requires review. We find that classification with BERT (forcing at least one label) has the smallest 0/1 loss (13.1%) among methods considered (18.9%-21.6%). As expected, it is much easier to correctly predict answer texts that correspond to a single label (7.1% loss) than those that correspond to multiple labels (\u223c50% loss). Because BERT predicts zero labels for only 1.5% of the answers, forcing at least one label, while successful and recommended, ultimately does not lower the 0/1 loss by much. Our work has important implications for social scientists: 1) We have shown multi-label classification with BERT works in the German language for open-ends. 2) For mildly multi-label classification tasks, the loss now appears small enough to allow for fully automatic classification. Previously, the loss was more substantial, usually requiring semi-automatic approaches. 3) Multilabel classification with BERT requires only a single model. The leading competitor, ECC, is an iterative approach that iterates through individual single label predictions."}, "cited_paper_content": {"title": "Scikit-Learn: Machine Learning In Python", "abstract": "Scikit-learn is a Python module integrating a wide range of state-of-the-art machine learning algorithms for medium-scale supervised and unsupervised problems. This package focuses on bringing machine learning to non-specialists using a general-purpose high-level language. Emphasis is put on ease of use, performance, documentation, and API consistency. It has minimal dependencies and is distributed under the simplified BSD license, encouraging its use in both academic and commercial settings. Source code, binaries, and documentation can be downloaded from http://scikit-learn.sourceforge.net."}, "keywords": ["binary relevance", "\"scikit-learn\" library"], "citation_intent": "method"}
{"citing_id": "2304.11685v1", "cited_id": "1812.00194", "section_title": "V. Discussion", "citation": "In the report, they note several of the same observations regarding race and age. Similar findings were made in #REFR .", "text_before_citation": ["They do see a performance increase by fine-tuning a system on their child database.", "These results are comparable with the results observed in this work, but this dataset has the benefit of being synthetic.", "It was also observed how subjects of Black and Asian race in general perform worse than ones of White and Latino-Hispanic.", "It was further seen that all races have a performance decrease as they get younger. In #OTHEREFR , Grother et al.", "have performed a vendor test with a specific focus on the performance and bias of commercial face recognition systems concerning demographics."], "text_after_citation": ["Overall the results indicate that facial recognition systems are not robust to younger subjects and that racial and gender bias is a general problem across age groups.", "In figure 15 , pairs of subjects from ages 7-10 with a high non-mated score can be seen.", "As seen, the pairs of subjects has the same gender and race.", "Similarly, pairs of subjects from the youngest age group (ages 1-4) with a high non-mated score can be seen in figure 16 .", "In this particular age group, false matches across gender and race have also been observed."], "citing_paper_content": {"title": "Child Face Recognition At Scale: Synthetic Data Generation And Performance Benchmark", "abstract": "We address the need for a large-scale database of children's faces by using generative adversarial networks (GANs) and face age progression (FAP) models to synthesize a realistic dataset referred to as HDA-SynChildFaces. To this end, we proposed a processing pipeline that initially utilizes StyleGAN3 to sample adult subjects, which are subsequently progressed to children of varying ages using InterFaceGAN. Intra-subject variations, such as facial expression and pose, are created by further manipulating the subjects in their latent space. Additionally, the presented pipeline allows to evenly distribute the races of subjects, allowing to generate a balanced and fair dataset with respect to race distribution. The created HDA-SynChildFaces consists of 1,652 subjects and a total of 188,832 images, each subject being present at various ages and with many different intra-subject variations. Subsequently, we evaluates the performance of various facial recognition systems on the generated database and compare the results of adults and children at different ages. The study reveals that children consistently perform worse than adults, on all tested systems, and the degradation in performance is proportional to age. Additionally, our study uncovers some biases in the recognition systems, with Asian and Black subjects and females performing worse than White and Latino Hispanic subjects and males."}, "cited_paper_content": {"title": "Racial Faces In The Wild: Reducing Racial Bias By Information Maximization Adaptation Network", "abstract": "Racial bias is an important issue in biometric, but has not been thoroughly studied in deep face recognition. In this paper, we first contribute a dedicated dataset called Racial Faces in-the-Wild (RFW) database, on which we firmly validated the racial bias of four commercial APIs and four state-of-the-art (SOTA) algorithms. Then, we further present the solution using deep unsupervised domain adaptation and propose a deep information maximization adaptation network (IMAN) to alleviate this bias by using Caucasian as source domain and other races as target domains. This unsupervised method simultaneously aligns global distribution to decrease race gap at domain-level, and learns the discriminative target representations at cluster level. A novel mutual information loss is proposed to further enhance the discriminative ability of network output without label information. Extensive experiments on RFW, GBU, and IJB-A databases show that IMAN successfully learns features that generalize well across different races and across different databases."}, "keywords": ["age", "race"], "citation_intent": "result"}
{"citing_id": "2304.03088v1", "cited_id": "1307.5640", "section_title": "Introduction", "citation": "Approaches that exploit online-sampling of the uncertainties for the scenario approximation of SMPC problems are known as Scenario MPC #REFR .", "text_before_citation": ["By introducing chance constraints, i.e., requiring state or output constraints to be satisfied with a prespecified probability level, SMPC allows for a systematic trade-off between control performance and constraint satisfaction.", "This is particularly important for MPC of uncertain systems when optimal performance requires operation near constraint boundaries in applications where rare or transient constraint violations are acceptable, such as in electric grids #OTHEREFR or finance #OTHEREFR .", "For safety-critical applications, safety guarantees are enabled by employing failsafe or robust backup plans #OTHEREFR .", "A major challenge in SMPC is to reformulate the chance constraint into a deterministic expression for tractability of the OCP.", "Sampling-based approaches provide an appealing remedy, as they are easy to implement, independent of the underlying probability distribution, and also allow for nonlinearity of the uncertainties in the system dynamics (Lorenzen et al., 2017) ."], "text_after_citation": ["While online-sampling comes with reduced sample complexity, offline-sampling allows for a reduced online computational load by removing redundant constraints offline, as well as for a guarantee of recursive feasibility by introducing a constraint on the first predicted step, as proposed by Lorenzen et al. (2017) for known system matrices subject to parametric uncertainties.", "On the other hand, sampling-based approaches can only guarantee chance constraint satisfaction with confidence, and the computational load increases drastically with the dimension of the system #OTHEREFR .", "In the data-driven setting, a reformulation of the chance constraint may be achieved by leveraging polynomial chaos expansion #OTHEREFR or employing stochastic tubes #OTHEREFR .", "However, both works consider systems subject to additive stochastic disturbances, and assume the available (measured) data to be exact."], "citing_paper_content": {"title": "Offline Uncertainty Sampling In Data-Driven Stochastic Mpc", "abstract": "In this work, we exploit an offline-sampling based strategy for the constrained data-driven predictive control of an unknown linear system subject to random measurement noise. The strategy uses only past measured, potentially noisy data in a non-parametric system representation and does not require any prior model identification. The approximation of chance constraints using uncertainty sampling leads to efficient constraint tightening. Under mild assumptions, robust recursive feasibility and closed-loop constraint satisfaction is shown. In a simulation example, we provide evidence for the improved control performance of the proposed control scheme in comparison to a purely robust data-driven predictive control approach."}, "cited_paper_content": {"title": "The Scenario Approach For Stochastic Model Predictive Control With Bounds On Closed-Loop Constraint Violations", "abstract": "Many practical applications in control require that constraints on the inputs and states of the system are respected, while some performance criterion is optimized. In the presence of model uncertainties or disturbances, it is often sufficient to satisfy the state constraints for at least a prescribed share of the time, such as in building climate control or load mitigation for wind turbines. For such systems, this paper presents a new method of Scenario-Based Model Predictive Control (SCMPC). The basic idea is to optimize the control inputs over a finite horizon, subject to robust constraint satisfaction under a finite number of random scenarios of the uncertainty and/or disturbances. Previous SCMPC approaches have suffered from a substantial gap between the rate of constraint violations specified in the optimal control problem and that actually observed in closed-loop operation of the controlled system. This paper identifies the two theoretical explanations for this gap. First, accounting for the special structure of the optimal control problem leads to a substantial reduction of the problem dimension. Second, the probabilistic constraints have to be interpreted as average-in-time, rather than pointwise-in-time. Based on these insights, a novel SCMPC method can be devised for general linear systems with additive and multiplicative disturbances, for which the number of scenarios is significantly reduced. The presented method retains the essential advantages of the general SCMPC approach, namely a low computational complexity and the ability to handle arbitrary probability distributions. Moreover, the computational complexity can be adjusted by a sample-and-remove strategy."}, "keywords": ["Scenario MPC"], "citation_intent": "background"}
{"citing_id": "2304.13013v1", "cited_id": "1804.04235", "section_title": "Preliminaries And Related Work", "citation": "While our analysis and methods build directly on Shazeer and Stern #REFR (AdaFactor), there are important differences.", "text_before_citation": ["These instabilities may slow learning, or even destabilize training completely.", "Various solutions have been proposed, including freezing the embedding layer #OTHEREFR , adding additional layer normalization #OTHEREFR , or reparametrizing the weights #OTHEREFR .", "In our work we investigate instabilities which arise during CLIP training.", "Unlike the instabilities observed in #OTHEREFR , we find these are not caused by attention entropy collapse.", "Instead, our results indicate that spikes arise when the second moment estimator is out of date for the networks early layers."], "text_after_citation": ["In contrast with Shazeer and Stern #OTHEREFR , who only observe instabilities without warmup, we observe instabilities despite a long warmup period.", "Moreover, in contrast with Shazeer and Stern #OTHEREFR we find that an out-of-date second moment estimator is primarily an issue for the (patch) embedding layer, and measure how well loss spikes are predicted by this event.", "Finally, we note that reserachers have moved away from AdaFactor in its original formulation for large-scale training #OTHEREFR , finding AdaFactor to under-perform AdamW #OTHEREFR .", "We believe this is due to the factored second moment or absence of first moment.", "This is why our focus is AdamW #OTHEREFR which is the de facto standard optimizer for transformers."], "citing_paper_content": {"title": "Stable And Low-Precision Training For Large-Scale Vision-Language Models", "abstract": "We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models. 1) Towards accelerating training, we introduce SwitchBack , a linear layer for int8 quantized training which provides a speed-up of 13-25% while matching the performance of bfloat16 training within 0.1 percentage points for the 1B parameter CLIP ViT-Huge-the largest int8 training to date. Our main focus is int8 as GPU support for float8 is rare, though we also analyze float8 training through simulation. While SwitchBack proves effective for float8, we show that standard techniques are also successful if the network is trained and initialized so that large feature magnitudes are discouraged, which we accomplish via layer-scale initialized with zeros. 2) Towards stable training, we analyze loss spikes and find they consistently occur 1-8 iterations after the squared gradients become underestimated by their AdamW second moment estimator. As a result, we recommend an AdamW-Adafactor hybrid, which we refer to as StableAdamW because it avoids loss spikes when training a CLIP ViT-Huge model and outperforms gradient clipping."}, "cited_paper_content": {"title": "Adafactor: Adaptive Learning Rates With Sublinear Memory Cost", "abstract": "In several recently proposed stochastic optimization methods (e.g. RMSProp, Adam, Adadelta), parameter updates are scaled by the inverse square roots of exponential moving averages of squared past gradients. Maintaining these per-parameter second-moment estimators requires memory equal to the number of parameters. For the case of neural network weight matrices, we propose maintaining only the per-row and per-column sums of these moving averages, and estimating the per-parameter second moments based on these sums. We demonstrate empirically that this method produces similar results to the baseline. Secondly, we show that adaptive methods can produce larger-than-desired updates when the decay rate of the second moment accumulator is too slow. We propose update clipping and a gradually increasing decay rate scheme as remedies. Combining these methods and dropping momentum, we achieve comparable results to the published Adam regime in training the Transformer model on the WMT 2014 English-German machine translation task, while using very little auxiliary storage in the optimizer. Finally, we propose scaling the parameter updates based on the scale of the parameters themselves."}, "keywords": ["AdaFactor"], "citation_intent": "method"}
{"citing_id": "2303.09706v1", "cited_id": "1503.02531", "section_title": "Input Image", "citation": "In addition, we also compare the semi-supervised settings following #REFR upon the same network, and the results are reported in Table 6 .", "text_before_citation": ["Ground Truth Ours Full Fully-Supervised APB w/o. Knowledge Enhancement Table 6 .", "Comparing the different training paradigms, i.e., supervised, semi-supervised and unsupervised settings.", "Table 5 , the result indicates that using the operation in Eq. 3 works best. Semi-supervised setting."], "text_after_citation": ["Specifically, we conduct two semi-supervised training schemes: 1) Semi-supervised v1 refers to training the APB using 1 4 of randomly sampled labeled data on BDD-A and then training the entire network using pseudo-labels generated from the remaining raw images; 2) Semi-supervised v2 refers to the reversed process.", "However, as is shown in Table 6 , we observe drastic drops in the result of the network in both Semi-supervised v1 and v2 compared with fully-supervised APB and are even inferior to our model trained in an unsupervised way.", "The poor performance can be explained by only using a small portion of the dataset tend to fool the model into learning a more restricted central bias, especially in self-driving.", "Our unsupervised method can leverage the information transferred from natural scenes by uncertainty mining, which is able to include more generalized information from non-traffic scenes to reduce bias.", "Figure 5 shows visual comparisons of our model's variants on the BDD-A test set."], "citing_paper_content": {"title": "Unsupervised Self-Driving Attention Prediction Via Uncertainty Mining And Knowledge Embedding", "abstract": "Predicting attention regions of interest is an important yet challenging task for self-driving systems. Existing methodologies rely on large-scale labeled traffic datasets that are labor-intensive to obtain. Besides, the huge domain gap between natural scenes and traffic scenes in current datasets also limits the potential for model training. To address these challenges, we are the first to introduce an unsupervised way to predict self-driving attention by uncertainty modeling and driving knowledge integration. Our approach's Uncertainty Mining Branch (UMB) discovers commonalities and differences from multiple generated pseudo-labels achieved from models pre-trained on natural scenes by actively measuring the uncertainty. Meanwhile, our Knowledge Embedding Block (KEB) bridges the domain gap by incorporating driving knowledge to adaptively refine the generated pseudo-labels. Quantitative and qualitative results with equivalent or even more impressive performance compared to fully-supervised state-of-the-art approaches across all three public datasets demonstrate the effectiveness of the proposed method and the potential of this direction. The code will be made publicly available."}, "cited_paper_content": {"title": "Distilling The Knowledge In A Neural Network", "abstract": "A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel."}, "keywords": ["semi-supervised settings"], "citation_intent": "result"}
{"citing_id": "2303.15395v1", "cited_id": "1806.00138", "section_title": "Vii. Numerical Results", "citation": "As a further research, we plan to find a theoretical methods of evaluating the expectations in #REFR and 7and evaluating the number of samples required to find these expectations reliably.", "text_before_citation": ["One can observe a gap of \u2248 3 dB between the achievability and converse bounds.", "With a further K a increase, this gap becomes smaller, and when K a approaches n, projection-based achievability bound shows worse results.", "To show this, we compare the ML-based and projection-based achievability for k = 100, n = 1000, L = 64", "and P e = 10 \u22123 . This scenario has been taken from #OTHEREFR .", "We also find that FASURA #OTHEREFR , a state-of-the-art practical scheme, demonstrates the energy efficiency very close to both ML-and projection-based bounds for K a \u2264 400."], "text_after_citation": ["Projection-based ML-based Single-user ach, #OTHEREFR Single-user converse, #OTHEREFR Multi-user converse, #OTHEREFR Fig. 2. Same-codebook achievability bound for no-CSI setting.", "Frame length n = 1000, the number of information bits k = 100.", "Base station is equipped with L = 64 antennas, Pe = 10 \u22123 .", "Note that the channel (1) is permutation invariant, thus we can perform such type of decoding.DRAFTMarch 28, 2023", "Note that the papers are devoted to a single antenna case.March 28, 2023 DRAFT"], "citing_paper_content": {"title": "Unsourced Random Access With The Mimo Receiver: Projection Decoding Analysis", "abstract": "We consider unsourced random access with MIMO receiver-a crucial communication scenario for future 5G/6G wireless networks. We perform a projection-based decoder analysis and derive energy efficiency achievability bounds when channel state information is unknown at transmitters and the receiver (no-CSI scenario). The comparison to the maximum-likelihood (ML) achievability bounds by Gao et al. (2023) is performed. We show that there is a region where the new bound outperforms the ML bound. The latter fact should not surprise the reader as both decoding criteria are suboptimal when considering per-user probability of error (PUPE). Moreover, transition to projection decoding allows for significant dimensionality reduction, which greatly reduces the computation time."}, "cited_paper_content": {"title": "A Coupled Compressive Sensing Scheme For Unsourced Multiple Access", "abstract": "This article introduces a novel paradigm for the unsourced multiple-access communication problem. This divide-and-conquer approach leverages recent advances in compressive sensing and forward error correction to produce a computationally efficient algorithm. Within the proposed framework, every active device first partitions its data into several sub-blocks, and subsequently adds redundancy using a systematic linear block code. Compressive sensing techniques are then employed to recover sub-blocks, and the original messages are obtained by connecting pieces together using a low-complexity tree-based algorithm. Numerical results suggest that the proposed scheme outperforms other existing practical coding schemes. Measured performance lies approximately $4.3$~dB away from the Polyanskiy achievability limit, which is obtained in the absence of complexity constraints."}, "keywords": ["number"], "citation_intent": "method"}
{"citing_id": "2304.03856v1", "cited_id": "1903.03063", "section_title": "I. Introduction", "citation": "A high number of devices attempting to connect to the network at the same time is a medium access control problem; standard random access (RA) systems are incapable of handling such a huge number of requests #REFR , and pure RA methods such as ALOHA have serious performance limitations.", "text_before_citation": ["In B5G, the base station (BS) needs to simultaneously support devices with a variety of capabilities and deployments as the number of active devices continues to increase as well as access requests for multiple services, e.g., vehicles, sensors, mobiles, etc., and applications in 5G and B5G, such as massive machine-type communication (mMTC) and crowded mobile broadband (cMBB) #OTHEREFR .", "Hence, to support massive connections a promising candidate for the next generation of multiple access techniques, non-orthogonal multiple access (NOMA) has the ability to serve in the same resource element (RE), more than one device.", "In comparison to standard orthogonal multiple access (OMA), the NOMA increases the system throughput, improves user fairness, reduces latency, and enables large connections #OTHEREFR , sharing the same orthogonal resource (time and frequency) by coding superposition in transmitter side and successive interference cancellation (SIC) in receiver side #OTHEREFR ."], "text_after_citation": ["The adoption of NOMA and SIC considerably enhances performance, allowing two or more users per time slot #OTHEREFR , #OTHEREFR .", "In (over-)crowded scenarios, the number of user connection attempts considerably outnumbers the number of available pilot sequences.", "As a result, establishing collision resolution techniques became crucial for enabling efficient communication.", "The strongest user collision resolution (SUCRe) protocol is a well-known decentralized grant-based random access (RA) protocol for crowded massive multiple-input multiple-output (MIMO) systems, which takes advantage of MIMO properties #OTHEREFR , giving preference to users with good channel conditions while harming edge users.", "Hence, the SUCRe protocol has undergone several evolutions; e.g., in #OTHEREFR a graphbased pilot access (SUCR-GBPA) protocol is proposed, enabling all users who lost contention resolution to choose a new pilot at random."], "citing_paper_content": {"title": "Improving Random Access With Noma In Mmtc Xl-Mimo", "abstract": "The extra-large multiple-input multiple-output (XL-MIMO) architecture has been recognized as a technology for giving support for the massive MTC (mMTC), providing very high-data rates in high-user density scenarios. However, the large dimension of the array increases the Rayleigh distance (d Rayl), in addition to obstacles and scatters causing spatial non-stationarities and distinct visibility regions (VRs) across the XL array extension. We investigate the random access (RA) problem in crowded XL-MIMO scenarios; the proposed grant-based random access (GB-RA) protocol combining the advantage of non-orthogonal multiple access (NOMA) and strongest user collision resolutions in extra-large arrays (SUCRe-XL) named NOMA-XL can allow access of two or three colliding users in the same XL subarray (SA) selecting the same pilot sequence. The received signal processing in a SA basis changes the d Rayl , enabling the far-field planar wavefront propagation condition, while improving the system performance. The proposed NOMA-XL GB-RA protocol is able to provide a reduction in the number of attempts to access the mMTC network while improving the average sum rate, as the number of SA increases."}, "cited_paper_content": {"title": "From 5G To 6G: Has The Time For Modern Random Access Come?", "abstract": "This short paper proposes the use of modern random access for IoT applications in 6G. A short overview of recent advances in uncoordinated medium access is provided, highlighting the gains that can be achieved by leveraging smart protocol design intertwined with advanced signal processing techniques at the receiver. The authors\u2019 vision on the benefits such schemes can yield for beyond-5G systems is presented, with the aim to trigger further discussion."}, "keywords": ["standard random access"], "citation_intent": "background"}
{"citing_id": "2305.00104v1", "cited_id": "1904.08779", "section_title": "Data Augmentations", "citation": "We also utilize the Specaugment #REFR , which involves applying spectrogram masking with a maximum time mask length of 192 frames and a maximum frequency mask length of 48 bins.", "text_before_citation": ["Our audio experiments include various augmentation techniques including Mixup #OTHEREFR and Cutmix #OTHEREFR , which are com- Figure 3 : Audio-specific Cutmix which contains only a temporal axis cut.", "monly used in audio data augmentation #OTHEREFR .", "Additionally, we introduce an audio-specific data augmentation method called audio Cutmix as shown in Figure 3 .", "Audio cutmix is similar to image Cutmix #OTHEREFR , but has an temporal axis cut because the frequency axis contains quantitative information that is not invariant as in images.", "Throughout the entire experiment training process, including the process for training the baseline models, half of the audio clips are augmented using Mixup and the other half are augmented using our audio Cutmix technique."], "text_after_citation": ["Another augmentation technique we use is random rolling, which randomly shifts each audio clip in time by a certain number of samples.", "This random shift causes the entire waveform to appear as if it was played earlier or later in time, which prevents the model from overfitting to the training data.", "We adopted the identical data augmentation pipeline as MViT #OTHEREFR for our image classification experiment, which encompasses Mixup #OTHEREFR , Cutmix #OTHEREFR , Random erasing #OTHEREFR , and Randaugment #OTHEREFR ."], "citing_paper_content": {"title": "Mmvit: Multiscale Multiview Vision Transformers", "abstract": "We present Multiscale Multiview Vision Transformers (MMViT), which introduces multiscale feature maps and multiview encodings to transformer models. Our model encodes different views of the input signal and builds several channelresolution feature stages to process the multiple views of the input at different resolutions in parallel. At each scale stage, we use a cross-attention block to fuse information across different views. This enables the MMViT model to acquire complex high-dimensional representations of the input at different resolutions. The proposed model can serve as a backbone model in multiple domains. We demonstrate the effectiveness of MMViT on audio and image classification tasks, achieving state-of-theart results."}, "cited_paper_content": {"title": "Specaugment: A Simple Data Augmentation Method For Automatic Speech Recognition", "abstract": "We present SpecAugment, a simple data augmentation method for speech recognition. SpecAugment is applied directly to the feature inputs of a neural network (i.e., filter bank coefficients). The augmentation policy consists of warping the features, masking blocks of frequency channels, and masking blocks of time steps. We apply SpecAugment on Listen, Attend and Spell networks for end-to-end speech recognition tasks. We achieve state-of-the-art performance on the LibriSpeech 960h and Swichboard 300h tasks, outperforming all prior work. On LibriSpeech, we achieve 6.8% WER on test-other without the use of a language model, and 5.8% WER with shallow fusion with a language model. This compares to the previous state-of-the-art hybrid system of 7.5% WER. For Switchboard, we achieve 7.2%/14.6% on the Switchboard/CallHome portion of the Hub5'00 test set without the use of a language model, and 6.8%/14.1% with shallow fusion, which compares to the previous state-of-the-art hybrid system at 8.3%/17.3% WER."}, "keywords": ["spectrogram masking", "Specaugment"], "citation_intent": "method"}
{"citing_id": "2303.03278v1", "cited_id": "1712.01765", "section_title": "Human Evaluation Setup", "citation": "We use best-worst-scaling (BWS) for evaluating the informativeness of the generated summaries, as this method is \"a less labor-intensive alternative to paired comparisons that has been shown to produce more reliable results than rating scales\" #REFR .", "text_before_citation": ["the percentage of summaries rated as 3-star) as the faithfulness score, and also report the distribution of summaries rated as 1, 2, and 3 stars.", "Details on qualification, payment and other aspects of the evaluation can be found in Appendix A.4.", "Informativeness.", "We also evaluate the generated summaries in terms of informativeness.", "We consider summary to be informative if its content is important and relevant, but it does not necessarily need to be long."], "text_after_citation": ["Accordingly, for each dataset, we select 200 random articles with the corresponding summaries from five systems in random order.", "We ask three annotators to select the most informative (\"best\") and the least informative (\"worst\") among the five.", "A rating per system is computed as the percentage of times it is chosen as best minus the percentage of times it is selected as worst.", "A value of 100 means that the system has been unanimously picked as \"best\", whereas a value of -100 means that the system has been unanimously picked as \"worst\".", "Additional details, as well as the screenshot of the annotation interface, are in Appendix A.4."], "citing_paper_content": {"title": "Faithfulness-Aware Decoding Strategies For Abstractive Summarization", "abstract": "Despite significant progress in understanding and improving faithfulness in abstractive summarization, the question of how decoding strategies affect faithfulness is less studied. We present a systematic study of the effect of generation techniques such as beam search and nucleus sampling on faithfulness in abstractive summarization. We find a consistent trend where beam search with large beam sizes produces the most faithful summaries while nucleus sampling generates the least faithful ones. We propose two faithfulness-aware generation methods to further improve faithfulness over current generation techniques: (1) ranking candidates generated by beam search using automatic faithfulness metrics and (2) incorporating lookahead heuristics that produce a faithfulness score on the future summary. We show that both generation methods significantly improve faithfulness across two datasets as evaluated by four automatic faithfulness metrics and human evaluation. To reduce computational cost, we demonstrate a simple distillation approach that allows the model to generate faithful summaries with just greedy decoding. 1 * * Work conducted during an internship at Amazon."}, "cited_paper_content": {"title": "Best-Worst Scaling More Reliable Than Rating Scales: A Case Study On Sentiment Intensity Annotation", "abstract": "Rating scales are a widely used method for data annotation; however, they present several challenges, such as difficulty in maintaining inter- and intra-annotator consistency. Best-worst scaling (BWS) is an alternative method of annotation that is claimed to produce high-quality annotations while keeping the required number of annotations similar to that of rating scales. However, the veracity of this claim has never been systematically established. Here for the first time, we set up an experiment that directly compares the rating scale method with BWS. We show that with the same total number of annotations, BWS produces significantly more reliable results than the rating scale."}, "keywords": ["generated summaries", "rating scales"], "citation_intent": "method"}
{"citing_id": "2303.13186v1", "cited_id": "1706.03762", "section_title": "Overview", "citation": "The multi-modal fusion module leverages attention mechanism #REFR to fuse proposal features F p , gestural features F g , and word features F l , thereby generating the confidence score of each bounding-box.", "text_before_citation": ["To better leverage the features among language, gesture, and the scene point cloud, an attention mechanism #OTHEREFR is employed in our work.", "The proposal generation module is the same as that used in 3DVG-Transformer #OTHEREFR and it generates a bounding-box from object proposal while extracting context-aware features.", "The proposal features are represented by F p \u2208 R M \u00d7H , for M proposals with H-dimensional features.", "The gestural encoding module uses a PointNet++ #OTHEREFR to extract the features F g \u2208 R M \u00d7H of the human agent's point cloud.", "Similar to ScanRefer #OTHEREFR and 3DVG-Transformer #OTHEREFR , the language encoding module aggregates the word embeddings into the language features F l \u2208 R L\u00d7H and global language features using a GRU #OTHEREFR cell and a self-attention module."], "text_after_citation": ["Specifically, this study centers on the combination of gestural information with the proposal and word information, aiming to disambiguate referring expressions and accurately identify the referred object."], "citing_paper_content": {"title": "Scaneru: Interactive 3D Visual Grounding Based On Embodied Reference Understanding", "abstract": "Aiming to link natural language descriptions to specific regions in a 3D scene represented as 3D point clouds, 3D visual grounding is a very fundamental task for human-robot interaction. The recognition errors can significantly impact the overall accuracy and then degrade the operation of AI systems. Despite their effectiveness, existing methods suffer from the difficulty of low recognition accuracy in cases of multiple adjacent objects with similar appearances. To address this issue, this work intuitively introduces the humanrobot interaction as a cue to facilitate the development of 3D visual grounding. Specifically, a new task termed Embodied Reference Understanding (ERU) is first designed for this concern. Then a new dataset called ScanERU is constructed to evaluate the effectiveness of this idea. Different from existing datasets, our ScanERU is the first to cover semi-synthetic scene integration with textual, real-world visual, and synthetic gestural information. Additionally, this paper formulates a heuristic framework based on attention mechanisms and human body movements to enlighten the research of ERU. Experimental results demonstrate the superiority of the proposed method, especially in the recognition of multiple identical objects. Our codes and dataset 1 are ready to be available publicly."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["proposal features", "multi-modal fusion module"], "citation_intent": "method"}
{"citing_id": "2304.09374v1", "cited_id": "2002.05709", "section_title": "I. Introduction", "citation": "Especially, contrastive learning method has shown great advances in image classification with massive unlabeled data. It aims to use similarities and differences between images. SimCLR #REFR is a prominent work for this approach.", "text_before_citation": ["Up until now, however, its success is mainly for upstream tasks, not for the downstream.", "It requires yet another labeled data to finetune the pre-trained models.", "In other words, it still depends on supervised learning that requires much manual effort.", "NLP downstream tasks and their performances are limited by disadvantages of supervised learning.", "In image processing, on the other hand, numerous approaches to utilize SSL have been fruitful."], "text_after_citation": ["As its pre-trained model applies to downstream task of classification, it outperforms even supervised learning models.", "But, it also need two stage process of upstream and downstream task.", "To take one step further, some works make downstream task also self-supervised, to minimize human intervention.", "They include SCAN #OTHEREFR , RUC #OTHEREFR , SelfMatch #OTHEREFR , yielding tangible results.", "There is not much work to extend SSL or unsupervised learning method to downstream task in natural language processing field. A few has tried contrastive learning in the field."], "citing_paper_content": {"title": "Shuffle & Divide: Contrastive Learning For Long Text", "abstract": "We propose a self-supervised learning method for long text documents based on contrastive learning. A key to our method is Shuffle and Divide (SaD), a simple text augmentation algorithm that sets up a pretext task required for contrastive updates to BERT-based document embedding. SaD splits a document into two sub-documents containing randomly shuffled words in the entire documents. The sub-documents are considered positive examples, leaving all other documents in the corpus as negatives. After SaD, we repeat the contrastive update and clustering phases until convergence. It is naturally a time-consuming, cumbersome task to label text documents, and our method can help alleviate human efforts, which are most expensive resources in AI. We have empirically evaluated our method by performing unsupervised text classification on the 20 Newsgroups, Reuters-21578, BBC, and BBCSport datasets. In particular, our method pushes the current state-of-the-art, SS-SB-MT, on 20 Newsgroups by 20.94% in accuracy. We also achieve the state-of-the-art performance on Reuters-21578 and exceptionally-high accuracy performances (over 95%) for unsupervised classification on the BBC and BBCSport datasets."}, "cited_paper_content": {"title": "A Simple Framework For Contrastive Learning Of Visual Representations", "abstract": "This paper presents SimCLR: a simple framework for contrastive learning of visual representations. We simplify recently proposed contrastive self-supervised learning algorithms without requiring specialized architectures or a memory bank. In order to understand what enables the contrastive prediction tasks to learn useful representations, we systematically study the major components of our framework. We show that (1) composition of data augmentations plays a critical role in defining effective predictive tasks, (2) introducing a learnable nonlinear transformation between the representation and the contrastive loss substantially improves the quality of the learned representations, and (3) contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. By combining these findings, we are able to considerably outperform previous methods for self-supervised and semi-supervised learning on ImageNet. A linear classifier trained on self-supervised representations learned by SimCLR achieves 76.5% top-1 accuracy, which is a 7% relative improvement over previous state-of-the-art, matching the performance of a supervised ResNet-50. When fine-tuned on only 1% of the labels, we achieve 85.8% top-5 accuracy, outperforming AlexNet with 100X fewer labels."}, "keywords": ["contrastive learning method"], "citation_intent": "method"}
{"citing_id": "2304.05360v1", "cited_id": "1401.6848", "section_title": "Explicit De Finetti-Style Theorems", "citation": "Compared with the earlier information-theoretic results #REFR and 3, the bound in Corollary 2.4 is both more general and stronger.", "text_before_citation": ["I(X k 1 ; Y ) = H(X k 1 ) \u2212 H(X k 1 |Y ) \u2264 H(X k 1 ) \u2264 k log |A|,", "where H(X) = \u2212 x\u2208B P (x) log P (x) denotes the entropy of a random variable with probability mass function P on a discrete alphabet B.", "Therefore, Theorem 2.1 immediately yields: is an exchangeable vector of random variables X i taking values in a discrete alphabet A.", "For every 1 \u2264 k \u2264 n \u2212 1 there exists a probability measure \u00b5 = \u00b5 k,n on M 1 (A), such that:", "D P X k 1 M k,\u00b5 \u2264 k(k \u2212 1) 2(n \u2212 k + 1) H(X 1 ) \u2264 k(k \u2212 1) 2(n \u2212 k + 1) log |A|."], "text_after_citation": ["Moreover, it can be used to recover the classical infinite version of de Finetti's theorem for compact spaces, under some conditions.", "Corollary 2.5 (Classical de Finetti theorem for compact spaces) Let G be a compact metrisable space equipped with its Baire \u03c3-algebra G.", "Suppose the process {X k ; k \u2265 1} is exchangeable and the X k take values in G and are G-measurable. If for every k we have I(X k\u22121"], "citing_paper_content": {"title": "A Third Information-Theoretic Approach To Finite De Finetti Theorems", "abstract": "A new finite form of de Finetti's representation theorem is established using elementary information-theoretic tools. The distribution of the first k random variables in an exchangeable vector of n \u2265 k random variables is close to a mixture of product distributions. Closeness is measured in terms of the relative entropy and an explicit bound is provided. This bound is tighter than those obtained via earlier information-theoretic proofs, and its utility extends to random variables taking values in general spaces. The core argument employed has its origins in the quantum information-theoretic literature."}, "cited_paper_content": {"title": "Am With Multiple Merlins", "abstract": "We introduce and study a new model of interactive proofs: AM(k), or Arthur-Merlin with k non-communicating Merlins. Unlike with the better-known MIP, here the assumption is that each Merlin receives an independent random challenge from Arthur. One motivation for this model (which we explore in detail) comes from the close analogies between it and the quantum complexity class QMA(k), but the AM(k) model is also natural in its own right. We illustrate the power of multiple Merlins by giving an AM(2) protocol for 3SAT, in which the Merlins' challenges and responses consist of only n^{1/2+o(1)} bits each. Our protocol has the consequence that, assuming the Exponential Time Hypothesis (ETH), any algorithm for approximating a dense CSP with a polynomial-size alphabet must take n^{(log n)^{1-o(1)}} time. Algorithms nearly matching this lower bound are known, but their running times had never been previously explained. Brandao and Harrow have also recently used our 3SAT protocol to show quasipolynomial hardness for approximating the values of certain entangled games. In the other direction, we give a simple quasipolynomial-time approximation algorithm for free games, and use it to prove that, assuming the ETH, our 3SAT protocol is essentially optimal. More generally, we show that multiple Merlins never provide more than a polynomial advantage over one: that is, AM(k)=AM for all k=poly(n). The key to this result is a subsampling theorem for free games, which follows from powerful results by Alon et al. and Barak et al. on subsampling dense CSPs, and which says that the value of any free game can be closely approximated by the value of a logarithmic-sized random subgame."}, "keywords": ["earlier information-theoretic results"], "citation_intent": "result"}
{"citing_id": "2303.17764v1", "cited_id": "1611.07725", "section_title": "Revisiting Distillation For Catastrophic Forgetting", "citation": "It, however, is observed in #REFR that there is tendency of classifying test samples to new classes by LwF.", "text_before_citation": ["EQUATION", "where C = C o \u222a C n , y i is the i th value of the one-hot ground truth y, and p i is the i th value of predicted class probability p.", "The goal of L dis is to preserve knowledge obtained from previous data, which is expressed as", "EQUATION", "where p * is the soft label of x generated by the old model."], "text_after_citation": ["Thus, iCaRL utilized herd selection to better approximate the class mean vector of old classes, where samples that are close to the center of old classes are selected.", "Recall that our goal is to obtain a robust model trained in the continual learning manner.", "To gain robustness, adversarial training is inevitable, which requires augmenting datasets with adversarial examples in every training iteration.", "Following the definition of continual learning, we can derive the loss function of Robust Continual Learning (RCL).", "With adversarial training, we should replace the input x in Equation 1and 2with its adversarial counterpart x adv , which is solved by"], "citing_paper_content": {"title": "Towards Adversarially Robust Continual Learning", "abstract": "Recent studies show that models trained by continual learning can achieve the comparable performances as the standard supervised learning and the learning flexibility of continual learning models enables their wide applications in the real world. Deep learning models, however, are shown to be vulnerable to adversarial attacks. Though there are many studies on the model robustness in the context of standard supervised learning, protecting continual learning from adversarial attacks has not yet been investigated. To fill in this research gap, we are the first to study adversarial robustness in continual learning and propose a novel method called Task-Aware Boundary Augmentation (TABA) to boost the robustness of continual learning models. With extensive experiments on CIFAR-10 and CIFAR-100, we show the efficacy of adversarial training and TABA in defending adversarial attacks."}, "cited_paper_content": {"title": "Icarl: Incremental Classifier And Representation Learning", "abstract": "A major open problem on the road to artificial intelligence is the development of incrementally learning systems that learn about more and more concepts over time from a stream of data. In this work, we introduce a new training strategy, iCaRL, that allows learning in such a class-incremental way: only the training data for a small number of classes has to be present at the same time and new classes can be added progressively. iCaRL learns strong classifiers and a data representation simultaneously. This distinguishes it from earlier works that were fundamentally limited to fixed data representations and therefore incompatible with deep learning architectures. We show by experiments on CIFAR-100 and ImageNet ILSVRC 2012 data that iCaRL can learn many classes incrementally over a long period of time where other strategies quickly fail."}, "keywords": ["new classes"], "citation_intent": "background"}
{"citing_id": "2304.09730v1", "cited_id": "2003.09504", "section_title": "Conclusion", "citation": "In the future, we first aim to focus on embedding the graph information #REFR in the optimization process of the proposed approach.", "text_before_citation": ["In conclusion, the high-dimensional nature and imbalanced classes of hyperspectral images pose challenges for traditional machine learning algorithms. One-class classifiers are useful in cases", "where the training data is from a single class only, but they still face challenges in handling the curse of dimensionality.", "To address these challenges, we leverage S-SVDD for one-class classification of hyperspectral images.", "Our experiments on two benchmark HSI datasets show that the proposed approach can effectively tackle the curse of dimensionality and the imbalanced nature of HSI data."], "text_after_citation": ["We also consider utilizing neighbouring pixels' spectral features for a future study, since they are usually correlated and give better scattering information, allowing enhanced performance."], "citing_paper_content": {"title": "Hyperspectral Image Analysis With Subspace Learning-Based One-Class Classification", "abstract": "Hyperspectral image (HSI) classification is an important task in many applications, such as environmental monitoring, medical imaging, and land use/land cover (LULC) classification. Due to the significant amount of spectral information from recent HSI sensors, analyzing the acquired images is challenging using traditional Machine Learning (ML) methods. As the number of frequency bands increases, the required number of training samples increases exponentially to achieve a reasonable classification accuracy, also known as the curse of dimensionality. Therefore, separate band selection or dimensionality reduction techniques are often applied before performing any classification task over HSI data. In this study, we investigate recently proposed subspace learning methods for one-class classification (OCC). These methods map high-dimensional data to a lower-dimensional feature space that is optimized for one-class classification. In this way, there is no separate dimensionality reduction or feature selection procedure needed in the proposed classification framework. Moreover, one-class classifiers have the ability to learn a data description from the category of a single class only. Considering the imbalanced labels of the LULC classification problem and rich spectral information (high number of dimensions), the proposed classification approach is well-suited for HSI data. Overall, this is a pioneer study focusing on subspace learning-based one-class classification for HSI data. We analyze the performance of the proposed subspace learning one-class classifiers in the proposed pipeline. Our experiments validate that the proposed approach helps tackle the curse of dimensionality along with the imbalanced nature of HSI data."}, "cited_paper_content": {"title": "Ellipsoidal Subspace Support Vector Data Description", "abstract": "In this paper, we propose a novel method for transforming data into a low-dimensional space optimized for one-class classification. The proposed method iteratively transforms data into a new subspace optimized for ellipsoidal encapsulation of target class data. We provide both linear and non-linear formulations for the proposed method. The method takes into account the covariance of the data in the subspace; hence, it yields a more generalized solution as compared to Subspace Support Vector Data Description for a hypersphere. We propose different regularization terms expressing the class variance in the projected space. We compare the results with classic and recently proposed one-class classification methods and achieve better results in the majority of cases. The proposed method is also noticed to converge much faster than recently proposed Subspace Support Vector Data Description."}, "keywords": ["graph information"], "citation_intent": "method"}
{"citing_id": "2304.12202v1", "cited_id": "1706.03741", "section_title": "Introduction", "citation": "OpenAI's latest conversational agent, Chat-GPT 1 (OpenAI, 2022), a successor of Instruct-GPT (Ouyang et al., 2022) -also known as GPT-3.5models, is an instruction-following transformerbased language model, which has been further trained (aligned) with reinforcement learning from human feedback (RLHF) #REFR .", "text_before_citation": ["Recent advances in Large Language Models (LLMs) #OTHEREFR Chowdhery et al., 2022) , also known as Foundation Models (Bommasani et al., 2021) , have challenged the traditional supervised learning paradigm of fine-tuning by demonstrating emergent zero-shot Natural Language Understanding (NLU) capabilities #OTHEREFR ) through scaling the model's size in billions of parameters #OTHEREFR ."], "text_after_citation": ["ChatGPT demonstrates unprecedented emergent capabilities in zero-shot Question-Answering (QA) 1 https://chat.openai.com/chat capabilities that cover common sense knowledge, but also extend to specialized domains such as problem solving, programming/debugging, and law, as presented by many users in the web.", "Recently, Bommarito and Katz (2022) audited several variants of OpenAI's GPT 2/3/3.5 models in legal bar exam questions, and found that the most advanced -at the time-model ('text-davinci-003') achieves an accuracy of 50.3% on a complete practice exam, significantly in excess of the 25% baseline guessing rate, while it performs at a passing rate in two legal areas (Evidence and Torts). In a follow-up work, #OTHEREFR", "(2023) assessed the model's performance in accounting certification exams, where the model significantly under-performs human capabilities with a correct rate of 14.4%.", "Following the work of Bommarito and Katz, we evaluate the latest OpenAI's GPT-3.5 model (Ouyang et al., 2022 ) ('gpt-3.5-turbo', v.", "March 2023 , the first available ChatGPT, on legal text classification tasks from the LexGLUE #OTHEREFR benchmark in a zeroshot fashion providing examples in a templated instruction-following format, similar to those used by Chung et al. (2022) ."], "citing_paper_content": {"title": "Chatgpt May Pass The Bar Exam Soon, But Has A Long Way To Go For The Lexglue Benchmark", "abstract": "Following the hype around OpenAI's Chat-GPT conversational agent, the last straw in the recent development of Large Language Models (LLMs) that demonstrate emergent unprecedented zero-shot capabilities, we audit the latest OpenAI's GPT-3.5 model, 'gpt-3.5-turbo', the first available ChatGPT model, in the LexGLUE benchmark in a zeroshot fashion providing examples in a templated instruction-following format. The results indicate that ChatGPT achieves an average micro-F1 score of 49.0% across LexGLUE tasks, surpassing the baseline guessing rates. Notably, the model performs exceptionally well in some datasets, achieving micro-F1 scores of 62.8% and 70.1% in the ECtHR B and LEDGAR datasets, respectively. The code base and model predictions are available for review on https://github.com/coastalcph/ zeroshot_lexglue."}, "cited_paper_content": {"title": "Deep Reinforcement Learning From Human Preferences", "abstract": "For sophisticated reinforcement learning (RL) systems to interact usefully with real-world environments, we need to communicate complex goals to these systems. In this work, we explore goals defined in terms of (non-expert) human preferences between pairs of trajectory segments. We show that this approach can effectively solve complex RL tasks without access to the reward function, including Atari games and simulated robot locomotion, while providing feedback on less than one percent of our agent's interactions with the environment. This reduces the cost of human oversight far enough that it can be practically applied to state-of-the-art RL systems. To demonstrate the flexibility of our approach, we show that we can successfully train complex novel behaviors with about an hour of human time. These behaviors and environments are considerably more complex than any that have been previously learned from human feedback."}, "keywords": ["OpenAI's latest conversational", "reinforcement learning"], "citation_intent": "method"}
{"citing_id": "2303.02400v1", "cited_id": "1902.10811", "section_title": "Related Work", "citation": "Robustness under distribution shifts Generalization capabilities of existing image classifiers have been a crucial problem #REFR , currently addressed from a few different viewpoints.", "text_before_citation": ["In general, transformer-based models rely on an abundance of training data to ensure proper generalization.", "This requirement was relaxed in DeiT #OTHEREFR , enabling learning on medium-sized datasets.", "Further development introduced novel transformer-based architectures, such as BeiT #OTHEREFR , Swin #OTHEREFR and RegNets #OTHEREFR , which realize specific refinements to boost performance.", "Overall, it has been proven that ViTs are more robust compared to classic CNN image classifiers #OTHEREFR .", "In our work, we verify the degree this claim holds by testing CNN and transformer-based classifiers on the uncurated fine-grained setting."], "text_after_citation": ["Artificial corruptions #OTHEREFR or natural shifts #OTHEREFR on curated data have already exposed biases and architectural vulnerabilities.", "Adversarial robustness #OTHEREFR ] is a related field where models are tested against adversarial examples, which introduce imperceptible though influential perturbations on images.", "Contrary to such attempts, we concentrated around naturally occurring distribution shifts stemming from uncurated image data.", "Regarding architectural choices, many studies perform robustness tests attempting to resolve the CNN vs Transformer contest #OTHEREFR , while other ventures focus on interpreting and understanding model robustness #OTHEREFR .", "In our approach, by experimenting with both CNN and transformer-based architectures we adopt such research attempts to the uncurated setting."], "citing_paper_content": {"title": "Fine-Grained Imagenet Classification In The Wild", "abstract": "Image classification has been one of the most popular tasks in Deep Learning, seeing an abundance of impressive implementations each year. However, there is a lot of criticism tied to promoting complex architectures that continuously push performance metrics higher and higher. Robustness tests can uncover several vulnerabilities and biases which go unnoticed during the typical model evaluation stage. So far, model robustness under distribution shifts has mainly been examined within carefully curated datasets. Nevertheless, such approaches do not test the real response of classifiers in the wild, e.g. when uncurated web-crawled image data of corresponding classes are provided. In our work, we perform fine-grained classification on closely related categories, which are identified with the help of hierarchical knowledge. Extensive experimentation on a variety of convolutional and transformer-based architectures reveals model robustness in this novel setting. Finally, hierarchical knowledge is again employed to evaluate and explain misclassifications, providing an information-rich evaluation scheme adaptable to any classifier."}, "cited_paper_content": {"title": "Do Imagenet Classifiers Generalize To Imagenet?", "abstract": "We build new test sets for the CIFAR-10 and ImageNet datasets. Both benchmarks have been the focus of intense research for almost a decade, raising the danger of overfitting to excessively re-used test sets. By closely following the original dataset creation processes, we test to what extent current classification models generalize to new data. We evaluate a broad range of models and find accuracy drops of 3% - 15% on CIFAR-10 and 11% - 14% on ImageNet. However, accuracy gains on the original test sets translate to larger gains on the new test sets. Our results suggest that the accuracy drops are not caused by adaptivity, but by the models' inability to generalize to slightly \"harder\" images than those found in the original test sets."}, "keywords": ["existing image classifiers"], "citation_intent": "background"}
{"citing_id": "2303.17228v1", "cited_id": "1603.00831", "section_title": "Experimental Setup", "citation": "MOT17 #REFR is a multiple object tracking dataset that contains 7 training sequences and 7 test sequences.", "text_before_citation": ["Something-Something V2 (SSv2) #OTHEREFR is another largescale action recognition dataset which focus more on temporal modeling.", "The labels are like \"Pulling something from left to right\", so it is crucial to learn motion information.", "The training set contains 168.9K training videos and the validation set contains 24.7K validation videos.", "We use segment-based sampling from #OTHEREFR to sample 32 frames with 224 \u00d7 224 resolution.", "The augmentation and regularization in SSv2 include random augmentation #OTHEREFR , repeated augmentation #OTHEREFR , random erasing #OTHEREFR , Mixup #OTHEREFR , and CutMix #OTHEREFR , which follow the practice in MViT #OTHEREFR ."], "text_after_citation": ["The total frame number is only 11k, so it is not enough to train our S-ViT model.", "We use the CrowdHuman #OTHEREFR dataset and the MOTSynth #OTHEREFR dataset to expand the training data.", "CrowdHuman contains 19.4k images in crowd human scenarios, and MOTSynth contains 764 synthetic video sequences with 1.3m frames generated from Grand Theft Auto V.", "We conduct our experiments with combinations of different data sources and discuss the influence in Sec. #OTHEREFR"], "citing_paper_content": {"title": "Streaming Video Model", "abstract": "Figure 1. Illustration of the proposed streaming video model with a comparison to conventional frame-based architecture and clip-based architecture. (a) The two-stage streaming video model gracefully serves different types of video tasks through a unified architecture. The output of the temporal-aware (T-aware) spatial encoder serves the frame-based tasks, such as MOT, while the output of the temporal decoder serves the sequence-based tasks, such as action recognition. (b) Frame-based architecture, which uses single image model to independently extract spatial features for each frame, is widely used in the frame-based video tasks. (c) Clip-based architecture, which uses video model to produce the spatiotemporal features for an entire clip, is widely used in the sequence-based video tasks."}, "cited_paper_content": {"title": "Mot16: A Benchmark For Multi-Object Tracking", "abstract": "Standardized benchmarks are crucial for the majority of computer vision applications. Although leaderboards and ranking tables should not be over-claimed, benchmarks often provide the most objective measure of performance and are therefore important guides for reseach. ::: Recently, a new benchmark for Multiple Object Tracking, MOTChallenge, was launched with the goal of collecting existing and new data and creating a framework for the standardized evaluation of multiple object tracking methods. The first release of the benchmark focuses on multiple people tracking, since pedestrians are by far the most studied object in the tracking community. This paper accompanies a new release of the MOTChallenge benchmark. Unlike the initial release, all videos of MOT16 have been carefully annotated following a consistent protocol. Moreover, it not only offers a significant increase in the number of labeled boxes, but also provides multiple object classes beside pedestrians and the level of visibility for every single object of interest."}, "keywords": ["7 training sequences", "multiple object"], "citation_intent": "background"}
{"citing_id": "2304.04932v1", "cited_id": "1807.04271", "section_title": "Motivation: Approximate Sampling-And-Query Access", "citation": "Consider for instance the inner product estimation from #REFR , a technique also used in several other \"quantum-inspired\" algorithms, which dequantizes the SWAP test (a basic quantum algorithm that can be used to estimate the inner product between two pure states).", "text_before_citation": [", n} we can query the entry u(i)), for the sampling operation we can only get one sample from a distributionp u : {1, . . .", ", n} \u2192 [0, 1] that is close to p u .", "More precisely, the only guarantee is that the total variation distance betweenp u and p u is at most \u03b5.", "(Note that standard sampling-and-query access corresponds to the case \u03b5 = 0.) This variant corresponds to the setting where query access to u can be easily implemented (e.g., when the input is stored in RAM) but perfect sampling access is not available or simply too costly to be implemented.", "It happens that this seemingly minor change is problematic for known dequantization techniques."], "text_after_citation": ["Given two unit-norm vectors u, v \u2208 R n , where u is given via sampling-and-query access, the approach typically works as follows: sample an index i \u2208 {1, . . .", ", n} according to the distribution p u and output the value v (i) u (i) . Note that the expectation of this value is", "EQUATION", "It is easy to show that the variance of this estimator is small, and thus taking the mean for a few samples gives a good approximation of the inner product (u, v) .", "Note that this test, however, is not \"robust\" due to the division by u(i)."], "citing_paper_content": {"title": "Robust Dequantization Of The Quantum Singular Value Transformation And Quantum Machine Learning Algorithms", "abstract": "Several quantum algorithms for linear algebra problems, and in particular quantum machine learning problems, have been \"dequantized\" in the past few years. These dequantization results typically hold when classical algorithms can access the data via lengthsquared sampling. This assumption, which is standard in the field of randomized linear algebra, means that for a unit-norm vector u \u2208 C n , we can sample from the distribution p u : {1,. .. , n} \u2192 [0, 1] defined as p u (i) = |u(i)| 2 for each i \u2208 {1,. .. , n}. Since this distribution corresponds to the distribution obtained by measuring the quantum state |u in the computational basis, length-squared sampling access gives a reasonable classical analogue to the kind of quantum access considered in many quantum algorithms for linear algebra problems. In this work we investigate how robust these dequantization results are. We introduce the notion of approximate length-squared sampling, where classical algorithms are only able to sample from a distribution close to the ideal distribution in total variation distance. While quantum algorithms are natively robust against small perturbations, current techniques in dequantization are not. Our main technical contribution is showing how many techniques from randomized linear algebra can be adapted to work under this weaker assumption as well. We then use these techniques to show that the recent low-rank dequantization framework by Chia, Gily\u00e9n, Li, Lin, Tang and Wang (JACM 2022) and the dequantization framework for sparse matrices by Gharibian and Le Gall (STOC 2022), which are both based on the Quantum Singular Value Transformation, can be generalized to the case of approximate length-squared sampling access to the input. We also apply these results to obtain a robust dequantization of many quantum machine learning algorithms, including quantum algorithms for recommendation systems, supervised clustering and low-rank matrix inversion."}, "cited_paper_content": {"title": "A Quantum-Inspired Classical Algorithm For Recommendation Systems", "abstract": "We give a classical analogue to Kerenidis and Prakash\u2019s quantum recommendation system, previously believed to be one of the strongest candidates for provably exponential speedups in quantum machine learning. Our main result is an algorithm that, given an m \u00d7 n matrix in a data structure supporting certain l2-norm sampling operations, outputs an l2-norm sample from a rank-k approximation of that matrix in time O(poly(k)log(mn)), only polynomially slower than the quantum algorithm. As a consequence, Kerenidis and Prakash\u2019s algorithm does not in fact give an exponential speedup over classical algorithms. Further, under strong input assumptions, the classical recommendation system resulting from our algorithm produces recommendations exponentially faster than previous classical systems, which run in time linear in m and n. The main insight of this work is the use of simple routines to manipulate l2-norm sampling distributions, which play the role of quantum superpositions in the classical setting. This correspondence indicates a potentially fruitful framework for formally comparing quantum machine learning algorithms to classical machine learning algorithms."}, "keywords": ["\"quantum-inspired\" algorithms"], "citation_intent": "method"}
{"citing_id": "2304.12406v1", "cited_id": "1604.01685", "section_title": "Segmentation", "citation": "Cityscapes #REFR is a street-view dataset with high quality annotations, containing 2975 training images and 500 validation images, with a total of 19 object classes.", "text_before_citation": ["Datasets.", "We evaluate on semantic, instance, and panoptic segmentation using 3 datasets: ADE-20K #OTHEREFR is a semantic segmentation dataset containing 150 categories across 20K training images and 2K validation images."], "text_after_citation": ["COCO 2017 #OTHEREFR is an instance segmentation dataset, containing 118K training and 5K validation images."], "citing_paper_content": {"title": "Autofocusformer: Image Segmentation Off The Grid", "abstract": "Real world images often have highly imbalanced content density. Some areas are very uniform, e.g., large patches of blue sky, while other areas are scattered with many small objects. Yet, the commonly used successive grid downsampling strategy in convolutional deep networks treats all areas equally. Hence, small objects are represented in very few spatial locations, leading to worse results in tasks such as segmentation. Intuitively, retaining more pixels representing small objects during downsampling helps to preserve important information. To achieve this, we propose AutoFocusFormer (AFF), a local-attention transformer image recognition backbone, which performs adaptive downsampling by learning to retain the most important pixels for the task. Since adaptive downsampling generates a set of pixels irregularly distributed on the image plane, we abandon the classic grid structure. Instead, we develop a novel point-based local attention block, facilitated by a balanced clustering module and a learnable neighborhood merging module, which yields representations for our point-based versions of state-of-the-art segmentation heads. Experiments show that our AutoFocusFormer (AFF) improves significantly over baseline models of similar sizes. * Work done while Chen Ziwen was an intern at Apple Inc. Image Remaining tokens Stage 2 Prediction AFF Swin Remaining tokens Stage 4 Comparison between on-grid model Swin and off-grid model AFF. AFF downsamples non automatically focusing on more textured, important image regions, and successfully captu the background. The red pixels indicate the locations of the remaining tokens."}, "cited_paper_content": {"title": "The Cityscapes Dataset For Semantic Urban Scene Understanding", "abstract": "Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations, 20 000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark."}, "keywords": ["street-view dataset"], "citation_intent": "background"}
{"citing_id": "2305.02474v1", "cited_id": "1711.05225", "section_title": "Introduction", "citation": "Similarly, deep convolutional neural networks have been shown to achieve superior performance in detecting pneumonia and other pathologies from chest X-rays, compared to practicing radiologists #REFR .", "text_before_citation": ["Academic hospitals are increasingly dedicating resources to bring machine learning (ML) to the bedside and to addressing issues encountered by clinical staff.", "These resources are being utilized across a range of applications including clinical decision support, early warning, treatment recommendation, risk prediction, image informatics, telediagnosis, drug discovery, and intelligent health knowledge systems.", "There are various examples of ML being applied to medical data, including prediction of sepsis #OTHEREFR , in-hospital mortality, prolonged length-of-stay, patient deterioration, and unplanned readmission #OTHEREFR .", "In particular, sepsis is one of the leading causes of in-hospital deaths.", "A large-scale study demonstrated the impact of an early warning system to reduce the lead time for detecting the onset of sepsis, and hence allowing more time for clinicians to prescribe antibiotics #OTHEREFR ."], "text_after_citation": ["These results highlight the potential of ML models when they are strongly integrated into clinical workflows.", "When deployed successfully, data-driven models can free time for clinicians #OTHEREFR , improve clinical outcomes #OTHEREFR , reduce costs #OTHEREFR , and provide improved quality care for patients.", "However, most studies remain preliminary, limited to small datasets, and/or implemented in select health sub-systems.", "Integrating with clinical workflows remains crucial #OTHEREFR but, despite recent computational advances and an explosion of health data, deploying ML in healthcare responsibly and reliably faces several operational and engineering challenges, including:", "\u2022 Standardizing data formats,"], "citing_paper_content": {"title": "Mlhops: Machine Learning For Healthcare Operations", "abstract": "Machine Learning Health Operations (MLHOps) is the combination of processes for reliable, efficient, usable, and ethical deployment and maintenance of machine learning models in healthcare settings. This paper provides both a survey of work in this area and guidelines for developers and clinicians to deploy and maintain their own models in clinical practice. We cover the foundational concepts of general machine learning operations, describe the initial setup of MLHOps pipelines (including data sources, preparation, engineering, and tools). We then describe long-term monitoring and updating (including data distribution shifts and model updating) and ethical considerations (including bias, fairness, interpretability, and privacy). This work therefore provides guidance across the full pipeline of MLHOps from conception to initial and ongoing deployment."}, "cited_paper_content": {"title": "Chexnet: Radiologist-Level Pneumonia Detection On Chest X-Rays With Deep Learning", "abstract": "We develop an algorithm that can detect pneumonia from chest X-rays at a level exceeding practicing radiologists. Our algorithm, CheXNet, is a 121-layer convolutional neural network trained on ChestX-ray14, currently the largest publicly available chest X-ray dataset, containing over 100,000 frontal-view X-ray images with 14 diseases. Four practicing academic radiologists annotate a test set, on which we compare the performance of CheXNet to that of radiologists. We find that CheXNet exceeds average radiologist performance on the F1 metric. We extend CheXNet to detect all 14 diseases in ChestX-ray14 and achieve state of the art results on all 14 diseases."}, "keywords": ["pneumonia", "deep convolutional neural"], "citation_intent": "background"}
{"citing_id": "2304.09718v1", "cited_id": "1910.05513", "section_title": "B. Model-Based Reinforcement Learning", "citation": "Both (a) and (b) essentially imply a sort of intrinsic robustness of the ODE flow \u03c6 t (z 0 ) to perturbations on z 0 #REFR .", "text_before_citation": ["\u2212 \u2212 \u2192 E(T ) generated by H \u03b6 \u03b6 \u03b6 for some time t \u2208 [0, T ] and propagator E.", "The ansatz is a good choice because of the following two properties of ODE paths: (a) they do not intersect and (b) if paths \u03c6", "(A) 0 , \u03c6 (B) 0 start close compared to path \u03c6 (C) 0 , then paths \u03c6 (A) t , \u03c6 (B) t remain close compared to path \u03c6 (C) t .", "Both properties are well known #OTHEREFR for ODEs and become very useful when we try to predict the trajectories from noisy quantum data by imposing strong priors on the space of learnable Hamiltonians.", "Property (b) is a consequence of Gronwall's inequality #OTHEREFR and essentially can be interpreted as: ODE flows that start off closer (w.r.t. the initial condition) stay closer (w.r.t. the final condition)."], "text_after_citation": ["They constrain the trajectories predicted by the model M \u03b6 \u03b6 \u03b6 to be intrinsically robust to small noise in the states s k and inaccuracies in the learned system Hamiltonian H", "(L) 0 (\u03b6 \u03b6 \u03b6).", "We call the SAC equipped with this differentiable ODE model the learnable Hamiltonian model-based SAC (LH-MBSAC) as listed in Algorithm 2.", "Crucially, we note that LH-MBSAC generalizes the SAC by allowing the policy to interact with the ODE model and the physical system.", "LH-MBSAC gracefully falls back to the model-free SAC in the absence of a model with low prediction error that is measured from the performance of the model's predictions on an unseen validation set of interaction data."], "citing_paper_content": {"title": "Sample-Efficient Model-Based Reinforcement Learning For Quantum Control", "abstract": "We propose a model-based reinforcement learning (RL) approach for noisy time-dependent gate optimization with improved sample complexity over model-free RL. Sample complexity is the number of controller interactions with the physical system. Leveraging an inductive bias, inspired by recent advances in neural ordinary differential equations (ODEs), we use an auto-differentiable ODE parametrised by a learnable Hamiltonian ansatz to represent the model approximating the environment whose time-dependent part, including the control, is fully known. Control alongside Hamiltonian learning of continuous time-independent parameters is addressed through interactions with the system. We demonstrate an order of magnitude advantage in the sample complexity of our method over standard model-free RL in preparing some standard unitary gates with closed and open system dynamics, in realistic numerical experiments incorporating single shot measurements, arbitrary Hilbert space truncations and uncertainty in Hamiltonian parameters. Also, the learned Hamiltonian can be leveraged by existing control methods like GRAPE for further gradient-based optimization with the controllers found by RL as initializations. Our algorithm that we apply on nitrogen vacancy (NV) centers and transmons in this paper is well suited for controlling partially characterised one and two qubit systems."}, "cited_paper_content": {"title": "On Robustness Of Neural Ordinary Differential Equations", "abstract": "Neural ordinary differential equations (ODEs) have been attracting increasing attention in various research domains recently. There have been some works studying optimization issues and approximation capabilities of neural ODEs, but their robustness is still yet unclear. In this work, we fill this important gap by exploring robustness properties of neural ODEs both empirically and theoretically. We first present an empirical study on the robustness of the neural ODE-based networks (ODENets) by exposing them to inputs with various types of perturbations and subsequently investigating the changes of the corresponding outputs. In contrast to conventional convolutional neural networks (CNNs), we find that the ODENets are more robust against both random Gaussian perturbations and adversarial attack examples. We then provide an insightful understanding of this phenomenon by exploiting a certain desirable property of the flow of a continuous-time ODE, namely that integral curves are non-intersecting. Our work suggests that, due to their intrinsic robustness, it is promising to use neural ODEs as a basic block for building robust deep network models. To further enhance the robustness of vanilla neural ODEs, we propose the time-invariant steady neural ODE (TisODE), which regularizes the flow on perturbed data via the time-invariant property and the imposition of a steady-state constraint. We show that the TisODE method outperforms vanilla neural ODEs and also can work in conjunction with other state-of-the-art architectural methods to build more robust deep networks."}, "keywords": ["perturbations", "ODE flow"], "citation_intent": "background"}
{"citing_id": "2305.01165v1", "cited_id": "1708.09832", "section_title": "Differentiation Of Forged And Authentic Paa Images", "citation": "The fool rate of the iterative refinement-based PAA image generation is 49.375%, which is much higher than the fool rate of ESRGAN, which is only 20% #REFR .", "text_before_citation": ["We employ a random cropping and shuffling method to create a dataset consisting of 40 PAA images, both authetic and forged.", "We then sent this dataset to 14 experts and asked them to differentiate between real and forged images. As shown in Fig.", "3(c) , the labels Y and N indicate whether an image is classified as a real and forged PAA image, respectively.", "Each red point or black cross represents a result from an expert, and the y-axis refers to the number of samples classified as either Y or N.", "Our results showed that the probability of authentic PAA images of human lips being classified as real is 48.04%, while the probability of placing the forged images into the real pool is 46.79%."], "text_after_citation": ["To quantitatively compare the forged PAA images and other images with high-resolution PAA images as the standard, we utilized the Fr\u00e9chet Inception Distance #OTHEREFR , which measures the distance between two images based on the mean and covariance matrices from the Inception V3 model.", "Despite not being trained on authentic PAA images, the Inception V3 model can still differentiate between blood vessel image sets using the Fr\u00e9chet Inception Distance.", "The Fr\u00e9chet Inception Distance between forged and authentic PAA images is", "EQUATION", "where , stand for the input image and the ground-truth, , are mean matrices of the input and the ground-truth, () is the trace of a matrix, and \u03a3 , \u03a3 are covariance matrices of the input and ground-truth, respectively."], "citing_paper_content": {"title": "Self-Similarity-Based Super-Resolution Of Photoacoustic Angiography From Hand-Drawn Doodles", "abstract": "Deep-learning-based super-resolution photoacoustic angiography (PAA) is a powerful tool that restores blood vessel images from under-sampled images to facilitate disease diagnosis. Nonetheless, due to the scarcity of training samples, PAA super-resolution models often exhibit inadequate generalization capabilities, particularly in the context of continuous monitoring tasks. To address this challenge, we propose a novel approach that employs a super-resolution PAA method trained with forged PAA images. We start by generating realistic PAA images of human lips from hand-drawn curves using a diffusion-based image generation model. Subsequently, we train a self-similarity-based super-resolution model with these forged PAA images. Experimental results show that our method outperforms the super-resolution model trained with authentic PAA images in both original-domain and cross-domain tests. Specially, our approach boosts the quality of super-resolution reconstruction using the images forged by the deep learning model, indicating that the collaboration between deep learning models can facilitate generalization, despite limited initial dataset. This approach shows promising potential for exploring zero-shot learning neural networks for vision tasks."}, "cited_paper_content": {"title": "Model-Based Learning For Accelerated, Limited-View 3-D Photoacoustic Tomography", "abstract": "Recent advances in deep learning for tomographic reconstructions have shown great potential to create accurate and high quality images with a considerable speed up. In this paper, we present a deep neural network that is specifically designed to provide high resolution 3-D images from restricted photoacoustic measurements. The network is designed to represent an iterative scheme and incorporates gradient information of the data fit to compensate for limited view artifacts. Due to the high complexity of the photoacoustic forward operator, we separate training and computation of the gradient information. A suitable prior for the desired image structures is learned as part of the training. The resulting network is trained and tested on a set of segmented vessels from lung computed tomography scans and then applied to in-vivo photoacoustic measurement data."}, "keywords": ["image generation"], "citation_intent": "result"}
{"citing_id": "2304.11033v1", "cited_id": "1801.10228", "section_title": "Blockchain Implementation.", "citation": "The widely used Hyperledger Fabric #REFR was not considered, as it is based on PoA-based consensus.", "text_before_citation": ["We do not depend on smart contracts, which made our selection of a concrete blockchain more flexible. Potential choices included Bitcoin #OTHEREFR and Ethereum #OTHEREFR ."], "text_after_citation": ["Between the two, Ethereum offers multiple advantages for our use case, namely a configurable hash difficulty and support for the creation of private chains [31] . Therefore, we use it for our implementation.", "As the client, we use the official implementation Go Ethereum (Geth).", "Note that, at the time of implementation, Ethereum still used PoW consensus. Recently, Ethereum switched to PoS consensus #OTHEREFR .", "When setting up a private network, it would therefore have to be configured to use PoW instead. Alternatively, a different PoW-based blockchain can be used."], "citing_paper_content": {"title": "Decentralized Inverse Transparency With Blockchain", "abstract": "Employee data can be used to facilitate work, but their misusage may pose risks for individuals. Inverse transparency therefore aims to track all usages of personal data, allowing individuals to monitor them to ensure accountability for potential misusage. This necessitates a trusted log to establish an agreed-upon and non-repudiable timeline of events. The unique properties of blockchain facilitate this by providing immutability and availability. For power asymmetric environments such as the workplace, permissionless blockchain is especially beneficial as no trusted third party is required. Yet, two issues remain: (1) In a decentralized environment, no arbiter can facilitate and attest to data exchanges. Simple peer-to-peer sharing of data, conversely, lacks the required non-repudiation. (2) With data governed by privacy legislation such as the GDPR, the core advantage of immutability becomes a liability. After a rightful request, an individual's personal data need to be rectified or deleted, which is impossible in an immutable blockchain. To solve these issues, we present Kovacs, a decentralized data exchange and usage logging system for inverse transparency built on blockchain. Its new-usage protocol ensures non-repudiation, and therefore accountability, for inverse transparency. Its one-time pseudonym generation algorithm guarantees unlinkability and enables proof of ownership, which allows data subjects to exercise their legal rights regarding their personal data. With our implementation, we show the viability of our solution. The decentralized communication impacts performance and scalability, but exchange duration and storage size are still reasonable. More importantly, the provided information security meets high requirements. We conclude that Kovacs realizes decentralized inverse transparency through secure and GDPR-compliant use of permissionless blockchain. CCS Concepts: \u2022 Computer systems organization \u2192 Peer-to-peer architectures; \u2022 Security and privacy \u2192 Distributed systems security; Privacy-preserving protocols; Cryptography."}, "cited_paper_content": {"title": "Hyperledger Fabric: A Distributed Operating System For Permissioned Blockchains", "abstract": "Hyperledger Fabric is a modular and extensible open-source system for deploying and operating permissioned blockchains. Fabric is currently used in more than 400 prototypes and proofs-of-concept of distributed ledger technology, as well as several production systems, across different industries and use cases. Starting from the premise that there are no\"one-size-fits-all\"solutions, Fabric is the first truly extensible blockchain system for running distributed applications. It supports modular consensus protocols, which allows the system to be tailored to particular use cases and trust models. Fabric is also the first blockchain system that runs distributed applications written in general-purpose programming languages, without systemic dependency on a native cryptocurrency. This stands in sharp contrast to existing blockchain platforms for running smart contracts that require code to be written in domain-specific languages or rely on a cryptocurrency. Furthermore, it uses a portable notion of membership for realizing the permissioned model, which may be integrated with industry-standard identity management. To support such flexibility, Fabric takes a novel approach to the design of a permissioned blockchain and revamps the way blockchains cope with non-determinism, resource exhaustion, and performance attacks. This paper describes Fabric, its architecture, the rationale behind various design decisions, its security model and guarantees, its most prominent implementation aspects, as well as its distributed application programming model. We further evaluate Fabric by implementing and benchmarking a Bitcoin-inspired digital currency. We show that Fabric achieves end-to-end throughput of more than 3500 transactions per second in certain popular deployment configurations, with sub-second latency."}, "keywords": ["PoA-based consensus", "Hyperledger Fabric"], "citation_intent": "method"}
{"citing_id": "2304.03931v1", "cited_id": "1911.05076", "section_title": "Introduction", "citation": "Overall, distortions produced when using Euclidean geometry for non-Euclidean geometric structures are overwhelming, causing the loss of semantic information, and hence resulting in inferior performance #REFR .", "text_before_citation": ["In fact, data in countless applications intrinsically has non-Euclidean geometric structures #OTHEREFR .", "Several studies show that non-Euclidean geometric structures can be better captured by particular forms of Riemannian geometry #OTHEREFR .", "For example, the hyperbolic geometry has a natural expressive ability for the hierarchical structure and is hence used successfully for fine-grained images #OTHEREFR .", "The spherical geometry is shown as a suitable choice for face images that have the cyclical structure #OTHEREFR .", "In addition to the geometric structures discussed above, natural data may be diverse and irregular in structure, e.g., data exhibits hierarchical forms in some regions and cyclical forms in others #OTHEREFR ."], "text_after_citation": ["In this paper, we study how to attain suitable non-Euclidean geometry to capture the intrinsic geometric structures of data during continual learning.", "To achieve our goal, we have to face two challenges (see Fig. 1 ).", "(1) Non-stationary stream of data will inevitably increase the complexity of intrinsic geometric structures.", "In other words, fixing the geometry of the underlying space cannot always match new and unseen data in continual learning.", "For example, more and more complex hierarchies in a data stream bring more leaf nodes, requiring a faster growing space volume with the radius, which conflicts with a fixed geometry #OTHEREFR ."], "citing_paper_content": {"title": "Exploring Data Geometry For Continual Learning", "abstract": "Continual learning aims to efficiently learn from a nonstationary stream of data while avoiding forgetting the knowledge of old data. In many practical applications, data complies with non-Euclidean geometry. As such, the commonly used Euclidean space cannot gracefully capture non-Euclidean geometric structures of data, leading to inferior results. In this paper, we study continual learning from a novel perspective by exploring data geometry for the non-stationary stream of data. Our method dynamically expands the geometry of the underlying space to match growing geometric structures induced by new data, and prevents forgetting by keeping geometric structures of old data into account. In doing so, making use of the mixed curvature space, we propose an incremental search scheme, through which the growing geometric structures are encoded. Then, we introduce an angular-regularization loss and a neighbor-robustness loss to train the model, capable of penalizing the change of global geometric structures and local geometric structures. Experiments show that our method achieves better performance than baseline methods designed in Euclidean space."}, "cited_paper_content": {"title": "Constant Curvature Graph Convolutional Networks", "abstract": "Interest has been rising lately towards methods representing data in non-Euclidean spaces, e.g. hyperbolic or spherical, that provide specific inductive biases useful for certain real-world data properties, e.g. scale-free, hierarchical or cyclical. However, the popular graph neural networks are currently limited in modeling data only via Euclidean geometry and associated vector space operations. Here, we bridge this gap by proposing mathematically grounded generalizations of graph convolutional networks (GCN) to (products of) constant curvature spaces. We do this by i) introducing a unified formalism that can interpolate smoothly between all geometries of constant curvature, ii) leveraging gyro-barycentric coordinates that generalize the classic Euclidean concept of the center of mass. Our class of models smoothly recover their Euclidean counterparts when the curvature goes to zero from either side. Empirically, we outperform Euclidean GCNs in the tasks of node classification and distortion minimization for symbolic data exhibiting non-Euclidean behavior, according to their discrete curvature."}, "keywords": ["non-Euclidean geometric structures"], "citation_intent": "background"}
{"citing_id": "2304.07699v1", "cited_id": "1909.02027", "section_title": "Related Works", "citation": "First, it enhances representations by incorporating strong prior knowledge from a network pre-trained on external data in the intention domain (i.e., CLINC dataset #REFR ) and adding a masked language modeling (MLM) task.", "text_before_citation": ["We have conducted a pilot study using the CDAC+ #OTHEREFR algorithm, which first captures pairwise sentence relationships with the guidance of labeled data and then refines cluster assignments with the DEC loss.", "Another of our works, DeepAligned #OTHEREFR , initializes intent representations under the supervision of labeled data and then iteratively performs clustering and representation learning, aligning cluster centroids between adjacent iterations to obtain consistent self-supervised signals.", "DCSC #OTHEREFR improves the pretraining stage by applying contrastive losses to both labeled and unlabeled data.", "It mainly uses the SwAV #OTHEREFR algorithm for unsupervised learning, which requires each sample to predict the swapped view and uses Sinkhorn-Knopp #OTHEREFR to produce soft cluster assignments.", "MTP-CLNN #OTHEREFR is the current state-of-the-art method, which has two key features."], "text_after_citation": ["Second, it adapts the SCAN algorithm #OTHEREFR to the semi-supervised setting, creating positive pairs with K-nearest neighbors or samples with the same labels for contrastive learning.", "However, this method relies heavily on the selected external data, and its performance drops dramatically in a purely unsupervised scenario #OTHEREFR .", "Open set recognition (OSR) #OTHEREFR and open intent detection #OTHEREFR , #OTHEREFR are two similar fields in computer vision and natural language processing, respectively.", "Both fields focus on detecting unseen image-based or text-based classes during testing. However, they have limitations in distinguishing fine-grained new classes.", "CD-OSR #OTHEREFR extends OSR to new class discovery, but it doesn't make use of unlabeled data during training."], "citing_paper_content": {"title": "Usnid: A Framework For Unsupervised And Semi-Supervised New Intent Discovery", "abstract": "New intent discovery is of great value to natural language processing, allowing for a better understanding of user needs and providing friendly services. However, most existing methods struggle to capture the complicated semantics of discrete text representations when limited or no prior knowledge of labeled data is available. To tackle this problem, we propose a novel framework called USNID for unsupervised and semi-supervised new intent discovery, which has three key technologies. First, it takes full use of unsupervised or semi-supervised data to mine shallow semantic similarity relations and provide well-initialized representations for clustering. Second, it designs a centroid-guided clustering mechanism to address the issue of cluster allocation inconsistency and provide high-quality self-supervised targets for representation learning. Third, it captures high-level semantics in unsupervised or semi-supervised data to discover fine-grained intent-wise clusters by optimizing both cluster-level and instance-level objectives. We also propose an effective method for estimating the cluster number in open-world scenarios without knowing the number of new intents beforehand. USNID performs exceptionally well on several intent benchmark datasets, achieving new state-of-the-art results in unsupervised and semi-supervised new intent discovery and demonstrating robust performance with different cluster numbers. Index Terms-new intent discovery, unsupervised and semi-supervised clustering, self-supervised learning, deep neural network."}, "cited_paper_content": {"title": "An Evaluation Dataset For Intent Classification And Out-Of-Scope Prediction", "abstract": "Task-oriented dialog systems need to know when a query falls outside their range of supported intents, but current text classification corpora only define label sets that cover every example. We introduce a new dataset that includes queries that are out-of-scope---i.e., queries that do not fall into any of the system's supported intents. This poses a new challenge because models cannot assume that every query at inference time belongs to a system-supported intent class. Our dataset also covers 150 intent classes over 10 domains, capturing the breadth that a production task-oriented agent must handle. We evaluate a range of benchmark classifiers on our dataset along with several different out-of-scope identification schemes. We find that while the classifiers perform well on in-scope intent classification, they struggle to identify out-of-scope queries. Our dataset and evaluation fill an important gap in the field, offering a way of more rigorously and realistically benchmarking text classification in task-driven dialog systems."}, "keywords": ["intention domain", "masked language modeling"], "citation_intent": "method"}
{"citing_id": "2303.12068v1", "cited_id": "1409.0473", "section_title": "The History Of Attention", "citation": "This notion of context is the motivation behind the introduction of the attention mechanism in 2015 #REFR .", "text_before_citation": [], "text_after_citation": ["Before this, language translation was mostly relying on encoder-decoder architectures: recurrent neural networks (RNNs) #OTHEREFR and in particular long short-term memory (LSTMs) networks were used to model the relationship among words #OTHEREFR .", "Specifically, each word of an input sentence is processed by the encoder sequentially.", "At each step, the past and present information are summarized and encoded into a fixed-length vector.", "In the end, the encoder has processed every word, and outputs a final fixed-length vector, which summarizes all input information.", "This final vector is then decoded, and finally translates the input information into the target language."], "citing_paper_content": {"title": "Transformers And Visual Transformers", "abstract": "Transformers were initially introduced for natural language processing (NLP) tasks, but fast they were adopted by most deep learning fields, including computer vision. They measure the relationships between pairs of input tokens (words in the case of text strings, parts of images for visual Transformers), termed attention. The cost is exponential with the number of tokens. For image classification, the most common Transformer Architecture uses only the Transformer Encoder in order to transform the various input tokens. However, there are also numerous other applications in which the decoder part of the traditional Transformer Architecture is also used. Here, we first introduce the Attention mechanism (Section 1), and then the Basic Transformer Block including the Vision Transformer (Section 2). Next, we discuss some improvements of visual Transformers to account for small datasets or less computation (Section 3). Finally, we introduce Visual Transformers applied to tasks other than image classification, such as detection, segmentation, generation and training without labels (Section 4) and other domains, such as video or multimodality using text or audio data (Section 5)."}, "cited_paper_content": {"title": "Neural Machine Translation By Jointly Learning To Align And Translate", "abstract": "Neural machine translation is a recently proposed approach to machine translation. Unlike the traditional statistical machine translation, the neural machine translation aims at building a single neural network that can be jointly tuned to maximize the translation performance. The models proposed recently for neural machine translation often belong to a family of encoder-decoders and consists of an encoder that encodes a source sentence into a fixed-length vector from which a decoder generates a translation. In this paper, we conjecture that the use of a fixed-length vector is a bottleneck in improving the performance of this basic encoder-decoder architecture, and propose to extend this by allowing a model to automatically (soft-)search for parts of a source sentence that are relevant to predicting a target word, without having to form these parts as a hard segment explicitly. With this new approach, we achieve a translation performance comparable to the existing state-of-the-art phrase-based system on the task of English-to-French translation. Furthermore, qualitative analysis reveals that the (soft-)alignments found by the model agree well with our intuition."}, "keywords": ["attention mechanism"], "citation_intent": "background"}
{"citing_id": "2303.17573v1", "cited_id": "1603.02754", "section_title": "Model Training, Evaluation And Explanation", "citation": "To model our dataset, we applied a standard set of regressor models like XGBoost, CatBoost, SVM, etc., and found XG-Boost #REFR to be the best-performing one.", "text_before_citation": [], "text_after_citation": ["XGBoost -a widely used algorithm for regression, classification, and ranking problems -stands for eXtreme Gradient Boosting, and it implements a gradient boosting decision trees algorithm.", "XGBoost-regressor is an implementation of the XGBoost algorithm for regression problems.", "It works by building a series of decision trees where each tree tries to correct the errors made by the previous tree.", "In the end, the algorithm combines the results of all trees to make the final prediction.", "We conducted an extensive hyperparameter search of the XGBoost regressor by experimenting with different learning rates, max depth of the tree, number of estimators to use, etc."], "citing_paper_content": {"title": "Using Ai To Measure Parkinson'S Disease Severity At Home", "abstract": "We present an artificial intelligence system to remotely assess the motor performance of individuals with Parkinson's disease (PD). Participants performed a motor task (i.e., tapping fingers) in front of a webcam, and data from 250 global participants were rated by three expert neurologists following the Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS). The neurologists' ratings were highly reliable, with an intra-class correlation coefficient (ICC) of 0.88. We developed computer algorithms to obtain objective measurements that align with the MDS-UPDRS guideline and are strongly correlated with the neurologists' ratings. Our machine learning model trained on these measures outperformed an MDS-UPDRS certified rater, with a mean absolute error (MAE) of 0.59 compared to the rater's MAE of 0.79. However, the model performed slightly worse than the expert neurologists (0.53 MAE). The methodology can be replicated for similar motor tasks, providing the possibility of evaluating individuals with PD and other movement disorders remotely, objectively, and in areas with limited access to neurological care."}, "cited_paper_content": {"title": "Xgboost: A Scalable Tree Boosting System", "abstract": "Tree boosting is a highly effective and widely used machine learning method. In this paper, we describe a scalable end-to-end tree boosting system called XGBoost, which is used widely by data scientists to achieve state-of-the-art results on many machine learning challenges. We propose a novel sparsity-aware algorithm for sparse data and weighted quantile sketch for approximate tree learning. More importantly, we provide insights on cache access patterns, data compression and sharding to build a scalable tree boosting system. By combining these insights, XGBoost scales beyond billions of examples using far fewer resources than existing systems."}, "keywords": ["SVM", "XGBoost"], "citation_intent": "method"}
{"citing_id": "2305.01275v1", "cited_id": "1606.00915", "section_title": "Scribbles", "citation": "Using the best pseudo labels from scribble prompts, DeepLab-v2 #REFR can reach 75.9% and 76.6% mIoU scores on the validation and test sets, as shown in Tab. 3.", "text_before_citation": ["2, we find that sampling 20% scribble pixels outperforms inputting all scribble pixels in an object by 6.4%.", "Besides, iteratively inputting scribble pixels of a class can further improve the performance by 3.3%.", "We analyze that iterative input is more effective for scribbles and points than image-level labels due to accurate point locations.", "Finally, when inputting the scribble pixels of one class, the scribble pixels of other classes can be regarded as negative points.", "We can see that adding negative points can further improve the quality of pseudo labels."], "text_after_citation": [], "citing_paper_content": {"title": "Segment Anything Is A Good Pseudo-Label Generator For Weakly Supervised Semantic Segmentation", "abstract": "Weakly supervised semantic segmentation with weak labels is a long-lived illposed problem. Mainstream methods mainly focus on improving the quality of pseudo labels. In this report, we attempt to explore the potential of 'prompt to masks' from the powerful class-agnostic large segmentation model, i.e., segmentanything. Specifically, different weak labels are used as prompts to the segmentanything model, generating precise class masks. The class masks are utilized to generate pseudo labels to train the segmentation networks. We have conducted extensive experiments on PASCAL VOC 2012 dataset. Experiments demonstrate that segment-anything can serve as a good pseudo-label generator. The code will be made publicly available. * * denotes equal contribution. Preprint. Under review."}, "cited_paper_content": {"title": "Deeplab: Semantic Image Segmentation With Deep Convolutional Nets, Atrous Convolution, And Fully Connected Crfs", "abstract": "In this work we address the task of semantic image segmentation with Deep Learning and make three main contributions that are experimentally shown to have substantial practical merit. First , we highlight convolution with upsampled filters, or \u2018atrous convolution\u2019, as a powerful tool in dense prediction tasks. Atrous convolution allows us to explicitly control the resolution at which feature responses are computed within Deep Convolutional Neural Networks. It also allows us to effectively enlarge the field of view of filters to incorporate larger context without increasing the number of parameters or the amount of computation. Second , we propose atrous spatial pyramid pooling (ASPP) to robustly segment objects at multiple scales. ASPP probes an incoming convolutional feature layer with filters at multiple sampling rates and effective fields-of-views, thus capturing objects as well as image context at multiple scales. Third , we improve the localization of object boundaries by combining methods from DCNNs and probabilistic graphical models. The commonly deployed combination of max-pooling and downsampling in DCNNs achieves invariance but has a toll on localization accuracy. We overcome this by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF), which is shown both qualitatively and quantitatively to improve localization performance. Our proposed \u201cDeepLab\u201d system sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 79.7 percent mIOU in the test set, and advances the results on three other datasets: PASCAL-Context, PASCAL-Person-Part, and Cityscapes. All of our code is made publicly available online."}, "keywords": ["best pseudo labels", "DeepLab-v2"], "citation_intent": "method"}
{"citing_id": "2303.16372v2", "cited_id": "1808.06651", "section_title": "Definition 6.3 (Input Lipschitzness)", "citation": "We extend the analysis of one such algorithm, the Projected Noisy Stochastic Gradient Descent (PNSGD) #REFR , restated in Algorithm 1, and establish that the algorithm satisfies (\u03b1, )-R\u00e9nyi mDP with a lower noise magnitude than that required for (\u03b1, )-R\u00e9nyi DP. Proposition 6.5.", "text_before_citation": ["Standard arguments on the composition of privacy mechanisms renders the differentially private variant of SGD (q , q\u03b4)\u2212differentially private at each step, where q = L/N is the lot size.", "We note that, by replacing the global sensitivity assumption with input Lipschitzness, the Gaussian mechanism ensures ( L , \u03b4)-mDP for appropriately scaled noise, as in Proposition 6.4.", "Thus, DP-SGD inherently incorporates metric differential privacy, with privacy accounting based on standard composition theorems #OTHEREFR .", "Metric Differential Privacy in PN-SGD When analysing the privacy of learning algorithms that require iterative updates on an intermediate solution, it is common practice to ensure privacy at each iteration and argue about the cumulative loss of privacy via composition theorems.", "Another popular direction is the theoretical analysis of Noisy Stochastic Gradient Descent to formalize privacy amplifications under certain assumptions and obtain bounds on the degradation of privacy across iterations."], "text_after_citation": ["Let K \u2282 R d be a convex set and let {f (., x)} x\u2208cX be a family of convex, \u03b2-smooth functions over K, where the gradients are L input -input Lipschitz. Furthermore, assume X is a bounded set.", "Then, for any \u03b7 \u2264 2/\u03b2 and \u03b1 > 1, initializing w 0 \u2208 K and dataset S \u2208 X n , PNSGD run with", "\u03c3 2 \u2265 2\u03b1L 2 input diam(X ) L (n\u2212t+1)", "satisfies ( L , \u03b1) metric differential privacy.", "Proof. See Appendix 9.8 Claim 6.6."], "citing_paper_content": {"title": "Non-Asymptotic Lower Bounds For Training Data Reconstruction", "abstract": "We investigate semantic guarantees of private learning algorithms for their resilience to training Data Reconstruction Attacks (DRAs) by informed adversaries. To this end, we derive non-asymptotic minimax lower bounds on the adversary's reconstruction error against learners that satisfy differential privacy (DP) and metric differential privacy (mDP). Furthermore, we demonstrate that our lower bound analysis for the latter also covers the high dimensional regime, wherein, the input data dimensionality may be larger than the adversary's query budget. Motivated by the theoretical improvements conferred by metric DP, we extend the privacy analysis of popular deep learning algorithms such as DP-SGD[1] and Projected Noisy SGD[2] to cover the broader notion of metric differential privacy."}, "cited_paper_content": {"title": "Privacy Amplification By Iteration", "abstract": "Many commonly used learning algorithms work by iteratively updating an intermediate solution using one or a few data points in each iteration. Analysis of differential privacy for such algorithms often involves ensuring privacy of each step and then reasoning about the cumulative privacy cost of the algorithm. This is enabled by composition theorems for differential privacy that allow releasing of all the intermediate results. In this work, we demonstrate that for contractive iterations, not releasing the intermediate results strongly amplifies the privacy guarantees. We describe several applications of this new analysis technique to solving convex optimization problems via noisy stochastic gradient descent. For example, we demonstrate that a relatively small number of non-private data points from the same distribution can be used to close the gap between private and non-private convex optimization. In addition, we demonstrate that we can achieve guarantees similar to those obtainable using the privacy-amplification-by-sampling technique in several natural settings where that technique cannot be applied."}, "keywords": ["lower noise magnitude"], "citation_intent": "background"}
{"citing_id": "2304.11072v1", "cited_id": "1807.04320", "section_title": "Rq2: Can Our Classifier Learn Vulnerability In A Biased Setting?", "citation": "Nevertheless, using either weighted loss or Focal Loss, the accuracy of our model drops by almost 1% compared to #REFR , indicating that these previous models are highly biased with an assumption of non-vulnerability.", "text_before_citation": ["However, \u03b3 is used as an exponent in Equation 7, so when we run RoBERTa-GCN w/ WL, \u03b3 is set to 0 in order to ignore its effect.", "Observing the last two rows in Table 5 , we find that, initially, we tested our model using Weighted Loss (WL) only, and later we tested with Focal Loss, which is a combination of weighted loss \u03b1 with the hyperparameter \u03b3.", "In both cases, we observe that our model has surpassed previous models in terms of Precision, Recall, and F1 score, indicating lower false positive and false negative rates with WL and with FL.", "However, these numbers improved slightly when compared with performance between weighted loss and Focal Loss, with the exception of the precision metric.", "From the results, it is observed that with Focal Loss, we achieved an improvement of 11.18% in Precision, 1.06% in Recall, and 0.61% on our F1 score compared to the weighted loss."], "text_after_citation": ["Compared to the next best model #OTHEREFR , our model with Focal Loss shows an improvement of 22.91% in Precision, 23.53% in Recall, and 18.04% on F1 score.", "Thus, Focal Loss improves precision, recall and F1 on imbalanced data while not causing a negating impact on a balanced dataset only by adjusting the parameters \u03b1 and \u03b3 RQ3: Is our classifier generalized enough to detect vulnerabilities in N-day and zero-day program sam- We evaluated our classifier's performance based on its ability to accurately predict vulnerability with 273 Nday real-world sample programs. These sample programs are never used during training.", "We also used 4 zero-day examples in order to evaluate our classifier on predicting zero-day vulnerabilities as well.", "The classifier predicts the vulnerability classif the vulnerability exists in the code and predicts non-vulnerable when the vulnerability does not exist.", "Out of these 273 N-day and 4, zero-day code samples, some vulnerability classes exist that are not part of our VulF dataset from table 1."], "citing_paper_content": {"title": "An Unbiased Transformer Source Code Learning With Semantic Vulnerability Graph", "abstract": "Over the years, open-source software systems have become prey to threat actors. Even highly-adopted software has been crippled by unforeseeable attacks, leaving millions of devices exposed. Even as open-source communities act quickly to patch the breach, code vulnerability screening should be an integral part of agile software development from the beginning. Unfortunately, current vulnerability screening techniques are ineffective at identifying novel vulnerabilities or providing developers with code vulnerability and classification. Furthermore, the datasets used for vulnerability learning often exhibit distribution shifts from the real-world testing distribution due to novel attack strategies deployed by adversaries and as a result, the machine learning model's performance may be hindered or biased. To address these issues, we propose a joint interpolated multitasked unbiased vulnerability classifier comprising a transformer \"RoBERTa\" and graph convolution neural network (GCN). We present a training process utilizing a semantic vulnerability graph (SVG) representation from source code, created by integrating edges from a sequential flow, control flow, and data flow, as well as a novel flow dubbed Poacher Flow (PF). Poacher flow edges reduce the gap between dynamic and static program analysis and handle complex long-range dependencies. Moreover, our approach reduces biases of classifiers regarding unbalanced datasets by integrating Focal Loss objective function along with SVG. Remarkably, experimental results show that our classifier outperforms state-of-the-art results on vulnerability detection with fewer false negatives and false positives. After testing our model across multiple datasets, it shows an improvement of at least 2.41% and 18.75% in the bestcase scenario. Evaluations using N-day program samples demonstrate that our proposed approach achieves a 93% accuracy and was able to detect 4, zero-day vulnerabilities from popular GitHub repositories. Our code and data are"}, "cited_paper_content": {"title": "Automated Vulnerability Detection In Source Code Using Deep Representation Learning", "abstract": "Increasing numbers of software vulnerabilities are discovered every year whether they are reported publicly or discovered internally in proprietary code. These vulnerabilities can pose serious risk of exploit and result in system compromise, information leaks, or denial of service. We leveraged the wealth of C and C++ open-source code available to develop a large-scale function-level vulnerability detection system using machine learning. To supplement existing labeled vulnerability datasets, we compiled a vast dataset of millions of open-source functions and labeled it with carefully-selected findings from three different static analyzers that indicate potential exploits. The labeled dataset is available at: this https URL. Using these datasets, we developed a fast and scalable vulnerability detection tool based on deep feature representation learning that directly interprets lexed source code. We evaluated our tool on code from both real software packages and the NIST SATE IV benchmark dataset. Our results demonstrate that deep feature representation learning on source code is a promising approach for automated software vulnerability detection."}, "keywords": ["vulnerability"], "citation_intent": "result"}
{"citing_id": "2303.18190v1", "cited_id": "2004.01670", "section_title": "Qualitative Language Model Risk Assessment", "citation": "How harms present is often highly language-dependent, and so each language needs its own dataset, but the distribution of languages represented in harm detection data is skewed #REFR .", "text_before_citation": ["Further, automated systems project an unknown set of values onto the result. How their creators define e.g.", "\"toxicity\" and represent it through data is often not transparent.", "Thus, not only is it hard to discover when novel forms of harm slip past undetected, it is also uncertain how well their classifications match the goal of an assessment.", "Automated systems are frequently limited to well-resourced languages.", "The efficacy of harm detection classifiers are limited by the amount of language-specific data."], "text_after_citation": ["Automated systems degrade over time.", "Forms of linguistic expression evolve, but a classifier is frozen in time when it is trained (or, specifically, when its training data was gathered).", "For example, some APIs would consistently mark any message containing the term \"toot\" as profane, causing errors first apparent when applied to Mastodon.", "Automating evaluation stops assessors from learning.", "A way to become better at assessing LM risks is to granularly understand their data, and output behaviours."], "citing_paper_content": {"title": "Assessing Language Model Deployment With Risk Cards", "abstract": "This paper introduces RiskCards, a framework for structured assessment and documentation of risks associated with an application of language models. As with all language, text generated by language models can be harmful, or used to bring about harm. Automating language generation adds both an element of scale and also more subtle or emergent undesirable tendencies to the generated text. Prior work establishes a wide variety of language model harms to many different actors: existing taxonomies identify categories of harms posed by language models; benchmarks establish automated tests of these harms; and documentation standards for models, tasks and datasets encourage transparent reporting. However, there is no risk-centric framework for documenting the complexity of a landscape in which some risks are shared across models and contexts, while others are specific, and where certain conditions may be required for risks to manifest as harms. RiskCards address this methodological gap by providing a generic framework for assessing the use of a given language model in a given scenario. Each RiskCard makes clear the routes for the risk to manifest harm, their placement in harm taxonomies, and example prompt-output pairs. While RiskCards are designed to be open-source, dynamic and participatory, we present a \"starter set\" of RiskCards taken from a broad literature survey, each of which details a concrete risk presentation. Language model RiskCards initiate a community knowledge base which permits the mapping of risks and harms to a specific model or its application scenario, ultimately contributing to a better, safer and shared understanding of the risk landscape. CCS Concepts: \u2022 Computing methodologies \u2192 Natural language processing; \u2022 Security and privacy \u2192 Human and societal aspects of security and privacy."}, "cited_paper_content": {"title": "Directions In Abusive Language Training Data: Garbage In, Garbage Out", "abstract": "Data-driven analysis and detection of abusive online content covers many different tasks, phenomena, contexts, and methodologies. This paper systematically reviews abusive language dataset creation and content in conjunction with an open website for cataloguing abusive language data. This collection of knowledge leads to a synthesis providing evidence-based recommendations for practitioners working with this complex and highly diverse data."}, "keywords": ["language", "harm detection data"], "citation_intent": "background"}
{"citing_id": "2304.09915v1", "cited_id": "1908.07919", "section_title": "C. Dual Context Network With Transformer", "citation": "We remove the fifth stage and the last three pooling layers to preserve a high-resolution feature map, which has been proven to be effective in segmentation #REFR . Therefore, H 1 = H/4, W 1 = W/4.", "text_before_citation": ["In this section, we first introduce the architecture of the proposed DCN-T.", "Since the classification results are from the derived tri-spectral image set, the final prediction can be obtained by voting. The details will be presented later.", "1) Architecture: Combining the proposed DCM with an ImageNet pretrained backbone network, we obtain a complete segmentation network named DCN-T, where \"T\" represents the transformer used in DCM.", "After reviewing the literature in the remote sensing community #OTHEREFR , it has been found that VGG-16 #OTHEREFR is one of the most frequently utilized backbones.", "As a result, we adopt the VGG-16 network as the primary backbone for our study."], "text_after_citation": ["We add a 3 \u00d7 3 standard convolution to reduce the dimension to C = 256 in the feature F.", "In the DCM, since we use one transformer encoder to obtain RACs, and then use one transformer encoder and a single-layer transformer decoder to separately encode the relationships between homogeneous areas and aggregate global contexts to form 2-D features. Therefore, N D = N E = 1.", "In addition, the MLP ratio is always set to 2 for all transformer layers.", "Finally, the input and output of the DCM, i.e., F and F GAC , are concatenated together.", "The prediction is obtained after a convolutional layer and a bilinear upsampling layer."], "citing_paper_content": {"title": "Dcn-T: Dual Context Network With Transformer For Hyperspectral Image Classification", "abstract": "Hyperspectral image (HSI) classification is challenging due to spatial variability caused by complex imaging conditions. Prior methods suffer from limited representation ability, as they train specially designed networks from scratch on limited annotated data. We propose a tri-spectral image generation pipeline that transforms HSI into high-quality trispectral images, enabling the use of off-the-shelf ImageNet pretrained backbone networks for feature extraction. Motivated by the observation that there are many homogeneous areas with distinguished semantic and geometric properties in HSIs, which can be used to extract useful contexts, we propose an end-to-end segmentation network named DCN-T. It adopts transformers to effectively encode regional adaptation and global aggregation spatial contexts within and between the homogeneous areas discovered by similarity-based clustering. To fully exploit the rich spectrums of the HSI, we adopt an ensemble approach where all segmentation results of the tri-spectral images are integrated into the final prediction through a voting scheme. Extensive experiments on three public benchmarks show that our proposed method outperforms state-of-the-art methods for HSI classification. The code will be released at https://github.com/DotWang/DCN-T."}, "cited_paper_content": {"title": "Deep High-Resolution Representation Learning For Visual Recognition", "abstract": "High-resolution representations are essential for position-sensitive vision problems, such as human pose estimation, semantic segmentation, and object detection. Existing state-of-the-art frameworks first encode the input image as a low-resolution representation through a subnetwork that is formed by connecting high-to-low resolution convolutions \\emph{in series} (e.g., ResNet, VGGNet), and then recover the high-resolution representation from the encoded low-resolution representation. Instead, our proposed network, named as High-Resolution Network (HRNet), maintains high-resolution representations through the whole process. There are two key characteristics: (i) Connect the high-to-low resolution convolution streams \\emph{in parallel}; (ii) Repeatedly exchange the information across resolutions. The benefit is that the resulting representation is semantically richer and spatially more precise. We show the superiority of the proposed HRNet in a wide range of applications, including human pose estimation, semantic segmentation, and object detection, suggesting that the HRNet is a stronger backbone for computer vision problems. All the codes are available at~{\\url{this https URL}}."}, "keywords": ["segmentation", "high-resolution feature map"], "citation_intent": "method"}
{"citing_id": "2304.03833v1", "cited_id": "1804.02717", "section_title": "Ablate Reference State Initialization In Dmfd", "citation": "An improvement we made over the original DMfD algorithm is to disable matching with expert states, known as RSI-IR, first proposed in #REFR .", "text_before_citation": [], "text_after_citation": ["We justify this improvement in this ablation, labeled ABL5 in Fig. 7 . As shown in Fig.", "10b , removing RSI and IR has a net positive effect throughout training, and around 10% on the final policy performance.", "This means that matching expert states exactly via imitation reward does not help, even during the initial stages of training when the policy is randomly initialized.", "We believe this is because RSI helps when there are hard-to-reach intermediate states that the policy cannot reach during the initial stages of training.", "This is true for dynamic or long-horizon tasks, such as karate chops and roundhouse kicks."], "citing_paper_content": {"title": "Bridging Action Space Mismatch In Learning From Demonstrations", "abstract": "Learning from demonstrations (LfD) methods guide learning agents to a desired solution using demonstrations from a teacher. While some LfD methods can handle small mismatches in the action spaces of the teacher and student, here we address the case where the teacher demonstrates the task in an action space that can be substantially different from that of the student-thereby inducing a large action space mismatch. We bridge this gap with a framework, Morphological Adaptation in Imitation Learning (MAIL), that allows training an agent from demonstrations by other agents with significantly different morphologies (from the student or each other). MAIL is able to learn from suboptimal demonstrations, so long as they provide some guidance towards a desired solution. We demonstrate MAIL on challenging household cloth manipulation tasks and introduce a new DRY CLOTH task-cloth manipulation in 3D task with obstacles. In these tasks, we train a visual control policy for a robot with one end-effector using demonstrations from a simulated agent with two end-effectors. MAIL shows up to 27% improvement over LfD and non-LfD baselines. It is deployed to a real Franka Panda robot, and can handle multiple variations in cloth properties (color, thickness, size, material) and pose (rotation and translation). We further show generalizability to transfers from n-to-m end-effectors, in the context of a simple rearrangement task."}, "cited_paper_content": {"title": "Deepmimic: Example-Guided Deep Reinforcement Learning Of Physics-Based Character Skills", "abstract": "A longstanding goal in character animation is to combine data-driven specification of behavior with a system that can execute a similar behavior in a physical simulation, thus enabling realistic responses to perturbations and environmental variation. We show that well-known reinforcement learning (RL) methods can be adapted to learn robust control policies capable of imitating a broad range of example motion clips, while also learning complex recoveries, adapting to changes in morphology, and accomplishing user-specified goals. Our method handles keyframed motions, highly-dynamic actions such as motion-captured flips and spins, and retargeted motions. By combining a motion-imitation objective with a task objective, we can train characters that react intelligently in interactive settings, e.g., by walking in a desired direction or throwing a ball at a user-specified target. This approach thus combines the convenience and motion quality of using motion clips to define the desired style and appearance, with the flexibility and generality afforded by RL methods and physics-based animation. We further explore a number of methods for integrating multiple clips into the learning process to develop multi-skilled agents capable of performing a rich repertoire of diverse skills. We demonstrate results using multiple characters (human, Atlas robot, bipedal dinosaur, dragon) and a large variety of skills, including locomotion, acrobatics, and martial arts."}, "keywords": ["improvement", "expert states"], "citation_intent": "method"}
{"citing_id": "2303.16447v1", "cited_id": "2001.06659", "section_title": "Mvas Versus Mvps", "citation": "B-MVPS #REFR achieves the best scores in 4 objects due to the usage of calibrated light information.", "text_before_citation": ["We use 15-view azimuth maps for optimization and leave out 5 views for testing, following PS-NeRF #OTHEREFR .", "The azimuth maps are computed from the normal maps estimated by the self-calibrated photometric stereo method SDPS #OTHEREFR .", "Evaluation metrics We use Chamfer distance (CD) and F-score for geometry accuracy #OTHEREFR , and mean angular error (MAE) for normal accuracy #OTHEREFR .", "For CD and F-score, we only consider visible points by casting rays for all pixels and finding the first ray-mesh intersections #OTHEREFR .", "Table 2 reports the geometry accuracy of the recovered DiLiGenT-MV surfaces."], "text_after_citation": ["UA-MVPS #OTHEREFR distorts the surface reconstruction by not considering the multi-view consistency.", "MVAS outperforms PS-NeRF #OTHEREFR in 3 objects without modeling the rendering process. Figure 7 visually compares recovered \"Buddha\" and \"Reading\" objects.", "Despite not having the best numerical scores, our method produces comparable results.", "Lower scores for these objects are mainly due to our method's sensitivity to inaccurate silhouette masks provided by DiLiGenT-MV #OTHEREFR .", "We project the GT surface onto the image plane and find up to 10-pixel inconsistency between the projected region and the GT mask. Thus, the silhouette loss Eq."], "citing_paper_content": {"title": "Multi-View Azimuth Stereo Via Tangent Space Consistency", "abstract": "Figure 1. 3D reconstruction from calibrated multi-view azimuth maps (3 out of 31 are shown). An azimuth angle indicates the surface normal's orientation in the image plane, and an azimuth map records the azimuth angles across the entire surface. We show that azimuth maps can be effectively used for shape and normal recovery. Color images are for reference only and are not used in shape optimization."}, "cited_paper_content": {"title": "Multi-View Photometric Stereo: A Robust Solution And Benchmark Dataset For Spatially Varying Isotropic Materials", "abstract": "We present a method to capture both 3D shape and spatially varying reflectance with a multi-view photometric stereo (MVPS) technique that works for general isotropic materials. Our algorithm is suitable for perspective cameras and nearby point light sources. Our data capture setup is simple, which consists of only a digital camera, some LED lights, and an optional automatic turntable. From a single viewpoint, we use a set of photometric stereo images to identify surface points with the same distance to the camera. We collect this information from multiple viewpoints and combine it with structure-from-motion to obtain a precise reconstruction of the complete 3D shape. The spatially varying isotropic bidirectional reflectance distribution function (BRDF) is captured by simultaneously inferring a set of basis BRDFs and their mixing weights at each surface point. In experiments, we demonstrate our algorithm with two different setups: a studio setup for highest precision and a desktop setup for best usability. According to our experiments, under the studio setting, the captured shapes are accurate to 0.5 millimeters and the captured reflectance has a relative root-mean-square error (RMSE) of 9%. We also quantitatively evaluate state-of-the-art MVPS on a newly collected benchmark dataset, which is publicly available for inspiring future research."}, "keywords": ["calibrated light information"], "citation_intent": "background"}
{"citing_id": "2303.12993v1", "cited_id": "1905.02249", "section_title": "Ablation Study On Defense Settings", "citation": "As shown in Table 4 , UDA and ReMixMatch can still have similar robustness against backdoor attacks compared with MixMatch #REFR under our proposed ASD.", "text_before_citation": ["However, it can wrongly introduce a large number of poisoned samples into D C and result in the failure of our ASD, especially under WaNet and CLB.", "Hence, it is necessary to control the speed to build D C and constrain the number of samples in D C during stage 1. Different semi-supervised learning methods.", "We treat the samples in D P as unlabeled and apply semi-supervised learning to learn from both data pools.", "In this experiment, we show our ASD can work well with various semisupervised learning, e.g., UDA #OTHEREFR and ReMixMatch #OTHEREFR .", "We keep all settings unchanged."], "text_after_citation": ["More details about these three semisupervised learning methods are in Appendix J."], "citing_paper_content": {"title": "Backdoor Defense Via Adaptively Splitting Poisoned Dataset", "abstract": "Backdoor defenses have been studied to alleviate the threat of deep neural networks (DNNs) being backdoor attacked and thus maliciously altered. Since DNNs usually adopt some external training data from an untrusted third party, a robust backdoor defense strategy during the training stage is of importance. We argue that the core of training-time defense is to select poisoned samples and to handle them properly. In this work, we summarize the training-time defenses from a unified framework as splitting the poisoned dataset into two data pools. Under our framework, we propose an adaptively splitting datasetbased defense (ASD). Concretely, we apply loss-guided split and meta-learning-inspired split to dynamically update two data pools. With the split clean data pool and polluted data pool, ASD successfully defends against backdoor attacks during training. Extensive experiments on multiple benchmark datasets and DNN models against six state-ofthe-art backdoor attacks demonstrate the superiority of our ASD. Our code is available at https://github.com/ KuofengGao/ASD."}, "cited_paper_content": {"title": "Mixmatch: A Holistic Approach To Semi-Supervised Learning", "abstract": "Semi-supervised learning has proven to be a powerful paradigm for leveraging unlabeled data to mitigate the reliance on large labeled datasets. In this work, we unify the current dominant approaches for semi-supervised learning to produce a new algorithm, MixMatch, that works by guessing low-entropy labels for data-augmented unlabeled examples and mixing labeled and unlabeled data using MixUp. We show that MixMatch obtains state-of-the-art results by a large margin across many datasets and labeled data amounts. For example, on CIFAR-10 with 250 labels, we reduce error rate by a factor of 4 (from 38% to 11%) and by a factor of 2 on STL-10. We also demonstrate how MixMatch can help achieve a dramatically better accuracy-privacy trade-off for differential privacy. Finally, we perform an ablation study to tease apart which components of MixMatch are most important for its success."}, "keywords": ["backdoor attacks", "MixMatch"], "citation_intent": "result"}
{"citing_id": "2303.03951v1", "cited_id": "1610.01644", "section_title": "Probing Strategies", "citation": "The advocates of linear probing #REFR argue that the probe model should be simple, e.g.", "text_before_citation": ["The goal of probing is to evaluate the \"extractability\" or \"readability\" of a property from the representation.", "The standard approach is to train a separate model, called a probe, to predict the property p i given the fixed representations z i .", "Specifically, the probing dataset D is split into a train and test set, the probe is trained on the train set, and its performance is evaluated on the test set.", "Good test performance is taken as evidence that the representation contains information about the property.", "Low performance indicates that the property is either not present in the representations or not usable. The idea of usability is prominent in the literature."], "text_after_citation": ["a logistic regression (or linear regression for continuous properties), since this means that the information can be easily extracted and used in subsequent processing (e.g. in subsequent layers). The advocates for more complex probes #OTHEREFR", "(2020b) argue they are better since the information about the property may be non-linearly encoded in the representation.", "Yet, the good performance of non-linear probes may come from overfitting (memorization of spurious correlations).", "To overcome these limitations, various control tasks have been proposed: comparing the performance to a majority baseline, random representations, randomization of the properties, or the use of minimum description length.", "Despite its limitations, in this paper we use linear probing since the results are more interpretable."], "citing_paper_content": {"title": "Probing Graph Representations", "abstract": "Today we have a good theoretical understanding of the representational power of Graph Neural Networks (GNNs). For example, their limitations have been characterized in relation to a hierarchy of Weisfeiler-Lehman (WL) isomorphism tests. However, we do not know what is encoded in the learned representations. This is our main question. We answer it using a probing framework to quantify the amount of meaningful information captured in graph representations. Our findings on molecular datasets show the potential of probing for understanding the inductive biases of graph-based models. We compare different families of models and show that transformer-based models capture more chemically relevant information compared to models based on message passing. We also study the effect of different design choices such as skip connections and virtual nodes. We advocate for probing as a useful diagnostic tool for evaluating graph-based models."}, "cited_paper_content": {"title": "Understanding Intermediate Layers Using Linear Classifier Probes", "abstract": "Neural network models have a reputation for being black boxes. We propose a new method to better understand the roles and dynamics of the intermediate layers. This has direct consequences on the design of such models and it enables the expert to be able to justify certain heuristics (such as adding auxiliary losses in middle layers). Our method uses linear classifiers, referred to as ``probes'', where a probe can only use the hidden units of a given intermediate layer as discriminating features. Moreover, these probes cannot affect the training phase of a model, and they are generally added after training. They allow the user to visualize the state of the model at multiple steps of training. We demonstrate how this can be used to develop a better intuition about models and to diagnose potential problems."}, "keywords": ["probe model"], "citation_intent": "background"}
{"citing_id": "2303.04839v1", "cited_id": "1512.00567", "section_title": "Introduction", "citation": "A large dataset with approximately 129,450 images has been utilized to develop the skin cancer classification model with Inception v3 #REFR .", "text_before_citation": ["Computer-aided diagnosis of skin diseases has become more popular since the introduction of Inception v3 #OTHEREFR that achieved a performance accuracy of 93.3% #OTHEREFR in classifying various cancerous skin conditions."], "text_after_citation": ["However, gathering such a large amount of data is not feasible for some skin conditions such as Rosacea.", "Although many skin conditions can lead to fatal consequences, cancer has been considered the most serious of all and has motivated the gathering of more data over time.", "As a result, many Teledermatology #OTHEREFR websites have a substantial amount of skin cancer images.", "On the other hand, there is very limited data for non-fatal chronic skin conditions such as Rosacea. Deep Convolutional Neural Networks (DCNNs) e.g.", "Inception v3, perform relatively well provided with a large training dataset #OTHEREFR ."], "citing_paper_content": {"title": "High Fidelity Synthetic Face Generation For Rosacea Skin Condition From Limited Data", "abstract": "Similar to the majority of deep learning applications, diagnosing skin diseases using computer vision and deep learning often requires a large volume of data. However, obtaining sufficient data for particular types of facial skin conditions can be difficult due to privacy concerns. As a result, conditions like Rosacea are often understudied in computer-aided diagnosis. The limited availability of data for facial skin conditions has led to the investigation of alternative methods for computer-aided diagnosis. In recent years, Generative Adversarial Networks (GANs), mainly variants of StyleGANs, have demonstrated promising results in generating synthetic facial images. In this study, for the first time, a small dataset of Rosacea with 300 full-face images is utilized to further investigate the possibility of generating synthetic data. The preliminary experiments show how fine-tuning the model and varying experimental settings significantly affect the fidelity of the Rosacea features. It is demonstrated that R 1 Regularization strength helps achieve high-fidelity details. Additionally, this study presents qualitative evaluations of synthetic/generated faces by expert dermatologists and non-specialist participants. The quantitative evaluation is presented using a few validation metric(s). Furthermore a number of limitations and future directions are discussed. Code and generated dataset are"}, "cited_paper_content": {"title": "Rethinking The Inception Architecture For Computer Vision", "abstract": "Convolutional networks are at the core of most state of-the-art computer vision solutions for a wide variety of tasks. Since 2014 very deep convolutional networks started to become mainstream, yielding substantial gains in various benchmarks. Although increased model size and computational cost tend to translate to immediate quality gains for most tasks (as long as enough labeled data is provided for training), computational efficiency and low parameter count are still enabling factors for various use cases such as mobile vision and big-data scenarios. Here we are exploring ways to scale up networks in ways that aim at utilizing the added computation as efficiently as possible by suitably factorized convolutions and aggressive regularization. We benchmark our methods on the ILSVRC 2012 classification challenge validation set demonstrate substantial gains over the state of the art: 21:2% top-1 and 5:6% top-5 error for single frame evaluation using a network with a computational cost of 5 billion multiply-adds per inference and with using less than 25 million parameters. With an ensemble of 4 models and multi-crop evaluation, we report 3:5% top-5 error and 17:3% top-1 error on the validation set and 3:6% top-5 error on the official test set."}, "keywords": ["skin cancer classification", "Inception v3"], "citation_intent": "method"}
{"citing_id": "2304.07911v1", "cited_id": "1904.08030", "section_title": "Study Of M2Gnn (Rq2)", "citation": "Secondly, the increase of interest capsules indeed improves the recall list since local businesses contain multiple aspects, while there is a saturation limit which has a similar conclusion in #REFR . Thirdly, the optimal number of is 6~8.", "text_before_citation": ["The lower part of Table 3 shows that M2GNNw/oL 1 sharply underperforms M2GNN without regularizing tag embeddings via L 1 , which proves the necessity of the skip-gram regularization. Impact of hyperparameters.", "We investigate the influence of three hyperparameters: layer numbers, capsule numbers, and exponent before softmax, all of which greatly affect multi-interest extraction and transfer.", ", max and are searched in {1, 2, 3}, {2, 4, 6, 8} and {2, 4, 6, 8}, respectively.", "Figure 6 shows the results in the DPBJ dataset and it has a similar conclusion in the Amazon dataset, which is omitted. From the observations above, we make several conclusions.", "Firstly, = 2 results in the best performance, which means 2-order connections integrate more similar tags into users and stacking more layers continually brings in noisy tags, which is harmful to the recommendation."], "text_after_citation": ["The reason is that a proper exponent helps the softmax function to identify more important interests and assign bigger scores."], "citing_paper_content": {"title": "M2Gnn: Metapath And Multi-Interest Aggregated Graph Neural Network For Tag-Based Cross-Domain Recommendation", "abstract": "Cross-domain recommendation (CDR) is an effective way to alleviate the data sparsity problem. Content-based CDR is one of the most promising branches since most kinds of products can be described by a piece of text, especially when cold-start users or items have few interactions. However, two vital issues are still under-explored: (1) From the content modeling perspective, sufficient long-text descriptions are usually scarce in a real recommender system, more often the lightweight textual features, such as a few keywords or tags, are more accessible, which is improperly modeled by existing methods. (2) From the CDR perspective, not all inter-domain interests are helpful to infer intra-domain interests. Caused by domain-specific features, there are part of signals benefiting for recommendation in the source domain but harmful for that in the target domain. Therefore, how to distill useful interests is crucial. To tackle the above two problems, we propose a metapath and multi-interest aggregated graph neural network (M2GNN). Specifically, to model the tag-based contents, we construct a heterogeneous information network to hold the semantic relatedness between users, items, and tags in all domains. The metapath schema is predefined according to domain-specific knowledge, with one metapath for one domain. User representations are learned by GNN with a hierarchical aggregation framework, where the intra-metapath aggregation firstly filters out trivial tags and the inter-metapath aggregation further filters out useless interests. Offline experiments and online A/B tests demonstrate that M2GNN achieves significant improvements over the state-of-the-art methods and current industrial recommender system in Dianping, respectively. Further analysis shows that M2GNN offers an interpretable recommendation."}, "cited_paper_content": {"title": "Multi-Interest Network With Dynamic Routing For Recommendation At Tmall", "abstract": "Industrial recommender systems usually consist of the matching stage and the ranking stage, in order to handle the billion-scale of users and items. The matching stage retrieves candidate items relevant to user interests, while the ranking stage sorts candidate items by user interests. Thus, the most critical ability is to model and represent user interests for either stage. Most of the existing deep learning-based models represent one user as a single vector which is insufficient to capture the varying nature of user's interests. In this paper, we approach this problem from a different view, to represent one user with multiple vectors encoding the different aspects of the user's interests. We propose the Multi-Interest Network with Dynamic routing (MIND) for dealing with user's diverse interests in the matching stage. Specifically, we design a multi-interest extractor layer based on capsule routing mechanism, which is applicable for clustering historical behaviors and extracting diverse interests. Furthermore, we develop a technique named label-aware attention to help learn a user representation with multiple vectors. Through extensive experiments on several public benchmarks and one large-scale industrial dataset from Tmall, we demonstrate that MIND can achieve superior performance than state-of-the-art methods for recommendation. Currently, MIND has been deployed for handling major online traffic at the homepage on Mobile Tmall App."}, "keywords": ["interest capsules"], "citation_intent": "result"}
{"citing_id": "2304.11697v1", "cited_id": "1711.08488", "section_title": "B. Multi-Modal Fusion For Object Detection 1) Multi-Modal Object Detection:", "citation": "Frustum PointNets #REFR extract the 3D bounding frustum of an object by extruding 2D bounding boxes from image detectors.", "text_before_citation": ["To date, several studies have investigated multi-modal fusion for 2D and 3D object detection."], "text_after_citation": ["PointFusion #OTHEREFR combines a CNN and a PointNet #OTHEREFR architecture respectively to process images and raw point clouds then predict 3D boxes.", "PointPainting #OTHEREFR projects LiDAR points into the output of an image-only semantic segmentation network, and appends the class scores to each point.", "All these fusion methods of RGB and LiDAR achieve high average precision on the benchmarks, however, the coupling or interrelation of two modalities will cause the whole system to fail easily once part of the sensors break down.", "Besides, the methods above only provide a deterministic predict result, making it risky to carry out in the real application.", "2) Adaptive fusion: Several new studies have proposed self-adaptive techniques in computer vision."], "citing_paper_content": {"title": "Informative Data Selection With Uncertainty For Multi-Modal Object Detection", "abstract": "Noise has always been nonnegligible trouble in object detection by creating confusion in model reasoning, thereby reducing the informativeness of the data. It can lead to inaccurate recognition due to the shift in the observed pattern, that requires a robust generalization of the models. To implement a general vision model, we need to develop deep learning models that can adaptively select valid information from multi-modal data. This is mainly based on two reasons. Multi-modal learning can break through the inherent defects of single-modal data, and adaptive information selection can reduce chaos in multi-modal data. To tackle this problem, we propose a universal uncertainty-aware multi-modal fusion model. It adopts a multi-pipeline loosely coupled architecture to combine the features and results from point clouds and images. To quantify the correlation in multimodal information, we model the uncertainty, as the inverse of data information, in different modalities and embed it in the bounding box generation. In this way, our model reduces the randomness in fusion and generates reliable output. Moreover, we conducted a completed investigation on the KITTI 2D object detection dataset and its derived dirty data. Our fusion model is proven to resist severe noise interference like Gaussian, motion blur, and frost, with only slight degradation. The experiment results demonstrate the benefits of our adaptive fusion. Our analysis on the robustness of multi-modal fusion will provide further insights for future research."}, "cited_paper_content": {"title": "Frustum Pointnets For 3D Object Detection From Rgb-D Data", "abstract": "In this work, we study 3D object detection from RGB-D data in both indoor and outdoor scenes. While previous methods focus on images or 3D voxels, often obscuring natural 3D patterns and invariances of 3D data, we directly operate on raw point clouds by popping up RGB-D scans. However, a key challenge of this approach is how to efficiently localize objects in point clouds of large-scale scenes (region proposal). Instead of solely relying on 3D proposals, our method leverages both mature 2D object detectors and advanced 3D deep learning for object localization, achieving efficiency as well as high recall for even small objects. Benefited from learning directly in raw point clouds, our method is also able to precisely estimate 3D bounding boxes even under strong occlusion or with very sparse points. Evaluated on KITTI and SUN RGB-D 3D detection benchmarks, our method outperforms the state of the art by remarkable margins while having real-time capability."}, "keywords": ["image detectors", "Frustum PointNets"], "citation_intent": "background"}
{"citing_id": "2304.05147v1", "cited_id": "1911.01547", "section_title": "Intelligence", "citation": "In general, there are two different interpretations: intelligence as either a collection of task-specific skills or a general learning ability #REFR ], which reflect the distinction between crystallised and fluid abilities, respectively.", "text_before_citation": ["Intelligence is a controversial and elusive concept subject to philosophical debate #OTHEREFR", "2007] , best understood as a nomological network of constructs #OTHEREFR .", "Etymologically, intelligence comes from Latin \"intelligere\", which means \"to understand\".", "It can be defined as \"the global capacity of the individual to act purposefully, to think rationally, and to deal effectively with the environment\" #OTHEREFR ], or the property that \"measures an agent's ability to achieve goals in a wide range of environments\" #OTHEREFR ."], "text_after_citation": ["Problems about intelligence include, for instance, its definition and modelling, such as devising the structure of intelligence #OTHEREFR", "2011] ; its relation with action; its measurement and evaluation; its analysis; and its construction and development.", "Concerning the theories of intelligence, there are two main traditions #OTHEREFR", "2011] : the psychometric tradition, based on the number and nature of basic cognitive abilities or factors; and the developmental or holistic perspective, based on acquired intellect.", "The problem of the measure of intelligence #OTHEREFR is of course related to what representation or model of intelligence is considered, and is complicated by the need of distinguishing between causality and correlation, selecting a representative set of environments for evaluation, etc. Carrol defines an ability (i.e."], "citing_paper_content": {"title": "Artificial Collective Intelligence Engineering: A Survey Of Concepts And Perspectives", "abstract": "Collectiveness is an important property of many systems-both natural and artificial. By exploiting a large number of individuals, it is often possible to produce effects that go far beyond the capabilities of the smartest individuals, or even to produce intelligent collective behaviour out of not-so-intelligent individuals. Indeed, collective intelligence, namely the capability of a group to act collectively in a seemingly intelligent way, is increasingly often a design goal of engineered computational systems-motivated by recent techno-scientific trends like the Internet of Things, swarm robotics, and crowd computing, just to name a few. For several years, the collective intelligence observed in natural and artificial systems has served as a source of inspiration for engineering ideas, models, and mechanisms. Today, artificial and computational collective intelligence are recognised research topics, spanning various techniques, kinds of target systems, and application domains. However, there is still a lot of fragmentation in the research panorama of the topic within computer science, and the verticality of most communities and contributions makes it difficult to extract the core underlying ideas and frames of reference. The challenge is to identify, place in a common structure, and ultimately connect the different areas and methods addressing intelligent collectives. To address this gap, this paper considers a set of broad scoping questions providing a map of collective intelligence research, mostly by the point of view of computer scientists and engineers. Accordingly, it covers preliminary notions, fundamental concepts, and the main research perspectives, identifying opportunities and challenges for researchers on artificial and computational collective intelligence engineering."}, "cited_paper_content": {"title": "On The Measure Of Intelligence", "abstract": "To make deliberate progress towards more intelligent and more human-like artificial systems, we need to be following an appropriate feedback signal: we need to be able to define and evaluate intelligence in a way that enables comparisons between two systems, as well as comparisons with humans. Over the past hundred years, there has been an abundance of attempts to define and measure intelligence, across both the fields of psychology and AI. We summarize and critically assess these definitions and evaluation approaches, while making apparent the two historical conceptions of intelligence that have implicitly guided them. We note that in practice, the contemporary AI community still gravitates towards benchmarking intelligence by comparing the skill exhibited by AIs and humans at specific tasks such as board games and video games. We argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to \"buy\" arbitrary levels of skills for a system, in a way that masks the system's own generalization power. We then articulate a new formal definition of intelligence based on Algorithmic Information Theory, describing intelligence as skill-acquisition efficiency and highlighting the concepts of scope, generalization difficulty, priors, and experience. Using this definition, we propose a set of guidelines for what a general AI benchmark should look like. Finally, we present a benchmark closely following these guidelines, the Abstraction and Reasoning Corpus (ARC), built upon an explicit set of priors designed to be as close as possible to innate human priors. We argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans."}, "keywords": ["intelligence"], "citation_intent": "background"}
{"citing_id": "2304.08868v1", "cited_id": "1802.04741", "section_title": "I. Introduction", "citation": "We note that the syndrome does not depend on the codeword and, therefore, we do not require the NN to have a special structure, it can be arbitrary, but the best results were obtained with recurrent NNs #REFR .", "text_before_citation": ["The next idea was to replace the activation functions, the architecture is called a hyper-network #OTHEREFR , #OTHEREFR . Later, Cammerer et al.", "proposed to replace node and edge message updates with trainable functions, thus allowing NN to learn a generalized message passing algorithm #OTHEREFR .", "Another approach proposed in #OTHEREFR is to consider the syndrome-based decoding algorithm that is suitable for any linear codes.", "The basic syndrome-based decoding algorithm implies the use of the mapping (syndrome to the coset leader), which has the exponential (in the number of parity-check bits) size.", "The idea of #OTHEREFR is to approximate this table with a NN."], "text_after_citation": ["Later a syndrome-based approach was adapted to the transformer and denoising diffusion architectures #OTHEREFR , #OTHEREFR .", "We also note the papers (see, e.g. #OTHEREFR ) devoted to DNN-based code construction.", "For additional literature and a more detailed overview, we refer the reader to #OTHEREFR .", "The papers above focus on the performance of hard-output decoding, i.e.", "the decoder is required to return the estimated information word."], "citing_paper_content": {"title": "Soft-Output Deep Neural Network-Based Decoding", "abstract": "Deep neural network (DNN)-based channel decoding is widely considered in the literature. The existing solutions are investigated for the case of hard output, i.e. when the decoder returns the estimated information word. At the same time, soft-output decoding is of critical importance for iterative receivers and decoders. In this paper, we focus on the soft-output DNN-based decoding problem. We start with the syndrome-based approach proposed by Bennatan et al. (2018) and modify it to provide soft output in the AWGN channel. The new decoder can be considered as an approximation of the MAP decoder with smaller computation complexity. We discuss various regularization functions for joint DNN-MAP training and compare the resulting distributions for [64, 45] BCH code. Finally, to demonstrate the soft-output quality we consider the turboproduct code with [64, 45] BCH codes as row and column codes. We show that the resulting DNN-based scheme is very close to the MAP-based performance and significantly outperforms the solution based on the Chase decoder. We come to the conclusion that the new method is prospective for the challenging problem of DNN-based decoding of long codes consisting of short component codes."}, "cited_paper_content": {"title": "Deep Learning For Decoding Of Linear Codes - A Syndrome-Based Approach", "abstract": "We present a novel framework for applying deep neural networks (DNN) to soft decoding of linear codes at arbitrary block lengths. Unlike other approaches, our framework allows unconstrained DNN design, enabling the free application of powerful designs that were developed in other contexts. Our method is robust to overfitting that inhibits many competing methods, which follows from the exponentially large number of codewords required for their training. We achieve this by transforming the channel output before feeding it to the network, extracting only the syndrome of the hard decisions and the channel output reliabilities. We prove analytically that this approach does not involve any intrinsic performance penalty, and guarantees the generalization of performance obtained during training. Our best results are obtained using a recurrent neural network (RNN) architecture combined with simple preprocessing by permutation. We provide simulation results that demonstrate performance that sometimes approaches that of the ordered statistics decoding (OSD) algorithm."}, "keywords": ["recurrent NNs"], "citation_intent": "result"}
{"citing_id": "2305.01310v1", "cited_id": "1705.10253", "section_title": "Continuization Results", "citation": "If c 1 \u2265 2, then d c1 \u2264 1 2 , i.e., for size 1, the solution has value d c1 \u2264 #REFR 2 which implies that the solution has a competitive ratio of at least 2. Thus assume that c 1 = 1.", "text_before_citation": ["There are even discrete instances where every continuization of the instance has a competitive ratio smaller than the initial instance.", "For i \u2208 {2, 5, 6, 7, 8, 9, 10, 11}, the density is chosen such that i", "\u2022 d i = v i = v i\u22121 = (i \u2212 1)d i\u22121 .", "We show that every incremental solution (c 1 , c 2 , . . .", ") has a competitive ratio of at least 1.446 for this problem instance."], "text_after_citation": ["If c 2 \u2265 5, we can, without loss of generality, assume that c 2 \u2265 12.", "Otherwise we can improve the solution by choosing c 2 = 4 instead.", "Then, the value of the solution for size 4 is max{1, 3", "\u2022 d c2 } = max{1, 3 \u2022 16473", "107200 } = 1, while the optimal solution has value 4d 4 = 17 10 , i.e., the competitive ratio of the solution is at least 1.7."], "citing_paper_content": {"title": "Incremental Maximization Via Continuization", "abstract": "We consider the problem of finding an incremental solution to a cardinality-constrained maximization problem that not only captures the solution for a fixed cardinality, but also describes how to gradually grow the solution as the cardinality bound increases. The goal is to find an incremental solution that guarantees a good competitive ratio against the optimum solution for all cardinalities simultaneously. The central challenge is to characterize maximization problems where this is possible, and to determine the best-possible competitive ratio that can be attained. A lower bound of 2.18 and an upper bound of \u03d5+1 \u2248 2.618 are known on the competitive ratio for monotone and accountable objectives [Bernstein et al., Math. Prog., 2022], which capture a wide range of maximization problems. We introduce a continuization technique and identify an optimal incremental algorithm that provides strong evidence that \u03d5 + 1 is the best-possible competitive ratio. Using this continuization, we obtain an improved lower bound of 2.246 by studying a particular recurrence relation whose characteristic polynomial has complex roots exactly beyond the lower bound. Based on the optimal continuous algorithm combined with a scaling approach, we also provide a 1.772-competitive randomized algorithm. We complement this by a randomized lower bound of 1.447 via Yao's principle."}, "cited_paper_content": {"title": "General Bounds For Incremental Maximization", "abstract": "We propose a theoretical framework to capture incremental solutions to cardinality constrained maximization problems. The defining characteristic of our framework is that the cardinality/support of the solution is bounded by a value $k\\in\\mathbb{N}$ that grows over time, and we allow the solution to be extended one element at a time. We investigate the best-possible competitive ratio of such an incremental solution, i.e., the worst ratio over all $k$ between the incremental solution after $k$ steps and an optimum solution of cardinality $k$. We define a large class of problems that contains many important cardinality constrained maximization problems like maximum matching, knapsack, and packing/covering problems. We provide a general $2.618$-competitive incremental algorithm for this class of problems, and show that no algorithm can have competitive ratio below $2.18$ in general. In the second part of the paper, we focus on the inherently incremental greedy algorithm that increases the objective value as much as possible in each step. This algorithm is known to be $1.58$-competitive for submodular objective functions, but it has unbounded competitive ratio for the class of incremental problems mentioned above. We define a relaxed submodularity condition for the objective function, capturing problems like maximum (weighted) ($b$-)matching and a variant of the maximum flow problem. We show that the greedy algorithm has competitive ratio (exactly) $2.313$ for the class of problems that satisfy this relaxed submodularity condition. Note that our upper bounds on the competitive ratios translate to approximation ratios for the underlying cardinality constrained problems."}, "keywords": ["competitive ratio"], "citation_intent": "background"}
{"citing_id": "2303.04603v1", "cited_id": "1505.04597", "section_title": "Evaluation Metrics", "citation": "We first train a U-Net #REFR on GAMMA and then evaluate on the testing set of the degraded REFUGE dataset. VSD.", "text_before_citation": ["3 , higher values of PSNR or SSIM do not necessarily indicate better preservation of structural details, as already pointed out elsewhere #OTHEREFR .", "To assess the fundus image enhancement performance with respect to structural details and clinically-relevant applications, we propose the following evaluation metrics: FIQA.", "Similar to #OTHEREFR , we train a ResNet101 #OTHEREFR on EyeQ for quality assessment and use the predictions to calculate a fundus image quality assessment score (FIQA).", "We evaluate on the original testing set of EyeQ labeled as \"Usable\". OCSD.", "We use the OC segmentation Dice (OCSD) as a metric of assessing the capability of preserving anatomical structures."], "text_after_citation": ["We use the vessel segmentation Dice (VSD) as another metric of assessing the capability of preserving anatomical structures.", "Iter-Net #OTHEREFR is used to segment vessels on the testing set of the degraded DRIVE dataset. FCNR.", "Inspired by the contrast noise ratio #OTHEREFR , we propose a fundus contrast noise ratio (FCNR) to evaluate the contrast quality of the vessel areas.", "We first obtain a region of interest (ROI) R using a disk-shaped dilation operation with a radius of 3 pixels based on the vessel segmentation mask.", "We formulate the vessel area as V while the background area as B. The FCNR is defined as"], "citing_paper_content": {"title": "Learning Enhancement From Degradation: A Diffusion Model For Fundus Image Enhancement", "abstract": "The quality of a fundus image can be compromised by numerous factors, many of which are challenging to be appropriately and mathematically modeled. In this paper, we introduce a novel diffusion model based framework, named Learning Enhancement from Degradation (LED), for enhancing fundus images. Specifically, we first adopt a data-driven degradation framework to learn degradation mappings from unpaired high-quality to low-quality images. We then apply a conditional diffusion model to learn the inverse enhancement process in a paired manner. The proposed LED is able to output enhancement results that maintain clinically important features with better clarity. Moreover, in the inference phase, LED can be easily and effectively integrated with any existing fundus image enhancement framework. We evaluate the proposed LED on several downstream tasks with respect to various clinically-relevant metrics, successfully demonstrating its superiority over existing state-of-the-art methods both quantitatively and qualitatively."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["degraded REFUGE dataset", "U-Net"], "citation_intent": "method"}
{"citing_id": "2304.06708v1", "cited_id": "1906.03327", "section_title": "Implementation Details", "citation": "We find that the baseline already performs competitively on our benchmarks, despite the relatively small size of SMiT compared to other datasets such as HowTo100M #REFR , due to the quality and diversity of the manually annotated captions.", "text_before_citation": ["Spoken Moments in Time (SMiT) pretraining dataset.", "The SMiT #OTHEREFR training set consists of 481K pairs of 3 seconds video clips with corresponding captions.", "It is a subset of Moments in Time (MiT) #OTHEREFR .", "Our work falls under the umbrella of transfer learning: we pretrain on SMiT and then use the resulting features to solve different downstream tasks in a zero-shot or fine-tuned manner.", "Pretraining is either done as in regular contrastive learning ('baseline') or with our VFC framework."], "text_after_citation": ["We encourage the community to consider SMiT as a powerful pretraining dataset. PaLM.", "We use PaLM-540B #OTHEREFR with beam size 4, output sequence length 512, and temperature of 0.7.", "The negative captions are generated in an autogressive way and are therefore of arbitrary length.", "We post-process them by removing text after any newline character and by filtering out candidates which contain the same verbs as the original caption.", "Training details. Most hyper-parameters follow CLIP4CLIP #OTHEREFR ."], "citing_paper_content": {"title": "Verbs In Action: Improving Verb Understanding In Video-Language Models", "abstract": "two brown horses eating grass LLM Original caption Video Add prompt two brown horses running on the grass two brown horses fighting on the grass two brown horses lying on the grass eating grass Verb phrase Hard verb negative captions two brown horses sleeping on the grass two brown horses playing on the grass eating grass cleaning camera standing squatting Verb phrase loss two brown horses eating grass two brown horses lying on the grass a woman standing in a post office person squatting at the gym a man cleaning his camera Hard negative loss Generated hard negative caption Negative captions in the batch Negative verb phrases in the batch Batch Figure 1. Verb-Focused Contrastive (VFC) learning: (Left): Given a video and its corresponding caption, we leverage a Large Language Model (LLM) to output (1) hard negative captions, where only the verb has been changed while keeping the remaining context, and (2) verb phrases which succinctly describe the action in the video. (Right): To encourage better verb reasoning, we subsequently enforce (1) a calibrated hard negative loss, using our generated hard negative captions and other captions in the batch, and (2) a fine-grained, verb phrase loss. We show that VFC improves verb understanding of video-language models compared to the standard contrastive loss."}, "cited_paper_content": {"title": "Howto100M: Learning A Text-Video Embedding By Watching Hundred Million Narrated Video Clips", "abstract": "Learning text-video embeddings usually requires a dataset of video clips with manually provided captions. However, such datasets are expensive and time consuming to create and therefore difficult to obtain on a large scale. In this work, we propose instead to learn such embeddings from video data with readily available natural language annotations in the form of automatically transcribed narrations. The contributions of this work are three-fold. First, we introduce HowTo100M: a large-scale dataset of 136 million video clips sourced from 1.22M narrated instructional web videos depicting humans performing and describing over 23k different visual tasks. Our data collection procedure is fast, scalable and does not require any additional manual annotation. Second, we demonstrate that a text-video embedding trained on this data leads to state-of-the-art results for text-to-video retrieval and action localization on instructional video datasets such as YouCook2 or CrossTask. Finally, we show that this embedding transfers well to other domains: fine-tuning on generic Youtube videos (MSR-VTT dataset) and movies (LSMDC dataset) outperforms models trained on these datasets alone. Our dataset, code and models will be publicly available at: www.di.ens.fr/willow/research/howto100m/."}, "keywords": ["manually annotated captions"], "citation_intent": "result"}
{"citing_id": "2303.01276v3", "cited_id": "1604.01685", "section_title": "Datasets", "citation": "Cityscapes dataset #REFR is another benchmark dataset for SSS, which focuses on urban scenarios and it consists of 2,975 annotated training images, 500 validation images and 1,525 testing images from 19 classes.", "text_before_citation": ["Pascal VOC 2012 dataset #OTHEREFR is a standard semisupervised semantic segmentation (SSS) benchmark dataset, which consists of over 13,000 images from 21 classes.", "It contains 1,464 fully annotated images for training, 1,449 images for validation and 1,456 images for testing.", "Previous works use SBD #OTHEREFR to render the labelled images and extend the number of labelled data to 10,582.", "The rendered labelled images are of low quality and some of them are accompanied by noise.", "Therefore, most of the previous works validate their SSS methods with sampled labelled images from the high-quality training images and rendered training images, respectively."], "text_after_citation": [], "citing_paper_content": {"title": "Conflict-Based Cross-View Consistency For Semi-Supervised Semantic Segmentation", "abstract": "Semi-supervised semantic segmentation (SSS) has recently gained increasing research interest as it can reduce the requirement for large-scale fully-annotated training data. The current methods often suffer from the confirmation bias from the pseudo-labelling process, which can be alleviated by the co-training framework. The current co-training-based SSS methods rely on hand-crafted perturbations to prevent the different sub-nets from collapsing into each other, but these artificial perturbations cannot lead to the optimal solution. In this work, we propose a new conflict-based cross-view consistency (CCVC) method based on a two-branch co-training framework which aims at enforcing the two sub-nets to learn informative features from irrelevant views. In particular, we first propose a new cross-view consistency (CVC) strategy that encourages the two sub-nets to learn distinct features from the same input by introducing a feature discrepancy loss, while these distinct features are expected to generate consistent prediction scores of the input. The CVC strategy helps to prevent the two sub-nets from stepping into the collapse. In addition, we further propose a conflict-based pseudo-labelling (CPL) method to guarantee the model will learn more useful information from conflicting predictions, which will lead to a stable training process. We validate our new CCVC approach on the SSS benchmark datasets where our method achieves new state-of-the-art performance. Our code is available at https://github.com/xiaoyao3302/CCVC."}, "cited_paper_content": {"title": "The Cityscapes Dataset For Semantic Urban Scene Understanding", "abstract": "Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations, 20 000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark."}, "keywords": ["2,975 annotated training", "Cityscapes"], "citation_intent": "background"}
{"citing_id": "2304.07358v1", "cited_id": "1905.08750", "section_title": "Performance", "citation": "It coincides with the performance of centralized benchmark (44) and the asymptotic performance of #REFR .", "text_before_citation": ["Proof.", "The proof revolves around introducing a long-term model, analogous to #OTHEREFR , which can be shown to be accurate for small step-sizes and under conditions 3-4. Details are omitted due to space limitations. We verify the accuracy of numerically in Section 5.", "We note that for moderately large step-sizes \u00b5, expression (59) will yield more accurate estimates of the steadystate performance than (62). Relation (62) on the other hand is more tractable.", "We can interpret U T H U in (62) as the projection of the Hessian onto the space spanned by U, while U T R s U is the projection of the noise covariance onto the same space.", "In this sense, Tr U T H U \u22121 U T R s U is a measure of the inverse signal-to-noise ratio after restricting the signal to the space of feasible solutions in #OTHEREFR ."], "text_after_citation": ["We will verify in Section 5 that (59) is accurate for finite step-sizes \u00b5, and that the proposed bias-corrected algorithm outperforms the approximate solution of #OTHEREFR ."], "citing_paper_content": {"title": "Exact Subspace Diffusion For Decentralized Multitask Learning", "abstract": "Classical paradigms for distributed learning, such as federated or decentralized gradient descent, employ consensus mechanisms to enforce homogeneity among agents. While these strategies have proven effective in i.i.d. scenarios, they can result in significant performance degradation when agents follow heterogeneous objectives or data. Distributed strategies for multitask learning, on the other hand, induce relationships between agents in a more nuanced manner, and encourage collaboration without enforcing consensus. We develop a generalization of the exact diffusion algorithm for subspace constrained multitask learning over networks, and derive an accurate expression for its mean-squared deviation when utilizing noisy gradient approximations. We verify numerically the accuracy of the predicted performance expressions, as well as the improved performance of the proposed approach over alternatives based on approximate projections."}, "cited_paper_content": {"title": "Adaptation And Learning Over Networks Under Subspace Constraints\u2014Part I: Stability Analysis", "abstract": "This paper considers optimization problems over networks where agents have individual objectives to meet, or individual parameter vectors to estimate, subject to subspace constraints that require the objectives across the network to lie in low-dimensional subspaces. This constrained formulation includes consensus optimization as a special case, and allows for more general task relatedness models such as smoothness. While such formulations can be solved via projected gradient descent, the resulting algorithm is not distributed. Starting from the centralized solution, we propose an iterative and distributed implementation of the projection step, which runs in parallel with the stochastic gradient descent update. We establish that, for small step-sizes $\\mu$, the proposed distributed adaptive strategy leads to small estimation errors on the order of $\\mu$. We also examine steady-state performance. The results reveal explicitly the influence of the gradient noise, data characteristics, and subspace constraints, on the network performance. The results also show that in the small step-size regime, the iterates generated by the distributed algorithm achieve the centralized steady-state performance. Finally, we apply the proposed strategy to distributed adaptive beamforming."}, "keywords": ["centralized benchmark"], "citation_intent": "result"}
{"citing_id": "2303.17294v1", "cited_id": "1807.10418", "section_title": "Optimizing Process", "citation": "Besides, we apply the co-activity similarity loss L cas 1 #REFR on the enhanced feature E a and the suppressed coarse T-CAS S coarse to model better definitephase feature representation.", "text_before_citation": ["The loss L supp mil is defined as the cross-entropy loss between the video label and the prediction p(j),", "EQUATION", "where y f (j) contains only foreground activities, i.e., the background class y f (C + 1) = 0 for suppressing the background activity.", "And the final suppressed MIL loss can be formulated as,", "EQUATION"], "text_after_citation": ["Following the work #OTHEREFR , we utilize a L1-norm loss L norm to make foreground weights more polarized, denoted as", "L norm = T i=1 |A ness i |,", "where |\u2022| is a L1-norm function.", "In addition, we introduce a guidance loss L guide to make the distribution of action-ness scores A ness opposite to the background class probability in S, i.e.,", "L guide = T i=1 |1 \u2212 A ness i \u2212 s c+1 |,"], "citing_paper_content": {"title": "Jcdnet: Joint Of Common And Definite Phases Network For Weakly Supervised Temporal Action Localization", "abstract": "Weakly-supervised temporal action localization aims to localize action instances in untrimmed videos with only video-level supervision. We witness that different actions record common phases, e.g., the run-up in the HighJump and LongJump. These different actions are defined as conjoint actions, whose rest parts are definite phases, e.g., leaping over the bar in a HighJump. Compared with the common phases, the definite phases are more easily localized in existing researches. Most of them formulate this task as a Multiple Instance Learning paradigm, in which the common phases are tended to be confused with the background, and affect the localization completeness of the conjoint actions. To tackle this challenge, we propose a Joint of Common and Definite phases Network (JCDNet) by improving feature discriminability of the conjoint actions. Specifically, we design a Class-Aware Discriminative module to enhance the contribution of the common phases in classification by the guidance of the coarse definite-phase features. Besides, we introduce a temporal attention module to learn robust action-ness scores via modeling temporal dependencies, distinguishing the common phases from the background. Extensive experiments on three datasets (THUMOS14, ActivityNetv1.2, and a conjoint-action subset) demonstrate that JCDNet achieves competitive performance against the state-of-the-art methods."}, "cited_paper_content": {"title": "W-Talc: Weakly-Supervised Temporal Activity Localization And Classification", "abstract": "Most activity localization methods in the literature suffer from the burden of frame-wise annotation requirement. Learning from weak labels may be a potential solution towards reducing such manual labeling effort. Recent years have witnessed a substantial influx of tagged videos on the Internet, which can serve as a rich source of weakly-supervised training data. Specifically, the correlations between videos with similar tags can be utilized to temporally localize the activities. Towards this goal, we present W-TALC, a Weakly-supervised Temporal Activity Localization and Classification framework using only video-level labels. The proposed network can be divided into two sub-networks, namely the Two-Stream based feature extractor network and a weakly-supervised module, which we learn by optimizing two complimentary loss functions. Qualitative and quantitative results on two challenging datasets - Thumos14 and ActivityNet1.2, demonstrate that the proposed method is able to detect activities at a fine granularity and achieve better performance than current state-of-the-art methods."}, "keywords": ["co-activity similarity loss"], "citation_intent": "method"}
{"citing_id": "2304.03977v1", "cited_id": "1806.08887", "section_title": "More Related Works", "citation": "In 2018, a work called sparse manifold transform #REFR builds upon the above two areas.", "text_before_citation": ["SSL Methods Not Based on Deep Learning.", "Our work has also been inspired by the classical approaches before deep learning, especially sparse modeling and manifold learning.", "Some earlier works approach unsupervised learning mainly from the perspective of sparsity #OTHEREFR .", "In particular, a work focuses on lossy coding #OTHEREFR has inspired many of the recent SSL learning methods #OTHEREFR , as well as our work to promote covariance in the representation of data through maximizing the coding rate.", "Manifold learning #OTHEREFR and spectral clustering #OTHEREFR propose to model the geometric structure of high dimensional objects in the signal space."], "text_after_citation": ["The work proposes to use sparsity to handle locality in the data space to build support and construct representations that assign similar values to similar points on the support.", "One may note that this work already shares a similar idea with the current joint-embedding self-supervised learning in the deep-learning community."], "citing_paper_content": {"title": "Emp-Ssl: Towards Self-Supervised Learning In One Training Epoch Under Review", "abstract": "Recently, self-supervised learning (SSL) has achieved tremendous success in learning image representation. Despite the empirical success, most self-supervised learning methods are rather \"inefficient\" learners, typically taking hundreds of training epochs to fully converge. In this work, we show that the key towards efficient self-supervised learning is to increase the number of crops from each image instance. Leveraging one of the state-of-the-art SSL method, we introduce a simplistic form of self-supervised learning method called Extreme-Multi-Patch Self-Supervised-Learning (EMP-SSL) that does not rely on many heuristic techniques for SSL such as weight sharing between the branches, feature-wise normalization, output quantization, and stop gradient, etc, and reduces the training epochs by two orders of magnitude. We show that the proposed method is able to converge to 85.1% on CIFAR-10, 58.5% on CIFAR-100, 38.1% on Tiny ImageNet and 58.5% on ImageNet-100 in just one epoch. Furthermore, the proposed method achieves 91.5% on CIFAR-10, 70.1% on CIFAR-100, 51.5% on Tiny ImageNet and 78.9% on ImageNet-100 with linear probing in less than ten training epochs. In addition, we show that EMP-SSL shows significantly better transferability to out-of-domain datasets compared to baseline SSL methods. We will release the code in https://github.com/tsb0601/EMP-SSL. * Equal contribution 2 For example, all representations collapse to the same point."}, "cited_paper_content": {"title": "The Sparse Manifold Transform", "abstract": "We present a signal representation framework called the sparse manifold transform that combines key ideas from sparse coding, manifold learning, and slow feature analysis. It turns non-linear transformations in the primary sensory signal space into linear interpolations in a representational embedding space while maintaining approximate invertibility. The sparse manifold transform is an unsupervised and generative framework that explicitly and simultaneously models the sparse discreteness and low-dimensional manifold structure found in natural scenes. When stacked, it also models hierarchical composition. We provide a theoretical description of the transform and demonstrate properties of the learned representation on both synthetic data and natural videos."}, "keywords": ["sparse manifold transform"], "citation_intent": "background"}
{"citing_id": "2305.02224v2", "cited_id": "1702.06491", "section_title": "Conclusions And Future Work", "citation": "It is also evident there is diversity in fact-checking practices and that robustness of the workflow depends on the collaborative and ongoing efforts of the team members, confirming findings from our earlier study #REFR .", "text_before_citation": ["In this paper we have reported findings from a series of semi-structured interviews with news professionals working in fact-checking.", "We have seen how the fact-checking workflow consists of series of stages.", "Each stage has its particular challenges for fact-checkers working under pressure to determine quickly the veracity of a claim and publish the result."], "text_after_citation": ["Interviewees identified ways in which computational tools could assist throughout the fact-checking workflow: surfacing claims, deciding on which are check-worthy and assembling the evidence.", "Interviewees were skeptical about the prospects for automating the process.", "They take for granted that the fact-checker is the human-in-the-loop and that they must be the final arbiter when a claim is assessed.", "As with the introduction of advanced decision-making tools in other fields, there are important professional and organisational reasons why this should remain the case in fact-checking #OTHEREFR .", "Hence, like several recent studies (e.g., #OTHEREFR , #OTHEREFR and #OTHEREFR"], "citing_paper_content": {"title": "Some Observations On Fact-Checking Work With Implications For Computational Support", "abstract": "Social media and user-generated content (UGC) have become increasingly important features of journalistic work in a number of different ways. However, the growth of misinformation means that news organisations have had devote more and more resources to determining its veracity and to publishing corrections if it is found to be misleading. In this work, we present the results of interviews with eight members of fact-checking teams from two organisations. Team members described their fact-checking processes and the challenges they currently face in completing a fact-check in a robust and timely way. The former reveals, inter alia, significant differences in fact-checking practices and the role played by collaboration between team members. We conclude with a discussion of the implications for the development and application of computational tools, including where computational tool support is currently lacking and the importance of being able to accommodate different fact-checking practices."}, "cited_paper_content": {"title": "Supporting The Use Of User Generated Content In Journalistic Practice", "abstract": "Social media and user-generated content (UGC) are increasingly important features of journalistic work in a number of different ways. However, their use presents major challenges, not least because information posted on social media is not always reliable and therefore its veracity needs to be checked before it can be considered as fit for use in the reporting of news. We report on the results of a series of in-depth ethnographic studies of journalist work practices undertaken as part of the requirements gathering for a prototype of a social media verification 'dashboard' and its subsequent evaluation. We conclude with some reflections upon the broader implications of our findings for the design of tools to support journalistic work."}, "keywords": ["fact-checking practices"], "citation_intent": "result"}
{"citing_id": "2304.14463v1", "cited_id": "1910.02653", "section_title": "Formulation", "citation": "In addition to the methods of CHECKMATE and MOCCASIN, we include results for the rounding algorithm proposed in #REFR under the \"LP+rounding\" column of the table.", "text_before_citation": ["Table 2 provides numerical results for a range of different computation graphs.", "In Table 2 we have selected the memory budget values for each graph to be the 80% and 90% #OTHEREFR to evaluate CHECKMATE, where CM 1 is FCN with VGG layers and CM 2 is the ResNet50 model, n: number of nodes, m: number of edges, M : memory budget, TDI: total duration increase in percentage, peak mem: peak memory of the resulting rematerialization sequence.", "The column 'Time (s)' indicates the elapsed time in seconds until the best solution. Dashes \"-\" indicate that no solution is found.", "The best solution in each row is shown in bold font.", "Full version of this of the initial peak memory without rematerialization."], "text_after_citation": ["This method consists of relaxing the MILP into a linear program (LP) and then rounding the solution (see #OTHEREFR for further details on this algorithm).", "Note that solution produced by the rounding algorithm is not guaranteed to satisfy the memory budget constraint.", "This could be seen in Table 2 where in most cases the peak memory for the relaxation and rounding approach is higher than the memory budget M .", "Table 2 shows that the random layered graphs and real-world graphs are the most challenging ones among the graph set.", "The solve times for these graphs are higher than the CM graphs, which is consistent with higher edge densities and more complex edge connectivities of the RL and RW graphs."], "citing_paper_content": {"title": "Moccasin: Efficient Tensor Rematerialization For Neural Networks", "abstract": "The deployment and training of neural networks on edge computing devices pose many challenges. The low memory nature of edge devices is often one of the biggest limiting factors encountered in the deployment of large neural network models. Tensor rematerialization or recompute is a way to address high memory requirements for neural network training and inference. In this paper we consider the problem of execution time minimization of compute graphs subject to a memory budget. In particular, we develop a new constraint programming formulation called MOCCASIN with only O(n) integer variables, where n is the number of nodes in the compute graph. This is a significant improvement over the works in the recent literature that propose formulations with O(n 2) Boolean variables. We present numerical studies that show that our approach is up to an order of magnitude faster than recent work especially for large-scale graphs."}, "cited_paper_content": {"title": "Checkmate: Breaking The Memory Wall With Optimal Tensor Rematerialization", "abstract": "Modern neural networks are increasingly bottlenecked by the limited capacity of on-device GPU memory. Prior work explores dropping activations as a strategy to scale to larger neural networks under memory constraints. However, these heuristics assume uniform per-layer costs and are limited to simple architectures with linear graphs, limiting their usability. In this paper, we formalize the problem of trading-off DNN training time and memory requirements as the tensor rematerialization optimization problem, a generalization of prior checkpointing strategies. We introduce Checkmate, a system that solves for optimal schedules in reasonable times (under an hour) using off-the-shelf MILP solvers, then uses these schedules to accelerate millions of training iterations. Our method scales to complex, realistic architectures and is hardware-aware through the use of accelerator-specific, profile-based cost models. In addition to reducing training cost, Checkmate enables real-world networks to be trained with up to 5.1$\\times$ larger input sizes."}, "keywords": ["CHECKMATE"], "citation_intent": "method"}
{"citing_id": "2304.06036v1", "cited_id": "2002.00538", "section_title": "I. Introduction", "citation": "It has been found that movement execution (ME) generates stronger amplitude co-relates in EEG signals, hence decoding ME gave promising results when compared to decoding movement imagination (MI) #REFR .", "text_before_citation": ["For instance, it does not require surgical intervention, it is painless, cheap, portable, and accurate #OTHEREFR .", "Hence EEG is widely used in various biomedical applications including seizure detection #OTHEREFR , stress assessment #OTHEREFR , depression disorder detection #OTHEREFR , and schizophrenia #OTHEREFR .", "The interaction of the human body with its environment depends significantly on controlled movements of the upper limbs.", "However, spinal cord injury (SCI) and other neuro-muscular diseases can effect this control and limb movement.", "It is crucial to restoring upper limb movements for people with SCI so they can independently take care of their daily activities."], "text_after_citation": ["Deep learning (DL) methods are getting a lot of attention recently in a diverse set of data processing tasks.", "The same holds for upper limb movement classification using EEG data.", "Features like minimal pre-processing of data, automatic low-and high-level feature extraction, and superior learning capabilities make DL an attractive tool in the machine learning domain #OTHEREFR .", "Classification of the MI of two movements was performed using a convolutional neural network (CNN), along with common spatial patterns (CSP) as a feature extraction mechanism beforehand.", "The strategy was able to classify palm extension and hand grasp movements #OTHEREFR ."], "citing_paper_content": {"title": "Upper Limb Movement Execution Classification Using Electroencephalography For Brain Computer Interface", "abstract": "An accurate classification of upper limb movements using electroencephalography (EEG) signals is gaining significant importance in recent years due to the prevalence of brain-computer interfaces. The upper limbs in the human body are crucial since different skeletal segments combine to make a range of motion that helps us in our trivial daily tasks. Decoding EEG-based upper limb movements can be of great help to people with spinal cord injury (SCI) or other neuro-muscular diseases such as amyotrophic lateral sclerosis (ALS), primary lateral sclerosis, and periodic paralysis. This can manifest in a loss of sensory and motor function, which could make a person reliant on others to provide care in day-today activities. We can detect and classify upper limb movement activities, whether they be executed or imagined using an EEG-based brain-computer interface (BCI). Toward this goal, we focus our attention on decoding movement execution (ME) of the upper limb in this study. For this purpose, we utilize a publicly available EEG dataset that contains EEG signal recordings from fifteen subjects acquired using a 61-channel EEG device. We propose a method to classify four ME classes for different subjects using spectrograms of the EEG data through pre-trained deep learning (DL) models. Our proposed method of using EEG spectrograms for the classification of ME has shown significant results, where the highest average classification accuracy (for four ME classes) obtained is 87.36%, with one subject achieving the best classification accuracy of 97.03%. Clinical relevance-This research shows that movement execution of upper limbs is classified with significant accuracy by employing a spectrogram of the EEG signals and a pre-trained deep learning model which is fine-tuned for the downstream task."}, "cited_paper_content": {"title": "Decoding Movement Imagination And Execution From Eeg Signals Using Bci-Transfer Learning Method Based On Relation Network", "abstract": "A brain-computer interface (BCI) is used not only to control external devices for healthy people but also to rehabilitate motor functions for motor-disabled patients. Decoding movement intention is one of the most significant aspects for performing arm movement tasks using brain signals. Decoding movement execution (ME) from electroencephalogram (EEG) signals have shown high performance in previous works, however movement imagination (MI) paradigm-based intention decoding has so far failed to achieve sufficient accuracy. In this study, we focused on a robust MI decoding method with transfer learning for the ME and MI paradigm. We acquired EEG data related to arm reaching for 3D directions. We proposed a BCI-transfer learning method based on a Relation network (BTRN) architecture. Decoding performances showed the highest performance compared to conventional works. We confirmed the possibility of the BTRN architecture to contribute to continuous decoding of MI using ME datasets."}, "keywords": ["EEG signals"], "citation_intent": "result"}
{"citing_id": "2304.04151v1", "cited_id": "1706.03762", "section_title": "History Encoder", "citation": "To better capture long range spatial-temporal dependencies in users' historical check-in sequences, we stack Transformer encoder layers #REFR for constructing the history encoder.", "text_before_citation": ["All these vectors are then linearly projected into -dimensional embeddings \u2208 R , \u2208 R , and \u2208 R .", "In this way, for the user , the historical check-in sequence { } =1 can be further denoted as", "({ } =1 , { } =1 , { } =1 , { } =1 ).", "Note that since checkin data requires a certain order of precedence, learnable positional embedding is also added into inputs for history encoder.", "Compared with previous RNN-based methods, Transformer architecture #OTHEREFR can not only avoid recurrence, allowing parallel computing to reduce training time, but also migrate performance degradation problem with regard to long-term dependencies in RNNs."], "text_after_citation": ["Each Transformer encoder layer involves a multi-head self-attention module and a point-wise feed-forward network.", "We also keep the residual connection and layer normalization employed in Transformer encoder layers.", "Dividing the attention mechanism into multiple heads to form multiple sub-spaces allows the model to focus on different aspects of information.", "For each attention head, self-attention result for a check-in can be computed as", "EQUATION"], "citing_paper_content": {"title": "Timestamps As Prompts For Geography-Aware Location Recommendation", "abstract": "Figure 1: An illustration of how TPG performs the next location recommendation (denoted by the purple line) and interval predictions (denoted by red lines) by using temporal prompts. Different colored markers denote different categories of POIs."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["history encoder", "Transformer encoder layers"], "citation_intent": "method"}
{"citing_id": "2303.16079v1", "cited_id": "1802.06132", "section_title": "Limitations Of Existing Approaches", "citation": "A similar limitation has been reported for the simultaneous gradient descent-ascent (SGDA) approach #REFR .", "text_before_citation": ["First, we discuss the slow convergence issue on smooth strongly convex-concave problems highlighted in #OTHEREFR .", "For instance, consider a convex-concave quadratic problem ( , ) = ( /2) 2 + \u2212 ( /2) 2 . The worst-case scenario is\u02c6( ) = ( / )", "for each and the optimal solution is\u02c6( ) = \u2212( / ) for each .", "It is intuitive that both\u02c6( ) and\u02c6( ) should not be too sensitive to follow their change by (3).", "In fact, it has been theoretically derived that, for linear convergence, the learning rate must be set as , \u2208 ( /( + 2 )) and the required number of iterations to find near-optimal solution is \u03a9(1 + 2 /( )); refer to #OTHEREFR for details."], "text_after_citation": ["The same limitation is expected to exist in ZO-Min-Max because it is regarded as an approximation of the SGDA approach.", "The adaptation of the learning rates in ADV-CMA-ES can mitigate the difficulty in tuning learning rates. However, it cannot avoid the slow convergence problem.", "The situation is worse if the objective function is convex-concave but not strongly convex-concave.", "For example, consider 7 with = = 1 and = .", "This objective function is similar to , but the coefficients are regarded as = (1/2) 2 and = (1/2) 2 , i.e., decreasing as the solution approaches the global min-max saddle point ( * , * =\u02c6( * ))."], "citing_paper_content": {"title": "Covariance Matrix Adaptation Evolutionary Strategy With Worst-Case Ranking Approximation For Min-Max Optimization And Its Application To Berthing Control Tasks", "abstract": "In this study, we consider a continuous min-max optimization problem min \u2208X max \u2208Y (,) whose objective function is a black-box. We propose a novel approach to minimize the worst-case objective function () = max (,) directly using a covariance matrix adaptation evolution strategy (CMA-ES) in which the rankings of solution candidates are approximated by our proposed worst-case ranking approximation (WRA) mechanism. We develop two variants of WRA combined with CMA-ES and approximate gradient ascent as numerical solvers for the inner maximization problem. Numerical experiments show that our proposed approach outperforms several existing approaches when the objective function is a smooth strongly convex-concave function and the interaction between and is strong. We investigate the advantages of the proposed approach for problems where the objective function is not limited to smooth strongly convex-concave functions. The effectiveness of the proposed approach is demonstrated in the robust berthing control problem with uncertainty. CCS Concepts: \u2022 Mathematics of computing \u2192 Continuous optimization."}, "cited_paper_content": {"title": "Interaction Matters: A Note On Non-Asymptotic Local Convergence Of Generative Adversarial Networks", "abstract": "Motivated by the pursuit of a systematic computational and algorithmic understanding of Generative Adversarial Networks (GANs), we present a simple yet unified non-asymptotic local convergence theory for smooth two-player games, which subsumes several discrete-time gradient-based saddle point dynamics. The analysis reveals the surprising nature of the off-diagonal interaction term as both a blessing and a curse. On the one hand, this interaction term explains the origin of the slow-down effect in the convergence of Simultaneous Gradient Ascent (SGA) to stable Nash equilibria. On the other hand, for the unstable equilibria, exponential convergence can be proved thanks to the interaction term, for four modified dynamics proposed to stabilize GAN training: Optimistic Mirror Descent (OMD), Consensus Optimization (CO), Implicit Updates (IU) and Predictive Method (PM). The analysis uncovers the intimate connections among these stabilizing techniques, and provides detailed characterization on the choice of learning rate. As a by-product, we present a new analysis for OMD proposed in Daskalakis, Ilyas, Syrgkanis, and Zeng [2017] with improved rates."}, "keywords": ["simultaneous gradient descent-ascent"], "citation_intent": "result"}
{"citing_id": "2305.01289v1", "cited_id": "1903.00868", "section_title": "Cubature Rule For Unitary Jacobi Ensembles", "citation": "The corresponding specialization of the cubature rule in Theorem 2 was presented in turn in #REFR Section 9] .", "text_before_citation": ["\u03ba \u00b1 := 1 2 1\u2264r\u2264d 1 \u2212 |a r | 1 + |a r | \u00b11 + 1 2 1\u2264r\u2264d 1 \u2212 |\u00e3 r| 1 + |\u00e3 r| \u00b11 .", "Moreover, it is clear from Eqs.", "(3.1d), (3.1e) that the boundary value \u03be (m+n) 0 = 0 is attained iff \u01eb \u2212 =\u01eb \u2212 = 0 and the boundary value \u03be (m+n) m+n\u22121 = \u03c0 is attained iff \u01eb + =\u01eb + = 0 (since \u03c0 0 u a (\u03b8)d\u03b8 = \u03c0 for |a| < 1). Remark 4.", "In the above integration formulas the degree of exactness is optimal if D (3.2b) reaches the Gaussian value 2m+1, which is achieved whend = 0 and\u01eb \u00b1 = 1. This special case of the quadrature rule in Eqs.", "(3.1a)-(3.1g) can be inferred from #OTHEREFR for \u01eb \u00b1 = 0, and from #OTHEREFR for general \u01eb \u00b1 \u2208 {0, 1} (cf. also #OTHEREFR Section 8] )."], "text_after_citation": ["(2\u03c0) n n! \u03c0 0 \u2022 \u2022 \u2022 \u03c0 0 f cos(\u03be) \u03c1 \u01eb (\u03be)d\u03be 1 \u2022 \u2022 \u2022 d\u03be n = (3.4a) 1 N (m,n) \u01eb \u03bb\u2208\u039b (m,n) 1 2 (1\u2212\u01eb+)(1\u2212\u01eb+)\u03b4 m\u2212\u03bb 1 +(1\u2212\u01eb\u2212)(1\u2212\u01eb\u2212)\u03b4 \u03bbn f cos \u03be (m,n) \u03bb \u03c1 \u01eb \u03be (m,n) \u03bb", "for f cos(\u03be) := f cos(\u03be 1 ), . . .", ", cos(\u03be n ) , where f (x 1 , . . .", ", x n ) = f (x) \u2208 P (D,n) with D = 2m +\u01eb + +\u01eb \u2212 \u2212 1 and", "N (m,n) \u01eb := 2(m + n \u2212 1) + \u01eb + + \u01eb \u2212 +\u01eb + +\u01eb \u2212 n . (3.4b)"], "citing_paper_content": {"title": "Cubature Rules For Unitary Jacobi Ensembles", "abstract": "We present Chebyshev type cubature rules for the exact integration of rational symmetric functions with poles on prescribed coordinate hyperplanes. Here the integration is with respect to the densities of unitary Jacobi ensembles stemming from the Haar measures of the orthogonal and the compact symplectic Lie groups."}, "cited_paper_content": {"title": "Exact Cubature Rules For Symmetric Functions", "abstract": "We employ a multivariate extension of the Gauss quadrature formula, originally due to Berens, Schmid and Xu [BSX95], so as to derive cubature rules for the integration of symmetric functions over hypercubes (or infinite limiting degenerations thereof) with respect to the densities of unitary random matrix ensembles. Our main application concerns the explicit implementation of a class of cubature rules associated with the Bernstein-Szego polynomials, which permit the exact integration of symmetric rational functions with prescribed poles at coordinate hyperplanes against unitary circular Jacobi distributions stemming from the Haar measures on the symplectic and the orthogonal groups."}, "keywords": ["cubature rule"], "citation_intent": "background"}
{"citing_id": "2304.04017v2", "cited_id": "1707.02880", "section_title": "A. Automatic Photo Retouching", "citation": "HDRNet #REFR obtains transformations on the low-resolution input and applies upsampled transformations to the full-resolution image with bilateral grid processing.", "text_before_citation": ["StarEnhancer #OTHEREFR introduces multiple-style enhancement with the ability to transform inputs to an unseen style.", "CSRNet #OTHEREFR designs a lightweight framework containing a base network and a conditioning network for extracting global features and performing photo retouching, respectively.", "Pienet #OTHEREFR constructs preference vectors with metric learning and adaptively enhances images according to user-provided preferable styles.", "Operator Prediction Methods.", "DeepLPF #OTHEREFR regresses the parameters of spatially localized filters and automatically applies those filters to enhance inputs."], "text_after_citation": ["Enhancement curve learning methods #OTHEREFR , #OTHEREFR , #OTHEREFR estimate retouching curves to tone global properties of inputs rather than directly mapping.", "RCTNet #OTHEREFR first estimates the transformation of the representative colors and then enhances inputs according to the similarity between inputs and representative colors.", "3D LUT based methods #OTHEREFR , #OTHEREFR utilize variants of 3D lookup tables (3D LUTs) with deep learning, achieving real-time and flexible photo enhancement performance.", "In contrast to most of the methods above, 3D LUT HRP [2] focuses on portrait retouching and achieves region awareness through the human-region priority strategy.", "Our work distinguishes from 3D LUT HRP #OTHEREFR by emphasizing the significance of the user guidance and investigating interactive region-aware portrait retouching, whcih provides the flexibility to retouch different instances according to users' intents."], "citing_paper_content": {"title": "Region-Aware Portrait Retouching With Sparse Interactive Guidance", "abstract": "Portrait retouching aims to improve the aesthetic quality of input portrait photos and especially requires humanregion priority. The deep learning-based methods largely elevate the retouching efficiency and provide promising retouched results. However, existing portrait retouching methods focus on automatic retouching, which treats all human-regions equally and ignores users' preferences for specific individuals, thus suffering from limited flexibility in interactive scenarios. In this work, we emphasize the importance of users' intents and explore the interactive portrait retouching task. Specifically, we propose a regionaware retouching framework with two branches: an automatic branch and an interactive branch. The automatic branch involves an encoding-decoding process, which searches region candidates and performs automatic region-aware retouching without user guidance. The interactive branch encodes sparse user guidance into a priority condition vector and modulates latent features with a region selection module to further emphasize the userspecified regions. Experimental results show that our interactive branch effectively captures users' intents and generalizes well to unseen scenes with sparse user guidance, while our automatic branch also outperforms the state-of-the-art retouching methods due to improved region-awareness."}, "cited_paper_content": {"title": "Deep Bilateral Learning For Real-Time Image Enhancement", "abstract": "Performance is a critical challenge in mobile image processing. Given a reference imaging pipeline, or even human-adjusted pairs of images, we seek to reproduce the enhancements and enable real-time evaluation. For this, we introduce a new neural network architecture inspired by bilateral grid processing and local affine color transforms. Using pairs of input/output images, we train a convolutional neural network to predict the coefficients of a locally-affine model in bilateral space. Our architecture learns to make local, global, and content-dependent decisions to approximate the desired image transformation. At runtime, the neural network consumes a low-resolution version of the input image, produces a set of affine transformations in bilateral space, upsamples those transformations in an edge-preserving fashion using a new slicing node, and then applies those upsampled transformations to the full-resolution image. Our algorithm processes high-resolution images on a smartphone in milliseconds, provides a real-time viewfinder at 1080p resolution, and matches the quality of state-of-the-art approximation techniques on a large class of image operators. Unlike previous work, our model is trained off-line from data and therefore does not require access to the original operator at runtime. This allows our model to learn complex, scene-dependent transformations for which no reference implementation is available, such as the photographic edits of a human retoucher."}, "keywords": ["upsampled transformations"], "citation_intent": "method"}
{"citing_id": "2304.05015v1", "cited_id": "1907.13372", "section_title": "Introduction", "citation": "For example, #REFR selects the most common samples with the lowest diversity for replay, believing that the most representative samples will elevate the effectiveness of replay.", "text_before_citation": ["By doing so, the model can be trained with samples from both current and previous classes, resulting in better generalization.", "However, since the number of selected samples in the memory is much smaller than those within the new classes, the selected samples are easy to be ignored or cause overfitting when training due to the small number.", "Careful selection of the samples is required, which naturally brings the question: How to select the best samples for replay?", "Some attempts have been made to answer the question, aiming to seek the most effective samples for replay.", "Researchers propose different criteria that are mostly manually designed based on some heuristic factors like diversity #OTHEREFR ."], "text_after_citation": ["However, the most common samples may not always be the samples being forgotten in later stages.", "#OTHEREFR proposes to save both the low-diversity samples near the distribution center and high-diversity samples near the classification boundaries.", "However, new challenges arise since the memory length is limited, so it is challenging to find the optimal quotas for the two kinds of samples to promote replay effectiveness to the greatest extent.", "Moreover, most of the existing methods are designed based on a single factor, the selection performance, however, can be influenced by many factors with complicated relationships.", "For example, besides diversity, memory sample selection should also be class-dependent because the hard classes need more samples to replay in order to alleviate the more severe catastrophic forgetting issue."], "citing_paper_content": {"title": "Continual Semantic Segmentation With Automatic Memory Sample Selection", "abstract": "Continual Semantic Segmentation (CSS) extends static semantic segmentation by incrementally introducing new classes for training. To alleviate the catastrophic forgetting issue in CSS, a memory buffer that stores a small number of samples from the previous classes is constructed for replay. However, existing methods select the memory samples either randomly or based on a single-factor-driven handcrafted strategy, which has no guarantee to be optimal. In this work, we propose a novel memory sample selection mechanism that selects informative samples for effective replay in a fully automatic way by considering comprehensive factors including sample diversity and class performance. Our mechanism regards the selection operation as a decision-making process and learns an optimal selection policy that directly maximizes the validation performance on a reward set. To facilitate the selection decision, we design a novel state representation and a dual-stage action space. Our extensive experiments on Pascal-VOC 2012 and ADE 20K datasets demonstrate the effectiveness of our approach with state-of-the-art (SOTA) performance achieved, outperforming the second-place one by 12.54% for the 6stage setting on Pascal-VOC 2012."}, "cited_paper_content": {"title": "Incremental Learning Techniques For Semantic Segmentation", "abstract": "Deep learning architectures exhibit a critical drop of performance due to catastrophic forgetting when they are required to incrementally learn new tasks. Contemporary incremental learning frameworks focus on image classification and object detection while in this work we formally introduce the incremental learning problem for semantic segmentation in which a pixel-wise labeling is considered. To tackle this task we propose to distill the knowledge of the previous model to retain the information about previously learned classes, whilst updating the current model to learn the new ones. We propose various approaches working both on the output logits and on intermediate features. In opposition to some recent frameworks, we do not store any image from previously learned classes and only the last model is needed to preserve high accuracy on these classes. The experimental evaluation on the Pascal VOC2012 dataset shows the effectiveness of the proposed approaches."}, "keywords": ["replay"], "citation_intent": "background"}
{"citing_id": "2303.10344v1", "cited_id": "1606.00373", "section_title": "Loss Function And Training Details", "citation": "We optimize PanoTransformer by minimizing a pixelwise reverse Huber loss #REFR between the predicted panoramas and corresponding ground truth.", "text_before_citation": [], "text_after_citation": ["Since using a standard L1 loss function to learn a binary light mask heavily penalizes even small shifts of a light source position, the reverse Huber loss takes advantage of L1 loss and L2 loss as below:", "L B = \uf8f1 \uf8f2 \uf8f3 |y \u2212\u0177|, |y \u2212\u0177| \u2264 T (y \u2212\u0177) 2 + T 2 2T , |y \u2212\u0177| \u2265 T (4)", "where y is the ground truth value and\u0177 is the prediction.", "The threshold T is set to 0.2 in our experiments.", "To generate more realistic details, an extra adversarial loss is also involved in the training process."], "citing_paper_content": {"title": "Local-To-Global Panorama Inpainting For Locale-Aware Indoor Lighting Prediction", "abstract": "Figure 1. We propose a locale-aware indoor illumination prediction method that can generate a full and texture-rich HDR panorama (the third column in each group) at any locale in the scene, enabling spatially-varying and consistent shading after virtual object insertion (the first and second columns)."}, "cited_paper_content": {"title": "Deeper Depth Prediction With Fully Convolutional Residual Networks", "abstract": "This paper addresses the problem of estimating the depth map of a scene given a single RGB image. To model the ambiguous mapping between monocular images and depth maps, we leverage on deep learning capabilities and present a fully convolutional architecture encompassing residual learning. The proposed model is deeper than the current state of the art, but contains fewer parameters and requires less training data, while still outperforming all current CNN approaches aimed at the same task. We further present a novel way to efficiently learn feature map up-sampling within the network. For optimization we introduce the reverse Huber loss, particularly suited for the task at hand and driven by the value distributions commonly present in depth maps. The predictions are given by a single architecture, trained end-to-end, that does not rely on post-processing techniques, such as CRFs or other additional refinement steps."}, "keywords": ["predicted panoramas"], "citation_intent": "method"}
{"citing_id": "2303.02244v1", "cited_id": "2002.07764", "section_title": "B. Survey Measures", "citation": "In the email, we pledged to donate US$ 2 per completed response to the United Nations World Food Programme #REFR .", "text_before_citation": ["We also merged records where individuals used different names but the same email address, or different email addresses but the same or an extremely similar name for different commits.", "When judgment was needed, we made decisions as conservatively as possible. We ultimately obtained a list of 1,555 unique individuals.", "In case of multiple email addresses per person, we selected only one, preferring personal email addresses over professional ones because the person's commits to the projects were partially already quite old and the person might have moved organizations since.", "We sent an email to all 1,555 developers to invite them to our survey #OTHEREFR (which was part of a larger data collection effort for multiple studies).", "We sent the invitation with a personalized link to the survey."], "text_after_citation": ["We also sent two reminders #OTHEREFR , including one that linked to an official university web page confirming the authenticity of the survey as some developers were concerned that the invitation might be a scam.", "Because 165 emails bounced, we reached 1,390 developers (89.4% deliverable emails), of which 194 developers started the survey, and 124 completed it.", "We dropped all respondents that provided implausible values for their age (\u226410 or \u2265100), leaving us with a preliminary sample of 121 developers.", "Although this number may seem low, it is of a similar magnitude as the total number even from studies that assessed personality not based on self-reports but on mining vast corpora of emails.", "Calefato et al., for instance, used about 1.35 million emails from 46,304 developers but only obtained full personality profiles for 211 of them (<0.5%), of which only 118 had made source code commits #OTHEREFR ."], "citing_paper_content": {"title": "The Type To Take Out A Loan? A Study Of Developer Personality And Technical Debt", "abstract": "Background: Technical debt (TD) has been widely discussed in software engineering research, and there is an emerging literature linking it to developer characteristics. However, developer personality has not yet been studied in this context. Aims and Method: We explore the relationship between various personality traits (Five Factor Model, regulatory focus, and narcissism) of developers and the introduction and removal of TD. To this end, we complement an existing TD dataset with novel selfreport personality data gathered by surveying developers, and analyze 2,145 commits from 19 developers. Results: We find that conscientiousness, emotional stability, openness to experience, and prevention focus are negatively associated with TD. There were no significant results for extraversion, agreeableness, promotion focus, or narcissism. Conclusions: We take our results as first evidence that developer personality has a systematic influence on the introduction and removal of TD. This has implications not only for future research, which could, for example, study the effects of personality on downstream consequences of TD like defects, but also for software engineering practitioners who may, for example, consider developer personality in staffing decisions."}, "cited_paper_content": {"title": "Sampling In Software Engineering Research: A Critical Review And Guidelines", "abstract": "Representative sampling appears rare in software engineering research. Not all studies need representative samples, but a general lack of representative sampling undermines a scientific field. This study therefore investigates the state of sampling in recent, high-quality software engineering research. The key findings are: (1) random sampling is rare; (2) sophisticated sampling strategies are very rare; (3) sampling, representativeness and randomness do not appear well-understood. To address these problems, the paper synthesizes existing knowledge of sampling into a succinct primer and proposes extensive guidelines for improving the conduct, presentation and evaluation of sampling in software engineering research. It is further recommended that while researchers should strive for more representative samples, disparaging non-probability sampling is generally capricious and particularly misguided for predominately qualitative research."}, "keywords": ["email"], "citation_intent": "background"}
{"citing_id": "2303.00894v1", "cited_id": "1811.07871", "section_title": "Introduction", "citation": "By incentivizing incorrect behavior, misspecified objectives can lead to useless or even dangerous outcomes #REFR .", "text_before_citation": ["Standard AI and machine learning algorithms require the designer to specify a cost or reward function.", "This objective incentivizes desired behavior and penalizes mistakes, teaching the system how to perform the task.", "While such objectives are easy to manually specify for problems with clear win conditions, such as games #OTHEREFR and tasks with clear goals, such as image classification #OTHEREFR , they can be challenging to formalize for more nuanced tasks #OTHEREFR . For example, Lee et al.", "#OTHEREFR find that humans struggle to define an objective that incentivizes bipedal locomotion, despite being experts in both machine learning and walking."], "text_after_citation": ["Ensuring that AI systems optimize objectives that align with our own is a crucial part of building safe and beneficial AI.", "Reward learning techniques enable AI systems to learn their objectives by observing and interacting with humans instead of requiring their designers to specify these objectives manually #OTHEREFR .", "Humans can train reward learning systems using a variety of feedback modalities, including demonstrations #OTHEREFR , pairwise comparisons #OTHEREFR , natural language #OTHEREFR , numeric values #OTHEREFR , corrections #OTHEREFR , and proxy rewards #OTHEREFR .", "Reward learning from pairwise comparisons in particular has proven remarkably effective across a variety of tasks, including complex physical maneuvers for continuous control systems #OTHEREFR and text summarization for language language models #OTHEREFR .", "In the future, it may even be possible to use reward learning to train AI systems to assist humans in researching safe AI #OTHEREFR ."], "citing_paper_content": {"title": "Active Reward Learning From Multiple Teachers", "abstract": "Reward learning algorithms utilize human feedback to infer a reward function, which is then used to train an AI system. This human feedback is often a preference comparison, in which the human teacher compares several samples of AI behavior and chooses which they believe best accomplishes the objective. While reward learning typically assumes that all feedback comes from a single teacher, in practice these systems often query multiple teachers to gather sufficient training data. In this paper, we investigate this disparity, and find that algorithmic evaluation of these different sources of feedback facilitates more accurate and efficient reward learning. We formally analyze the value of information (VOI) when reward learning from teachers with varying levels of rationality, and define and evaluate an algorithm that utilizes this VOI to actively select teachers to query for feedback. Surprisingly, we find that it is often more informative to query comparatively irrational teachers. By formalizing this problem and deriving an analytical solution, we hope to facilitate improvement in reward learning approaches to aligning AI behavior with human values."}, "cited_paper_content": {"title": "Scalable Agent Alignment Via Reward Modeling: A Research Direction", "abstract": "One obstacle to applying reinforcement learning algorithms to real-world problems is the lack of suitable reward functions. Designing such reward functions is difficult in part because the user only has an implicit understanding of the task objective. This gives rise to the agent alignment problem: how do we create agents that behave in accordance with the user's intentions? We outline a high-level research direction to solve the agent alignment problem centered around reward modeling: learning a reward function from interaction with the user and optimizing the learned reward function with reinforcement learning. We discuss the key challenges we expect to face when scaling reward modeling to complex and general domains, concrete approaches to mitigate these challenges, and ways to establish trust in the resulting agents."}, "keywords": ["misspecified objectives"], "citation_intent": "background"}
{"citing_id": "2304.00858v1", "cited_id": "1911.12409", "section_title": "The Multi-View Auto-Encoder Backbone", "citation": "Instead of a two-stage encoder for feature extraction #REFR , our FoCoViL makes full use of a single auto-encoder in an end-to-end fashion that already achieves promising results.", "text_before_citation": ["To maintain the action representation, we employ an effective sequential autoencoder as the backbone network sharing among the multi-view actions (see Fig. 2(c) ).", "From #OTHEREFR , the encoder usually plays a more important role than the decoder to integrate representative features.", "We thus consider a portable decoder by feeding the empty frame (zero vector) to every step of the decoder, such that the model only focuses on the hidden representation delivered from the encoded output."], "text_after_citation": ["The structure consists of a three-layer bi-directional encoder f e to derive the latent representation, a linear projection net g that is specifically designed for contrastive learning, and a single-layer uni-directional decoder f d for reconstruction purposes.", "Both f e and f d are under the Gated Recurrent Unit (GRU) architecture to process frame-wise information.", "For each action X u i , the reconstruction loss L r is defined as:", "EQUATION", "where T is the total number of frames in the action video."], "citing_paper_content": {"title": "Focalized Contrastive View-Invariant Learning For Self-Supervised Skeleton-Based Action Recognition", "abstract": "Learning view-invariant representation is a key to improving feature discrimination power for skeleton-based action recognition. Existing approaches cannot effectively remove the impact of viewpoint due to the implicit view-dependent representations. In this work, we propose a self-supervised framework called Focalized Contrastive View-invariant Learning (FoCoViL), which significantly suppresses the view-specific information on the representation space where the viewpoints are coarsely aligned. By maximizing mutual information with an effective contrastive loss between multi-view sample pairs, FoCoViL associates actions with common view-invariant properties and simultaneously separates the dissimilar ones. We further propose an adaptive focalization method based on pairwise similarity to enhance contrastive learning for a clearer cluster boundary in the learned space. Different from many existing self-supervised representation learning work that rely heavily on supervised classifiers, FoCoViL performs well on both unsupervised and supervised classifiers with superior recognition performance. Extensive experiments also show that the proposed contrastive-based focalization generates a more discriminative latent representation."}, "cited_paper_content": {"title": "Predict&Cluster: Unsupervised Skeleton Based Action Recognition", "abstract": "We propose a novel system for unsupervised skeleton-based action recognition. Given inputs of body keypoints sequences obtained during various movements, our system associates the sequences with actions. Our system is based on an encoder-decoder recurrent neural network, where the encoder learns a separable feature representation within its hidden states formed by training the model to perform prediction task. We show that according to such unsupervised training the decoder and the encoder self-organize their hidden states into a feature space which clusters similar movements into the same cluster and distinct movements into distant clusters. Current state-of-the-art methods for action recognition are strongly supervised, i.e., rely on providing labels for training. Unsupervised methods have been proposed, however, they require camera and depth inputs (RGB+D) at each time step. In contrast, our system is fully unsupervised, does not require labels of actions at any stage, and can operate with body keypoints input only. Furthermore, the method can perform on various dimensions of body keypoints (2D or 3D) and include additional cues describing movements. We evaluate our system on three extensive action recognition benchmarks with different number of actions and examples. Our results outperform prior unsupervised skeleton-based methods, unsupervised RGB+D based methods on cross-view tests and while being unsupervised have similar performance to supervised skeleton-based action recognition."}, "keywords": ["single auto-encoder"], "citation_intent": "method"}
{"citing_id": "2303.00171v1", "cited_id": "1603.02754", "section_title": "Gbdt", "citation": "For the given user and TTS pronunciations, the phoneme embedding sequences are concatenated and used as input to train a GBDT model using XGBoost #REFR with logistic loss.", "text_before_citation": ["We train a Gradient Boosted Decision Tree (GBDT) #OTHEREFR classifier using phoneme embeddings as input."], "text_after_citation": ["The annotations are binary labels, where 0 represents both pronunciations are the same, and 1 otherwise."], "citing_paper_content": {"title": "Dtw-Siamesenet: Dynamic Time Warped Siamese Network For Mispronunciation Detection And Correction", "abstract": "Personal Digital Assistants (PDAs)-such as Siri, Alexa and Google Assistant, to name a few-play an increasingly important role to access information and complete tasks spanning multiple domains, and by diverse groups of users. A textto-speech (TTS) module allows PDAs to interact in a natural, human-like manner, and play a vital role when the interaction involves people with visual impairments or other disabilities. To cater to the needs of a diverse set of users, inclusive TTS is important to recognize and pronounce correctly text in different languages and dialects. Despite great progress in speech synthesis, the pronunciation accuracy of named entities in a multilingual setting still has a large room for improvement. Existing approaches to correct named entity (NE) mispronunciations, like retraining Grapheme-to-Phoneme (G2P) models, or maintaining a TTS pronunciation dictionary, require expensive annotation of the ground truth pronunciation, which is also time consuming. In this work, we present a highly-precise, PDA-compatible pronunciation learning framework for the task of TTS mispronunciation detection and correction. In addition, we also propose a novel mispronunciation detection model called DTW-SiameseNet, which employs metric learning with a Siamese architecture for Dynamic Time Warping (DTW) with triplet loss. We demonstrate that a locale-agnostic, privacypreserving solution to the problem of TTS mispronunciation detection is feasible. We evaluate our approach on a real-world dataset, and a corpus of NE pronunciations of an anonymized audio dataset of person names recorded by participants from 10 different locales. Human evaluation shows our proposed approach improves pronunciation accuracy on average by \u2248 6% compared to strong phoneme-based and audio-based baselines."}, "cited_paper_content": {"title": "Xgboost: A Scalable Tree Boosting System", "abstract": "Tree boosting is a highly effective and widely used machine learning method. In this paper, we describe a scalable end-to-end tree boosting system called XGBoost, which is used widely by data scientists to achieve state-of-the-art results on many machine learning challenges. We propose a novel sparsity-aware algorithm for sparse data and weighted quantile sketch for approximate tree learning. More importantly, we provide insights on cache access patterns, data compression and sharding to build a scalable tree boosting system. By combining these insights, XGBoost scales beyond billions of examples using far fewer resources than existing systems."}, "keywords": ["TTS pronunciations", "XGBoost"], "citation_intent": "method"}
{"citing_id": "2304.08747v1", "cited_id": "cs/0702015", "section_title": "I. Introduction", "citation": "Furthermore, the results of #REFR showed that there is a trade-off between the minimum repair bandwidth and the storage capacity of the network.", "text_before_citation": ["The rapid development of distributed storage systems raised the question of how failed nodes in these systems can be efficiently repaired.", "A large body of literature on erasure codes for distributed storage addressed this question over the past decade.", "One approach to assess repair efficiency is to measure the so-called repair bandwidth, which is the amount of information downloaded from other nodes for the repair.", "This approach, first introduced in #OTHEREFR , assumes the nodes form a homogeneous network and determines the repair bandwidth by considering the information flow in course of the repair.", "In particular, #OTHEREFR derived a bound on the smallest number of symbols required for the repair of a single failed node, known as the cut-set bound on the repair bandwidth."], "text_after_citation": ["In this paper we shall focus on codes with minimum storage overhead and optimal repair bandwidth, namely, maximum distance separable (MDS) codes with optimal repair bandwidth.", "Such codes are termed minimum-storage regenerating (MSR) codes in the literature.", "The problem of MSR codes has received much attention and affords several variants.", "In its basic form, the problem concerns repair of a single failed node, but it has been generalized to repair of multiple failed nodes.", "There are generally two different models for repairing multiple failed nodes: the centralized repair model #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR and the cooperative repair model #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR ."], "citing_paper_content": {"title": "Rack-Aware Minimum-Storage Regenerating Codes With Optimal Access", "abstract": "We derive a lower bound on the amount of information accessed to repair failed nodes within a single rack from any number of helper racks in the rack-aware storage model that allows collective information processing in the nodes that share the same rack. Furthermore, we construct a family of rack-aware minimum-storage regenerating (MSR) codes with the property that the number of symbols accessed for repairing a single failed node attains the bound with equality for all admissible parameters. Constructions of rack-aware optimal-access MSR codes were only known for limited parameters. We also present a family of Reed-Solomon (RS) codes that only require accessing a relatively small number of symbols to repair multiple failed nodes in a single rack. In particular, for certain code parameters, the RS construction attains the bound on the access complexity with equality and thus has optimal access."}, "cited_paper_content": {"title": "Network Coding For Distributed Storage Systems", "abstract": "Peer-to-peer distributed storage systems provide reliable access to data through redundancy spread over nodes across the Internet. A key goal is to minimize the amount of bandwidth used to maintain that redundancy. Storing a file using an erasure code, in fragments spread across nodes, promises to require less redundancy and hence less maintenance bandwidth than simple replication to provide the same level of reliability. However, since fragments must be periodically replaced as nodes fail, a key question is how to generate a new fragment in a distributed way while transferring as little data as possible across the network. In this paper, we introduce a general technique to analyze storage architectures that combine any form of coding and replication, as well as presenting two new schemes for maintaining redundancy using erasure codes. First, we show how to optimally generate MDS fragments directly from existing fragments in the system. Second, we introduce a new scheme called regenerating codes which use slightly larger fragments than MDS but have lower overall bandwidth use. We also show through simulation that in realistic environments, regenerating codes can reduce maintenance bandwidth use by 25% or more compared with the best previous design - a hybrid of replication and erasure codes - while simplifying system architecture."}, "keywords": ["minimum repair", "storage capacity"], "citation_intent": "result"}
{"citing_id": "2305.01518v1", "cited_id": "1310.0606", "section_title": "Replicability", "citation": "This definition is consistent with that given by the American Statistical Association [6] and with other work in statistics #REFR .", "text_before_citation": ["From this follows the desideratum that scientific predictions also agree well with experimental observations made across sufficiently similar circumstances. Defining this rigorously is not straightforward.", "In 2019 the USA's National Academies established a Committee on Reproducibility and Replicability in Science.", "Their report #OTHEREFR is essential reading for those interested in this topic.", "Though their focus is primarily on scientific hypotheses, their definition is a good starting point for this discussion.", "Conclusion 3-1 on page 36 states: \"Replicability is obtaining consistent results across studies aimed at answering the same scientific question, each of which has obtained its own data\"."], "text_after_citation": ["A distinct concept is that of repeatability: a repeatable prediction approach produces predictions without variation across independent tests carried out by repeating the entire process, including data collection, on the same individual or sampling unit #OTHEREFR . This is important but is not examined here.", "Replicability is also used in contrast to reproducibility, defined as \"obtaining consistent results using the same input data, computational steps, methods, and code, and conditions of analysis\" #OTHEREFR .", "Usage of these terms is often inconsistent and is plainly reversed in computer science -a sarcastic twist in the parallel evolution of the same concepts in siloed fields.", "A thread of literature debates and documents related terminologies and their usage #OTHEREFR .", "For predictions, I propose to modify the NAS definition to say:"], "citing_paper_content": {"title": "Defining Replicability Of Prediction Rules", "abstract": "In this article I propose an approach for defining replicability for prediction rules. Motivated by a recent NAS report, I start from the perspective that replicability is obtaining consistent results across studies suitable to address the same prediction question, each of which has obtained its own data. I then discuss concept and issues in defining key elements of this statement. I focus specifically on the meaning of \"consistent results\" in typical utilization contexts, and propose a multi-agent framework for defining replicability, in which agents are neither partners nor adversaries. I recover some of the prevalent practical approaches as special cases. I hope to provide guidance for a more systematic assessment of replicability in machine learning."}, "cited_paper_content": {"title": "Deciding Whether Follow-Up Studies Have Replicated Findings In A Preliminary Large-Scale\"Omics' Study\"", "abstract": "We propose a formal method to declare that findings from a primary study have been replicated in a follow-up study. Our proposal is appropriate for primary studies that involve large-scale searches for rare true positives (i.e., needles in a haystack). Our proposal assigns an r value to each finding; this is the lowest false discovery rate at which the finding can be called replicated. Examples are given and software is available."}, "keywords": ["American Statistical Association"], "citation_intent": "result"}
{"citing_id": "2304.04364v3", "cited_id": "1812.04948", "section_title": "Preliminaries", "citation": "Specifically, EG3D starts with randomly sampled GAN latent codes, and then a feature generator based on Style-GAN #REFR converts latent codes into 2D features and maps them into 3D tri-planes.", "text_before_citation": ["EG3D.", "We first briefly review the network architecture of a SOTA 3D network, EG3D #OTHEREFR .", "Our generator G 3 , G 3 and G 3 are finetuned on EG3D."], "text_after_citation": ["An MLP decoder predicts features for 3D point projections on tri-planes to generate color and density.", "Finally, volume rendering is used to generate an image on the orientation of the camera pose. This process can be formulated as follows:", "EQUATION", "where W 3 \u2208 R 1\u00d714\u00d7512 is the lantent code. P \u2208 R 1\u00d725 is the camera pose.", "is the weight parameters of the EG3D generator G 3 . I is the generated images. CLIP-guided Loss."], "citing_paper_content": {"title": "Itportrait: Image-Text Coupled 3D Portrait Domain Adaptation", "abstract": "Artistic portraits have many applications [10, 12, 22, 24, 34] in our daily lives, especially in industries related to animation, art, and the metaverse. As shown in Fig. 1, artistic portraits can be regarded as a portrait domain adaptation task, which refers to transforming the artistic style, cross-species identity, and expression shape change. The current domain adaption methods are mainly divided into two categories: guided by the vision-based method (artistic-image [6, 21, 34]), or guided by the language-based method (text-description [23, 36]). Combining image-guided and text-driven guidance can not only transfer the precise and detailed style of the reference image but also have the text-driven flexible editing ability. Therefore, Image-Text coupled guidance has better style control-ability and artistic merit [8, 27]. However, the potential of Vision-Language (Image-Text) multi-modal guidance is under-explored."}, "cited_paper_content": {"title": "A Style-Based Generator Architecture For Generative Adversarial Networks", "abstract": "We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces."}, "keywords": ["Style-GAN converts latent"], "citation_intent": "method"}
{"citing_id": "2304.11823v1", "cited_id": "1712.05526", "section_title": "Conclusion", "citation": "For Blended #REFR , we blend the poisoned samples with a Hello-Ketty image and the blended ratio is 0.1.", "text_before_citation": ["We evaluate our method on PreAct-ResNet18 #OTHEREFR and VGG19-BN #OTHEREFR networks.", "We compare our method with SOTA defense methods on three datasets and the two networks with a 10% poisining ratio and 5% clean samples for defense.", "To study the effectiveness of our method under different poisoning ratios, we compare with SOTA defense methods on CIFAR-10 dataset and PreAct-ResNet18 network on 5% and 1% poisoning ratios.", "Attack Details. We introduce some details about the backdoor attacks here.", "For BadNets-A2O and BadNets-A2A [12], we patch a 3 \u00d7 3 white square in the lower right corner of the images for CIFAR-10 and GTSRB datasets, and 6 \u00d7 6 white square for Tiny ImageNet."], "text_after_citation": ["Defense Details.", "The seven SOTA defense methods can be divided into two types based on what the defender is given.", "AC [4] and ABL #OTHEREFR assumes that the defender is given a poisoned dataset, while the remaining six defense methods assumes that the defender can acquire a subset of clean samples and a backdoored model.", "The learning rate for all methods is set to 0.01, and the batch size is set to 256.", "The threshold for ANP #OTHEREFR is set to 0.4 since we find that the recommended threshold 0.2 fails to remove backdoors."], "citing_paper_content": {"title": "Enhancing Fine-Tuning Based Backdoor Defense With Sharpness-Aware Minimization", "abstract": "Backdoor defense, which aims to detect or mitigate the effect of malicious triggers introduced by attackers, is becoming increasingly critical for machine learning security and integrity. Fine-tuning based on benign data is a natural defense to erase the backdoor effect in a backdoored model. However, recent studies show that, given limited benign data, vanilla fine-tuning has poor defense performance. In this work, we provide a deep study of fine-tuning the backdoored model from the neuron perspective and find that backdoorrelated neurons fail to escape the local minimum in the fine-tuning process. Inspired by observing that the backdoorrelated neurons often have larger norms, we propose FT-SAM, a novel backdoor defense paradigm that aims to shrink the norms of backdoor-related neurons by incorporating sharpness-aware minimization with fine-tuning. We demonstrate the effectiveness of our method on several benchmark datasets and network architectures, where it achieves stateof-the-art defense performance. Overall, our work provides a promising avenue for improving the robustness of machine learning models against backdoor attacks."}, "cited_paper_content": {"title": "Targeted Backdoor Attacks On Deep Learning Systems Using Data Poisoning", "abstract": "Deep learning models have achieved high performance on many tasks, and thus have been applied to many security-critical scenarios. For example, deep learning-based face recognition systems have been used to authenticate users to access many security-sensitive applications like payment apps. Such usages of deep learning systems provide the adversaries with sufficient incentives to perform attacks against these systems for their adversarial purposes. In this work, we consider a new type of attacks, called backdoor attacks, where the attacker's goal is to create a backdoor into a learning-based authentication system, so that he can easily circumvent the system by leveraging the backdoor. Specifically, the adversary aims at creating backdoor instances, so that the victim learning system will be misled to classify the backdoor instances as a target label specified by the adversary. In particular, we study backdoor poisoning attacks, which achieve backdoor attacks using poisoning strategies. Different from all existing work, our studied poisoning strategies can apply under a very weak threat model: (1) the adversary has no knowledge of the model and the training set used by the victim system; (2) the attacker is allowed to inject only a small amount of poisoning samples; (3) the backdoor key is hard to notice even by human beings to achieve stealthiness. We conduct evaluation to demonstrate that a backdoor adversary can inject only around 50 poisoning samples, while achieving an attack success rate of above 90%. We are also the first work to show that a data poisoning attack can create physically implementable backdoors without touching the training process. Our work demonstrates that backdoor poisoning attacks pose real threats to a learning system, and thus highlights the importance of further investigation and proposing defense strategies against them."}, "keywords": ["poisoned samples"], "citation_intent": "method"}
{"citing_id": "2303.08984v1", "cited_id": "1511.08458", "section_title": "C. The Previous Rpcnn Model", "citation": "Afterwards, we train the plots with Convolutional Neural Network, which is proved to be powerful in image-driven pattern recognition #REFR .", "text_before_citation": ["However, the problems of a high False Positive rate and unstable performance need to be addressed.", "The workflow of the RPCNN model is sketched in Fig. 4 .", "First, we take two classes of time windows either close to or far from the interlocks, where the window length is a tunable model parameter.", "Explicitly, the interlock samples are taken as sliding windows from 1 s to 15 s before interlocks.", "Second, we transform each 1-dimensional time series of the 376 channels to a 2-dimensional Recurrence Plot, which could be interpreted as a pairwise distance measure of the time series and is capable of extracting finer dynamical patterns."], "text_after_citation": ["The output is a probability score inside the range [0, 1] indicating the likelihood of a sample belonging to the positive (i.e. close to interlock) class.", "More details such as model architecture are published in #OTHEREFR .", "Typical binary classification metrics in a confusion matrix are defined and applied in our setting.", "A True Positive (TP) means an interlock sample -a sample less than 15 s before an interlock -being classified as an interlock.", "A False Positive (FP) means a stable sample -a sample at least 10 minutes away from an interlock -being mistaken as an interlock."], "citing_paper_content": {"title": "Forecasting Particle Accelerator Interruptions Using Logistic Lasso Regression", "abstract": "Unforeseen particle accelerator interruptions, also known as interlocks, lead to abrupt operational changes despite being necessary safety measures. These may result in substantial loss of beam time and perhaps even equipment damage. We propose a simple yet powerful binary classification model aiming to forecast such interruptions, in the case of the High Intensity Proton Accelerator complex at the Paul Scherrer Institut. The model is formulated as logistic regression penalized by least absolute shrinkage and selection operator, based on a statistical two sample test to distinguish between unstable and stable states of the accelerator. The primary objective for receiving alarms prior to interlocks is to allow for countermeasures and reduce beam time loss. Hence, a continuous evaluation metric is developed to measure the saved beam time in any period, given the assumption that interlocks could be circumvented by reducing the beam current. The best-performing interlock-to-stable classifier can potentially increase the beam time by around 5 min in a day. Possible instrumentation for fast adjustment of the beam current is also listed and discussed."}, "cited_paper_content": {"title": "An Introduction To Convolutional Neural Networks", "abstract": "The field of machine learning has taken a dramatic twist in recent times, with the rise of the Artificial Neural Network (ANN). These biologically inspired computational models are able to far exceed the performance of previous forms of artificial intelligence in common machine learning tasks. One of the most impressive forms of ANN architecture is that of the Convolutional Neural Network (CNN). CNNs are primarily used to solve difficult image-driven pattern recognition tasks and with their precise yet simple architecture, offers a simplified method of getting started with ANNs. ::: This document provides a brief introduction to CNNs, discussing recently published papers and newly formed techniques in developing these brilliantly fantastic image recognition models. This introduction assumes you are familiar with the fundamentals of ANNs and machine learning."}, "keywords": ["Convolutional Neural Network"], "citation_intent": "method"}
{"citing_id": "2304.01910v1", "cited_id": "1810.04805", "section_title": "Bert Finetuning", "citation": "In this section we study BERT #REFR finetuning, and use the tools developed in Section 4 to clearly differentiate the behavior of BERT-Large from BERT-Base.", "text_before_citation": [], "text_after_citation": ["For our experiment, we finetune pretrained checkpoints of both models 1,000 times each on the MRPC #OTHEREFR task.", "MRPC contains 5,801 sentence pairs labeled yes/no for whether the second sentence paraphrases the first, and has a training set of 3,668 examples, validation set of 408 examples, and test set of 1,725 examples.", "Previous works #OTHEREFR report and investigate training instability for BERT finetuning, noting particular instability for BERT-Large.", "In our experiment, we find that both BERT-Large and BERT-Base have substantial variance in validation-set performance between runs.", "BERT-Base has a standard deviation of 0.80% and BERT-Large has 2.24%, which seems to imply that both models are unstable, with BERT-Large being only somewhat moreso."], "citing_paper_content": {"title": "Calibrated Chaos: Variance Between Runs Of Neural Network Training Is Harmless And Inevitable", "abstract": "Typical neural network trainings have substantial variance in test-set performance between repeated runs, impeding hyperparameter comparison and training reproducibility. We present the following results towards understanding this variation. (1) Despite having significant variance on their test-sets, we demonstrate that standard CIFAR-10 and ImageNet trainings have very little variance in their performance on the test-distributions from which their test-sets are sampled, suggesting that variance is less of a practical issue than previously thought. (2) We present a simplifying statistical assumption which closely approximates the structure of the test-set accuracy distribution. (3) We argue that test-set variance is inevitable in the following two senses. First, we show that variance is largely caused by high sensitivity of the training process to initial conditions, rather than by specific sources of randomness like the data order and augmentations. Second, we prove that variance is unavoidable given the observation that ensembles of trained networks are well-calibrated. (4) We conduct preliminary studies of distribution-shift, fine-tuning, data augmentation and learning rate through the lens of variance between runs."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["BERT finetuning", "BERT-Base"], "citation_intent": "method"}
{"citing_id": "2303.15109v1", "cited_id": "1409.0575", "section_title": "Compatibility Of Our Methods To Other Types Of Attacks", "citation": "The results show that direction tuning attacks are perfectly Table 5 : The ASR (%) comparison on six models on ImageNet #REFR .", "text_before_citation": ["In this subsection, we combine our methods with a feature importanceaware attack (i.e., FIA #OTHEREFR ), three input transformation methods (i.e., DIM #OTHEREFR , TIM #OTHEREFR and SIM #OTHEREFR ), SGM #OTHEREFR and More Bayesian (MB) #OTHEREFR to verify the compatibility of our methods (including both direction tuning attack and network pruning method).", "Table 4 shows the ASR on six victim models when the gradient-based attacks are combined with the feature importance-aware attack (i.e., FIA #OTHEREFR ).", "The results show that our DTA and VDTA can assist FIA to achieve the greatest ASR in comparison with other gradient-based attacks (i.e., MI/NIFGSM and their variance tuning-based version).", "Besides, our DTA also spends less time consumption in comparison with VMI/VNIFGSM, in which the gradient of input in our method is computed 100 times (i.e., K\u2022T = 10\u00d710), while that in variance tuning-based gradient attacks is 200 times (i.e., N \u2022 T = 20 \u00d7 10).", "Table 5 shows the ASR on six victim models when our methods, including DTA, VDTA, and NP, are combined with three input transformation methods, respectively."], "text_after_citation": ["The gradient-based attacks are enhanced by DIM #OTHEREFR , SIM #OTHEREFR and TIM #OTHEREFR , respectively.", "The adversarial examples are generated by the surrogate model, ResNet50 #OTHEREFR . * denotes the ASR under the white-box setting. Avg. means the average value except *.", "compatible with each input transformation method, and the network pruning method can further improve the transferability of adversarial examples.", "In particular, the combination of both DTA+NP and DIM #OTHEREFR not only has the highest transferability but also has less time consumption in comparison with variance tuning-based gradient attacks.", "Note that, in this evaluation, VDTA computes the gradient of input 2000 times (i.e., K \u2022 N \u2022 T = 10 \u00d7 20 \u00d7 10)."], "citing_paper_content": {"title": "Improving The Transferability Of Adversarial Examples Via Direction Tuning", "abstract": "ing the oscillating component. By doing so, our direction tuning attack can achieve better convergence and enhance the transferability of the generated adversarial examples. In addition, a network pruning method is proposed to smooth the decision boundary, thereby further decreasing the update oscillation and enhancing the transferability of the generated adversarial examples. The experiment results on ImageNet demonstrate that the average attack success rate (ASR) of the adversarial examples generated by our method can be improved from 87.9% to 94.5% on five victim models without defenses, and from 69.1% to 76.2% on eight advanced defense methods, in comparison with that of latest gradient-based attacks."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["direction tuning attacks", "ImageNet"], "citation_intent": "result"}
{"citing_id": "2304.12294v1", "cited_id": "2003.08934", "section_title": "Nerf Decoder", "citation": "With the emitted color c and volume density \u03c3 predicted by the rendering network, novel views can be synthesized via volume rendering, which is implemented with differential ray marching as in NeRF #REFR .", "text_before_citation": ["4constructed from the encoder f \u03b8 is fed into the NeRF decoder g \u03c6 for predicting NeRF color and density, as formulated in Eq. (1). Rendering network.", "We follow prior works #OTHEREFR to construct a MLP-based rendering network.", "Similarly, we also include the texture priors by concatenating the color information sampled on all input views with the given position.", "But unlike the typical MLP decoder that processes all points on a ray independently, we further explore introducing cross-point interactions by fusing the rendered information along a ray via a Transformer.", "We adopt IBRNet's #OTHEREFR ray Transformer in our implementation for convenience. Volume rendering."], "text_after_citation": ["Specifically, to estimate the color C of a pixel, radiance needs to be accumulated across all sampled shading points on the corresponding ray that passes through the pixel,", "EQUATION", "where c i , \u03c3 i refer to the color and density of the i-th sampled 3D point on the ray.", "T i is the volume transmittance and \u03b4 i denotes the distances between adjacent points.", "K is the total number of sampled 3D points on a ray."], "citing_paper_content": {"title": "Explicit Correspondence Matching For Generalizable Neural Radiance Fields", "abstract": "We present a new generalizable NeRF method that is able to directly generalize to new unseen scenarios and perform novel view synthesis with as few as two source views. The key to our approach lies in the explicitly modeled correspondence matching information, so as to provide the geometry prior to the prediction of NeRF color and density for volume rendering. The explicit correspondence matching is quantified with the cosine similarity between image features sampled at the 2D projections of a 3D point on different views, which is able to provide reliable cues about the surface geometry. Unlike previous methods where image features are extracted independently for each view, we consider modeling the cross-view interactions via Transformer cross-attention, which greatly improves the feature matching quality. Our method achieves state-of-the-art results on different evaluation settings, with the experiments showing a strong correlation between our learned cosine feature similarity and volume density, demonstrating the effectiveness and superiority of our proposed method. Code is at https://github.com/donydchen/matchnerf."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["volume rendering"], "citation_intent": "method"}
{"citing_id": "2303.12869v1", "cited_id": "1910.10683", "section_title": "A. Fine-Tuning", "citation": "According to #REFR , a way to improve the model's performance is by increasing the number of steps in the training.", "text_before_citation": ["We fine-tune our models based on two criteria: a) Sequence Length: After analyzing in detail the outputs generated by previous works, we observed that some of the code sequences produced by the models were incomplete compared to the target ones.", "Consequently, we tokenized the training and validation sets with SentencePiece model #OTHEREFR .", "We then computed the largest sequence data, and used its length for both the inputs and the targets.", "b) Number of steps: Since we increased the length of sequences in our model, we increased the number of fine-tuning steps."], "text_after_citation": ["We apply both criteria initializing the fine-tuning from CoTexT checkpoints 2CC and 1CC, respectively.", "CoTexT-1CC is pretrained on unimodal data (only code), and CoTexT-2CC is pretrained on bimodal data (both code and natural language).", "Results of these experiments are shown in Table III ."], "citing_paper_content": {"title": "Jacotext: A Pretrained Model For Java Code-Text Generation", "abstract": "Pretrained transformer-based models have shown high performance in natural language generation task. However, a new wave of interest has surged: automatic programming language generation. This task consists of translating natural language instructions to a programming code. Despite the fact that well-known pretrained models on language generation have achieved good performance in learning programming languages, effort is still needed in automatic code generation. In this paper, we introduce JaCoText, a model based on Transformers neural network. It aims to generate java source code from natural language text. JaCoText leverages advantages of both natural language and code generation models. More specifically, we study some findings from the state of the art and use them to (1) initialize our model from powerful pretrained models, (2) explore additional pretraining on our java dataset, (3) carry out experiments combining the unimodal and bimodal data in the training, and (4) scale the input and output length during the fine-tuning of the model. Conducted experiments on CONCODE dataset show that JaCoText achieves new state-of-the-art results."}, "cited_paper_content": {"title": "Exploring The Limits Of Transfer Learning With A Unified Text-To-Text Transformer", "abstract": "Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new \"Colossal Clean Crawled Corpus\", we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code."}, "keywords": ["training"], "citation_intent": "background"}
{"citing_id": "2303.12582v1", "cited_id": "1907.05019", "section_title": "Training Settings", "citation": "This is similar to studies #REFR Team et al., ai) that have leveraged upsampling for under-represented languages during pre-training large language models in machine translation.", "text_before_citation": ["Relatively Equal Model Parameters: The XLS-R model consists of 315, 703, 690 model parameters while the Wav2Vec2.0-Large model has 315, 693, 962 -a requirement enforced in order to ensure that neither model has an edge over the other based on their size.", "Handling Class Imbalance: The AfroDigits dataset currently has very small audio samples for our focus languages, with an unequal balance of digits for each language (see Figures 5 -16) .", "In order to prevent the model from overfitting on the classes with many samples, we implemented weighted sampling #OTHEREFR .", "With weighted sampling in the data loading process, different from the normal sampling which favors the majority classes #OTHEREFR , the labels are chosen with a probability inversely proportional to their size in the training set.", "This means that at each training step and for each language, the labels with few samples are more likely to be chosen for training the model."], "text_after_citation": ["Training Setup: All audio samples were resampled to 16kHz for the finetuning experiments.", "We froze the encoders of each model and finetuned for 100 epochs.", "We used the Adam optimizer Kingma & Ba (2015), with a learning rate of 3e \u2212 5 for both models.", "We did not do any search for optimal hyperparameters but instead used the recommended settings from the authors.", "We ran our finetuning experiments with five different seeds, then we took the average over the different runs as well as the standard deviation."], "citing_paper_content": {"title": "Afrodigits: A Community-Driven Spoken Digit Dataset For African Languages", "abstract": "The advancement of speech technologies has been remarkable, yet its integration with African languages remains limited due to the scarcity of African speech corpora. To address this issue, we present AfroDigits, a minimalist, community-driven dataset of spoken digits for African languages, currently covering 38 African languages. As a demonstration of the practical applications of AfroDigits, we conduct audio digit classification experiments on six African languages [Igbo (ibo), Yoruba (yor), Rundi (run), Oshiwambo (kua), Shona (sna), and Oromo (gax)] using the Wav2Vec2.0-Large and XLS-R models. Our experiments reveal a useful insight on the effect of mixing African speech corpora during finetuning. AfroDigits is the first published audio digit dataset for African languages and we believe it will, among other things, pave the way for Afro-centric speech applications such as the recognition of telephone numbers, and street numbers. We release the dataset and platform publicly at https://huggingface.co/datasets/ chrisjay/crowd-speech-africa and https://huggingface.co/ spaces/chrisjay/afro-speech respectively."}, "cited_paper_content": {"title": "Massively Multilingual Neural Machine Translation In The Wild: Findings And Challenges", "abstract": "We introduce our efforts towards building a universal neural machine translation (NMT) system capable of translating between any language pair. We set a milestone towards this goal by building a single massively multilingual NMT model handling 103 languages trained on over 25 billion examples. Our system demonstrates effective transfer learning ability, significantly improving translation quality of low-resource languages, while keeping high-resource language translation quality on-par with competitive bilingual baselines. We provide in-depth analysis of various aspects of model building that are crucial to achieving quality and practicality in universal NMT. While we prototype a high-quality universal translation system, our extensive empirical analysis exposes issues that need to be further addressed, and we suggest directions for future research."}, "keywords": ["under-represented languages", "machine translation"], "citation_intent": "result"}
{"citing_id": "2303.10902v2", "cited_id": "1503.02531", "section_title": "Backbone Classifier", "citation": "The reason for using a soft label is that a soft label usually provides more information #REFR .", "text_before_citation": ["K", "EQUATION", "Note that p i is a soft pseudo label rather than a hard pseudo label."], "text_after_citation": ["By using the proposed test time self-distillation, the network could map the uniformity for the current samples to improve the quality of representations.", "Although we use an entropy filter to drop noisy labels when computing prototypes, there are still some mistake predictions inevitable.", "We propose that, for a reliable sample, the outputs of the linear fully connected layer and prototype-based classifier should be similar.", "Hence, we adopt the consistency filter to identify the mistake predictions.", "Specially, if the linear classifier and prototype-based classifier produce the same predictions, i.e."], "citing_paper_content": {"title": "Feature Alignment And Uniformity For Test Time Adaptation", "abstract": "Test time adaptation (TTA) aims to adapt deep neural networks when receiving out of distribution test domain samples. In this setting, the model can only access online unlabeled test samples and pre-trained models on the training domains. We first address TTA as a feature revision problem due to the domain gap between source domains and target domains. After that, we follow the two measurements alignment and uniformity to discuss the test time feature revision. For test time feature uniformity, we propose a test time self-distillation strategy to guarantee the consistency of uniformity between representations of the current batch and all the previous batches. For test time feature alignment, we propose a memorized spatial local clustering strategy to align the representations among the neighborhood samples for the upcoming batch. To deal with the common noisy label problem, we propound the entropy and consistency filters to select and drop the possible noisy labels. To prove the scalability and efficacy of our method, we conduct experiments on four domain generalization benchmarks and four medical image segmentation tasks with various backbones. Experiment results show that our method not only improves baseline stably but also outperforms existing state-of-the-art test time adaptation methods."}, "cited_paper_content": {"title": "Distilling The Knowledge In A Neural Network", "abstract": "A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel."}, "keywords": ["soft label"], "citation_intent": "background"}
{"citing_id": "2303.13528v1", "cited_id": "1605.05419", "section_title": "Introduction", "citation": "Scripting languages are well suited to such tasks because researchers can focus on high-level needs and worry less about memory management, code efficiency, and other technical details #REFR .", "text_before_citation": ["These opportunities have motivated the creation of interdisciplinary training programs, courses, workshops, and tutorials to teach computing skills in a life-sciences context 6,2,@ 7,8-12 .", "In some circumstances, it is sufficient for researchers to understand computing concepts and learn to use existing tools; in other settings, learning to write computer code is invaluable #OTHEREFR .", "A 2011 survey of scientists from many disciplines (other than computer science) found that researchers spent 35% of their time, on average, writing code #OTHEREFR .", "Computer programming makes it possible to complete tasks not supported by existing tools, interface with software libraries, adapt algorithms based on custom needs, tidy data, and more #OTHEREFR [16] #OTHEREFR .", "In these applied scenarios, computer programs are often small #OTHEREFR , and the code might only be used for one particular project."], "text_after_citation": ["Python, a scripting language, has gained much acceptance among scientists #OTHEREFR and programming educators #OTHEREFR , perhaps due to its relatively simple syntax 19 and the availability of libraries supporting common tasks #OTHEREFR .", "However, learning to program is a daunting challenge for many researchers.", "Decades of research have sought to characterize common errors and identify effective ways for novices to learn programming skills #OTHEREFR ; much remains to be discovered.", "Recent advances in artificial intelligence have shown promise for converting natural-language descriptions of programming tasks to functional code 31, #OTHEREFR .", "The first such large language models (LLMs) fine-tuned to generate code that captured widespread interest were OpenAI's Codex and DeepMind's AlphaCode #OTHEREFR ."], "citing_paper_content": {"title": "Many Bioinformatics Programming Tasks Can Be Automated With Chatgpt", "abstract": "Computer programming is a fundamental tool for life scientists, allowing them to carry out many essential research tasks. However, despite a variety of educational efforts, learning to write code can be a challenging endeavor for both researchers and students in life science disciplines. Recent advances in artificial intelligence have made it possible to translate human-language prompts to functional code, raising questions about whether these technologies can aid (or replace) life scientists' efforts to write code. Using 184 programming exercises from an introductory-bioinformatics course, we evaluated the extent to which one such model-OpenAI's ChatGPT-can successfully complete basic-to moderatelevel programming tasks. On its first attempt, ChatGPT solved 139 (75.5%) of the exercises. For the remaining exercises, we provided natural-language feedback to the model, prompting it to try different approaches. Within 7 or fewer attempts, ChatGPT solved 179 (97.3%) of the exercises. These findings have important implications for life-sciences research and education. For many programming tasks, researchers no longer need to write code from scratch. Instead, machine-learning models may produce usable solutions. Instructors may need to adapt their pedagogical approaches and assessment techniques to account for these new capabilities that are available to the general public."}, "cited_paper_content": {"title": "An Introduction To Programming For Bioscientists: A Python-Based Primer", "abstract": "Computing has revolutionized the biological sciences over the past several decades, such that virtually all contemporary research in the biosciences utilizes computer programs. The computational advances have come on many fronts, spurred by fundamental developments in hardware, software, and algorithms. These advances have influenced, and even engendered, a phenomenal array of bioscience fields, including molecular evolution and bioinformatics; genome-, proteome-, transcriptome- and metabolome-wide experimental studies; structural genomics; and atomistic simulations of cellular-scale molecular assemblies as large as ribosomes and intact viruses. In short, much of post-genomic biology is increasingly becoming a form of computational biology. The ability to design and write computer programs is among the most indispensable skills that a modern researcher can cultivate. Python has become a popular programming language in the biosciences, largely because (i) its straightforward semantics and clean syntax make it a readily accessible first language; (ii) it is expressive and well-suited to object-oriented programming, as well as other modern paradigms; and (iii) the many available libraries and third-party toolkits extend the functionality of the core language into virtually every biological domain (sequence and structure analyses, phylogenomics, workflow management systems, etc.). This primer offers a basic introduction to coding, via Python, and it includes concrete examples and exercises to illustrate the language's usage and capabilities; the main text culminates with a final project in structural bioinformatics. A suite of Supplemental Chapters is also provided. Starting with basic concepts, such as that of a 'variable', the Chapters methodically advance the reader to the point of writing a graphical user interface to compute the Hamming distance between two DNA sequences."}, "keywords": ["Scripting languages"], "citation_intent": "background"}
{"citing_id": "2304.00445v1", "cited_id": "1703.09197", "section_title": "Introduction", "citation": "West and O'Shea #REFR apply a convolutional long short term deep neural network (CLDNN), which significantly improves classification accuracy.", "text_before_citation": ["Traditional AMC can be divided into two categories: likelihood-based (LB) methods #OTHEREFR and f eature-based (FB) methods #OTHEREFR .", "However, LB methods rely on prior knowledge about channel and signal.", "FB methods select hand-crafted features, then conduct the classification using the machine learning algorithm, such as support vector machines #OTHEREFR and random forests #OTHEREFR . FB methods highly depend on expert knowledge.", "O'Shea, Corgan, and Clanc #OTHEREFR pioneer a CNN model for AMC, initiating the application of deep learning in AMC. It outperforms traditional methods that rely on manual features.", "A model based on LSTM is proposed by #OTHEREFR . Huang et al. #OTHEREFR apply GRU to classify the signals."], "text_after_citation": ["Moreover, the dual-stream structure of CNN-LSTM is proposed in #OTHEREFR to efficiently classify signals.", "It uses the information of I/Q channel and amplitude/phase to achieve better performance.", "I and Q channel are intergrated in #OTHEREFR to learn the correlations of signals in parallel. Furthermore, Huynh-The et al.", "#OTHEREFR implement different asymmetric convolution kernels and skip connections to learn spatial correlations. Liang et al.", "#OTHEREFR combine the attention mechanism and complex-valued neural network to better represent the signal."], "citing_paper_content": {"title": "Amc-Net: An Effective Network For Automatic Modulation Classification", "abstract": "Automatic modulation classification (AMC) is a crucial stage in the spectrum management, signal monitoring, and control of wireless communication systems. The accurate classification of the modulation format plays a vital role in the subsequent decoding of the transmitted data. End-to-end deep learning methods have been recently applied to AMC, outperforming traditional feature engineering techniques. However, AMC still has limitations in low signal-to-noise ratio (SNR) environments. To address the drawback, we propose a novel AMC-Net that improves recognition by denoising the input signal in the frequency domain while performing multi-scale and effective feature extraction. Experiments on two representative datasets demonstrate that our model performs better in efficiency and effectiveness than the most current methods."}, "cited_paper_content": {"title": "Deep Architectures For Modulation Recognition", "abstract": "We survey the latest advances in machine learning with deep neural networks by applying them to the task of radio modulation recognition. Results show that radio modulation recognition is not limited by network depth and further work should focus on improving learned synchronization and equalization. Advances in these areas will likely come from novel architectures designed for these tasks or through novel training methods."}, "keywords": ["deep neural network"], "citation_intent": "method"}
{"citing_id": "2303.02347v1", "cited_id": "1606.06160", "section_title": "Quantization-Aware Training (Qat)", "citation": "DoReFa-Net #REFR proposed to optimize the clipping value and the scaling factor of the uniform quantizers for weights and activations separately.", "text_before_citation": [], "text_after_citation": ["It was validated on image classification tasks under multiple bit-widths, but only with the rather simple AlexNet architecture.", "Most QAT works quantize weights and activations simultaneously by optimizing the uniform quantization parameters #OTHEREFR , layer-wise or channel-wise mixed-precision quantization #OTHEREFR , or leveraging nonuniform quantization such as Logarithmic quantizer #OTHEREFR .", "Most recent QAT works #OTHEREFR used \"Straight-Through Estimator\" (STE) #OTHEREFR to estimate the gradient of the nondifferentiable quantization function, while another work #OTHEREFR softened the linear quantization operation in order to match the true gradient with STE.", "#OTHEREFR adopted a primitive quantizer design based on a uniform quantizer for gradients (without scaling and other optimization), and large performance drops were observed when training with low-bit gradients.", "SBM #OTHEREFR adopted fixed-point 8-bit gradient quantization but only focused on improving the quantization schemes in the forward pass."], "citing_paper_content": {"title": "Metagrad: Adaptive Gradient Quantization With Hypernetworks", "abstract": "A popular track of network compression approach is Quantization aware Training (QAT), which accelerates the forward pass during the neural network training and inference. However, not much prior efforts have been made to quantize and accelerate the backward pass during training, even though that contributes around half of the training time. This can be partly attributed to the fact that errors of low-precision gradients during backward cannot be amortized by the training objective as in the QAT setting. In this work, we propose to solve this problem by incorporating the gradients into the computation graph of the next training iteration via a hypernetwork. Various experiments on CIFAR-10 dataset with different CNN network architectures demonstrate that our hypernetwork-based approach can effectively reduce the negative effect of gradient quantization noise and successfully quantizes the gradients to INT4 with only 0.64 accuracy drop for VGG-16 on CIFAR-10."}, "cited_paper_content": {"title": "Dorefa-Net: Training Low Bitwidth Convolutional Neural Networks With Low Bitwidth Gradients", "abstract": "We propose DoReFa-Net, a method to train convolutional neural networks that have low bitwidth weights and activations using low bitwidth parameter gradients. In particular, during backward pass, parameter gradients are stochastically quantized to low bitwidth numbers before being propagated to convolutional layers. As convolutions during forward/backward passes can now operate on low bitwidth weights and activations/gradients respectively, DoReFa-Net can use bit convolution kernels to accelerate both training and inference. Moreover, as bit convolutions can be efficiently implemented on CPU, FPGA, ASIC and GPU, DoReFa-Net opens the way to accelerate training of low bitwidth neural network on these hardware. Our experiments on SVHN and ImageNet datasets prove that DoReFa-Net can achieve comparable prediction accuracy as 32-bit counterparts. For example, a DoReFa-Net derived from AlexNet that has 1-bit weights, 2-bit activations, can be trained from scratch using 6-bit gradients to get 46.1\\% top-1 accuracy on ImageNet validation set. The DoReFa-Net AlexNet model is released publicly."}, "keywords": ["uniform quantizers"], "citation_intent": "background"}
{"citing_id": "2303.15404v1", "cited_id": "1608.00163", "section_title": "Introduction", "citation": "Although there is no firm answer concerning these rules #REFR , it is commonly admitted that the definition of a community relates to a difference in connection density between its interior and its boundary.", "text_before_citation": ["Nodes of these networks are often arranged in closely tight groups called communities.", "These communities delineate the organizational supports of function, property, purpose or categories.", "They thus highlight a structure of the network providing an organizational understanding behind the topology.", "Formally, the goal is to identify a node partition of the nodes of the network.", "A community structure is a partition of the vertices of a graph defined according rules structuring the vertex distribution."], "text_after_citation": ["The density of connection between nodes inside a community must be higher than the density of connection across communities.", "Such community obtained by this method is called the topological community #OTHEREFR .", "Community detection algorithms capture this difference of connection density for detecting communities in a network #OTHEREFR .", "The quality of a community structure is evaluated by a measure assessing this partitioning rule.", "A recognized standard is the modularity introduced by Newmman #OTHEREFR ."], "citing_paper_content": {"title": "Chromatic Community Structure Detection", "abstract": "The detection of community structure is probably one of the hottest trends in complex network research as it reveals the internal organization of people, molecules or processes behind social, biological or computer networks.. . The issue is to provide a network partition representative of this organization so that each community presumably gathers nodes sharing a common mission, purpose or property. Usually the identification is based on the difference between the connectivity density of the interior and the boundary of a community. Indeed, nodes sharing a common purpose or property are expected to interact closely. Although this rule appears mostly relevant, some fundamental scientific problems like disease module detection highlight the inability to determine significantly the communities under this connectivity rule. The main reason is that the connectivity density is not correlated to a shared property or purpose. Therefore, another paradigm is required for properly formalize this issue in order to meaningfully detect these communities. In this article we study the community formation from this new principle. Considering colors formally figures the shared properties, the issue is thus to maximize group of nodes with the same color within communities.. We study this novel community framework by introducing new measurement called chromarity assessing the quality of the community structure regarding this constraint. Next we propose an algorithm solving the community structure detection based on this new community formation paradigm."}, "cited_paper_content": {"title": "Community Detection In Networks: A User Guide", "abstract": "Community detection in networks is one of the most popular topics of modern network science. Communities, or clusters, are usually groups of vertices having higher probability of being connected to each other than to members of other groups, though other patterns are possible. Identifying communities is an ill-defined problem. There are no universal protocols on the fundamental ingredients, like the definition of community itself, nor on other crucial issues, like the validation of algorithms and the comparison of their performances. This has generated a number of confusions and misconceptions, which undermine the progress in the field. We offer a guided tour through the main aspects of the problem. We also point out strengths and weaknesses of popular methods, and give directions to their use."}, "keywords": ["community"], "citation_intent": "background"}
{"citing_id": "2303.06588v1", "cited_id": "1808.09781", "section_title": "Experimental Setup 5.1 Experimental Settings", "citation": "SASRec #REFR is trained with adam optimizer and a learning rate of 0.001. The number of layers and attention heads is 2.", "text_before_citation": ["For sequential recommendation baselines, we keep a consistent batch size of 4096 and the maximum interaction length to be 50.", "We employ a leave-one-out strategy for validation and testing, and the full item set is used for evaluation. We use early stopping patience of 10.", "We consider the sequential recommendation task as a multiclass classification task and use the cross-entropy loss for training the models."], "text_after_citation": ["The dropout rate is 0.5 and gelu is used as an activation function.", "HGN #OTHEREFR and SINE #OTHEREFR are trained with embedding size 64, learning rate 0.001, and adam optimizer.", "LigthSANs #OTHEREFR uses the latent interest dimension to be 5, 2 attention heads and 2 transformer layers.", "Training is done with a learning rate of 0.001 and adam optimizer.", "GCSAN #OTHEREFR has 2 transformer encoder layers, 2 attention heads, 64 is the hidden state features size, feed-forward layers' hidden size is 256, weight is set to be 0.6 and the number of layers in graph neural network is 1, adam is used for optimization with a learning rate of 0.001."], "citing_paper_content": {"title": "Mobilerec: A Large-Scale Dataset For Mobile Apps Recommendation", "abstract": "Recommender systems have become ubiquitous in our digital lives, from recommending products on e-commerce websites to suggesting movies and music on streaming platforms. Existing recommendation datasets, such as Amazon Product Reviews and MovieLens, greatly facilitated the research and development of recommender systems in their respective domains. While the number of mobile users and applications (aka apps) has increased exponentially over the past decade, research in mobile app recommender systems has been significantly constrained, primarily due to the lack of high-quality benchmark datasets, as opposed to recommendations for products, movies, and news. To facilitate research for app recommendation systems, we introduce a large-scale dataset, called MobileRec. We constructed MobileRec from users' activity on the Google play store. MobileRec contains 19.3 million user interactions (i.e., user reviews on apps) with over 10K unique apps across 48 categories. MobileRec records the sequential activity of a total of 0.7 million distinct users. Each of these users has interacted with no fewer than five distinct apps, which stands in contrast to previous datasets on mobile apps that recorded only a single interaction per user. Furthermore, MobileRec presents users' ratings as well as sentiments on installed apps, and each app contains rich metadata such as app name, category, description, and overall rating, among others. We demonstrate that MobileRec can serve as an excellent testbed for app recommendation through a comparative study of several state-of-the-art recommendation approaches. The quantitative results can act as a baseline for other researchers to compare their results against."}, "cited_paper_content": {"title": "Self-Attentive Sequential Recommendation", "abstract": "Sequential dynamics are a key feature of many modern recommender systems, which seek to capture the 'context' of users' activities on the basis of actions they have performed recently. To capture such patterns, two approaches have proliferated: Markov Chains (MCs) and Recurrent Neural Networks (RNNs). Markov Chains assume that a user's next action can be predicted on the basis of just their last (or last few) actions, while RNNs in principle allow for longer-term semantics to be uncovered. Generally speaking, MC-based methods perform best in extremely sparse datasets, where model parsimony is critical, while RNNs perform better in denser datasets where higher model complexity is affordable. The goal of our work is to balance these two goals, by proposing a self-attention based sequential model (SASRec) that allows us to capture long-term semantics (like an RNN), but, using an attention mechanism, makes its predictions based on relatively few actions (like an MC). At each time step, SASRec seeks to identify which items are 'relevant' from a user's action history, and use them to predict the next item. Extensive empirical studies show that our method outperforms various state-of-the-art sequential models (including MC/CNN/RNN-based approaches) on both sparse and dense datasets. Moreover, the model is an order of magnitude more efficient than comparable CNN/RNN-based models. Visualizations on attention weights also show how our model adaptively handles datasets with various density, and uncovers meaningful patterns in activity sequences."}, "keywords": ["layers"], "citation_intent": "method"}
{"citing_id": "2304.01823v1", "cited_id": "1602.04505", "section_title": "Contracting A Single Crossedge.", "citation": "Its proof is the same as the proof of #REFR Lemma 4 .28], which directly translates to the locally finite case.", "text_before_citation": ["If there exists N such that s \u2032 \u2208 Y \u2032 N then for all n N we must have s \u2032 \u2208 Y n by definition of .", "Up to extracting another infinite subsequence, we can assume that either s \u2032 \u2208 Y \u2032 n for all n or s \u2032 \u2208 Z \u2032 n for all n. As a result,", "((Y \u2032 n ) \u2227 , S \u2032 n , (Z \u2032 n ) \u2227 )", "n\u2208N is an infinite decreasing sequence of separations of order 3 in T , contradicting the fact that T is a well-quasi-ordered set.", "We conclude this subsection with the following result relating the degeneracy of minimal separations in G and G \u2032 ."], "text_after_citation": ["To be more precise, we also need the additional assumption that T \u2032 is a region tangle to make the proof work, which is given by Lemma 3.25.", "Lemma 3.26 (Lemma 4.28 and Corollary 4.29 in #OTHEREFR ).", "Either G \u2032 is 4-connected and T \u2032 min = {(\u2205, \u2205, V (G \u2032 ))}, or T \u2032 min = {(Y, S, Z) \u2228 : (Y, S, Z) \u2208 T min and S \u2228 is a separator of G \u2032 }.", "In the latter case, for all (Y, S, Z) \u2208 T min , (Y, S, Z) is non-degenerate if and only if (Y, S, Z) \u2228 is non-degenerate.", "3.6. Contracting all the crossedges."], "citing_paper_content": {"title": "The Structure Of Quasi-Transitive Graphs Avoiding A Minor With Applications To The Domino Problem", "abstract": "An infinite graph is quasi-transitive if its vertex set has finitely many orbits under the action of its automorphism group. In this paper we obtain a structure theorem for locally finite quasi-transitive graphs avoiding a minor, which is reminiscent of the Robertson-Seymour Graph Minor Structure Theorem. We prove that every locally finite quasi-transitive graph G avoiding a minor has a treedecomposition whose torsos are finite or planar; moreover the tree-decomposition is canonical, i.e. invariant under the action of the automorphism group of G. As applications of this result, we prove the following. \u2022 Every locally finite quasi-transitive graph attains its Hadwiger number, that is, if such a graph contains arbitrarily large clique minors, then it contains an infinite clique minor. This extends a result of Thomassen (1992) who proved it in the 4-connected case and suggested that this assumption could be omitted. In particular, this shows that a Cayley graph excludes a finite minor if and only if it avoids the countable clique as a minor. \u2022 Locally finite quasi-transitive graphs avoiding a minor are accessible (in the sense of Thomassen and Woess), which extends known results on planar graphs to any proper minor-closed family. \u2022 Minor-excluded finitely generated groups are accessible (in the group-theoretic sense) and finitely presented, which extends classical results on planar groups. \u2022 The domino problem is decidable in a minor-excluded finitely generated group if and only if the group is virtually free, which proves the minorexcluded case of a conjecture of Ballier and Stein (2018)."}, "cited_paper_content": {"title": "Quasi-4-Connected Components", "abstract": "We introduce a new decomposition of a graphs into quasi-4-connected components, where we call a graph quasi-4-connected if it is 3-connected and it only has separations of order 3 that remove a single vertex. Moreover, we give a cubic time algorithm computing the decomposition of a given graph. Our decomposition into quasi-4-connected components refines the well-known decompositions of graphs into biconnected and triconnected components. We relate our decomposition to Robertson and Seymour's theory of tangles by establishing a correspondence between the quasi-4-connected components of a graph and its tangles of order 4."}, "keywords": ["locally finite case", "Lemma"], "citation_intent": "background"}
{"citing_id": "2303.06982v1", "cited_id": "2002.05709", "section_title": "Related Works", "citation": "This is in line with other studies in the image community #REFR . Chung et al.", "text_before_citation": ["As SSL models are becoming mainstream, many researchers have tried to understand what is happening under the hood of these pre-trained models either by learning a probing tasks #OTHEREFR or analyzing the representations directly #OTHEREFR are two most popular approaches.", "The authors of #OTHEREFR study the WAV2VEC2.0 model #OTHEREFR representations directly rather than training additional classifiers as probes.", "They show that the pre-trained model follows an auto-encoder style behavior i.e., intermediate layers provide richer information of higherlevel classes (phone/word information) than the initial and last layers."], "text_after_citation": ["#OTHEREFR studies the similarity between the representations of three SSL pre-training techniques i.e., contrastive predictive coding (CPC), auto-regressive predictive coding (APC), and masked predictive coding (MPC).", "Their findings suggest that it is the learning objective which controls the representation similarity than architectural choices such as building blocks and directionality.", "Similarly, in a comprehensive review of self-supervised speech representation learning by #OTHEREFR", "(2022) #OTHEREFR , the authors posited that the choice of training criterion has a greater impact on the performance gains compared to the architecture or the directionality of input.", "Given pre-training criterion has the highest impact on the model's performance, in this paper we investigate the impact of the masked prediction loss on the encoded information type in various layers of HuBERT model #OTHEREFR ."], "citing_paper_content": {"title": "Analysing The Masked Predictive Coding Training Criterion For Pre-Training A Speech Representation Model", "abstract": "Recent developments in pre-trained speech representation utilizing self-supervised learning (SSL) have yielded exceptional results on a variety of downstream tasks. One such technique, known as masked predictive coding (MPC), has been employed by some of the most high-performing models. In this study, we investigate the impact of MPC loss on the type of information learnt at various layers in the HuBERT model, using nine probing tasks. Our findings indicate that the amount of content information learned at various layers of the HuBERT model has a positive correlation to the MPC loss. Additionally, it is also observed that any speaker-related information learned at intermediate layers of the model, is an indirect consequence of the learning process, and therefore cannot be controlled using the MPC loss. These findings may serve as inspiration for further research in the speech community, specifically in the development of new pre-training tasks or the exploration of new pre-training criterion's that directly preserves both speaker and content information at various layers of a learnt model."}, "cited_paper_content": {"title": "A Simple Framework For Contrastive Learning Of Visual Representations", "abstract": "This paper presents SimCLR: a simple framework for contrastive learning of visual representations. We simplify recently proposed contrastive self-supervised learning algorithms without requiring specialized architectures or a memory bank. In order to understand what enables the contrastive prediction tasks to learn useful representations, we systematically study the major components of our framework. We show that (1) composition of data augmentations plays a critical role in defining effective predictive tasks, (2) introducing a learnable nonlinear transformation between the representation and the contrastive loss substantially improves the quality of the learned representations, and (3) contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. By combining these findings, we are able to considerably outperform previous methods for self-supervised and semi-supervised learning on ImageNet. A linear classifier trained on self-supervised representations learned by SimCLR achieves 76.5% top-1 accuracy, which is a 7% relative improvement over previous state-of-the-art, matching the performance of a supervised ResNet-50. When fine-tuned on only 1% of the labels, we achieve 85.8% top-5 accuracy, outperforming AlexNet with 100X fewer labels."}, "keywords": ["Chung", "image community"], "citation_intent": "result"}
{"citing_id": "2304.13292v1", "cited_id": "1810.04805", "section_title": "Results And Discussions", "citation": "The larger improvement in scores for intent classification may be correlated with the fact that for our data augmentation experiment on paraphrasing and machine translation, we were only able to Table 6 : Accuracy results for intent classification on the validation set. Baseline: mBERT #REFR .", "text_before_citation": ["Our findings regarding the performance of larger models are also observed in the test set.", "Table 6 presents the evaluation results for both slot filling and intent classification tasks across all three target languages. Our mT0 models strongly outperform the baseline models.", "Specifically, our mT0 models outperformed the baseline models in all target languages for the intent classification task, highlighting the effectiveness of larger models for intent classification.", "Moreover, our mT0 models also outperform the baseline models in two of the target languages for slot filling task, further indicating the superiority of larger models for sentence-level classification tasks.", "The improvement in scores for intent classification is more evident than for slot filling."], "text_after_citation": ["LS: LASER #OTHEREFR +sBERT #OTHEREFR . LL: LASER+LaBSE #OTHEREFR . LX: LASER+XLM-R #OTHEREFR . Underline: Best-performing models for each setting.", "Bold: Best F 1 score across all the experiments and settings.", "augment data for intent classification, resulting in a larger improvement in performance for this task compared to slot filling.", "It is worth noting that we use the validation set for model selection, which resulted in higher scores than those achieved on the test set.", "This is because the validation data is similar to the data used during training, while the test data is entirely new and unseen."], "citing_paper_content": {"title": "Zero-Shot Slot And Intent Detection In Low-Resource Languages", "abstract": "Intent detection and slot filling are critical tasks in spoken and natural language understanding for task-oriented dialog systems. In this work we describe our participation in the slot and intent detection for low-resource language varieties (SID4LR; Aepli et al. (2023)). We investigate the slot and intent detection (SID) tasks using a wide range of models and settings. Given the recent success of multitaskprompted finetuning of large language models, we also test the generalization capability of the recent encoder-decoder model mT0 (Muennighoff et al., 2022) on new tasks (i.e., SID) in languages they have never intentionally seen. We show that our best model outperforms the baseline by a large margin (up to +30 F 1 points) in both SID tasks."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["intent classification", "machine translation"], "citation_intent": "result"}
{"citing_id": "2304.04544v1", "cited_id": "1306.0187", "section_title": "Image Motion Deblurring", "citation": "First we have found that MALA with subgradient clearly has the worst performance among all the methods, a finding agreeing with #REFR .", "text_before_citation": ["Next we test the algorithms with the additional Metropolis (accept-reject) step included.", "More precisely we implement the following algorithms: MALA with subgradient, the PMALA method in #OTHEREFR , a variant of PMALA with Chambolle2004 replaced by K-step Chambolle-Pock, and the proposed PDFP based algorithm denoted as MALA-PDFP.", "The results of all the methods are compared in Table 6 .3, and we reinstate that thanks to the Metropolis step, the samples are asymptotically unbiased.", "For the stability of PMALA and MALA-PDFP, step size \u03b4 should be no larger than parameter \u03c1.", "Following #OTHEREFR we fix \u03b4 = \u03c1 and the values of them (that are shown in Table 6 .3) are chosen such that the acceptance rates of all the algorithms are around 50% [31, #OTHEREFR for fair comparison."], "text_after_citation": ["Moreover, in both MALA-PDFP and PMALA-CP, we can see that the results of K = 5 are rather close to those of K = 100 and PMALA where in both cases the subproblem is solved accurately.", "Notably in Table 6 .3 the run time of MALA-PDFP for K = 100 is similar or less than that for K = 5, this is because in this experiments \u03c1 is much smaller than Table 6 .2 and the stopping criteria x n,k+1 \u2212 x n,k < 10 \u22125 is met even k < 5.", "More interestingly, however, PMALA-CP with K = 1 yields substantially worse results (in terms of ESS and ESJD) than the algorithms that solve the subprobem accurately, while MALA-PDFP with K = 1 produces results that are comparable to those.", "While this is an interesting indicator that the 1-step MALA-PDFP may be an effective and efficient sampling algorithm, further investigation and more comprehensive tests of the method are needed. Table 6 .3: Comparison of the Metropolis-adjusted Langevin algorithms."], "citing_paper_content": {"title": "Approximate Primal-Dual Fixed-Point Based Langevin Algorithms For Non-Smooth Convex Potentials *", "abstract": "The Langevin algorithms are frequently used to sample the posterior distributions in Bayesian inference. In many practical problems, however, the posterior distributions often consist of non-differentiable components, posing challenges for the standard Langevin algorithms, as they require to evaluate the gradient of the energy function in each iteration. To this end, a popular remedy is to utilize the proximity operator, and as a result one needs to solve a proximity subproblem in each iteration. The conventional practice is to solve the subproblems accurately, which can be exceedingly expensive, as the subproblem needs to be solved in each iteration. We propose an approximate primal-dual fixed-point algorithm for solving the subproblem, which only seeks an approximate solution of the subproblem and therefore reduces the computational cost considerably. We provide theoretical analysis of the proposed method and also demonstrate its performance with numerical examples."}, "cited_paper_content": {"title": "Proximal Markov Chain Monte Carlo Algorithms", "abstract": "This paper presents a new Metropolis-adjusted Langevin algorithm (MALA) that uses convex analysis to simulate efficiently from high-dimensional densities that are log-concave, a class of probability distributions that is widely used in modern high-dimensional statistics and data analysis. The method is based on a new first-order approximation for Langevin diffusions that exploits log-concavity to construct Markov chains with favourable convergence properties. This approximation is closely related to Moreau-Yoshida regularisations for convex functions and uses proximity mappings instead of gradient mappings to approximate the continuous-time process. The proposed method complements existing MALA methods in two ways. First, the method is shown to have very robust stability properties and to converge geometrically for many target densities for which other MALA are not geometric, or only if the step size is sufficiently small. Second, the method can be applied to high-dimensional target densities that are not continuously differentiable, a class of distributions that is increasingly used in image processing and machine learning and that is beyond the scope of existing MALA and HMC algorithms. To use this method it is necessary to compute or to approximate efficiently the proximity mappings of the logarithm of the target density. For several popular models, including many Bayesian models used in modern signal and image processing and machine learning, this can be achieved with convex optimisation algorithms and with approximations based on proximal splitting techniques, which can be implemented in parallel. The proposed method is demonstrated on two challenging high-dimensional and non-differentiable models related to image resolution enhancement and low-rank matrix estimation that are not well addressed by existing MCMC methodology."}, "keywords": ["subgradient"], "citation_intent": "result"}
{"citing_id": "2303.15782v1", "cited_id": "2003.08934", "section_title": "Related Work", "citation": "Recent advances in differential rendering have enabled learning of shapes, as well as other scene properties such as appearance, only from images and without the need for 3D supervision #REFR .", "text_before_citation": ["coordinate-based multi-layer perceptrons #OTHEREFR , have become a popular method for reconstruction in recent years.", "These methods encode continuous functions that model various scene properties, such as Signed Distance #OTHEREFR , radiance #OTHEREFR , and occupancy #OTHEREFR .", "Variations of these include hybrid discrete-continuous representations that employ an external data structure, i.e.", "a grid or an octree, to partition the implicit function #OTHEREFR .", "The encoded shape can then be extracted via sphere tracing #OTHEREFR after querying the implicit function repeatedly."], "text_after_citation": ["Our approach falls into the paradigm of using neural fields for articulated object reconstruction and further learns a complete system for detection, pose estimation, and articulated shape reconstruction from a single observation.", "Implicit Reconstruction of Non-Rigid Objects: Going beyond static scenes with rigid objects, #OTHEREFR handle dynamic scenes while #OTHEREFR focus on reconstructing humans by leveraging their strong shape and kinematic prior as well as the amount of readily available datasets.", "[42] propose a general reconstruction framework to reconstruct any nonrigid entity (i.e.", "humans, animals, or objects) given only an RGB-video without requiring a category-specific shape template, while #OTHEREFR focus on point cloud input data and split the prediction into a canonical shape and a deformation field.", "One downside of general reconstruction methods is that they do not leverage the rigidity and kinematic constraints of articulated objects."], "citing_paper_content": {"title": "Carto: Category And Joint Agnostic Reconstruction Of Articulated Objects", "abstract": "Figure 1. Visualization of CARTO on unseen object instances. We first use CARTO to jointly detect all objects in the scene and then articulate them while keeping the predicted shape code constant."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["3D supervision", "differential rendering"], "citation_intent": "background"}
{"citing_id": "2303.03975v1", "cited_id": "2003.07082", "section_title": "Corpus Development Process", "citation": "For example, we used Stanza #REFR to filter some web-scraped data for those containing an animate noun marked as an indirect object and provided this to the linguists.", "text_before_citation": ["\u2022 1,000 single animate noun AGME \u2022 500 single pronoun AGME \u2022 500 with two or more AGMEs Linguists were given details of the various categories and attributes listed in section A and asked to find sentences such that each such category is well represented (depending on the relative ease of finding such sentences).", "Linguists were also asked to prioritize diversity of animate nouns where possible.", "They were allowed to pull examples sentences from natural text or construct them from scratch as they saw fit.", "However, except for a small number of toy examples, we asked that they include only sentences that were natural in both English and their target language, and could reasonably appear in some imaginable context.", "We provided samples of web-scraped data that had been filtered with various heuristics to help identify sentences fitting some of the harder-tosatisfy criteria."], "text_after_citation": ["In some cases these sentences were used directly, and in others they were modified slightly to fit the requirements.", "Throughout the process, we prioritized diversity of sentence structure, domain and vocabulary.", "Rather than produce a representative sample, our intention was to produce a corpus that would chal-lenge any tested systems on a wide range of phenomena.", "3 Evaluation with GATE"], "citing_paper_content": {"title": "Gate: A Challenge Set For Gender-Ambiguous Translation Examples", "abstract": "Although recent years have brought significant progress in improving translation of unambiguously gendered sentences, translation of ambiguously gendered input remains relatively unexplored. When source gender is ambiguous, machine translation models typically default to stereotypical gender roles, perpetuating harmful bias. Recent work has led to the development of \"gender rewriters\" that generate alternative gender translations on such ambiguous inputs, but such systems are plagued by poor linguistic coverage. To encourage better performance on this task we present and release GATE, a linguistically diverse corpus of gender-ambiguous source sentences along with multiple alternative target language translations. We also provide tools for evaluation and system analysis when using GATE and use them to evaluate our translation rewriter system."}, "cited_paper_content": {"title": "Stanza: A Python Natural Language Processing Toolkit For Many Human Languages", "abstract": "We introduce Stanza, an open-source Python natural language processing toolkit supporting 66 human languages. Compared to existing widely used toolkits, Stanza features a language-agnostic fully neural pipeline for text analysis, including tokenization, multi-word token expansion, lemmatization, part-of-speech and morphological feature tagging, dependency parsing, and named entity recognition. We have trained Stanza on a total of 112 datasets, including the Universal Dependencies treebanks and other multilingual corpora, and show that the same neural architecture generalizes well and achieves competitive performance on all languages tested. Additionally, Stanza includes a native Python interface to the widely used Java Stanford CoreNLP software, which further extends its functionalities to cover other tasks such as coreference resolution and relation extraction. Source code, documentation, and pretrained models for 66 languages are available at https://stanfordnlp.github.io/stanza."}, "keywords": ["linguists"], "citation_intent": "method"}
{"citing_id": "2303.09105v1", "cited_id": "1710.06081", "section_title": "Cosine Similarity Encourager", "citation": "Although directly applying this algorithm can achieve good results, it is incompatible with SAM due to the varying scales of gradient norm #REFR .", "text_before_citation": ["where", "x 0 t = x t .", "Once the update for every model is complete, we calculate the final update using a larger step size \u03b1 as", "x t+1 = clip xnat, (x t + \u03b1 \u2022 (x n t \u2212 x t ))", "."], "text_after_citation": ["To solve this problem, we normalize the gradient at each update by their 2 norm.", "We discover that the modified version actually maximizes the cosine similarity between gradients (proof in Appendix E).", "Thus, we call it Cosine Similarity Encourager (CSE), which can be further combined with MI as MI-CSE.", "MI-CSE involves an inner momentum term to accumulate the gradients of each model. We provide the pseudocode in Appendix E."], "citing_paper_content": {"title": "Rethinking Model Ensemble In Transfer-Based Adversarial Attacks", "abstract": "Deep learning models are vulnerable to adversarial examples. Transfer-based adversarial attacks attract tremendous attention as they can identify the weaknesses of deep learning models in a black-box manner. An effective strategy to improve the transferability of adversarial examples is attacking an ensemble of models. However, previous works simply average the outputs of different models, lacking an in-depth analysis on how and why model ensemble can strongly improve the transferability. In this work, we rethink the ensemble in adversarial attacks and define the common weakness of model ensemble with the properties of the flatness of loss landscape and the closeness to the local optimum of each model. We empirically and theoretically show that these two properties are strongly correlated with the transferability and propose a Common Weakness Attack (CWA) to generate more transferable adversarial examples by promoting these two properties. Experimental results on both image classification and object detection tasks validate the effectiveness of our approach to improve the adversarial transferability, especially when attacking adversarially trained models."}, "cited_paper_content": {"title": "Boosting Adversarial Attacks With Momentum", "abstract": "Deep neural networks are vulnerable to adversarial examples, which poses security concerns on these algorithms due to the potentially severe consequences. Adversarial attacks serve as an important surrogate to evaluate the robustness of deep learning models before they are deployed. However, most of existing adversarial attacks can only fool a black-box model with a low success rate. To address this issue, we propose a broad class of momentum-based iterative algorithms to boost adversarial attacks. By integrating the momentum term into the iterative process for attacks, our methods can stabilize update directions and escape from poor local maxima during the iterations, resulting in more transferable adversarial examples. To further improve the success rates for black-box attacks, we apply momentum iterative algorithms to an ensemble of models, and show that the adversarially trained models with a strong defense ability are also vulnerable to our black-box attacks. We hope that the proposed methods will serve as a benchmark for evaluating the robustness of various deep models and defense methods. With this method, we won the first places in NIPS 2017 Non-targeted Adversarial Attack and Targeted Adversarial Attack competitions."}, "keywords": ["good results", "gradient norm"], "citation_intent": "method"}
{"citing_id": "2303.04222v1", "cited_id": "2004.01670", "section_title": "Limitations", "citation": "Class Imbalance Our dataset is more balanced on the binary label (24% sexism) than many previous datasets for hate and abuse detection #REFR .", "text_before_citation": [], "text_after_citation": ["However, there is substantial imbalance in categories and vectors of sexism.", "Thus, it is hard to confirm whether confusion between categories and vectors is due to inherent features of the data or due to the class imbalance.", "We encourage future work to examine the effect on performance and cross-vector confusion when balancing the dataset.", "That said, we explicitly made the decision to not re-balance our dataset because different types of sexism (especially at the finegrained level) do have very different base rates 'in the wild'.", "Dealing with class imbalance is then part-and-parcel of the problem we seek to address."], "citing_paper_content": {"title": "Semeval-2023 Task 10: Explainable Detection Of Online Sexism", "abstract": "Online sexism is a widespread and harmful phenomenon. Automated tools can assist the detection of sexism at scale. Binary detection, however, disregards the diversity of sexist content, and fails to provide clear explanations for why something is sexist. To address this issue, we introduce SemEval Task 10 on the Explainable Detection of Online Sexism (EDOS). We make three main contributions: i) a novel hierarchical taxonomy of sexist content, which includes granular vectors of sexism to aid explainability; ii) a new dataset of 20,000 social media comments with fine-grained labels, along with larger unlabelled datasets for model adaptation; and iii) baseline models as well as an analysis of the methods, results and errors for participant submissions to our task. Content warning: We show illustrative examples of sexist language to describe the taxonomy and analyse error types."}, "cited_paper_content": {"title": "Directions In Abusive Language Training Data: Garbage In, Garbage Out", "abstract": "Data-driven analysis and detection of abusive online content covers many different tasks, phenomena, contexts, and methodologies. This paper systematically reviews abusive language dataset creation and content in conjunction with an open website for cataloguing abusive language data. This collection of knowledge leads to a synthesis providing evidence-based recommendations for practitioners working with this complex and highly diverse data."}, "keywords": ["24% sexism", "abuse detection"], "citation_intent": "result"}
{"citing_id": "2304.11104v1", "cited_id": "1811.04551", "section_title": "World Models", "citation": "DreamerV2 is composed of the following components: an image encoder \u223c ( | , \u210e ) that learns a posterior latent representation conditional on the current observation and recurrent state \u210e , the recurrent state space model (RSSM) #REFR which is a mixture of deterministic and stochastic categorical latents, and the image, reward and discount predictors.", "text_before_citation": ["To learn a world model for behaviour learning and look-ahead shielding we leverage DreamerV2 #OTHEREFR , which was used to master Atari games in the ALE #OTHEREFR ."], "text_after_citation": ["The RSSM consists of two main components: the recurrent model", "\u210e = (\u210e \u22121 , \u22121 , \u22121 )", ", which computes the next deterministic latents given the past state \u22121 = (\u210e \u22121 , \u22121 ) and action \u22121 , and the transition predictor\u02c6\u223c (\u02c6| \u210e ), which is used as the prior distribution over the stochastic latents conditional on the deterministic latents.", "The image predictor or decoder\u02c6\u223c (\u02c6| \u210e , ) is trained to predict the current observation with a reconstruction loss.", "The image predictor provides useful self-supervised gradients that help the world model learn a structured latent space for effective policy optimisation #OTHEREFR ."], "citing_paper_content": {"title": "Approximate Shielding Of Atari Agents For Safe Exploration", "abstract": "Balancing exploration and conservatism in the constrained setting is an important problem if we are to use reinforcement learning for meaningful tasks in the real world. In this paper, we propose a principled algorithm for safe exploration based on the concept of shielding. Previous approaches to shielding assume access to a safety-relevant abstraction of the environment or a high-fidelity simulator. Instead, our work is based on latent shielding-another approach that leverages world models to verify policy roll-outs in the latent space of a learned dynamics model. Our novel algorithm builds on this previous work, using safety critics and other additional features to improve the stability and farsightedness of the algorithm. We demonstrate the effectiveness of our approach by running experiments on a small set of Atari games with state dependent safety labels. We present preliminary results that show our approximate shielding algorithm effectively reduces the rate of safety violations, and in some cases improves the speed of convergence and quality of the final agent."}, "cited_paper_content": {"title": "Learning Latent Dynamics For Planning From Pixels", "abstract": "Planning has been very successful for control tasks with known environment dynamics. To leverage planning in unknown environments, the agent needs to learn the dynamics from interactions with the world. However, learning dynamics models that are accurate enough for planning has been a long-standing challenge, especially in image-based domains. We propose the Deep Planning Network (PlaNet), a purely model-based agent that learns the environment dynamics from images and chooses actions through fast online planning in latent space. To achieve high performance, the dynamics model must accurately predict the rewards ahead for multiple time steps. We approach this using a latent dynamics model with both deterministic and stochastic transition components. Moreover, we propose a multi-step variational inference objective that we name latent overshooting. Using only pixel observations, our agent solves continuous control tasks with contact dynamics, partial observability, and sparse rewards, which exceed the difficulty of tasks that were previously solved by planning with learned models. PlaNet uses substantially fewer episodes and reaches final performance close to and sometimes higher than strong model-free algorithms."}, "keywords": ["stochastic categorical latents", "posterior latent representation"], "citation_intent": "method"}
{"citing_id": "2303.14365v1", "cited_id": "1308.6027", "section_title": "Introduction", "citation": "The total variation of a function is defined as |Du|(\u2126) = sup{ \u2126 u\u2207 \u2022 gdx : g \u2208 (C 1 0 (\u2126)) #REFR and |g(x)| \u2264 1 in \u2126}, which is not differentiable as L 2 and H 1 -norm.", "text_before_citation": ["Usually, the unknown conductivity \u03c3 is not smooth and does not lie in H 1 0 (\u2126 C ).", "In this situation, the H 1 -regularization is not suitable.", "For reconstructing nonsmooth parameter, it is known that total variation regularization is a good choice.", "Comparing with H 1 regularization, total regularization can deal with non-smooth parameters.", "Here we give some reference about the total-variation regularization, such as #OTHEREFR ."], "text_after_citation": ["This brings us difficulty in solving the optimization problem by iterative method based on gradient.", "One way to deal with this difficulty is replacing T V term |Du| by its smoothness |\u2207u| 2 + \u03bd 2 for small \u03bd > 0 #OTHEREFR .", "Another way is employing the splitting method #OTHEREFR .", "To deal with the discontinuity of the parameter \u03c3, we introduce the total variation regularization to the inverse eddy current problem.", "In the present work, we treat the T V regularization with splitting method."], "citing_paper_content": {"title": "Iterative Methods For An Inverse Eddy Current Problem With Total Variation Regularization", "abstract": "Conductivity reconstruction in an inverse eddy current problem is considered in the present paper. With the electric field measurement on part of domain boundary, we formulate the reconstruction problem to a constrained optimization problem with total variation regularization. Existence and stability are proved for the solution to the optimization problem. The finite element method is employed to discretize the optimization problem. The gradient Lipschitz properties of the objective functional are established for the the discrete optimization problems. We propose the alternating direction method of multipliers to solve the discrete problem. Based on the the gradient Lipschitz property, we prove the convergence by extending the admissible set to the whole finite element space. Finally, we show some numerical experiments to illustrate the efficiency of the proposed methods."}, "cited_paper_content": {"title": "Detection And Classification From Electromagnetic Induction Data", "abstract": "In this paper we introduce an efficient algorithm for identifying conductive objects using induction data derived from eddy currents. Our method consists of first extracting geometric features from the induction data and then matching them to precomputed data for known objects from a given dictionary. The matching step relies on fundamental properties of conductive polarization tensors and new invariants introduced in this paper. A new shape identification scheme is introduced and studied. We test it numerically in the presence of measurement noise. Stability and resolution capabilities of the proposed identification algorithm are quantified in numerical simulations."}, "keywords": ["H 1 -norm"], "citation_intent": "background"}
{"citing_id": "2303.16133v1", "cited_id": "1405.0312", "section_title": "A. Cococon Categories & Examples", "citation": "Since the COCOCON is created from COCO Captions #REFR , the original captions predominantly contain animals defined in COCO objects.", "text_before_citation": ["Adjectives used as modifiers to a noun in the caption, such as color (red chair), height (tall building), size (small), and material (tiled wall).", "\u2022 Food. Food-related concepts including fruits, vegetables, and cooked items.", "\u2022 Sport.", "Sport-related objects and references to the sport itself are included in this category.", "\u2022 Animal. Includes all mentions of animals."], "text_after_citation": ["\u2022 Location.", "Includes broadly defined areas (e.g., bathroom, hotel, library), finer visual elements (e.g., floor, sidewalk), and spatial references (e.g., inside, outside, on table)", "\u2022 Action. Comprises transitive (e.g. flying kite) as well as intransitive actions (e.g. sitting, standing) performed by persons and animals.", "\u2022 Person.", "Concepts from one of the categories: man/male/guy, woman/female/lady, boy, girl."], "citing_paper_content": {"title": "Exposing And Addressing Cross-Task Inconsistency In Unified Vision-Language Models", "abstract": "As general purpose vision models get increasingly effective at a wide set of tasks, it is imperative that they be consistent across the tasks they support. Inconsistent AI models are considered brittle and untrustworthy by human users and are more challenging to incorporate into larger systems that take dependencies on their outputs. Measuring consistency between very heterogeneous tasks that might include outputs in different modalities is challenging since it is difficult to determine if the predictions are consistent with one another. As a solution, we introduce a benchmark dataset, COCOCON, where we use contrast sets created by modifying test instances for multiple tasks in small but semantically meaningful ways to change the gold label, and outline metrics for measuring if a model is consistent by ranking the original and perturbed instances across tasks. We find that state-of-the-art systems suffer from a surprisingly high degree of inconsistent behavior across tasks, especially for more heterogeneous tasks. Finally, we propose using a rank correlation-based auxiliary objective computed over large automatically created cross-task contrast sets to improve the multi-task consistency of large unified models, while retaining their original accuracy on downstream tasks."}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["COCO Captions", "COCO objects"], "citation_intent": "background"}
{"citing_id": "2303.15833v1", "cited_id": "2002.08546", "section_title": "Codag: Complementary Domain Adaptation And Generalization", "citation": "To apply this approach to the SHOT algorithm #REFR with the current target domain D t , we first initialize the DA model f DA with the parameters of the DG model trained with previously experienced domains, or \u03b8 * DG,t\u22121 , treating it as a source model.", "text_before_citation": ["where \u03b8 DG,0 is the DG model parameters, L ce is the crossentropy loss for a classification setting, and R represents the RandMix augmentation #OTHEREFR .", "Generalized initialization with DG for DA In general, unsupervised source-free domain adaptation approaches involve initially training a source model with the source domain data and further updating the source model with unlabeled data from a new target domain.", "However, in our proposed framework, the DA model utilizes the parameters of the previous DG model for its initialization.", "This allows the DA model to leverage the DG model's generalization ability to learn domain-invariant features and reduce domainspecific factors.", "As a result, we achieve efficient adaptation to a new target domain, even when there is a large gap between previously experienced domains and the new target domain (see Section 4.2 for experimental results)."], "text_after_citation": ["Then, we freeze the classifier head in f DA and only update the feature extractor part of it using information maximization and self-supervised pseudolabeling with data from the current target domain D t .", "Accordingly, the loss to adapt to D t is written as,", "EQUATION", "where \u03b8 DA,t is the parameters of the DA model initialized with the optimal parameters of the DG model trained on the previous domain D t\u22121 , or \u03b8 * DG,t\u22121 , and L shot is the loss of the SHOT algorithm.", "Pseudo-label generation with DA for DG In our proposed framework, we simply use an empirical risk minimization (ERM) method along with an enhanced data augmentation method for training the DG model."], "citing_paper_content": {"title": "Complementary Domain Adaptation And Generalization For Unsupervised Continual Domain Shift Learning", "abstract": "Continual domain shift poses a significant challenge in real-world applications, particularly in situations where labeled data is not available for new domains. The challenge of acquiring knowledge in this problem setting is referred to as unsupervised continual domain shift learning. Existing methods for domain adaptation and generalization have limitations in addressing this issue, as they focus either on adapting to a specific domain or generalizing to unseen domains, but not both. In this paper, we propose Complementary Domain Adaptation and Generalization (CoDAG), a simple yet effective learning framework that combines domain adaptation and generalization in a complementary manner to achieve three major goals of unsupervised continual domain shift learning: adapting to a current domain, generalizing to unseen domains, and preventing forgetting of previously seen domains. Our approach is modelagnostic, meaning that it is compatible with any existing domain adaptation and generalization algorithms. We evaluate CoDAG on several benchmark datasets and demonstrate that our model outperforms state-of-the-art models in all datasets and evaluation metrics, highlighting its effectiveness and robustness in handling unsupervised continual domain shift learning."}, "cited_paper_content": {"title": "Do We Really Need To Access The Source Data? Source Hypothesis Transfer For Unsupervised Domain Adaptation", "abstract": "Unsupervised domain adaptation (UDA) aims to leverage the knowledge learned from a labeled source dataset to solve similar tasks in a new unlabeled domain. Prior UDA methods typically require to access the source data when learning to adapt the model, making them risky and inefficient for decentralized private data. In this work we tackle a novel setting where only a trained source model is available and investigate how we can effectively utilize such a model without source data to solve UDA problems. To this end, we propose a simple yet generic representation learning framework, named \\emph{Source HypOthesis Transfer} (SHOT). Specifically, SHOT freezes the classifier module (hypothesis) of the source model and learns the target-specific feature extraction module by exploiting both information maximization and self-supervised pseudo-labeling to implicitly align representations from the target domains to the source hypothesis. In this way, the learned target model can directly predict the labels of target data. We further investigate several techniques to refine the network architecture to parameterize the source model for better transfer performance. To verify its versatility, we evaluate SHOT in a variety of adaptation cases including closed-set, partial-set, and open-set domain adaptation. Experiments indicate that SHOT yields state-of-the-art results among multiple domain adaptation benchmarks."}, "keywords": ["previously experienced domains", "current target domain"], "citation_intent": "method"}
{"citing_id": "2304.04527v1", "cited_id": "1602.01783", "section_title": "A. Convergence Speed", "citation": "Further, we also visualize the different components of the QoE metric from equation #REFR to understand how ALISA performs better than other ABR algorithms.", "text_before_citation": ["We have also compared ALISA with Pensieve, an RL-based ALISA achieves a higher QoE on all metrics over all different configurations.", "ALISA obtains up to 25% higher QoE than RB, 230% higher QoE than BB, 30% higher QoE than BOLA, 25% higher QoE than RobustMPC and 20% higher QoE compared to Pensieve when tested under lossless conditions. This performance translates to lossy conditions as well.", "We note that ALISA is able to obtain up to 25%, 28%, 48%, and 48% higher QoE compared to Pensieve under losses of 0.1%, 0.5%, 1%, and 2%, respectively.", "We summarize the remained of our testing QoE metrics for a random packet loss percentage of 0.1%, 0.5%, 1%, and 2% in Table III, Table IV, Table V  and Table VI , respectively.", "These results indicate that ALISA achieves a significantly better performance than many other fixed-rule-based ABR algorithms and also Pensieve."], "text_after_citation": ["Figure 5 presents the total reward achieved by various ABR algorithms with QoE lin metric for each trace when the network is emulated with 0.1% packet loss.", "Our results show that the ALISA algorithm achieves a higher average QoE of 44.58 as compared to other ABR algorithms.", "Figure 6 presents the average total reward achieved by various ABR algorithms with QoE lin metric for each trace when the network is emulated with 1% packet loss.", "Our results show that the ALISA algorithm achieves a higher average QoE of 34.87 as compared to other ABR algorithms.", "Figure 7 (a) and Figure 7 (b) shows how ALISA can consistently achieve higher bitrates than other methods for random sample traces. This increases the first component of QoE."], "citing_paper_content": {"title": "Deep Reinforcement Learning With Importance Weighted A3C For Qoe Enhancement In Video Delivery Services", "abstract": "Adaptive bitrate (ABR) algorithms are used to adapt the video bitrate based on the network conditions to improve the overall video quality of experience (QoE). Recently, reinforcement learning (RL) and asynchronous advantage actor-critic (A3C) methods have been used to generate adaptive bit rate algorithms and they have been shown to improve the overall QoE as compared to fixed rule ABR algorithms. However, a common issue in the A3C methods is the lag between behaviour policy and target policy. As a result, the behaviour and the target policies are no longer synchronized which results in suboptimal updates. In this work, we present ALISA: An Actor-Learner Architecture with Importance Sampling for efficient learning in ABR algorithms. ALISA incorporates importance sampling weights to give more weightage to relevant experience to address the lag issues with the existing A3C methods. We present the design and implementation of ALISA, and compare its performance to state-of-the-art video rate adaptation algorithms including vanilla A3C implemented in the Pensieve framework and other fixed-rule schedulers like BB, BOLA, and RB. Our results show that ALISA improves average QoE by up to 25%-48% higher average QoE than Pensieve, and even more when compared to fixed-rule schedulers."}, "cited_paper_content": {"title": "Asynchronous Methods For Deep Reinforcement Learning", "abstract": "We propose a conceptually simple and lightweight framework for deep reinforcement learning that uses asynchronous gradient descent for optimization of deep neural network controllers. We present asynchronous variants of four standard reinforcement learning algorithms and show that parallel actor-learners have a stabilizing effect on training allowing all four methods to successfully train neural network controllers. The best performing method, an asynchronous variant of actor-critic, surpasses the current state-of-the-art on the Atari domain while training for half the time on a single multi-core CPU instead of a GPU. Furthermore, we show that asynchronous actor-critic succeeds on a wide variety of continuous motor control problems as well as on a new task of navigating random 3D mazes using a visual input."}, "keywords": ["QoE metric", "different components"], "citation_intent": "method"}
{"citing_id": "2303.15946v1", "cited_id": "1205.2618", "section_title": "Baselines", "citation": "To demonstrate the benefit of our approach we compare it against the following baselines: BPRMF #REFR Matrix factorisation optimised by the BPR loss function.", "text_before_citation": [], "text_after_citation": ["iALS #OTHEREFR matrix factorization learned by implicit alternating least squares.", "PureSVD #OTHEREFR Compute item embeddings through a singular value decomposition of the user-item interaction matrix, which will be then used to infer user representations.", "FISM #OTHEREFR Learn item embeddings through optimisation process creating user representations as a weighted combination of items in their profile.", "Additional user and item biases as well as an agreement term are considered in the score estimation.", "NGCF #OTHEREFR Work that introduces graph convolution to the collaborative filtering scenario, it uses dense layer and inner product to enrich the knowledge injected in the user item embeddings during the convolution process."], "citing_paper_content": {"title": "Item Graph Convolution Collaborative Filtering For Inductive Recommendations", "abstract": "Graph Convolutional Networks (GCN) have been recently employed as core component in the construction of recommender system algorithms, interpreting user-item interactions as the edges of a bipartite graph. However, in the absence of side information, the majority of existing models adopt an approach of randomly initialising the user embeddings and optimising them throughout the training process. This strategy makes these algorithms inherently transductive, curtailing their ability to generate predictions for users that were unseen at training time. To address this issue, we propose a convolution-based algorithm, which is inductive from the user perspective, while at the same time, depending only on implicit user-item interaction data. We propose the construction of an item-item graph through a weighted projection of the bipartite interaction network and to employ convolution to inject higher order associations into item embeddings, while constructing user representations as weighted sums of the items with which they have interacted. Despite not training individual embeddings for each user our approach achieves state-of-the-art recommendation performance with respect to transductive baselines on four real-world datasets, showing at the same time robust inductive performance."}, "cited_paper_content": {"title": "Bpr: Bayesian Personalized Ranking From Implicit Feedback", "abstract": "Item recommendation is the task of predicting a personalized ranking on a set of items (e.g. websites, movies, products). In this paper, we investigate the most common scenario with implicit feedback (e.g. clicks, purchases). There are many methods for item recommendation from implicit feedback like matrix factorization (MF) or adaptive k-nearest-neighbor (kNN). Even though these methods are designed for the item prediction task of personalized ranking, none of them is directly optimized for ranking. In this paper we present a generic optimization criterion BPR-Opt for personalized ranking that is the maximum posterior estimator derived from a Bayesian analysis of the problem. We also provide a generic learning algorithm for optimizing models with respect to BPR-Opt. The learning method is based on stochastic gradient descent with bootstrap sampling. We show how to apply our method to two state-of-the-art recommender models: matrix factorization and adaptive kNN. Our experiments indicate that for the task of personalized ranking our optimization method outperforms the standard learning techniques for MF and kNN. The results show the importance of optimizing models for the right criterion."}, "keywords": ["BPRMF Matrix factorisation"], "citation_intent": "method"}
{"citing_id": "2304.01203v2", "cited_id": "1812.05905", "section_title": "C.3. Online Gcrl", "citation": "For the adaptive entropy regularizer #REFR , we regularize policy to have target entropy \u2212|A|, where the entropy regularizer weight is initialized to be 1 and optimized in log-space with a learning rate of 3 \u00d7 10 \u22124 .", "text_before_citation": ["For d \u03b8 , we use a 128-512-2048 projector followed by an IQE-maxmean head with 64 components, each of size 32.", "We use x-512-512-8 network for policy, where x is the input size and 8 parametrizes a tanh-transformed diagonal Normal distribution. L transition is optimized with a weight of 0.1.", "Our learning rates are 0.01 for \u03bb, 1 \u00d7 10 \u22124 for the model parameters, and 3 \u00d7 10 \u22125 for the policy parameters. We use a batch size of 256 in training.", "We prefill the replay buffer with 200 episodes from a random actor, and then iteratively perform (1) generating 10 rollouts and (2) optimizing QRL objective for 500 gradients steps.", "We use N (0, 0.3 2 )-perturbed action noise in exploration."], "text_after_citation": ["Since the environment has much shorter horizon (each episodes ends at 50 timesteps), we instead use a different affine-transformed softplus for maximizing d \u03b8 , where \u03c6(x) \u2212softplus(15 \u2212 x, \u03b2 = 0.1).", "QRL (Image-based Observations).", "All settings are the same as QRL for state-based observations, except a few changes:", "\u2022 We use the convolutional backbone followed by a x-512-128 network for encoder f .", "\u2022 We optimize L transition with an increased weight of 10 (since the dynamics aren't fully deterministic)."], "citing_paper_content": {"title": "Optimal Goal-Reaching Reinforcement Learning Via Quasimetric Learning", "abstract": "In goal-reaching reinforcement learning (RL), the optimal value function has a particular geometry, called quasimetric structure. This paper introduces Quasimetric Reinforcement Learning (QRL), a new RL method that utilizes quasimetric models to learn optimal value functions. Distinct from prior approaches, the QRL objective is specifically designed for quasimetrics, and provides strong theoretical recovery guarantees. Empirically, we conduct thorough analyses on a discretized MountainCar environment, identifying properties of QRL and its advantages over alternatives. On offline and online goal-reaching benchmarks, QRL also demonstrates improved sample efficiency and performance, across both state-based and image-based observations."}, "cited_paper_content": {"title": "Soft Actor-Critic Algorithms And Applications", "abstract": "Model-free deep reinforcement learning (RL) algorithms have been successfully applied to a range of challenging sequential decision making and control tasks. However, these methods typically suffer from two major challenges: high sample complexity and brittleness to hyperparameters. Both of these challenges limit the applicability of such methods to real-world domains. In this paper, we describe Soft Actor-Critic (SAC), our recently introduced off-policy actor-critic algorithm based on the maximum entropy RL framework. In this framework, the actor aims to simultaneously maximize expected return and entropy. That is, to succeed at the task while acting as randomly as possible. We extend SAC to incorporate a number of modifications that accelerate training and improve stability with respect to the hyperparameters, including a constrained formulation that automatically tunes the temperature hyperparameter. We systematically evaluate SAC on a range of benchmark tasks, as well as real-world challenging tasks such as locomotion for a quadrupedal robot and robotic manipulation with a dexterous hand. With these improvements, SAC achieves state-of-the-art performance, outperforming prior on-policy and off-policy methods in sample-efficiency and asymptotic performance. Furthermore, we demonstrate that, in contrast to other off-policy algorithms, our approach is very stable, achieving similar performance across different random seeds. These results suggest that SAC is a promising candidate for learning in real-world robotics tasks."}, "keywords": ["learning rate", "adaptive entropy regularizer"], "citation_intent": "method"}
{"citing_id": "2304.13812v1", "cited_id": "1804.07802", "section_title": "I. Introduction", "citation": "Significant research has also been done on quantization-aware training methods #REFR , where the loss in accuracy due to bit precision reduction is minimized.", "text_before_citation": ["Quantization aims to shrink the memory footprint of deep neural networks by reducing the number of bits used to store the values for the learnable parameters and activations.", "This is not only ideal for application scenarios where memory resources may be restricted such as embedded systems or microcontroller environments, but with the selected weight representation you could potentially facilitate faster inference using cheaper arithmetic operations #OTHEREFR .", "With the reduction in parameter bit precision, however, it is typical that a quantized neural network will perform worse in terms of accuracy than its non-quantized counterpart using gradient-based learning methods.", "However, these drops in accuracy are usually considered minimal and worth the given benefit in memory reduction and inference speed-up.", "There exists much literature describing various techniques of quantization and successful results thereof, including works utilizing stochastic rounding to select weight values beneficial to gradient training #OTHEREFR and applications on modern deep architectures #OTHEREFR , as well as quantization methods that reduce the number of multiplication operations required during training time #OTHEREFR ."], "text_after_citation": ["Some of these quantization-aware training methods utilize a straightthrough gradient estimator (STE) #OTHEREFR to more appropriately select weights during network training that minimizes accuracy and further reduces computational burden #OTHEREFR .", "As quantization methods are used for neural network reduction, there inevitably exist discrepancies between the performances of original and compressed neural networks.", "In this work, we propose a computationally tractable approach to compute the guaranteed output error caused by quantization.", "A merged neural network is constructed to generate the output differences between two neural networks, and then reachability analysis on the merged neural network can be performed to obtain the guaranteed error.", "The remainder of the paper is organized as follows: Preliminaries are given in Section II."], "citing_paper_content": {"title": "Guaranteed Quantization Error Computation For Neural Network Model Compression", "abstract": "Neural network model compression techniques can address the computation issue of deep neural networks on embedded devices in industrial systems. The guaranteed output error computation problem for neural network compression with quantization is addressed in this paper. A merged neural network is built from a feedforward neural network and its quantized version to produce the exact output difference between two neural networks. Then, optimization-based methods and reachability analysis methods are applied to the merged neural network to compute the guaranteed quantization error. Finally, a numerical example is proposed to validate the applicability and effectiveness of the proposed approach."}, "cited_paper_content": {"title": "Value-Aware Quantization For Training And Inference Of Neural Networks", "abstract": "We propose a novel value-aware quantization which applies aggressively reduced precision to the majority of data while separately handling a small amount of large data in high precision, which reduces total quantization errors under very low precision. We present new techniques to apply the proposed quantization to training and inference. The experiments show that our method with 3-bit activations (with 2% of large ones) can give the same training accuracy as full-precision one while offering significant (41.6% and 53.7%) reductions in the memory cost of activations in ResNet-152 and Inception-v3 compared with the state-of-the-art method. Our experiments also show that deep networks such as Inception-v3, ResNet-101 and DenseNet-121 can be quantized for inference with 4-bit weights and activations (with 1% 16-bit data) within 1% top-1 accuracy drop."}, "keywords": ["quantization-aware training methods"], "citation_intent": "background"}
{"citing_id": "2303.15367v1", "cited_id": "1701.09133", "section_title": "Bernoulli-Domination.", "citation": "It also requires Lemma 3.7, the proof of which is adapted from that of #REFR Lemma 7] and is deferred to the Appendix.", "text_before_citation": [", \u2113 \u03c3 (v s ) \u2264 p\u2113} |C (G)| \u2264 |C (G \\ J)| \u2022 (p\u2113) s |C (G \\ J)| \u2022 \u2113 s = p s .", "The upper bound for the numerator in the above inequality comes from two facts that rely on J being an independent set.", "First, given a colouring \u03c3 0 \u2208 C (G \\ J), the number of extensions of \u03c3 0 to a colouring \u03c3 \u2208 C (G) is v\u2208J \u2113 \u03c3 0 (v).", "Second, given such an extension \u03c3, we have \u2113 \u03c3 (v) = \u2113 \u03c3 0 (v) for every v \u2208 J. We conclude that (X v ) v\u2208I are Ber(p)-dominated.", "Leveraging Theorem 3.6, and the fact that neighbourhoods induce independent sets, we can prove the following exponential upper bound on the likelihood of short lists."], "text_after_citation": ["Lemma 3.7.", "Let G be a triangle-free graph, let v \u2208 V (G) and let \u03c3 0 be a proper k-colouring of G \\ N [v] (for some k), with at least one extension to G.", "Then if \u03c3 is the uniformly random extension of \u03c3 0 to G, writing", "\u2113 := E [\u2113 \u03c3 (v)], we have P [\u2113 \u03c3 (v) \u2264 (1 \u2212 \u03b4)\u2113] \u2264 e \u2212 \u03b4 2 \u2113 2 ,", "for all \u03b4 \u2208 (0, 1)."], "citing_paper_content": {"title": "Uniformly Random Colourings Of Sparse Graphs", "abstract": "We analyse uniformly random proper k-colourings of sparse graphs with maximum degree \u2206 in the regime \u2206 < k ln k. This regime corresponds to the lower side of the shattering threshold for random graph colouring, a paradigmatic example of the shattering threshold for random Constraint Satisfaction Problems. We prove a variety of results about the solution space geometry of colourings of fixed graphs, generalising work of Achlioptas, Coja-Oghlan [ACO08], and Molloy [Mol12] on random graphs, and justifying the performance of stochastic local search algorithms in this regime. Our central proof relies only on elementary techniques, namely the firstmoment method and a quantitative induction, yet it strengthens list-colouring results due to Vu [Vu02], and more recently Davies, Kang, P., and Sereni [DKPS20], and generalises state-of-the-art bounds from Ramsey theory in the context of sparse graphs. It further yields an approximately tight lower bound on the number of colourings, also known as the partition function of the Potts model, with implications for efficient approximate counting. The research leading to these results was partially supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)-428212407 (E. Hurley)."}, "cited_paper_content": {"title": "The List Chromatic Number Of Graphs With Small Clique Number", "abstract": "We prove that every triangle-free graph with maximum degree \u0394 has list chromatic number at most ( 1 + o ( 1 ) ) \u0394 ln \u2061 \u0394 . This matches the best-known upper bound for graphs of girth at least 5. We also provide a new proof that for any r \u2265 4 every K r -free graph has list-chromatic number at most 200 r \u0394 ln \u2061 ln \u2061 \u0394 ln \u2061 \u0394 ."}, "keywords": ["Lemma"], "citation_intent": "background"}
{"citing_id": "2303.06241v1", "cited_id": "1706.06083", "section_title": "B. Results On Cifar-10", "citation": "During training the adversarial examples are constructed with = 0.0157 (4/255) as suggested by #REFR .", "text_before_citation": ["Next, we evaluated our approach on the CIFAR-10 dataset using FGSM adversarial training.", "We trained the CIFAR-10 model using Wide-Resnet 32-10 model and standard hyperparameters used by Madry et al. #OTHEREFR ."], "text_after_citation": ["Same value is used to measure robust accuracy during test time. The results are shown in Table I .", "Conventional adversarial training on the CIFAR-10 dataset took around 50 hours to converge.", "By using our approach, we were able to decrease the training time by a factor of 1.98 and bring it down to 25 hours.", "The robust accuracy of the model decreased marginally by 0.64%.", "This shows that we were able to achieve a similar level of robustness by using a subset of training data."], "citing_paper_content": {"title": "Do We Need Entire Training Data For Adversarial Training?", "abstract": "Deep Neural Networks (DNNs) are being used to solve a wide range of problems in many domains including safetycritical domains like self-driving cars and medical imagery. DNNs suffer from vulnerability against adversarial attacks. In the past few years, numerous approaches have been proposed to tackle this problem by training networks using adversarial training. Almost all the approaches generate adversarial examples for the entire training dataset, thus increasing the training time drastically. We show that we can decrease the training time for any adversarial training algorithm by using only a subset of training data for adversarial training. To select the subset, we filter the adversarially-prone samples from the training data. We perform a simple adversarial attack on all training examples to filter this subset. In this attack, we add a small perturbation to each pixel and a few grid lines to the input image. We perform adversarial training on the adversarially-prone subset and mix it with vanilla training performed on the entire dataset. Our results show that when our method-agnostic approach is plugged into FGSM [9], we achieve a speedup of 3.52x on MNIST and 1.98x on the CIFAR-10 dataset with comparable robust accuracy. We also test our approach on state-of-the-art Free adversarial training [24] and achieve a speedup of 1.2x in training time with a marginal drop in robust accuracy on the ImageNet dataset."}, "cited_paper_content": {"title": "Towards Deep Learning Models Resistant To Adversarial Attacks", "abstract": "Recent work has demonstrated that neural networks are vulnerable to adversarial examples, i.e., inputs that are almost indistinguishable from natural data and yet classified incorrectly by the network. In fact, some of the latest findings suggest that the existence of adversarial attacks may be an inherent weakness of deep learning models. To address this problem, we study the adversarial robustness of neural networks through the lens of robust optimization. This approach provides us with a broad and unifying view on much of the prior work on this topic. Its principled nature also enables us to identify methods for both training and attacking neural networks that are reliable and, in a certain sense, universal. In particular, they specify a concrete security guarantee that would protect against any adversary. These methods let us train networks with significantly improved resistance to a wide range of adversarial attacks. They also suggest the notion of security against a first-order adversary as a natural and broad security guarantee. We believe that robustness against such well-defined classes of adversaries is an important stepping stone towards fully resistant deep learning models."}, "keywords": ["adversarial examples"], "citation_intent": "method"}
{"citing_id": "2304.08660v1", "cited_id": "1902.07381", "section_title": "A. Image-Based Place Recognition", "citation": "In #REFR , a method to improve feature matching with depth estimation has been introduced, opening a potential to enhance place recognition with estimated depth.", "text_before_citation": ["This approach has exhibited superior robustness to changes in appearance.", "To efficiently translate place-distinctive image encoding into localization descriptors, the negative and positive samples were trained in pairs using weak supervision in NetVLAD #OTHEREFR with a triplet loss #OTHEREFR . Radenovic et al.", "discovered more consistent and distinctive feature representation from images, such as contrastive loss with generalized mean (GeM) pooling #OTHEREFR .", "The study considers every image with sufficiently large number of co-observed 3D points or similar features for a training pair.", "This differs from weak supervision with triplet loss, which selects pairs by the image location and their descriptor distance."], "text_after_citation": ["In recent studies, methods unifying local and global features have been proposed for expansion over large-scale visual place recognition #OTHEREFR to geometrically verify that the local feature matches after global image searching."], "citing_paper_content": {"title": "(Lc) 2 : Lidar-Camera Loop Constraints For Cross-Modal Place Recognition", "abstract": "Localization has been a challenging task for autonomous navigation. A loop detection algorithm must overcome environmental changes for the place recognition and relocalization of robots. Therefore, deep learning has been extensively studied for the consistent transformation of measurements into localization descriptors. Street view images are easily accessible; however, images are vulnerable to appearance changes. LiDAR can robustly provide precise structural information. However, constructing a point cloud database is expensive, and point clouds exist only in limited places. Different from previous works that train networks to produce shared embedding directly between the 2D image and 3D point cloud, we transform both data into 2.5D depth images for matching. In this work, we propose a novel cross-matching method, called (LC) 2 , for achieving LiDAR localization without a prior point cloud map. To this end, LiDAR measurements are expressed in the form of range images before matching them to reduce the modality discrepancy. Subsequently, the network is trained to extract localization descriptors from disparity and range images. Next, the best matches are employed as a loop factor in a pose graph. Using public datasets that include multiple sessions in significantly different lighting conditions, we demonstrated that LiDAR-based navigation systems could be optimized from image databases and vice versa."}, "cited_paper_content": {"title": "Look No Deeper: Recognizing Places From Opposing Viewpoints Under Varying Scene Appearance Using Single-View Depth Estimation", "abstract": "Visual place recognition (VPR) - the act of recognizing a familiar visual place - becomes difficult when there is extreme environmental appearance change or viewpoint change. Particularly challenging is the scenario where both phenomena occur simultaneously, such as when returning for the first time along a road at night that was previously traversed during the day in the opposite direction. While such problems can be solved with panoramic sensors, humans solve this problem regularly with limited field-of-view vision and without needing to constantly turn around. In this paper, we present a new depth- and temporal-aware visual place recognition system that solves the opposing viewpoint, extreme appearance-change visual place recognition problem. Our system performs sequence-to-single frame matching by extracting depth-filtered keypoints using a state-of-the-art depth estimation pipeline, constructing a keypoint sequence over multiple frames from the reference dataset, and comparing these keypoints to the keypoints extracted from a single query image. We evaluate the system on a challenging benchmark dataset and show that it consistently outperforms state-of-the-art techniques. We also develop a range of diagnostic simulation experiments that characterize the contribution of depth-filtered keypoint sequences with respect to key domain parameters including the degree of appearance change and camera motion."}, "keywords": ["place recognition"], "citation_intent": "method"}
{"citing_id": "2303.14029v1", "cited_id": "2002.11049", "section_title": "Table I: Code Context Extraction Comparison", "citation": "The well-established 'Easy to Find' #REFR features such as 'TODO', 'FIXME', and 'XXX' were excluded in the manual validation.", "text_before_citation": ["Context in prior work #OTHEREFR Code Context in SoCCMiner #OTHEREFR public #OTHEREFR with 'Sense2Vec' #OTHEREFR .", "Sense2Vec captures contextually similar words with its deep word embedding.", "For example, for feature 'problematic', Sense2Vec returns multiple semantically similar words such as 'troublesome', 'undesirable', 'contentious', 'separate issue', etc., We use these extrapolated features (words) to evaluate our code comments for SATD.", "Sense2Vec extrapolated 'Easy to Find' features such as ('sabotaging', 'damaging', 'counter-productive', etc.,) are passed to the first round of automated SATD annotation.", "The first author manually evaluates the results by reading the comments."], "text_after_citation": ["During the validation process, the first author creates the heuristics for 'Hard to Find' #OTHEREFR This iterative process involved multiple iterations of annotation, manual evaluation, creating heuristics (pattern matching), and creating features for the 'Hard to Find' SATD.", "SATD Annotation: Features that attract too many false positives are removed from the SATD feature list.", "For example, 'cause problem' will capture the comment '// Empty bboxes can cause problems' which sound right at first.", "Later, after manual scanning, several header comments had the following comment content '//This won't cause problems'.", "Such features are removed from the SATD feature list to avoid overfitting features."], "citing_paper_content": {"title": "Pentacet Data -23 Million Contextual Code Comments And 500,000 Satd Comments", "abstract": "Most Self-Admitted Technical Debt (SATD) research utilizes explicit SATD features such as 'TODO' and 'FIXME' for SATD detection. A closer look reveals several SATD research uses simple SATD ('Easy to Find') code comments without the contextual data (preceding and succeeding source code context). This work addresses this gap through PENTACET (or 5C dataset) data. PENTACET is a large Curated Contextual Code Comments per Contributor and the most extensive SATD data. We mine 9,096 Open Source Software Java projects with a total of 435 million LOC. The outcome is a dataset with 23 million code comments, preceding and succeeding source code context for each comment, and more than 500,000 comments labeled as SATD, including both 'Easy to Find' and 'Hard to Find' SATD. We believe PENTACET data will further SATD research using Artificial Intelligence techniques."}, "cited_paper_content": {"title": "Identifying Self-Admitted Technical Debts With Jitterbug: A Two-Step Approach", "abstract": "Keeping track of and managing the self-admitted technical debts (SATDs) is important to maintaining a healthy software project. This requires much time and effort from human experts to identify these SATDs manually. Currently, automated solutions do not have high enough precision and recall in identifying SATDs to fully automate the process. To solve the above problems, we propose a two-step framework called Jitterbug for identifying SATDs by first finding the \"easy to find\" SATDs automatically with close to 100% precision via a novel pattern recognition technique, then applying machine learning techniques to assist human experts in manually identifying the rest \"hard to find\" SATDs with reduced human effort. Our simulation studies on ten software projects show that Jitterbug can identify SATDs more efficiently (with less human effort) than the prior state of the art methods."}, "keywords": ["manual validation"], "citation_intent": "method"}
{"citing_id": "2305.00656v1", "cited_id": "1908.06148", "section_title": "B. Convolution Neural Network", "citation": "The output of the embedding layer is fed into a convolution layer with kernels of any size. Several convolution layers can be stacked #REFR .", "text_before_citation": ["Various kernels can be used to extract various features from the input data points.", "In general, kernel filters are designed to be smaller than the input to allow for the sharing of kernel weights across input dimensions.", "In file fragment classification, the input to the CNN is either a vector of 4096 or 512 dimensions, equivalent to fragments of 4KB or 512 bytes, respectively.", "The input is then transmitted to an embedding layer that converts the input vector into a continuous (4096,32) or (512,32) dimensional vector, dependent on the fragment size.", "The embedding layer is a preliminary step to prepare the input vectors for subseque CNN layers."], "text_after_citation": ["As a result of stacking standard convolution layers when developing a Deep Neural Network (DNN), in which the model can have a large number of parameters."], "citing_paper_content": {"title": "File Fragment Classification Using Light-Weight Convolutional Neural Networks", "abstract": "In digital forensics, file fragment classification is an important step toward completing file carving process. There exist several techniques to identify the type of file fragments without relying on meta-data, such as using features like header/footer and N-gram to identify the fragment type. Recently, convolutional neural network (CNN) models have been used to build classification models to achieve this task. However, the number of parameters in CNNs tends to grow exponentially as the number of layers increases. This results in a dramatic increase in training and inference time. In this paper, we propose lightweight file fragment classification models based on depthwise separable CNNs. The evaluation results show that our proposed models provide faster inference time with comparable accuracy as compared to the state-of-art CNN based models. In particular, our models were able to achieve an accuracy of 79% on the FFT-75 dataset with nearly 100K parameters and 164M FLOPs, which is 4x smaller and 6x faster than the state-of-the-art classifier in the literature."}, "cited_paper_content": {"title": "Fifty: Large-Scale File Fragment Type Identification Using Neural Networks", "abstract": "We present FiFTy, a modern file type identification tool for memory forensics and data carving. In contrast to previous approaches based on hand-crafted features, we design a compact neural network architecture, which uses a trainable embedding space, akin to successful natural language processing models. Our approach dispenses with explicit feature extraction which is a bottleneck in legacy systems. We evaluate the proposed method on a novel dataset with 75 file types - the most diverse and balanced dataset reported to date. FiFTy consistently outperforms all baselines in terms of speed, accuracy and individual misclassification rates. We achieved an average accuracy of 77.5% with processing speed of approx 38 sec/GB, which is better and more than an order of magnitude faster than the previous state-of-the-art tool - Sceadan (69% at 9 min/GB). Our tool and the corresponding dataset are available publicly online."}, "keywords": ["Several convolution layers", "embedding layer"], "citation_intent": "method"}
{"citing_id": "2303.10030v1", "cited_id": "1902.11156", "section_title": "Main Result", "citation": "Compared to the stability result in #REFR (see Theorem 2.5), in our result we observe a square-root dependence of the reconstruction error bound on the noise level \u03c4 for small noise levels.", "text_before_citation": ["To see how this compares to the the existing dimension-dependent recovery guarantee (2.6) in [ARR14; LS17; JKS17] we first reformulate (3.2) as", "X * \u2212 X 0 F (log(\u03c9L)) 1/2 X 0 F \u03c4 \u2022 \u03c4.", "Ignoring logarithmic factors, for noise levels", "L KN \u00b5 2 max X 0 F \u226a \u03c4 \u2264 X 0 F ,", "this significantly improves over the dimension-dependent recovery guarantee (2.6)."], "text_after_citation": ["In contrast, the reconstruction error bound in Theorem 2.5 becomes constant whenever the noise level \u03c4 is smaller than a certain threshold.", "The bound (3.1) in Theorem 3.1 becomes worse when the noise level \u03c4 becomes smaller.", "This reflects the instability result in #OTHEREFR , see Theorem 2.4, which shows the existence of an alternative solution which amplifies the output error by a dimension factor."], "citing_paper_content": {"title": "How Robust Is Randomized Blind Deconvolution Via Nuclear Norm Minimization Against Adversarial Noise?", "abstract": "In this paper, we study the problem of recovering two unknown signals from their convolution, which is commonly referred to as blind deconvolution. Reformulation of blind deconvolution as a low-rank recovery problem has led to multiple theoretical recovery guarantees in the past decade due to the success of the nuclear norm minimization heuristic. In particular, in the absence of noise, exact recovery has been established for sufficiently incoherent signals contained in lower-dimensional subspaces. However, if the convolution is corrupted by additive bounded noise, the stability of the recovery problem remains much less understood. In particular, existing reconstruction bounds involve large dimension factors and therefore fail to explain the empirical evidence for dimension-independent robustness of nuclear norm minimization. Recently, theoretical evidence has emerged for ill-posed behavior of low-rank matrix recovery for sufficiently small noise levels. In this work, we develop improved recovery guarantees for blind deconvolution with adversarial noise which exhibit square-root scaling in the noise level. Hence, our results are consistent with existing counterexamples which speak against linear scaling in the noise level as demonstrated for related low-rank matrix recovery problems."}, "cited_paper_content": {"title": "On The Convex Geometry Of Blind Deconvolution And Matrix Completion", "abstract": "Low-rank matrix recovery from structured measurements has been a topic of intense study in the last decade and many important problems like matrix completion and blind deconvolution have been formulated in this framework. An important benchmark method to solve these problems is to minimize the nuclear norm, a convex proxy for the rank. A common approach to establish recovery guarantees for this convex program relies on the construction of a so-called approximate dual certificate. However, this approach provides only limited insight in various respects. Most prominently, the noise bounds exhibit seemingly suboptimal dimension factors. In this paper we take a novel, more geometric viewpoint to analyze both the matrix completion and the blind deconvolution scenario. We find that for both these applications the dimension factors in the noise bounds are not an artifact of the proof, but the problems are intrinsically badly conditioned. We show, however, that bad conditioning only arises for very small noise levels: Under mild assumptions that include many realistic noise levels we derive near-optimal error estimates for blind deconvolution under adversarial noise."}, "keywords": ["reconstruction error"], "citation_intent": "result"}
{"citing_id": "2303.10151v1", "cited_id": "1609.04802", "section_title": "Super-Resolution", "citation": "The first application of GANs to super-resolution was in SRGAN, which outperformed prior methods and achieved state-of-the-art results #REFR .", "text_before_citation": [], "text_after_citation": ["The authors attributed this, in part, to their use of a perceptual loss function that accounted for perceptual similarity instead of just similarity in pixel space.", "At that time, a common problem with superresolution was the presence of artifacts when upsampling.", "ESRGAN addressed this issue by identifying that batch normalization layers tended to create unwanted artifacts #OTHEREFR .", "They also improved the perceptual loss function and used Residual-in-Residual Dense Blocks to generate more realistic images consistently.", "Additionally, they later proposed REAL-ESRGAN, which incorporated a u-net discriminator and spectral normalization #OTHEREFR ."], "citing_paper_content": {"title": "Toward Super-Resolution For Appearance-Based Gaze Estimation", "abstract": "Gaze tracking is a valuable tool with a broad range of applications in various fields, including medicine, psychology, virtual reality, marketing, and safety. Therefore, it is essential to have gaze tracking software that is cost-efficient and high-performing. Accurately predicting gaze remains a difficult task, particularly in real-world situations where images are affected by motion blur, video compression, and noise. Super-resolution has been shown to improve image quality from a visual perspective. This work examines the usefulness of super-resolution for improving appearancebased gaze tracking. We show that not all SR models preserve the gaze direction. We propose a two-step framework based on SwinIR super-resolution model. The proposed method consistently outperforms the state-of-the-art, particularly in scenarios involving low-resolution or degraded images. Furthermore, we examine the use of superresolution through the lens of self-supervised learning for gaze prediction. Self-supervised learning aims to learn from unlabelled data to reduce the amount of required labeled data for downstream tasks. We propose a novel architecture called \"SuperVision\" by fusing an SR backbone network to a ResNet18 (with some skip connections). The proposed SuperVision method uses 5x less labeled data and yet outperforms, by 15%, the state-of-the-art method of GazeTR which uses 100% of training data. We will make our code publicly available upon publication."}, "cited_paper_content": {"title": "Photo-Realistic Single Image Super-Resolution Using A Generative Adversarial Network", "abstract": "Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central problem remains largely unsolved: how do we recover the finer texture details when we super-resolve at large upscaling factors? The behavior of optimization-based super-resolution methods is principally driven by the choice of the objective function. Recent work has largely focused on minimizing the mean squared reconstruction error. The resulting estimates have high peak signal-to-noise ratios, but they are often lacking high-frequency details and are perceptually unsatisfying in the sense that they fail to match the fidelity expected at the higher resolution. In this paper, we present SRGAN, a generative adversarial network (GAN) for image super-resolution (SR). To our knowledge, it is the first framework capable of inferring photo-realistic natural images for 4x upscaling factors. To achieve this, we propose a perceptual loss function which consists of an adversarial loss and a content loss. The adversarial loss pushes our solution to the natural image manifold using a discriminator network that is trained to differentiate between the super-resolved images and original photo-realistic images. In addition, we use a content loss motivated by perceptual similarity instead of similarity in pixel space. Our deep residual network is able to recover photo-realistic textures from heavily downsampled images on public benchmarks. An extensive mean-opinion-score (MOS) test shows hugely significant gains in perceptual quality using SRGAN. The MOS scores obtained with SRGAN are closer to those of the original high-resolution images than to those obtained with any state-of-the-art method."}, "keywords": ["resolution", "GANs"], "citation_intent": "method"}
{"citing_id": "2305.01918v2", "cited_id": "1810.04805", "section_title": "Sample Pair Generation:", "citation": "Training details: We use the base version of the pre-trained language model BERT #REFR and RoBERTa as our backbone models.", "text_before_citation": ["For CLAIF, we use unpaired sentences from the training set of STS-B as original sentences to construct sentence pairs from scratch and randomly sample two other sentences for each original sentence to construct two sentence pairs with a similarity score of 0.", "For CLHAIF, following previous studies #OTHEREFR , we use the SNLI and MNLI datasets to construct sentence pairs and add a AI feedback similarity score for each sentence pair.", "We only use the AI feedback scores for positive pairs in our experiments of CLHAIF.", "Besides, to demonstrate the scalability of CLAIF, we use sentence pairs constructed from STS-B and from NLI datasets for the training of CLAIF, which we called CLAIF scaled .", "We list statistics of some datasets used for different methods in Table 2 ."], "text_after_citation": ["We use the development set of STS-B as our validation set.", "In CLAIF, we use the mean pooling strategy to get sentence embeddings for BERT and RoBERTa.", "For CLHAIF, we take the same pooling strategy as the corresponding baseline. Other implementation details are in Appendix A."], "citing_paper_content": {"title": "Improving Contrastive Learning Of Sentence Embeddings From Ai Feedback", "abstract": "Contrastive learning has become a popular approach in natural language processing, particularly for the learning of sentence embeddings. However, the discrete nature of natural language makes it difficult to ensure the quality of positive and negative sample pairs generated through data augmentation methods. Although supervised contrastive learning can produce more accurate sample pairs with human feedback labels, it still lacks fine-grained training signals. In this paper, we propose to improve Contrastive Learning of sentence embeddings from AI Feedback (CLAIF). Our method utilizes AI feedback from large pretrained language models (LLMs) to construct sample pairs with fine-grained sample similarity scores to improve contrastive learning. Besides, we combine human feedback and AI feedback to provide better supervision signals for supervised contrastive learning of sentence embeddings. Experimental results show that our method achieves state-of-the-art performance on several semantic textual similarity (STS) and transfer learning tasks compared to other unsupervised and supervised contrastive learning methods. 1"}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["pre-trained language model"], "citation_intent": "method"}
{"citing_id": "2304.04667v2", "cited_id": "1901.01412", "section_title": "Preliminaries", "citation": "Using notations in #REFR , a search problem P is a binary relation, and we say that S is a solution to instance x iff (x, S) \u2208 P .", "text_before_citation": ["Min-Cut data structure.", "Following notations from [AKT20a], a Min-Cut data structure for a graph family F is a data structure that given a graph G \u2208 F, after a preprocessing phase which makes t pq (m) Max-Flow calls on G and spends t po (m) time outside of these calls, can answer Min-Cut queries for any two nodes s, t \u2208 V in amortized query time (or output sensitive time) t mc (k st ), where k st denotes the output size (number of edges or nodes in the output cut or separator).", "In particular, it means that the preprocessing time is at most t pq (m) \u2022\u00d4(m) + t po (m) using [CKL + 22], and the query time is O(k st \u2022 t mc (k st )).", "We may also be interested in Min-Cut data structures that support queries of Max-Flow values instead of cuts, and we denote their query time by t mf (m).", "Non-reducibility."], "text_after_citation": ["Let SOL(x) = {S : (x, S) \u2208 P } denote the set of solutions for instance x.", "We say that P is a total function if every instance x has at least one solution, i.e. SOL(x) = \u2205.", "Let \u22a5 be the \"don't know\" symbol and assume \u22a5 / \u2208 SOL(x) for all x.", "In particular, in our context of Min-Cut data structures, x is a graph, and SOL(x) is the set of all Min-Cut data structures with t mc (m) =\u00d4(1) for x where m is the number of edges in x.", "Next, we define the nondeterministic complexity of a total function."], "citing_paper_content": {"title": "(Almost) Ruling Out Seth Lower Bounds For All-Pairs Max-Flow", "abstract": "The All-Pairs Max-Flow problem has gained significant popularity in the last two decades, and many results are known regarding its fine-grained complexity. Despite this, wide gaps remain in our understanding of the time complexity for several basic variants of the problem. In this paper, we aim to bridge this gap by providing algorithms, conditional lower bounds, and non-reducibility results. Notably, we show that for most problem settings, deterministic reductions based on the Strong Exponential Time Hypothesis (SETH) cannot rule out n 4\u2212o(1) time algorithms under a hypothesis called NSETH. We present results for the following two cases, where some of the bounds assume that the matrix multiplication exponent \u03c9 = 2. Node-Capacities. For directed graphs with general node-capacities, we show that for any fixed \u03b5 > 0, a deterministic SETH-based reduction cannot rule out O(n 3+\u03b5) time algorithms under NSETH, which is tight with the known conditional lower bound of n 3\u2212o(1) devised for sparse graphs under SETH [Abboud et al., ToC 2021]. The same non-reducibility bound immediately applies to undirected graphs with unit node-capacities, which can be compared to an m 3/2\u2212o(1) + n 2 SETH-based lower bound following from a combination of previous works. Edge-Capacities. For directed graphs with unit edge-capacities, we similarly rule out an \u221a m \u2022 n 5/2+\u03b5\u2212o(1) SETH-based lower bound under NSETH, which can be contrasted with the known mn 1\u2212o(1) SETH-based lower bound [Krauthgamer and Trabelsi, TALG 2018]. While it remains open whether deterministic SETH-based reductions can rule out an n 4\u2212o(1) time algorithm for general edge-capacities, a consequence of our results is that a subquartic fine-grained reduction from the directed edge-capacitated setting to the directed node-capacitated setting would rule out such deterministic SETH-based reductions under NSETH. As a step towards ruling out even mn 1+\u03b5\u2212o(1) SETH lower bounds for undirected graphs with unit node-capacities, we design a new randomized O m 2+o(1) time combinatorial algorithm. This is an improvement over the recent O m 11/5+o(1) time algorithm [Huang et al., STOC 2023] and matching their m 2\u2212o(1) lower bound (up to subpolynomial factors), thus essentially settling the time complexity for this setting of the problem. More generally, our main technical contribution is the insight that st-cuts can be verified quickly, and that in most settings, st-flows can be shipped succinctly (i.e., with respect to the flow support). This is a key idea in our non-reducibility results, and it may be of independent interest."}, "cited_paper_content": {"title": "New Algorithms And Lower Bounds For All-Pairs Max-Flow In Undirected Graphs", "abstract": "We investigate the time-complexity of the All-Pairs Max-Flow problem: Given a graph with $n$ nodes and $m$ edges, compute for all pairs of nodes the maximum-flow value between them. If Max-Flow (the version with a given source-sink pair $s,t$) can be solved in time $T(m)$, then an $O(n^2) \\cdot T(m)$ is a trivial upper bound. But can we do better? ::: For directed graphs, recent results in fine-grained complexity suggest that this time bound is essentially optimal. In contrast, for undirected graphs with edge capacities, a seminal algorithm of Gomory and Hu (1961) runs in much faster time $O(n)\\cdot T(m)$. Under the plausible assumption that Max-Flow can be solved in near-linear time $m^{1+o(1)}$, this half-century old algorithm yields an $nm^{1+o(1)}$ bound. Several other algorithms have been designed through the years, including $\\tilde{O}(mn)$ time for unit-capacity edges (unconditionally), but none of them break the $O(mn)$ barrier. Meanwhile, no super-linear lower bound was shown for undirected graphs. ::: We design the first hardness reductions for All-Pairs Max-Flow in undirected graphs, giving an essentially optimal lower bound for the node-capacities setting. For edge capacities, our efforts to prove similar lower bounds have failed, but we have discovered a surprising new algorithm that breaks the $O(mn)$ barrier for graphs with unit-capacity edges! Assuming $T(m)=m^{1+o(1)}$, our algorithm runs in time $m^{3/2 +o(1)}$ and outputs a cut-equivalent tree (similarly to the Gomory-Hu algorithm). Even with current Max-Flow algorithms we improve state-of-the-art in many density regimes."}, "keywords": ["binary relation"], "citation_intent": "background"}
{"citing_id": "2303.01471v1", "cited_id": "1503.01243", "section_title": "B.3 Qhd For Quadratic Model Functions", "citation": "This rate is at par with the convergence rate of the continuous-time model of Nesterov's method #REFR .", "text_before_citation": ["The Schr\u00f6dinger equation describing QHD is often too complicated to solve analytically.", "Fortunately, we can find a closed-form solution of QHD when f is a quadratic function.", "In the following calculation, we consider the one-dimensional case f (x) = 1 2 x 2 for simplicity.", "It is worth noting that the same method also applies to general finite-dimensional quadratic forms f (x) = 1 2 x T Ax with a positive semidefinite matrix A #OTHEREFR .", "It turns out that, if we choose the same time-dependent parameters as in Nesterov's accelerated gradient descent, the convergence rate is E[f ] \u223c\u03a8t = O(t \u22123 )."], "text_after_citation": ["Remark 4.", "Our result does not mean QHD can not achieve faster convergence for quadratic objective functions.", "In fact, one can use a linear function for \u03b2 t , and the convergence rate can be exponentially fast.", "Here, our goal is to compare QHD to the classical ODE model of Nesterov's method.", "We choose the following time-dependent parameters in the QHD Hamiltonian (A.23):"], "citing_paper_content": {"title": "Quantum Hamiltonian Descent *", "abstract": "Gradient descent is a fundamental algorithm in both theory and practice for continuous optimization. Identifying its quantum counterpart would be appealing to both theoretical and practical quantum applications. A conventional approach to quantum speedups in optimization relies on the quantum acceleration of intermediate steps of classical algorithms, while keeping the overall algorithmic trajectory and solution quality unchanged. We propose Quantum Hamiltonian Descent (QHD), which is derived from the path integral of dynamical systems referring to the continuous-time limit of classical gradient descent algorithms, as a truly quantum counterpart of classical gradient methods where the contribution from classically-prohibited trajectories can significantly boost QHD's performance for non-convex optimization. Moreover, QHD is described as a Hamiltonian evolution efficiently simulatable on both digital and analog quantum computers. By embedding the dynamics of QHD into the evolution of the socalled Quantum Ising Machine (including D-Wave and others), we empirically observe that the D-Wave-implemented QHD outperforms a selection of state-of-the-art gradient-based classical solvers and the standard quantum adiabatic algorithm, based on the time-to-solution metric, on non-convex constrained quadratic programming instances up to 75 dimensions. Finally, we propose a \"three-phase picture\" to explain the behavior of QHD, especially its difference from the quantum adiabatic algorithm."}, "cited_paper_content": {"title": "A Differential Equation For Modeling Nesterov'S Accelerated Gradient Method: Theory And Insights", "abstract": "We derive a second-order ordinary differential equation (ODE) which is the limit of Nesterov's accelerated gradient method. This ODE exhibits approximate equivalence to Nesterov's scheme and thus can serve as a tool for analysis. We show that the continuous time ODE allows for a better understanding of Nesterov's scheme. As a byproduct, we obtain a family of schemes with similar convergence rates. The ODE interpretation also suggests restarting Nesterov's scheme leading to an algorithm, which can be rigorously proven to converge at a linear rate whenever the objective is strongly convex."}, "keywords": ["Nesterov's method"], "citation_intent": "result"}
{"citing_id": "2303.06419v1", "cited_id": "1706.06083", "section_title": "Method", "citation": "One such approach to solving this maximization would be the straight-forward application of projected gradient descent (PGD) #REFR .", "text_before_citation": ["In order to achieve robustness to this class of perturbations we use a min-max optimization approach common in the adversarial robustness literature, but modified for our problem setting:", "EQUATION", "We use \u00d7 to denote element-wise product throughout.", "The above formulation uses a weighting \u03b1 to trade off between the standard task loss, the first term, and the adversarial loss incurred by the neural network being non-robust in the human specified shortcut directions, the second term.", "We can leverage the many advances in adversarial robustness in order to approximately solve the maximization problem posed in the second half of our loss term."], "text_after_citation": ["Given an input, x, a mask m, and a positive value \u03ba, PGD uses a first-order optimization approach to arrive at an input x * that approximately solves the maximization in the second term of our loss.", "We refer to the approach of using PGD in our loss formulation as PGD-Ex.", "Given the non-convexity of this maximization, however, no guarantees can be made about the quality of the approximate solution x * .", "Reliance on a local attack method in our optimization has the potential to bias our approaches to the specific attack chosen, as observed in the adversarial robustness literature #OTHEREFR .", "Both local explanations and local adversarial attacks share this weakness."], "citing_paper_content": {"title": "Robust Learning From Explanations", "abstract": "Machine learning from explanations (MLX) is an approach to learning that uses human-provided annotations of relevant features for each input to ensure that model predictions are right for the right reasons. Existing MLX approaches rely heavily on a specific model interpretation approach and require strong parameter regularization to align model and human explanations, leading to suboptimal performance. We recast MLX as an adversarial robustness problem, where human explanations specify a lower dimensional manifold from which perturbations can be drawn, and show both theoretically and empirically how this approach alleviates the need for strong parameter regularization. We consider various approaches to achieving robustness, leading to improved performance over prior MLX methods. Finally, we combine robustness with an earlier MLX method, yielding state-of-the-art results on both synthetic and real-world benchmarks."}, "cited_paper_content": {"title": "Towards Deep Learning Models Resistant To Adversarial Attacks", "abstract": "Recent work has demonstrated that neural networks are vulnerable to adversarial examples, i.e., inputs that are almost indistinguishable from natural data and yet classified incorrectly by the network. In fact, some of the latest findings suggest that the existence of adversarial attacks may be an inherent weakness of deep learning models. To address this problem, we study the adversarial robustness of neural networks through the lens of robust optimization. This approach provides us with a broad and unifying view on much of the prior work on this topic. Its principled nature also enables us to identify methods for both training and attacking neural networks that are reliable and, in a certain sense, universal. In particular, they specify a concrete security guarantee that would protect against any adversary. These methods let us train networks with significantly improved resistance to a wide range of adversarial attacks. They also suggest the notion of security against a first-order adversary as a natural and broad security guarantee. We believe that robustness against such well-defined classes of adversaries is an important stepping stone towards fully resistant deep learning models."}, "keywords": ["projected gradient descent"], "citation_intent": "method"}
{"citing_id": "2304.03495v1", "cited_id": "1707.09700", "section_title": "Predcls", "citation": "SGCls SGDet R@50 / 100 mR@50/100 F@50 / 100 R@50 / 100 mR@50/100 F@50 / 100 R@50 / 100 mR@50/100 F@50 / 100 IMP+ \u2021 #REFR 61. Table S4 .", "text_before_citation": [], "text_after_citation": ["mR@100 on the SGDet setting for head, body, and tail classes.", "\u2020 denotes that the bi-level sampling is applied on the model to achieve these results. Bold numbers indicate the best performances."], "citing_paper_content": {"title": "Devil'S On The Edges: Selective Quad Attention For Scene Graph Generation", "abstract": "Scene graph generation aims to construct a semantic graph structure from an image such that its nodes and edges respectively represent objects and their relationships. One of the major challenges for the task lies in the presence of distracting objects and relationships in images; contextual reasoning is strongly distracted by irrelevant objects or backgrounds and, more importantly, a vast number of irrelevant candidate relations. To tackle the issue, we propose the Selective Quad Attention Network (SQUAT) that learns to select relevant object pairs and disambiguate them via diverse contextual interactions. SQUAT consists of two main components: edge selection and quad attention. The edge selection module selects relevant object pairs, i.e., edges in the scene graph, which helps contextual reasoning, and the quad attention module then updates the edge features using both edge-to-node and edge-to-edge cross-attentions to capture contextual information between objects and object pairs. Experiments demonstrate the strong performance and robustness of SQUAT, achieving the state of the art on the Visual Genome and Open Images v6 benchmarks."}, "cited_paper_content": {"title": "Scene Graph Generation From Objects, Phrases And Region Captions", "abstract": "Object detection, scene graph generation and region captioning, which are three scene understanding tasks at different semantic levels, are tied together: scene graphs are generated on top of objects detected in an image with their pairwise relationship predicted, while region captioning gives a language description of the objects, their attributes, relations, and other context information. In this work, to leverage the mutual connections across semantic levels, we propose a novel neural network model, termed as Multi-level Scene Description Network (denoted as MSDN), to solve the three vision tasks jointly in an end-to-end manner. Objects, phrases, and caption regions are first aligned with a dynamic graph based on their spatial and semantic connections. Then a feature refining structure is used to pass messages across the three levels of semantic tasks through the graph. We benchmark the learned model on three tasks, and show the joint learning across three tasks with our proposed method can bring mutual improvements over previous models. Particularly, on the scene graph generation task, our proposed method outperforms the state-of-art method with more than 3% margin."}, "keywords": ["Table S4"], "citation_intent": "background"}
{"citing_id": "2303.15749v1", "cited_id": "1412.6980", "section_title": "Implementation Details", "citation": "For the training of patch embedder g(x), we use an initial learning rate of 1e-5 with Adam #REFR optimizer for 10000 iterations with batch size being 100.", "text_before_citation": ["For Camelyon16, we tile the WSIs into 256\u00d7256 patches on 20\u00d7 magnification using the official code of #OTHEREFR , while for the HCC dataset the patches are 384\u00d7384 on 40\u00d7 magnification following the pathologists' advice.", "For both datasets, we use an ImageNet pre-trained ResNet50 7 to initialize g(x) (except DS-MIL #OTHEREFR ).", "The instance embedding process is the same of #OTHEREFR , which means for each patch, it will be firstly embedded into a 1024-dimension vector, and then be projected to a 512-dimension hidden space for further bag-level training.", "For the training of bag classifier f (x), we use an initial learning rate of 2e-4 with Adam #OTHEREFR optimizer for 200 epochs with batch size being 1.", "Camelyon16 results are reported on the official test split, while the HCC dataset uses a 7:1:2 split for training, validation and test."], "text_after_citation": ["Three metrics are used for evaluation, namely area under curve (AUC), F1 score, and slide-level accuracy (Acc).", "Experiments are all conducted on a Nvidia Tesla M40 (12GB)."], "citing_paper_content": {"title": "Iteratively Coupled Multiple Instance Learning From Instance To Bag Classifier For Whole Slide Image Classification", "abstract": "Whole Slide Image (WSI) classification remains a challenge due to their extremely high resolution and the absence of fine-grained labels. Presently, WSIs are usually classified as a Multiple Instance Learning (MIL) problem when only slide-level labels are available. MIL methods involve a patch embedding process and a bag-level classification process, but they are prohibitively expensive to be trained end-to-end. Therefore, existing methods usually train them separately, or directly skip the training of the embedder. Such schemes hinder the patch embedder's access to slide-level labels, resulting in inconsistencies within the entire MIL pipeline. To overcome this issue, we propose a novel framework called Iteratively Coupled MIL (ICMIL), which bridges the loss back-propagation process from the bag-level classifier to the patch embedder. In ICMIL, we use category information in the bag-level classifier to guide the patch-level fine-tuning of the patch feature extractor. The refined embedder then generates better instance representations for achieving a more accurate bag-level classifier. By coupling the patch embedder and bag classifier at a low cost, our proposed framework enables information exchange between the two processes, benefiting the entire MIL classification model. We tested our framework on two datasets using three different backbones, and our experimental results demonstrate consistent performance improvements over state-of-the-art MIL methods. Code will be made available upon acceptance."}, "cited_paper_content": {"title": "Adam: A Method For Stochastic Optimization", "abstract": "We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm."}, "keywords": ["patch embedder g(x", "Adam optimizer"], "citation_intent": "method"}
{"citing_id": "2303.11715v1", "cited_id": "1810.04805", "section_title": "I. Introduction", "citation": "Second, applying NLP methods to learn a question answering model usually requires a large-scale training set #REFR .", "text_before_citation": ["Question answering is an important topic in the field of natural language processing, and many works have been proposed recently #OTHEREFR .", "However, there are some challenges in using NLP algorithms for log question answering.", "First, there exists a domain shift between general natural language and log data.", "Log data includes domain-specific symbols, such as IP addresses and modular identifiers #OTHEREFR .", "General NLP techniques consider these symbols to be out-of-domain terms and replace them with a special token, however they are crucial for the log domain and cannot be ignored."], "text_after_citation": ["However, to the best of our knowledge, there is no public question answering dataset.", "It is difficult to implement an effective question answering system with limited data.", "In this work, we propose a question answering system for unstructured logs, namely LogQA, which aims to answer questions in the form of natural language based on large-scale unstructured log corpus.", "LogQA has two key components: Log Retriever and Log Reader. Log Retriever retrieves some relevant and helpful raw logs.", "The goal of Log Reader is to predict exact answers based on retrieved logs."], "citing_paper_content": {"title": "Logqa: Question Answering In Unstructured Logs", "abstract": "Modern systems produce a large volume of logs to record run-time status and events. System operators use these raw logs to track a system in order to obtain some useful information to diagnose system anomalies. One of the most important problems in this area is to help operators find the answers to log-based questions efficiently and user-friendly. In this work, we propose LogQA, which aims at answering logbased questions in the form of natural language based on largescale unstructured log corpora. Our system presents the answer to a question directly instead of returning a list of relevant snippets, thus offering better user-friendliness and efficiency. LogQA represents the first approach to solve question answering in lod domain. LogQA has two key components: Log Retriever and Log Reader. Log Retriever aims at retrieving relevant logs w.r.t. a given question, while Log Reader is responsible for inferring the final answer. Given the lack of a public dataset for log questing answering, we manually labelled a QA dataset of three open-source log corpus and will make them publicly available. We evaluated our proposed model on these datasets by comparing its performance with 6 other baseline methods. Our experimental results demonstrate that LogQA has outperformed other baseline methods."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["question answering model"], "citation_intent": "method"}
{"citing_id": "2303.01092v1", "cited_id": "1911.12580", "section_title": "Related Work", "citation": "Stable learning #REFR learns a set of global sample weights that could remove the confounding bias for all the potential treatments from data distribution.", "text_before_citation": ["Distribution shift in supervised learning.", "Distribution shift problem has been studied in many literature #OTHEREFR .", "Most works aim to learn a representation that performs well on different source domains simultaneously #OTHEREFR Mahajan et al., 2021; #OTHEREFR , following the idea of causal invariance #OTHEREFR Arjovsky et al., 2019) .", "Structural equation models are often assumed for theoretical analysis #OTHEREFR Mahajan et al., 2021) .", "Distributionally robust optimization optimizes a model's worst-case performance over some uncertainty set directly (Krueger et al., 2021; #OTHEREFR ."], "text_after_citation": ["Disentangled representation learning #OTHEREFR Tr\u00e4uble et al., 2021; #OTHEREFR aims to learn representations where distinct and informative factors of variations in data are separated.", "Theoretical understanding of contrastive learning.", "A number of recent works also aim to theoretically explain the success of contrastive learning in IID settings.", "One way to explain it is through the mutual information between positive samples #OTHEREFR . #OTHEREFR", "(2019) directly analyze the generalization of InfoNCE loss based on the assumption that positive samples are drawn from the same latent classes. In the same setting, #OTHEREFR"], "citing_paper_content": {"title": "", "abstract": "Self-Supervised Learning (SSL) is a paradigm that leverages unlabeled data for model training. Empirical studies show that SSL can achieve promising performance in distribution shift scenarios, where the downstream and training distributions differ. However, the theoretical understanding of its transferability remains limited. In this paper, we develop a theoretical framework to analyze the transferability of self-supervised contrastive learning, by investigating the impact of data augmentation on it. Our results reveal that the downstream performance of contrastive learning depends largely on the choice of data augmentation. Moreover, we show that contrastive learning fails to learn domain-invariant features, which limits its transferability. Based on these theoretical insights, we propose a novel method called Augmentation-robust Contrastive Learning (ArCL), which guarantees to learn domain-invariant features and can be easily integrated with existing contrastive learning algorithms. We conduct experiments on several datasets and show that ArCL significantly improves the transferability of contrastive learning. * Equal Contribution. This work was partially done when Xuyang was visiting Qing Yuan Research Institute."}, "cited_paper_content": {"title": "Stable Learning Via Sample Reweighting", "abstract": "We consider the problem of learning linear prediction models with model misspecification bias. In such case, the collinearity among input variables may inflate the error of parameter estimation, resulting in instability of prediction results when training and test distributions do not match. In this paper we theoretically analyze this fundamental problem and propose a sample reweighting method that reduces collinearity among input variables. Our method can be seen as a pretreatment of data to improve the condition of design matrix, and it can then be combined with any standard learning method for parameter estimation and variable selection. Empirical studies on both simulation and real datasets demonstrate the effectiveness of our method in terms of more stable performance across different distributed data."}, "keywords": ["Stable learning"], "citation_intent": "background"}
{"citing_id": "2304.05314v1", "cited_id": "1602.03786", "section_title": "A. Optimal Control Problems", "citation": "In case that the solution of Problem 2 does not exist, we can derive an alternative trajectory for the CAVs by solving a two-level optimization problem numerically that includes piecing together the constrained and unconstrained arcs until the solution does not violate any constraints #REFR .", "text_before_citation": ["By solving Problem 2, the optimal exit time t f i along with the optimal trajectory #OTHEREFR and control law #OTHEREFR are obtained for CAV-i for t \u2208 [t 0", "[a i , b i , c i , d i ] using", "i , t f i ]", ".", "If a feasible solution to Problem 2 exists, then the solution is a cubic polynomial that guarantees none of the constraints are activated."], "text_after_citation": [], "citing_paper_content": {"title": "Coordination For Connected Automated Vehicles At Merging Roadways In Mixed Traffic Environment", "abstract": "In this paper, we present a two-level optimal control framework to address motion coordination of connected automated vehicles (CAVs) in the presence of human-driven vehicles (HDVs) in merging scenarios. Our framework combines an unconstrained trajectory solution of a low-level energyoptimal control problem with an upper-level optimization problem that yields the minimum travel time for CAVs. We predict the future trajectories of the HDVs using Newell's car-following model. To handle potential deviations of HDVs' actual behavior from the one predicted, we provide a risktriggered re-planning mechanism for the CAVs based on timeto-conflict. The effectiveness of the proposed control framework is demonstrated via simulations with heterogeneous human driving behaviors and via experiments in a scaled environment."}, "cited_paper_content": {"title": "A Decentralized Energy-Optimal Control Framework For Connected Automated Vehicles At Signal-Free Intersections", "abstract": "We address the problem of optimally controlling connected and automated vehicles (CAVs) crossing an urban intersection without any explicit traffic signaling, so as to minimize energy consumption subject to a throughput maximization requirement. We show that the solution of the throughput maximization problem depends only on the hard safety constraints imposed on CAVs and its structure enables a decentralized optimal control problem formulation for energy minimization. We present a complete analytical solution of these decentralized problems and derive conditions under which feasible solutions satisfying all safety constraints always exist. The effectiveness of the proposed solution is illustrated through simulation which shows substantial dual benefits of the proposed decentralized framework by allowing CAVs to conserve momentum and fuel while also improving travel time."}, "keywords": ["alternative trajectory", "two-level optimization problem"], "citation_intent": "background"}
{"citing_id": "2304.00152v1", "cited_id": "1703.04977", "section_title": "Method", "citation": "Kendall and Gal #REFR use the negative log-likelihood of the prediction model as the loss function to be minimized in pixel-wise tasks.", "text_before_citation": ["The objective of our work is to jointly estimate the disparity and its uncertainty.", "An important benefit of this joint formulation is that the multi-task network learns to predict more accurate disparities than the standalone disparity estimator when the uncertainty subnetwork is added.", "Given a stereo image pair X = {x l , x r }, with image dimensions H \u00d7 W , and the corresponding ground truth disparity d, the predictiond of a stereo-matching network f \u03b8 can be represented asd = f \u03b8 (x l , x r ).", "For each pixel i, the error (i) of the prediction is calculated using the L1 loss."], "text_after_citation": ["We take the formulation a step further by requiring that the network generate a distribution of uncertainties that matches the distribution of errors.", "To this end, we propose to minimize the divergence D between the distributions of predicted uncertainty and actual disparity error.", "In the following subsections, we present aleatoric uncertainty estimation (Section 3.1), the proposed KL divergence loss (Section 3.2), our network architecture (Section 3.3), and the combined loss function (Section 3.4)."], "citing_paper_content": {"title": "Learning The Distribution Of Errors In Stereo Matching For Joint Disparity And Uncertainty Estimation", "abstract": "We present a new loss function for joint disparity and uncertainty estimation in deep stereo matching. Our work is motivated by the need for precise uncertainty estimates and the observation that multi-task learning often leads to improved performance in all tasks. We show that this can be achieved by requiring the distribution of uncertainty to match the distribution of disparity errors via a KL divergence term in the network's loss function. A differentiable soft-histogramming technique is used to approximate the distributions so that they can be used in the loss. We experimentally assess the effectiveness of our approach and observe significant improvements in both disparity and uncertainty prediction on large datasets. Our code is available at https://github.com/lly00412/SEDNet.git."}, "cited_paper_content": {"title": "What Uncertainties Do We Need In Bayesian Deep Learning For Computer Vision?", "abstract": "There are two major types of uncertainty one can model. Aleatoric uncertainty captures noise inherent in the observations. On the other hand, epistemic uncertainty accounts for uncertainty in the model - uncertainty which can be explained away given enough data. Traditionally it has been difficult to model epistemic uncertainty in computer vision, but with new Bayesian deep learning tools this is now possible. We study the benefits of modeling epistemic vs. aleatoric uncertainty in Bayesian deep learning models for vision tasks. For this we present a Bayesian deep learning framework combining input-dependent aleatoric uncertainty together with epistemic uncertainty. We study models under the framework with per-pixel semantic segmentation and depth regression tasks. Further, our explicit uncertainty formulation leads to new loss functions for these tasks, which can be interpreted as learned attenuation. This makes the loss more robust to noisy data, also giving new state-of-the-art results on segmentation and depth regression benchmarks."}, "keywords": ["pixel-wise tasks", "loss function"], "citation_intent": "method"}
{"citing_id": "2304.08211v1", "cited_id": "1908.00080", "section_title": "Flexible-Model Neural Accelerators", "citation": "For example, Google offers a low-cost and lowpower version of the Google TPU called EdgeTPU #REFR , which can run dedicated neural networks with 8-bit precision.", "text_before_citation": ["In this section we discus neural accelerators that can support multiple models without hardware changes.", "Efforts in accelerating the execution of TensorFlow Lite models using flexible hardware have focused on deploying systolic architectures for tensor operations to obtain very high throughput."], "text_after_citation": ["The systolic array size in the EdgeTPU has 64 x 64 multiply-add cells obtaining 4 TOPS at 480MHz, and it is much smaller than the TPU cloud configurations.", "Layers with floatingpoint precision will run on a external CPU that would act as the host for the EdgeTPU device.", "Xilinx has also focused on inference including support for TensorFlow, with the Xilinx DPU #OTHEREFR unit.", "It is composed of a register configuration unit, the data controller and convolution computing modules optimized for the FPGA hardware resources.", "The original hardware is specialized for convolutional neural networks, although alternative architectures and hardware configurations are made available for other model types."], "citing_paper_content": {"title": "Dynamically Reconfigurable Variable-Precision Sparse-Dense Matrix Acceleration In Tensorflow Lite", "abstract": "In this paper, we present a dynamically reconfigurable hardware accelerator called FADES (Fused Architecture for DEnse and Sparse matrices). The FADES design offers multiple configuration options that trade off parallelism and complexity using a dataflow model to create four stages that read, compute, scale and write results. FADES is mapped to the programmable logic (PL) and integrated with the TensorFlow Lite inference engine running on the processing system (PS) of a heterogeneous SoC device. The accelerator is used to compute the tensor operations, while the dynamically reconfigurable approach can be used to switch precision between int8 and float modes. This dynamic reconfiguration enables better performance by allowing more cores to be mapped to the resource-constrained device and lower power consumption compared with supporting both arithmetic precisions simultaneously. We compare the proposed hardware with a high-performance systolic architecture for dense matrices obtaining 25% better performance in dense mode with half the DSP blocks in the same technology. In sparse mode, we show that the core can outperform dense mode even at low sparsity levels, and a single-core achieves up to 20x acceleration over the software-optimized NEON RUY library."}, "cited_paper_content": {"title": "Machine Learning At The Network Edge: A Survey", "abstract": "Devices comprising the Internet of Things, such as sensors and small cameras, usually have small memories and limited computational power. The proliferation of such resource-constrained devices in recent years has led to the generation of large quantities of data. These data-producing devices are appealing targets for machine learning applications but struggle to run machine learning algorithms due to their limited computing capability. They typically offload data to external computing systems (such as cloud servers) for further processing. The results of the machine learning computations are communicated back to the resource-scarce devices, but this worsens latency, leads to increased communication costs, and adds to privacy concerns. Therefore, efforts have been made to place additional computing devices at the edge of the network, i.e close to the IoT devices where the data is generated. Deploying machine learning systems on such edge devices alleviates the above issues by allowing computations to be performed close to the data sources. This survey describes major research efforts where machine learning has been deployed at the edge of computer networks."}, "keywords": ["dedicated neural networks"], "citation_intent": "method"}
{"citing_id": "2304.01534v1", "cited_id": "1608.03983", "section_title": "Experiment Setup", "citation": "The image features are then encoded at different spatial resolutions, resulting in tensors of shape #REFR .", "text_before_citation": ["In practical applications, autonomous software suppliers often want to upgrade their models using existing customer data without accessing their customers' data directly.", "UC 3 represented a typical way of using public traffic data collected in vehicular networks for training, which often involves a large number of clients with small amounts of data and significant differences between datasets.", "Lastly, UC 4 focused on federated learning for clients with different numbers of cameras, which can happen when data is collected from different vehicle models by different manufacturers. BEVT architecture.", "We begin by feeding our input images X i,k \u2208 R L k \u00d7H\u00d7W \u00d73 through a 3-layer ResNet34 encoder.", "To ensure consistency across inputs, we use the AMCM to resize all inputs to have L k = 4."], "text_after_citation": ["Each client is locally trained for one epoch with a batch size of 4 using AdamW [30] optimizer. Baselines.", "Given the limited research on federated learning on BEVT, we conducted the first trial of training fedAvg on this platform.", "However, to validate the effectiveness of FedCaP, we also incorporated recent research findings on federated transformer learning as a baseline on BEVT.", "In addition to showcasing the results of local training on each client, we compared FedCaP with the following baselines:", "R L k \u00d764\u00d764\u00d7128 , R L k \u00d732\u00d732\u00d7256 , R L k \u00d716\u00d716\u00d7512 ."], "citing_paper_content": {"title": "Fedbevt: Federated Learning Bird'S Eye View Perception Transformer In Road Traffic Systems", "abstract": "Bird's eye view (BEV) perception is becoming increasingly important in the field of autonomous driving. It uses multi-view camera data to learn a transformer model that directly projects the perception of the road environment onto the BEV perspective. However, training a transformer model often requires a large amount of data, and as camera data for road traffic is often private, it is typically not shared. Federated learning offers a solution that enables clients to collaborate and train models without exchanging data. In this paper, we propose FedBEVT, a federated transformer learning approach for BEV perception. We address two common data heterogeneity issues in FedBEVT: (i) diverse sensor poses, and (ii) varying sensor numbers in perception systems. We present federated learning with camera-attentive personalization (FedCaP) and adaptive multi-camera masking (AMCM) to enhance the performance in real-world scenarios. To evaluate our method in real-world settings, we create a dataset consisting of four typical federated use cases. Our findings suggest that FedBEVT outperforms the baseline approaches in all four use cases, demonstrating the potential of our approach for improving BEV perception in autonomous driving. We will make all codes and data publicly available."}, "cited_paper_content": {"title": "Sgdr: Stochastic Gradient Descent With Warm Restarts", "abstract": "Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets, where we demonstrate new state-of-the-art results at 3.14% and 16.21%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at https://github.com/loshchil/SGDR"}, "keywords": ["image features"], "citation_intent": "background"}
{"citing_id": "2303.06808v1", "cited_id": "1910.07517", "section_title": "Rq2: Can Existing Data Augmentation Methods Produce Robust Code Models?", "citation": "This phenomenon is consistent with the conclusion drawn by previous work #REFR that simply increasing the training data is not sufficient for improving the robustness of code models.", "text_before_citation": ["Adversarial robustness reflects how models handle the data with noise, which is an important characteristic that should be evaluated by the models.", "Table 5 presents the attack success rate of two state-of-the-art attack methods on our trained models.", "First, the results demonstrate that data augmentation can not always enhance the robustness of models.", "Compared to the No Aug models, only four out of eight models trained by using data augmentation have higher robustness."], "text_after_citation": ["Also importantly, in some cases, although data augmentation can help train a more robust model, the robustness improvement is insignificant, e.g., the greatest improvement is 5.98% (GraphCodeBERT-Refactor-BigCloneBench-ALERT). Then, we compare each data augmentation method.", "In CodeBERT, RS performs the best and has a relatively better robustness improvement in five (out of 12) cases, and reduces ASR by up to 5.10% under MHM attack and 3.69% under ALERT attack compared to No Aug.", "Besides, the second best one, RD reduces ASR by up to 5.67% in MHM and 2.66% in ALERT compared to No Aug.", "In GraphCodeBERT, RS still is the best choice for robustness improvement in five (out of 12) cases, which deduces ASR by up to 4.10% in MHM and 3.02% in ALERT compared to No Aug.", "Interestingly, based on pre-trained PL models, compared to Refactor, methods like RS that could sightly break the syntax of programs can produce more accurate and robust code models to solve downstream tasks."], "citing_paper_content": {"title": "Boosting Source Code Learning With Data Augmentation: An Empirical Study", "abstract": "The next era of program understanding is being propelled by the use of machine learning to solve software problems. Recent studies have shown surprising results of source code learning, which applies deep neural networks (DNNs) to various critical software tasks, e.g., bug detection and clone detection. This success can be greatly attributed to the utilization of massive high-quality training data, and in practice, data augmentation, which is a technique used to produce additional training data, has been widely adopted in various domains, such as computer vision. However, in source code learning, data augmentation has not been extensively studied, and existing practice is limited to simple syntax-preserved methods, such as code refactoring. Essentially, source code is often represented in two ways, namely, sequentially as text data and structurally as graph data, when it is used as training data in source code learning. Inspired by these analogy relations, we take an early step to investigate whether data augmentation methods that are originally used for text and graphs are effective in improving the training quality of source code learning. To that end, we first collect and categorize data augmentation methods in the literature. Second, we conduct a comprehensive empirical study on four critical tasks and 11 DNN architectures to explore the effectiveness of 12 data augmentation methods (including code refactoring and 11 other methods for text and graph data). Our results identify the data augmentation methods that can produce more accurate and robust models for source code learning, including those based on mixup (e.g., SenMixup for texts and Manifold-Mixup for graphs), and those that slightly break the syntax of source code (e.g., random swap and random deletion for texts)."}, "cited_paper_content": {"title": "Adversarial Examples For Models Of Code", "abstract": "Neural models of code have shown impressive performance for tasks such as predicting method names and identifying certain kinds of bugs. In this paper, we show that these models are vulnerable to adversarial examples, and introduce a novel approach for attacking trained models of code with adversarial examples. The main idea is to force a given trained model to make an incorrect prediction as specified by the adversary by introducing small perturbations that do not change the program's semantics. To find such perturbations, we present a new technique for Discrete Adversarial Manipulation of Programs (DAMP). DAMP works by deriving the desired prediction with respect to the model's inputs while holding the model weights constant and following the gradients to slightly modify the code. ::: To defend a model against such attacks, we propose placing a defensive model (Anti-DAMP) in front of it. Anti-DAMP detects unlikely mutations and masks them before feeding the input to the downstream model. ::: We show that our DAMP attack is effective across three neural architectures: code2vec, GGNN, and GNN-FiLM, in both Java and C#. We show that DAMP has up to 89% success rate in changing a prediction to the adversary's choice (\"targeted attack\"), and a success rate of up to 94% in changing a given prediction to any incorrect prediction (\"non-targeted attack\"). By using Anti-DAMP, the success rate of the attack drops drastically for both targeted and non-targeted attacks, with a minor penalty of 2% relative degradation in accuracy while not performing under attack."}, "keywords": ["code models"], "citation_intent": "result"}
{"citing_id": "2304.13958v1", "cited_id": "1409.0473", "section_title": "Learning And Reasoning Poverty Estimation", "citation": "The data merging can be done in various ways, such as early fusion, late fusion, naive concatenation, attention mechanism #REFR , or gating-based fusion.", "text_before_citation": ["However, as we have limited training data, training visual or textual embeddings from scratch will, in all likelihood, not yield desirable results.", "Thus, we propose to use a transfer learning approach wherein we use pre-trained models trained on large corpora and fine-tune them to our task-specific dataset.", "For the visual inputs of satellite images, we can utilize pre-trained CNN-based models, such as VGG Net [Simonyan and Zisserman, 2014], or ResNet #OTHEREFR .", "For the textual inputs, we prefer the language models, such as BERT [Devlin et al., 2019] or RoBERTa #OTHEREFR .", "Each of the representations goes through a non-linear transformation before the data merging operation."], "text_after_citation": ["We treat our problem as a classification task wherein we have four ground-truth labels at the district level -'advanced', 'catching up', 'falling behind', and 'lagged'.", "This is calculated based on aggregating the individual MDPI scores at the district level.", "Correlation, Association, and Causality Analysing the relationship of the input variables with the target output can be a challenging step but it is important for strategic actions.", "Such insights are important to determine the driving factors (causative), factors that exhibit linear relationships (correlative), and factors that co-occur (associa- Figure 2 : Proposed multi-input deep learning model for aggregating and processing data from proxy and traditional data sources. tive).", "We propose to use the Bayesian Networks for identifying causative factors, Pearson's correlation to determine the correlative factors, and the Hypergeometric test to discover the associative factors."], "citing_paper_content": {"title": "Learning And Reasoning Multifaceted And Longitudinal Data For Poverty Estimates And Livelihood Capabilities Of Lagged Regions In Rural India", "abstract": "Poverty is a multifaceted phenomenon linked to the lack of capabilities of households to earn a sustainable livelihood, increasingly being assessed using multidimensional indicators. Its spatial pattern depends on social, economic, political, and regional variables. Artificial intelligence has shown immense scope in analyzing the complexities and nuances of poverty. The proposed project aims to examine the poverty situation of rural India for the period of 1990-2022 based on the quality of life and livelihood indicators. The districts will be classified into 'advanced', 'catching up', 'falling behind', and 'lagged' regions. The project proposes to integrate multiple data sources, including conventional national-level large sample household surveys, census surveys, and proxy variables like daytime, and nighttime data from satellite images, and communication networks, to name a few, to provide a comprehensive view of poverty at the district level. The project also intends to examine causation and longitudinal analysis to examine the reasons for poverty. Poverty and inequality could be widening in developing countries due to demographic and growth-agglomerating policies. Therefore, targeting the lagging regions and the vulnerable population is essential to eradicate poverty and improve the quality of life to achieve the goal of 'zero poverty'. Thus, the study also focuses on the districts with a higher share of the marginal section of the population compared to the national average to trace the performance of development indicators and their association with poverty in these regions."}, "cited_paper_content": {"title": "Neural Machine Translation By Jointly Learning To Align And Translate", "abstract": "Neural machine translation is a recently proposed approach to machine translation. Unlike the traditional statistical machine translation, the neural machine translation aims at building a single neural network that can be jointly tuned to maximize the translation performance. The models proposed recently for neural machine translation often belong to a family of encoder-decoders and consists of an encoder that encodes a source sentence into a fixed-length vector from which a decoder generates a translation. In this paper, we conjecture that the use of a fixed-length vector is a bottleneck in improving the performance of this basic encoder-decoder architecture, and propose to extend this by allowing a model to automatically (soft-)search for parts of a source sentence that are relevant to predicting a target word, without having to form these parts as a hard segment explicitly. With this new approach, we achieve a translation performance comparable to the existing state-of-the-art phrase-based system on the task of English-to-French translation. Furthermore, qualitative analysis reveals that the (soft-)alignments found by the model agree well with our intuition."}, "keywords": ["data merging", "naive concatenation"], "citation_intent": "method"}
{"citing_id": "2304.02541v1", "cited_id": "1801.09536", "section_title": "Introduction", "citation": "In contrast to semantic word embeddings #REFR , we show that intrinsic and extrinsic metrics for phonetic word embeddings generally correlate with each other.", "text_before_citation": ["They range from intuitive baselines to more complex techniques using metric and contrastive learning.", "More importantly, however, we include an evaluation suite for testing the performance of phonetic embeddings. The motivations for this are two-fold. First, prior works are inconsistent in evaluating their models.", "This prevents the field from observing long-term improvements of such embeddings and from making fair comparisons across different approaches.", "Secondly, when a practitioner is deciding which phonetic word embedding method to use, the go-to approach is to first apply the embeddings (generally fast) and then train a downstream model on those embeddings (compute and time intensive).", "Instead, intrinsic embedding evaluation metrics (cheap)-if shown to correlate well with extrinsic metrics-could provide useful signals in embedding method selection prior to training of downstream models (expensive)."], "text_after_citation": ["While some work on evaluating acoustic word embeddings exists #OTHEREFR , this work specializes in phonetic word embeddings for text, not speech.", "Our contributions are threefold:", "\u2022 a survey of existing phonetic embeddings, \u2022 four novel methods for phonetic word embedding, ranging from simple baselines to complex models, and \u2022 an evaluation suite for such embeddings."], "citing_paper_content": {"title": "Pwesuite: Phonetic Word Embeddings And Tasks They Facilitate", "abstract": "Word embeddings that map words into a fixeddimensional vector space are the backbone of modern NLP. Most word embedding methods encode semantic information. However, phonetic information, which is important for some tasks, is often overlooked. In this work, we develop several novel methods which leverage articulatory features to build phonetically informed word embeddings, and present a set of phonetic word embeddings to encourage their community development, evaluation and use. While several methods for learning phonetic word embeddings already exist, there is a lack of consistency in evaluating their effectiveness. Thus, we also proposes several ways to evaluate both intrinsic aspects of phonetic word embeddings, such as word retrieval and correlation with sound similarity, and extrinsic performances, such as rhyme and cognate detection and sound analogies. We hope that our suite of tasks will promote reproducibility and provide direction for future research on phonetic word embeddings."}, "cited_paper_content": {"title": "A Survey Of Word Embeddings Evaluation Methods", "abstract": "Word embeddings are real-valued word representations able to capture lexical semantics and trained on natural language corpora. Models proposing these representations have gained popularity in the recent years, but the issue of the most adequate evaluation method still remains open. This paper presents an extensive overview of the field of word embeddings evaluation, highlighting main problems and proposing a typology of approaches to evaluation, summarizing 16 intrinsic methods and 12 extrinsic methods. I describe both widely-used and experimental methods, systematize information about evaluation datasets and discuss some key challenges."}, "keywords": ["phonetic word embeddings", "semantic word embeddings"], "citation_intent": "result"}
{"citing_id": "2303.14733v1", "cited_id": "1408.1588", "section_title": "Introduction", "citation": "Also, the author in #REFR considers the synchronization of multi-dimensional mechanical and electrical systems where the subsystems interact via positive semidefinite matrix weights.", "text_before_citation": ["Corresponding to a matrix-weighted consensus system, one may define a matrix-weighted graph and a matrix-weighted Laplacian.", "It has been shown that algebraic graph properties of the matrix-weighted Laplacian determine the consensus and clustering behaviors of the whole system #OTHEREFR .", "Several applications of the matrix-weighted consensus algorithm can be found in the literature.", "For example, the authors in #OTHEREFR propose multi-dimensional opinion dynamics, where the in-teractions and inter-logical dependencies between several considered topics are captured by matrix weights, respectively.", "In bearing-based formation control and network localization #OTHEREFR , the interactions between neighboring agents are modeled by an orthogonal projection matrix obtained from the directional (or bearing) vector between neighboring agents."], "text_after_citation": ["It is noteworthy that most existing works on matrixweighted consensus assumed that the agents update their states synchronously or following a deterministic update sequence.", "Continuous-time matrix-weighted consensus with switching graph topologies was studied in #OTHEREFR .", "Discrete-time matrix-weighted consensus with fixed or switching topologies was studied in #OTHEREFR .", "Matrixweighted consensus with hybrid continuous-discrete time updates was examined in #OTHEREFR .", "The randomized consensus algorithms, which refer to a family of consensus algorithms in which the edges between agents are randomly selected according to some stochastic model for every discrete instant, have received a lot of attention in the literature #OTHEREFR ."], "citing_paper_content": {"title": "Randomized Matrix Weighted Consensus \u22c6", "abstract": "In this paper, a randomized gossip-type matrix-weighted consensus algorithm is proposed for both leaderless and leaderfollower topologies. Under some mild assumptions, the proposed pairwise asynchronous update algorithm achieves a consensus in expectation. Moreover, the probability distribution, the weighting matrices, and the updating step size jointly determine the upper bound of the \u01eb-convergence time of the algorithm. The theoretical result is verified by several simulation examples."}, "cited_paper_content": {"title": "Synchronization Under Matrix-Weighted Laplacian", "abstract": "Synchronization in a group of linear time-invariant systems is studied where the coupling between each pair of systems is characterized by a different output matrix. Simple methods are proposed to generate a (separate) linear coupling gain for each pair of systems, which ensures that all the solutions converge to a common trajectory."}, "keywords": ["positive semidefinite matrix", "synchronization"], "citation_intent": "background"}
{"citing_id": "2303.01064v1", "cited_id": "1906.02192", "section_title": "Evaluation Results And Discussion", "citation": "For example, the record in validation dataset with 'celex_id' of '31999L0010' extracts 'foodstuff' but gets 'foodstuff' and 'processed foodstuff' with a reordered categories list. #REFR Ko et al.", "text_before_citation": ["And Legal-BERT-FP and BERT-base-uncased gain competitive f-scores in validation and testing dataset.", "At epoch 5, Legal-BERT-FP performs slightly better than BERT in validation and slightly worse in testing.", "The extracted answers are not independent of the positions of the category in categories list.", "The Table 6 : validation and testing results (%) in the concept multi-label classification for multi-answer questioning downstream task with proposed classification metric extracted answers might be different if the same title and same domain-tree randomly reorders the categories in the categories list.", "However, in classification, the order of the categories is not important."], "text_after_citation": ["(2020) also mentions the position bias in extractive question answering.", "The other issue in the proposed multi-label classification task is the long input problem since BERT has limited input length.", "Table 7 shows validation and testing in domain multi-label classification with the downstream task utilizing the multi-answer questioning task proposed in methodology.", "The domain multi-label classification has the constant domain label group which is different from concept multi-label classification having various lengths of label groups.", "The evaluation metrics include training loss, validation loss, and overall metrics -precision, recall, f1 score and accuracy in seqeval."], "citing_paper_content": {"title": "Adopting The Multi-Answer Questioning Task With An Auxiliary Metric For Extreme Multi-Label Text Classification Utilizing The Label Hierarchy", "abstract": "Extreme multi-label text classification utilizes the label hierarchy to partition extreme labels into multiple label groups, turning the task into simple multi-group multi-label classification tasks. Current research encodes labels as a vector with fixed length which needs establish multiple classifiers for different label groups. The problem is how to build only one classifier without sacrificing the label relationship in the hierarchy. This paper adopts the multi-answer questioning task for extreme multi-label classification. This paper also proposes an auxiliary classification evaluation metric. This study adopts the proposed method and the evaluation metric to the legal domain. The utilization of legal Berts and the study on task distribution are discussed. The experiment results show that the proposed hierarchy and multianswer questioning task can do extreme multilabel classification for EURLEX dataset. And in minor/fine-tuning the multi-label classification task, the domain adapted BERT models could not show apparent advantages in this experiment. The method is also theoretically applicable to zero-shot learning. * corresponding author 1 https://analyticsindiamag.com/ what-is-extreme-multilabel-text-classification/ 4 sklearn.preprocessing. MultiLabelBinarizer\\T1\\ textemdashscikit-learn1.0.2documentation 5 https:"}, "cited_paper_content": {"title": "Large-Scale Multi-Label Text Classification On Eu Legislation", "abstract": "We consider Large-Scale Multi-Label Text Classification (LMTC) in the legal domain. We release a new dataset of 57k legislative documents from EURLEX, annotated with ~4.3k EUROVOC labels, which is suitable for LMTC, few- and zero-shot learning. Experimenting with several neural classifiers, we show that BIGRUs with label-wise attention perform better than other current state of the art methods. Domain-specific WORD2VEC and context-sensitive ELMO embeddings further improve performance. We also find that considering only particular zones of the documents is sufficient. This allows us to bypass BERT's maximum text length limit and fine-tune BERT, obtaining the best results in all but zero-shot learning cases."}, "keywords": ["reordered categories list", "example"], "citation_intent": "background"}
{"citing_id": "2303.11954v1", "cited_id": "1906.01537", "section_title": "Bo For Function Composition", "citation": "Astudillo and Frazier #REFR model the constituent functions of the composition by a single multioutput function f(x) = (f 1 (x), . . .", "text_before_citation": ["(\u2022), K (n) (\u2022, \u2022)) where \u00b5 (n) (\u2022)", "is the posterior mean function and K (n) (\u2022, \u2022) is the posterior covariance function, see #OTHEREFR for more details.", "The acquisition function then uses this posterior model to identify the next query location x n+1 .", "In doing so, vanilla BO ignores the values of the member functions in the composition h.", "BO for composite function, on the other hand, takes advantage of the available information about h, and its easy-to-compute nature."], "text_after_citation": [", f M (x)) and then model the uncertainty in f(x) using a multi-output Gaussian process to optimize h(f(x)).", "Since the prior over f is modelled as a MOGP, the proposed method tries to capture the correlations between different components of the multi-output function f(x).", "Note that the proposed EI and PI-based acquisition functions are required to be computed using Monte Carlo sampling.", "Furthermore, a sample from the posterior distribution is obtained by first sampling an n variate normal distribution, then scaling it by the lower Cholesky factor and then centering it with the mean of the posterior GP. Two problems arise due to this: 1.", "Such simulation based averaging approach increases the time complexity of the procedure linearly with the number of samples taken for averaging and 2."], "citing_paper_content": {"title": "Bayesian Optimization For Function Compositions With Applications To Dynamic Pricing", "abstract": "Bayesian Optimization (BO) is used to find the global optima of black box functions. In this work, we propose a practical BO method of function compositions where the form of the composition is known but the constituent functions are expensive to evaluate. By assuming an independent Gaussian process (GP) model for each of the constituent black-box function, we propose EI and UCB based BO algorithms and demonstrate their ability to outperform vanilla BO and the current state-of-art algorithms. We demonstrate a novel application of the proposed methods to dynamic pricing in revenue management when the underlying demand function is expensive to evaluate."}, "cited_paper_content": {"title": "Bayesian Optimization Of Composite Functions", "abstract": "We consider optimization of composite objective functions, i.e., of the form $f(x)=g(h(x))$, where $h$ is a black-box derivative-free expensive-to-evaluate function with vector-valued outputs, and $g$ is a cheap-to-evaluate real-valued function. While these problems can be solved with standard Bayesian optimization, we propose a novel approach that exploits the composite structure of the objective function to substantially improve sampling efficiency. Our approach models $h$ using a multi-output Gaussian process and chooses where to sample using the expected improvement evaluated on the implied non-Gaussian posterior on $f$, which we call expected improvement for composite functions (\\ei). Although \\ei\\ cannot be computed in closed form, we provide a novel stochastic gradient estimator that allows its efficient maximization. We also show that our approach is asymptotically consistent, i.e., that it recovers a globally optimal solution as sampling effort grows to infinity, generalizing previous convergence results for classical expected improvement. Numerical experiments show that our approach dramatically outperforms standard Bayesian optimization benchmarks, reducing simple regret by several orders of magnitude."}, "keywords": ["constituent functions", "single multioutput function"], "citation_intent": "background"}
{"citing_id": "2303.16809v1", "cited_id": "1311.2037", "section_title": "Vi. Conclusion", "citation": "Though the main benefit may be further reduction in overall communication cost, it is not clear whether an advantage over pairwise approaches can be achieved when an average pairwise intersection is large compared to the total intersection (\u2229 i S i ) #REFR .", "text_before_citation": ["As a result, it is easier to formally analyze its performance, and, indeed, we have shown that it completes in time bounded by the network diameter (or logarithmic in network size for the \"small-world\" networks that reasonably model blockchain networks).", "We have also validated our analytical findings against a novel event-based simulator that we have developed.", "We run the simulator on real-world transaction pool statistics drawn from our own measurement campaign.", "In our simulations, SREP incurs only tens of gigabytes of overall bandwidth overhead to synchronize networks with ten thousand nodes, which is several times better than the current approach in the literature.", "For future work, we propose to consider multi-party set reconciliation #OTHEREFR , #OTHEREFR in the context of transaction pool sync."], "text_after_citation": [], "citing_paper_content": {"title": "Srep: Out-Of-Band Sync Of Transaction Pools For Large-Scale Blockchains", "abstract": "Synchronization of transaction pools (mempools) has shown potential for improving the performance and block propagation delay of state-of-the-art blockchains. Indeed, various heuristics have been proposed in the literature to this end, all of which incorporate exchanges of unconfirmed transactions into their block propagation protocol. In this work, we take a different approach, maintaining transaction synchronization outside (and independently) of the block propagation channel. In the process, we formalize the synchronization problem within a graph theoretic framework and introduce a novel algorithm (SREP-Set Reconciliation-Enhanced Propagation) with quantifiable guarantees. We analyze the algorithm's performance for various realistic network topologies, and show that it converges on any connected graph in a number of steps that is bounded by the diameter of the graph. We confirm our analytical findings through extensive simulations that include comparison with MempoolSync, a recent approach from the literature. Our simulations show that SREP incurs reasonable overall bandwidth overhead and, unlike MempoolSync, scales gracefully with the size of the network."}, "cited_paper_content": {"title": "Simple Multi-Party Set Reconciliation", "abstract": "As users migrate information to cloud storage, many distributed cloud-based services use multiple loosely consistent replicas of user information to avoid the high overhead of more tightly coupled synchronization. Periodically, the information must be synchronized, or reconciled. One can place this problem in the theoretical framework of {\\em set reconciliation}: two parties $A_1$ and $A_2$ each hold a set of keys, named $S_1$ and $S_2$ respectively, and the goal is for both parties to obtain $S_1 \\cup S_2$. Typically, set reconciliation is interesting algorithmically when sets are large but the set difference $|S_1-S_2|+|S_2-S_1|$ is small. In this setting the focus is on accomplishing reconciliation efficiently in terms of communication; ideally, the communication should depend on the size of the set difference, and not on the size of the sets. ::: In this paper, we extend recent approaches using Invertible Bloom Lookup Tables (IBLTs) for set reconciliation to the multi-party setting. In this setting there are three or more parties $A_1,A_2,\\ldots,A_n$ holding sets of keys $S_1,S_2,\\ldots,S_n$ respectively, and the goal is for all parties to obtain $\\cup_i S_i$. This could of course be done by pairwise reconciliations, but we seek more effective methods. ::: Our methodology uses network coding techniques in conjunction with IBLTs, allowing efficiency in network utilization along with efficiency obtained by passing messages of size $O(|\\cup_i S_i - \\cap_i S_i|)$. Further, our approach can function even if the number of parties is not exactly known in advance, and in many cases can be used to determine which parties contain keys not in the joint union. By connecting reconciliation with network coding, we can allow for substantially more efficient reconciliation methods that apply to a number of natural distributed computing problems."}, "keywords": ["overall communication cost", "average pairwise intersection"], "citation_intent": "background"}
{"citing_id": "2304.14271v1", "cited_id": "1712.01887", "section_title": "D. Federated Learning And Autonomous Driving", "citation": "To further overcome this performance gap, authors in #REFR proposed modifications to the existing approach through deep gradient compression.", "text_before_citation": ["However, the challenge is to choose the favorable lower-limit value, as similar to soft-filter pruning, the quantization and selection of the wrong lower-limit value can directly impact the overall model aggregation, which may provide an overall reduced model size but decreases the accuracy.", "To overcome the previous challenge, stochastic gradient descent with k-sparsification is proposed in #OTHEREFR , by reducing the data and model size and also improving convergence through error compensation for the transmission taking place between edge and server.", "A similar approach is used in #OTHEREFR , the method proposes to fix the sparsity rate.", "The communication or transmission of the gradient is only enabled for a fraction of the gradient with the highest magnitudes and keeping the unused gradient in the container.", "The sparsity rate used by the authors is p = 0.001, and this approach has relatively less impact the learned model overall accuracy and performance."], "text_after_citation": ["Deep gradient compression uses approaches such as: momentum correction, local gradient clipping, for the convolutional neural network and recurrent neural network.", "Results show that gradients are compressed by ratio of 270-660 following a hierarchical approach, without slowing down the model convergence.", "Sparsification methods were initially proposed with the function of improving and promoting distributed and parallel training among the cloud and data-centers.", "However, these methods lacked model convergence and aggregation as a scope which is currently a most essential metrics for the federated and distributed machine learning.", "Similarly, attention should be given to the number of edge devices participating in the transmission and the server participating in collaborative training."], "citing_paper_content": {"title": "A Survey On Approximate Edge Ai For Energy Efficient Autonomous Driving Services", "abstract": "Autonomous driving services rely heavily on sensors such as cameras, LiDAR, radar, and communication modules. A common practice of processing the sensed data is using a high-performance computing unit placed inside the vehicle, which deploys AI models and algorithms to act as the brain or administrator of the vehicle. The vehicular data generated from average hours of driving can be up to 20 Terabytes depending on the data rate and specification of the sensors. Given the scale and fast growth of services for autonomous driving, it is essential to improve the overall energy and environmental efficiency, especially in the trend towards vehicular electrification (e.g., battery-powered). Although the areas have seen significant advancements in sensor technologies, wireless communications, computing and AI/ML algorithms, the challenge still exists in how to apply and integrate those technology innovations to achieve energy efficiency. This survey reviews and compares the connected vehicular applications, vehicular communications, approximation and Edge AI techniques. The focus is on energy efficiency by covering newly proposed approximation and enabling frameworks. To the best of our knowledge, this survey is the first to review the latest approximate Edge AI frameworks and publicly available datasets in energy-efficient autonomous driving. The insights and vision from this survey can be beneficial for the collaborative driving service development on low-power and memory-constrained systems and also for the energy optimization of autonomous vehicles."}, "cited_paper_content": {"title": "Deep Gradient Compression: Reducing The Communication Bandwidth For Distributed Training", "abstract": "Large-scale distributed training requires significant communication bandwidth for gradient exchange that limits the scalability of multi-node training, and requires expensive high-bandwidth network infrastructure. The situation gets even worse with distributed training on mobile devices (federated learning), which suffers from higher latency, lower throughput, and intermittent poor connections. In this paper, we find 99.9% of the gradient exchange in distributed SGD is redundant, and propose Deep Gradient Compression (DGC) to greatly reduce the communication bandwidth. To preserve accuracy during compression, DGC employs four methods: momentum correction, local gradient clipping, momentum factor masking, and warm-up training. We have applied Deep Gradient Compression to image classification, speech recognition, and language modeling with multiple datasets including Cifar10, ImageNet, Penn Treebank, and Librispeech Corpus. On these scenarios, Deep Gradient Compression achieves a gradient compression ratio from 270x to 600x without losing accuracy, cutting the gradient size of ResNet-50 from 97MB to 0.35MB, and for DeepSpeech from 488MB to 0.74MB. Deep gradient compression enables large-scale distributed training on inexpensive commodity 1Gbps Ethernet and facilitates distributed training on mobile."}, "keywords": ["deep gradient compression"], "citation_intent": "method"}
{"citing_id": "2305.02414v1", "cited_id": "1512.04385", "section_title": "Introduction", "citation": "Combining this with #REFR After rearranging terms, we conclude that Lemma 3.1 holds when G is connected. Suppose G is not connected.", "text_before_citation": ["Observe that no edge on the boundary of a face of length three is also in a face of length four, or else a triangle and 4-cycle are adjacent.", "Because G is not a triangle and no two triangles are adjacent, no two length-3 faces share an edge.", "Therefore, we conclude that at least 3|X| edges of G are on the boundary of a face in X.", "Moreover, every edge on the boundary of a length-4 face is on the boundary of at most two length-4 faces, so at least 2|Y | edges of G are on the boundary of a face in Y .", "Thus, we have |E(G)| \u2265 3|X| + 2|Y |, which implies #OTHEREFR ."], "text_after_citation": ["Let J be the set of components of G with less than four vertices, and let K be the set of components of G with at least four vertices.", "Note that for all j \u2208 J, |E(j)| \u2264 |V (j)|.", "Therefore, if |K| \u2265 1, Lemma 3.1 holds by induction on |J|. Otherwise, |K| = 0.", "Note that n \u2264 15(n \u2212 2)/7 if n \u2265 4, and in this case |E(G)| \u2264 |V (G)|.", "Therefore, if the total number of vertices in the components in J is at least four, Lemma 3.1 holds."], "citing_paper_content": {"title": "The Independence Ratio Of 4-Cycle-Free Planar Graphs", "abstract": "We prove that every n-vertex planar graph G with no triangle sharing an edge with a 4-cycle has independence ratio n/\u03b1(G) \u2264 4 \u2212 \u03b5 for \u03b5 = 1/30. This result implies that the same bound holds for 4-cycle-free planar graphs and planar graphs with no adjacent triangles and no triangle sharing an edge with a 5-cycle. For the latter case we strengthen the bound to \u03b5 = 2/9."}, "cited_paper_content": {"title": "Extremal C4-Free/C5-Free Planar Graphs", "abstract": "We study the topic of \"extremal\" planar graphs, defining $\\mathrm{ex_{_{\\mathcal{P}}}}(n,H)$ to be the maximum number of edges possible in a planar graph on $n$ vertices that does not contain a given graph $H$ as a subgraph. In particular,we examine the case when $H$ is a small cycle,obtaining $\\mathrm{ex_{_{\\mathcal{P}}}}(n,C_{4}) \\leq \\frac{15}{7}(n-2)$ for all $n \\geq 4$ and $\\mathrm{ex_{_{\\mathcal{P}}}}(n,C_{5}) \\leq \\frac{12n-33}{5}$ for all $n \\geq 11$, and showing that both of these bounds are tight."}, "keywords": ["Lemma", "G"], "citation_intent": "background"}
{"citing_id": "2303.10945v1", "cited_id": "2003.12267", "section_title": "Related Work", "citation": "ADGAN #REFR ameliorates this issue by extracting expressive textures vectors from different semantic entities to synthesize the target image.", "text_before_citation": ["(a) To investigate the domain gap between source and OOD datasets, we obtain the high-level feature by a person ReID model #OTHEREFR and use t-SNE for visualization. (b) The domain generalization of NTED #OTHEREFR .", "Typical pose transfer model performs reasonably well on the source domain, however, the generated results could be easily violated with OOD input.", "We first propose an Open-World Pose Transfer (OWPT) framework to investigate the domain generalization of a pre-defined model toward OOD appearance and skeleton.", "The former presents a coarse-to-fine structure which utilizes coarse shape or foreground masks to ensure generalization on arbitrary poses.", "However, these methods are not efficient at inference and require additional computing power."], "text_after_citation": ["CASD #OTHEREFR introduces attention-based methods to distribute the semantic vectors to the target poses.", "Compared to these parser-based methods, NTED #OTHEREFR applies sparse attention-based operation to extract the semantic textures without the assistance of the external parser.", "Although these methods have been validated well on the DeepFashion dataset #OTHEREFR , there is no relevant research to extend these pre-trained models to OOD dataset.", "Our work first explores the performance of these models on OOD data.", "However, the performance still remains a large gap between photo-realistic image due to the pre-trained models overfitting on DeepFashion dataset #OTHEREFR ."], "citing_paper_content": {"title": "Open-World Pose Transfer Via Sequential Test-Time Adaption", "abstract": "Figure 1. Visualization of open-world pose transfer (OWPT). With open-world references, it can be observed that typical pose transfer method (e.g. NTED [32]) exhibits a twisty pattern. In this sense, we call for a solid model that can handle open-world instances beyond a specific dataset."}, "cited_paper_content": {"title": "Controllable Person Image Synthesis With Attribute-Decomposed Gan", "abstract": "This paper introduces the Attribute-Decomposed GAN, a novel generative model for controllable person image synthesis, which can produce realistic person images with desired human attributes (e.g., pose, head, upper clothes and pants) provided in various source inputs. The core idea of the proposed model is to embed human attributes into the latent space as independent codes and thus achieve flexible and continuous control of attributes via mixing and interpolation operations in explicit style representations. Specifically, a new architecture consisting of two encoding pathways with style block connections is proposed to decompose the original hard mapping into multiple more accessible subtasks. In source pathway, we further extract component layouts with an off-the-shelf human parser and feed them into a shared global texture encoder for decomposed latent codes. This strategy allows for the synthesis of more realistic output images and automatic separation of un-annotated attributes. Experimental results demonstrate the proposed method's superiority over the state of the art in pose transfer and its effectiveness in the brand-new task of component attribute transfer."}, "keywords": ["target image", "expressive textures vectors"], "citation_intent": "background"}
{"citing_id": "2304.01585v1", "cited_id": "1802.00761", "section_title": "Networks", "citation": "Attribute representations is a method of describing the data semantically #REFR . An attribute vector a represents a set of soft-biometrics.", "text_before_citation": ["Prior to concatenation, the outputs of each convolutional blocks are processed by a fully connected layer or LSTM layer depending on the network type.", "Concatenation is followed by a two-layered fully connected MLP and a classifier layer for tCNN-IMU M LP .", "In the case of tCNN-IMU LST M , the concatenation is followed by two LSTM layers and a classifier layer.", "The networks use a softmax classifier for person identification, whereas a sigmoid layer for soft-biometrics identification #OTHEREFR .", "Soft-biometrics of individuals describe or categorise an individual or a group of individuals #OTHEREFR , e.g., Gender Identity, Age, Weight and Height."], "text_after_citation": ["A similar combination of soft-biometrics could represent different persons with similar features.", "The Nearest Neighbour Approach (NNA) is used for soft-biometrics-based identification.", "The NNA calculates the distance between a prediction attribute vector a and an attribute representation A, with all the different combinations of soft-biometrics.", "The person identity is assigned to the one related with the least distance from A.", "NNA is performed by computing a certain similarity between a A and the vector a from the network; typically, the cosine similarity #OTHEREFR and the Probabilistic Retrieval Model (PRM) similarity #OTHEREFR ."], "citing_paper_content": {"title": "Multi-Channel Time-Series Person And Soft-Biometric Identification", "abstract": "Multi-channel time-series datasets are popular in the context of human activity recognition (HAR). On-body device (OBD) recordings of human movements are often preferred for HAR applications not only for their reliability but as an approach for identity protection, e.g., in industrial settings. Contradictory, the gait activity is a biometric, as the cyclic movement is distinctive and collectable. In addition, the gait cycle has proven to contain soft-biometric information of human groups, such as age and height. Though general human movements have not been considered a biometric, they might contain identity information. This work investigates person and soft-biometrics identification from OBD recordings of humans performing different activities using deep architectures. Furthermore, we propose the use of attribute representation for soft-biometric identification. We evaluate the method on four datasets of multi-channel time-series HAR, measuring the performance of a person and soft-biometrics identification and its relation concerning performed activities. We find that person identification is not limited to gait activity. The impact of activities on the identification performance was found to be training and dataset specific. Soft-biometric based attribute representation shows promising results and emphasis the necessity of larger datasets."}, "cited_paper_content": {"title": "Learning Attribute Representation For Human Activity Recognition", "abstract": "Attribute representations became relevant in image recognition and word spotting, providing support under the presence of unbalance and disjoint datasets. However, for human activity recognition using sequential data from on-body sensors, human-labeled attributes are lacking. This paper introduces a search for attributes that represent favorably signal segments for recognizing human activities. It presents three deep architectures, including temporal-convolutions and an IMU centered design, for predicting attributes. An empiric evaluation of random and learned attribute representations, and as well as the networks is carried out on two datasets, outperforming the state-of-the art."}, "keywords": ["soft-biometrics", "Attribute representations"], "citation_intent": "background"}
{"citing_id": "2303.08989v1", "cited_id": "1910.11333", "section_title": "Introduction", "citation": "For instance, to simulate Google's Sycamore #REFR , which has 53 qubits, we would require 128 PB of memory.", "text_before_citation": ["In a quantum computer, all operations follow quantum mechanics: preparing qubits (a quantum version of classical bits), applying unitary gates, and measuring qubits to get classical data.", "The goal of quantum circuit simulation is to reproduce the classical result obtained by these quantum operations only with a classical computer.", "There exist various types of quantum simulators #OTHEREFR 29] .", "For general circuits dominated by non-Clifford gates, the two types of simulators: 1) state vector and 2) tensor network methods are widely used. We choose these simulation methods according to the objectives.", "The state vector simulations require 2 n complex values on memory, where n is the number of qubits."], "text_after_citation": ["Since the total memory capacity of the current largest supercomputers is in the order of a few PB, state vector methods are limited by memory capacity.", "One advantage of state vector methods is that the computational complexity for a circuit of depth d is O(2 n \u00d7d) and scales linearly with d.", "Furthermore, there are some studies to reduce the required memory size, such as by splitting the circuit #OTHEREFR .", "On the other hand, the tensor contraction method #OTHEREFR can simulate several thousands of qubits with low-depth layers at a slightly higher computational cost.", "Therefore, tensor contraction is the method of choice for many recent studies that aim to validate quantum supremacy in both quantum computing and high performance computing #OTHEREFR ."], "citing_paper_content": {"title": "Quantum Circuit Simulation By Sgemm Emulation On Tensor Cores And Automatic Precision Selection", "abstract": "Quantum circuit simulation provides the foundation for the development of quantum algorithms and the verification of quantum supremacy. Among the various methods for quantum circuit simulation, tensor network contraction has been increasing in popularity due to its ability to simulate a larger number of qubits. During tensor contraction, the input tensors are reshaped to matrices and computed by a GEMM operation, where these GEMM operations could reach up to 90% of the total calculation time. GEMM throughput can be improved by utilizing mixed-precision hardware such as Tensor Cores, but straightforward implementation results in insufficient fidelity for deep and large quantum circuits. Prior work has demonstrated that compensated summation with special care of the rounding mode can fully recover the FP32 precision of SGEMM even when using TF32 or FP16 Tensor Cores. The exponent range is a critical issue when applying such techniques to quantum circuit simulation. While TF32 supports almost the same exponent range as FP32, FP16 supports a much smaller exponent range. In this work, we use the exponent range statistics of input tensor elements to select which Tensor Cores we use for the GEMM. We evaluate our method on Random Circuit Sampling (RCS), including Sycamore's quantum circuit, and show that the throughput is 1.86 times higher at maximum while maintaining accuracy."}, "cited_paper_content": {"title": "Supplementary Information For\"Quantum Supremacy Using A Programmable Superconducting Processor\"", "abstract": "The promise of quantum computers is that certain computational tasks might be executed exponentially faster on a quantum processor than on a classical processor1. A fundamental challenge is to build a high-fidelity processor capable of running quantum algorithms in an exponentially large computational space. Here we report the use of a processor with programmable superconducting qubits2\u20137 to create quantum states on 53 qubits, corresponding to a computational state-space of dimension 253 (about 1016). Measurements from repeated experiments sample the resulting probability distribution, which we verify using classical simulations. Our Sycamore processor takes about 200 seconds to sample one instance of a quantum circuit a million times\u2014our benchmarks currently indicate that the equivalent task for a state-of-the-art classical supercomputer would take approximately 10,000 years. This dramatic increase in speed compared to all known classical algorithms is an experimental realization of quantum supremacy8\u201314 for this specific computational task, heralding a much-anticipated computing paradigm. Quantum supremacy is demonstrated using a programmable superconducting processor known as Sycamore, taking approximately 200 seconds to sample one instance of a quantum circuit a million times, which would take a state-of-the-art supercomputer around ten thousand years to compute."}, "keywords": ["53 qubits"], "citation_intent": "background"}
{"citing_id": "2305.01828v1", "cited_id": "1707.00291", "section_title": "Los Probability Models", "citation": "For the UMi and UMa scenarios, 3GPP SCM had a higher probability of predicting LOS channel conditions at larger distances (several hundred meters) than NYUSIM #REFR .", "text_before_citation": ["The first step in generating a channel model for a UE is determining whether the channel condition is LOS/NLOS.", "We extend the ChannelCondition class developed in #OTHEREFR with five different classes, namely, NYURmaChannelConditionModel, NYUUma-ChannelConditionModel, NYUUmiChannelConditionModel, NYU-InHChannelConditionModel, and NYUInFChannelConditionModel, each handling a different scenario for NYUSIM.", "All the newly introduced NYUSIM classes mentioned above derive from the same base class, called NYUChannelCondition, which extends the Cha-nnelConditionModel interface.", "The NYUSIM LOS probability models for outdoor scenarios were developed based on radio propagation measurements conducted at 28 and 73 GHz in New York City #OTHEREFR .", "The Tx-Rx locations were selected from measurement data, and ray tracing was employed to determine whether the path between the Tx and Rx is in LOS/NLOS #OTHEREFR ."], "text_after_citation": ["However, in a real-world UMi and UMa scenario, getting LOS propagation over several hundred meters is very challenging #OTHEREFR .", "Thus, the UMi and UMa, LOS probability models, implemented in NYUUmiChannelConditionModel and NYU-UmaChannelConditionModel, use the NYU (squared) model based on Tables 1 and 2 in #OTHEREFR respectively and is given in (1) and #OTHEREFR .", "EQUATION", "EQUATION", "In (1) 1 = 22m, 2 = 100m, in (2) 1 = 20m, 2 = 160m, is defined in Table 7 .4.2 in #OTHEREFR and \u210e is the height of the UE in meters."], "citing_paper_content": {"title": "Ns-3 Implementation Of Sub-Terahertz And Millimeter Wave Drop-Based Nyu Channel Model (Nyusim)", "abstract": "The next generation of wireless networks will use sub-THz frequencies alongside mmWave frequencies to enable multi-Gbps and low latency applications. To enable different verticals and use cases, engineers must take a holistic approach to build, analyze, and study different parts of the network and the interplay among the lower and higher layers of the protocol stack. It is of paramount importance to accurately characterize the radio propagation in diverse scenarios such as urban microcell (UMi), urban macrocell (UMa), rural macrocell (RMa), indoor hotspot (InH), and indoor factory (InF) for a wide range of frequencies. The 3GPP statistical channel model (SCM) is oversimplified and restricted to the frequency range of 0.5-100 GHz. Thus, to overcome these limitations, this paper presents a detailed implementation of the drop-based NYU channel model (NYUSIM) for the frequency range of 0.5-150 GHz for the UMi, UMa, RMa, InH, and InF scenarios. NYUSIM allows researchers to design and evaluate new algorithms and protocols for future sub-THz wireless networks in ns-3. CCS CONCEPTS \u2022 Networks \u2192 Network simulation; Mobile networks;."}, "cited_paper_content": {"title": "Investigation And Comparison Of 3Gpp And Nyusim Channel Models For 5G Wireless Communications", "abstract": "Channel models describe how wireless channel parameters behave in a given scenario, and help evaluate link- and system-level performance. A proper channel model should be able to faithfully reproduce the channel parameters obtained in field measurements and accurately predict the spatial and temporal channel impulse response along with large-scale fading. This paper compares two popular channel models for next generation wireless communications: the 3rd Generation Partnership Project (3GPP) TR 38.900 Release 14 channel model and the statistical spatial channel model NYUSIM developed by New York University (NYU). The two channel models employ different modeling approaches in many aspects, such as the line-of-sight probability, path loss, and clustering methodology. Simulations are performed using the two channel models to analyze the channel eigenvalue distribution and spectral efficiency leveraging the analog/digital hybrid beamforming methods found in the literature. Simulation results show that the 3GPP model produces different eigenvalue and spectral efficiency distributions for mmWave bands, as compared to the outcome from NYUSIM that is based on massive amounts of real-world measured data in New York City. This work shows NYUSIM is more accurate for realistic simulations than 3GPP in urban environments."}, "keywords": ["LOS channel conditions"], "citation_intent": "background"}
{"citing_id": "2304.08983v1", "cited_id": "1805.07944", "section_title": "Iv. Numerical Example", "citation": "If the attack identification is performed with all the measurements, then the number of cases is equal to #REFR 4 = 4845.", "text_before_citation": [", 10, 1 2 x 3 \u2212 1 2 sin x 2 , i = 11, .", ". .", ", 20, and the noise v(t) \u2208 R 20 be given such that v(t) \u2264 0.01.", "In this example, the partial information \u03a6 i (x) of the state x obtained from each individual output y i is no more than h i (x), i.e., \u03a6 i = h i , \u2200i, as it can be computed from (5) that L f h i = \u2212h i for i \u2264 10, and L f h i = 0 for i \u2265 11.", "Then, it can be verified that the state x(t) can be restored from any 12 elements of the measurements {y i (t)} #OTHEREFR i=1 , so that the attack identification and the resilient state estimation is possible when up to 4 measurements are compromised."], "text_after_citation": ["But, the attack identification can also be performed with the local measurements {y i (t)} 10 i=1 and {y i (t)} 20 i=11 , separately, although the state x(t) is not observable from either of them.", "This is because both the collections {\u03a6 i } 10", "i=1 and {\u03a6 i } 20"], "citing_paper_content": {"title": "", "abstract": "A resilient state estimation scheme for uniformly observable nonlinear systems, based on a method for local identification of sensor attacks, is presented. The estimation problem is combinatorial in nature, and so many methods require substantial computational and storage resources as the number of sensors increases. To reduce the complexity, the proposed method performs the attack identification with local subsets of the measurements, not with the set of all measurements. A condition for nonlinear attack identification is introduced as a relaxed version of existing redundant observability condition. It is shown that an attack identification can be performed even when the state cannot be recovered from the measurements. As a result, although a portion of measurements are compromised, they can be locally identified and excluded from the state estimation, and thus the true state can be recovered. Simulation results demonstrate the effectiveness of the proposed scheme."}, "cited_paper_content": {"title": "Detection Of Sensor Attack And Resilient State Estimation For Uniformly Observable Nonlinear Systems Having Redundant Sensors", "abstract": "This paper presents a detection algorithm for sensor attacks and a resilient state estimation scheme for a class of uniformly observable nonlinear systems. An adversary is supposed to corrupt a subset of sensors with the possibly unbounded signals, while the system has sensor redundancy. We design an individual high-gain observer for each measurement output so that only the observable portion of the system state is obtained. Then, a nonlinear error correcting problem is solved by collecting all the information from those partial observers and exploiting redundancy. A computationally efficient, on-line monitoring scheme is presented for attack detection. Based on the attack detection scheme, an algorithm for resilient state estimation is provided. The simulation results demonstrate the effectiveness of the proposed algorithm."}, "keywords": ["attack identification"], "citation_intent": "background"}
{"citing_id": "2304.03026v1", "cited_id": "1804.08489", "section_title": "A. Related Work", "citation": "In #REFR , authors studied the reliability of command and control channel between UAVs and a traditional cellular network with massive MIMO.", "text_before_citation": ["A high mobility UAV communication relay was analyzed in #OTHEREFR .", "The authors studied the throughput maximization problems by designing UAV trajectory and power allocations.", "Authors in #OTHEREFR , #OTHEREFR investigated a Max-Min SNR based UAV trajectory optimization problem.", "Authors in #OTHEREFR jointly designed multiple UAVs' trajectories and transmit powers to maximize the throughput of a group of ground users.", "In #OTHEREFR , the authors considered a UAV that collects data from multiple wireless sensor networks and jointly optimized UAV's trajectory and communication scheduling to maximize the minimum average data collection rate."], "text_after_citation": ["Besides, a deep reinforcement learning-based approach of optimizing multiple cellular-connected UAVs to minimize their interference with ground users was analyzed in #OTHEREFR .", "Authors in #OTHEREFR studied the uplink inter-cell interference coordination design for a cellular network simultaneously serving UAVs and ground users.", "Typically, the maximum weighted sum-rate of UAVs and users is jointly optimized based on power allocations and uplink cell associations.", "Authors in #OTHEREFR maximized the sum rate of UAV uplink transmission by optimizing the precoding vectors at the multi-antenna UAVs.", "A deep reinforcement learning-based intelligent navigation task of a cellular-connected UAV network was investigated in #OTHEREFR which aims to minimize the weighted sum of time cost and expected outage duration alongside UAV trajectory."], "citing_paper_content": {"title": "Coverage Analysis And Trajectory Optimization For Aerial Users With Dedicated Cellular Infrastructure", "abstract": "In this paper, we consider a novel cellular network for aerial users, which is composed of dedicated base stations (BSs), whose antennas are directed towards aerial users, and traditional terrestrial BSs (TBSs). Besides, the dedicated BSs are deployed on roadside furniture, such as lampposts and traffic lights, to achieve multiple features while occupying less space. Therefore, the locations of dedicated BSs and TBSs are modeled by a Poisson-line-Cox-process (PLCP) and Poisson point process (PPP), respectively. For the proposed network, we first compute the aerial coverage probability and show that the deployment of dedicated BSs improves the coverage probability in both high dense areas and rural areas. We then consider a cellular-connected UAV that has a flying mission and optimize its trajectory to maximize the minimal achievable signal-to-interference-plus-noise ratio (SINR) (Max-Min SINR). To obtain the Max-Min SINR and minimal time trajectory that satisfies the Max-Min SINR, we proposed two algorithms that are practical in large-scale networks. Finally, our results show that the optimal density of dedicated BSs which maximizes Max-Min SINR decreases with the increase of the road densities."}, "cited_paper_content": {"title": "Understanding Uav Cellular Communications: From Existing Networks To Massive Mimo", "abstract": "The purpose of this article is to bestow the reader with a timely study of UAV cellular communications, bridging the gap between the 3GPP standardization status quo and the more forward-looking research. Special emphasis is placed on the downlink command and control (C&C) channel to aerial users, whose reliability is deemed of paramount technological importance for the commercial success of UAV cellular communications. Through a realistic side-by-side comparison of two network deployments -- a present-day cellular infrastructure versus a next-generation massive MIMO system -- a plurality of key facts are cast light upon, with the three main ones summarized as follows: (i) UAV cell selection is essentially driven by the secondary lobes of a base station's radiation pattern, causing UAVs to associate to far-flung cells; (ii) over a 10 MHz bandwidth, and for UAV heights of up to 300 m, massive MIMO networks can support 100 kbps C&C channels in 74% of the cases when the uplink pilots for channel estimation are reused among base station sites, and in 96% of the cases without pilot reuse across the network; (iii) supporting UAV C&C channels can considerably affect the performance of ground users on account of severe pilot contamination, unless suitable power control policies are in place."}, "keywords": ["traditional cellular network", "massive MIMO"], "citation_intent": "background"}
{"citing_id": "2303.12357v1", "cited_id": "1706.06083", "section_title": "Introduction", "citation": "Projected gradient descent attack #REFR is a widely-used attack method that applies small steps of maximizing the loss objective iteratively and clipping the values of intermediate results after each step (projection to the norm ball) to ensure that they are in a constrained neighbourhood of the original inputs.", "text_before_citation": ["The time series that appear more distinguished to our eyes are captured by larger Wasserstein distances.", "The example in the upper figure may not be a successful adversarial example under 1 distance as it is distant in the 1 space. However, it is almost imperceptible to human evaluation.", "In this case, Wasserstein distance is a better measurement of adversarial examples.", "In this work, we study the adversarial attack on time series in the Wasserstein space for the first time.", "Our goal is to generate adversarial examples that have small Wasserstein perturbation so it is more indistinguishable and natural to human, e.g., physician who examines ECG data."], "text_after_citation": ["Similarly, we propose a Wasserstein PGD method to search for adversarial examples in the Wasserstein space for univariant time series.", "Wasserstein distance cannot be calculated directly without solving an optimization subproblem and has no closed-form solution in most cases, which limits its applications.", "At present, there are only two cases that the Wasserstein distance can be directly calculated, one is the case of the dimension of inputs being 1, and the other is the inputs following Gaussian distribution.", "For the univariant time series, we can take advantage of its 1 characteristic and use the closedform Wasserstein distance to apply the projection of intermediate results of each step onto the Wasserstein ball with gradient descent method.", "The direct projection of intermediate results onto the Wasserstein ball using gradient descent could be time-consuming and converge to a sub-optimal solution."], "citing_paper_content": {"title": "Wasserstein Adversarial Examples On Univariant Time Series Data", "abstract": "Adversarial examples are crafted by adding indistinguishable perturbations to normal examples in order to fool a well-trained deep learning model to misclassify. In the context of computer vision, this notion of indistinguishability is typically bounded by \u221e or other norms. However, these norms are not appropriate for measuring indistinguishiability for time series data. In this work, we propose adversarial examples in the Wasserstein space for time series data for the first time and utilize Wasserstein distance to bound the perturbation between normal examples and adversarial examples. We introduce Wasserstein projected gradient descent (WPGD), an adversarial attack method for perturbing univariant time series data. We leverage the closed-form solution of Wasserstein distance in the 1D space to calculate the projection step of WPGD efficiently with the gradient descent method. We further propose a two-step projection so that the search of adversarial examples in the Wasserstein space is guided and constrained by Euclidean norms to yield more effective and imperceptible perturbations. We empirically evaluate the proposed attack on several time series datasets in the healthcare domain. Extensive results demonstrate that the Wasserstein attack is powerful and can successfully attack most of the target classifiers with a high attack success rate. To better study the nature of Wasserstein adversarial example, we evaluate a strong defense mechanism named Wasserstein smoothing for potential certified robustness defense. Although the defense can achieve some accuracy gain, it still has limitations in many cases and leaves space for developing a stronger certified robustness method to Wasserstein adversarial examples on univariant time series data."}, "cited_paper_content": {"title": "Towards Deep Learning Models Resistant To Adversarial Attacks", "abstract": "Recent work has demonstrated that neural networks are vulnerable to adversarial examples, i.e., inputs that are almost indistinguishable from natural data and yet classified incorrectly by the network. In fact, some of the latest findings suggest that the existence of adversarial attacks may be an inherent weakness of deep learning models. To address this problem, we study the adversarial robustness of neural networks through the lens of robust optimization. This approach provides us with a broad and unifying view on much of the prior work on this topic. Its principled nature also enables us to identify methods for both training and attacking neural networks that are reliable and, in a certain sense, universal. In particular, they specify a concrete security guarantee that would protect against any adversary. These methods let us train networks with significantly improved resistance to a wide range of adversarial attacks. They also suggest the notion of security against a first-order adversary as a natural and broad security guarantee. We believe that robustness against such well-defined classes of adversaries is an important stepping stone towards fully resistant deep learning models."}, "keywords": ["Projected gradient descent", "loss objective"], "citation_intent": "method"}
{"citing_id": "2303.00498v1", "cited_id": "1706.03762", "section_title": "C. Adaptive Hybrid Spatial-Temporal Learning Block", "citation": "We also apply multi-head attention #REFR to capture the relation semantics between nodes from different learning subspaces.", "text_before_citation": ["Dynamic Graph Learning.", "DGL implements Graph Attention Network (GAT) #OTHEREFR to perform dynamic spatial feature aggregation, owing to GAT's ability to capture the dynamic influences of neighbors.", "We believe that short-term dynamic influences mostly occur between nearby cell towers, so we deploy the exponential distance-decay matrix #OTHEREFR as the prior graph structure of GAT.", "We call it the distance adjacency matrix, denoted as A dis \u2208 R N \u00d7N . As shown in Fig.", "2 , GAT achieves weighted feature aggregation by calculating the attention scores of the central node and neighbor nodes."], "text_after_citation": ["Moreover, the multi-head attention can be computed in parallel for reducing time complexity.", "For simplicity, taking a graph node v i as an example, the graph attention mechanism of DGL can be formulated as follows:", "EQUATION", "EQUATION", "where a ij \u2208 R represents the computed attention score between v i and v j , where j \u2208 N i ."], "citing_paper_content": {"title": "Adaptive Hybrid Spatial-Temporal Graph Neural Network For Cellular Traffic Prediction", "abstract": "Cellular traffic prediction is an indispensable part for intelligent telecommunication networks. Nevertheless, due to the frequent user mobility and complex network scheduling mechanisms, cellular traffic often inherits complicated spatialtemporal patterns, making the prediction incredibly challenging. Although recent advanced algorithms such as graph-based prediction approaches have been proposed, they frequently model spatial dependencies based on static or dynamic graphs and neglect the coexisting multiple spatial correlations induced by traffic generation. Meanwhile, some works lack the consideration of the diverse cellular traffic patterns, result in suboptimal prediction results. In this paper, we propose a novel deep learning network architecture, Adaptive Hybrid Spatial-Temporal Graph Neural Network (AHSTGNN), to tackle the cellular traffic prediction problem. First, we apply adaptive hybrid graph learning to learn the compound spatial correlations among cell towers. Second, we implement a Temporal Convolution Module with multi-periodic temporal data input to capture the nonlinear temporal dependencies. In addition, we introduce an extra Spatial-Temporal Adaptive Module to conquer the heterogeneity lying in cell towers. Our experiments on two real-world cellular traffic datasets show AHSTGNN outperforms the state-of-the-art by a significant margin, illustrating the superior scalability of our method for spatial-temporal cellular traffic prediction."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["nodes", "multi-head attention"], "citation_intent": "method"}
{"citing_id": "2304.02091v1", "cited_id": "0912.2371", "section_title": "Subgraph Problems", "citation": "In fact, one of the fastest methods for Subgraph Isomorphism works via an arithmetic circuit for evaluating the homomorphism polynomial #REFR , allowing for the randomized detection of a subgraph H with |V (H)| = k and of treewidth w in time O(2 k n w+1 ).", "text_before_citation": ["In our first application in this section, we note a generalisation of the above results. The second result regards subgraph isomorphism.", "Given graphs G and H, the problem of checking whether H is a subgraph of G parameterized by |V (H)| can be either FPT, as when H is a path, or W[1]-hard, as when H is a clique.", "More generally, Subgraph Isomorphism is FPT by |V (H)| if H comes from a family of graphs with bounded treewidth (originally shown using the colour-coding technique of Alon et al.", "#OTHEREFR ), and there is good evidence that no more general such class exists #OTHEREFR .", "In general, the parameterized complexity of subgraph isomorphism problems has been extensively and meticulously investigated #OTHEREFR ."], "text_after_citation": ["Furthermore, the exponent w + 1 here is optimal, up to plausible conjectures #OTHEREFR .", "We observe that this running time is compatible with an additional constraint that the copy of H found in G should be independent in a given linear matroid."], "citing_paper_content": {"title": "Determinantal Sieving", "abstract": "We introduce a new, remarkably powerful tool to the toolbox of algebraic FPT algorithms, determinantal sieving. Given a polynomial P (x 1 ,. .. , x n) over a field F of characteristic 2, on a set of variables X = {x 1 ,. .. , x n }, and a linear matroid M = (X, I) over F of rank k, in 2 k evaluations of P we can sieve for those terms in the monomial expansion of P which are multilinear and whose support is a basis for M. The known tools of multilinear detection and constrained multilinear detection then correspond to the case where M is a uniform matroid, respectively the truncation of a disjoint union of uniform matroids. More generally, let the odd support of a monomial m be the set of variables which have odd degree in m. Using 2 k evaluations of P , we can sieve for those terms m whose odd support spans M. Applying this framework to well-known efficiently computable polynomial families allows us to simplify, generalize and improve on a range of algebraic FPT algorithms, such as:"}, "cited_paper_content": {"title": "Faster Algorithms For Finding And Counting Subgraphs", "abstract": "In this paper we study a natural generalization of both {\\sc $k$-Path} and {\\sc $k$-Tree} problems, namely, the {\\sc Subgraph Isomorphism} problem. In the {\\sc Subgraph Isomorphism} problem we are given two graphs $F$ and $G$ on $k$ and $n$ vertices respectively as an input, and the question is whether there exists a subgraph of $G$ isomorphic to $F$. We show that if the treewidth of $F$ is at most $t$, then there is a randomized algorithm for the {\\sc Subgraph Isomorphism} problem running in time $\\cO^*(2^k n^{2t})$. To do so, we associate a new multivariate {Homomorphism polynomial} of degree at most $k$ with the {\\sc Subgraph Isomorphism} problem and construct an arithmetic circuit of size at most $n^{\\cO(t)}$ for this polynomial. Using this polynomial, we also give a deterministic algorithm to count the number of homomorphisms from $F$ to $G$ that takes $n^{\\cO(t)}$ time and uses polynomial space. For the counting version of the {\\sc Subgraph Isomorphism} problem, where the objective is to count the number of distinct subgraphs of $G$ that are isomorphic to $F$, we give a deterministic algorithm running in time and space $\\cO^*({n \\choose k/2}n^{2p})$ or ${n\\choose k/2}n^{\\cO(t \\log k)}$. We also give an algorithm running in time $\\cO^{*}(2^{k}{n \\choose k/2}n^{5p})$ and taking space polynomial in $n$. Here $p$ and $t$ denote the pathwidth and the treewidth of $F$, respectively. Thus our work not only improves on known results on {\\sc Subgraph Isomorphism} but it also extends and generalize most of the known results on {\\sc $k$-Path} and {\\sc $k$-Tree}."}, "keywords": ["Subgraph Isomorphism"], "citation_intent": "method"}
{"citing_id": "2303.15901v1", "cited_id": "2003.05991", "section_title": "Introduction", "citation": "A DAE is a type of DNN that is trained to detect and discard noise from input data and reconstruct it back to its original form #REFR .", "text_before_citation": ["Several defence strategies have been proposed to overcome this problem, including defensive distillation, which has been successful in defending DNNs from adversarial attacks in run-time settings #OTHEREFR .", "Nevertheless, one of the drawbacks of the defensive distillation method is that it remains susceptible to data poisoning attacks, in which adversaries aim to impair the model's performance by inserting erroneous data entries into the training set.", "These harmful data points could be carefully crafted to be close to the model's decision boundary to circumvent the countermeasures offered by defensive distillation.", "This study presents a novel method for robustifying a distilled network against data poisoning adversarial threats by integrating a denoising autoencoder (DAE) #OTHEREFR in the defensive distillation cycle.", "Defensive distillation involves training two DNNs, the instructor model (the first model), and distilling its knowledge into the student model (the second or distilled model) to make it robust adversarial examples and previously unseen input #OTHEREFR ."], "text_after_citation": ["This paper is motivated by the fact that the instructor model is not immune to data poisoning adversarial attacks.", "Although the student model has more latitude to reject input modifications because it leverages the \"distilled\" version of the training data, where the training examples are transformed by a temperature parameter T #OTHEREFR .", "Thus, minimising the instructor model's susceptibility to data poisoning attacks is pivotal for developing a reliable, distilled DNN.", "To achieve this, we designed a DAE to detect and reconstruct poisonous adversarial inputs in the training data.", "The defensive distillation method already offers a strong foundation, but when combined with a DAE, the protection mechanism against data poisoning adversarial attacks is significantly strengthened."], "citing_paper_content": {"title": "Denoising Autoencoder-Based Defensive Distillation As An Adversarial Robustness Algorithm", "abstract": "Adversarial attacks significantly threaten the robustness of deep neural networks (DNNs). Despite the multiple defensive methods employed, they are nevertheless vulnerable to poison attacks, where attackers meddle with the initial training data. In order to defend DNNs against such adversarial attacks, this work proposes a novel method that combines the defensive distillation mechanism with a denoising autoencoder (DAE). This technique tries to lower the sensitivity of the distilled model to poison attacks by spotting and reconstructing poisonous adversarial inputs in the training data. We added carefully created adversarial samples to the initial training data to assess the proposed method's performance. Our experimental findings demonstrate that our method successfully identified and reconstructed the poisonous inputs while also considering enhancing the DNN's resilience. The proposed approach provides a potent and robust defence mechanism for DNNs in various applications where data poisoning attacks are a concern. Thus, the defensive distillation technique's limitation posed by poisonous adversarial attacks is overcome."}, "cited_paper_content": {"title": "Autoencoders", "abstract": "An autoencoder is a specific type of a neural network, which is mainlydesigned to encode the input into a compressed and meaningful representation, andthen decode it back such that the reconstructed input is similar as possible to theoriginal one. This chapter surveys the different types of autoencoders that are mainlyused today. It also describes various applications and use-cases of autoencoders."}, "keywords": ["DNN", "input data"], "citation_intent": "background"}
{"citing_id": "2303.05648v1", "cited_id": "1906.06365", "section_title": "B. Machine Learning-Based Methods", "citation": "In #REFR , a function-aggregation method is proposed, which assumes that one utility function can capture only a part of users' preferences, and the whole preferences can be infinitely approximated by introducing enough utility functions.", "text_before_citation": ["Recently, many machine learning-based methods have been proposed to specially address choice problems.", "These methods do not need an extra transformation process and avoid possible preference distortions.", "For example, the pointer neural network (PNN) #OTHEREFR , which is an encoder-decoder network based on recurrent neural network and attention mechanism, is utilized as the mapping function between the available items in the markets and users'", "choices.", "The available items are fed into PNN one by one, and PNN would point to the item that is predicted as the best choice #OTHEREFR ."], "text_after_citation": ["The function-aggregation method adopts multiple utility functions simultaneously, and the total utility is equal to their weighted sum.", "The restricted Boltzmann machine is adopted in #OTHEREFR to model context effects.", "It is proved that the restricted Boltzmann machine can be translated into the MNL formulation, together with an additional term accounting for the comparisons among items."], "citing_paper_content": {"title": "Pacos: Modeling Users' Interpretable And Context-Dependent Choices In Preference Reversals", "abstract": "Choice problems refer to selecting the best choices from several items, and learning users' preferences in choice problems is of great significance in understanding the decision making mechanisms and providing personalized services. Existing works typically assume that people evaluate items independently. In practice, however, users' preferences depend on the market in which items are placed, which is known as context effects; and the order of users' preferences for two items may even be reversed, which is referred to preference reversals. In this work, we identify three factors contributing to context effects: users' adaptive weights, the inter-item comparison, and display positions. We propose a context-dependent preference model named Pacos as a unified framework for addressing three factors simultaneously, and consider two design methods including an additive method with high interpretability and an ANN-based method with high accuracy. We study the conditions for preference reversals to occur and provide an theoretical proof of the effectiveness of Pacos in addressing preference reversals. Experimental results show that the proposed method has better performance than prior works in predicting users' choices, and has great interpretability to help understand the cause of preference reversals."}, "cited_paper_content": {"title": "Predicting Choice With Set-Dependent Aggregation", "abstract": "Providing users with alternatives to choose from is an essential component in many online platforms, making the accurate prediction of choice vital to their success. A renewed interest in learning choice models has led to significant progress in modeling power, but most current methods are either limited in the types of choice behavior they capture, cannot be applied to large-scale data, or both. ::: Here we propose a learning framework for predicting choice that is accurate, versatile, theoretically grounded, and scales well. Our key modeling point is that to account for how humans choose, predictive models must capture certain set-related invariances. Building on recent results in economics, we derive a class of models that can express any behavioral choice pattern, enjoy favorable sample complexity guarantees, and can be efficiently trained end-to-end. Experiments on three large choice datasets demonstrate the utility of our approach."}, "keywords": ["users' preferences", "whole preferences"], "citation_intent": "method"}
{"citing_id": "2304.12486v2", "cited_id": "1902.06705", "section_title": "Threat Model", "citation": "We define this threat model according to Carlini et al #REFR by defining the goals, capabilities, and knowledge of the target AI system, that the adversary has.", "text_before_citation": ["In order to have correct metrics, and to assess the validity of our approach in a real-world context, it is essential that we define a threat model using a precise taxonomy."], "text_after_citation": ["Let C(x) : X \u2192 Y be a classifier, and x \u2208 X \u2282 R d an input to the classifier.", "Let y true \u2208 Y = {1, 2, ..., N } be the ground truth of the input x, i.e. the true class label of x among N classes.", "We call y pred = C(x) the predicted label for x.", "In our study, we perform untargeted attacks, ie.", "the adversary's goal is to generate an adversarial example x adv for x that is misclassified by C."], "citing_paper_content": {"title": "Evaluating Adversarial Robustness On Document Image Classification", "abstract": "Adversarial attacks and defenses have gained increasing interest on computer vision systems in recent years, but as of today, most investigations are limited to natural images. However, many artificial intelligence models actually handle documentary data, which is very different from real world images. Hence, in this work, we try to apply the adversarial attack philosophy on documentary data and to protect models against such attacks. Our methodology is to implement untargeted gradient-based, transfer-based and score-based attacks and evaluate the impact of defenses such as adversarial training, JPEG input compression and grey-scale input transformation on the robustness of ResNet50 and EfficientNetB0 model architectures. To the best of our knowledge, no such work has been conducted by the community in order to study the impact of these attacks on the document image classification task."}, "cited_paper_content": {"title": "On Evaluating Adversarial Robustness", "abstract": "Correctly evaluating defenses against adversarial examples has proven to be extremely difficult. Despite the significant amount of recent work attempting to design defenses that withstand adaptive attacks, few have succeeded; most papers that propose defenses are quickly shown to be incorrect. ::: We believe a large contributing factor is the difficulty of performing security evaluations. In this paper, we discuss the methodological foundations, review commonly accepted best practices, and suggest new methods for evaluating defenses to adversarial examples. We hope that both researchers developing defenses as well as readers and reviewers who wish to understand the completeness of an evaluation consider our advice in order to avoid common pitfalls."}, "keywords": ["target AI system"], "citation_intent": "method"}
{"citing_id": "2304.02742v1", "cited_id": "1703.10593", "section_title": "A. Image-To-Image Translation", "citation": "One of the most representative GAN-based methods is CycleGAN #REFR , which uses cyclic consistency losses to constrain the entire model into two generative networks.", "text_before_citation": ["The primary idea behind the image-to-image translation task is to establish a mapping between images from two different domains.", "In medical image translation, the main connection between two domains is the underlying anatomical structure information.", "The previous works on the image translation are mainly based on GANs #OTHEREFR and VAEs #OTHEREFR ."], "text_after_citation": ["It can translate a source domain image to the target domain and vice versa.", "Another representative work is the Geometry-consistent generative adversarial network (GcGAN) #OTHEREFR , which provides unilateral unsupervised mapping.", "GcGAN maps the original image into two different predefined geometric transformations and generates two images in a new domain under the corresponding geometric consistency constraints.", "However, GAN-based models have some inherent drawbacks since they are trained by two networks (generator and discriminator) playing against each other, which can render the training difficult and unstable.", "For VAE-based methods, one of the most representative works on unpaired image translation is UNIT #OTHEREFR , which assumes that two domains share a common latent space and the corresponding images in both domains are mapped to the same latent code."], "citing_paper_content": {"title": "Zero-Shot Medical Image Translation Via Frequency-Guided Diffusion Models", "abstract": "Recently, the diffusion model has emerged as a superior generative model that can produce highquality images with excellent realism. There is a growing interest in applying diffusion models to image translation tasks. However, for medical image translation, the existing diffusion models are deficient in accurately retaining structural information since the structure details of source domain images are lost during the forward diffusion process and cannot be fully recovered through learned reverse diffusion, while the integrity of anatomical structures is extremely important in medical images. Training and conditioning diffusion models using paired source and target images with matching anatomy can help. However, such paired data are very difficult and costly to obtain, and may also reduce the robustness of the developed model to outof-distribution testing data. We propose a frequency-guided diffusion model (FGDM) that employs frequency-domain filters to guide the diffusion model for structure-preserving image translation. Based on its design, FGDM allows zeroshot learning, as it can be trained solely on the data from the target domain, and used directly for source-to-target domain translation without any exposure to the sourcedomain data during training. We trained FGDM solely on the head-and-neck CT data, and evaluated it on both headand-neck and lung cone-beam CT (CBCT)-to-CT translation tasks. FGDM outperformed the state-of-the-art methods (GAN-based, VAE-based, and diffusion-based) in all metrics, showing its significant advantages in zero-shot medical image translation."}, "cited_paper_content": {"title": "Unpaired Image-To-Image Translation Using Cycle-Consistent Adversarial Networks", "abstract": "Image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between an input image and an output image using a training set of aligned image pairs. However, for many tasks, paired training data will not be available. We present an approach for learning to translate an image from a source domain $X$ to a target domain $Y$ in the absence of paired examples. Our goal is to learn a mapping $G: X \\rightarrow Y$ such that the distribution of images from $G(X)$ is indistinguishable from the distribution $Y$ using an adversarial loss. Because this mapping is highly under-constrained, we couple it with an inverse mapping $F: Y \\rightarrow X$ and introduce a cycle consistency loss to push $F(G(X)) \\approx X$ (and vice versa). Qualitative results are presented on several tasks where paired training data does not exist, including collection style transfer, object transfiguration, season transfer, photo enhancement, etc. Quantitative comparisons against several prior methods demonstrate the superiority of our approach."}, "keywords": ["representative GAN-based methods"], "citation_intent": "method"}
{"citing_id": "2304.01585v1", "cited_id": "1802.00761", "section_title": "Networks", "citation": "The networks use a softmax classifier for person identification, whereas a sigmoid layer for soft-biometrics identification #REFR .", "text_before_citation": ["OBDs from each human limb is allotted a branch of four convolutional layers.", "The convolutional blocks extract descriptive local features from the input OBD data, while the subsequent layers assimilate the global view of the extracted features.", "Prior to concatenation, the outputs of each convolutional blocks are processed by a fully connected layer or LSTM layer depending on the network type.", "Concatenation is followed by a two-layered fully connected MLP and a classifier layer for tCNN-IMU M LP .", "In the case of tCNN-IMU LST M , the concatenation is followed by two LSTM layers and a classifier layer."], "text_after_citation": ["Soft-biometrics of individuals describe or categorise an individual or a group of individuals #OTHEREFR , e.g., Gender Identity, Age, Weight and Height.", "Attribute representations is a method of describing the data semantically #OTHEREFR . An attribute vector a represents a set of soft-biometrics.", "A similar combination of soft-biometrics could represent different persons with similar features.", "The Nearest Neighbour Approach (NNA) is used for soft-biometrics-based identification.", "The NNA calculates the distance between a prediction attribute vector a and an attribute representation A, with all the different combinations of soft-biometrics."], "citing_paper_content": {"title": "Multi-Channel Time-Series Person And Soft-Biometric Identification", "abstract": "Multi-channel time-series datasets are popular in the context of human activity recognition (HAR). On-body device (OBD) recordings of human movements are often preferred for HAR applications not only for their reliability but as an approach for identity protection, e.g., in industrial settings. Contradictory, the gait activity is a biometric, as the cyclic movement is distinctive and collectable. In addition, the gait cycle has proven to contain soft-biometric information of human groups, such as age and height. Though general human movements have not been considered a biometric, they might contain identity information. This work investigates person and soft-biometrics identification from OBD recordings of humans performing different activities using deep architectures. Furthermore, we propose the use of attribute representation for soft-biometric identification. We evaluate the method on four datasets of multi-channel time-series HAR, measuring the performance of a person and soft-biometrics identification and its relation concerning performed activities. We find that person identification is not limited to gait activity. The impact of activities on the identification performance was found to be training and dataset specific. Soft-biometric based attribute representation shows promising results and emphasis the necessity of larger datasets."}, "cited_paper_content": {"title": "Learning Attribute Representation For Human Activity Recognition", "abstract": "Attribute representations became relevant in image recognition and word spotting, providing support under the presence of unbalance and disjoint datasets. However, for human activity recognition using sequential data from on-body sensors, human-labeled attributes are lacking. This paper introduces a search for attributes that represent favorably signal segments for recognizing human activities. It presents three deep architectures, including temporal-convolutions and an IMU centered design, for predicting attributes. An empiric evaluation of random and learned attribute representations, and as well as the networks is carried out on two datasets, outperforming the state-of-the art."}, "keywords": ["soft-biometrics identification", "softmax classifier"], "citation_intent": "method"}
{"citing_id": "2303.02223v1", "cited_id": "2003.00803", "section_title": "Related Work", "citation": "Due to these reasons, techniques such as ML and deep learning created a new research direction in the financial literature #REFR .", "text_before_citation": ["Statistical models make some simplifying assumptions to be able to obtain some theoretical guarantees and thus have a simple structure and potentially lover performance compared to machine learning techniques.", "Machine learning models are more general, make fewer simplifying assumptions, and offer better model-fitting capabilities."], "text_after_citation": ["To forecast the future value of financial assets and find the reason for these assets' behavior, numerous machine learning (ML) techniques were employed, such as SVM #OTHEREFR , Support Vector Regression (SVR) #OTHEREFR , Random Forest (RF) #OTHEREFR , and convolutional neural networks (CNN) #OTHEREFR .", "These works show that it is possible to use ML techniques to make more accurate forecasts and as alternative approaches to conventional techniques for investigating the behavior of financial assets. #OTHEREFR", "(2019) applied several ML techniques such as single hidden layer feed-forward neural network, MLP, autoencoders, and bagof-features algorithms to forecast the mid-price movement of the stock price.", "The results indicated that the mentioned ML techniques could forecast price movement. #OTHEREFR", "(2019) organized the stock price prediction techniques into four categories, namely statistical methods, pattern recognition, machine learning, and sentiment analysis."], "citing_paper_content": {"title": "Feature Selection For Forecasting *", "abstract": "This work investigates the importance of feature selection for improving the forecasting performance of machine learning algorithms for financial data. Artificial neural networks (ANN), convolutional neural networks (CNN), long-short term memory (LSTM) networks, as well as linear models were applied for forecasting purposes. The Feature Selection with Annealing (FSA) algorithm was used to select the features from about 1000 possible predictors obtained from 26 technical indicators with specific periods and their lags. In addition to this, the Boruta feature selection algorithm was applied as a baseline feature selection method. The dependent variables consisted of daily logarithmic returns and daily trends of ten financial data sets, including cryptocurrency and different stocks. Experiments indicate that the FSA algorithm increased the performance of ML models regardless of the problem type. The FSA hybrid machine learning models showed better performance in 10 out of 10 data sets for regression and 8 out of 10 data sets for classification. None of the hybrid Boruta models outperformed the hybrid FSA models. However, the BOR-CNN model performance was comparable to the best model for 4 out of 10 data sets for regression estimates. BOR-LR and BOR-CNN models showed comparable performance with the best hybrid FSA models in 2 out of 10 datasets for classification. FSA was observed to improve the model performance in both better performance metrics as well as a decreased computation time by providing a lower dimensional input feature space."}, "cited_paper_content": {"title": "Ascertaining Price Formation In Cryptocurrency Markets With Deeplearning", "abstract": "The cryptocurrency market is amongst the fastest-growing of all the financial markets in the world. Unlike traditional markets, such as equities, foreign exchange and commodities, cryptocurrency market is considered to have larger volatility and illiquidity. This paper is inspired by the recent success of using deep learning for stock market prediction. In this work, we analyze and present the characteristics of the cryptocurrency market in a high-frequency setting. In particular, we applied a deep learning approach to predict the direction of the mid-price changes on the upcoming tick. We monitored live tick-level data from $8$ cryptocurrency pairs and applied both statistical and machine learning techniques to provide a live prediction. We reveal that promising results are possible for cryptocurrencies, and in particular, we achieve a consistent $78\\%$ accuracy on the prediction of the mid-price movement on live exchange rate of Bitcoins vs US dollars."}, "keywords": ["deep learning"], "citation_intent": "background"}
{"citing_id": "2303.04178v1", "cited_id": "1807.03819", "section_title": "Resources Needed For Picante", "citation": "To generate 2 22 reduced samples, 2 #REFR /n matrices must be reduced (one n \u00d7 n matrix produces 2n reduced samples, see \u00a7 4.1).", "text_before_citation": ["The total cost of PICANTE is the sum of the resources needed to preprocess data, train the model, and recover the secret.", "Data preprocessing is the most resource intensive part of PICANTE."], "text_after_citation": ["As the dimension increases, the number of matrices scales down linearly.", "To avoid the exponential cost of BKZ-reduction #OTHEREFR , we fix the block size at most \u03b2 = 20 so that the preprocessing step scales polynomially with n and log q.", "See discussion of the parameter choices for preprocessing in \u00a7 5.4 below.", "In practice, to save resources, we choose smaller \u03b2 for larger dimensions.", "Table 5 reports the preprocessing resources (in cpu hours) for each n. Our preprocessing is fully parallelizable."], "citing_paper_content": {"title": "Salsa Picante: A Machine Learning Attack On Lwe With Binary Secrets", "abstract": "The Learning With Errors (LWE) problem is one of the major hard problems in post-quantum cryptography. For example, 1) the only Key Exchange Mechanism KEM standardized by NIST [14] is based on LWE; and 2) current publicly available Homomorphic Encryption (HE) libraries are based on LWE. NIST KEM schemes use random secrets, but homomorphic encryption schemes use binary or ternary secrets, for efficiency reasons. In particular, sparse binary secrets have been proposed, but not standardized [2], for HE. Prior work SALSA [49] demonstrated a new machine learning attack on sparse binary secrets for the LWE problem in small dimensions (up to n = 128) and low Hamming weights (up to h = 4). However, this attack assumed access to millions of LWE samples, and was not scaled to higher Hamming weights or dimensions. Our attack, PICANTE, reduces the number of samples required to just m = 4n samples. Moreover, it can recover secrets with much larger dimensions (up to 350) and Hamming weights (roughly n/10, or h = 33 for n = 300). To achieve this, we introduce a preprocessing step which allows us to generate the training data from a linear number of samples and changes the distribution of the training data to improve transformer training. We also improve the distinguisher/secret recovery methods of SALSA and introduce a novel cross-attention recovery mechanism which allows us to read-off the secret directly from the trained models."}, "cited_paper_content": {"title": "Universal Transformers", "abstract": "Recurrent neural networks (RNNs) sequentially process data by updating their state with each new data point, and have long been the de facto choice for sequence modeling tasks. However, their inherently sequential computation makes them slow to train. Feed-forward and convolutional architectures have recently been shown to achieve superior results on some sequence modeling tasks such as machine translation, with the added advantage that they concurrently process all inputs in the sequence, leading to easy parallelization and faster training times. Despite these successes, however, popular feed-forward sequence models like the Transformer fail to generalize in many simple tasks that recurrent models handle with ease, e.g. copying strings or even simple logical inference when the string or formula lengths exceed those observed at training time. We propose the Universal Transformer (UT), a parallel-in-time self-attentive recurrent sequence model which can be cast as a generalization of the Transformer model and which addresses these issues. UTs combine the parallelizability and global receptive field of feed-forward sequence models like the Transformer with the recurrent inductive bias of RNNs. We also add a dynamic per-position halting mechanism and find that it improves accuracy on several tasks. In contrast to the standard Transformer, under certain assumptions, UTs can be shown to be Turing-complete. Our experiments show that UTs outperform standard Transformers on a wide range of algorithmic and language understanding tasks, including the challenging LAMBADA language modeling task where UTs achieve a new state of the art, and machine translation where UTs achieve a 0.9 BLEU improvement over Transformers on the WMT14 En-De dataset."}, "keywords": ["2n reduced samples", "one n \u00d7"], "citation_intent": "background"}
{"citing_id": "2304.06868v1", "cited_id": "1910.11664", "section_title": "Results And Discussion", "citation": "This result aligns with the intuition that the Fourier tempogram is the most similar to the CQT representation in #REFR as it tends to have upper harmonics but not sub-harmonics.", "text_before_citation": ["Since calibration results correlate with model performance, i.e.", "a model with more linear and less disperse calibration curves leads to more consistent estimations #OTHEREFR , we look at calibration results for each model, depicted in Figure 3 .", "The Fourier tempogram shows more consistent calibration results across data distributions and leads to smoother curves for both synthetic and real data."], "text_after_citation": ["On the contrary, the autocorrelation tempogram results present the most variation, meaning that for similar tempo values, the model struggles to assign the same tempo output.", "Surprisingly, the hybrid tempogram does worse than the Fourier tempogram, which suggests that the multiplication of the autocorrelation and Fourier tempograms removes harmonic information that the model relies on to perform tempo estimation.", "The experiments with the different data distributions show that the model does not benefit from the wider range of a log-uniform distribution, and instead is able to extrapolate from a log-normal distribution centered in the lower tempi end (70 bpm).", "When we look at this result closely, an explanation may be found in Figure 2 .", "Given a tempo distribution expressed in linear BPM values (Figure 2 top), the model will be fed a slightly different tempo distribution ( Figure  2 bottom) because of the logarithmic effect of the input tempogram."], "citing_paper_content": {"title": "Tempo Vs. Pitch: Understanding Self-Supervised Tempo Estimation", "abstract": "Self-supervision methods learn representations by solving pretext tasks that do not require human-generated labels, alleviating the need for time-consuming annotations. These methods have been applied in computer vision, natural language processing, environmental sound analysis, and recently in music information retrieval, e.g. for pitch estimation. Particularly in the context of music, there are few insights about the fragility of these models regarding different distributions of data, and how they could be mitigated. In this paper, we explore these questions by dissecting a self-supervised model for pitch estimation adapted for tempo estimation via rigorous experimentation with synthetic data. Specifically, we study the relationship between the input representation and data distribution for self-supervised tempo estimation."}, "cited_paper_content": {"title": "Spice: Self-Supervised Pitch Estimation", "abstract": "We propose a model to estimate the fundamental frequency in monophonic audio, often referred to as pitch estimation. We acknowledge the fact that obtaining ground truth annotations at the required temporal and frequency resolution is a particularly daunting task. Therefore, we propose to adopt a self-supervised learning technique, which is able to estimate (relative) pitch without any form of supervision. The key observation is that pitch shift maps to a simple translation when the audio signal is analysed through the lens of the constant-Q transform (CQT). We design a self-supervised task by feeding two shifted slices of the CQT to the same convolutional encoder, and require that the difference in the outputs is proportional to the corresponding difference in pitch. In addition, we introduce a small model head on top of the encoder, which is able to determine the confidence of the pitch estimate, so as to distinguish between voiced and unvoiced audio. Our results show that the proposed method is able to estimate pitch at a level of accuracy comparable to fully supervised models, both on clean and noisy audio samples, yet it does not require access to large labeled datasets"}, "keywords": ["Fourier tempogram"], "citation_intent": "result"}
{"citing_id": "2304.04445v1", "cited_id": "1202.2336", "section_title": "Prdos", "citation": "To ensure that this oracle has query time O(log h), we generalize the argument of Wulff-Nilsen #REFR from full oracles to partial ones.", "text_before_citation": ["The two paths P u,u , P v,v have the property that", "w(P u,u ), w(P v,v ) \u2264 h \u2022 d G (u, v) .", "Intuitively, given a query (u, v), the partial PRDO provides a u \u2212 v path P u,v (rather than paths P u,u , P v,v ) if and only if the Thorup-Zwick oracle does so while employing only the first h levels of the oracle.", "We note that using the first h levels of the TZ oracle indeed guarantees almost all the aforementioned properties.", "The only problem is that the query time of such partial oracle is O(h), rather than O(log h)."], "text_after_citation": ["In the second step, to find a low-stretch path between u and v , we construct an interactive emulator for the set S.", "An interactive emulator is an oracle that similarly to PRDOs, provides approximate shortest paths between two queried vertices.", "However, as opposed to a PRDO, the provided path does not use actual edges in the original graph, but rather employs virtual edges that belong to some sparse low-stretch emulator of the graph.", "Our construction of interactive emulator is based on the distance oracle of Mendel and Naor #OTHEREFR . Recall that the Mendel-Naor oracle is not path-reporting.", "Built for N -vertex graph and a parameter k 1 , this oracle provides stretch O(k 1 ), has size O(N 1+ 1 k 1 ), and has query time O(1)."], "citing_paper_content": {"title": "Path-Reporting Distance Oracles With Near-Logarithmic Stretch And Linear Size", "abstract": "Given an n-vertex undirected graph G = (V, E, w), and a parameter k \u2265 1, a pathreporting distance oracle (or PRDO) is a data structure of size S(n, k), that given a query (u, v) \u2208 V 2 , returns an f (k)-approximate shortest u \u2212 v path P in G within time q(k) + O(|P |). Here S(n, k), f (k) and q(k) are arbitrary (hopefully slowly-growing) functions. A distance oracle that only returns an approximate estimated(u, v) of the distance d G (u, v) between the queried vertices is called a non-path-reporting distance oracle. A landmark PRDO due to Thorup and Zwick [54] has S(n, k) = O(k \u2022 n 1+ 1 k), f (k) = 2k \u2212 1 and q(k) = O(k). Wulff-Nilsen [57] devised an improved query algorithm for this oracle with q(k) = O(log k). The size of this oracle is \u2126(n log n) for all k. Elkin and Pettie [28] devised a PRDO with S(n, k) = O(log k \u2022 n 1+ 1 k), f (k) = O(k log 4/3 7) and q(k) = O(log k). Neiman and Shabat [44] recently devised an improved PRDO with S(n, k) = O(n 1+ 1 k), f (k) = O(k log 4/3 4) and q(k) = O(log k). These oracles (of [28, 44]) can be much sparser than O(n log n) (the oracle of [44] can have linear size), but their stretch is polynomially larger than the optimal bound of 2k \u2212 1. On the other hand, a long line of non-path-reporting distance oracles culminated in a celebrated result by Chechik [12], in which S(n, k) = O(n 1+ 1 k), f (k) = 2k \u2212 1 and q(k) = O(1). In this paper we make a dramatic progress in bridging the gap between path-reporting and non-path-reporting distance oracles. In particular, we devise a PRDO with size S(n, k) = O k log log n log n"}, "cited_paper_content": {"title": "Approximate Distance Oracles With Improved Query Time", "abstract": "Given an undirected graph $G$ with $m$ edges, $n$ vertices, and non-negative edge weights, and given an integer $k\\geq 2$, we show that a $(2k-1)$-approximate distance oracle for $G$ of size $O(kn^{1 + 1/k})$ and with $O(\\log k)$ query time can be constructed in $O(\\min\\{kmn^{1/k},\\sqrt km + kn^{1 + c/\\sqrt k}\\})$ time for some constant $c$. This improves the $O(k)$ query time of Thorup and Zwick. Furthermore, for any $0<\\epsilon \\leq 1$, we give an oracle of size $O(kn^{1 + 1/k})$ that answers $((2 + \\epsilon)k)$-approximate distance queries in $O(1/\\epsilon)$ time. At the cost of a $k$-factor in size, this improves the $128k$ approximation achieved by the constant query time oracle of Mendel and Naor and approaches the best possible tradeoff between size and stretch, implied by a widely believed girth conjecture of Erd\\H{o}s. We can match the $O(n^{1 + 1/k})$ size bound of Mendel and Naor for any constant $\\epsilon>0$ and $k = O(\\log n/\\log\\log n)$."}, "keywords": ["full oracles"], "citation_intent": "method"}
{"citing_id": "2305.00724v1", "cited_id": "1811.03508", "section_title": "Results And Discussion", "citation": "This is somewhat contrary to the results obtained in LDP paper #REFR , but it is apparently an advantage of RF, since it considers each feature separately, while calculating tree splits.", "text_before_citation": ["Additionally, for each value, we also calculate the absolute average difference between its result and the best result for a given hyperparameter on each dataset.", "The lower the absolute difference, the better, since it means that a given hyperparameter value, on average, gives the best results among all its possible values. results are presented in Table 3 .", "For the number of bins, we can clearly select 50 bins as the optimal value.", "While 30 bins gave the best results the same number of times, on average they performed worse compared to the optimal hyperparameter value.", "Similarly, for normalization it is evident that we do not need to perform any kind of normalization, since using no normalization obtained the best results on majority of datasets and on average."], "text_after_citation": ["For aggregation method, the results are very close, both for number of wins and average difference compared to the best result.", "In this case, the choice does not matter that much, and we choose the simpler histogram method.", "The linear scale obtained much better results on average than the log scale, so the choice is obvious.", "Overall, this means that we can confidently recommend default values for all LDP hyperparameters, and tuning them is not particularly helpful.", "This dramatically decreases the computational cost, while having little effect on accuracy on average, which is a desirable tradeoff in a baseline method. Table 3 ."], "citing_paper_content": {"title": "Strengthening Structural Baselines For Graph Classification Using Local Topological Profile", "abstract": "We present the analysis of the topological graph descriptor Local Degree Profile (LDP), which forms a widely used structural baseline for graph classification. Our study focuses on model evaluation in the context of the recently developed fair evaluation framework, which defines rigorous routines for model selection and evaluation for graph classification, ensuring reproducibility and comparability of the results. Based on the obtained insights, we propose a new baseline algorithm called Local Topological Profile (LTP), which extends LDP by using additional centrality measures and local vertex descriptors. The new approach provides the results outperforming or very close to the latest GNNs for all datasets used. Specifically, state-of-the-art results were obtained for 4 out of 9 benchmark datasets. We also consider computational aspects of LDP-based feature extraction and model construction to propose practical improvements affecting execution speed and scalability. This allows for handling modern, large datasets and extends the portfolio of benchmarks used in graph representation learning. As the outcome of our work, we obtained LTP as a simple to understand, fast and scalable, still robust baseline, capable of outcompeting modern graph classification models such as Graph Isomorphism Network (GIN). We provide open-source implementation at GitHub."}, "cited_paper_content": {"title": "A Simple Yet Effective Baseline For Non-Attribute Graph Classification", "abstract": "Graphs are complex objects that do not lend themselves easily to typical learning tasks. Recently, a range of approaches based on graph kernels or graph neural networks have been developed for graph classification and for representation learning on graphs in general. As the developed methodologies become more sophisticated, it is important to understand which components of the increasingly complex methods are necessary or most effective. As a first step, we develop a simple yet meaningful graph representation, and explore its effectiveness in graph classification. We test our baseline representation for the graph classification task on a range of graph datasets. Interestingly, this simple representation achieves similar performance as the state-of-the-art graph kernels and graph neural networks for non-attributed graph classification. Its performance on classifying attributed graphs is slightly weaker as it does not incorporate attributes. However, given its simplicity and efficiency, we believe that it still serves as an effective baseline for attributed graph classification. Our graph representation is efficient (linear-time) to compute. We also provide a simple connection with the graph neural networks. Note that these observations are only for the task of graph classification while existing methods are often designed for a broader scope including node embedding and link prediction. The results are also likely biased due to the limited amount of benchmark datasets available. Nevertheless, the good performance of our simple baseline calls for the development of new, more comprehensive benchmark datasets so as to better evaluate and analyze different graph learning methods. Furthermore, given the computational efficiency of our graph summary, we believe that it is a good candidate as a baseline method for future graph classification (or even other graph learning) studies."}, "keywords": ["LDP paper"], "citation_intent": "result"}
{"citing_id": "2304.01046v1", "cited_id": "2002.04326", "section_title": "Dataset", "citation": "The dataset we used was the ReClor #REFR training and validation sets, which contain 5138 samples.", "text_before_citation": [], "text_after_citation": ["Each data point has a context, a question, and 4 answer choices.", "Exactly 1 answer choice is correct; each of the other choices is either logically or factually inaccurate.", "The ReClor questions are sourced from past GMAT and LSAT exams and high-quality practice exams."], "citing_paper_content": {"title": "Polytuplet Loss: A Reverse Approach To Training Reading Comprehension And Logical Reasoning Models", "abstract": "Throughout schooling, students are tested on reading comprehension and logical reasoning. Students have developed various strategies for completing such exams, some of which are generally thought to outperform others. One such strategy involves emphasizing relative accuracy over absolute accuracy and can theoretically produce the correct answer without full knowledge of the information required to solve the question. This paper examines the effectiveness of applying such a strategy to train transfer learning models to solve reading comprehension and logical reasoning questions. The models were evaluated on the ReClor dataset, a challenging reading comprehension and logical reasoning benchmark. While previous studies targeted logical reasoning skills, we focus on a general training method and model architecture. We propose the polytuplet loss function, an extension of the triplet loss function, to ensure prioritization of learning the relative correctness of answer choices over learning the true accuracy of each choice. Our results indicate that models employing polytuplet loss outperform existing baseline models. Although polytuplet loss is a promising alternative to other contrastive loss functions, further research is required to quantify the benefits it may present. 1"}, "cited_paper_content": {"title": "Reclor: A Reading Comprehension Dataset Requiring Logical Reasoning", "abstract": "Recent powerful pre-trained language models have achieved remarkable performance on most of the popular datasets for reading comprehension. It is time to introduce more challenging datasets to push the development of this field towards more comprehensive reasoning of text. In this paper, we introduce a new Reading Comprehension dataset requiring logical reasoning (ReClor) extracted from standardized graduate admission examinations. As earlier studies suggest, human-annotated datasets usually contain biases, which are often exploited by models to achieve high accuracy without truly understanding the text. In order to comprehensively evaluate the logical reasoning ability of models on ReClor, we propose to identify biased data points and separate them into EASY set while the rest as HARD set. Empirical results show that the state-of-the-art models have an outstanding ability to capture biases contained in the dataset with high accuracy on EASY set. However, they struggle on HARD set with poor performance near that of random guess, indicating more research is needed to essentially enhance the logical reasoning ability of current models."}, "keywords": ["ReClor training", "dataset"], "citation_intent": "method"}
{"citing_id": "2304.10701v1", "cited_id": "1904.02868", "section_title": "Related Work 2.1 Data Valuation", "citation": "To reduce the computation cost, Monte-Carlo and gradient-based techniques are proposed to approximate data SV when using large dataset to train the complex models #REFR . Jia et al.", "text_before_citation": ["And most of efforts have been devoted to the last two directions.", "For strategies of data valuation, the commonly-used approach is based on marginal contribution (MC).", "As the basic method for data valuation, LOO is used to evaluate the value of training data point X i by observational change of model performance when leaving out that data point from the training dataset.", "To overcome inaccuracy and strict desirability of LOO, SV originated from Cooperative Game Theory (CGT) is widely used to measure the contribution of data #OTHEREFR .", "Considering the joining sequence of each training data point, SV needs to calculate the marginal performance of all possible subsets in which the time complexity is exponential."], "text_after_citation": ["#OTHEREFR propose an efficient approximation approach to estimate the exact SV for K-nearest-neighor (KNN).", "Other works also apply these to other fields such as collaboration machine learning #OTHEREFR .", "Though several works devotes to accelerate the computation of traditional SV, it is still impossible to use these approaches in practical due to their inevitable expensive computation.", "When it comes to performance metric of data valuation, some recent studies start to consider data-driven metric rather than model-driven metric. Tay et al.", "#OTHEREFR propose Cosine Gradient Shapley Value (CGSV) which leverages cosine similarity between gradient vectors of the model to evaluate the contribution of each participant in collaborative machine learning."], "citing_paper_content": {"title": "Matching-Based Data Valuation For Generative Model", "abstract": "Data valuation is critical in machine learning, as it helps enhance model transparency and protect data properties. Existing data valuation methods have primarily focused on discriminative models, neglecting deep generative models that have recently gained considerable attention. Similar to discriminative models, there is an urgent need to assess data contributions in deep generative models as well. However, previous data valuation approaches mainly relied on discriminative model performance metrics and required model retraining. Consequently, they cannot be applied directly and efficiently to recent deep generative models, such as generative adversarial networks and diffusion models, in practice. To bridge this gap, we formulate the data valuation problem in generative models from a similarity-matching perspective. Specifically, we introduce Generative Model Valuator (GMValuator), the first model-agnostic approach for any generative models, designed to provide data valuation for generation tasks. We have conducted extensive experiments to demonstrate the effectiveness of the proposed method. To the best of their knowledge, GMValuator is the first work that offers a training-free, post-hoc data valuation strategy for deep generative models. Keywords Generative Model \u2022 Data Valuation 1 Introduction Recently, applications of generative models have exploded across many fields. For example, ChatGPT [1] is capable of automatically generating code by utilizing a generative model, and Stable Diffusion [2] can generate photo-realistic images. As a form of unsupervised machine learning, generative models learn the probabilities distribution of the training datasets from which they sample the synthetic data, which is similar to the original training dataset. Data is regarded as the fuel for the model artificial intelligence (i.e., deep learning) [3]. Obtaining high-quality generative models requires a large amount of data, and including informative data samples related to the task is also an important ingredient in the training process of deep generative models. From a data privacy perspective, personal data is protected by various regulations including General Data Protection Regulation (GDPR) [4] and become a valuable asset. Privacy Calculus [5] requires assessing the trade-off between potential benefits and expected costs or risks of disclosing private data. Recent AI Generated Content further arouse people's attention[6]. Therefore, measuring the value of the collected data in generative model training is vital to achieving the training objectives. Unfortunately, there is no existing studies devoting to data valuation for generative model. Most of the previous data valuation methods are for discriminate model [7]. For example, some performance metrics (e.g., loss value, accuracy) of the machine learning model are always used to calculate the score function while the strategy of data valuation is Leave-one-out (LOO), Shapley Value (SV) and Banzhaf Index (BI), etc [8, 9, 10]."}, "cited_paper_content": {"title": "Data Shapley: Equitable Valuation Of Data For Machine Learning", "abstract": "As data becomes the fuel driving technological and economic growth, a fundamental challenge is how to quantify the value of data in algorithmic predictions and decisions. For example, in healthcare and consumer markets, it has been suggested that individuals should be compensated for the data that they generate, but it is not clear what is an equitable valuation for individual data. In this work, we develop a principled framework to address data valuation in the context of supervised machine learning. Given a learning algorithm trained on $n$ data points to produce a predictor, we propose data Shapley as a metric to quantify the value of each training datum to the predictor performance. Data Shapley value uniquely satisfies several natural properties of equitable data valuation. We develop Monte Carlo and gradient-based methods to efficiently estimate data Shapley values in practical settings where complex learning algorithms, including neural networks, are trained on large datasets. In addition to being equitable, extensive experiments across biomedical, image and synthetic data demonstrate that data Shapley has several other benefits: 1) it is more powerful than the popular leave-one-out or leverage score in providing insight on what data is more valuable for a given learning task; 2) low Shapley value data effectively capture outliers and corruptions; 3) high Shapley value data inform what type of new data to acquire to improve the predictor."}, "keywords": ["Monte-Carlo", "large dataset"], "citation_intent": "method"}
{"citing_id": "2304.06275v1", "cited_id": "1803.08024", "section_title": "Cross-Modal Retrieval", "citation": "To capture fine-grained interplay between modalities, SCAN #REFR introduce stacked cross attention to enable attention with context from both image and sentence.", "text_before_citation": ["The extra label information is employed to boost the discrimination of common representations.", "However, the lack of annotations limited the practicability in the real world.", "Our proposed approach falls in the category of unsupervised methods.", "For example, Wang #OTHEREFR uses a twobranch neural networks to learn the joint embeddings of multimodal data.", "Inspired by hard negative mining, VSE++ #OTHEREFR uses hard negatives to improve the retrieval performance."], "text_after_citation": ["Recently, motivated by the powerful learning ability of graph model, GSMN #OTHEREFR and SGRAF #OTHEREFR construct graph structure for multimodal data to benefit the learning of finegrained correspondence.", "Although existing works have achieved remarkable results, they usually depend on the correct correspondence among multimodal data and cannot tackle the noise issue.", "Thus, it is significant to explore how to learn cross-modal retrieval with noisy correspondence, but which is rarely touched in previous studies."], "citing_paper_content": {"title": "Noisy Correspondence Learning With Meta Similarity Correction", "abstract": "Despite the success of multimodal learning in crossmodal retrieval task, the remarkable progress relies on the correct correspondence among multimedia data. However, collecting such ideal data is expensive and timeconsuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain mismatched pairs. Training on such noisy correspondence datasets causes performance degradation because the cross-modal retrieval methods can wrongly enforce the mismatched data to be similar. To tackle this problem, we propose a Meta Similarity Correction Network (MSCN) to provide reliable similarity scores. We view a binary classification task as the meta-process that encourages the MSCN to learn discrimination from positive and negative meta-data. To further alleviate the influence of noise, we design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises, including Flickr30K, MS-COCO, and Conceptual Captions. Our code is publicly available. 1"}, "cited_paper_content": {"title": "Stacked Cross Attention For Image-Text Matching", "abstract": "In this paper, we study the problem of image-text matching. Inferring the latent semantic alignment between objects or other salient stuff (e.g. snow, sky, lawn) and the corresponding words in sentences allows to capture fine-grained interplay between vision and language, and makes image-text matching more interpretable. Prior work either simply aggregates the similarity of all possible pairs of regions and words without attending differentially to more and less important words or regions, or uses a multi-step attentional process to capture limited number of semantic alignments which is less interpretable. In this paper, we present Stacked Cross Attention to discover the full latent alignments using both image regions and words in a sentence as context and infer image-text similarity. Our approach achieves the state-of-the-art results on the MS-COCO and Flickr30K datasets. On Flickr30K, our approach outperforms the current best methods by 22.1% relatively in text retrieval from image query, and 18.2% relatively in image retrieval with text query (based on Recall@1). On MS-COCO, our approach improves sentence retrieval by 17.8% relatively and image retrieval by 16.6% relatively (based on Recall@1 using the 5K test set). Code has been made available at: https://github.com/kuanghuei/SCAN."}, "keywords": ["stacked cross attention"], "citation_intent": "background"}
{"citing_id": "2304.00948v1", "cited_id": "2002.05227", "section_title": "Ii. Related Work", "citation": "In #REFR , the Riemannian Brownian motion model was constructed over the manifold by using VAE (R-VAE).", "text_before_citation": ["A normal prior, on the other hand, causes the approximate posterior to be over-regularized, leading to a less effective learned latent representation of the input.", "Manifold learning.", "The goal of manifold learning approaches is to infer latent representations that reflect the inherent geometric structure of data.", "Such methods are mainly used for data visualization and dimensionality reduction and recently are used to improve the performance of generative models #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR .", "Traditional non-linear manifold learning methods like local tangent space alignment #OTHEREFR and linear embedding #OTHEREFR maintain local structures and extract the features of data manifolds."], "text_after_citation": ["The Riemannian structure overcomes the identifiability problem by providing a meaningful representation which is not affected by reparametrizations.", "However, Riemannian manifold-based models cannot accurately find the shortest path between two samples that are far away from each other.", "Generally, Riemannian geometry methods use Ordinary Differential Equation (ODE) to find the shortest interpolation curve in the latent space.", "ODE is computationally expensive and cannot find the shortest interpolation curve more consistently than straight-length methods.", "The interpolation among data samples is widely used to attain a smooth transformation from one sample to another."], "citing_paper_content": {"title": "Vtae: Variational Transformer Autoencoder With Manifolds Learning", "abstract": "Deep generative models have demonstrated successful applications in learning non-linear data distributions through a number of latent variables and these models use a non-linear function (generator) to map latent samples into the data space. On the other hand, the non-linearity of the generator implies that the latent space shows an unsatisfactory projection of the data space, which results in poor representation learning. This weak projection, however, can be addressed by a Riemannian metric, and we show that geodesics computation and accurate interpolations between data samples on the Riemannian manifold can substantially improve the performance of deep generative models. In this paper, a Variational spatial-Transformer AutoEncoder (VTAE) is proposed to minimize geodesics on a Riemannian manifold and improve representation learning. In particular, we carefully design the variational autoencoder with an encoded spatial-Transformer to explicitly expand the latent variable model to data on a Riemannian manifold, and obtain global context modelling. Moreover, to have smooth and plausible interpolations while traversing between two different objects' latent representations, we propose a geodesic interpolation network different from the existing models that use linear interpolation with inferior performance. Experiments on benchmarks show that our proposed model can improve predictive accuracy and versatility over a range of computer vision tasks, including image interpolations, and reconstructions."}, "cited_paper_content": {"title": "Variational Autoencoders With Riemannian Brownian Motion Priors", "abstract": "Variational Autoencoders (VAEs) represent the given data in a low-dimensional latent space, which is generally assumed to be Euclidean. This assumption naturally leads to the common choice of a standard Gaussian prior over continuous latent variables. Recent work has, however, shown that this prior has a detrimental effect on model capacity, leading to subpar performance. We propose that the Euclidean assumption lies at the heart of this failure mode. To counter this, we assume a Riemannian structure over the latent space, which constitutes a more principled geometric view of the latent codes, and replace the standard Gaussian prior with a Riemannian Brownian motion prior. We propose an efficient inference scheme that does not rely on the unknown normalizing factor of this prior. Finally, we demonstrate that this prior significantly increases model capacity using only one additional scalar parameter."}, "keywords": ["manifold", "Riemannian Brownian motion"], "citation_intent": "method"}
{"citing_id": "2303.11101v2", "cited_id": "1911.05722", "section_title": "Byol Method:", "citation": "BYOL method is highly sensitive to the exponential moving average (EMA) value of a momentum encoder #REFR .", "text_before_citation": ["We used an Adam optimizer #OTHEREFR with a learning rate of 1e-3 and 2 regularization parameter 1e-6."], "text_after_citation": ["Therefore, we followed the same EMA scheduler as in the original implementation #OTHEREFR , of which the momentum starts from 0.996 to 1.", "We should also note that we did not use the EMA scheduler for the experiments with short epochs.", "SwAV method: We mostly followed the same settings as the SimCLR method, such as an SGD optimizer with a learning rate of 0.1 and weight decay of 1e-4.", "In the case of pretraining on ImageNet, we used the temperature value of 0.1 and an epsilon value of 0.02, while we froze the 3,000-dimensional prototypes for one epoch.", "For every other cases, we froze the 100-dimensional prototypes for ten epochs."], "citing_paper_content": {"title": "Coreset Sampling From Open-Set For Fine-Grained Self-Supervised Learning", "abstract": "Deep learning in general domains has constantly been extended to domain-specific tasks requiring the recognition of fine-grained characteristics. However, real-world applications for fine-grained tasks suffer from two challenges: a high reliance on expert knowledge for annotation and necessity of a versatile model for various downstream tasks in a specific domain (e.g., prediction of categories, bounding boxes, or pixel-wise annotations). Fortunately, the recent self-supervised learning (SSL) is a promising approach to pretrain a model without annotations, serving as an effective initialization for any downstream tasks. Since SSL does not rely on the presence of annotation, in general, it utilizes the large-scale unlabeled dataset, referred to as an open-set. In this sense, we introduce a novel Open-Set Self-Supervised Learning problem under the assumption that a large-scale unlabeled open-set is available, as well as the fine-grained target dataset, during a pretraining phase. In our problem setup, it is crucial to consider the distribution mismatch between the open-set and target dataset. Hence, we propose SimCore algorithm to sample a coreset, the subset of an open-set that has a minimum distance to the target dataset in the latent space. We demonstrate that SimCore significantly improves representation learning performance through extensive experimental settings, including eleven fine-grained datasets and seven open-sets in various downstream tasks."}, "cited_paper_content": {"title": "Momentum Contrast For Unsupervised Visual Representation Learning", "abstract": "We present Momentum Contrast (MoCo) for unsupervised visual representation learning. From a perspective on contrastive learning as dictionary look-up, we build a dynamic dictionary with a queue and a moving-averaged encoder. This enables building a large and consistent dictionary on-the-fly that facilitates contrastive unsupervised learning. MoCo provides competitive results under the common linear protocol on ImageNet classification. More importantly, the representations learned by MoCo transfer well to downstream tasks. MoCo can outperform its supervised pre-training counterpart in 7 detection/segmentation tasks on PASCAL VOC, COCO, and other datasets, sometimes surpassing it by large margins. This suggests that the gap between unsupervised and supervised representation learning has been largely closed in many vision tasks."}, "keywords": ["momentum encoder"], "citation_intent": "method"}
{"citing_id": "2305.00656v1", "cited_id": "1908.06148", "section_title": "E. Discussion", "citation": "This is in contrast to the results reported in FiFTy #REFR , where the inference time of FiFTy can be significantly larger.", "text_before_citation": ["In summary, the results show that all three models have similar inference times, with minimal variance between runs."], "text_after_citation": ["Our models are optimized for efficient inference by utilizing cutting-edge techniques to minimize computational overhead and reduce latency.", "As a result, we were able to achieve inference times that are comparable to state-of-the-art models while still maintaining high accuracy.", "This demonstrates the effectiveness of our proposed models in real-world applications where fast and efficient inference is crucial.", "Similar to previous work #OTHEREFR , we observe that files with high entropy are difficult to classify because there is no statistical trace that the convolutional kernel can extract.", "Moreover, many files are container types that contain other files as embedded objects, e.g., pdf files that can contain embedded jpg images. As a result, classifiers behave erratically."], "citing_paper_content": {"title": "File Fragment Classification Using Light-Weight Convolutional Neural Networks", "abstract": "In digital forensics, file fragment classification is an important step toward completing file carving process. There exist several techniques to identify the type of file fragments without relying on meta-data, such as using features like header/footer and N-gram to identify the fragment type. Recently, convolutional neural network (CNN) models have been used to build classification models to achieve this task. However, the number of parameters in CNNs tends to grow exponentially as the number of layers increases. This results in a dramatic increase in training and inference time. In this paper, we propose lightweight file fragment classification models based on depthwise separable CNNs. The evaluation results show that our proposed models provide faster inference time with comparable accuracy as compared to the state-of-art CNN based models. In particular, our models were able to achieve an accuracy of 79% on the FFT-75 dataset with nearly 100K parameters and 164M FLOPs, which is 4x smaller and 6x faster than the state-of-the-art classifier in the literature."}, "cited_paper_content": {"title": "Fifty: Large-Scale File Fragment Type Identification Using Neural Networks", "abstract": "We present FiFTy, a modern file type identification tool for memory forensics and data carving. In contrast to previous approaches based on hand-crafted features, we design a compact neural network architecture, which uses a trainable embedding space, akin to successful natural language processing models. Our approach dispenses with explicit feature extraction which is a bottleneck in legacy systems. We evaluate the proposed method on a novel dataset with 75 file types - the most diverse and balanced dataset reported to date. FiFTy consistently outperforms all baselines in terms of speed, accuracy and individual misclassification rates. We achieved an average accuracy of 77.5% with processing speed of approx 38 sec/GB, which is better and more than an order of magnitude faster than the previous state-of-the-art tool - Sceadan (69% at 9 min/GB). Our tool and the corresponding dataset are available publicly online."}, "keywords": ["inference time", "FiFTy"], "citation_intent": "result"}
{"citing_id": "2303.09668v1", "cited_id": "1409.0575", "section_title": "B. Implementation Details", "citation": "For the feature extractor, HG-FEN with ResNet-50 as the backbone and Imagenet-pretrained model #REFR as the initialized weights.", "text_before_citation": ["All the experiments are implemented using PyTorch and run on a desktop with 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz and a single NVIDIA GeForce RTX 3090 GPU.", "we directly apply the publicly available detector of YOLOX #OTHEREFR trained by #OTHEREFR for MOT17, MOT20, and ablation study on MOT17."], "text_after_citation": ["We train our HG-FEN model on MOT17 and MOT20, its parameters are updated using the Adam optimizer #OTHEREFR with weight decay of 5 \u00d7 10 \u22124 .", "During the training procedure, the initial learning rate is 3.5\u00d710 \u22124 , input batch size is set as 64 and the resolution of every image is 256 \u00d7 128. Total training 120 epochs.", "For the tracker, we set four tracklet states, including tentative, confirmed, deleted and lost state.", "To initialize a new tracklet, in the first frame of each video sequence we set the tracklet with a confidence score larger than the new tracklet threshold of 0.7 to the confirm state, while subsequent frames in the case where the condition is met will be set to the tentative.", "For the lost tracklets, we keep them for 30 frames in case it appears again. We set \u03bb = 0.95 in Eq. #OTHEREFR ."], "citing_paper_content": {"title": "Rt-Track: Robust Tricks For Multi-Pedestrian Tracking", "abstract": "Object tracking is divided into single-object tracking (SOT) and multi-object tracking (MOT). MOT aims to maintain the identities of multiple objects across a series of continuous video sequences. In recent years, MOT has made rapid progress. However, modeling the motion and appearance models of objects in complex scenes still faces various challenging issues. In this paper, we design a novel direction consistency method for smooth trajectory prediction (STP-DC) to increase the modeling of motion information and overcome the lack of robustness in previous methods in complex scenes. Existing methods use pedestrian reidentification (Re-ID) to model appearance, however, they extract more background information which lacks discriminability in occlusion and crowded scenes. We propose a hyper-grain feature embedding network (HG-FEN) to enhance the modeling of appearance models, thus generating robust appearance descriptors. We also proposed other robustness techniques, including CF-ECM for storing robust appearance information and SK-AS for improving association accuracy. To achieve state-of-the-art performance in MOT, we propose a robust tracker named Rttrack, incorporating various tricks and techniques. It achieves 79.5 MOTA, 76.0 IDF1 and 62.1 HOTA on the test set of MOT17.Rt-track also achieves 77.9 MOTA, 78.4 IDF1 and 63.3 HOTA on MOT20, surpassing all published methods."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["Imagenet-pretrained model"], "citation_intent": "method"}
{"citing_id": "2303.03909v1", "cited_id": "1711.10275", "section_title": "B. Network Structure", "citation": "We then fuse them by performing convolutions, and utilize the same number of layers of 3D sparse deconvolutions #REFR as those convolutions used in the instance detection module to resume the point-level features hierarchically.", "text_before_citation": [", where C represents the number of instance categories, U represents the number of instances,x and\u1ef9 represent coordinate quantization offsets, z represents the height, l, m and q represent 3D bounding box size and \u03c9 represents orientation.", "Each score in heatmap\u00ca cij represents the confidence at location (i, j) belonging to the instance class c.", "For more details about CenterHead, please refer to #OTHEREFR .", "3) Upsample Fusion Module: In upsample fusion module, we aim to resume point-level features by fusing the spatio-temporal features and the instance features.", "We directly concatenate the spatio-temporal features and the instance features as the input."], "text_after_citation": ["To strengthen the instance information in the final features and maintain more details learned in different layers, we propagate the instance features from multi-resolution point clouds and form the instance pyramid.", "After concatenating with the corresponding spatio-temporal features, we inject them into the upsample fusion module.", "Finally, the upsample fusion module outputs the point-", "wise features F = {f i \u2208 R 3 } M i=1", ", which are then used to determine the moving label of LiDAR points after passing a softmax function."], "citing_paper_content": {"title": "Insmos: Instance-Aware Moving Object Segmentation In Lidar Data", "abstract": "Identifying moving objects is a crucial capability for autonomous navigation, consistent map generation, and future trajectory prediction of objects. In this paper, we propose a novel network that addresses the challenge of segmenting moving objects in 3D LiDAR scans. Our approach not only predicts point-wise moving labels but also detects instance information of main traffic participants. Such a design helps determine which instances are actually moving and which ones are temporarily static in the current scene. Our method exploits a sequence of point clouds as input and quantifies them into 4D voxels. We use 4D sparse convolutions to extract motion features from the 4D voxels and inject them into the current scan. Then, we extract spatio-temporal features from the current scan for instance detection and feature fusion. Finally, we design an upsample fusion module to output point-wise labels by fusing the spatio-temporal features and predicted instance information. We evaluated our approach on the LiDAR-MOS benchmark based on SemanticKITTI and achieved better moving object segmentation performance compared to state-of-theart methods, demonstrating the effectiveness of our approach in integrating instance information for moving object segmentation. Furthermore, our method shows superior performance on the Apollo dataset with a pre-trained model on SemanticKITTI, indicating that our method generalizes well in different scenes. The code and pre-trained models of our method will be released at https://github.com/nubot-nudt/InsMOS."}, "cited_paper_content": {"title": "3D Semantic Segmentation With Submanifold Sparse Convolutional Networks", "abstract": "Convolutional networks are the de-facto standard for analyzing spatio-temporal data such as images, videos, and 3D shapes. Whilst some of this data is naturally dense (e.g., photos), many other data sources are inherently sparse. Examples include 3D point clouds that were obtained using a LiDAR scanner or RGB-D camera. Standard \"dense\" implementations of convolutional networks are very inefficient when applied on such sparse data. We introduce new sparse convolutional operations that are designed to process spatially-sparse data more efficiently, and use them to develop spatially-sparse convolutional networks. We demonstrate the strong performance of the resulting models, called submanifold sparse convolutional networks (SSCNs), on two tasks involving semantic segmentation of 3D point clouds. In particular, our models outperform all prior state-of-the-art on the test set of a recent semantic segmentation competition."}, "keywords": ["point-level features", "3D sparse deconvolutions"], "citation_intent": "method"}
{"citing_id": "2304.12579v1", "cited_id": "1907.04595", "section_title": "Experiments", "citation": "The reasons behind a larger learning rate resulting in a smaller generalization error have been explored in #REFR ; Barrett & Dherin (2020).", "text_before_citation": ["Further discussions on why a small learning rate leads to a larger generalization error can be found in #OTHEREFR ; Barrett & Dherin (2020).", "The complexity of learning trajectory correlates with the generalization error under mild learning rate In Figure 3 , we carry out experiments under various settings.", "Each data point in the figure represents the average of three repeated experiments.", "The results demonstrate that both the generalization error and C(J t ) increase as the level of label noise is raised (Figure 3 Left).", "The another experiments measure C(J t ) and generalization error for different learning rate and discover that C(J t ) can capture the trend generalization error."], "text_after_citation": ["Additionally, Appendix A.7 discusses why a larger learning rate can lead to a smaller C(J t )."], "citing_paper_content": {"title": "Learning Trajectories Are Generalization Indicators", "abstract": "This paper aims to investigate the relation between learning trajectories of the Deep Neural Networks (DNNs) and their corresponding generalization capabilities when being optimized with gradient descent and stochastic gradient descent algorithms. In this paper, we construct Linear Approximation Function to model the trajectory information and we propose a new generalization bound with richer trajectory information based on this modelling. Our proposed generalization bound relies on the complexity of learning trajectory and the ratio between the bias and diversity of training set. Experimental results indicate that the proposed method effectively captures the generalization trend across various iterations, learning rates, and label noise levels across entire training process."}, "cited_paper_content": {"title": "Towards Explaining The Regularization Effect Of Initial Large Learning Rate In Training Neural Networks", "abstract": "Stochastic gradient descent with a large initial learning rate is widely used for training modern neural net architectures. Although a small initial learning rate allows for faster training and better test performance initially, the large learning rate achieves better generalization soon after the learning rate is annealed. Towards explaining this phenomenon, we devise a setting in which we can prove that a two layer network trained with large initial learning rate and annealing provably generalizes better than the same network trained with a small learning rate from the start. The key insight in our analysis is that the order of learning different types of patterns is crucial: because the small learning rate model first memorizes low-noise, hard-to-fit patterns, it generalizes worse on hard-to-generalize, easier-to-fit patterns than its large learning rate counterpart. This concept translates to a larger-scale setting: we demonstrate that one can add a small patch to CIFAR-10 images that is immediately memorizable by a model with small initial learning rate, but ignored by the model with large learning rate until after annealing. Our experiments show that this causes the small learning rate model's accuracy on unmodified images to suffer, as it relies too much on the patch early on."}, "keywords": ["smaller generalization error", "larger learning rate"], "citation_intent": "background"}
{"citing_id": "2304.10248v1", "cited_id": "0906.0483", "section_title": "Introduction", "citation": "However, this is no longer true when these vectors are not orthogonal-in fact, subtracting a best rank-1 approximation from X can even yield a tensor of higher rank #REFR .", "text_before_citation": ["One example is in latent variable model learning #OTHEREFR , where the vectors x i appearing in the decomposition are directly related to the model's parameters (for instance, each x i is the mean of a Gaussian component in a mixture model).", "If the vectors x i in (1) were orthogonal, then one could retrieve them by resorting to a greedy deflation procedure, first J. H. de M.", "Goulart's work was supported by the ANR LabEx CIMI (ANR-11-LABX-0040) within the French Programme \"Investissements d'Avenir.\".", "introduced by Hotelling for matrices #OTHEREFR : a best rank-1 approximation of X is computed and then subtracted from X, and the process is repeated r times.", "Algorithmically, each such approximation can be computed by power iteration #OTHEREFR ."], "text_after_citation": ["In some applications, this can in principle be circumvented by transforming X in such a way that it becomes a rank-r symmetric orthogonal decomposition, as long as r \u2264 n.", "For instance, in latent variable model learning the eigendecomposition of a matrix of second-order statistics can be exploited to obtain a whitening matrix W \u2208 R r\u00d7n such that the vectors", "x i = W x i \u2208 R r , i = 1, .", ". . , r, are pairwise orthogonal.", "An analysis of an algorithm employing this technique coupled with tensor power iteration was carried out in #OTHEREFR , including a robust estimation result quantifying the performance in the case one observes Y = X + E, in terms of the spectral norm of the perturbation E."], "citing_paper_content": {"title": "Hotelling Deflation On Large Symmetric Spiked Tensors", "abstract": "-Cet article \u00e9tudie l'algorithme de d\u00e9flation appliqu\u00e9 \u00e0 l'estimation d'un mod\u00e8le de rang faible sym\u00e9trique contenu dans un tenseur de grandes dimensions corrompu par un bruit additif gaussien. Plus pr\u00e9cis\u00e9ment, nous fournissons une caract\u00e9risation pr\u00e9cise de la performance en grandes dimensions de la d\u00e9flation en termes des alignements des vecteurs obtenus par approximations successives de rang 1 et de leurs poids, en supposant des corr\u00e9lations (fixes) non-triviales entre les composantes du mod\u00e8le. Notre analyse permet de comprendre le m\u00e9canisme de d\u00e9flation en pr\u00e9sence de bruit et peut \u00eatre exploit\u00e9e pour concevoir des m\u00e9thodes d'estimation plus efficaces."}, "cited_paper_content": {"title": "Subtracting A Best Rank-1 Approximation May Increase Tensor Rank", "abstract": "It has been shown that a best rank-R approximation of an order-k tensor may not exist when R>1 and k>2. This poses a serious problem to data analysts using tensor decompositions. It has been observed numerically that, generally, this issue cannot be solved by consecutively computing and subtracting best rank-1 approximations. The reason for this is that subtracting a best rank-1 approximation generally does not decrease tensor rank. In this paper, we provide a mathematical treatment of this property for real-valued 2x2x2 tensors, with symmetric tensors as a special case. Regardless of the symmetry, we show that for generic 2x2x2 tensors (which have rank 2 or 3), subtracting a best rank-1 approximation results in a tensor that has rank 3 and lies on the boundary between the rank-2 and rank-3 sets. Hence, for a typical tensor of rank 2, subtracting a best rank-1 approximation increases the tensor rank."}, "keywords": ["best rank-1 approximation"], "citation_intent": "background"}
{"citing_id": "2304.04358v1", "cited_id": "1702.05379", "section_title": "Introduction", "citation": "As a result, Wikipedia articles become the best bet for most users when searching answers for factual queries on the Web #REFR .", "text_before_citation": ["Information acquisition is one of the fundamental daily needs of human beings.", "Acquiring information from the Web is undoubtedly a convenient and efficient way.", "However, with the exponential growth of the Web, information on the Web becomes scattered and evolves quickly, making it challenging for users to acquire the expected information quickly."], "text_after_citation": ["The reason is that Wikipedia articles provide credible content in which most claims can be supported by references from reputable sources.", "While Wikipedia is a good source of answers for factual queries, the need for manual editing (crowd-sourcing and editor checking) curbs its growth of coverage on a broader range of information needs. What if Wikipedia articles could be automatically generated?", "In this paper, we introduce a new task, WEBBRAIN, exploring the capacity of generating short factual articles for queries via a large web corpus.", "Given a factual query, the goal of the task is to enable a system to mine supporting evidence from the Web and generate a short factual article in which the claims are supported by the mined evidence (defined in Section 3.1).", "One of the potential generation targets for WEBBRAIN is the first section of a new Wiki page, based on which we can further explore generating long factual articles (e.g., a complete Wiki page)."], "citing_paper_content": {"title": "Webbrain: Learning To Generate Factually Correct Articles For Queries By Grounding On Large Web Corpus", "abstract": "In this paper, we introduce a new NLP task-generating short factual articles with references for queries by mining supporting evidence from the Web. In this task, called WEBBRAIN, the ultimate goal is to generate a fluent, informative, and factually-correct short article (e.g., a Wikipedia article) for a factual query unseen in Wikipedia. To enable experiments on WEBBRAIN, we construct a large-scale dataset WebBrain-Raw by extracting English Wikipedia articles and their crawlable Wikipedia references. WebBrain-Raw is ten times larger than the previous biggest peer dataset, which can greatly benefit the research community. From WebBrain-Raw, we construct two task-specific datasets: WebBrain-R and WebBrain-G, which are used to train in-domain retriever and generator, respectively. Besides, we empirically analyze the performances of the current state-ofthe-art NLP techniques on WEBBRAIN and introduce a new framework ReGen, which enhances the generation factualness by improved evidence retrieval and task-specific pre-training for generation. Experiment results show that ReGen outperforms all baselines in both automatic and human evaluations. Our code and datasets are released at https://github.com/qhjqhj00/WebBrain."}, "cited_paper_content": {"title": "Why We Read Wikipedia", "abstract": "Wikipedia is one of the most popular sites on the Web, with millions of users relying on it to satisfy a broad range of information needs every day. Although it is crucial to understand what exactly these needs are in order to be able to meet them, little is currently known about why users visit Wikipedia. The goal of this paper is to fill this gap by combining a survey of Wikipedia readers with a log-based analysis of user activity. Based on an initial series of user surveys, we build a taxonomy of Wikipedia use cases along several dimensions, capturing users' motivations to visit Wikipedia, the depth of knowledge they are seeking, and their knowledge of the topic of interest prior to visiting Wikipedia. Then, we quantify the prevalence of these use cases via a large-scale user survey conducted on live Wikipedia with almost 30,000 responses. Our analyses highlight the variety of factors driving users to Wikipedia, such as current events, media coverage of a topic, personal curiosity, work or school assignments, or boredom. Finally, we match survey responses to the respondents' digital traces in Wikipedia's server logs, enabling the discovery of behavioral patterns associated with specific use cases. For instance, we observe long and fast-paced page sequences across topics for users who are bored or exploring randomly, whereas those using Wikipedia for work or school spend more time on individual articles focused on topics such as science. Our findings advance our understanding of reader motivations and behavior on Wikipedia and can have implications for developers aiming to improve Wikipedia's user experience, editors striving to cater to their readers' needs, third-party services (such as search engines) providing access to Wikipedia content, and researchers aiming to build tools such as recommendation engines."}, "keywords": ["Wikipedia articles"], "citation_intent": "background"}
{"citing_id": "2303.07287v1", "cited_id": "1906.05247", "section_title": "Application In Multi-Armed Bandit Problem", "citation": "Many previous UCB algorithms based on non-asymptotic inference in the literature assume that the sub-G parameter is a preset constant, see the algorithm in #REFR for instance.", "text_before_citation": ["(2022) ), so Algorithm 1 can be seen as an optimal algorithm.", "Compared with the traditional vanilla UCB, we do improve the constant.", "When Y k \u223c N (\u00b5 k , 1), the constant factor in regret bound in #OTHEREFR", "(2002) is 256, which is larger than 16(2 + \u221a 2) 2 in our theorem.", "When the UCB has unknown sub-G parameters, Theorem 4 first studies a feasible UCB algorithm with sub-G parameter plugging estimation."], "text_after_citation": ["Next, we give an simulation for Theorem 4 in two sub-G cases to verify the performance of estimated norms. Similar to #OTHEREFR", "(2019) ; , we design the three methods as follows:", "1.", "Use our method \u03d5(Y T k (t) ) with Estimated Norm in Theorem 4;", "2."], "citing_paper_content": {"title": "Tight Non-Asymptotic Inference Via Sub-Gaussian Intrinsic Moment Norm", "abstract": "In non-asymptotic statistical inferences, variance-type parameters of sub-Gaussian distributions play a crucial role. However, direct estimation of these parameters based on the empirical moment generating function (MGF) is infeasible. To this end, we recommend using a sub-Gaussian intrinsic moment norm [Buldygin and Kozachenko (2000), Theorem 1.3] through maximizing a series of normalized moments. Importantly, the recommended norm can not only recover the exponential moment bounds for the corresponding MGFs, but also lead to tighter Hoeffiding's sub-Gaussian concentration inequalities. In practice, we propose an intuitive way of checking sub-Gaussian data with a finite sample size by the sub-Gaussian plot. Intrinsic moment norm can be robustly estimated via a simple plug-in approach. Our theoretical results are applied to non-asymptotic analysis, including the multi-armed bandit."}, "cited_paper_content": {"title": "Bootstrapping Upper Confidence Bound", "abstract": "Upper Confidence Bound (UCB) method is arguably the most celebrated one used in online decision making with partial information feedback. Existing techniques for constructing confidence bounds are typically built upon various concentration inequalities, which thus lead to over-exploration. In this paper, we propose a non-parametric and data-dependent UCB algorithm based on the multiplier bootstrap. To improve its finite sample performance, we further incorporate second-order correction into the above construction. In theory, we derive both problem-dependent and problem-independent regret bounds for multi-armed bandits under a much weaker tail assumption than the standard sub-Gaussianity. Numerical results demonstrate significant regret reductions by our method, in comparison with several baselines in a range of multi-armed and linear bandit problems."}, "keywords": ["non-asymptotic inference"], "citation_intent": "background"}
{"citing_id": "2303.02861v1", "cited_id": "1902.00751", "section_title": "Introduction", "citation": "There thus has been a growing interest in developing parameter-efficient methods for model tuning #REFR , where the goal is to learn only a small number of additional parameters per task while achieving performance comparable to full finetuning.", "text_before_citation": ["Finetuning pretrained language models (PLMs) has led to significant improvements across various downstream NLP tasks #OTHEREFR .", "However, the conventional paradigm of full task-specific finetuning (FT) is difficult to scale to multiple tasks, given that modern PLMs can have hundreds of millions (or even billions) of parameters."], "text_after_citation": ["Prompt tuning (PT), which prepends tunable continuous prompt vectors to the input, has emerged as a promising approach for parameter-efficient transfer learning with PLMs #OTHEREFR .", "PT freezes the PLM parameters and only learns a small set of task-specific prompt vectors.", "However, despite their impressive performance, there is still a large gap between prompt tuning and full finetuning .", "Additionally, this approach is sensitive to initialization and often requires more training time than finetuning #OTHEREFR .", "Recent work has proposed to address these issues by transferring prompt vectors from various tasks #OTHEREFR ."], "citing_paper_content": {"title": "Multitask Prompt Tuning Enables Parameter-Efficient Transfer Learning", "abstract": "Prompt tuning, in which a base pretrained model is adapted to each task via conditioning on learned prompt vectors, has emerged as a promising approach for efficiently adapting large language models to multiple downstream tasks. However, existing methods typically learn soft prompt vectors from scratch, and it has not been clear how to exploit the rich cross-task knowledge with prompt vectors in a multitask learning setting. We propose multitask prompt tuning (MPT), which first learns a single transferable prompt by distilling knowledge from multiple task-specific source prompts. We then learn multiplicative low rank updates to this shared prompt to efficiently adapt it to each downstream target task. Extensive experiments on 23 NLP datasets demonstrate that our proposed approach outperforms the state-of-the-art methods, including the full finetuning baseline in some cases, despite only tuning 0.035% as many task-specific parameters. 1 * Work done during an internship at MIT-IBM Watson AI Lab."}, "cited_paper_content": {"title": "Parameter-Efficient Transfer Learning For Nlp", "abstract": "Fine-tuning large pre-trained models is an effective transfer mechanism in NLP. However, in the presence of many downstream tasks, fine-tuning is parameter inefficient: an entire new model is required for every task. As an alternative, we propose transfer with adapter modules. Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones. The parameters of the original network remain fixed, yielding a high degree of parameter sharing. To demonstrate adapter's effectiveness, we transfer the recently proposed BERT Transformer model to 26 diverse text classification tasks, including the GLUE benchmark. Adapters attain near state-of-the-art performance, whilst adding only a few parameters per task. On GLUE, we attain within 0.4% of the performance of full fine-tuning, adding only 3.6% parameters per task. By contrast, fine-tuning trains 100% of the parameters per task."}, "keywords": ["model tuning"], "citation_intent": "background"}
{"citing_id": "2303.16880v1", "cited_id": "1909.11500", "section_title": "", "citation": "Our model is also deeply related to the socalled hidden-manifold model #REFR , which has been used as an analytically solvable model of feedforward neural networks fitting datapoints that live on a low dimensional sub-manifold of their embedding space.", "text_before_citation": ["Most theoretical studies of (generalized) HMs assume simple distributions for the patterns [2, 3] , while in prac- * Corresponding author; matteo.negri@uniroma1.it tical applications the patterns are linearly or non-linearly encoded from and decoded to a different space [16] .", "In this work, we addressed this limitation by proposing a generative model for the patterns where each pattern is produced by the linear combinations of a fixed vocabulary of what we call factors weighted by pattern specific coefficients, followed by an elementwise non-linearity.", "We analyze the model in the high-dimensional regime using the replica method for the statistical physics of disordered systems.", "This data-generating process generalizes the structure of linear superposition proposed in [17] , where it was discussed in relation to the mapping between a Hopfield network and a restricted Boltzmann machine.", "A similar linear (but dense) mapping has been discussed in [18, #OTHEREFR ."], "text_after_citation": ["In fact, this lowdimensional latent structure is typical of many real-world datasets, e.g. the ones made of natural images.", "Here we do not modify the Hebb rule, as we will see that it is enough to produce a new behaviour of the model, in conjunction with the structure of correlation that we choose.", "In fact, we observe that if the correlations in the data are strong enough the model switches from a storage phase to a learning phase, in the sense that attractors appear corresponding to the factors in the data.", "We argue that this behaviour opens up a new paradigm for this model and shows that it may have some phenomenology in common with neural networks.", "Model definition."], "citing_paper_content": {"title": "The Hidden-Manifold Hopfield Model And A Learning Phase Transition", "abstract": "The Hopfield model has a long-standing tradition in statistical physics, being one of the few neural networks for which a theory is available. Extending the theory of Hopfield models for correlated data could help understand the success of deep neural networks, for instance describing how they extract features from data. Motivated by this, we propose and investigate a generalized Hopfield model that we name Hidden-Manifold Hopfield Model : we generate the couplings from P = \u03b1N examples with the Hebb rule using a non-linear transformation of D = \u03b1DN random vectors that we call factors, with N the number of neurons. Using the replica method, we obtain a phase diagram for the model that shows a phase transition where the factors hidden in the examples become attractors of the dynamics; this phase exists above a critical value of \u03b1 and below a critical value of \u03b1D. We call this behaviour learning transition."}, "cited_paper_content": {"title": "Modelling The Influence Of Data Structure On Learning In Neural Networks: The Hidden Manifold Model", "abstract": "The lack of crisp mathematical models that capture the structure of real-world data sets is a major obstacle to the detailed theoretical understanding of deep neural networks. Here, we introduce a generative model for data sets that we call the hidden manifold model (HMM). The idea is to have high-dimensional inputs lie on a lower-dimensional manifold, with labels that depend only on their position within this manifold, akin to a single layer decoder or generator in a generative adversarial network. We first demonstrate the effect of structured data sets by experimentally comparing the dynamics and the performance of two-layer neural networks trained on three different data sets: (i) an unstructured synthetic data set containing random i.i.d. inputs, (ii) a structured data set drawn from the HMM and (iii) a simple canonical data set containing MNIST images. We pinpoint two phenomena related to the dynamics of the networks and their ability to generalise that only appear when training on structured data sets, and we experimentally demonstrate that training networks on data sets drawn from the HMM reproduces both the phenomena seen during training on real dataset. Our main theoretical result is that we show that the learning dynamics in the hidden manifold model is amenable to an analytical treatment by proving a \"Gaussian Equivalence Theorem\", opening the way to further detailed theoretical studies. In particular, we show how the dynamics of stochastic gradient descent for a two-layer network is captured by a set of ordinary differential equations that track the generalisation error at all times."}, "keywords": ["socalled hidden-manifold model", "feedforward neural networks"], "citation_intent": "method"}
{"citing_id": "2303.03540v1", "cited_id": "2004.02843", "section_title": "I. Introduction", "citation": "CODEGNN #REFR is a graph neural network (GNN) that utilizes both sequential and AST representation of source code.", "text_before_citation": ["It is possible that this status quo is caused by the lack of information about the performance of optimizers in ML4SE tasks, as it is unclear how to infer the optimizer effectiveness for the ML4SE domain.", "This motivated us to study the efficiency of various optimizers on various ML4SE problems.", "In our work, we consider four models.", "CODE2SEQ #OTHEREFR and TREELSTM #OTHEREFR are recurrent neural networks (RNNs) that use Abstract Syntax Trees (AST) together with the source code itself.", "Both models are loosely based on the LSTM model, but CODE2SEQ derives a path-based representation of code from the AST, while TREELSTM uses the AST information as is."], "text_after_citation": ["Finally, CODETRANSFORMER #OTHEREFR is an augmented transformer model for code, which combines information from multiple relations between tokens to calculate attention.", "To test these models, we use two code-to-text generation problems as benchmarks: documentation generation and method name generation.", "We chose these two ML4SE related problems to check the consistency of the optimizer performance.", "Moreover, these two problems are often used as the benchmarking problems for ML4SE models #OTHEREFR , #OTHEREFR , #OTHEREFR - #OTHEREFR .", "For the documentation generation problem, we use the Python and Java parts of the CodeXGLUE dataset #OTHEREFR (JAVA-CODEXGLUE, PYTHON-CODEXGLUE), thus checking whether the programming language of the processed source code affects the optimizer performance."], "citing_paper_content": {"title": "Judging Adam: Studying The Performance Of Optimization Methods On Ml4Se Tasks", "abstract": "Solving a problem with a deep learning model requires researchers to optimize the loss function with a certain optimization method. The research community has developed more than a hundred different optimizers, yet there is scarce data on optimizer performance in various tasks. In particular, none of the benchmarks test the performance of optimizers on source code-related problems. However, existing benchmark data indicates that certain optimizers may be more efficient for particular domains. In this work, we test the performance of various optimizers on deep learning models for source code and find that the choice of an optimizer can have a significant impact on the model quality, with up to twofold score differences between some of the relatively well-performing optimizers. We also find that RAdam optimizer (and its modification with the Lookahead envelope) is the best optimizer that almost always performs well on the tasks we consider. Our findings show a need for a more extensive study of the optimizers in code-related tasks, and indicate that the ML4SE community should consider using RAdam instead of Adam as the default optimizer for coderelated deep learning tasks."}, "cited_paper_content": {"title": "Improved Code Summarization Via A Graph Neural Network", "abstract": "Automatic source code summarization is the task of generating natural language descriptions for source code. Automatic code summarization is a rapidly expanding research area, especially as the community has taken greater advantage of advances in neural network and AI technologies. In general, source code summarization techniques use the source code as input and outputs a natural language description. Yet a strong consensus is developing that using structural information as input leads to improved performance. The first approaches to use structural information flattened the AST into a sequence. Recently, more complex approaches based on random AST paths or graph neural networks have improved on the models using flattened ASTs. However, the literature still does not describe the using a graph neural network together with source code sequence as separate inputs to a model. Therefore, in this paper, we present an approach that uses a graph-based neural architecture that better matches the default structure of the AST to generate these summaries. We evaluate our technique using a data set of 2.1 million Java method-comment pairs and show improvement over four baseline techniques, two from the software engineering literature, and two from machine learning literature."}, "keywords": ["CODEGNN", "graph neural network"], "citation_intent": "background"}
{"citing_id": "2303.16502v1", "cited_id": "2002.12410", "section_title": "Unified Analysis", "citation": "Let {x k } k\u22650 be the iterates produced by SGD (Algorithm in (6)), where stochastic gradients are unbiased (i.e., satisfy #REFR ).", "text_before_citation": ["The main building block of the unified analysis from #OTHEREFR is the following parametric assumption on the gradient estimate g k and the problem itself. Assumption 1 (Assumption 4.1 from #OTHEREFR )."], "text_after_citation": ["Assume that there exist non-negative constants A, B, C, D 1 , D 2 \u2265 0, \u03c1 \u2208 (0, 1] and a (possibly) random non-negative sequence {\u03c3 #OTHEREFR k } k\u22650 such that the following two relations hold", "EQUATION", "E k \u03c3 2 k+1 \u2264 (1 \u2212 \u03c1)\u03c3 2 k + 2C f (x k ) \u2212 f (x * ) + D 2 . (9)", "The above assumption is motivated by the analysis of different SGD-type methods and can be derived for standard setups.", "The simplest example is Gradient Descent (GD) applied to the minimization problem of L-smooth function f ."], "citing_paper_content": {"title": "Unified Analysis Of Sgd-Type Methods", "abstract": "This note focuses on a simple approach to the unified analysis of SGD-type methods from [21] for strongly convex smooth optimization problems. The similarities in the analyses of different stochastic first-order methods are discussed along with the existing extensions of the framework. The limitations of the analysis and several alternative approaches are mentioned as well."}, "cited_paper_content": {"title": "On Biased Compression For Distributed Learning", "abstract": "In the last few years, various communication compression techniques have emerged as an indispensable tool helping to alleviate the communication bottleneck in distributed learning. However, despite the fact {\\em biased} compressors often show superior performance in practice when compared to the much more studied and understood {\\em unbiased} compressors, very little is known about them. In this work we study three classes of biased compression operators, two of which are new, and their performance when applied to (stochastic) gradient descent and distributed (stochastic) gradient descent. We show for the first time that biased compressors can lead to linear convergence rates both in the single node and distributed settings. Our {\\em distributed} SGD method enjoys the ergodic rate $\\mathcal{O}\\left(\\frac{\\delta L \\exp(-K) }{\\mu} + \\frac{(C + D)}{K\\mu}\\right)$, where $\\delta$ is a compression parameter which grows when more compression is applied, $L$ and $\\mu$ are the smoothness and strong convexity constants, $C$ captures stochastic gradient noise ($C=0$ if full gradients are computed on each node) and $D$ captures the variance of the gradients at the optimum ($D=0$ for over-parameterized models). Further, via a theoretical study of several synthetic and empirical distributions of communicated gradients, we shed light on why and by how much biased compressors outperform their unbiased variants. Finally, we propose a new highly performing biased compressor---combination of Top-$k$ and natural dithering---which in our experiments outperforms all other compression techniques."}, "keywords": ["stochastic gradients"], "citation_intent": "background"}
{"citing_id": "2303.12408v2", "cited_id": "2003.08934", "section_title": "Optimization", "citation": "The images of EgoNeRF are synthesized by applying the volume rendering equation along the camera ray #REFR and the optional environment map.", "text_before_citation": [], "text_after_citation": ["Specifically, the points x i = o + t i d along the camera ray from camera position o and ray direction d are accumulated to find the pixel value b\u0177", "C = N i=1 \u03c4 i (1 \u2212 e \u2212\u03c3(xi)\u03b4i )c(x i , d) + \u03c4 N +1 c env (d). (11) N = N c + N f", "is the number of samples as described in Sec. 4.1.", "\u03c3(x) and c(x, d) are obtained from our balanced feature grids in Eq. (5).", "Since the size of our feature grid is exponentially increasing along the r direction, we distribute N c coarse samples exponentially rather than uniformly. The second term in Eq. 11is fetched from the environment map"], "citing_paper_content": {"title": "Balanced Spherical Grid For Egocentric View Synthesis", "abstract": "Figure 1. We propose a practical solution to reconstruct large-scale scenes from a short egocentric video. (a) Our scalable capturing setup observes the holistic environment by casually swiping a selfie stick with an omnidirectional camera attached. (b) Then we optimize our balanced spherical feature grids which are tailored for the outward-looking setup."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["camera ray", "volume"], "citation_intent": "method"}
{"citing_id": "2305.00817v1", "cited_id": "1911.10692", "section_title": "Loss Function Based Mitigation.", "citation": "In support of this review of prior work on racial bias mitigation a summary table of related work is provided to compare overall relative performance on the RFW dataset #REFR (Table 3 ).", "text_before_citation": ["L = L + \u00d7 L + \u00d7 L \u210e L = , ( , ) 2 , 0 < ( , ) < (7)", "where is the upper bound of the penalty interval, and is the number of feature representation vectors pairs whose cosine similarity lies within the interval (0, ). domains.", "In another example in face recognition, #OTHEREFR introduces Cross-Domain Triplet (CDT) loss based on the triplet loss #OTHEREFR and uses similarity metrics from one domain to learn compact feature clusters of identities by incorporating them into another domain.", "Relative performance for both CDT and MFR on the RFW dataset are shown in Table 3 .", "This section presents a brief overview of face representation learning, including the potential sources of biases and mitigation studies within this stage of the face recognition processing pipeline (Fig. 3) ."], "text_after_citation": [], "citing_paper_content": {"title": "Racial Bias Within Face Recognition: A Survey", "abstract": "Facial recognition is one of the most academically studied and industrially developed areas within computer vision where we readily find associated applications deployed globally. This widespread adoption has uncovered significant performance variation across subjects of different racial profiles leading to focused research attention on racial bias within face recognition spanning both current causation and future potential solutions. In support, this study provides an extensive taxonomic review of research on racial bias within face recognition exploring every aspect and stage of the face recognition processing pipeline. Firstly, we discuss the problem definition of racial bias, starting with race definition, grouping strategies, and the societal implications of using race or race-related groupings. Secondly, we divide the common face recognition processing pipeline into four stages: image acquisition, face localisation, face representation, face verification and identification, and review the relevant corresponding literature associated with each stage. The overall aim is to provide comprehensive coverage of the racial bias problem with respect to each and every stage of the face recognition processing pipeline whilst also highlighting the potential pitfalls and limitations of contemporary mitigation strategies that need to be considered within future research endeavours or commercial applications alike."}, "cited_paper_content": {"title": "Mitigate Bias In Face Recognition Using Skewness-Aware Reinforcement Learning", "abstract": "Racial equality is an important theme of international human rights law, but it has been largely obscured when the overall face recognition accuracy is pursued blindly. More facts indicate racial bias indeed degrades the fairness of recognition system and the error rates on non-Caucasians are usually much higher than Caucasians. To encourage fairness, we introduce the idea of adaptive margin to learn balanced performance for different races based on large margin losses. A reinforcement learning based race balance network (RL-RBN) is proposed. We formulate the process of finding the optimal margins for non-Caucasians as a Markov decision process and employ deep Q-learning to learn policies for an agent to select appropriate margin by approximating the Q-value function. Guided by the agent, the skewness of feature scatter between races can be reduced. Besides, we provide two ethnicity aware training datasets, called BUPT-Globalface and BUPT-Balancedface dataset, which can be utilized to study racial bias from both data and algorithm aspects. Extensive experiments on RFW database show that RL-RBN successfully mitigates racial bias and learns more balanced performance for different races."}, "keywords": ["racial bias"], "citation_intent": "background"}
{"citing_id": "2305.02093v1", "cited_id": "1402.5886", "section_title": "Experimental Results", "citation": "UFODT performs even better than VFDT and EFDT on LED, Heart and Fico datasets. 2018] as in #REFR .", "text_before_citation": ["We compute the average number of queries per online session to compare the costs of algorithms.", "We also evaluate the generalization power of classifiers via holdout test sets.", "Additionally, (in the appendix F) we measure the prediction performance on training sets during learning together with various other aspects of our framework. Datasets.", "We have used three stationary datasets in our experiments that are standard binary classification datasets taken from UCI repository #OTHEREFR .", "Furthermore, we conduct experiments on the ProPublica recidivism (Compas) dataset #OTHEREFR and the Fair Isaac (Fico) credit risk dataset [FICO et al., Figure 2 : Test utilities during training: UFODT reaches test utilities comparable with those from VFDT and EFDT but with significantly lower costs."], "text_after_citation": ["In Compas dataset, we predict the individuals arrested after two years of release, and in Fico we predict if an individual will default on a loan.", "For concept drifting experiments, we adopt the non-stationary Stagger dataset #OTHEREFR , where each data has three nominal attributes and the target concept will change abruptly at some point.", "For extensions to continuous features (as well as for feature selection in the appendix), we use Prima Indians Diabetes Dataset [Smith et al., 1988] Algorithms.", "The VFDT algorithm #OTHEREFR ] is used as a classical baseline ODT model.", "We also compare our method with the EFDT algorithm #OTHEREFR ."], "citing_paper_content": {"title": "Efficient Online Decision Tree Learning With Active Feature Acquisition", "abstract": "Constructing decision trees online is a classical machine learning problem. Existing works often assume that features are readily available for each incoming data point. However, in many real world applications, both feature values and the labels are unknown a priori and can only be obtained at a cost. For example, in medical diagnosis, doctors have to choose which tests to perform (i.e., making costly feature queries) on a patient in order to make a diagnosis decision (i.e., predicting labels). We provide a fresh perspective to tackle this practical challenge. Our framework consists of an active planning oracle embedded in an online learning scheme for which we investigate several information acquisition functions. Specifically, we employ a surrogate information acquisition function based on adaptive submodularity to actively query feature values with a minimal cost, while using a posterior sampling scheme to maintain a low regret for online prediction. We demonstrate the efficiency and effectiveness of our framework via extensive experiments on various real-world datasets. Our framework also naturally adapts to the challenging setting of online learning with concept drift and is shown to be competitive with baseline models while being more flexible."}, "cited_paper_content": {"title": "Near Optimal Bayesian Active Learning For Decision Making", "abstract": "How should we gather information to make effective decisions? We address Bayesian active learning and experimental design problems, where we sequentially select tests to reduce uncertaintyaboutasetofhypotheses. Instead ofminimizinguncertaintyperse,weconsidera set of overlappingdecision regions of these hypotheses. Our goal is to drive uncertainty into a single decision region as quickly as possible. We identify necessary and sucient conditionsforcorrectlyidentifyingadecisionregion that contains all hypotheses consistent with observations. We develop a novel Hyperedge Cutting (HEC) algorithm for this problem, and prove that is competitive with the intractable optimal policy. Our ecient implementation of the algorithm relies on computingsubsetsofthecompletehomogeneoussymmetric polynomials. Finally, we demonstrate its eectiveness on two practical applications: approximate comparison-based learning and activelocalizationusingarobotmanipulator."}, "keywords": ["Fico datasets"], "citation_intent": "result"}
{"citing_id": "2304.09595v1", "cited_id": "1912.02292", "section_title": "C.1 Analysis Of The Relationship Between Gnn Size And Error", "citation": "Second, the training error does not approach zero, indicating that even a large model is not over-parameterized #REFR .", "text_before_citation": ["This inequality suggests that smaller model sizes typically lead to lower test errors when training a model from scratch.", "This forms the basis for the effectiveness of our delta tuning method.", "We conducted experiments on six small molecular datasets, varying model sizes in two ways: (1) by varying embedding dimensions while fixing the MLP middle dimension at twice the embedding dimensions, and (2) by fixing embedding dimensions at 300 and varying the middle dimensions.", "The results are presented in Figures 6 and 7 . For all datasets, we made two observations.", "First, the test error initially decreases and then increases as the model size grows."], "text_after_citation": ["These observations suggest that our tasks fall within the scope of the classical regime and reducing the size of the parameter space, unless too small, can help improve generalization ability.", "It is important to note that the phenomenon in question occurs exclusively in the context of training from scratch.", "To leverage the knowledge gained from pre-training, employing a larger model is necessary.", "To enhance the generalization ability of this large model, the full fine-tuning approach can be replaced with delta tuning. As illustrated in Fig.", "3(a) , embedding dimensions between 200 to 300 exhibit the highest performance for full fine-tuning, albeit with inferior test error when training from scratch."], "citing_paper_content": {"title": "Adaptergnn: Efficient Delta Tuning Improves Generalization Ability In Graph Neural Networks", "abstract": "Fine-tuning pre-trained models has recently yielded remarkable performance gains in graph neural networks (GNNs). In addition to pre-training techniques, inspired by the latest work in the natural language fields, more recent work has shifted towards applying effective fine-tuning approaches, such as parameter-efficient tuning (delta tuning). However, given the substantial differences between GNNs and transformer-based models, applying such approaches directly to GNNs proved to be less effective. In this paper, we present a comprehensive comparison of delta tuning techniques for GNNs and propose a novel delta tuning method specifically designed for GNNs, called AdapterGNN. AdapterGNN preserves the knowledge of the large pre-trained model and leverages highly expressive adapters for GNNs, which can adapt to downstream tasks effectively with only a few parameters, while also improving the model's generalization ability on the downstream tasks. Extensive experiments show that AdapterGNN achieves higher evaluation performance (outperforming full fine-tuning by 1.4% and 5.5% in the chemistry and biology domains respectively, with only 5% of its parameters tuned) and lower generalization gaps compared to full fine-tuning. Moreover, we empirically show that a larger GNN model can have a worse generalization ability, which differs from the trend observed in large language models. We have also provided a theoretical justification for delta tuning can improve the generalization ability of GNNs by applying generalization bounds."}, "cited_paper_content": {"title": "Deep Double Descent: Where Bigger Models And More Data Hurt", "abstract": "We show that a variety of modern deep learning tasks exhibit a \"double-descent\" phenomenon where, as we increase model size, performance first gets worse and then gets better. Moreover, we show that double descent occurs not just as a function of model size, but also as a function of the number of training epochs. We unify the above phenomena by defining a new complexity measure we call the effective model complexity and conjecture a generalized double descent with respect to this measure. Furthermore, our notion of model complexity allows us to identify certain regimes where increasing (even quadrupling) the number of train samples actually hurts test performance."}, "keywords": ["large model"], "citation_intent": "background"}
{"citing_id": "2304.06244v1", "cited_id": "2002.04881", "section_title": "Related Works", "citation": "Further, they proposed to learn \"flat manifolds\" with VAEs #REFR , whose decoder essentially preserves the Euclidean distance between points in the latent space and the decoded points in the data space.", "text_before_citation": ["The encoders in commonly used traditional codecs such as H.264 #OTHEREFR and HEVC #OTHEREFR are also equipped with an exhaustive search procedure to select the optimal block partitioning and coding modes for image frame.", "More recently, the idea of iterative and optimization-based encoding is becoming increasingly prominent in nonlinear transform coding #OTHEREFR , as well as computer vision in the form of implicit neural representations #OTHEREFR .", "It is therefore interesting to see whether ideas from vector quantization and implicit neural representations may prove fruitful for further reducing the decoding complexity in NTC.", "Manifold learning A distantly related line of work is in metric learning with deep generative models, where the idea is to learn a latent representation of the data such that distance in the latent space preserves the similarity in the data space. Chen et al.", "#OTHEREFR proposes the use of the Riemannian distance metric induced by a decoding transform of a latent variable to measure similarity in the data space."], "text_after_citation": ["Their method is based on regularizing the Jacobian of the VAE to be constant, essentially resulting in a linear (affine) decoder with similar behavior to what we observed in learned synthesis transforms in Section 3.1."], "citing_paper_content": {"title": "Asymmetrically-Powered Neural Image Compression With Shallow Decoders", "abstract": "Neural image compression methods have seen increasingly strong performance in recent years. However, they suffer orders of magnitude higher computational complexity compared to traditional codecs, which stands in the way of real-world deployment. This paper takes a step forward in closing this gap in decoding complexity by adopting shallow or even linear decoding transforms. To compensate for the resulting drop in compression performance, we exploit the often asymmetrical computation budget between encoding and decoding, by adopting more powerful encoder networks and iterative encoding. We theoretically formalize the intuition behind, and our experimental results establish a new frontier in the trade-off between rate-distortion and decoding complexity for neural image compression. Specifically, we achieve rate-distortion performance competitive with the established mean-scale hyperprior architecture of Minnen et al. (2018), while reducing the overall decoding complexity by 80 %, or over 90 % for the synthesis transform alone. Our code can be found at https: //github.com/mandt-lab/shallow-ntc. 1"}, "cited_paper_content": {"title": "Learning Flat Latent Manifolds With Vaes", "abstract": "Measuring the similarity between data points often requires domain knowledge. This can in parts be compensated by relying on unsupervised methods such as latent-variable models, where similarity/distance is estimated in a more compact latent space. Prevalent is the use of the Euclidean metric, which has the drawback of ignoring information about similarity of data stored in the decoder, as captured by the framework of Riemannian geometry. Alternatives---such as approximating the geodesic---are often computationally inefficient, rendering the methods impractical. We propose an extension to the framework of variational auto-encoders allows learning flat latent manifolds, where the Euclidean metric is a proxy for the similarity between data points. This is achieved by defining the latent space as a Riemannian manifold and by regularising the metric tensor to be a scaled identity matrix. Additionally, we replace the compact prior typically used in variational auto-encoders with a recently presented, more expressive hierarchical one---and formulate the learning problem as a constrained optimisation problem. We evaluate our method on a range of data-sets, including a video-tracking benchmark, where the performance of our unsupervised approach nears that of state-of-the-art supervised approaches, while retaining the computational efficiency of straight-line-based approaches."}, "keywords": ["latent space"], "citation_intent": "background"}
{"citing_id": "2303.05194v1", "cited_id": "2002.05709", "section_title": "Cross-Domain Contrastive Loss", "citation": "Although the CDC loss could in principle be applied to encoder features directly, we first project the dense features to a dedicated 128-dimensional embedding space, as per standard practice #REFR .", "text_before_citation": ["The cross-domain contrastive (CDC) loss is the central component of CMA.", "It incentivizes the encoder to learn features which discriminatively reflect the semantics, but are invariant to the visual condition.", "To this end, spatial target image features represent anchors, which are pulled towards spatially corresponding reference image features-the positives.", "The anchors and positives are assumed to represent similar semantics in distinct visual conditions.", "Simultaneously, the anchors are contrasted to other target image features-the negatives-to prevent mode collapse."], "text_after_citation": ["The projection head PROJ consists of two 1\u00d71 convolutions with a ReLU non-linearity inbetween. As shown in Fig.", "2 , the embeddings of the trainable model PROJ \u03b8 \u2022 ENC \u03b8 serve as anchors, while-as proposed in #OTHEREFR -positives and negatives are obtained by an exponential moving average model PROJ ema \u2022 ENC ema to improve their consistency.", "Furthermore, we use a queue to accumulate negatives #OTHEREFR .", "This enables the use of a large number of negatives during instance discrimination, which encourages the learning of meaningful representations by making the discrimination more challenging.", "Finally, the positives are spatially warped to align them with the anchors, as detailed in Sec. 3.2."], "citing_paper_content": {"title": "Contrastive Model Adaptation For Cross-Condition Robustness In Semantic Segmentation", "abstract": "Standard unsupervised domain adaptation methods adapt models from a source to a target domain using labeled source data and unlabeled target data jointly. In model adaptation, on the other hand, access to the labeled source data is prohibited, i.e., only the source-trained model and unlabeled target data are available. We investigate normal-to-adverse condition model adaptation for semantic segmentation, whereby image-level correspondences are available in the target domain. The target set consists of unlabeled pairs of adverse-and normalcondition street images taken at GPS-matched locations. Our method-CMA-leverages such image pairs to learn condition-invariant features via contrastive learning. In particular, CMA encourages features in the embedding space to be grouped according to their condition-invariant semantic content and not according to the condition under which respective inputs are captured. To obtain accurate cross-domain semantic correspondences, we warp the normal image to the viewpoint of the adverse image and leverage warp-confidence scores to create robust, aggregated features. With this approach, we achieve state-of-the-art semantic segmentation performance for model adaptation on several normal-to-adverse adaptation benchmarks, such as ACDC and Dark Zurich. We also evaluate CMA on a newly procured adverse-condition generalization benchmark and report favorable results compared to standard unsupervised domain adaptation methods, despite the comparative handicap of CMA due to source data inaccessibility. Code is available at https://github.com/brdav/cma."}, "cited_paper_content": {"title": "A Simple Framework For Contrastive Learning Of Visual Representations", "abstract": "This paper presents SimCLR: a simple framework for contrastive learning of visual representations. We simplify recently proposed contrastive self-supervised learning algorithms without requiring specialized architectures or a memory bank. In order to understand what enables the contrastive prediction tasks to learn useful representations, we systematically study the major components of our framework. We show that (1) composition of data augmentations plays a critical role in defining effective predictive tasks, (2) introducing a learnable nonlinear transformation between the representation and the contrastive loss substantially improves the quality of the learned representations, and (3) contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. By combining these findings, we are able to considerably outperform previous methods for self-supervised and semi-supervised learning on ImageNet. A linear classifier trained on self-supervised representations learned by SimCLR achieves 76.5% top-1 accuracy, which is a 7% relative improvement over previous state-of-the-art, matching the performance of a supervised ResNet-50. When fine-tuned on only 1% of the labels, we achieve 85.8% top-5 accuracy, outperforming AlexNet with 100X fewer labels."}, "keywords": ["dense features", "dedicated 128-dimensional embedding"], "citation_intent": "method"}
{"citing_id": "2303.07109v1", "cited_id": "1912.01603", "section_title": "Transformer", "citation": "This is in contrast to related works #REFR Chen et al., 2022) that require the full world model during inference. 2.", "text_before_citation": ["The Transformer-XL architecture #OTHEREFR is much more computationally efficient than vanilla transformers at inference time and introduces relative positional encodings, which remove the dependence on absolute time steps.", "Our contributions: The contributions of this work can be summarized as follows:", "1.", "We present a new autoregressive world model based on the Transformer-XL #OTHEREFR ) architecture and a model-free agent trained in latent imagination.", "Running our policy is computationally efficient, as the transformer is not needed at inference time."], "text_after_citation": ["Our world model is provided with information on how much reward has already been emitted by feeding back predicted rewards into the world model. As shown in our ablation study, this improves performance. 3.", "We rewrite the balanced KL divergence loss of #OTHEREFR", "(2021) to allow us to fine-tune the relative weight of the involved entropy and cross-entropy terms. 4.", "We introduce a new thresholded entropy loss that stabilizes the policy's entropy during training and hereby simplifies the selection of hyperparameters that behave well across different games. 5.", "We propose a new effective sampling procedure for the growing dataset of experience, which balances the training distribution to shift the focus towards the latest experience."], "citing_paper_content": {"title": "Transformer-Based World Models Are Happy With 100K Interactions", "abstract": "Deep neural networks have been successful in many reinforcement learning settings. However, compared to human learners they are overly data hungry. To build a sample-efficient world model, we apply a transformer to real-world episodes in an autoregressive manner: not only the compact latent states and the taken actions but also the experienced or predicted rewards are fed into the transformer, so that it can attend flexibly to all three modalities at different time steps. The transformer allows our world model to access previous states directly, instead of viewing them through a compressed recurrent state. By utilizing the Transformer-XL architecture, it is able to learn long-term dependencies while staying computationally efficient. Our transformer-based world model (TWM) generates meaningful, new experience, which is used to train a policy that outperforms previous model-free and model-based reinforcement learning algorithms on the Atari 100k benchmark."}, "cited_paper_content": {"title": "Dream To Control: Learning Behaviors By Latent Imagination", "abstract": "To select effective actions in complex environments, intelligent agents need to generalize from past experience. World models can represent knowledge about the environment to facilitate such generalization. While learning world models from high-dimensional sensory inputs is becoming feasible through deep learning, there are many potential ways for deriving behaviors from them. We present Dreamer, a reinforcement learning agent that solves long-horizon tasks purely by latent imagination. We efficiently learn behaviors by backpropagating analytic gradients of learned state values through trajectories imagined in the compact state space of a learned world model. On 20 challenging visual control tasks, Dreamer exceeds existing approaches in data-efficiency, computation time, and final performance."}, "keywords": ["full world model"], "citation_intent": "result"}
{"citing_id": "2304.12904v1", "cited_id": "1910.01108", "section_title": "Experiments", "citation": "We use DistilBERT #REFR as the pre-trained language model for the first stage models and for reranking Electra-large and T5-3B.", "text_before_citation": ["Note that this could be extended to other models, but we chose the three that we had most familiarity with.", "In more detail, we use DPR-cls #OTHEREFR as our dense retriever, SPLADE-max #OTHEREFR as our learned sparse retriever and RankT5-3b encoder only #OTHEREFR , monoElectra-large #OTHEREFR as our cross encoding rerankers.", "We train models on either the \"titled\" or the official MS MARCO-passage dataset. Titled passages are generated using \"Title [SEP] Passage\". We use only contrastive training (i.e.", "no distillation) and hard-negatives #OTHEREFR to train the first stage models and the negatives from SPLADE to train the reranker.", "First-stage retrieval training uses 8 queries per batch, and 32 negatives per query, and trains for 3 epochs."], "text_after_citation": ["Statistical significance is computed only when we directly compare the original vs \"title\" datasets (i.e.", "only Tables 1 and 4 and only on the same model with different corpus), with Student's t-test and \u2264 0.05."], "citing_paper_content": {"title": "The Tale Of Two Ms Marco -And Their Unfair Comparisons", "abstract": "The MS MARCO-passage dataset has been the main large-scale dataset open to the IR community and it has fostered successfully the development of novel neural retrieval models over the years. But, it turns out that two different corpora of MS MARCO are used in the literature, the official one and a second one where passages were augmented with titles, mostly due to the introduction of the Tevatron code base. However, the addition of titles actually leaks relevance information, while breaking the original guidelines of the MS MARCO-passage dataset. In this work, we investigate the differences between the two corpora and demonstrate empirically that they make a significant difference when evaluating a new method. In other words, we show that if a paper does not properly report which version is used, reproducing fairly its results is basically impossible. Furthermore, given the current status of reviewing, where monitoring state-of-the-art results is of great importance, having two different versions of a dataset is a large problem. This is why this paper aims to report the importance of this issue so that researchers can be made aware of this problem and appropriately report their results."}, "cited_paper_content": {"title": "Distilbert, A Distilled Version Of Bert: Smaller, Faster, Cheaper And Lighter", "abstract": "As Transfer Learning from large-scale pre-trained models becomes more prevalent in Natural Language Processing (NLP), operating these large models in on-the-edge and/or under constrained computational training or inference budgets remains challenging. In this work, we propose a method to pre-train a smaller general-purpose language representation model, called DistilBERT, which can then be fine-tuned with good performances on a wide range of tasks like its larger counterparts. While most prior work investigated the use of distillation for building task-specific models, we leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster. To leverage the inductive biases learned by larger models during pre-training, we introduce a triple loss combining language modeling, distillation and cosine-distance losses. Our smaller, faster and lighter model is cheaper to pre-train and we demonstrate its capabilities for on-device computations in a proof-of-concept experiment and a comparative on-device study."}, "keywords": ["pre-trained language model"], "citation_intent": "method"}
{"citing_id": "2303.17078v1", "cited_id": "1905.11075", "section_title": "Discussion", "citation": "In the past, many advances have been driven in the field of fluid mechanics #REFR , and this is likely to continue.", "text_before_citation": ["Understanding how these solution operators vary with system parameters is an important avenue of ongoing research #OTHEREFR .", "Similarly, machine learning may be used to accelerate traditional scientific computing workflows, for example by flexible super-resolution or learning of improved solution stencils.", "However, here are several challenges with these approaches, foremost the fact that traditional numerical algorithms are extremely mature and scaleable, so that machine learning solutions are expected to compete with decades of progress.", "In all of the cases explored in this perspective, progress will be accelerated by a diverse and robust set of benchmark problems with which to assess new solutions #OTHEREFR .", "In addition, we must stress that these techniques are primarily tools to be used by human experts for scientific discovery."], "text_after_citation": ["For example, understanding sensitivities with resolvent analysis #OTHEREFR , using physics informed neural networks (PINNs) #OTHEREFR for RANS modeling #OTHEREFR , and using wall measurements to estimate turbulent flow fields #OTHEREFR are all exciting avenues of research.", "Interestingly, there are also efforts to understand neural networks using techniques from PDEs #OTHEREFR .", "Although there is a desire for automated machine learning algorithms, when applied to science and engineering applications, this is still primarily a human endeavor.", "However, progress in the field of PDEs, enabled by machine learning, is undeniable.", "Despite this progress, there is still much we don't know about PDEs."], "citing_paper_content": {"title": "Machine Learning For Partial Differential Equations", "abstract": "Partial differential equations (PDEs) are among the most universal and parsimonious descriptions of natural physical laws, capturing a rich variety of phenomenology and multi-scale physics in a compact and symbolic representation. This review will examine several promising avenues of PDE research that are being advanced by machine learning, including: 1) the discovery of new governing PDEs and coarse-grained approximations for complex natural and engineered systems, 2) learning effective coordinate systems and reduced-order models to make PDEs more amenable to analysis, and 3) representing solution operators and improving traditional numerical algorithms. In each of these fields, we summarize key advances, ongoing challenges, and opportunities for further development."}, "cited_paper_content": {"title": "Machine Learning For Fluid Mechanics", "abstract": "The field of fluid mechanics is rapidly advancing, driven by unprecedented volumes of data from field measurements, experiments and large-scale simulations at multiple spatiotemporal scales. Machine learning offers a wealth of techniques to extract information from data that could be translated into knowledge about the underlying fluid mechanics. Moreover, machine learning algorithms can augment domain knowledge and automate tasks related to flow control and optimization. This article presents an overview of past history, current developments, and emerging opportunities of machine learning for fluid mechanics. It outlines fundamental machine learning methodologies and discusses their uses for understanding, modeling, optimizing, and controlling fluid flows. The strengths and limitations of these methods are addressed from the perspective of scientific inquiry that considers data as an inherent part of modeling, experimentation, and simulation. Machine learning provides a powerful information processing framework that can enrich, and possibly even transform, current lines of fluid mechanics research and industrial applications."}, "keywords": ["fluid mechanics"], "citation_intent": "background"}
{"citing_id": "2304.05894v1", "cited_id": "1803.01616", "section_title": "Sdsbm Works With Little Data.", "citation": "When the number of observations is high, the \"no coupling\" baseline #REFR reaches the performances of SDSBM.", "text_before_citation": ["4b ., we vary the number of observations available for each item from 100 to 10.000, distributed over 100 epochs.", "Thus, in the most challenging situation, there is only one observation per epoch used to determine dynamic memberships over 3 clusters. In the expression of , is set to 0.05. We see Fig. 4b .", "that for both patterns, the predictive power of SDSBM remains high in such conditions.", "Moreover, the RMSE on the true dynamic memberships in this case is fairly low, and decreases rapidly as the number of observations increases.", "This is due to SDSBM linking the time slices together during the training phase: the smoothing constraint makes it so that every time slices indirectly benefits from the training data of its neighbours."], "text_after_citation": ["This is because as the number of observations in each slice goes to infinity, the models need less to rely on temporal neighbours.", "However, even for 10.000 observations per item (100 observations per epoch), SDSBM yields better results. As an illustration, the results in Fig.", "1 have been obtained using only 5 observations per epoch.", "4.1.3 SDSBM handles highly stochastic interaction patterns.", "Finally, we control the deterministic character of the block-interaction matrix by varying ."], "citing_paper_content": {"title": "Dynamic Mixed Membership Stochastic Block Model For Weighted Labeled Networks", "abstract": "Most real-world networks evolve over time. Existing literature proposes models for dynamic networks that are either unlabeled or assumed to have a single membership structure. On the other hand, a new family of Mixed Membership Stochastic Block Models (MMSBM) allows to model static labeled networks under the assumption of mixed-membership clustering. In this work, we propose to extend this later class of models to infer dynamic labeled networks under a mixed membership assumption. Our approach takes the form of a temporal prior on the model's parameters. It relies on the single assumption that dynamics are not abrupt. We show that our method significantly differs from existing approaches, and allows to model more complex systems-dynamic labeled networks. We demonstrate the robustness of our method with several experiments on both synthetic and real-world datasets. A key interest of our approach is that it needs very few training data to yield good results. The performance gain under challenging conditions broadens the variety of possible applications of automated learning tools-as in social sciences, which comprise many fields where small datasets are a major obstacle to the introduction of machine learning methods."}, "cited_paper_content": {"title": "Tensorial And Bipartite Block Models For Link Prediction In Layered Networks And Temporal Networks", "abstract": "Many real-world complex systems are well represented as multilayer networks; predicting interactions in those systems is one of the most pressing problems in predictive network science. To address this challenge, we introduce two stochastic block models for multilayer and temporal networks; one of them uses nodes as its fundamental unit, whereas the other focuses on links. We also develop scalable algorithms for inferring the parameters of these models. Because our models describe all layers simultaneously, our approach takes full advantage of the information contained in the whole network when making predictions about any particular layer. We illustrate the potential of our approach by analyzing two empirical datasets---a temporal network of email communications, and a network of drug interactions for treating different cancer types. We find that modeling all layers simultaneously does result, in general, in more accurate link prediction. However, the most predictive model depends on the dataset under consideration; whereas the node-based model is more appropriate for predicting drug interactions, the link-based model is more appropriate for predicting email communication."}, "keywords": ["SDSBM", "\"no coupling\" baseline"], "citation_intent": "background"}
{"citing_id": "2303.11577v1", "cited_id": "1903.00104", "section_title": "Unsteady Lid-Driven Flow Problem", "citation": "According to #REFR , we work on inferring the newly defined effective reaction rate k f = v A k f,r as well as a r .", "text_before_citation": ["EQUATION", "where C A and C B are the concentration of solutes A and B respectively, q = 0.5 is the Darcy velocity, \u03c8 = 0.4 is the porosity, D = 1 \u00d7 10 \u22128 is the diffusion coefficient, v A = a r and v B = \u22121 are the stoichiometric coefficients.", "The constant k f,r is the chemical reaction rate and a r is the order of the chemical reaction, both of which are difficult to measure and will be inferred by the observations of C A .", "To obtain the observations, numerical simulations are conducted using a secondorder finite difference method for spatial discretization and the second-order backward difference method for implicitly marching in time.", "In simulations, the mesh size is \u2206x = 0.0125 and the time step size is \u2206t = 0.005."], "text_after_citation": ["We set the exact values to k f = 1.577 and a r = 2. The high-fidelity observations at Table 3 .", "It is shown that the inferred results are highly accurate with a small relative error and variance.", "Our results are significantly outperforming the reported results in #OTHEREFR .", "It is worth noting that only 50 residual points are used for training the networks, while 30000 residual points are used in #OTHEREFR .", "In order to take a further look at the prediction details, Fig."], "citing_paper_content": {"title": "Feature-Adjacent Multi-Fidelity Physics-Informed Machine Learning For Partial Differential Equations", "abstract": "Physics-informed neural networks have emerged as an alternative method for solving partial differential equations. However, for complex problems, the training of such networks can still require high-fidelity data which can be expensive to generate. To reduce or even eliminate the dependency on high-fidelity data, we propose a novel multi-fidelity architecture which is based on a feature space shared by the lowand high-fidelity solutions. In the feature space, the projections of the low-fidelity and high-fidelity solutions are adjacent by constraining their relative distance. The feature space is represented with an encoder and its mapping to the original solution space is effected through a decoder. The proposed multi-fidelity approach is validated on forward and inverse problems for steady and unsteady problems described by partial differential equations."}, "cited_paper_content": {"title": "A Composite Neural Network That Learns From Multi-Fidelity Data: Application To Function Approximation And Inverse Pde Problems", "abstract": "Currently the training of neural networks relies on data of comparable accuracy but in real applications only a very small set of high-fidelity data is available while inexpensive lower fidelity data may be plentiful. We propose a new composite neural network (NN) that can be trained based on multi-fidelity data. It is comprised of three NNs, with the first NN trained using the low-fidelity data and coupled to two high-fidelity NNs, one with activation functions and another one without, in order to discover and exploit nonlinear and linear correlations, respectively, between the low-fidelity and the high-fidelity data. We first demonstrate the accuracy of the new multi-fidelity NN for approximating some standard benchmark functions but also a 20-dimensional function that is not easy to approximate with other methods, e.g. Gaussian process regression. Subsequently, we extend the recently developed physics-informed neural networks (PINNs) to be trained with multi-fidelity data sets (MPINNs). MPINNs contain four fully-connected neural networks, where the first one approximates the low-fidelity data, while the second and third construct the correlation between the low- and high-fidelity data and produce the multi-fidelity approximation, which is then used in the last NN that encodes the partial differential equations (PDEs). Specifically, by decomposing the correlation into a linear and nonlinear part, the present model is capable of learning both the linear and complex nonlinear correlations between the low- and high-fidelity data adaptively. By training the MPINNs, we can: (1) obtain the correlation between the low- and high-fidelity data, (2) infer the quantities of interest based on a few scattered data, and (3) identify the unknown parameters in the PDEs. In particular, we employ the MPINNs to learn the hydraulic conductivity field for unsaturated flows as well as the reactive models for reactive transport. The results demonstrate that MPINNs can achieve relatively high accuracy based on a very small set of high-fidelity data. Despite the relatively low dimension and limited number of fidelities (two-fidelity levels) for the benchmark problems in the present study, the proposed model can be readily extended to very high-dimensional regression and classification problems involving multi-fidelity data."}, "keywords": ["reaction rate"], "citation_intent": "background"}
{"citing_id": "2303.15100v2", "cited_id": "2003.00104", "section_title": "Tokenization Analysis -Datasets", "citation": "To further explore the number of word pieces per entity type, we isolate the unique entities #REFR .", "text_before_citation": ["We choose the biomedical (ADE) dataset to explore the effect of tokenization in a special domain.", "The ADE dataset contains entities of Drugs and Adverse Effects (AE) and has labels for the relations between them.", "The tokenizer of cased BERT #OTHEREFR and bioclinical BERT (b-BERT) #OTHEREFR is based on the WordPiece algorithm, while ALBERT #OTHEREFR adopts the SentencePiece algorithm. In Tab.", "2, we present the effect of tokenization on the average sentence length, in terms of word pieces (subwords), for each dataset.", "The sentence length increases by approximately 12 tokens, up to 58%, after the tokenization in the biomedical domain."], "text_after_citation": ["Then, we find the unique words that are part of each entity type and tokenize the unique entities and words using the different tokenizers to notice the difference in the length and the addition of the word pieces. In Tab.", "1 the last column represents the average tokenized word length per entity type, and the Out type describes the words that are not part of an entity of interest.", "In the ADE dataset, the length of the drug and AE entities increases substantially, and the drug entities are split into more word pieces.", "Particularly, a word that is part of a drug entity is split into approximately 4 word pieces, on average, when using the tokenizer of cased BERT and b-BERT.", "The tokenizer of ALBERT tends to split the entities of interest into fewer pieces."], "citing_paper_content": {"title": "An Information Extraction Study: Take In Mind The Tokenization!", "abstract": "Current research on the advantages and trade-offs of using characters, instead of tokenized text, as input for deep learning models, has evolved substantially. New token-free models remove the traditional tokenization step; however, their efficiency remains unclear. Moreover, the effect of tokenization is relatively unexplored in sequence tagging tasks. To this end, we investigate the impact of tokenization when extracting information from documents and present a comparative study and analysis of subword-based and character-based models. Specifically, we study Information Extraction (IE) from biomedical texts. The main outcome is twofold: tokenization patterns can introduce inductive bias that results in state-of-the-art performance, and the character-based models produce promising results; thus, transitioning to token-free IE models is feasible."}, "cited_paper_content": {"title": "Arabert: Transformer-Based Model For Arabic Language Understanding", "abstract": "The Arabic language is a morphologically rich language with relatively few resources and a less explored syntax compared to English. Given these limitations, Arabic Natural Language Processing (NLP) tasks like Sentiment Analysis (SA), Named Entity Recognition (NER), and Question Answering (QA), have proven to be very challenging to tackle. Recently, with the surge of transformers based models, language-specific BERT based models have proven to be very efficient at language understanding, provided they are pre-trained on a very large corpus. Such models were able to set new standards and achieve state-of-the-art results for most NLP tasks. In this paper, we pre-trained BERT specifically for the Arabic language in the pursuit of achieving the same success that BERT did for the English language. The performance of AraBERT is compared to multilingual BERT from Google and other state-of-the-art approaches. The results showed that the newly developed AraBERT achieved state-of-the-art performance on most tested Arabic NLP tasks. The pretrained araBERT models are publicly available on https://github.com/aub-mind/arabert hoping to encourage research and applications for Arabic NLP."}, "keywords": ["word pieces", "entity type"], "citation_intent": "method"}
{"citing_id": "2304.02991v1", "cited_id": "1505.04597", "section_title": "Base 2D/3D Architecture", "citation": "The 2D branch processes images to obtain a pixel-wise prediction given x 2D and it consists of a standard 2D U-Net #REFR .", "text_before_citation": ["We build our contributions upon the two independent branches (2D and 3D) architecture proposed in #OTHEREFR ."], "text_after_citation": ["On the other hand, the 3D branch takes in input point clouds to estimate the class of each point of x 3D and it is implemented as a 3D sparse convolutional network #OTHEREFR .", "Thanks to the fact that 2D-3D correspondences are known, 3D points can be projected into the image plane to supervise the 2D branch, as supervision is provided only for the sparse 3D points.", "We denote the 3D semantic labels projected into 2D with the symbol y 3D\u21922D .", "As argued by #OTHEREFR , such design choice allows one to take advantage of the strengths of each input modality, and final predictions can be obtained by averaging the outputs of the two branches to achieve an effective ensemble.", "In our work, we adopt the same framework, and we give an intuitive explanation of why this design choice is particularly effective."], "citing_paper_content": {"title": "Exploiting The Complementarity Of 2D And 3D Networks To Address Domain-Shift In 3D Semantic Segmentation", "abstract": "2D Receptive Field Re-Projected into 3D Figure 1. 3D (top) and 2D (bottom) networks processing point clouds and images of the same scene extract features that contain complementary information. Indeed, 2D and 3D effective receptive fields [34] centered on a point n focus on different portions of the scene, i.e., 2D or 3D neighborhoods respectively. Thus, corresponding features have different content by construction. We exploit this property to reduce the domain gap in 3D semantic segmentation."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["standard 2D U-Net"], "citation_intent": "method"}
{"citing_id": "2303.12711v1", "cited_id": "1610.05683", "section_title": "Sampling From The Von Mises-Fisher Distribution", "citation": "The gradient for this sampling procedure can be computed using the reparameterization trick for acceptance-rejection sampling schemes as proposed by #REFR and further defined for the vMF distribution by .", "text_before_citation": [".", "The next step is to construct a Householder reflection matrix H, defined as H = I \u2212 2hh T , where H = I is the identity matrix and h = e1\u2212\u00b5 e1\u2212\u00b5 , with modal vector e 1 = (1, 0, \u2022 \u2022 \u2022 , 0).", "Applying this Householder transform to z essentially reflects it across the hyperplane that lies between \u00b5 and e 1 , resulting in z = Hz , a direction vector sampled from the vMF distribution.", "Algorithm 1 vMF Sampling 1: input: dimension m, mean \u00b5, concentration \u03ba 2: Acceptance-rejection sampling:", "\u03c9 \u223c g(\u03c9 | \u03ba, m) \u221d exp(\u03c9\u03ba) 1 \u2212 \u03c9 2 1 2 (m\u22123) 3: Sample v from Uniform distribution: v \u223c U S m\u22122 4: Householder transform: z \u2190 \u03c9; \u221a 1 \u2212 \u03c9 2 v H \u2190 Householder (e 1 , \u00b5) 5: return : Hz"], "text_after_citation": [], "citing_paper_content": {"title": "Geometry-Aware Latent Representation Learning For Modeling Disease Progression Of Barrett'S Esophagus", "abstract": "First and foremost, I would like to thank my supervisors, Erik Bekkers, Onno de Boer and Sybren Meijer. Erik for his valuable supervision and feedback, and for always supplying me with many interesting research ideas. His insights introduced me to many new ideas that have also fueled my enthusiasm for research in general. Furthermore, Onno and Sybren for introducing me to the field of histopathology and offering me an interesting look into the medical world. Working with you all made me see how technology and AI can directly impact clinical practice, and affirmed for me the importance of the medical applications of my work. I would also like to thank Sharvaree Vadgama for our weekly paper meetings at the beginning of this project, which greatly helped me in learning the technical preliminaries necessary to conduct this study. Moreover, I of course want to thank my family for supporting me unconditionally. And finally, I thank my friends from the former master AI room, who were always there struggling with me during this period and who made working on this thesis an almost enjoyable experience."}, "cited_paper_content": {"title": "Reparameterization Gradients Through Acceptance-Rejection Sampling Algorithms", "abstract": "Variational inference using the reparameterization trick has enabled large-scale approximate Bayesian inference in complex probabilistic models, leveraging stochastic optimization to sidestep intractable expectations. The reparameterization trick is applicable when we can simulate a random variable by applying a differentiable deterministic function on an auxiliary random variable whose distribution is fixed. For many distributions of interest (such as the gamma or Dirichlet), simulation of random variables relies on acceptance-rejection sampling. The discontinuity introduced by the accept-reject step means that standard reparameterization tricks are not applicable. We propose a new method that lets us leverage reparameterization gradients even when variables are outputs of a acceptance-rejection sampling algorithm. Our approach enables reparameterization on a larger class of variational distributions. In several studies of real and synthetic data, we show that the variance of the estimator of the gradient is significantly lower than other state-of-the-art methods. This leads to faster convergence of stochastic gradient variational inference."}, "keywords": ["reparameterization trick", "acceptance-rejection sampling schemes"], "citation_intent": "method"}
{"citing_id": "2303.03382v1", "cited_id": "1906.02107", "section_title": "Why Should We Care About Threshold Networks?", "citation": "Specifically, these networks have significantly lower memory footprint, less computational complexity, and consume less energy #REFR .", "text_before_citation": ["Neural networks with threshold activations are highly desirable due to the following reasons:", "\u2022 Since the threshold activation (1) is restricted to take values in {0, s}, threshold neural network models are far more suitable for hardware implementations #OTHEREFR ."], "text_after_citation": ["\u2022 Modern neural networks have extremely large number of full precision trainable parameters so that several computational barriers emerge during hardware implementations. One approach to 7with the non-convex training heuristic STE.", "We also indicate the time taken to solve the convex programs with markers.", "For non-convex STE, we repeat the training with 5 different initializations.", "In each case, our convex training algorithms achieve lower objective than all the non-convex heuristics (see Appendix B.5 for details).", "mitigate these issues is reducing the network size by grouping the parameters via a hash function #OTHEREFR ."], "citing_paper_content": {"title": "Globally Optimal Training Of Neural Net-Works With Threshold Activation Functions", "abstract": "Threshold activation functions are highly preferable in neural networks due to their efficiency in hardware implementations. Moreover, their mode of operation is more interpretable and resembles that of biological neurons. However, traditional gradient based algorithms such as Gradient Descent cannot be used to train the parameters of neural networks with threshold activations since the activation function has zero gradient except at a single non-differentiable point. To this end, we study weight decay regularized training problems of deep neural networks with threshold activations. We first show that regularized deep threshold network training problems can be equivalently formulated as a standard convex optimization problem, which parallels the LASSO method, provided that the last hidden layer width exceeds a certain threshold. We also derive a simplified convex optimization formulation when the dataset can be shattered at a certain layer of the network. We corroborate our theoretical results with various numerical experiments."}, "cited_paper_content": {"title": "Latent Weights Do Not Exist: Rethinking Binarized Neural Network Optimization", "abstract": "Optimization of Binarized Neural Networks (BNNs) currently relies on real-valued latent weights to accumulate small update steps. In this paper, we argue that these latent weights cannot be treated analogously to weights in real-valued networks. Instead their main role is to provide inertia during training. We interpret current methods in terms of inertia and provide novel insights into the optimization of BNNs. We subsequently introduce the first optimizer specifically designed for BNNs, Binary Optimizer (Bop), and demonstrate its performance on CIFAR-10 and ImageNet. Together, the redefinition of latent weights as inertia and the introduction of Bop enable a better understanding of BNN optimization and open up the way for further improvements in training methodologies for BNNs."}, "keywords": ["networks", "less computational complexity"], "citation_intent": "background"}
{"citing_id": "2303.05496v1", "cited_id": "1703.06103", "section_title": "Real-World Knowledge Graph Reasoning", "citation": "We also include Hit@10 as an additional metric following knowledge graph embedding literature and GraIL #REFR .", "text_before_citation": ["SpaLoc also has comparable performance with KE methods on the FB15K-237 datasets, outperforming GraIL by a large margin.", "Comparing SpaLoc with node embedding-based methods (TransE) and GNN-based methods (GraIL) that only consider binary edges, we see that our hyperedge-based model enables better relation prediction that requires reasoning about other entities.", "The necessity of hyperedges is further supported by Appendix E, where we show that setting the maximum arity of SpaLoc to 2 (i.e., removing hyperedges) significantly degrades the performance.", "Notably, in contrast to other methods for the transductive setting that store entity embeddings for all knowledge graph nodes, SpaLoc directly uses the inductive learning setting.", "That is, SpaLoc does Table 5 : Results of transductive link prediction on real-world knowledge graphs."], "text_after_citation": ["The scores and standard errors (appendix Table 8 not store entity information about each knowledge graph node.", "SpaLoc performs pretty well in the classification setting and ranking setting when negative candidate sets are small (Table 5 ).", "However, we find inductive reasoning methods, including SpaLoc, GraIL #OTHEREFR , and TACT #OTHEREFR , perform poorly in the traditional knowledge graph ranking setting #OTHEREFR (i.e., ranking the positive triplets against negative triplets of all replacement of head/tail) entities.", "The ranking scores are close to zero when the negative candidates set are large.", "We attribute this phenomenon to the fact that the inductive models do not store entity information."], "citing_paper_content": {"title": "Sparse And Local Networks For Hypergraph Reasoning", "abstract": "Reasoning about the relationships between entities from input facts (e.g., whether Ari is a grandparent of Charlie) generally requires explicit consideration of other entities that are not mentioned in the query (e.g., the parents of Charlie). In this paper, we present an approach for learning to solve problems of this kind in large, real-world domains, using sparse and local hypergraph neural networks (SpaLoc). SpaLoc is motivated by two observations from traditional logic-based reasoning: relational inferences usually apply locally (i.e., involve only a small number of individuals), and relations are usually sparse (i.e., only hold for a small percentage of tuples in a domain). We exploit these properties to make learning and inference efficient in very large domains by (1) using a sparse tensor representation for hypergraph neural networks, (2) applying a sparsification loss during training to encourage sparse representations, and (3) subsampling based on a novel information sufficiency-based sampling process during training. SpaLoc achieves state-of-the-art performance on several real-world, large-scale knowledge graph reasoning benchmarks, and is the first framework for applying hypergraph neural networks on real-world knowledge graphs with more than 10k nodes."}, "cited_paper_content": {"title": "Modeling Relational Data With Graph Convolutional Networks", "abstract": "Knowledge bases play a crucial role in many applications, for example question answering and information retrieval. Despite the great effort invested in creating and maintaining them, even the largest representatives (e.g., Yago, DBPedia or Wikidata) are highly incomplete. We introduce relational graph convolutional networks (R-GCNs) and apply them to two standard knowledge base completion tasks: link prediction (recovery of missing facts, i.e.~subject-predicate-object triples) and entity classification (recovery of missing attributes of entities). R-GCNs are a generalization of graph convolutional networks, a recent class of neural networks operating on graphs, and are developed specifically to deal with highly multi-relational data, characteristic of realistic knowledge bases. Our methods achieve competitive performance on standard benchmarks for both tasks, demonstrating especially promising results on the challenging FB15k-237 subset of Freebase."}, "keywords": ["knowledge graph"], "citation_intent": "method"}
{"citing_id": "2303.10766v1", "cited_id": "1504.00325", "section_title": "Introduction", "citation": "The experimental results on the standard MSCOCO image captioning dataset #REFR show an improved or comparable performance comparing with several state-of-the-art methods in fair conditions.", "text_before_citation": ["This allows the captioning system to exploit both spatial and semantic features simultaneously, and makes end-to-end training of the system possible.", "Finally, a multi-modal reward function is proposed for DRL in a two-phase supervised-reinforced training process of the captioning system.", "This reward function consists of CIDEr, as the language reward, and an image-caption similarity as vision reward.", "The language reward measures the similarity between the generated and ground truth captions.", "The vision reward measures the similarity between the generated caption and the image contents, obtained from an embedding network for mapping both caption and image into a common latent space."], "text_after_citation": ["Detailed experimental analysis reveal the effectiveness of the visual relationships information on improving the performance of the captioning system based on common evaluation metrics.", "Moreover, the results indicate that using the proposed multi-modal reward for DRL with SCST algorithm outperforms the original version of this algorithm.", "In summary, the main contributions of this study are as follows:", "\u2022 Scene graph-based visual relationship triplets are introduced as high-level semantic features to improve the image captioning results.", "\u2022 An AoA-based deep neural architecture is proposed for fusion of visual relationships and spatial features extracted from images in the decoder of the captioning system."], "citing_paper_content": {"title": "Multi-Modal Reward For Visual Relationships-Based Image Captioning", "abstract": "Deep neural networks have achieved promising results in automatic image captioning due to their effective representation learning and context-based content generation capabilities. As a prominent type of deep features used in many of the recent image captioning methods, the well-known bottomup features provide a detailed representation of different objects of the image in comparison with the feature maps directly extracted from the raw image. However, the lack of high-level semantic information about the relationships between these objects is an important drawback of bottom-up features, despite their expensive and resource-demanding extraction procedure. To take advantage of visual relationships in caption generation, this paper proposes a deep neural network architecture for image captioning based on fusing the visual relationships information extracted from an image's scene graph with the spatial feature maps of the image. A multi-modal reward function is then introduced for deep reinforcement learning of the proposed network using a combination of language and vision similarities in a common embedding space. The results of extensive experimentation on the MSCOCO dataset show the effectiveness of using visual relationships in the proposed captioning method. Moreover, the results clearly indicate that the proposed multi-modal reward in deep reinforcement learning leads to better model optimization, outperforming several state-of-the-art image captioning algorithms, while using light and easy to extract image features. A detailed experimental study of the components constituting the proposed method is also presented."}, "cited_paper_content": {"title": "Microsoft Coco Captions: Data Collection And Evaluation Server", "abstract": "In this paper we describe the Microsoft COCO Caption dataset and evaluation server. When completed, the dataset will contain over one and a half million captions describing over 330,000 images. For the training and validation images, five independent human generated captions will be provided. To ensure consistency in evaluation of automatic caption generation algorithms, an evaluation server is used. The evaluation server receives candidate captions and scores them using several popular metrics, including BLEU, METEOR, ROUGE and CIDEr. Instructions for using the evaluation server are provided."}, "keywords": ["captioning dataset"], "citation_intent": "result"}
{"citing_id": "2303.17597v2", "cited_id": "1807.01697", "section_title": "Evaluation Metrics", "citation": "We follow #REFR and use the mean CE (mCE) as the primary metric in comparing models' robustness.", "text_before_citation": ["Corruption Error (CE)."], "text_after_citation": ["To normalize the severity effects, we choose Cen-terPoint #OTHEREFR and MinkUNet #OTHEREFR as the baseline models for the 3D detectors and segmentors, respectively. The CE and mCE scores are calculated as follows:", "EQUATION", "where Acc i,l denotes task-specific accuracy scores, i.e., mIoU, AP, NDS, or APH(L2), on corruption type i at severity level l.", "N = 8 is the total number of corruption types. Resilience Rate (RR).", "We define mean RR (mRR) as the relative robustness indicator for measuring how much accuracy can a model retain when evaluated on the corruption sets. The RR and mRR scores are calculated as follows."], "citing_paper_content": {"title": "Robo3D: Towards Robust And Reliable 3D Perception Against Corruptions", "abstract": "Figure 1: Taxonomy of the Robo3D benchmark. We simulate eight corruption types from three categories: 1) Severe weather conditions, such as fog, rain, and snow; 2) External disturbances that are caused by motion blur or result in the missing of LiDAR beams; and 3) Internal sensor failure, including LiDAR crosstalk, possible incomplete echo, and cross-sensor scenarios. Each corruption is further split into three levels (light, moderate, and heavy) based on its severity."}, "cited_paper_content": {"title": "Benchmarking Neural Network Robustness To Common Corruptions And Surface Variations", "abstract": "In this paper we establish rigorous benchmarks for image classifier robustness. Our first benchmark, ImageNet-C, standardizes and expands the corruption robustness topic, while showing which classifiers are preferable in safety-critical applications. Then we propose a new dataset called ImageNet-P which enables researchers to benchmark a classifier's robustness to common perturbations. Unlike recent robustness research, this benchmark evaluates performance on common corruptions and perturbations not worst-case adversarial perturbations. We find that there are negligible changes in relative corruption robustness from AlexNet classifiers to ResNet classifiers. Afterward we discover ways to enhance corruption and perturbation robustness. We even find that a bypassed adversarial defense provides substantial common perturbation robustness. Together our benchmarks may aid future work toward networks that robustly generalize."}, "keywords": ["models' robustness"], "citation_intent": "method"}
{"citing_id": "2303.12818v1", "cited_id": "1809.00846", "section_title": "Regularization And Batchnorm", "citation": "The algorithm in #REFR produces accuracies that are comparable to BatchNorm, but unlike the case with BatchNorm, they found that introducing dropout layers improved performance further.", "text_before_citation": ["In the 2018 paper #OTHEREFR , population statistics were used instead of batch statistics, and a regularization term was included for the parameter.", "In these experiments, it was noted that for batches of size larger than 32, the population statistics function as well as the batch statistics.", "It was also noted that BatchNorm introduces Gaussian noise into the mean and variance parameters."], "text_after_citation": ["They also state that BatchNorm has very similar effects to norm regularization where the -norm is defined as", "\u2016 \u2016 = (\ufe01 \u2211\ufe01 =1 | | )\ufe01 1/ .", "If norm regularization was sufficient, there would be no need for an optimization such as BatchNorm, since regularization is less computationally intensive.", "However, this regularization claim is not consistent with other empirical studies, such as that presented in the 2018 paper #OTHEREFR ."], "citing_paper_content": {"title": "An Empirical Analysis Of The Shift And Scale Parameters In Batchnorm", "abstract": "Batch Normalization (BatchNorm) is a technique that improves the training of deep neural networks, especially Convolutional Neural Networks (CNN). It has been empirically demonstrated that BatchNorm increases performance, stability, and accuracy, although the reasons for such improvements are unclear. BatchNorm includes a normalization step as well as trainable shift and scale parameters. In this paper, we empirically examine the relative contribution to the success of BatchNorm of the normalization step, as compared to the re-parameterization via shifting and scaling. To conduct our experiments, we implement two new optimizers in PyTorch, namely, a version of BatchNorm that we refer to as AffineLayer, which includes the re-parameterization step without normalization, and a version with just the normalization step, that we call BatchNorm-minus. We compare the performance of our AffineLayer and BatchNorm-minus implementations to standard BatchNorm, and we also compare these to the case where no batch normalization is used. We experiment with four ResNet architectures (ResNet18, ResNet34, ResNet50, and ResNet101) over a standard image dataset and multiple batch sizes. Among other findings, we provide empirical evidence that the success of BatchNorm may derive primarily from improved weight initialization."}, "cited_paper_content": {"title": "Towards Understanding Regularization In Batch Normalization", "abstract": "Batch Normalization (BN) improves both convergence and generalization in training neural networks. This work understands these phenomena theoretically. We analyze BN by using a basic block of neural networks, consisting of a kernel layer, a BN layer, and a nonlinear activation function. This basic network helps us understand the impacts of BN in three aspects. First, by viewing BN as an implicit regularizer, BN can be decomposed into population normalization (PN) and gamma decay as an explicit regularization. Second, learning dynamics of BN and the regularization show that training converged with large maximum and effective learning rate. Third, generalization of BN is explored by using statistical mechanics. Experiments demonstrate that BN in convolutional neural networks share the same traits of regularization as the above analyses."}, "keywords": ["BatchNorm"], "citation_intent": "result"}
{"citing_id": "2303.15904v1", "cited_id": "1905.04804", "section_title": "Arxiv:2303.15904V1 [Cs.Cv] 28 Mar 2023", "citation": "Results are reported using ResNet-50 as backbone on the YTVIS 2019 #REFR benchmark. Video Mask: using YTVIS video mask labels.", "text_before_citation": ["To further enforce temporal consistency through the video clip, TK-Loss is employed in a cyclic manner instead of using dense frame-wise connections. This greatly reduces memory cost with negligible performance drop.", "We extensively evaluate our MaskFreeVIS on four largescale VIS benchmarks, i.e., YouTube-VIS 2019/2021 #OTHEREFR , OVIS #OTHEREFR , and BDD100K MOTS #OTHEREFR .", "MaskFreeVIS achieves competitive VIS performance without using any video masks or even image mask labels on all datasets.", "Validated on various methods and backbones, MaskFreeVIS achieves 91.25% performance of its fully supervised counterparts, even outperforming a few recent fully-supervised methods #OTHEREFR on the popular YTVIS benchmark.", "Our simple yet effective design greatly narrows the performance gap between weakly-supervised and fully- Table 1 . Mask annotation requirement for state-of-the-art VIS methods."], "text_after_citation": ["Image Mask: using COCO #OTHEREFR image mask labels for image-based pretraining.", "Pseudo Video: using Pseudo Videos from COCO images for joint training #OTHEREFR . MaskFreeVIS achieves 91.5% (42.5 vs.", "46.4) of its fully-supervised baseline performance (Mask2Former) without using any masks during training. supervised video instance segmentation.", "It further demonstrates that expensive video masks, or even image masks, is not necessary for training high-performing VIS models.", "Our contributions are summarized as follows: (i) To utilize temporal information, we develop a new parameterfree Temporal KNN-patch Loss, which leverages temporal masks consistency using unsupervised one-to-k patch correspondence. We extensively analyze the TK-Loss through ablative experiments."], "citing_paper_content": {"title": "Mask-Free Video Instance Segmentation", "abstract": "Figure 1. Video instance segmentation (VIS) results of our MaskFreeVIS, trained without using any video or image mask annotation. By achieving a remarkable 42.5% mask AP on the YouTube-VIS val dataset, with a ResNet-50 backbone, our approach demonstrates that high-performing VIS can be learned even without any mask annotations."}, "cited_paper_content": {"title": "Video Instance Segmentation", "abstract": "In this paper we present a new computer vision task, named video instance segmentation. The goal of this new task is simultaneous detection, segmentation and tracking of instances in videos. In words, it is the first time that the image instance segmentation problem is extended to the video domain. To facilitate research on this new task, we propose a large-scale benchmark called YouTube-VIS, which consists of 2883 high-resolution YouTube videos, a 40-category label set and 131k high-quality instance masks. In addition, we propose a novel algorithm called MaskTrack R-CNN for this task. Our new method introduces a new tracking branch to Mask R-CNN to jointly perform the detection, segmentation and tracking tasks simultaneously. Finally, we evaluate the proposed method and several strong baselines on our new dataset. Experimental results clearly demonstrate the advantages of the proposed algorithm and reveal insight for future improvement. We believe the video instance segmentation task will motivate the community along the line of research for video understanding."}, "keywords": ["Video Mask"], "citation_intent": "method"}
{"citing_id": "2304.11033v1", "cited_id": "1901.03206", "section_title": "Mutability By", "citation": "Worse yet, the protocol introduces an additional issue in that it requires a majority of miners to act faithfully and actually perform the (legally mandated) mutations-something that it cannot guarantee by design #REFR .", "text_before_citation": ["create and formally prove an editable blockchain protocol #OTHEREFR .", "While any user can propose edits, the protocol ensures consensus-based voting on the proposals to prevent arbitrary edits.", "This also means that no trusted third party is required.", "The protocol is compatible with any consensus mechanism and even offers accountability of the performed edits.", "#OTHEREFR While this solution removes the need for a trusted third party, it does not solve the other issue of redactable and forgetting blockchains: every individual node in the network still needs to be trusted, as mutations are published as chain updates as well."], "text_after_citation": ["In contrast to that, our solution functions even with of adversarial network participants, as noted above.", "Furthermore, we do not depend on the honesty of the miners and, better still, do not require any changes of the blockchain software."], "citing_paper_content": {"title": "Decentralized Inverse Transparency With Blockchain", "abstract": "Employee data can be used to facilitate work, but their misusage may pose risks for individuals. Inverse transparency therefore aims to track all usages of personal data, allowing individuals to monitor them to ensure accountability for potential misusage. This necessitates a trusted log to establish an agreed-upon and non-repudiable timeline of events. The unique properties of blockchain facilitate this by providing immutability and availability. For power asymmetric environments such as the workplace, permissionless blockchain is especially beneficial as no trusted third party is required. Yet, two issues remain: (1) In a decentralized environment, no arbiter can facilitate and attest to data exchanges. Simple peer-to-peer sharing of data, conversely, lacks the required non-repudiation. (2) With data governed by privacy legislation such as the GDPR, the core advantage of immutability becomes a liability. After a rightful request, an individual's personal data need to be rectified or deleted, which is impossible in an immutable blockchain. To solve these issues, we present Kovacs, a decentralized data exchange and usage logging system for inverse transparency built on blockchain. Its new-usage protocol ensures non-repudiation, and therefore accountability, for inverse transparency. Its one-time pseudonym generation algorithm guarantees unlinkability and enables proof of ownership, which allows data subjects to exercise their legal rights regarding their personal data. With our implementation, we show the viability of our solution. The decentralized communication impacts performance and scalability, but exchange duration and storage size are still reasonable. More importantly, the provided information security meets high requirements. We conclude that Kovacs realizes decentralized inverse transparency through secure and GDPR-compliant use of permissionless blockchain. CCS Concepts: \u2022 Computer systems organization \u2192 Peer-to-peer architectures; \u2022 Security and privacy \u2192 Distributed systems security; Privacy-preserving protocols; Cryptography."}, "cited_paper_content": {"title": "Redactable Blockchain In The Permissionless Setting", "abstract": "Bitcoin is an immutable permissionless blockchain system that has been extensively used as a public bulletin board by many different applications that heavily relies on its immutability. However, Bitcoin's immutability is not without its fair share of demerits. Interpol exposed the existence of harmful and potentially illegal documents, images and links in the Bitcoin blockchain, and since then there have been several qualitative and quantitative analysis on the types of data currently residing in the Bitcoin blockchain. Although there is a lot of attention on blockchains, surprisingly the previous solutions proposed for data redaction in the permissionless setting are far from feasible, and require additional trust assumptions. Hence, the problem of harmful data still poses a huge challenge for law enforcement agencies like Interpol (Tziakouris, IEEE SP if a redaction gathers enough votes the operation is performed on the chain. As an extra feature, our protocol offers public verifiability and accountability for the redacted chain. Moreover, we provide formal security definitions and proofs showing that our protocol is secure against redactions that were not agreed by consensus. Additionally, we show the viability of our approach with a proof-of-concept implementation that shows only a tiny overhead in the chain validation of our protocol when compared to an immutable one."}, "keywords": ["protocol"], "citation_intent": "background"}
{"citing_id": "2303.01925v1", "cited_id": "1211.0358", "section_title": "Discussion And Limitations", "citation": "Another is to extend our framework and instead represent H with a deep GP prior #REFR .", "text_before_citation": ["The performance of the HGP on task 2, is somewhat underwhelming especially for larger K, relative to NODE.", "We believe some of this effect can be attributed to our choice of GP prior.", "We use a stationary GP prior, which is likely non-optimal for most Hamiltonian systems, which are typically nonstationary.", "The assumption of non-stationarity likely leads to the poor generalisation at new phase space points, since the GP prior will revert to zero mean functions there, which is not the case for the NN models.", "One option to improve the suitability of the GP prior is via the use of specifically designed kernels to represent symmetries in the system #OTHEREFR , or kernel structure learning #OTHEREFR ."], "text_after_citation": ["Deep GPs are adept at modelling complex, nonstationary functions, and so would be well suited to the task.", "Using the SVI scheme proposed by #OTHEREFR would make integration of deep GPs into our framework relatively straight forward, and would be interesting to undertake as part of future work.", "Control.", "The present method has significant potential at improving Bayesian online #OTHEREFR or policy-based RL (Yildiz et al., 2021) by incorporating Hamiltonian inductive biases.", "This requires expanding the model towards Hamiltonian systems with external forces, which relax the energy conservation assumption."], "citing_paper_content": {"title": "Learning Energy Conserving Dynamics Efficiently With Hamiltonian Gaussian Processes", "abstract": "Hamiltonian mechanics is one of the cornerstones of natural sciences. Recently there has been significant interest in learning Hamiltonian systems in a free-form way directly from trajectory data. Previous methods have tackled the problem of learning from many short, low-noise trajectories, but learning from a small number of long, noisy trajectories, whilst accounting for model uncertainty has not been addressed. In this work, we present a Gaussian process model for Hamiltonian systems with efficient decoupled parameterisation, and introduce an energy-conserving shooting method that allows robust inference from both short and long trajectories. We demonstrate the method's success in learning Hamiltonian systems in various data settings."}, "cited_paper_content": {"title": "Deep Gaussian Processes", "abstract": "In this paper we introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GP-LVM). We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples."}, "keywords": ["deep GP"], "citation_intent": "method"}
{"citing_id": "2303.10538v1", "cited_id": "1803.08475", "section_title": "Running Time Discussion", "citation": "As discussed in #REFR , running time is important but hard to compare since it is affected by many factors.", "text_before_citation": [], "text_after_citation": ["We report the clock time for solving all the test instances in Table 1 .", "For the UTSP (our method) and the state-of-the-art learning-based method Att-GCRN #OTHEREFR , we run the search algorithm on exactly the same environment (one Intel Xeon Gold 6326) for a fair comparison.", "And for other baselines, we directly refer to the results from #OTHEREFR .", "So the time there are only for indicative purpose since the computing hardware is not the same."], "citing_paper_content": {"title": "Unsupervised Learning For Solving The Travelling Salesman Problem", "abstract": "We propose UTSP, an unsupervised learning (UL) framework for solving the Travelling Salesman Problem (TSP). We train a Graph Neural Network (GNN) using a surrogate loss. The GNN outputs a heat map representing the probability for each edge to be part of the optimal path. We then apply local search to generate our final prediction based on the heat map. Our loss function consists of two parts: one pushes the model to find the shortest path and the other serves as a surrogate for the constraint that the route should form a Hamiltonian Cycle. Experimental results show that UTSP outperforms the existing data-driven TSP heuristics. Our approach is parameter efficient as well as data efficient: the model takes \u223c 10% of the number of parameters and \u223c 0.2% of training samples compared with reinforcement learning or supervised learning methods."}, "cited_paper_content": {"title": "Attention, Learn To Solve Routing Problems!", "abstract": "The recently presented idea to learn heuristics for combinatorial optimization problems is promising as it can save costly development. However, to push this idea towards practical implementation, we need better models and better ways of training. We contribute in both directions: we propose a model based on attention layers with benefits over the Pointer Network and we show how to train this model using REINFORCE with a simple baseline based on a deterministic greedy rollout, which we find is more efficient than using a value function. We significantly improve over recent learned heuristics for the Travelling Salesman Problem (TSP), getting close to optimal results for problems up to 100 nodes. With the same hyperparameters, we learn strong heuristics for two variants of the Vehicle Routing Problem (VRP), the Orienteering Problem (OP) and (a stochastic variant of) the Prize Collecting TSP (PCTSP), outperforming a wide range of baselines and getting results close to highly optimized and specialized algorithms."}, "keywords": ["running time"], "citation_intent": "background"}
{"citing_id": "2304.08994v1", "cited_id": "1804.01654", "section_title": "Introduction", "citation": "This bounding box is subdivided and serves as initial mesh, which is refined by an iterative mesh refinement as proposed in #REFR .", "text_before_citation": ["Transportation logistics and warehousing are a central part of every supply chain and play an important strategic role in the Industry 4.0 era #OTHEREFR .", "However, several challenges need to be faced by companies working in the logistics sector: clients demand cheaper, faster and more pre- Figure 1 .", "We take an RGB image as input and use Cube R-CNN's Cube Head #OTHEREFR to estimate a 3D bounding box."], "text_after_citation": ["For training and evaluation we present Parcel3D, a novel dataset of normal and damaged parcels with full 3D annotations.", "cisely scheduled deliveries while at the same time, cities and highways are congested, and environmental concerns are of rising importance.", "To tackle these challenges, process automation has huge potential #OTHEREFR .", "Key processes for automation in logistics are identification, digital measurement, damage detection and tampering recognition of packaging units, all of which we work towards with the approach presented in this work. Identification is necessary for process documentation and parcel tracking.", "Damage and tampering detection can be utilized to increase the safety and security along the supply chain #OTHEREFR ."], "citing_paper_content": {"title": "Parcel3D: Shape Reconstruction From Single Rgb Images For Applications In Transportation Logistics", "abstract": "We focus on enabling damage and tampering detection in logistics and tackle the problem of 3D shape reconstruction of potentially damaged parcels. As input we utilize single RGB images, which corresponds to use-cases where only simple handheld devices are available, e.g. for postmen during delivery or clients on delivery. We present a novel synthetic dataset, named Parcel3D, that is based on the Google Scanned Objects (GSO) dataset and consists of more than 13,000 images of parcels with full 3D annotations. The dataset contains intact, i.e. cuboid-shaped, parcels and damaged parcels, which were generated in simulations. We work towards detecting mishandling of parcels by presenting a novel architecture called CubeRefine R-CNN, which combines estimating a 3D bounding box with an iterative mesh refinement. We benchmark our approach on Parcel3D and an existing dataset of cuboidshaped parcels in real-world scenarios. Our results show, that while training on Parcel3D enables transfer to the real world, enabling reliable deployment in real-world scenarios is still challenging. CubeRefine R-CNN yields competitive performance in terms of Mesh AP and is the only model that directly enables deformation assessment by 3D mesh comparison and tampering detection by comparing viewpoint invariant parcel side surface representations. Dataset and code are available at https://a-nau.github.io/parcel3d."}, "cited_paper_content": {"title": "Pixel2Mesh: Generating 3D Mesh Models From Single Rgb Images", "abstract": "We propose an end-to-end deep learning architecture that produces a 3D shape in triangular mesh from a single color image. Limited by the nature of deep neural network, previous methods usually represent a 3D shape in volume or point cloud, and it is non-trivial to convert them to the more ready-to-use mesh model. Unlike the existing methods, our network represents 3D mesh in a graph-based convolutional neural network and produces correct geometry by progressively deforming an ellipsoid, leveraging perceptual features extracted from the input image. We adopt a coarse-to-fine strategy to make the whole deformation procedure stable, and define various of mesh related losses to capture properties of different levels to guarantee visually appealing and physically accurate 3D geometry. Extensive experiments show that our method not only qualitatively produces mesh model with better details, but also achieves higher 3D shape estimation accuracy compared to the state-of-the-art."}, "keywords": ["iterative mesh refinement"], "citation_intent": "method"}
{"citing_id": "2304.14122v1", "cited_id": "1903.12395", "section_title": "C. Comparison With State-Of-The-Arts", "citation": "Compared with recent DIL #REFR , our method still achieves better results on MARS and DukeMCMT-VID.", "text_before_citation": ["Meanwhile, with pyramid spatial-temporal learning, PSTA #OTHEREFR gains remarkable 91.5% and 98.3% Rank-1 accuracy on MARS and DukeMCMT-VID datasets.", "Different from PSTA, we propose a novel spatialtemporal complementary learning framework to obtain richer video representations.", "Thus, we obtain better performance than PSTA on MARS and DukeMCMT-VID, which validates the effectiveness of our proposed method.", "When compared with some methods that utilize Transformer, our method also shows great superiority in performance.", "For example, compared with STT #OTHEREFR , our method has 1.2% and 3.6% gains in mAP and Rank-1 accuracy."], "text_after_citation": ["Besides, TMT #OTHEREFR introduces a multi-view Transformer to extract comprehensive video representations. However, our method still performs better than TMT.", "We note that all these methods add Transformer layers after CNN backbones to obtain enhanced representations.", "Different from them, we deeply couple CNNs and Transformers, and propose a complementary content attention for spatial complementary learning.", "Besides, the hierarchical temporal aggregation is leveraged to progressively integrate temporal information.", "In this way, we utilize two kinds of typical visual features from the same videos and achieve spatial-temporal complementary learning for more informative representations."], "citing_paper_content": {"title": "Deeply-Coupled Convolution-Transformer With Spatial-Temporal Complementary Learning For Video-Based Person Re-Identification", "abstract": "Advanced deep Convolutional Neural Networks (CNNs) have shown great success in video-based person Re-Identification (Re-ID). However, they usually focus on the most obvious regions of persons with a limited global representation ability. Recently, it witnesses that Transformers explore the inter-patch relations with global observations for performance improvements. In this work, we take both sides and propose a novel spatial-temporal complementary learning framework named Deeply-Coupled Convolution-Transformer (DCCT) for high-performance video-based person Re-ID. Firstly, we couple CNNs and Transformers to extract two kinds of visual features and experimentally verify their complementarity. Further, in spatial, we propose a Complementary Content Attention (CCA) to take advantages of the coupled structure and guide independent features for spatial complementary learning. In temporal, a Hierarchical Temporal Aggregation (HTA) is proposed to progressively capture the inter-frame dependencies and encode temporal information. Besides, a gated attention is utilized to deliver aggregated temporal information into the CNN and Transformer branches for temporal complementary learning. Finally, we introduce a self-distillation training strategy to transfer the superior spatial-temporal knowledge to backbone networks for higher accuracy and more efficiency. In this way, two kinds of typical features from same videos are integrated mechanically for more informative representations. Extensive experiments on four public Re-ID benchmarks demonstrate that our framework could attain better performances than most state-of-the-art methods."}, "cited_paper_content": {"title": "Few-Shot Deep Adversarial Learning For Video-Based Person Re-Identification", "abstract": "Video-based person re-identification (re-ID) refers to matching people across camera views from arbitrary unaligned video footages. Existing methods rely on supervision signals to optimise a projected space under which the distances between inter/intra-videos are maximised/minimised. However, this demands exhaustively labelling people across camera views, rendering them unable to be scaled in large networked cameras. Also, it is noticed that learning effective video representations with view invariance is not explicitly addressed for which features exhibit different distributions otherwise. Thus, matching videos for person re-ID demands flexible models to capture the dynamics in time-series observations and learn view-invariant representations with access to limited labeled training samples. In this paper, we propose a novel few-shot deep learning approach to video-based person re-ID, to learn comparable representations that are discriminative and view-invariant. The proposed method is developed on the variational recurrent neural networks (VRNNs) and trained adversarially to produce latent variables with temporal dependencies that are highly discriminative yet view-invariant in matching persons. Through extensive experiments conducted on three benchmark datasets, we empirically show the capability of our method in creating view-invariant temporal features and state-of-the-art performance achieved by our method."}, "keywords": ["DukeMCMT-VID", "recent DIL"], "citation_intent": "result"}
{"citing_id": "2304.10060v1", "cited_id": "1805.10074", "section_title": "Main Results", "citation": "Recently, under the same capacity assumption, the work #REFR studies the mean square convergence of averaging online estimator with least square loss and multiple passes.", "text_before_citation": ["A small value of \u03b2 implies a fast polynomially decaying rate at least achieved by the eigenvalues {\u03bb k } k\u2208N .", "One can refer to Theorem 5 in our recent work #OTHEREFR , which provides a characterization of the relationship between the capacity assumption (9) in this paper and the decaying rate of integral operator eigenvalues.", "If the eigenvalues decay exponentially, the index \u03b2 can be arbitrarily close to zero.", "The polynomially decaying of the eigenvalue is typical for the Sobolev smooth kernels on domains in Euclidean spaces, and the parameter \u03b2 depends on the smoothness of the kernel K.", "While the exponentially decaying integral operator is typical for the analytic kernels on domains in Euclidean spaces."], "text_after_citation": ["Under the regularity condition (7) on the target function f \u03c1 and the capacity assumption (9) on the hypothesis space H K , we obtain the following sharp capacity dependent results for strong convergence in H K .", "Theorem 2. Let {f t } T +1", "t=1 be defined by algorithm #OTHEREFR with a windowing function W .", "Assume that W satisfies (2) and (3) with some p > 0, the regularity condition #OTHEREFR holds with r > 1 2 and the capacity assumption (9) holds with 0 < \u03b2 < 1. Choose step size \u03b7 = 1", "\u03b7 0 T 1\u22122r\u2212\u03b2 2r+\u03b2 with \u03b7 0 \u2265 max C W \u03ba 2 , 1 e + 2\u03ba 2 W \u2032 + (0) 2 ."], "citing_paper_content": {"title": "Optimality Of Robust Online Learning \u2020", "abstract": "In this paper, we study an online learning algorithm with a robust loss function L \u03c3 for regression over a reproducing kernel Hilbert space (RKHS). The loss function L \u03c3 involving a scaling parameter \u03c3 > 0 can cover a wide range of commonly used robust losses. The proposed algorithm is then a robust alternative for online least squares regression aiming to estimate the conditional mean function. For properly chosen \u03c3 and step size, we show that the last iterate of this online algorithm can achieve optimal capacity independent convergence in the mean square distance. Moreover, if additional information on the underlying function space is known, we also establish optimal capacity dependent rates for strong convergence in RKHS. To the best of our knowledge, both of the two results are new to the existing literature of online learning."}, "cited_paper_content": {"title": "Statistical Optimality Of Stochastic Gradient Descent On Hard Learning Problems Through Multiple Passes", "abstract": "We consider stochastic gradient descent (SGD) for least-squares regression with potentially several passes over the data. While several passes have been widely reported to perform practically better in terms of predictive performance on unseen data, the existing theoretical analysis of SGD suggests that a single pass is statistically optimal. While this is true for low-dimensional easy problems, we show that for hard problems, multiple passes lead to statistically optimal predictions while single pass does not; we also show that in these hard models, the optimal number of passes over the data increases with sample size. In order to define the notion of hardness and show that our predictive performances are optimal, we consider potentially infinite-dimensional models and notions typically associated to kernel methods, namely, the decay of eigenvalues of the covariance matrix of the features and the complexity of the optimal predictor as measured through the covariance matrix. We illustrate our results on synthetic experiments with non-linear kernel methods and on a classical benchmark with a linear model."}, "keywords": ["online estimator", "least square loss"], "citation_intent": "background"}
{"citing_id": "2304.11514v1", "cited_id": "1602.03602", "section_title": "I. Introduction", "citation": "Unmanned aerial vehicle (UAV), thanks to its low cost, high flexibility and high probability of line-of-propagation (LoP) links, has become an attractive means of improving air-ground transmission quality #REFR .", "text_before_citation": ["Nowadays, wireless networks serve a wide range of civilian and military applications and have become an essential part of our routine #OTHEREFR , #OTHEREFR .", "Improving the performance of wireless communication has become a popular research topic."], "text_after_citation": ["In general, UAVs can not only act as base stations #OTHEREFR , relays #OTHEREFR , etc, to transmit or forward signals, but also for data collection #OTHEREFR and positioning #OTHEREFR . Currently, UAV transmission technology has been widely explored.", "For instance, the authors in #OTHEREFR considered a cellular network deployment in which UAV-to-UAV launch-receive pairs used the identical spectrum as the uplink of cellular ground users, and the performances of underlay and overlay spectrum sharing mechanisms were analyzed and compared.", "In #OTHEREFR , to enhance the throughput of a single-cell multiuser orthogonal frequency division multiple access network with single UAV while guaranteeing the user fairness, an efficient method was proposed that outperformed the random and cellular schemes in terms of user fairness and sum rate.", "UAV-enabled relay under malicious jamming was considered in #OTHEREFR , and the successive convex approximation (SCA) algorithm was employed to maximize the end-to-end throughput.", "However, UAV network also presents some challenges that affect their performance."], "citing_paper_content": {"title": "Joint Beamforming And Phase Shift Design For Hybrid-Irs-And-Uav-Aided Directional Modulation Network", "abstract": "Recently, intelligent reflecting surface (IRS) and unmanned aerial vehicle (UAV) have been introduced into wireless communication systems to enhance the performance of air-ground transmission. To make a good balance between performance, cost, and power consumption, a hybrid-IRS-and-UAV-assisted directional modulation (DM) network is investigated in this paper, where the hybrid IRS consists of passive and active reflecting elements. To maximize the achievable rate, three optimization algorithms, called maximum signal-to-noise ratio (SNR)-fractional programming (FP) (Max-SNR-FP), maximum SNR-equal amplitude reflecting (EAR) (Max-SNR-EAR), and maximum SNR-majorization-minimization (MM) (Max-SNR-MM), are proposed to jointly design the beamforming vector and phase shift matrix (PSM) of hybrid IRS by alternately optimizing one and giving another. The Max-SNR-FP method employs the successive convex approximation and FP methods to derive the beamforming vector and hybrid IRS PSM. The Max-SNR-EAR method adopts the maximum signal-to-leakagenoise ratio method and the criteria of phase alignment and EAR to design them. In addition, the Max-SNR-MM method utilizes the MM criterion to derive the IRS PSM. Simulation results show that the rates harvested by the proposed three methods are slightly lower than those of active IRS with higher power consumption, which are 35 percent higher than those of no IRS and random phase IRS, while passive IRS achieves only about 17 percent rate gain over the latter. Moreover, compared to Max-SNR-FP, the proposed Max-SNR-EAR and Max-SNR-MM methods make an obvious complexity degradation at the price of a slight performance loss."}, "cited_paper_content": {"title": "Wireless Communications With Unmanned Aerial Vehicles: Opportunities And Challenges", "abstract": "Wireless communication systems that include unmanned aerial vehicles promise to provide cost-effective wireless connectivity for devices without infrastructure coverage. Compared to terrestrial communications or those based on high-altitude platforms, on-demand wireless systems with low-altitude UAVs are in general faster to deploy, more flexibly reconfigured, and likely to have better communication channels due to the presence of short-range line-of-sight links. However, the utilization of highly mobile and energy-constrained UAVs for wireless communications also introduces many new challenges. In this article, we provide an overview of UAV-aided wireless communications, by introducing the basic networking architecture and main channel characteristics, highlighting the key design considerations as well as the new opportunities to be exploited."}, "keywords": ["Unmanned aerial vehicle"], "citation_intent": "background"}
{"citing_id": "2303.15916v2", "cited_id": "1802.06739", "section_title": "Experiment 5: T-Sne Visualization Of Generated Data", "citation": "Similar to the previous dataset, the DPWGAN #REFR produces a distribution with an offset, resulting in a bad performance.", "text_before_citation": ["between the private and public data, which is reflected in the accuracy, as the public classifier cannot perform well on the private dataset. Furthermore, the two classes are not separated well.", "In contrast to that, Figure 1d shows that the generated data of the GSWGAN #OTHEREFR , although it is different from the original distribution, separated the two classes well.", "In addition, this model was able to perform well, although the distribution shows some differences.", "For the accuracy results related to this figure, the reader is referred to Table 2 .", "The last dataset is visualized in Figure 1g and Figure 1h ."], "text_after_citation": ["In contrast to that, the GSWGAN #OTHEREFR generated a dataset that shows a similar distribution within the T-SNE plot, resulting in high performance for the classifier trained on that dataset.", "One general finding across all datasets was that the DPWGAN #OTHEREFR produces data that is less similar to the original data.", "This can be explained by the fact that differential privacy is applied to both the discriminator and generator, resulting in a worse performance of the discriminator.", "This is not the case for the GSWGAN #OTHEREFR as the privacy constraints are only applied to the generator.", "The stronger discriminator seems to improve the quality of the generated data."], "citing_paper_content": {"title": "From Private To Public: Benchmarking Gans In The Context Of Private Time Series Classification", "abstract": "Deep learning has proven to be successful in various domains and for different tasks. However, when it comes to private data several restrictions are making it difficult to use deep learning approaches in these application fields. Recent approaches try to generate data privately instead of applying a privacy-preserving mechanism directly, on top of the classifier. The solution is to create public data from private data in a manner that preserves the privacy of the data. In this work, two very prominent GAN-based architectures were evaluated in the context of private time series classification. In contrast to previous work, mostly limited to the image domain, the scope of this benchmark was the time series domain. The experiments show that especially GSWGAN performs well across a variety of public datasets outperforming the competitor DPWGAN. An analysis of the generated datasets further validates the superiority of GSWGAN in the context of time series generation."}, "cited_paper_content": {"title": "Differentially Private Generative Adversarial Network", "abstract": "Generative Adversarial Network (GAN) and its variants have recently attracted intensive research interests due to their elegant theoretical foundation and excellent empirical performance as generative models. These tools provide a promising direction in the studies where data availability is limited. One common issue in GANs is that the density of the learned generative distribution could concentrate on the training data points, meaning that they can easily remember training samples due to the high model complexity of deep networks. This becomes a major concern when GANs are applied to private or sensitive data such as patient medical records, and the concentration of distribution may divulge critical patient information. To address this issue, in this paper we propose a differentially private GAN (DPGAN) model, in which we achieve differential privacy in GANs by adding carefully designed noise to gradients during the learning procedure. We provide rigorous proof for the privacy guarantee, as well as comprehensive empirical evidence to support our analysis, where we demonstrate that our method can generate high quality data points at a reasonable privacy level."}, "keywords": ["previous dataset", "DPWGAN"], "citation_intent": "result"}
{"citing_id": "2304.08979v1", "cited_id": "1809.02789", "section_title": "Evaluation Dataset", "citation": "The answers (yes/no) are marked by human annotators if certain Wikipedia pages contain sufficient information to address the questions. \u2022 OpenbookQA (OQA) #REFR is a multiple-choice reasoning dataset. The questions are derived from 1,326 core science facts.", "text_before_citation": ["We employ ten widely used benchmark QA datasets in our study, including BoolQ #OTHEREFR , Open-bookQA (OQA) #OTHEREFR , RACE #OTHEREFR , ARC #OTHEREFR , Common-senseQA (CQA) #OTHEREFR , SQuAD1 #OTHEREFR , SQuAD2 #OTHEREFR , Narra-tiveQA (NQA) #OTHEREFR , ELI5 #OTHEREFR , and TruthfulQA (TQA) #OTHEREFR .", "These datasets encompass a broad range of QA capabilities, such as reading comprehension (BoolQ, SQuDA1/2, RACE), reasoning (OQA, ARC), commonsense (CQA), full document comprehension (NQA, ELI5), and truthfulness (TQA).", "Furthermore, they comprise all four QA tasks #OTHEREFR , including yes/no (BoolQ), multiple-choice (OQA, RACE, ARC, CQA), extractive (SQuAD 1/2), and abstractive tasks (NQA, ELI5, TQA).", "They thus offer a solid foundation to comprehensively evaluate the ChatGPT's reliability in various realworld QA scenarios.", "Their details are outlined below and summarized in Table 2. \u2022 BoolQ #OTHEREFR is a yes/no reading comprehension dataset. The questions are derived from aggregated Google searches."], "text_after_citation": ["The answers consist of 4 candidates, of which only one is correct, requiring reasoning between questions and the given science facts and common knowledge.", "\u2022 RACE #OTHEREFR is a multiple-choice reading comprehension dataset.", "The questions are derived from English exams for Chinese students.", "The answers include 4 candidates, of which only one is correct, requiring reading comprehension of English passages.", "\u2022 ARC #OTHEREFR is a multiple-choice reasoning dataset."], "citing_paper_content": {"title": "In Chatgpt We Trust? Measuring And Characterizing The Reliability Of Chatgpt", "abstract": "The way users acquire information is undergoing a paradigm shift with the advent of ChatGPT. Unlike conventional search engines, ChatGPT retrieves knowledge from the model itself and generates answers for users. ChatGPT's impressive question-answering (QA) capability has attracted more than 100 million users within a short period of time but has also raised concerns regarding its reliability. In this paper, we perform the first large-scale measurement of ChatGPT's reliability in the generic QA scenario with a carefully curated set of 5,695 questions across ten datasets and eight domains. We find that ChatGPT's reliability varies across different domains, especially underperforming in law and science questions. We also demonstrate that system roles, originally designed by OpenAI to allow users to steer ChatGPT's behavior, can impact ChatGPT's reliability. We further show that ChatGPT is vulnerable to adversarial examples, and even a single character change can negatively affect its reliability in certain cases. We believe that our study provides valuable insights into ChatGPT's reliability and underscores the need for strengthening the reliability and security of large language models (LLMs)."}, "cited_paper_content": {"title": "Can A Suit Of Armor Conduct Electricity? A New Dataset For Open Book Question Answering", "abstract": "We present a new kind of question answering dataset, OpenBookQA, modeled after open book exams for assessing human understanding of a subject. The open book that comes with our questions is a set of 1329 elementary level science facts. Roughly 6000 questions probe an understanding of these facts and their application to novel situations. This requires combining an open book fact (e.g., metals conduct electricity) with broad common knowledge (e.g., a suit of armor is made of metal) obtained from other sources. While existing QA datasets over documents or knowledge bases, being generally self-contained, focus on linguistic understanding, OpenBookQA probes a deeper understanding of both the topic---in the context of common knowledge---and the language it is expressed in. Human performance on OpenBookQA is close to 92%, but many state-of-the-art pre-trained QA methods perform surprisingly poorly, worse than several simple neural baselines we develop. Our oracle experiments designed to circumvent the knowledge retrieval bottleneck demonstrate the value of both the open book and additional facts. We leave it as a challenge to solve the retrieval problem in this multi-hop setting and to close the large gap to human performance."}, "keywords": ["multiple-choice reasoning dataset"], "citation_intent": "method"}
{"citing_id": "2304.11546v1", "cited_id": "1904.02689", "section_title": "Guide Line", "citation": "Then, the relation between response range and grazing angle could be described in Equation #REFR .", "text_before_citation": ["However, in some case it is observed that the response is very scattered, and the response peak is often indistinguishable from other non-response points.", "Figure 3 helps to understand the relation between the response range and grazing angle.", "The black dotted line denotes the detected vectorized lane (hereinafter referred to as \"vector line\"), and the green dotted lines alongside indicate the range of pixels that generate a response on the heatmap.", "The greener the background is, the more positive the sample is. d is the radius of the range.", "The tangents of two guide lines are illustrated in orange and magenta, with grazing angles as \u03b11 and \u03b12, respectively."], "text_after_citation": ["The bigger the angle (close to 90 \u2022 ), the tighter the response range.", "As the grazing angle gets smaller, the range scatters rapidly. This reflects the results in Figure 1a .", "Therefore, the guide line principle is to \"reduce the number of lanes of small grazing angles\".", "EQUATION"], "citing_paper_content": {"title": "Canet: Curved Guide Line Network With Adaptive Decoder For Lane Detection", "abstract": "Lane detection is challenging due to the complicated onroad scenarios and line deformation from different camera perspectives. Lots of solutions were proposed, but can not deal with \"corner lanes\" well. To address this problem, this paper proposes a new top-down deep learning lane detection approach, CANET. A lane instance is first responded by the heatmap on the U-shaped \"curved guide line\" at global semantic level, thus the corresponding features of each lane are aggregated at the response point. Then CANET obtains the heatmap response of the entire lane through conditional convolution, and finally decodes the point set to describe lanes via adaptive decoder. The prototype is implemented with Pytorch, and evaluated against 3 well-known datasets extensively. The experimental results show that CANET reaches SOTA in different metrics. Our code will be released soon."}, "cited_paper_content": {"title": "Yolact: Real-Time Instance Segmentation", "abstract": "We present a simple, fully-convolutional model for real-time instance segmentation that achieves 29.8 mAP on MS COCO at 33.5 fps evaluated on a single Titan Xp, which is significantly faster than any previous competitive approach. Moreover, we obtain this result after training on only one GPU. We accomplish this by breaking instance segmentation into two parallel subtasks: (1) generating a set of prototype masks and (2) predicting per-instance mask coefficients. Then we produce instance masks by linearly combining the prototypes with the mask coefficients. We find that because this process doesn't depend on repooling, this approach produces very high-quality masks and exhibits temporal stability for free. Furthermore, we analyze the emergent behavior of our prototypes and show they learn to localize instances on their own in a translation variant manner, despite being fully-convolutional. Finally, we also propose Fast NMS, a drop-in 12 ms faster replacement for standard NMS that only has a marginal performance penalty."}, "keywords": ["grazing angle", "Equation"], "citation_intent": "background"}
{"citing_id": "2303.16129v1", "cited_id": "1909.11875", "section_title": "F. Security And Privacy", "citation": "Fortunately, several privacy-preserving distributed learning frameworks, such as FL #REFR , have been proposed to empower privacypreserving AIGC model fine-tuning and inference at mobile AIGC networks.", "text_before_citation": ["For example, AI-generated text can be used by malicious users to complete phishing emails, thus compromising the security and privacy of normal users #OTHEREFR .", "To ensure secure AIGC services, providers must choose trusted AIGC solutions and train AI models in a secure manner while providing secure hints and answers to AIGC service users.", "1) Privacy-preserving AIGC Service Provisioning: During the lifecycle of providing AIGC services, privacy information in large-scale datasets and user requests needs to be kept secure to prevent privacy breaches.", "In mobile AIGC networks, the generation and storage of data for AIGC model training occur at edge servers and mobile devices #OTHEREFR .", "Unlike resourceful cloud data centers, edge and mobile layers have limited defense capacities against various attacks."], "text_after_citation": ["In preserving user privacy in AIGC networks, FL is a distributed ML approach that allows users to transmit local models instead of data during model training #OTHEREFR - #OTHEREFR . Specifically, as illustrated in Fig.", "20 , there are two major approaches to employing FL in AIGC networks \u2022 Secure aggregation: While FL is being learned, the mobile devices send local updates to edge servers for global aggregation.", "During global aggregation, authenticated encryption allows the use of secret sharing mechanisms.", "\u2022 Differential privacy: Differential privacy can prevent FL servers from identifying the owners of a local update.", "Differential privacy is similar to secure aggregation in that it prevents FL servers from identifying owners of local updates."], "citing_paper_content": {"title": "Unleashing The Power Of Edge-Cloud Generative Ai In Mobile Networks: A Survey Of Aigc Services", "abstract": "Intelligence-Generated Content (AIGC) is an automated method for generating, manipulating, and modifying valuable and diverse data using AI algorithms creatively. This survey paper focuses on the deployment of AIGC applications, e.g., ChatGPT and Dall-E, at mobile edge networks, namely mobile AIGC networks, that provide personalized and customized AIGC services in real time while maintaining user privacy. We begin by introducing the background and fundamentals of generative models and the lifecycle of AIGC services at mobile AIGC networks, which includes data collection, training, finetuning, inference, and product management. We then discuss the collaborative cloud-edge-mobile infrastructure and technologies required to support AIGC services and enable users to access AIGC at mobile edge networks. Furthermore, we explore AIGCdriven creative applications and use cases for mobile AIGC networks. Additionally, we discuss the implementation, security, and privacy challenges of deploying mobile AIGC networks. Finally, we highlight some future research directions and open issues for the full realization of mobile AIGC networks."}, "cited_paper_content": {"title": "Federated Learning In Mobile Edge Networks: A Comprehensive Survey", "abstract": "In recent years, mobile devices are equipped with increasingly advanced sensing and computing capabilities. Coupled with advancements in Deep Learning (DL), this opens up countless possibilities for meaningful applications. Traditional Machine Learning (ML) approaches require the data to be centralized in a cloud server. However, this results in critical issues related to unacceptable latency and communication inefficiency. To this end, Mobile Edge Computing (MEC) has been proposed to bring intelligence closer to the edge, where data is produced. However, conventional enabling technologies for ML at mobile edge networks still require personal data to be shared with external parties, e.g., edge servers. Recently, in light of increasingly stringent data privacy legislation and growing privacy concerns, the concept of Federated Learning (FL) has been introduced. In FL, end devices use their local data to train an ML model required by the server. The end devices then send the model updates rather than raw data to the server for aggregation. FL can serve as an enabling technology in mobile edge networks since it enables the collaborative training of an ML model and also enables DL for mobile edge network optimization. However, in a large-scale and complex mobile edge network, heterogeneous devices with varying constraints are involved. This raises challenges of communication costs, resource allocation, and privacy and security in the implementation of FL at scale. In this survey, we begin with an introduction to the background and fundamentals of FL. Then, we highlight the aforementioned challenges of FL implementation and review existing solutions. Furthermore, we present the applications of FL for mobile edge network optimization. Finally, we discuss the important challenges, open issues and future research directions in FL."}, "keywords": ["privacypreserving AIGC model", "mobile AIGC networks"], "citation_intent": "background"}
{"citing_id": "2303.12408v2", "cited_id": "2003.08934", "section_title": "D. Inward-Facing Dataset", "citation": "EgoNeRF shows comparable results in the Synthetic-NeRF dataset #REFR , which contains 8 synthetic objects.", "text_before_citation": ["The spherical grid of EgoNeRF aligns nicely with outward-facing scenes, not inward-facing images of typical NeRF settings.", "We optionally report results from widely-used datasets for novel view synthesis in Tab. 3. Table 5 . Out of distribution test.", "r is the distance between the center of training camera trajectory and position of test view."], "text_after_citation": ["In mip-NeRF 360 dataset #OTHEREFR , which contains inward-facing objects but has unbounded background scenes, EgoNeRF outperforms other baselines except mip-NeRF 360."], "citing_paper_content": {"title": "Balanced Spherical Grid For Egocentric View Synthesis", "abstract": "Figure 1. We propose a practical solution to reconstruct large-scale scenes from a short egocentric video. (a) Our scalable capturing setup observes the holistic environment by casually swiping a selfie stick with an omnidirectional camera attached. (b) Then we optimize our balanced spherical feature grids which are tailored for the outward-looking setup."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["Synthetic-NeRF dataset", "8 synthetic objects"], "citation_intent": "result"}
{"citing_id": "2303.06507v1", "cited_id": "1708.00171", "section_title": "D. Learning A Varying Noise Model", "citation": "They show that an Inverse-Wishart distribution can be applied as the prior on the covariance, which results in a Student's t distribution over the noise #REFR .", "text_before_citation": ["We first review their methodology for uncorrelated noise prediction (i.e., \u03b8(\u03c8 * |D) = W(\u03c8 * |D)), then present how we adapt their methodology. Vega-Brown et al.", "#OTHEREFR formulate an estimator for \u03b8 * given a target feature, \u03c8 * .", "A joint posterior between \u03b8 * and \u03b8 1:N is formed using Bayes' rule and \u03b8 1:N is then marginalized out, resulting in", "p(\u03b8 * |\u03c8 * , D) \u221d N i=1 p(y i |x i , \u03b8 * , \u03c8 i , \u03c8 * ) p(\u03b8 * |\u03c8 * ), (15)", "where we note the uncorrelated factorization of the likelihood terms."], "text_after_citation": ["In our work we will apply an uninformative prior, p(\u03b8 * |\u03c8 * ) \u221d 1, which is an adaptation of their prior work, CELLO #OTHEREFR , #OTHEREFR .", "The key innovation of Vega-Brown et al.", "#OTHEREFR is their choice in modelling the likelihood in (15) (referred to as the extended likelihood), as", "EQUATION", "where h(\u03c8 i , \u03c8 * ) is a kernel function."], "citing_paper_content": {"title": "Towards Consistent Batch State Estimation Using A Time-Correlated Measurement Noise Model", "abstract": "In this paper, we present an algorithm for learning time-correlated measurement covariances for application in batch state estimation. We parameterize the inverse measurement covariance matrix to be block-banded, which conveniently factorizes and results in a computationally efficient approach for correlating measurements across the entire trajectory. We train our covariance model through supervised learning using the groundtruth trajectory. In applications where the measurements are time-correlated, we demonstrate improved performance in both the mean posterior estimate and the covariance (i.e., improved estimator consistency). We use an experimental dataset collected using a mobile robot equipped with a laser rangefinder to demonstrate the improvement in performance. We also verify estimator consistency in a controlled simulation using a statistical test over several trials."}, "cited_paper_content": {"title": "Probe-Gk: Predictive Robust Estimation Using Generalized Kernels", "abstract": "Many algorithms in computer vision and robotics make strong assumptions about uncertainty, and rely on the validity of these assumptions to produce accurate and consistent state estimates. In practice, dynamic environments may degrade sensor performance in predictable ways that cannot be captured with static uncertainty parameters. In this paper, we employ fast nonparametric Bayesian inference techniques to more accurately model sensor uncertainty. By setting a prior on observation uncertainty, we derive a predictive robust estimator, and show how our model can be learned from sample images, both with and without knowledge of the motion used to generate the data. We validate our approach through Monte Carlo simulations, and report significant improvements in localization accuracy relative to a fixed noise model in several settings, including on synthetic data, the KITTI dataset, and our own experimental platform."}, "keywords": ["covariance"], "citation_intent": "background"}
{"citing_id": "2303.13396v1", "cited_id": "1905.05055", "section_title": "C. Thresholds Used In Metrics", "citation": "For a given algorithm, varying the threshold values can result in distinct performance profiles, e.g., a precision-recall curve, and several thresholds may be used together for the purposes of evaluation and comparison, as is common practice in the object detection literature #REFR .", "text_before_citation": ["We provide Segment-to-text IoU (IoU st ) scores with several \u03c4 CLIP threshold values in Figure 12 and Table 6 .", "Selecting the threshold \u03c4 CLIP is more challenging, since there is no established consensus or user studies to rely on.", "Figure 13 shows histograms of CLIP similarity scores between groundtruth image segments and their corresponding ground-truth labels in Pascal Context and Pascal VOC datasets.", "Given the distributions, we select \u03c4 CLIP = 0.1 to be on the safe side to report Segment-to-text IoU scores in the main experiment.", "It is important to note that for our zero-guidance segmentation problem, the thresholds \u03c4 CLIP and \u03c4 SBERT are used in the label reassignment verification process (Section 4.2), which is part of the evaluation not the segmentation algorithm itself."], "text_after_citation": ["IoU threshold (\u03c4 IoU ).", "We use \u03c4 IoU = 0.5, which is commonly used in object detection tasks to determine if a predicted bounding box is \"correct\" compared to the ground"], "citing_paper_content": {"title": "Zero-Guidance Segmentation Using Zero Segment Labels", "abstract": "CLIP has enabled new and exciting joint vision-language applications, one of which is open-vocabulary segmentation, which can locate any segment given an arbitrary text query. In our research, we ask whether it is possible to discover semantic segments without any user guidance in the form of text queries or predefined classes, and label them using natural language automatically? We propose a novel problem zero-guidance segmentation and the first baseline that leverages two pre-trained generalist models, DINO and CLIP, to solve this problem without any finetuning or segmentation dataset. The general idea is to first segment an image into small over-segments, encode them into CLIP's visual-language space, translate them into text labels, and merge semantically similar segments together. The key challenge, however, is how to encode a visual segment into a segment-specific embedding that balances global and local context information, both useful for recognition. Our main contribution is a novel attention-masking technique that balances the two contexts by analyzing the attention layers inside CLIP. We also introduce several metrics for the evaluation of this new task. With CLIP's innate knowledge, our method can precisely locate the Mona Lisa painting among a museum crowd (Figure 1"}, "cited_paper_content": {"title": "Object Detection In 20 Years: A Survey", "abstract": "Object detection, as of one the most fundamental and challenging problems in computer vision, has received great attention in recent years. Its development in the past two decades can be regarded as an epitome of computer vision history. If we think of today's object detection as a technical aesthetics under the power of deep learning, then turning back the clock 20 years we would witness the wisdom of cold weapon era. This paper extensively reviews 400+ papers of object detection in the light of its technical evolution, spanning over a quarter-century's time (from the 1990s to 2019). A number of topics have been covered in this paper, including the milestone detectors in history, detection datasets, metrics, fundamental building blocks of the detection system, speed up techniques, and the recent state of the art detection methods. This paper also reviews some important detection applications, such as pedestrian detection, face detection, text detection, etc, and makes an in-deep analysis of their challenges as well as technical improvements in recent years."}, "keywords": ["object detection literature"], "citation_intent": "method"}
{"citing_id": "2303.17111v1", "cited_id": "1804.07723", "section_title": "Classification Module", "citation": "To process the masked image, we resort to the partial convolution operator (PConv) #REFR , whose convolution kernel is renormalized to be applied only on unmasked pixels.", "text_before_citation": ["Partial Convolution.", "Unlike prior work #OTHEREFR whose ultimate goal is to localize the forgery mask, we reuse the forgery mask to help HiFi-Net learn the optimal feature for classifying fine-grained forged attributes.", "Specifically, we generate a binary maskM, then overlayM with the input image as X M to obtain the masked image", "X mask \u2208 R 3\u00d7W0\u00d7H0 ."], "text_after_citation": ["The idea is to have feature maps only describe pixels at the manipulated region.", "PConv acts as conditioned dot product for each kernel, conditioned on the mask. Denoting W par as the convolution kernel, we have:", "EQUATION", "where the dot product is \"renormalized\" to account for zeros in the mask.", "At different layers, we update and propa- gate the new maskM according to the following equation:"], "citing_paper_content": {"title": "Hierarchical Fine-Grained Image Forgery Detection And Localization", "abstract": "Differences in forgery attributes of images generated in CNN-synthesized and image-editing domains are large, and such differences make a unified image forgery detection and localization (IFDL) challenging. To this end, we present a hierarchical fine-grained formulation for IFDL representation learning. Specifically, we first represent forgery attributes of a manipulated image with multiple labels at different levels. Then we perform fine-grained classification at these levels using the hierarchical dependency between them. As a result, the algorithm is encouraged to learn both comprehensive features and inherent hierarchical nature of different forgery attributes, thereby improving the IFDL representation. Our proposed IFDL framework contains three components: multi-branch feature extractor, localization and classification modules. Each branch of the feature extractor learns to classify forgery attributes at one level, while localization and classification modules segment the pixel-level forgery region and detect imagelevel forgery, respectively. Lastly, we construct a hierarchical fine-grained dataset to facilitate our study. We demonstrate the effectiveness of our method on 7 different benchmarks, for both tasks of IFDL and forgery attribute classification. Our source code and dataset can be found: github.com/CHELSEA234/HiFi-IFDL."}, "cited_paper_content": {"title": "Image Inpainting For Irregular Holes Using Partial Convolutions", "abstract": "Existing deep learning based image inpainting methods use a standard convolutional network over the corrupted image, using convolutional filter responses conditioned on both valid pixels as well as the substitute values in the masked holes (typically the mean value). This often leads to artifacts such as color discrepancy and blurriness. Post-processing is usually used to reduce such artifacts, but are expensive and may fail. We propose the use of partial convolutions, where the convolution is masked and renormalized to be conditioned on only valid pixels. We further include a mechanism to automatically generate an updated mask for the next layer as part of the forward pass. Our model outperforms other methods for irregular masks. We show qualitative and quantitative comparisons with other methods to validate our approach."}, "keywords": ["unmasked pixels"], "citation_intent": "method"}
{"citing_id": "2303.00280v2", "cited_id": "1707.00418", "section_title": "Related Work", "citation": "The authors in #REFR propose DNN architecture for solving the multilabel classification task, which incorporates the construction of label embeddings with feature and label interdependency awareness.", "text_before_citation": ["The work #OTHEREFR examines the same problem statement of multi-label classification in an event stream as we do.", "The authors' model targets capturing temporal and probabilistic dependencies between concurrent event types by encoding historical information with a transformer and then leveraging a conditional mixture of Bernoulli experts.", "Approaches to leveraging label dependencies.", "The authors of #OTHEREFR construct a model called C-Tran for a multi-label image classification task that leverages Transformer architecture that encourages capturing the dependencies among image features and target labels.", "The key idea is to train the model with label masking."], "text_after_citation": ["A label-correlation sensitive loss improves the efficiency of the constructed model.", "Another popular way to consider label relationships is to use Graph Neural Networks as a part of the pipeline.", "Namely, #OTHEREFR captures the correlation between the labels in the task of Multi-Label Text Classification by adopting Graph Attention Network (GAT).", "They predict the final set of labels combining feature vectors from BiLSTM and attended label features from GAT.", "Event sequences processing also tries to derive dependencies between different event types and consider specific attention mechanisms #OTHEREFR ."], "citing_paper_content": {"title": "Label Attention Network For Sequential Multi-Label Classification: You Were Looking At A Wrong Self-Attention", "abstract": "Most of the available user information can be represented as a sequence of timestamped events. Each event is assigned a set of categorical labels whose future structure is of great interest. For instance, our goal is to predict a group of items in the next customer's purchase or tomorrow's client transactions. This is a multi-label classification problem for sequential data. Modern approaches focus on transformer architecture for sequential data introducing self-attention for the elements in a sequence. In that case, we take into account events' time interactions but lose information on label inter-dependencies. Motivated by this shortcoming, we propose leveraging a self-attention mechanism over labels preceding the predicted step. As our approach is a Label-Attention NETwork, we call it LANET. Experimental evidence suggests that LANET outperforms the established models' performance and greatly captures interconnections between labels. For example, the micro-AUC of our approach is 0.9536 compared to 0.7501 for a vanilla transformer. We provide an implementation of LANET to facilitate its wider usage."}, "cited_paper_content": {"title": "Learning Deep Latent Spaces For Multi-Label Classification", "abstract": "Multi-label classification is a practical yet challenging task in machine learning related fields, since it requires the prediction of more than one label category for each input instance. We propose a novel deep neural networks (DNN) based model, Canonical Correlated AutoEncoder (C2AE), for solving this task. Aiming at better relating feature and label domain data for improved classification, we uniquely perform joint feature and label embedding by deriving a deep latent space, followed by the introduction of label-correlation sensitive loss function for recovering the predicted label outputs. Our C2AE is achieved by integrating the DNN architectures of canonical correlation analysis and autoencoder, which allows end-to-end learning and prediction with the ability to exploit label dependency. Moreover, our C2AE can be easily extended to address the learning problem with missing labels. Our experiments on multiple datasets with different scales confirm the effectiveness and robustness of our proposed method, which is shown to perform favorably against state-of-the-art methods for multi-label classification."}, "keywords": ["multilabel classification task", "label embeddings"], "citation_intent": "background"}
{"citing_id": "2304.08376v1", "cited_id": "1503.09016", "section_title": "1", "citation": "We will achieve this goal by designing a method for finding a nontrivial pair of subsets having equal sum and then, like in #REFR , applying the algorithm recursively to obtain 4, 8, 16, etc. disjoint subsets with equal sum.", "text_before_citation": ["5 Zero sum subsequences in Z n p", "In this section, we assume that our input is a sequence of vectors from Z n p .", "We also assume that p is an odd prime as for p = 2 a zero sum subsequence can be obtained from n + 1 vectors in the form of a zero linear combination.", "As subsequences can be represented as subsets of the index set, it will not be too misleading to use the term (sub)set for a (sub)sequence.", "Our strategy will be finding p pairwise disjoint subsets of input vectors having equal sums."], "text_after_citation": ["Note that a pair of disjoint subsets with equal sum can be interpreted as a representation of the zero vector by a linear combination of the input vectors with nonzero coefficients 1 or \u22121 only.", "Based on this, it will be convenient to use the term signed subsets and signed subset sums.", "A signed subset of a set S of vectors is formally a function from S to the set {0, 1, \u22121}.", "The support of such a signed subset is the set of elements on which the function takes nonzero values.", "With some sloppiness, we use the term signed subset sum to refer both to the signed subset and to the value of the signed sum."], "citing_paper_content": {"title": "Zero Sum Subsequences And Hidden Subgroups", "abstract": "We propose a method for solving the hidden subgroup problem in nilpotent groups. The main idea is iteratively transforming the hidden subgroup to its images in the quotient groups by the members of a central series, eventually to its image in the commutative quotient of the original group; and then using an abelian hidden subgroup algorithm to determine this image. Knowing this image allows one to descend to a proper subgroup unless the hidden subgroup is the full group. The transformation relies on finding zero sum subsequences of sufficiently large sequences of vectors over finite prime fields. We present a new deterministic polynomial time algorithm for the latter problem in the case when the size of the field is constant. The consequence is a polynomial time exact quantum algorithm for the hidden subgroup problem in nilpotent groups having constant nilpotency class and whose order only have prime factors also bounded by a constant."}, "cited_paper_content": {"title": "On Solving Systems Of Diagonal Polynomial Equations Over Finite Fields", "abstract": "We present an algorithm to solve a system of diagonal polynomial equations over finite fields when the number of variables is greater than some fixed polynomial of the number of equations whose degree depends only on the degree of the polynomial equations. Our algorithm works in time polynomial in the number of equations and the logarithm of the size of the field, whenever the degree of the polynomial equations is constant. As a consequence we design polynomial time quantum algorithms for two algebraic hidden structure problems: for the hidden subgroup problem in certain semidirect product p-groups of constant nilpotency class, and for the multi-dimensional univariate hidden polynomial graph problem when the degree of the polynomials is constant."}, "keywords": ["subsets", "algorithm"], "citation_intent": "method"}
{"citing_id": "2303.09375v2", "cited_id": "1912.04958", "section_title": "Avatar Generation Model", "citation": "The encoder E is based on the StyleGAN2 #REFR discriminator architecture and compresses the input image I rgb to a vector v of dimension 512.", "text_before_citation": ["This RGB texture allows us to explicitly save information about high-frequency details and original colors, which are hard to preserve when mapping the whole image to a vector of limited dimensionality (as discussed below).", "We additionally apply inpainting of small gaps with averaging neighbor pixels to fill the gaps in T rgb .", "We also save the binary map of sampled pixels to the B smp and the map of the sampled and inpainted pixels to the B fill .", "The main part of the neural texture is T gen .", "It has the number of channels L = 16 and is generated using the encoder-generator architecture T gen = G(E(I rgb ))."], "text_after_citation": ["The generator G( v) has the architecture of the StyleGAN2 generator and converts the vector v into a T gen neural texture with the number of channels L = 16 as in StylePeople #OTHEREFR .", "The final neural texture used in our method has a dimension of 256 \u00d7 256 \u00d7 21 and consists of the concatenation of: the generated texture T gen (256 \u00d7 256 \u00d7 16), the texture T rgb sampled from the RGB image (256 \u00d7 256 \u00d7 3) and the two binary segmentation maps (B smp and B fill ):", "EQUATION", "We note that such an approach with the explicit use of RGB channels as part of the neural texture was originally used in #OTHEREFR .", "We use the neural renderer \u03b8(R(F UV , T )) to translate the rasterized image R(F UV , T ) with L channels into I rend output RGB image."], "citing_paper_content": {"title": "Dinar: Diffusion Inpainting Of Neural Textures For One-Shot Human Avatars", "abstract": "We present DINAR, an approach for creating realistic rigged fullbody avatars from single RGB images. Similarly to previous works, our method uses neural textures combined with the SMPL-X body model to achieve photorealistic quality of avatars while keeping them easy to animate and fast to infer. To restore the texture, we use a latent diffusion model and show how such model can be trained in the neural texture space. The use of the diffusion model allows us to realistically reconstruct large unseen regions such as the back of a person given the frontal view. The models in our pipeline are trained using 2D images and videos only. In the experiments, our approach achieves state-of-the-art rendering quality and good generalization to new poses and viewpoints. In particular, the approach improves state-of-the-art on the SnapshotPeople public benchmark."}, "cited_paper_content": {"title": "Analyzing And Improving The Image Quality Of Stylegan", "abstract": "The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality."}, "keywords": ["StyleGAN2 discriminator architecture"], "citation_intent": "method"}
{"citing_id": "2303.12383v1", "cited_id": "1106.1819", "section_title": "Figure 4: Reusing Subtrees Of D-Dnnf", "citation": "Smoothing As discussed in Section 3.2, computing the cardinalities on an unsmooth d-DNNF causes computational overhead as the algorithm needs to keep track on the set of variables during the traversal #REFR .", "text_before_citation": [], "text_after_citation": ["To ensure smoothness, we add designated nodes for each unsmooth \u2228-node as seen in Figure 3 in a preprocessing step.", "Hereby, for each missing variable v \u2208 missingVariables \u03a6 in a child subtree \u03a6 of the \u2228-node, we add a disjunction v \u2228 \u00acv resulting in the following smooth disjunction on the right side.", "\u03a6\u2208children \u03a6 \u2192 \u03a6\u2208children (\u03a6 \u2227 v\u2208missingVariables \u03a6 v \u2228 \u00acv)", "The added subformulas are tautologies as v \u2228 \u00acv \u2261 holds for every variable v.", "Hence, when adding smoothing nodes to a subformula \u03a6, the result \u03a6 \u2227 equivalent to \u03a6."], "citing_paper_content": {"title": "Exploiting D-Dnnfs For Repetitive Counting Queries On Feature Models", "abstract": "Feature models are commonly used to specify the valid configurations of a product line. In industry, feature models are often complex due to a large number of features and constraints. Thus, a multitude of automated analyses have been proposed. Many of those rely on computing the number of valid configurations which typically depends on solving a #SAT problem, a computationally expensive operation. Further, most counting-based analyses require numerous #SAT computations on the same feature model. In particular, many analyses depend on multiple computations for evaluating the number of valid configurations that include certain features or conform to partial configurations. Instead of using expensive repetitive computations on highly similar formulas, we aim to improve the performance by reusing knowledge between these computations. In this work, we are the first to propose reusing d-DNNFs for performing efficient repetitive queries on features and partial configurations. Our empirical evaluation shows that our approach is up-to 8,300 times faster (99.99% CPU-time saved) than the state of the art of repetitively invoking #SAT solvers. Applying our tool ddnnife reduces runtimes from days to minutes compared to using #SAT solvers."}, "cited_paper_content": {"title": "A Knowledge Compilation Map", "abstract": "We propose a perspective on knowledge compilation which calls for analyzing different compilation approaches according to two key dimensions: the succinctness of the target compilation language, and the class of queries and transformations that the language supports in polytime. We then provide a knowledge compilation map, which analyzes a large number of existing target compilation languages according to their succinctness and their polytime transformations and queries. We argue that such analysis is necessary for placing new compilation approaches within the context of existing ones. We also go beyond classical, flat target compilation languages based on CNF and DNF, and consider a richer, nested class based on directed acyclic graphs (such as OBDDs), which we show to include a relatively large number of target compilation languages."}, "keywords": ["computational overhead", "traversal"], "citation_intent": "background"}
{"citing_id": "2304.01008v1", "cited_id": "1804.03641", "section_title": "Matching Prediction", "citation": "Owens and Efros #REFR also utilize this pretext task, but instead take temporal video frames and audio as input.", "text_before_citation": ["EQUATION", "where the pseudo-label z i is a one-hot vector representing whether the inputs are matched.", "Matching prediction is widely used for modeling audiovisual correspondence (AVC).", "AVC was introduced by L 3 -Net #OTHEREFR , which uses a fused representation from audio and video to make a binary prediction of whether the audioimage pair is from the same video clip.", "This strategy is adopted by AVE-Net #OTHEREFR by only using Euclidean distance alignment without fusion, leading to localization of the object that sounds within an image."], "text_after_citation": ["Also, they sample video frames from the same video to construct unaligned pairs.", "This makes the pretext task more difficult and thus learns betteraligned representation and more accurate localization.", "In order to achieve better audio-visual localization and separation, researchers have proposed more complex pretext tasks.", "One such approach is the mix-and-separate method #OTHEREFR , which combines audio signals from different videos to create an input mixture with known constituent source signals.", "The network is then trained to separate the audio sources based on corresponding video frames by predicting binary spectrogram masks for each sound source."], "citing_paper_content": {"title": "Self-Supervised Multimodal Learning: A Survey", "abstract": "Multimodal learning, which aims to understand and analyze information from multiple modalities, has achieved substantial progress in the supervised regime in recent years. However, the heavy dependence on data paired with expensive human annotations impedes scaling up models. Meanwhile, given the availability of large-scale unannotated data in the wild, self-supervised learning has become an attractive strategy to alleviate the annotation bottleneck. Building on these two directions, self-supervised multimodal learning (SSML) provides ways to leverage supervision from raw multimodal data. In this survey, we provide a comprehensive review of the state-of-the-art in SSML, which we categorize along three orthogonal axes: objective functions, data alignment, and model architectures. These axes correspond to the inherent characteristics of self-supervised learning methods and multimodal data. Specifically, we classify training objectives into instance discrimination, clustering, and masked prediction categories. We also discuss multimodal input data pairing and alignment strategies during training. Finally, we review model architectures including the design of encoders, fusion modules, and decoders, which are essential components of SSML methods. We review downstream multimodal application tasks, reporting the concrete performance of the state-of-the-art image-text models and multimodal video models, and also review real-world applications of SSML algorithms in diverse fields such as healthcare, remote sensing, and machine translation. Finally, we discuss challenges and future directions for SSML. A collection of related resources can be found at: https://github.com/ys-zong/awesome-self-supervised-multimodal-learning."}, "cited_paper_content": {"title": "Audio-Visual Scene Analysis With Self-Supervised Multisensory Features", "abstract": "The thud of a bouncing ball, the onset of speech as lips open\u2014when visual and audio events occur together, it suggests that there might be a common, underlying event that produced both signals. In this paper, we argue that the visual and audio components of a video signal should be modeled jointly using a fused multisensory representation. We propose to learn such a representation in a self-supervised way, by training a neural network to predict whether video frames and audio are temporally aligned. We use this learned representation for three applications: (a) sound source localization, i.e. visualizing the source of sound in a video; (b) audio-visual action recognition; and (c) on/off-screen audio source separation, e.g. removing the off-screen translator\u2019s voice from a foreign official\u2019s speech. Code, models, and video results are available on our webpage: http://andrewowens.com/multisensory."}, "keywords": ["pretext task", "temporal video frames"], "citation_intent": "background"}
{"citing_id": "2304.09793v1", "cited_id": "1610.08336", "section_title": "Camera Poses Estimation", "citation": "We first evaluate the performance of motion-compensation methods on rotational sequences #REFR by measuring the Root Mean Square (RMS) of angular velocity errors, as presented in Tab. 3.", "text_before_citation": [], "text_after_citation": ["CMax #OTHEREFR exhibits good performance for the 3-DoF rotational motion of event cameras, with the lowest time complexity among the evaluated methods.", "DMin #OTHEREFR extends CMax to high-dimensional feature spaces and applies entropy minimization to the projected events, resulting in an improvement in the performance of approximately 20%. However, DMin is computationally expensive.", "To address this issue, approximate DMin uses a truncated kernel to balance performance and efficiency.", "ST-PPP #OTHEREFR offers an alternative solution by employing a probabilistic model, achieving the highest performance among the evaluated methods, with a 39% improvement in the shapes sequence.", "Then, we assess the performance of both deep learning and motion-compensation methods on the outdoor day 1 sequence by measuring the APRE, ARRE, and AEE metrics. As shown in Tab."], "citing_paper_content": {"title": "Event-Based Simultaneous Localization And Mapping: A Comprehensive Survey", "abstract": "In recent decades, visual simultaneous localization and mapping (vSLAM) has gained significant interest in both academia and industry. It estimates camera motion and reconstructs the environment concurrently using visual sensors on a moving robot. However, conventional cameras are limited by hardware, including motion blur and low dynamic range, which can negatively impact performance in challenging scenarios like high-speed motion and high dynamic range illumination. Recent studies have demonstrated that event cameras, a new type of bio-inspired visual sensor, offer advantages such as high temporal resolution, dynamic range, low power consumption, and low latency. This paper presents a timely and comprehensive review of event-based vSLAM algorithms that exploit the benefits of asynchronous and irregular event streams for localization and mapping tasks. The review covers the working principle of event cameras and various event representations for preprocessing event data. It also categorizes event-based vSLAM methods into four main categories: feature-based, direct, motion-compensation, and deep learning methods, with detailed discussions and practical guidance for each approach. Furthermore, the paper evaluates the state-of-the-art methods on various benchmarks, highlighting current challenges and future opportunities in this emerging research area. A public repository will be maintained to keep track of the rapid developments in this field at https://github.com/kun150kun/ESLAM-survey."}, "cited_paper_content": {"title": "The Event-Camera Dataset And Simulator: Event-Based Data For Pose Estimation, Visual Odometry, And Slam", "abstract": "New vision sensors, such as the dynamic and active-pixel vision sensor (DAVIS), incorporate a conventional global-shutter camera and an event-based sensor in the same pixel array. These sensors have great potential for high-speed robotics and computer vision because they allow us to combine the benefits of conventional cameras with those of event-based sensors: low latency, high temporal resolution, and very high dynamic range. However, new algorithms are required to exploit the sensor characteristics and cope with its unconventional output, which consists of a stream of asynchronous brightness changes (called \u201cevents\u201d) and synchronous grayscale frames. For this purpose, we present and release a collection of datasets captured with a DAVIS in a variety of synthetic and real environments, which we hope will motivate research on new algorithms for high-speed and high-dynamic-range robotics and computer-vision applications. In addition to global-shutter intensity images and asynchronous events, we provide inertial measurements and ground-truth camera poses from a motion-capture system. The latter allows comparing the pose accuracy of ego-motion estimation algorithms quantitatively. All the data are released both as standard text files and binary files (i.e. rosbag). This paper provides an overview of the available data and describes a simulator that we release open-source to create synthetic event-camera data."}, "keywords": ["motion-compensation methods"], "citation_intent": "method"}
{"citing_id": "2304.11141v1", "cited_id": "1909.06982", "section_title": "Ii. The Proposed H2Tf", "citation": "Here, L(Y, X ) denotes the fidelity term and \u03c8(Z (k) ) represents the low-rank characterization of Z (k) (which denotes the k-th frontal (spatial) slice of Z \u2208 R h\u00d7w\u00d7b #REFR ).", "text_before_citation": ["A. The t-SVD framework", "We first introduce the general formulation of t-SVD.", "Suppose that the noisy HSI Y \u2208 R h\u00d7w\u00d7b admits Y = X + N , where X denotes the clean HSI and N denotes noise.", "To infer the underlying clean HSI X from the observed Y, t-SVD method generally formulates the following model:", "EQUATION"], "text_after_citation": ["\u03c6 \u03b8 (\u2022) : R h\u00d7w\u00d7b \u2192 R h\u00d7w\u00d7b denotes a transform with learnable parameters \u03b8, which transforms the low-rank representation Z into the original domain.", "Sometimes the transform \u03c6 \u03b8 (\u2022) may not be learnable (e.g., the fixed DFT #OTHEREFR ), and in those situations the optimization variable only includes Z.", "The philosophy of the t-SVD model (1) is to minimize the rank in the transformed domain, which can model the implicit low-rankness of HSI.", "There are naturally two key building blocks for exactly modeling the implicit low-rankness, i.e., the selection of the transform \u03c6 \u03b8 (\u2022) and the exact low-rank characterization \u03c8(\u2022) of the transformed frontal slice Z (k) .", "Most t-SVD-based methods focus on the design of different transforms \u03c6 \u03b8 (\u2022) (see examples in #OTHEREFR , #OTHEREFR , #OTHEREFR ), but all of them pay less attention to the exact characterization of the transformed frontal slice."], "citing_paper_content": {"title": "H2Tf For Hyperspectral Image Denoising: Where Hierarchical Nonlinear Transform Meets Hierarchical Matrix Factorization", "abstract": "Recently, tensor singular value decomposition (t-SVD) has emerged as a promising tool for hyperspectral image (HSI) processing. In the t-SVD, there are two key building blocks: (i) the low-rank enhanced transform and (ii) the accompanying low-rank characterization of transformed frontal slices. Previous t-SVD methods mainly focus on the developments of (i), while neglecting the other important aspect, i.e., the exact characterization of transformed frontal slices. In this letter, we exploit the potentiality in both building blocks by leveraging the Hierarchical nonlinear transform and the Hierarchical matrix factorization to establish a new Tensor Factorization (termed as H2TF). Compared to shallow counter partners, e.g., low-rank matrix factorization or its convex surrogates, H2TF can better capture complex structures of transformed frontal slices due to its hierarchical modeling abilities. We then suggest the H2TF-based HSI denoising model and develop an alternating direction method of multipliers-based algorithm to address the resultant model. Extensive experiments validate the superiority of our method over state-of-the-art HSI denoising methods."}, "cited_paper_content": {"title": "Framelet Representation Of Tensor Nuclear Norm For Third-Order Tensor Completion", "abstract": "The main aim of this paper is to develop a framelet representation of the tensor nuclear norm for third-order tensor completion. In the literature, the tensor nuclear norm can be computed by using tensor singular value decomposition based on the discrete Fourier transform matrix, and tensor completion can be performed by the minimization of the tensor nuclear norm which is the relaxation of the sum of matrix ranks from all Fourier transformed matrix frontal slices. These Fourier transformed matrix frontal slices are obtained by applying the discrete Fourier transform on the tubes of the original tensor. In this paper, we propose to employ the framelet representation of each tube so that a framelet transformed tensor can be constructed. Because of framelet basis redundancy, the representation of each tube is sparsely represented. When the matrix slices of the original tensor are highly correlated, we expect the corresponding sum of matrix ranks from all framelet transformed matrix frontal slices would be small, and the resulting tensor completion can be performed much better. The proposed minimization model is convex and global minimizers can be obtained. Numerical results on several types of multi-dimensional data (videos, multispectral images, and magnetic resonance imaging data) have tested and shown that the proposed method outperformed the other testing tensor completion methods."}, "keywords": ["k-th frontal (spatial)"], "citation_intent": "background"}
{"citing_id": "2304.01823v1", "cited_id": "1511.08777", "section_title": "Finite Presentability Of Minor-Excluded Groups. A Walk In A Graph G Is A Finite Sequence Of Vertices", "citation": "The following result was proved in #REFR Theorem 25] when G is a quasitransitive locally finite planar graph.", "text_before_citation": ["(v k , v k\u22121 . . . , v 2 , v 1 ). If W = (v 1 , . . .", ", v k ) and W \u2032 = (v \u2032 1 , . . .", ", v \u2032 \u2113 ) are two walks such that v k = v \u2032 1 , then their sum is the walk W \u2022 W \u2032 := (v 1 , . . . , v k = v \u2032 1 , . . . , v \u2032 \u2113 )", ".", "We will say that a set of closed walks W generates another set of closed walks W \u2032 if every element of W \u2032 can be obtained from elements of W by adding and deleting spurs and repetitions, and performing sums, reflections and rotations."], "text_after_citation": ["We reuse some of the arguments of the proof of #OTHEREFR Proposition 22] and combine them with our structure theorem to extend the result to graphs excluding K \u221e as a minor.", "Theorem 5.3.", "Let G be a locally finite graph excluding K \u221e as a minor and let \u0393 be a group acting quasi-transitively on G.", "Then the set of closed walks of G admits a \u0393-invariant generating set with finitely many \u0393-orbits.", "Proof."], "citing_paper_content": {"title": "The Structure Of Quasi-Transitive Graphs Avoiding A Minor With Applications To The Domino Problem", "abstract": "An infinite graph is quasi-transitive if its vertex set has finitely many orbits under the action of its automorphism group. In this paper we obtain a structure theorem for locally finite quasi-transitive graphs avoiding a minor, which is reminiscent of the Robertson-Seymour Graph Minor Structure Theorem. We prove that every locally finite quasi-transitive graph G avoiding a minor has a treedecomposition whose torsos are finite or planar; moreover the tree-decomposition is canonical, i.e. invariant under the action of the automorphism group of G. As applications of this result, we prove the following. \u2022 Every locally finite quasi-transitive graph attains its Hadwiger number, that is, if such a graph contains arbitrarily large clique minors, then it contains an infinite clique minor. This extends a result of Thomassen (1992) who proved it in the 4-connected case and suggested that this assumption could be omitted. In particular, this shows that a Cayley graph excludes a finite minor if and only if it avoids the countable clique as a minor. \u2022 Locally finite quasi-transitive graphs avoiding a minor are accessible (in the sense of Thomassen and Woess), which extends known results on planar graphs to any proper minor-closed family. \u2022 Minor-excluded finitely generated groups are accessible (in the group-theoretic sense) and finitely presented, which extends classical results on planar groups. \u2022 The domino problem is decidable in a minor-excluded finitely generated group if and only if the group is virtually free, which proves the minorexcluded case of a conjecture of Ballier and Stein (2018)."}, "cited_paper_content": {"title": "Planar Transitive Graphs", "abstract": "We prove that the first homology group of every planar locally finite transitive graph $G$ is finitely generated as an $\\Aut(G)$-module and we prove a similar result for the fundamental group of locally finite planar Cayley graphs. Corollaries of these results include Droms's theorem that planar groups are finitely presented and Dunwoody's theorem that planar locally finite transitive graphs are accessible."}, "keywords": ["planar graph"], "citation_intent": "background"}
{"citing_id": "2303.07625v1", "cited_id": "1405.0312", "section_title": "Introduction", "citation": "As a consequence, researchers are forced to leverage synthetic data generated from images (e.g., #REFR ) for transformation learning in deep planar tracking, which may result in inferior performance Figure 2 .", "text_before_citation": ["In particular, several benchmarks (e.g., #OTHEREFR ) have been specially developed for evaluating and comparing different planar trackers, which greatly facilitates related research and progress on this topic.", "Despite this, these benchmarks are severely limited in further pushing the frontier of planar object tracking.", "One of the major issues with existing benchmarks is their relatively small scales.", "Especially, in the deep learning era, to unleash the potential of deep planar tracking, it is desired to have a large-scale platform. Nevertheless, as displayed in Fig.", "2 , currently all planar tracking benchmarks consist of less than 300 sequences, which is insufficient for largescale learning of deep planar tracking."], "text_after_citation": ["Summary of planar object tracking datasets, containing POT-280 #OTHEREFR , POT-210 #OTHEREFR , TMT #OTHEREFR , UCSB #OTHEREFR , Metiao #OTHEREFR , POIC #OTHEREFR , and PlanarTrack.", "The circle diameter is in proportion to the number of frames of a dataset. Our PlanarTrack is the largest among all these benchmarks. due to domain gap between different tasks.", "Besides the small-scale issue, another problem is the less challenging scenarios for planar object tracking.", "Early planar tracking datasets (e.g., #OTHEREFR ) are constructed from the indoor laboratories with simple background, which cannot reflect the diverse and complicated scenarios of real world in performance evaluation.", "To deal with this, recent datasets (e.g., #OTHEREFR ) directly collect videos in the wild."], "citing_paper_content": {"title": "Planartrack: A Large-Scale Challenging Benchmark For Planar Object Tracking", "abstract": "Planar object tracking is a critical computer vision problem and has drawn increasing interest owing to its key roles in robotics, augmented reality, etc. Despite rapid progress, its further development, especially in the deep learning era, is largely hindered due to the lack of large-scale challenging benchmarks. Addressing this, we introduce PlanarTrack, a large-scale challenging planar tracking benchmark. Specifically, PlanarTrack consists of 1,000 videos with more than 490K images. All these videos are collected in complex unconstrained scenarios from the wild, which makes Planar-Track, compared with existing benchmarks, more challenging but realistic for real-world applications. To ensure the high-quality annotation, each frame in PlanarTrack is manually labeled using four corners with multiple-round careful inspection and refinement. To our best knowledge, Planar-Track, to date, is the largest and most challenging dataset dedicated to planar object tracking. In order to analyze the proposed PlanarTrack, we evaluate 10 planar trackers and conduct comprehensive comparisons and in-depth analysis. Our results, not surprisingly, demonstrate that current topperforming planar trackers degenerate significantly on the challenging PlanarTrack and more efforts are needed to improve planar tracking in the future. In addition, we further derive a variant named PlanarTrack BB for generic object tracking from PlanarTrack. Our evaluation of 10 excellent generic trackers on PlanarTrack BB manifests that, surprisingly, PlanarTrack BB is even more challenging than several popular generic tracking benchmarks and more attention should be paid to handle such planar objects, though they are rigid. All benchmarks and evaluations will be released at the project webpage. * Equal contributions. \u2020 Corresponding author. (a) Example of generic object tracking with rectangular bounding box"}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["deep planar tracking"], "citation_intent": "background"}
{"citing_id": "2303.16043v1", "cited_id": "1912.02814", "section_title": "Other Related Work", "citation": "Similar approaches have been presented for other problems, see e.g., Bamberger et al. #REFR for results on \u2206 + 1 coloring.", "text_before_citation": ["(1) Better randomized constructions for network decomposition have been known.", "In particular, the work of Linial and Saks #OTHEREFR presented an O(log 2 n) round randomized algorithm for computing decompositions with O(log n) colors and O(log n) weak diameter.", "Elkin and Neiman #OTHEREFR imported a parallel algorithm of Miller, Peng and Xu #OTHEREFR into the distributed setting and obtained an O(log 2 n) round randomized algorithm for computing decompositions with O(log n) colors and O(log n) strong diameter.", "(2) The deterministic MIS method described in Section 1.1 for using network decompositions in computing symmetry-breaking problems such as MIS would require large messages, as it gathers the topology of each cluster in a center.", "For MIS, one can work with O(log n)-bit messages, by using a derandomization method of Censor-Hillel, Parter, and Schwartzman #OTHEREFR , and that gives a deterministic algorithm with O(log n)-bit messages and round complexity O(cd)\u2022poly(log n)."], "text_after_citation": ["(3) Ghaffari, Harris, and Kuhn #OTHEREFR and Ghaffari, Kuhn and Maus #OTHEREFR showed that one can get a general derandomization method for the LOCAL model using network decompositions.", "This method transforms any poly(log n)-round randomized algorithm for any problem whose solution can be checked in poly(log n) rounds into a deterministic algorithm with round complexity O(cd + t) \u2022 poly(log n), assuming we have a deterministic (c, d) network decomposition algorithm with round complexity t."], "citing_paper_content": {"title": "Faster Deterministic Distributed Mis And Approximate Matching", "abstract": "We present an O(log 2 n) round deterministic distributed algorithm for the maximal independent set problem. By known reductions, this round complexity extends also to maximal matching, \u2206 + 1 vertex coloring, and 2\u2206 \u2212 1 edge coloring. These four problems are among the most central problems in distributed graph algorithms and have been studied extensively for the past four decades. This improved round complexity comes closer to the \u2126(log n) lower bound of maximal independent set and maximal matching [Balliu et al. FOCS '19]. The previous best known deterministic complexity for all of these problems was \u0398(log 3 n). Via the shattering technique, the improvement permeates also to the corresponding randomized complexities, e.g., the new randomized complexity of \u2206 + 1 vertex coloring is now O(log 2 log n) rounds. Our approach is a novel combination of the previously known (and seemingly orthogonal) two methods for developing fast deterministic algorithms for these problems, namely global derandomization via network decomposition (see e.g., [Rozhon, Ghaffari STOC'20; Ghaffari, Grunau, Rozhon SODA'21; Ghaffari et al. SODA'23]) and local rounding of fractional solutions (see e.g., [Fischer DISC'17; Harris FOCS'19; Fischer, Ghaffari, Kuhn FOCS'17; Ghaffari, Kuhn FOCS'21; Faour et al. SODA'23]). We consider a relaxation of the classic network decomposition concept, where instead of requiring the clusters in the same block to be non-adjacent, we allow each node to have a small number of neighboring clusters. We also show a deterministic algorithm that computes this relaxed decomposition faster than standard decompositions. We then use this relaxed decomposition to significantly improve the integrality of certain fractional solutions, before handing them to the local rounding procedure that now has to do fewer rounding steps. Randomized algorithms, and the pursuit of deterministic algorithms. In the 1980s, Luby [Lub86] and Alon, Babai, and Itai [ABI86] presented a simple and elegant randomized distributed algorithm that computes an MIS in O(log n) rounds, with high probability 1. Due to known reductions, this MIS algorithm led to O(log n) round randomized algorithms for many other key graph problems, including maximal matching, \u2206 + 1 vertex coloring, and (2\u2206 \u2212 1) edge coloring. These problems are often listed as the four fundamental symmetry-breaking problems in distributed graph algorithms and have a wide range of applications. The O(log n)-round randomized algorithm naturally led the researchers to seek a deterministic distributed algorithm with the same round complexity. In his celebrated work [Lin87, Lin92], Linial asked \"can it [MIS] always be found in polylogarithmic time [deterministically]? \" He even added that \"getting a deterministic polylog-time algorithm for MIS seems hard.\" Since then, this became known as Linial's MIS question and turned into one of the research foci in distributed graph algorithms."}, "cited_paper_content": {"title": "Efficient Deterministic Distributed Coloring With Small Bandwidth", "abstract": "We show that the $(degree+1)$-list coloring problem can be solved deterministically in $O(D \\cdot \\log n \\cdot\\log^2\\Delta)$ rounds in the \\CONGEST model, where $D$ is the diameter of the graph, $n$ the number of nodes, and $\\Delta$ the maximum degree. Using the recent polylogarithmic-time deterministic network decomposition algorithm by Rozho\\v{n} and Ghaffari [STOC 2020], this implies the first efficient (i.e., $\\poly\\log n$-time) deterministic \\CONGEST algorithm for the $(\\Delta+1)$-coloring and the $(\\mathit{degree}+1)$-list coloring problem. Previously the best known algorithm required $2^{O(\\sqrt{\\log n})}$ rounds and was not based on network decompositions. Our techniques also lead to deterministic $(\\mathit{degree}+1)$-list coloring algorithms for the congested clique and the massively parallel computation (MPC) model. For the congested clique, we obtain an algorithm with time complexity $O(\\log\\Delta\\cdot\\log\\log\\Delta)$, for the MPC model, we obtain algorithms with round complexity $O(\\log^2\\Delta)$ for the linear-memory regime and $O(\\log^2\\Delta + \\log n)$ for the sublinear memory regime."}, "keywords": ["problems"], "citation_intent": "method"}
{"citing_id": "2304.12666v1", "cited_id": "1911.04252", "section_title": "Effect Of Pretrained Weight", "citation": "In light of our observations that utilizing pretrained weights for both student and teacher networks leads to a improved performance, we note that the Noisy Student #REFR reported no improvement in performance when initializing both models simultaneously.", "text_before_citation": [], "text_after_citation": ["They claim that this approach may sometimes lead to getting stuck in a local optima, resulting in inferior performance compared to training the student model from scratch.", "We hypothesize that this contrasting result could be due to the fact that they used the same pretrained weight for both teacher and student.", "Since the student model has already inherited all the knowledge that could be learned from the teacher, distilling knowledge with the teacher may not bring additional performance gains.", "A recent study by Allen-Zhu and Li #OTHEREFR revealed that different networks trained with distinct random seeds learn different knowledge.", "Motivated by this observation, we train VGG-16 on CIFAR-100 with 8 different seeds following the same procedure as the baseline in Section 4.1 and utilize them to initialize teacher and student networks."], "citing_paper_content": {"title": "Bayesian Optimization Meets Self-Distillation", "abstract": "Bayesian optimization (BO) has contributed greatly to improving model performance by suggesting promising hyperparameter configurations iteratively based on observations from multiple training trials. However, only partial knowledge (i.e., the measured performances of trained models and their hyperparameter configurations) from previous trials is transferred. On the other hand, Self-Distillation (SD) only transfers partial knowledge learned by the task model itself. To fully leverage the various knowledge gained from all training trials, we propose the BOSS framework, which combines BO and SD. BOSS suggests promising hyperparameter configurations through BO and carefully selects pre-trained models from previous trials for SD, which are otherwise abandoned in the conventional BO process. BOSS achieves significantly better performance than both BO and SD in a wide range of tasks including general image classification, learning with noisy labels, semi-supervised learning, and medical image analysis tasks."}, "cited_paper_content": {"title": "Self-Training With Noisy Student Improves Imagenet Classification", "abstract": "We present a simple self-training method that achieves 88.4% top-1 accuracy on ImageNet, which is 2.0% better than the state-of-the-art model that requires 3.5B weakly labeled Instagram images. On robustness test sets, it improves ImageNet-A top-1 accuracy from 61.0% to 83.7%, reduces ImageNet-C mean corruption error from 45.7 to 28.3, and reduces ImageNet-P mean flip rate from 27.8 to 12.2. ::: To achieve this result, we first train an EfficientNet model on labeled ImageNet images and use it as a teacher to generate pseudo labels on 300M unlabeled images. We then train a larger EfficientNet as a student model on the combination of labeled and pseudo labeled images. We iterate this process by putting back the student as the teacher. During the generation of the pseudo labels, the teacher is not noised so that the pseudo labels are as accurate as possible. However, during the learning of the student, we inject noise such as dropout, stochastic depth and data augmentation via RandAugment to the student so that the student generalizes better than the teacher."}, "keywords": ["pretrained weights", "teacher networks"], "citation_intent": "result"}
{"citing_id": "2303.01515v1", "cited_id": "1704.00447", "section_title": "Related Work", "citation": "For pMRI: Variational network (VN) #REFR introduced gradient descent method by applying given sensitivities {s i }.", "text_before_citation": ["Most existing deep-learning (DL) based methods rendering end-to-end neural networks mapping from the partial k-space data to the reconstructed images #OTHEREFR .", "The common issue with this class of methods is that the DNNs require excessive amount of data to train, and the resulting networks perform similar to \"black-boxes\" which are difficult to interpret and modify.", "In recent years, a class of DL based methods improve over the end-to-end training by selecting the scheme of an iterative optimization algorithm and prescribe a phase number T , map each iteration of the scheme to one phase of the network.", "These methods are often known as the learned optimization algorithms (LOAs), #OTHEREFR .", "For instance, ADMM-Net #OTHEREFR , ISTA-Net + #OTHEREFR , and cascade network #OTHEREFR are regular MRI reconstruction."], "text_after_citation": ["MoDL #OTHEREFR proposed a recursive network by unrolling the conjugate gradient algorithm using a weight sharing strategy.", "Blind-PMRI-Net #OTHEREFR designed three network blocks to alternately update multi-channel images, sensitivity maps and the reconstructed MR image using an iterative algorithm based on half-quadratic splitting.", "The network in #OTHEREFR developed a Bayesian framework for joint MRI-PET reconstruction. VS-Net #OTHEREFR derived a variable splitting optimization method.", "However, existing methods still face the lack of accurate coil sensitivity maps and proper regularization in the pMRI problem.", "Recently, a method called DeepcomplexMRI #OTHEREFR developed an end-to-end learning without explicitly using coil sensitivity maps to recover channel-wise images, and then combine to a single channel image in testing."], "citing_paper_content": {"title": "Optimization-Based Deep Learning Methods For Magnetic Resonance Imaging Reconstruction And Synthesis", "abstract": ""}, "cited_paper_content": {"title": "Learning A Variational Network For Reconstruction Of Accelerated Mri Data", "abstract": "Purpose: To allow fast and high-quality reconstruction of clinical accelerated multi-coil MR data by learning a variational network that combines the mathematical structure of variational models with deep learning. Theory and Methods: Generalized compressed sensing reconstruction formulated as a variational model is embedded in an unrolled gradient descent scheme. All parameters of this formulation, including the prior model defined by filter kernels and activation functions as well as the data term weights, are learned during an offline training procedure. The learned model can then be applied online to previously unseen data. Results: The variational network approach is evaluated on a clinical knee imaging protocol. The variational network reconstructions outperform standard reconstruction algorithms in terms of image quality and residual artifacts for all tested acceleration factors and sampling patterns. Conclusion: Variational network reconstructions preserve the natural appearance of MR images as well as pathologies that were not included in the training data set. Due to its high computational performance, i.e., reconstruction time of 193 ms on a single graphics card, and the omission of parameter tuning once the network is trained, this new approach to image reconstruction can easily be integrated into clinical workflow."}, "keywords": ["Variational network"], "citation_intent": "method"}
{"citing_id": "2303.15495v1", "cited_id": "1912.02574", "section_title": "Ii. Related Works", "citation": "In this paper, they presented a stochastic optimization toolchain along with sensitivity analyses for choosing the optimal hyperparameters, and they solved the optimization problem by using a single-objective optimization task as well as a greedy algorithm, a genetic algorithm (GA), and a particle swarm optimization (PSO) algorithm #REFR .", "text_before_citation": ["Dubey #OTHEREFR introduced a public transportation decision support system for short-term and long-term prediction of arrival bus times.", "This study used the real-world historical data of two Nashville bus system routes.", "The approach of this research combined the clustering analysis and Kalman filters with a shared route segment model to produce more accurate arrival time predictions and, based on their results, compared to the basic arrival time prediction model that Nashville MTA was using, their system reduced arrival time prediction errors by 25% on average when predicting the arrival delay an hour ahead and 47% when predicting within a 15-minute future time window #OTHEREFR . S. Basak, F. Sun, S. Sengupta, and A.", "Dubey have conducted a similar study #OTHEREFR , using unsupervised clustering mechanisms to optimize transit on-time performance.", "As a local case study, they analyzed the monthly and seasonal delays of the Nashville metro region and clustered months with similar patterns."], "text_after_citation": ["According to the newest research in #OTHEREFR , dynamic datadriven application systems (DDDAS) that use real-time sensors and a data-driven decision support system can provide online model learning and multi-time-scale analytics to enhance the intelligence of the system.", "As part of their study, the authors analyzed an online bus arrival prediction system in Nashville using historical and real-time streaming data, which can be packaged as modular, distributed, and resilient micro-services.", "The long-term delay analysis service excludes noise from outliers in historical data to identify delay patterns associated with different hours, days, and seasons for specific time points and route segments.", "City planners can use the feedback data generated by these analytics services to improve bus schedules and increase rider satisfaction #OTHEREFR .", "In addition, another study by S. Nannapaneni and A."], "citing_paper_content": {"title": "A Novel Neural Network Approach For Predicting The Arrival Time Of Buses For Smart On-Demand Public Transit", "abstract": "Among the major public transportation systems in cities, bus transit has its problems, including more accuracy and reliability when estimating the bus arrival time for riders. This can lead to delays and decreased ridership, especially in cities where public transportation is heavily relied upon. A common issue is that the arrival times of buses do not match the schedules, resulting in latency for fixed schedules. According to the study in this paper on New York City bus data, there is an average delay of around eight minutes or 491 seconds mismatch between the bus arrivals and the actual scheduled time. This research paper presents a novel AI-based data-driven approach for estimating the arrival times of buses at each transit point (station). Our approach is based on a fully connected neural network and can predict the arrival time collectively across all bus lines in large metropolitan areas. Our neural-net data-driven approach provides a new way to estimate the arrival time of the buses, which can lead to a more efficient and smarter way to bring the bus transit to the general public. Our evaluation of the network bus system with more than 200 bus lines, and 2 million data points, demonstrates less than 40 seconds of estimated error for arrival times. The inference time per each validation set data point is less than 0.006 ms."}, "cited_paper_content": {"title": "Data-Driven Optimization Of Public Transit Schedule", "abstract": "Bus transit systems are the backbone of public transportation in the United States. An important indicator of the quality of service in such infrastructures is on-time performance at stops, with published transit schedules playing an integral role governing the level of success of the service. However there are relatively few optimization architectures leveraging stochastic search that focus on optimizing bus timetables with the objective of maximizing probability of bus arrivals at timepoints with delays within desired on-time ranges. In addition to this, there is a lack of substantial research considering monthly and seasonal variations of delay patterns integrated with such optimization strategies. To address these, this paper makes the following contributions to the corpus of studies on transit on-time performance optimization: (a) an unsupervised clustering mechanism is presented which groups months with similar seasonal delay patterns, (b) the problem is formulated as a single-objective optimization task and a greedy algorithm, a genetic algorithm (GA) as well as a particle swarm optimization (PSO) algorithm are employed to solve it, (c) a detailed discussion on empirical results comparing the algorithms are provided and sensitivity analysis on hyper-parameters of the heuristics are presented along with execution times, which will help practitioners looking at similar problems. The analyses conducted are insightful in the local context of improving public transit scheduling in the Nashville metro region as well as informative from a global perspective as an elaborate case study which builds upon the growing corpus of empirical studies using nature-inspired approaches to transit schedule optimization."}, "keywords": ["optimal hyperparameters", "particle swarm optimization"], "citation_intent": "method"}
{"citing_id": "2303.03443v1", "cited_id": "1810.01969", "section_title": "Forward Algorithm.", "citation": "The details of how it is implemented can be found under Algorithm A.1 of #REFR , but the main idea is to successively compute the distribution of the underlying state of H after each sample via Bayesian updates.", "text_before_citation": ["The Forward Algorithm is prevalent in the preprocessing, compression and decompression algorithms.", "Therefore, we provide an overview of what the algorithm does and how it works in this subsection.", "Given a Markov source H with \u2113 states, let the first j samples be Y 1 , . . . , Y j \u223c H n . For a setting y 1 , . . .", ", y j\u22121 , the Forward Algorithm computes the distribution of Y n given that (Y 1 , . . . , Y j\u22121 ) = (y 1 , . . . , y j\u22121 )."], "text_after_citation": ["Each update takes time O(\u2113 2 ) (since there are roughly this many possible transitions to consider), and we must do j of these updates, for a total runtime of O(j\u2113 2 ). This is implemented using dynamic programming."], "citing_paper_content": {"title": "Improving The Runtime Of Algorithmic Polarization Of Hidden Markov Models", "abstract": "We improve the runtime of the linear compression scheme for hidden Markov sources presented in a 2018 paper of Guruswami, Nakkiran, and Sudan. Under the previous scheme, compressing a message of length n takes O(n log n) runtime, and decompressing takes O(n 1+\u03b4) runtime for any fixed \u03b4 > 0. We present how to improve the runtime of the decoding scheme to O(n log n) by caching intermediate results to avoid repeating computation."}, "cited_paper_content": {"title": "Algorithmic Polarization For Hidden Markov Models", "abstract": "Using a mild variant of polar codes we design linear compression schemes compressing Hidden Markov sources (where the source is a Markov chain, but whose state is not necessarily observable from its output), and to decode from Hidden Markov channels (where the channel has a state and the error introduced depends on the state). We give the first polynomial time algorithms that manage to compress and decompress (or encode and decode) at input lengths that are polynomial $\\it{both}$ in the gap to capacity and the mixing time of the Markov chain. Prior work achieved capacity only asymptotically in the limit of large lengths, and polynomial bounds were not available with respect to either the gap to capacity or mixing time. Our results operate in the setting where the source (or the channel) is $\\it{known}$. If the source is $\\it{unknown}$ then compression at such short lengths would lead to effective algorithms for learning parity with noise -- thus our results are the first to suggest a separation between the complexity of the problem when the source is known versus when it is unknown."}, "keywords": ["Algorithm A.1"], "citation_intent": "method"}
{"citing_id": "2304.03706v1", "cited_id": "2004.02554", "section_title": "Related Work", "citation": "For the special case of additive valuations with binary marginals, the algorithm of Aziz #REFR improves upon the guarantees of Aleksandrov et al.", "text_before_citation": ["#OTHEREFR gave a randomized polynomial-time algorithm that outputs an EF1 allocation, while being envy-free ex-ante.", "Ensuing work by Aziz #OTHEREFR showed that there exists a similar randomized algorithm that additionally implements the well-known Probabilistic Serial fractional outcome described by Bogomolnaia and Moulin #OTHEREFR .", "This randomized algorithm also preserves a weak notion of efficiency. Babaioff et al.", "#OTHEREFR also study the case of additive valuations and find a distribution over ex-post proportional up to one item and #OTHEREFR 2 -MMS allocations that is ex-ante proportional, i.e., the expected value of each agent's bundle is at least a 1 n -fraction of her value for the set of all items.", "Best-of-both-worlds results have also been analyzed for other settings, including the case where the agents have binary marginals, and the case of additive valuations with arbitrary entitlements."], "text_after_citation": ["#OTHEREFR and is group-strategyproof, ex-ante fractionally-PO and envy-free, and ex-post fractionally PO and EF1.", "In a similar vein, for additive valuations with binary marginals, Halpern et al.", "#OTHEREFR independently showed that there is a distribution over expost Nash-welfare-maximizing allocations that also ex-ante maximizes the fractional Nash welfare, implying the same fairness guarantees as Aziz #OTHEREFR for this setting. For matroid rank valuations, Babaioff et al.", "#OTHEREFR present a randomized truthful mechanism that is ex-ante envy-free and ex-post Lorenz dominating (and thus ex-post Nash-welfare-maximizing and EFX for this specific class).", "For the case of agents with arbitrary entitlements, both Hoefer et al. #OTHEREFR and Aziz et al."], "citing_paper_content": {"title": "Breaking The Envy Cycle: Best-Of-Both-Worlds Guarantees For Subadditive Valuations", "abstract": "We study best-of-both-worlds guarantees for the fair division of indivisible items among agents with subadditive valuations. Our main result establishes the existence of a random allocation that is simultaneously ex-ante 1 2-envy-free, ex-post 1 2-EFX and ex-post EF1, for every instance with subadditive valuations. We achieve this result by a novel polynomial-time algorithm that randomizes the well-established envy cycles procedure in a way that provides ex-ante fairness. Notably, this is the first best-of-both-worlds fairness guarantee for subadditive valuations, even when considering only EF1 without EFX."}, "cited_paper_content": {"title": "Simultaneously Achieving Ex-Ante And Ex-Post Fairness", "abstract": "We present a polynomial-time algorithm that computes an ex-ante envy-free lottery over envy-free up to one item (EF1) deterministic allocations. It has the following advantages over a recently proposed algorithm: it does not rely on the linear programming machinery including separation oracles; it is SD-efficient (both ex-ante and ex-post); and the ex-ante outcome is equivalent to the outcome returned by the well-known probabilistic serial rule. As a result, we answer a question raised by Freeman, Shah, and Vaish (2020) whether the outcome of the probabilistic serial rule can be implemented by ex-post EF1 allocations. In the light of a couple of impossibility results that we prove, our algorithm can be viewed as satisfying a maximal set of properties. Under binary utilities, our algorithm is also ex-ante group-strategyproof and ex-ante Pareto optimal. Finally, we also show that checking whether a given random allocation can be implemented by a lottery over EF1 and Pareto optimal allocations is NP-hard."}, "keywords": ["additive valuations"], "citation_intent": "method"}
{"citing_id": "2305.01622v1", "cited_id": "1906.09788", "section_title": "C. Online Path Generation", "citation": "Note that #REFR was designed for trajectory optimization, and we slightly reformulate it to conduct path smoothing.", "text_before_citation": ["The output of Algorithm 2 consists of paths originating", "; 2 L \u2190 \u2205; 3 foreach G z of G do 4 D \u2190 \u2205; 5 l init \u2190 InitialGuessSearch(G z ) ; 6 {s i |i = 1, ..., N } \u2190 StationSampling(l init ); 7 for i = 1, ..., N do 8 D i \u2190 LateralClustering(s i ); 9 D \u2190 D \u222a D i 10 end 11 L z \u2190 DPSearch(D, {s i |i = 1, ..., N }, G z ); 12 L z \u2190 NonMaximumSuppress(L z ); 13 L \u2190 L \u222a L z ; 14 end 15 return L;", "from each entry lane.", "Due to the discretization of the grid, the path is not smooth enough for control.", "To this end, we utilize a local path smoothing #OTHEREFR based on Quadratic Programming (QP) to smooth the path."], "text_after_citation": ["An example of the path smoothing process is given in Fig. 4d ."], "citing_paper_content": {"title": "Flowmap: Path Generation For Automated Vehicles In Open Space Using Traffic Flow", "abstract": "There is extensive literature on perceiving road structures by fusing various sensor inputs such as lidar point clouds and camera images using deep neural nets. Leveraging the latest advance of neural architects (such as transformers) and bird-eye-view (BEV) representation, the road cognition accuracy keeps improving. However, how to cognize the \"road\" for automated vehicles where there is no well-defined \"roads\" remains an open problem. For example, how to find paths inside intersections without HD maps is hard since there is neither an explicit definition for \"roads\" nor explicit features such as lane markings. The idea of this paper comes from a proverb: it becomes a way when people walk on it. Although there are no \"roads\" from sensor readings, there are \"roads\" from tracks of other vehicles. In this paper, we propose FlowMap, a path generation framework for automated vehicles based on traffic flows. FlowMap is built by extending our previous work RoadMap [1], a lightweight semantic map, with an additional traffic flow layer. A path generation algorithm on traffic flow fields (TFFs) is proposed to generate human-like paths. The proposed framework is validated using real-world driving data and is amenable to generating paths for super complicated intersections without using HD maps."}, "cited_paper_content": {"title": "Safe Trajectory Generation For Complex Urban Environments Using Spatio-Temporal Semantic Corridor", "abstract": "Planning safe trajectories for autonomous vehicles in complex urban environments is challenging since there are numerous semantic elements (such as dynamic agents, traffic lights, and speed limits) to consider. These semantic elements may have different mathematical descriptions, such as obstacle, constraint, and cost. It is non-trivial to tune the effects from different combinations of semantic elements for a stable and generalizable behavior. In this letter, we propose a novel unified spatio-temporal semantic corridor (SSC) structure, which provides a level of abstraction for different types of semantic elements. The SSC consists of a series of mutually connected collision-free cubes with dynamical constraints posed by the semantic elements in the spatio-temporal domain. The trajectory generation problem then boils down to a general quadratic programming formulation. Thanks to the unified SSC representation, our framework can generalize to any combination of semantic elements. Moreover, our formulation provides a theoretical guarantee that the entire trajectory is safe and constraint-satisfied, by using the convex hull and hodograph properties of piecewise Bezier curve parameterization. We also release the code of our method to accommodate benchmarking."}, "keywords": ["path smoothing", "trajectory optimization"], "citation_intent": "method"}
{"citing_id": "2304.06537v1", "cited_id": "1706.04599", "section_title": "Introduction", "citation": "A model is called perfect calibrated if the predictive confidence of the model represents a good approximation of its actual probability of correctness #REFR .", "text_before_citation": ["With the development of deep neural networks, great progress has been made in image classification.", "In addition to performance, the uncertainty estimate of a given model is also receiving increasing attention, as the confidence of a model is expected to accurately reflect its performance. \u2020 Corresponding author."], "text_after_citation": ["Model calibration is particularly important in safety-critical applications, such as autonomous driving, medical diagnosis, and robotics #OTHEREFR .", "For example, if a prediction with low confidence is more likely to be wrong, we can take countermeasures to avoid unknown risks.", "Most existing calibration techniques assume that the distribution of training data is balanced, i.e., each class has a similar number of training instances, so that each class is treated equally #OTHEREFR . As shown in Fig.", "1 , the traditional calibration pipeline uses a balanced training set to train the classification model and a balanced validation set to obtain the calibration model, respectively.", "The target test set is in the same distribution as the training/validation set."], "citing_paper_content": {"title": "Transfer Knowledge From Head To Tail: Uncertainty Calibration Under Long-Tailed Distribution", "abstract": "How to estimate the uncertainty of a given model is a crucial problem. Current calibration techniques treat different classes equally and thus implicitly assume that the distribution of training data is balanced, but ignore the fact that real-world data often follows a long-tailed distribution. In this paper, we explore the problem of calibrating the model trained from a long-tailed distribution. Due to the difference between the imbalanced training distribution and balanced test distribution, existing calibration methods such as temperature scaling can not generalize well to this problem. Specific calibration methods for domain adaptation are also not applicable because they rely on unlabeled target domain instances which are not available. Models trained from a long-tailed distribution tend to be more overconfident to head classes. To this end, we propose a novel knowledge-transferring-based calibration method by estimating the importance weights for samples of tail classes to realize long-tailed calibration. Our method models the distribution of each class as a Gaussian distribution and views the source statistics of head classes as a prior to calibrate the target distributions of tail classes. We adaptively transfer knowledge from head classes to get the target probability density of tail classes. The importance weight is estimated by the ratio of the target probability density over the source probability density. Extensive experiments on CIFAR-10-LT, MNIST-LT, CIFAR-100-LT, and ImageNet-LT datasets demonstrate the effectiveness of our method."}, "cited_paper_content": {"title": "On Calibration Of Modern Neural Networks", "abstract": "Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions."}, "keywords": ["predictive confidence"], "citation_intent": "background"}
{"citing_id": "2303.05234v1", "cited_id": "1811.06186", "section_title": "Related Work 2.1. Gait Recognition", "citation": "Silhouette-based methods GaitSet #REFR aggregates temporal information in silhouette sequences using a statistical function to adapt to different frame rates.", "text_before_citation": ["Gait recognition methods can be classified into two main categories: appearance-based #OTHEREFR and modelbased methods #OTHEREFR depending on the type of data input.", "Appearance-based methods usually rely on silhouette sequences #OTHEREFR or their transformation, such as Gait Energy Image (GEI) #OTHEREFR and Chrono-Gait Image (CGI) #OTHEREFR .", "Model-based methods, on the other hand, represent the human body as mesh #OTHEREFR or a set of keypoints #OTHEREFR ."], "text_after_citation": ["GaitPart #OTHEREFR uses Focal Convolutional Layer to extract fine-grained local features and Micro-motion Template Builder with different window sizes to extract local temporal information.", "GaitGL #OTHEREFR uses 3D Convolution to extract global and local spatiotemporal information.", "LagrangeGait #OTHEREFR adds a local motion extractor and a viewpoint branch based on GaitGL to get more discriminative local temporal information.", "Pose-based methods PoseGait #OTHEREFR employs 3D pose information to generate multi-feature vectors and uses a CNN to extract the gait information in both spatial and temporal dimensions.", "GaitGraph #OTHEREFR and GaitGraph2 #OTHEREFR adopt Graph Convolutional Network for gait recognition, treating keypoints as nodes and limbs as edges to form a topology graph."], "citing_paper_content": {"title": "Gpgait: Generalized Pose-Based Gait Recognition", "abstract": "Recent works on pose-based gait recognition have demonstrated the potential of using such simple information to achieve results comparable to silhouette-based methods. However, the generalization ability of pose-based methods on different datasets is undesirably inferior to that of silhouette-based ones, which has received little attention but hinders the application of these methods in real-world scenarios. To improve the generalization ability of pose-based methods across datasets, we propose a Generalized Pose-based Gait recognition (GPGait) framework. First, a Human-Oriented Transformation (HOT) and a series of Human-Oriented Descriptors (HOD) are proposed to obtain a unified pose representation with discriminative multi-features. Then, given the slight variations in the unified representation after HOT and HOD, it becomes crucial for the network to extract local-global relationships between the keypoints. To this end, a Part-Aware Graph Convolutional Network (PAGCN) is proposed to enable efficient graph partition and local-global spatial feature extraction. Experiments on four public gait recognition datasets, CASIA-B, OUMVLP-Pose, Gait3D and GREW, show that our model demonstrates better and more stable cross-domain capabilities compared to existing skeletonbased methods, achieving comparable recognition results to silhouette-based ones. The code will be released."}, "cited_paper_content": {"title": "Gaitset: Regarding Gait As A Set For Cross-View Gait Recognition", "abstract": "As a unique biometric feature that can be recognized at a distance, gait has broad applications in crime prevention, forensic identification and social security. To portray a gait, existing gait recognition methods utilize either a gait template, where temporal information is hard to preserve, or a gait sequence, which must keep unnecessary sequential constraints and thus loses the flexibility of gait recognition. In this paper we present a novel perspective, where a gait is regarded as a set consisting of independent frames. We propose a new network named GaitSet to learn identity information from the set. Based on the set perspective, our method is immune to permutation of frames, and can naturally integrate frames from different videos which have been filmed under different scenarios, such as diverse viewing angles, different clothes/carrying conditions. Experiments show that under normal walking conditions, our single-model method achieves an average rank-1 accuracy of 95.0% on the CASIA-B gait dataset and an 87.1% accuracy on the OU-MVLP gait dataset. These results represent new state-of-the-art recognition accuracy. On various complex scenarios, our model exhibits a significant level of robustness. It achieves accuracies of 87.2% and 70.4% on CASIA-B under bag-carrying and coat-wearing walking conditions, respectively. These outperform the existing best methods by a large margin. The method presented can also achieve a satisfactory accuracy with a small number of frames in a test sample, e.g., 82.5% on CASIA-B with only 7 frames. The source code has been released at https://github.com/AbnerHqC/GaitSet."}, "keywords": ["GaitSet"], "citation_intent": "method"}
{"citing_id": "2303.04244v1", "cited_id": "1911.12409", "section_title": "Data Normalization", "citation": "Processing to remove location and orientation variability has been used in prior work #REFR , and to save space we present our version of normalization in the Appendix.", "text_before_citation": ["Before training, temporal windows of body pose data are normalized by centering, rotating and scale normalizing the data values.", "Centering and rotation use a body-specific procedure to make the 3D point data invariant to the body's location and facing direction in the capture space."], "text_after_citation": [], "citing_paper_content": {"title": "A Light-Weight Contrastive Approach For Aligning Human Pose Sequences", "abstract": "We present a simple unsupervised method for learning an encoder mapping short 3D pose sequences into embedding vectors suitable for sequence-to-sequence alignment by dynamic time warping. Training samples consist of temporal windows of frames containing 3D body points such as mocap markers or skeleton joints. A lightweight , 3layer encoder is trained using a contrastive loss function that encourages embedding vectors of augmented sample pairs to have cosine similarity 1, and similarity 0 with all other samples in a minibatch. When multiple scripted training sequences are available, temporal alignments inferred from an initial round of training are harvested to extract additional, cross-performance match pairs for a second phase of training to refine the encoder. In addition to being simple, the proposed method is fast to train, making it easy to adapt to new data using different marker sets or skeletal joint layouts. Experimental results illustrate ease of use, transferability, and utility of the learned embeddings for comparing and analyzing human behavior sequences."}, "cited_paper_content": {"title": "Predict&Cluster: Unsupervised Skeleton Based Action Recognition", "abstract": "We propose a novel system for unsupervised skeleton-based action recognition. Given inputs of body keypoints sequences obtained during various movements, our system associates the sequences with actions. Our system is based on an encoder-decoder recurrent neural network, where the encoder learns a separable feature representation within its hidden states formed by training the model to perform prediction task. We show that according to such unsupervised training the decoder and the encoder self-organize their hidden states into a feature space which clusters similar movements into the same cluster and distinct movements into distant clusters. Current state-of-the-art methods for action recognition are strongly supervised, i.e., rely on providing labels for training. Unsupervised methods have been proposed, however, they require camera and depth inputs (RGB+D) at each time step. In contrast, our system is fully unsupervised, does not require labels of actions at any stage, and can operate with body keypoints input only. Furthermore, the method can perform on various dimensions of body keypoints (2D or 3D) and include additional cues describing movements. We evaluate our system on three extensive action recognition benchmarks with different number of actions and examples. Our results outperform prior unsupervised skeleton-based methods, unsupervised RGB+D based methods on cross-view tests and while being unsupervised have similar performance to supervised skeleton-based action recognition."}, "keywords": ["orientation variability", "normalization"], "citation_intent": "method"}
{"citing_id": "2304.02309v1", "cited_id": "1803.07100", "section_title": "Efficient Learning", "citation": "While these results are not state-of-the-art (100% reported by #REFR ), this test proves that our model can cope with a larger dataset.", "text_before_citation": ["To remove this limitation, we perform a second set of tests where we choose for each class an optimized template stimulus (indicated by Optimized).", "This stimulus is determined by iterating through the training dataset and retaining the class-specific training picture that results in the highest training accuracy after a single training epoch.", "Table 4 shows the accuracy of the classifications for the different training schemes.", "Using the first images in the dataset for each class (condition First), we obtain a testing accuracy of 73%.", "This was in the range of the accuracies for training with individual avatars, which range from 66% to 79%."], "text_after_citation": ["When the model have access to every domain (MD-NRE-I), the test accuracy reaches a value of 92.15%, which exceeds the accuracy (89.02%) reported in the original paper on the FERG dataset #OTHEREFR .", "Moreover, the best transfer learning model (MD-NRE-Bonnie) reaches 84.42% accuracy.", "This is an encouraging transfer learning result for a classifier trained with only 12 images, and where all the expressions are from a single avatar type."], "citing_paper_content": {"title": "Multi-Domain Norm-Referenced Encoding Enables Data Efficient Transfer Learning Of Facial Expression Recognition *", "abstract": "People can innately recognize human facial expressions in unnatural forms, such as when depicted on the unusual faces drawn in cartoons or when applied to an animal's features. However, current machine learning algorithms struggle with out-of-domain transfer in facial expression recognition (FER). We propose a biologically-inspired mechanism for such transfer learning, which is based on norm-referenced encoding, where patterns are encoded in terms of difference vectors relative to a domain-specific reference vector. By incorporating domain-specific reference frames, we demonstrate high data efficiency in transfer learning across multiple domains. Our proposed architecture provides an explanation for how the human brain might innately recognize facial expressions on varying head shapes (humans, monkeys, and cartoon avatars) without extensive training. Norm-referenced encoding also allows the intensity of the expression to be read out directly from neural unit activity, similar to face-selective neurons in the brain. Our model achieves a classification accuracy of 92.15% on the FERG dataset with extreme data efficiency. We train our proposed mechanism with only 12 images, including a single image of each class (facial expression) and one image per domain (avatar). In comparison, the authors of the FERG dataset achieved a classification accuracy of 89.02% with their FaceExpr model, which was trained on 43,000 images."}, "cited_paper_content": {"title": "Vgan-Based Image Representation Learning For Privacy-Preserving Facial Expression Recognition", "abstract": "Reliable facial expression recognition plays a critical role in human-machine interactions. However, most of the facial expression analysis methodologies proposed to date pay little or no attention to the protection of a user's privacy. In this paper, we propose a Privacy-Preserving Representation-Learning Variational Generative Adversarial Network (PPRL-VGAN) to learn an image representation that is explicitly disentangled from the identity information. At the same time, this representation is discriminative from the standpoint of facial expression recognition and generative as it allows expression-equivalent face image synthesis. We evaluate the proposed model on two public datasets under various threat scenarios. Quantitative and qualitative results demonstrate that our approach strikes a balance between the preservation of privacy and data utility. We further demonstrate that our model can be effectively applied to other tasks such as expression morphing and image completion."}, "keywords": ["model"], "citation_intent": "result"}
{"citing_id": "2304.12486v2", "cited_id": "1904.04433", "section_title": "Conclusion And Future Works", "citation": "It seems to be the case with other blackbox attacks, since the perturbation is generated randomly using probabilistic laws that don't take into account the fact that document images are brighter than other images #REFR .", "text_before_citation": ["Compressing and then decompressing input images of a model using JPEG protocol improves the model robustness against adversarial images, but not consistently.", "On the other hand, the adversarial training of both models using the method of Kurakin et al. #OTHEREFR , strongly improves robustness of both models.", "This training method is quite easy to implement and does not affect a lot the test accuracy of models on our classification task.", "Therefore, this method seems far more effective on a document classification task than against an image classification task, as we can see in Dong et al. #OTHEREFR .", "The black-box attack we evaluated generates blurry examples that appear darker than legitimate document images."], "text_after_citation": ["There are many ways to improve the robustness of a model that would only use the visual modality of a document #OTHEREFR .", "However, state-of-the-art approaches to document classification take advantage of other information modalities, such as the layout of the document, and the text it contains 3 .", "Therefore, after this work on evaluating the robustness of visual models, it would be interesting to evaluate the transferability of the generated examples to a multimodal model such as DocFormer or LayoutLMv2, which use optical character recognition (OCR) and transformer layers.", "Furthermore, we could explore the possibility of designing adversarial attacks to which these models are more sensitive, for example by targeting OCR prediction errors #OTHEREFR that affect the textual modality and may also affect the robustness of such models #OTHEREFR .", "Dealing with the added modality given by text means that more approaches can be explored attacking only one modality or both."], "citing_paper_content": {"title": "Evaluating Adversarial Robustness On Document Image Classification", "abstract": "Adversarial attacks and defenses have gained increasing interest on computer vision systems in recent years, but as of today, most investigations are limited to natural images. However, many artificial intelligence models actually handle documentary data, which is very different from real world images. Hence, in this work, we try to apply the adversarial attack philosophy on documentary data and to protect models against such attacks. Our methodology is to implement untargeted gradient-based, transfer-based and score-based attacks and evaluate the impact of defenses such as adversarial training, JPEG input compression and grey-scale input transformation on the robustness of ResNet50 and EfficientNetB0 model architectures. To the best of our knowledge, no such work has been conducted by the community in order to study the impact of these attacks on the document image classification task."}, "cited_paper_content": {"title": "Efficient Decision-Based Black-Box Adversarial Attacks On Face Recognition", "abstract": "Face recognition has obtained remarkable progress in recent years due to the great improvement of deep convolutional neural networks (CNNs). However, deep CNNs are vulnerable to adversarial examples, which can cause fateful consequences in real-world face recognition applications with security-sensitive purposes. Adversarial attacks are widely studied as they can identify the vulnerability of the models before they are deployed. In this paper, we evaluate the robustness of state-of-the-art face recognition models in the decision-based black-box attack setting, where the attackers have no access to the model parameters and gradients, but can only acquire hard-label predictions by sending queries to the target model. This attack setting is more practical in real-world face recognition systems. To improve the efficiency of previous methods, we propose an evolutionary attack algorithm, which can model the local geometry of the search directions and reduce the dimension of the search space. Extensive experiments demonstrate the effectiveness of the proposed method that induces a minimum perturbation to an input face image with fewer queries. We also apply the proposed method to attack a real-world face recognition system successfully."}, "keywords": ["document images", "blackbox attacks"], "citation_intent": "background"}
{"citing_id": "2305.01732v1", "cited_id": "1812.11941", "section_title": "Related Work", "citation": "Depth from Single Image Convolutional networks within an encoder-decoder paradigm #REFR is the standard prototype architecture for dense depth prediction from a single image.", "text_before_citation": ["Synthetic dataset generation from computer games #OTHEREFR has been used extensively for training computer vision algorithms like semantic segmentation #OTHEREFR , object detection #OTHEREFR , and depth estimation #OTHEREFR . Mohammad et al.", "#OTHEREFR have also used the GTA-V game to generate a synthetic RGB-D dataset but with relative depth at the focus.", "The resolution of the dataset used for training is 256\u00d7256 which is much smaller than the publicly available datasets.", "Furthermore, they need a preprocessing phase, such as histogram equalization, to use the datasets before feeding them to the training.", "For training, they use a Resnet architecture and process the RGB image and GT depth with resolution 256 \u00d7 256."], "text_after_citation": ["The building blocks of such a network consist of convolutional and sub-sampling as their core elements.", "However, CNN as an encoder suffers from a local receptive field problem #OTHEREFR , leading to less global representation learning at higher resolutions.", "Several algorithms adapt different techniques to learn features at different resolutions to address this issue like dilated convolutions #OTHEREFR or parallel multi-scale feature aggression #OTHEREFR .", "Recently, transformer architectures such as vision transformer (ViT) #OTHEREFR or data-efficient image transformers (DeiT) #OTHEREFR have outperformed CNN architectures in image recognition #OTHEREFR , object detection #OTHEREFR , and semantic segmentation #OTHEREFR .", "Inspired by the success of transformers in various topics, Ren\u00e9 et al."], "citing_paper_content": {"title": "High-Resolution Synthetic Rgb-D Datasets For Monocular Depth Estimation", "abstract": "Accurate depth maps are essential in various applications, such as autonomous driving, scene reconstruction, point-cloud creation, etc. However, monocular-depth estimation (MDE) algorithms often fail to provide enough texture & sharpness, and also are inconsistent for homogeneous scenes. These algorithms mostly use CNN or vision transformer-based architectures requiring large datasets for supervised training. But, MDE algorithms trained on available depth datasets do not generalize well and hence fail to perform accurately in diverse real-world scenes. Moreover, the ground-truth depth maps are either lower resolution or sparse leading to relatively inconsistent depth maps. In general, acquiring a high-resolution ground truth dataset with pixel-level precision for accurate depth prediction is an expensive, and time-consuming challenge. In this paper, we generate a high-resolution synthetic depth dataset (HRSD) of dimension 1920 \u00d7 1080 from Grand Theft Auto (GTA-V), which contains 100,000 color images and corresponding dense ground truth depth maps. The generated datasets are diverse and have scenes from indoors to outdoors, from homogeneous surfaces to textures. For experiments and analysis, we train the DPT algorithm, a state-of-the-art transformer-based MDE algorithm on the proposed synthetic dataset, which significantly increases the accuracy of depth maps on different scenes by 9%. Since the synthetic datasets are of higher resolution, we propose adding a feature extraction module in the transformer's encoder and incorporating an attention-based loss, further improving the accuracy by 15 %."}, "cited_paper_content": {"title": "High Quality Monocular Depth Estimation Via Transfer Learning", "abstract": "Accurate depth estimation from images is a fundamental task in many applications including scene understanding and reconstruction. Existing solutions for depth estimation often produce blurry approximations of low resolution. This paper presents a convolutional neural network for computing a high-resolution depth map given a single RGB image with the help of transfer learning. Following a standard encoder-decoder architecture, we leverage features extracted using high performing pre-trained networks when initializing our encoder along with augmentation and training strategies that lead to more accurate results. We show how, even for a very simple decoder, our method is able to achieve detailed high-resolution depth maps. Our network, with fewer parameters and training iterations, outperforms state-of-the-art on two datasets and also produces qualitatively better results that capture object boundaries more faithfully. Code and corresponding pre-trained weights are made publicly available."}, "keywords": ["dense depth prediction"], "citation_intent": "background"}
{"citing_id": "2303.13582v1", "cited_id": "1804.08328", "section_title": "A.1. Experiments On Tanks And Temples", "citation": "As shown in Table A5 , SCADE trained with the same outof-domain prior that we used for the other datasets (which was trained on Taskonomy #REFR ) outperforms the baselines on the Tanks and Temples dataset as well. Moreover, Figure A11 shows qualitative results.", "text_before_citation": ["We conduct further experiments to test the robustness of SCADE.", "We evaluate on three scenes from the Tanks and Temples #OTHEREFR dataset, namely three large indoor rooms -Church, Courtroom and Auditorium scenes.", "The training set consists of 21, 26 and 21 sparse views for the Church, Courtroom and Auditorium scenes respectively, and the test set consists of 8 sparse views, so the amount of data is similar to that used in prior work #OTHEREFR .", "We also followed similar data preprocessing steps as prior work #OTHEREFR and ran SfM #OTHEREFR on all images to obtain camera poses for training."], "text_after_citation": ["As shown, SCADE is able to recover objects better than the baselines such as the table in the Church, the group of chairs in the Courtroom (second column), and the rows of seats in the Auditorium (clearer in the side-view seats on the second column).", "Moreover, results also show that SCADE avoids clouds of dust such as the lights on the wall of the Church (second column), painting on the wall of the Courtroom (last column) and details on the repetitive seats of the auditorium."], "citing_paper_content": {"title": "Scade: Nerfs From Space Carving With Ambiguity-Aware Depth Estimates", "abstract": "Figure 1. SCADE Overview. We present SCADE, a novel technique for NeRF reconstruction under sparse, unconstrained views for in-the-wild indoor scenes. We leverage on generalizable monocular depth priors and address to represent the inherent ambiguities of monocular depth by exploiting our ambiguity-aware depth estimates (left). Our approach accounts for multimodality of both distributions using our novel space carving loss that seeks to disambiguate and find the common mode to fuse the information between different views (middle). SCADE enables better photometric reconstruction especially in highly ambiguous scenes, e.g. non-opaque surfaces (right)."}, "cited_paper_content": {"title": "Taskonomy: Disentangling Task Transfer Learning", "abstract": "Do visual tasks have a relationship, or are they unrelated? For instance, could having surface normals simplify estimating the depth of an image? Intuition answers these questions positively, implying existence of a structure among visual tasks. Knowing this structure has notable values; it is the concept underlying transfer learning and provides a principled way for identifying redundancies across tasks, e.g., to seamlessly reuse supervision among related tasks or solve many tasks in one system without piling up the complexity. We proposes a fully computational approach for modeling the structure of space of visual tasks. This is done via finding (first and higher-order) transfer learning dependencies across a dictionary of twenty six 2D, 2.5D, 3D, and semantic tasks in a latent space. The product is a computational taxonomic map for task transfer learning. We study the consequences of this structure, e.g. nontrivial emerged relationships, and exploit them to reduce the demand for labeled data. For example, we show that the total number of labeled datapoints needed for solving a set of 10 tasks can be reduced by roughly 2/3 (compared to training independently) while keeping the performance nearly the same. We provide a set of tools for computing and probing this taxonomical structure including a solver that users can employ to devise efficient supervision policies for their use cases."}, "keywords": ["datasets", "Taskonomy"], "citation_intent": "result"}
{"citing_id": "2303.14572v1", "cited_id": "1407.6756", "section_title": "Application: Improved 3Sum In Preprocessed Universes", "citation": "Kopelowitz, Pettie and Porat #REFR gave a simple randomized reduction from 3SUM to O(log n) instances of 3SUM-Convolution via hashing.", "text_before_citation": ["As one immediate application, we can solve 3SUM with preprocessed universe, improving Chan and Lewenstein's previous solution which required O(n 13/7 ) query time #OTHEREFR , and also improving Corollary 5.3 regardless of the value of \u03c9: the query algorithm does not use fast matrix multiplication but uses FFT instead, though randomization is now needed in the preprocessing algorithm.", "Corollary 10.4.", "We can preprocess sets A, B, and C of n integers in O(n 2 ) Las Vegas randomized time, so that given any subsets A \u2032 \u2286 A, B \u2032 \u2286 B, and C \u2032 \u2286 C, we can solve All-Nums-3SUM on (A \u2032 , B \u2032 , C \u2032 ) in O(n 11/6 ) time.", "Proof."], "text_after_citation": ["The same approach works in the preprocessed universe setting, and transform the input into O(log n) instances where A, B, and C are indexed sets.", "During preprocessing, we apply Theorem 10.3 to A \u222a (\u2212B), producing subsets A (\u03bb) and a set R of pairs.", "During a query with given subsets A \u2032 \u2286 A, B \u2032 \u2286 B, and C \u2032 \u2286 C, we first examine each pair (a, \u2212b) \u2208 R and check whether a \u2208 A \u2032 , b \u2208 B \u2032 , and a + b \u2208 C \u2032 . This takes O(n 2 /s) time.", "Next, for each \u03bb, we compute (A (\u03bb) \u2229 A \u2032 ) + ((\u2212A (\u03bb) ) \u2229 B \u2032 ) by known FFT-based algorithms #OTHEREFR ; the running time is near-linear in the output size, which is bounded by |A (\u03bb) \u2212 A (\u03bb) |.", "For each output value c, we check whether c \u2208 C \u2032 ."], "citing_paper_content": {"title": "Fredman'S Trick Meets Dominance Product: Fine-Grained Complexity Of Unweighted Apsp, 3Sum Counting, And More", "abstract": "In this paper we carefully combine Fredman's trick [SICOMP'76] and Matou\u0161ek's approach for dominance product [IPL'91] to obtain powerful results in fine-grained complexity: \u2022 Under the hypothesis that APSP for undirected graphs with edge weights in {1, 2,. .. , n} requires n 3\u2212o(1) time (when \u03c9 = 2), we show a variety of conditional lower bounds, including an n 7/3\u2212o(1) lower bound for unweighted directed APSP and an n 2.2\u2212o(1) lower bound for computing the Minimum Witness Product between two n \u00d7 n Boolean matrices, even if \u03c9 = 2, improving upon their trivial n 2 lower bounds. Our techniques can also be used to reduce the unweighted directed APSP problem to other problems. In particular, we show that (when \u03c9 = 2), if unweighted directed APSP requires n 2.5\u2212o(1) time, then Minimum Witness Product requires n 7/3\u2212o(1) time."}, "cited_paper_content": {"title": "Higher Lower Bounds From The 3Sum Conjecture", "abstract": "The 3SUM conjecture has proven to be a valuable tool for proving conditional lower bounds on dynamic data structures and graph problems. This line of work was initiated by P\\v{a}tra\\c{s}cu (STOC 2010) who reduced 3SUM to an offline SetDisjointness problem. However, the reduction introduced by P\\v{a}tra\\c{s}cu suffers from several inefficiencies, making it difficult to obtain tight conditional lower bounds from the 3SUM conjecture. ::: In this paper we address many of the deficiencies of P\\v{a}tra\\c{s}cu's framework. We give new and efficient reductions from 3SUM to offline SetDisjointness and offline SetIntersection (the reporting version of SetDisjointness) which leads to polynomially higher lower bounds on several problems. Using our reductions, we are able to show the essential optimality of several algorithms, assuming the 3SUM conjecture. ::: - Chiba and Nishizeki's $O(m\\alpha)$-time algorithm (SICOMP 1985) for enumerating all triangles in a graph with arboricity/degeneracy $\\alpha$ is essentially optimal, for any $\\alpha$. ::: - Bj{\\o}rklund, Pagh, Williams, and Zwick's algorithm (ICALP 2014) for listing $t$ triangles is essentially optimal (assuming the matrix multiplication exponent is $\\omega=2$). ::: - Any static data structure for SetDisjointness that answers queries in constant time must spend $\\Omega(N^{2-o(1)})$ time in preprocessing, where $N$ is the size of the set system. ::: These statements were unattainable via P\\v{a}tra\\c{s}cu's reductions. ::: We also introduce several new reductions from 3SUM to pattern matching problems and dynamic graph problems. Of particular interest are new conditional lower bounds for dynamic versions of Maximum Cardinality Matching, which introduce a new technique for obtaining amortized lower bounds."}, "keywords": ["O(log n) instances"], "citation_intent": "background"}
{"citing_id": "2304.11400v1", "cited_id": "1704.02422", "section_title": "Results On Single-Coil Mri Reconstruction", "citation": "It is worth noting that even compared to the second best model DCCNN #REFR , PSNR is improved by 0.41dB. This effectively illustrates the excellence of EAMRI.", "text_before_citation": ["In Table 1 , we show the quantitative results for all model reconstructions on the Calgary single-coil brain dataset.", "According to the table, we can observe that EAMRI achieves the best results for the quantification metrics under acceleration factor 4 and requires fewer model parameters (123K)."], "text_after_citation": ["Moreover, in Fig.", "6 , we provide a visual comparison of the reconstruction results of these models.", "We can see that EAMRI has fewer bright spots in the heatmaps, which means less error between the EAMRI reconstructed image and the ground truth image.", "Meanwhile, according to the zoomed-in images of the selected areas, we can observe that our EAMRI can reconstruct more clean and accurate edges. This further validates the validity of EAMRI.", "Both the quantitative and the qualitative results for the single-coil MRI reconstruction demonstrate the effectiveness of EAMRI."], "citing_paper_content": {"title": "Fast Mri Reconstruction Via Edge Attention", "abstract": "Fast and accurate MRI reconstruction is a key concern in modern clinical practice. Recently, numerous Deep-Learning methods have been proposed for MRI reconstruction, however, they usually fail to reconstruct sharp details from the subsampled k-space data. To solve this problem, we propose a lightweight and accurate Edge Attention MRI Reconstruction Network (EAMRI) to reconstruct images with edge guidance. Specifically, we design an efficient Edge Prediction Network to directly predict accurate edges from the blurred image. Meanwhile, we propose a novel Edge Attention Module (EAM) to guide the image reconstruction utilizing the extracted edge priors, as inspired by the popular self-attention mechanism. EAM first projects the input image and edges into Q image , K edge , and V image , respectively. Then EAM pairs the Q image with K edge along the channel dimension, such that 1) it can search globally for the high-frequency image features that are activated by the edge priors; 2) the overall computation burdens are largely reduced compared with the traditional spatial-wise attention. With the help of EAM, the predicted edge priors can effectively guide the model to reconstruct high-quality MR images with accurate edges. Extensive experiments show that our proposed EAMRI outperforms other methods with fewer parameters and can recover more accurate edges."}, "cited_paper_content": {"title": "A Deep Cascade Of Convolutional Neural Networks For Dynamic Mr Image Reconstruction", "abstract": "Inspired by recent advances in deep learning, we propose a framework for reconstructing dynamic sequences of 2-D cardiac magnetic resonance (MR) images from undersampled data using a deep cascade of convolutional neural networks (CNNs) to accelerate the data acquisition process. In particular, we address the case where data are acquired using aggressive Cartesian undersampling. First, we show that when each 2-D image frame is reconstructed independently, the proposed method outperforms state-of-the-art 2-D compressed sensing approaches, such as dictionary learning-based MR image reconstruction, in terms of reconstruction error and reconstruction speed. Second, when reconstructing the frames of the sequences jointly, we demonstrate that CNNs can learn spatio-temporal correlations efficiently by combining convolution and data sharing approaches. We show that the proposed method consistently outperforms state-of-the-art methods and is capable of preserving anatomical structure more faithfully up to 11-fold undersampling. Moreover, reconstruction is very fast: each complete dynamic sequence can be reconstructed in less than 10 s and, for the 2-D case, each image frame can be reconstructed in 23 ms, enabling real-time applications."}, "keywords": ["second best model", "DCCNN"], "citation_intent": "result"}
{"citing_id": "2304.00858v1", "cited_id": "1905.00875", "section_title": "Self-Supervised Action Representation Learning", "citation": "To learn the contextual coherence in action representation, Lai and Xie #REFR matched pixelwise correspondence from the spatial-temporal color information. Han et al.", "text_before_citation": ["#OTHEREFR first proposed a convolutional auto-encoder to construct the latent space from the encoder output.", "By operating the high-level features, the model is functional in many areas, such as action interpolation and comparison.", "Later on, with a hierarchical RNN auto-encoder in both spatial and temporal domains, Wang et al.", "#OTHEREFR developed a high-quality representation space that is motivated for precise action modeling.", "As a vision-based learning task, the effectiveness of self-supervised representation is frequently explored in RGB-based action understanding."], "text_after_citation": ["#OTHEREFR exploited action representations from multiple modalities of RGB streams and optical flow.", "However, these RGB-based action recognition models usually learn contrastive representation from background or visual consistency #OTHEREFR .", "With the visual information unavailable, the challenge of learning self-supervised skeletonbased action representation mainly comes from the diverse pose information under different view observations #OTHEREFR .", "In self-supervised skeleton-based action representation learning, existing works such as #OTHEREFR mainly focus on preserving the action-dependent features as much as possible to identify samples #OTHEREFR .", "For example, an adversarial discriminator is used to assist the auto-encoder to rectify the reconstructed action for a more distinctive representation #OTHEREFR ."], "citing_paper_content": {"title": "Focalized Contrastive View-Invariant Learning For Self-Supervised Skeleton-Based Action Recognition", "abstract": "Learning view-invariant representation is a key to improving feature discrimination power for skeleton-based action recognition. Existing approaches cannot effectively remove the impact of viewpoint due to the implicit view-dependent representations. In this work, we propose a self-supervised framework called Focalized Contrastive View-invariant Learning (FoCoViL), which significantly suppresses the view-specific information on the representation space where the viewpoints are coarsely aligned. By maximizing mutual information with an effective contrastive loss between multi-view sample pairs, FoCoViL associates actions with common view-invariant properties and simultaneously separates the dissimilar ones. We further propose an adaptive focalization method based on pairwise similarity to enhance contrastive learning for a clearer cluster boundary in the learned space. Different from many existing self-supervised representation learning work that rely heavily on supervised classifiers, FoCoViL performs well on both unsupervised and supervised classifiers with superior recognition performance. Extensive experiments also show that the proposed contrastive-based focalization generates a more discriminative latent representation."}, "cited_paper_content": {"title": "Self-Supervised Learning For Video Correspondence Flow", "abstract": "The objective of this paper is self-supervised learning of feature embeddings that are suitable for matching correspondences along the videos, which we term correspondence flow. By leveraging the natural spatial-temporal coherence in videos, we propose to train a ``pointer'' that reconstructs a target frame by copying pixels from a reference frame. ::: We make the following contributions: First, we introduce a simple information bottleneck that forces the model to learn robust features for correspondence matching, and prevent it from learning trivial solutions, \\eg matching based on low-level colour information. Second, to tackle the challenges from tracker drifting, due to complex object deformations, illumination changes and occlusions, we propose to train a recursive model over long temporal windows with scheduled sampling and cycle consistency. Third, we achieve state-of-the-art performance on DAVIS 2017 video segmentation and JHMDB keypoint tracking tasks, outperforming all previous self-supervised learning approaches by a significant margin. Fourth, in order to shed light on the potential of self-supervised learning on the task of video correspondence flow, we probe the upper bound by training on additional data, \\ie more diverse videos, further demonstrating significant improvements on video segmentation."}, "keywords": ["action representation", "pixelwise correspondence"], "citation_intent": "background"}
{"citing_id": "2304.04726v1", "cited_id": "1803.05407", "section_title": "Representing Model Uncertainty", "citation": "Intuitively, ensembles and checkpoint averages also reflect the idea of different views and interpretations of the data and, therefore, provide a framework for uncertainty modeling. Stochastic Weight Averaging (SWA, #REFR ) and SWA-Gaussian (SWAG, Maddox et al. (2019)) both build on this idea.", "text_before_citation": ["The approach to uncertainty modeling that we consider is related to the well-established technique of model ensembling.", "Stochastic optimization procedures applied in training deep neural networks are non-deterministic and depend on hyper-parameters and initial seeds.", "Ensembles have been used as a pragmatic solution to average over several solutions, and the positive impact on model performance pushed ensembling into the standard toolbox of deep learning.", "Related to ensembling is the technique of checkpoint averaging (refer to e.g.", "#OTHEREFR , which is also known to improve performance."], "text_after_citation": ["SWA proposes using the first moments of the parameters of the solutions traversed by the optimizer during the optimization process, as mean estimates of the model parameters.", "Using such mean values have been argued to result in finding wider optima, providing better generalization to unseen data.", "On top of these mean estimations procured by SWA, SWAG then adds a low-rank plus diagonal approximation of covariances, which, when combined with the aforementioned mean estimations, provide us with corresponding Gaussian posterior approximations over model parameters.", "Posterior distribution approximations learned this way then represent our epistemic uncertainty about the model #OTHEREFR , meaning the uncertainty stemming from not knowing the perfect values of the model parameters, since we do not have infinite data to train on.", "During test time, instead of making estimates from a single model with deterministic parameters, we sample N different models from the approximated posteriors for each model parameter, and use the average of their prediction distributions as the model response."], "citing_paper_content": {"title": "Uncertainty-Aware Natural Language Inference With Stochastic Weight Averaging", "abstract": "This paper introduces Bayesian uncertainty modeling using Stochastic Weight Averaging-Gaussian (SWAG) in Natural Language Understanding (NLU) tasks. We apply the approach to standard tasks in natural language inference (NLI) and demonstrate the effectiveness of the method in terms of prediction accuracy and correlation with human annotation disagreements. We argue that the uncertainty representations in SWAG better reflect subjective interpretation and the natural variation that is also present in human language understanding. The results reveal the importance of uncertainty modeling, an often neglected aspect of neural language modeling, in NLU tasks."}, "cited_paper_content": {"title": "Averaging Weights Leads To Wider Optima And Better Generalization", "abstract": "Deep neural networks are typically trained by optimizing a loss function with an SGD variant, in conjunction with a decaying learning rate, until convergence. We show that simple averaging of multiple points along the trajectory of SGD, with a cyclical or constant learning rate, leads to better generalization than conventional training. We also show that this Stochastic Weight Averaging (SWA) procedure finds much broader optima than SGD, and approximates the recent Fast Geometric Ensembling (FGE) approach with a single model. Using SWA we achieve notable improvement in test accuracy over conventional SGD training on a range of state-of-the-art residual networks, PyramidNets, DenseNets, and Shake-Shake networks on CIFAR-10, CIFAR-100, and ImageNet. In short, SWA is extremely easy to implement, improves generalization, and has almost no computational overhead."}, "keywords": ["uncertainty modeling", "Stochastic Weight Averaging"], "citation_intent": "background"}
{"citing_id": "2303.09472v1", "cited_id": "1911.07783", "section_title": "B. Evaluation On Real-World Sr", "citation": "We evaluate all methods on the dataset provided in the challenge of Real-World Super-Resolution: NTIRE2020 Track1 and Tracks #REFR .", "text_before_citation": ["Specifically, we adopt the same loss functions of Real-ESRGAN #OTHEREFR , which further introduce perceptual loss and adversarial loss to the basic L 1 loss.", "We set the learning rate of the KDSR T to 2 \u00d7 10 \u22124 .", "We further validate the effectiveness of KDSR on Real-World datasets.", "For optimization, we use Adam with \u03b2 1 = 0.9, \u03b2 2 = 0.99.", "In both two stages of training, we set the batch size to 64, with the input patch size being 64."], "text_after_citation": ["In addition, we also validate our DiffIR on RealSRSet #OTHEREFR .", "Since NTIRE2020 Track1 and RealSRSet datasets provide a paired validation set, we use the LPIPS #OTHEREFR , DISTS #OTHEREFR , and PSNR for the evaluation.", "The quantitative results are shown in Tab. 6.", "We can see that DiffIR S2 outperforms SOTA real-world SR method KDSR S -GAN on LPIPS, DISTS, and PSNR, consuming fewer computational costs.", "In addition, we can see that DiffIR S2 outperforms classic real-world SR method Real-ESRGAN on LPIPS, DISTS, and PSNR, only consuming its 63% Mult-Adds."], "citing_paper_content": {"title": "Diffir: Efficient Diffusion Model For Image Restoration", "abstract": "Diffusion model (DM) has achieved SOTA performance by modeling the image synthesis process into a sequential application of a denoising network. However, different from image synthesis generating each pixel from scratch, most pixels of image restoration (IR) are given. Thus, for IR, traditional DMs running massive iterations on a large model to estimate whole images or feature maps is inefficient. To address this issue, we propose an efficient DM for IR (Dif-fIR), which consists of a compact IR prior extraction network (CPEN), dynamic IR transformer (DIRformer), and denoising network. Specifically, DiffIR has two training stages: pretraining and training DM. In pretraining, we input ground-truth images into CPEN S1 to capture a compact IR prior representation (IPR) to guide DIRformer. In the second stage, we train the DM to directly estimate the same IRP as pretrained CPEN S1 only using LQ images. We observe that since the IPR is only a compact vector, DiffIR can use fewer iterations than traditional DM to obtain accurate estimations and generate more stable and realistic results. Since the iterations are few, our DiffIR can adopt a joint optimization of CPEN S2 , DIRformer, and denoising network, which can further reduce the estimation error influence. We conduct extensive experiments on several IR tasks and achieve SOTA performance while consuming less computational costs."}, "cited_paper_content": {"title": "Aim 2019 Challenge On Real-World Image Super-Resolution: Methods And Results", "abstract": "This paper reviews the AIM 2019 challenge on real world super-resolution. It focuses on the participating methods and final results. The challenge addresses the real world setting, where paired true high and low-resolution images are unavailable. For training, only one set of source input images is therefore provided in the challenge. In Track 1: Source Domain the aim is to super-resolve such images while preserving the low level image characteristics of the source input domain. In Track 2: Target Domain a set of high-quality images is also provided for training, that defines the output domain and desired quality of the super-resolved images. To allow for quantitative evaluation, the source input images in both tracks are constructed using artificial, but realistic, image degradations. The challenge is the first of its kind, aiming to advance the state-of-the-art and provide a standard benchmark for this newly emerging task. In total 7 teams competed in the final testing phase, demonstrating new and innovative solutions to the problem."}, "keywords": ["dataset", "Resolution"], "citation_intent": "method"}
{"citing_id": "2304.01242v1", "cited_id": "2003.01332", "section_title": "Ii. Related Work", "citation": "HGT #REFR uses node-type and edge-type dependent parameters to characterize the heterogeneous attention over each edge based on metarelation, which achieves the state-of-the-art in the mission of knowledge embedding.", "text_before_citation": ["Knowledge graph embedding aims to project entities and relations in the knowledge graph into a low-dimensional space.", "Graph Neural Networks (GNNs), such as GCN #OTHEREFR and GAT #OTHEREFR , use information aggregation from nearby nodes to learn graph representations. Wang et al.", "#OTHEREFR proposed AM-GCN to improve the embedding capability of GCN.", "This is because GCN may struggle to learn complex correlation information between node features and topological structures. However, in real-world applications, knowledge graphs can be heterogeneous.", "RGCN #OTHEREFR is used as the graph embedding model in MedRec #OTHEREFR which achieve the state-of-theart performance in drug recommendation."], "text_after_citation": ["Current studies on knowledge-based recommendation suffer a lot from the problem of graph sparsity, which severely reduces the effectiveness of identifying valuable items for users.", "Meanwhile, heterogeneous graph neural network becomes popular for its capacity to incorporate both structured and unstructured information.", "But the use of heterogeneous network in recommendation systems has not been extensively investigated."], "citing_paper_content": {"title": "Enhancing Clinical Evidence Recommendation With Multi-Channel Heterogeneous Learning On Evidence Graphs", "abstract": "Clinical evidence encompasses the associations and impacts between patients, interventions (such as drugs or physiotherapy), problems, and outcomes. The goal of recommending clinical evidence is to provide medical practitioners with relevant information to support their decision-making processes and to generate new evidence. Our specific task focuses on recommending evidence based on clinical problems. However, the direct connections between certain clinical problems and related evidence are often sparse, creating a challenge of link sparsity. Additionally, to recommend appropriate evidence, it is essential to jointly exploit both topological relationships among evidence and textual information describing them. To address these challenges, we define two knowledge graphs: an Evidence Co-reference Graph and an Evidence Text Graph, to represent the topological and linguistic relations among evidential elements, respectively. We also introduce a multi-channel heterogeneous learning model and a fusional attention mechanism to handle the coreference-text heterogeneity in evidence recommendation. Our experiments demonstrate that our model outperforms state-of-the-art methods on open data."}, "cited_paper_content": {"title": "Heterogeneous Graph Transformer", "abstract": "Recent years have witnessed the emerging success of graph neural networks (GNNs) for modeling structured data. However, most GNNs are designed for homogeneous graphs, in which all nodes and edges belong to the same types, making them infeasible to represent heterogeneous structures. In this paper, we present the Heterogeneous Graph Transformer (HGT) architecture for modeling Web-scale heterogeneous graphs. To model heterogeneity, we design node- and edge-type dependent parameters to characterize the heterogeneous attention over each edge, empowering HGT to maintain dedicated representations for different types of nodes and edges. To handle dynamic heterogeneous graphs, we introduce the relative temporal encoding technique into HGT, which is able to capture the dynamic structural dependency with arbitrary durations. To handle Web-scale graph data, we design the heterogeneous mini-batch graph sampling algorithm---HGSampling---for efficient and scalable training. Extensive experiments on the Open Academic Graph of 179 million nodes and 2 billion edges show that the proposed HGT model consistently outperforms all the state-of-the-art GNN baselines by 9%--21% on various downstream tasks."}, "keywords": ["heterogeneous attention"], "citation_intent": "method"}
{"citing_id": "2303.16202v1", "cited_id": "1811.10541", "section_title": "Related Work", "citation": "HiPPI #REFR is a computationally efficient method that takes geometric relations into account while generalising permutation synchronization but is still limited in resolution.", "text_before_citation": ["When more than two shapes of the same class exist, stronger geometric cues can be leveraged to improve results by matching all of them simultaneously.", "Unfortunately, the already very high problem complexity increases even further the more shapes are used.", "Hence, existing multi-shape matching methods limit the total number of shapes and their resolution #OTHEREFR , work in spectral space #OTHEREFR , or relax the permutation constraints #OTHEREFR .", "Early multi-matching methods computed pair-wise matchings and subsequently used permutation synchronisation to establish cycle consistency #OTHEREFR .", "Still, permutation synchronisation requires the eigendecomposition of a matrix with quadratically increasing dimensions #OTHEREFR ."], "text_after_citation": ["Instead of looking at permutations directly, ZoomOut #OTHEREFR reduces the dimensionality of the problem by projecting it onto the spectral decomposition.", "This idea has been extended to take cycle consistency within the spectral space into account #OTHEREFR , which does not guarantee a point-wise consistent matching.", "To circumvent this issue, IsoMuSh #OTHEREFR jointly optimises point and functional correspondences.", "The method detangles the optimisation into smaller subproblems by using a so-called universe shape that all shapes are mapped to instead of each other, as Cao and Bernard do #OTHEREFR .", "Using a universe is similar to requiring a template shape, as many learning-based approaches do #OTHEREFR : Both synchronise all correspondences by matching them through a unified space."], "citing_paper_content": {"title": "Ccuantumm: Cycle-Consistent Quantum-Hybrid Matching Of Multiple Shapes", "abstract": "Jointly matching multiple, non-rigidly deformed 3D shapes is a challenging, N P-hard problem. A perfect matching is necessarily cycle-consistent: Following the pairwise point correspondences along several shapes must end up at the starting vertex of the original shape. Unfortunately, existing quantum shape-matching methods do not support multiple shapes and even less cycle consistency. This paper addresses the open challenges and introduces the first quantum-hybrid approach for 3D shape multimatching; in addition, it is also cycle-consistent. Its iterative formulation is admissible to modern adiabatic quantum hardware and scales linearly with the total number of input shapes. Both these characteristics are achieved by reducing the N-shape case to a sequence of three-shape matchings, the derivation of which is our main technical contribution. Thanks to quantum annealing, high-quality solutions with low energy are retrieved for the intermediate N Phard objectives. On benchmark datasets, the proposed approach significantly outperforms extensions to multi-shape matching of a previous quantum-hybrid two-shape matching method and is on-par with classical multi-matching methods. Our source code is available at 4dqv.mpiinf.mpg.de/CCuantuMM/."}, "cited_paper_content": {"title": "Hippi: Higher-Order Projected Power Iterations For Scalable Multi-Matching", "abstract": "The matching of multiple objects (e.g. shapes or images) is a fundamental problem in vision and graphics. In order to robustly handle ambiguities, noise and repetitive patterns in challenging real-world settings, it is essential to take geometric consistency between points into account. Computationally, the multi-matching problem is difficult. It can be phrased as simultaneously solving multiple (NP-hard) quadratic assignment problems (QAPs) that are coupled via cycle-consistency constraints. The main limitations of existing multi-matching methods are that they either ignore geometric consistency and thus have limited robustness, or they are restricted to small-scale problems due to their (relatively) high computational cost. We address these shortcomings by introducing a Higher-order Projected Power Iteration method, which is (i) efficient and scales to tens of thousands of points, (ii) straightforward to implement, (iii) able to incorporate geometric consistency, (iv) guarantees cycle-consistent multi-matchings, and (iv) comes with theoretical convergence guarantees. Experimentally we show that our approach is superior to existing methods."}, "keywords": ["permutation synchronization"], "citation_intent": "method"}
{"citing_id": "2303.11964v1", "cited_id": "1607.04247", "section_title": "Simulation Of The Stable Undershoot Via Domination By A Mixture Of Densities", "citation": "In the following two subsections we describe how to sample from the laws given by the densities \u03c8 The bivariate density \u03c8 #REFR s can be written as the product of a marginal log-concave density and an exponential density with deterministic shift and a random scale.", "text_before_citation": ["The random running time of Algorithm 5 has exponential moments.", "Moreover, the expected running time of Algorithm 5 is bounded above by", "\u03ba 5 (1 \u2212 \u03b1) \u22121 1 + \u03b1(1 \u2212 \u03b1) \u22121 log + 2 (s \u22121 ) + log + 2 1/(s \u03b1/(1\u2212\u03b1) \u2212 \u03b1 (1\u22122\u03b1)/(1\u2212\u03b1) (1 \u2212 \u03b1)) + \u03b1 log + 2 (s) + | log(\u03b1)| + (1 \u2212 \u03b1) \u22122 + log N ,", "where the constant \u03ba 5 > 0 depends on neither s = (\u03b8t) \u22121/\u03b1 w \u2208 (0, \u221e) (see line 1 of Algorithm 5), t > 0, \u03b8 \u2208 (0, \u221e) nor \u03b1 \u2208 (0, 1).", "The proof of Proposition 2.5 is given in Subsection 4.3.4 below."], "text_after_citation": ["The latter is easy to simulate and the former can also be simulated via the general Algorithm 11 by virtue of being log-concave.", "Require: Parameters \u03b1 \u2208 (0, 1) and s > 0", "1: Sample Y s with density y \u2192 exp(\u2212\u03c3 \u03b1 (y)s \u2212r ) via Algorithm 11 2: Sample E \u223c Exp(1) and return (s \u2212r + E/\u03c3 \u03b1 (Y s ), Y s )", "Proposition 2.6.", "Algorithm 6 samples from the density \u03c8 #OTHEREFR s ."], "citing_paper_content": {"title": "Fast Exact Simulation Of The First Passage Of A Tempered Stable Subordinator Across A Non-Increasing Function", "abstract": "We construct a fast exact algorithm for the simulation of the first-passage time, jointly with the undershoot and overshoot, of a tempered stable subordinator over an arbitrary non-increasing absolutely continuous function. We prove that the running time of our algorithm has finite exponential moments and provide bounds on its expected running time with explicit dependence on the characteristics of the process and the initial value of the function. The expected running time grows at most cubically in the stability parameter (as it approaches either 0 or 1) and is linear in the tempering parameter and the initial value of the function. Numerical performance, based on the implementation in the dedicated GitHub repository, exhibits a good agreement with our theoretical bounds. We provide numerical examples to illustrate the performance of our algorithm in Monte Carlo estimation."}, "cited_paper_content": {"title": "Accurate And Efficient Numerical Calculation Of Stable Densities Via Optimized Quadrature And Asymptotics", "abstract": "Stable distributions are an important class of infinitely-divisible probability distributions, of which two special cases are the Cauchy distribution and the normal distribution. Aside from a few special cases, the density function for stable distributions has no known analytic form, and is expressible only through the variate's characteristic function or other integral forms. In this paper we present numerical schemes for evaluating the density function for stable distributions, its gradient, and distribution function in various parameter regimes of interest, some of which had no pre-existing efficient method for their computation. The novel evaluation schemes consist of optimized generalized Gaussian quadrature rules for integral representations of the density function, complemented by various asymptotic expansions near various values of the shape and argument parameters. We report several numerical examples illustrating the efficiency of our methods. The resulting code has been made available online."}, "keywords": ["deterministic shift", "bivariate density"], "citation_intent": "background"}
{"citing_id": "2304.14298v1", "cited_id": "1805.01934", "section_title": "Table 13", "citation": "And numerical results are surprisingly good, i.e., 27.2 AP with SID #REFR , which outperforms baseline by 7.4 points. This implies the superiority of using RAW images.", "text_before_citation": ["And ResNet-50-FPN #OTHEREFR Faster R-CNN 22.1 35.6 23.5 GLADNet #OTHEREFR Faster R-CNN 15.4 24.9 16.4 Retinex-Net #OTHEREFR Faster R-CNN 19.6 31.1 21.5 EnlightenGAN #OTHEREFR Faster R-CNN 21.1 34.8 21.9 Zero-DCE #OTHEREFR Faster R-CNN 22.0 35.9 23.5", "Enhance + Denoise HE #OTHEREFR ) + SGN #OTHEREFR Faster R-CNN 25.1 39.8 26.9 GLADNet #OTHEREFR ) + SGN #OTHEREFR Faster R-CNN 24.1 39.1 25.3 Retinex-Net #OTHEREFR ) + SGN #OTHEREFR Faster R-CNN 25.5 41.0 27.4 EnlightenGAN #OTHEREFR itively expect performance improvement, but the accuracies stay the same (with histogram equalization #OTHEREFR and Zero-DCE #OTHEREFR ) or even decrease (with GLADNet #OTHEREFR , Retinex-Net #OTHEREFR and EnlightenGAN #OTHEREFR ).", "We guess the reason is that these enhancers only improve the overall brightness but cannot handle the noise.", "To verify it, we further introduce denoiser to the pipeline, and the overall accuracy significantly increases as expected, e.g., Zero-DCE #OTHEREFR plus SGN #OTHEREFR leads to 6.7 AP gain. Notice that these methods for comparison use camera outputs.", "Then, we also perform experiments with SID #OTHEREFR and REDI #OTHEREFR , which can restore sRGB images from low-light RAW images."], "text_after_citation": ["Though these enhancing and denoising steps boost the low-light instance segmentation performance remarkably, our method achieves the best quantitative results without extra preprocessing steps.", "Besides, the inference speed of the pro-posed method outperforms all other pipelines.", "And its speed is very close to the original Mask R-CNN #OTHEREFR .", "Moreover, qualitative results illustrated in Figure 12 show the proposed method can consistently recall most of the targets even in challenging scenarios."], "citing_paper_content": {"title": "Instance Segmentation In The Dark", "abstract": "Existing instance segmentation techniques are primarily tailored for high-visibility inputs, but their performance significantly deteriorates in extremely low-light environments. In this work, we take a deep look at instance segmentation in the dark and introduce several techniques that substantially boost the low-light inference accuracy. The proposed method is motivated by the observation that noise in low-light images introduces high-frequency disturbances to the feature maps of neural networks, thereby significantly degrading performance. To suppress this \"feature noise\", we propose a novel learning method that relies on an adaptive weighted downsampling layer, a smooth-oriented convolutional block, and disturbance suppression learning. These components effectively reduce feature noise during downsampling and convolution operations, enabling the model to learn disturbance-invariant features. Furthermore, we discover that high-bit-depth RAW images can better preserve richer scene information in low-light conditions compared to typical camera sRGB outputs, thus supporting the use of RAW-input algorithms. Our analysis indicates that high bit"}, "cited_paper_content": {"title": "Learning To See In The Dark", "abstract": "Imaging in low light is challenging due to low photon count and low SNR. Short-exposure images suffer from noise, while long exposure can induce blur and is often impractical. A variety of denoising, deblurring, and enhancement techniques have been proposed, but their effectiveness is limited in extreme conditions, such as video-rate imaging at night. To support the development of learning-based pipelines for low-light image processing, we introduce a dataset of raw short-exposure low-light images, with corresponding long-exposure reference images. Using the presented dataset, we develop a pipeline for processing low-light images, based on end-to-end training of a fully-convolutional network. The network operates directly on raw sensor data and replaces much of the traditional image processing pipeline, which tends to perform poorly on such data. We report promising results on the new dataset, analyze factors that affect performance, and highlight opportunities for future work. The results are shown in the supplementary video at this https URL"}, "keywords": ["RAW images"], "citation_intent": "result"}
{"citing_id": "2304.00173v1", "cited_id": "2003.07962", "section_title": "Downstream Las Decoder / Deliberation", "citation": "Figure 3 also ends with an LAS decoder, except this one can optionally attend to the continuous encoder features as well, as is done in previous deliberation work #REFR . Gradients do not flow back through embedded N-best.", "text_before_citation": ["A fitting baseline to this experiment is second-pass deliberation ASR #OTHEREFR .", "Typically, a deliberation system generates first-pass hypotheses using a fast decoder, like RNN-T, then embeds its Nbest hyps and attends to them with a second-pass full-context LAS decoder.", "We have therefore constructed a comparable deliberation baseline model shown in Figure 3 . This model is analogous to our full pipeline, i.e.", "Figures 1 & 2 put together, and is designed to have a similar total model size and encoder latency.", "It starts with the same frozen base encoder, then trains a first-pass RNN-T decoder to obtain the N-best hyps, which stands to be compared to the Lego-Features in terms of informativeness and modularity."], "text_after_citation": [], "citing_paper_content": {"title": "Lego-Features: Exporting Modular Encoder Features For Streaming And Deliberation Asr", "abstract": "In end-to-end (E2E) speech recognition models, a representational tight-coupling inevitably emerges between the encoder and the decoder. We build upon recent work that has begun to explore building encoders with modular encoded representations, such that encoders and decoders from different models can be stitched together in a zero-shot manner without further fine-tuning. While previous research only addresses full-context speech models, we explore the problem in a streaming setting as well. Our framework builds on top of existing encoded representations, converting them to modular features, dubbed as Lego-Features, without modifying the pre-trained model. The features remain interchangeable when the model is retrained with distinct initializations. Though sparse, we show that the Lego-Features are powerful when tested with RNN-T or LAS decoders, maintaining high-quality downstream performance. They are also rich enough to represent the first-pass prediction during twopass deliberation. In this scenario, they outperform the N-best hypotheses, since they do not need to be supplemented with acoustic features to deliver the best results. Moreover, generating the Lego-Features does not require beam search or auto-regressive computation. Overall, they present a modular, powerful and cheap alternative to the standard encoder output, as well as the N-best hypotheses."}, "cited_paper_content": {"title": "Deliberation Model Based Two-Pass End-To-End Speech Recognition", "abstract": "End-to-end (E2E) models have made rapid progress in automatic speech recognition (ASR) and perform competitively relative to conventional models. To further improve the quality, a two-pass model has been proposed to rescore streamed hypotheses using the non-streaming Listen, Attend and Spell (LAS) model while maintaining a reasonable latency. The model attends to acoustics to rescore hypotheses, as opposed to a class of neural correction models that use only first-pass text hypotheses. In this work, we propose to attend to both acoustics and first-pass hypotheses using a deliberation network. A bidirectional encoder is used to extract context information from first-pass hypotheses. The proposed deliberation model achieves 12% relative WER reduction compared to LAS rescoring in Google Voice Search (VS) tasks, and 23% reduction on a proper noun test set. Compared to a large conventional model, our best model performs 21% relatively better for VS. In terms of computational complexity, the deliberation decoder has a larger size than the LAS decoder, and hence requires more computations in second-pass decoding."}, "keywords": ["LAS decoder"], "citation_intent": "background"}
{"citing_id": "2303.00844v1", "cited_id": "1009.3525", "section_title": "Literature Review", "citation": "A similar result was derived in #REFR from a probabilistic point of view where the signal support is assumed to be formed by two subsets with different probability of occurrence.", "text_before_citation": ["Weights have been employed in sparse recovery methods for various purposes.", "For instance, in the seminal work #OTHEREFR , the authors propose to solve a sequence of (re)weighted \u2113 1 minimization problems to enhance sparse signal recovery.", "In our context, weights can generally be thought of as a way of incorporating prior information about the signal into a sparse recovery model.", "In adaptive LASSO #OTHEREFR , a data-driven but careful choice of weights is shown to admit near oracle properties.", "Works such as #OTHEREFR show that replacing the \u2113 1 -norm with its weighted version can improve recovery assuming that accurate (partial) support knowledge is provided."], "text_after_citation": ["Further studies of weighted \u2113 1 minimization and its impactful application in the context of function approximation from pointwise samples and uncertainty quantification include #OTHEREFR .", "The notion of weighted sparsity was formalized in #OTHEREFR .", "Weighted sparsity is related to structured sparsity (see #OTHEREFR ).", "In fact it allows one to promote structures (rather than being a structure itself).", "For example, in the context of highdimensional function approximation (see #OTHEREFR and references therein) weights are able to promote so-called sparsity in lower sets, which largely contributes to mitigating the curse of dimensionality in the sample complexity."], "citing_paper_content": {"title": "The Greedy Side Of The Lasso: New Algorithms For Weighted Sparse Recovery Via Loss Function-Based Orthogonal Matching Pursuit", "abstract": "We propose a class of greedy algorithms for weighted sparse recovery by considering new loss function-based generalizations of Orthogonal Matching Pursuit (OMP). Given a (regularized) loss function, the proposed algorithms alternate the iterative construction of the signal support via greedy index selection and a signal update based on solving a local data-fitting problem restricted to the current support. We show that greedy selection rules associated with popular weighted sparsity-promoting loss functions admit explicitly computable and simple formulas. Specifically, we consider \u2113 0-and \u2113 1-based versions of the weighted LASSO (Least Absolute Shrinkage and Selection Operator), the Square-Root LASSO (SR-LASSO) and the Least Absolute Deviations LASSO (LAD-LASSO). Through numerical experiments on Gaussian compressive sensing and high-dimensional function approximation, we demonstrate the effectiveness of the proposed algorithms and empirically show that they inherit desirable characteristics from the corresponding loss functions, such as SR-LASSO's noise-blind optimal parameter tuning and LAD-LASSO's fault tolerance. In doing so, our study sheds new light on the connection between greedy sparse recovery and convex relaxation."}, "cited_paper_content": {"title": "Analyzing Weighted $\\Ell_1$ Minimization For Sparse Recovery With Nonuniform Sparse Models", "abstract": "In this paper we introduce a nonuniform sparsity model and analyze the performance of an optimized weighted $\\ell_1$ minimization over that sparsity model. In particular, we focus on a model where the entries of the unknown vector fall into two sets, with entries of each set having a specific probability of being nonzero. We propose a weighted $\\ell_1$ minimization recovery algorithm and analyze its performance using a Grassmann angle approach. We compute explicitly the relationship between the system parameters-the weights, the number of measurements, the size of the two sets, the probabilities of being nonzero- so that when i.i.d. random Gaussian measurement matrices are used, the weighted $\\ell_1$ minimization recovers a randomly selected signal drawn from the considered sparsity model with overwhelming probability as the problem dimension increases. This allows us to compute the optimal weights. We demonstrate through rigorous analysis and simulations that for the case when the support of the signal can be divided into two different subclasses with unequal sparsity fractions, the optimal weighted $\\ell_1$ minimization outperforms the regular $\\ell_1$ minimization substantially. We also generalize the results to an arbitrary number of classes."}, "keywords": ["signal support"], "citation_intent": "result"}
{"citing_id": "2304.01447v1", "cited_id": "1802.05098", "section_title": "Iterated Prisoner'S Dilemma", "citation": "Additionally, we compared the performance of LOLA-OffPA2 with LOLA-DiCE #REFR , which is designed specifically for this game, and we reported the results in Table 3 .", "text_before_citation": ["The LOLA agents can shape the opponent's learning to encourage cooperation and, therefore, converge to TFT #OTHEREFR .", "We evaluate the methods' performances based on the Averaged Episode Reward (AER).", "Cooperate (\u22121, \u22121) (\u22123, 0) Defect (0, \u22123) (\u22123, \u22123)", "In Figure 4 , we depict the learning curves for LOLA-OffPA2 and the MADDPG-based methods.", "From this figure, we find that only LOLA-OffPA2 can solve the game, which once again highlights the importance of learning anticipation."], "text_after_citation": ["Although both methods demonstrate high values of AER, our LOLA-OffPA2 is significantly more efficient as its LATC value is much lower than that of LOLA-DiCE.", "Table 5 : Comparisons of LOLA-DiCE with our proposed LOLA-OffPA2 in the Exit-Room game, in terms of performance (normalized average return in different game levels) and efficiency (learning anticipation time complexity in different reasoning levels)."], "citing_paper_content": {"title": "Off-Policy Action Anticipation In Multi-Agent Reinforcement Learning", "abstract": "Learning anticipation in Multi-Agent Reinforcement Learning (MARL) is a reasoning paradigm where agents anticipate the learning steps of other agents to improve cooperation among themselves. As MARL uses gradient-based optimization, learning anticipation requires using Higher-Order Gradients (HOG), with so-called HOG methods. Existing HOG methods are based on policy parameter anticipation, i.e., agents anticipate the changes in policy parameters of other agents. Currently, however, these existing HOG methods have only been applied to differentiable games or games with small state spaces. In this work, we demonstrate that in the case of non-differentiable games with large state spaces, existing HOG methods do not perform well and are inefficient due to their inherent limitations related to policy parameter anticipation and multiple sampling stages. To overcome these problems, we propose Off-Policy Action Anticipation (OffPA2), a novel framework that approaches learning anticipation through action anticipation, i.e., agents anticipate the changes in actions of other agents, via off-policy sampling. We theoretically analyze our proposed OffPA2 and employ it to develop multiple HOG methods that are applicable to non-differentiable games with large state spaces. We conduct a large set of experiments and illustrate that our proposed HOG methods outperform the existing ones regarding efficiency and performance."}, "cited_paper_content": {"title": "Dice: The Infinitely Differentiable Monte-Carlo Estimator", "abstract": "The score function estimator is widely used for estimating gradients of stochastic objectives in Stochastic Computation Graphs (SCG), eg. in reinforcement learning and meta-learning. While deriving the first-order gradient estimators by differentiating a surrogate loss (SL) objective is computationally and conceptually simple, using the same approach for higher-order gradients is more challenging. Firstly, analytically deriving and implementing such estimators is laborious and not compliant with automatic differentiation. Secondly, repeatedly applying SL to construct new objectives for each order gradient involves increasingly cumbersome graph manipulations. Lastly, to match the first-order gradient under differentiation, SL treats part of the cost as a fixed sample, which we show leads to missing and wrong terms for higher-order gradient estimators. To address all these shortcomings in a unified way, we introduce DiCE, which provides a single objective that can be differentiated repeatedly, generating correct gradient estimators of any order in SCGs. Unlike SL, DiCE relies on automatic differentiation for performing the requisite graph manipulations. We verify the correctness of DiCE both through a proof and through numerical evaluation of the DiCE gradient estimates. We also use DiCE to propose and evaluate a novel approach for multi-agent learning. Our code is available at this https URL"}, "keywords": ["game", "LOLA-DiCE"], "citation_intent": "result"}
{"citing_id": "2305.02337v1", "cited_id": "1302.5843", "section_title": "C. Ising Model", "citation": "It is a natural starting point for Hamiltonian simulation since it can be used to formulate many computationally hard problems, such as spin glasses, Quadratic Unconstrained Binary Optimization problems (QUBOs), or graph partitioning #REFR .", "text_before_citation": ["This paper will primarily focus on the transverse-field Ising model as a representative of an important class of Hamiltonians.", "In its general form, it models the spins of particles and is described by the Hamiltonian", "EQUATION", "where J ij is the interaction strength between nearest-neighbor spin pairs at sites i, j and g j is an external field pointing perpendicular to the interactions at site j."], "text_after_citation": ["The following constrained version of this type of Hamiltonian will be used to illustrate all proposed concepts and methods throughout the remainder of this paper.", "Example 1. The L-site finite 1D Ising chain is defined by", "EQUATION", "where the parameters are site-independent, i.e., J ij := J and g := g.", "Using the product formula, a single Trotter under this model has the form"], "citing_paper_content": {"title": "Towards Hamiltonian Simulation With Decision Diagrams", "abstract": "This paper proposes a novel approach to Hamiltonian simulation using Decision Diagrams (DDs), which are an exact representation based on exploiting redundancies in representations of quantum states and operations. While the simulation of Hamiltonians has been studied extensively, scaling these simulations to larger or more complex systems is often challenging and may require approximations or new simulation methods altogether. DDs offer such an alternative that has not yet been applied to Hamiltonian simulation. In this work, we investigate the behavior of DDs for this task. To this end, we review the basics of DDs such as their construction and present how the relevant operations for Hamiltonian simulation are implemented in this data structure-leading to the first DD-based Hamiltonian simulation approach. Based on several series of evaluations and comparisons, we then discuss insights about the performance of this complementary approach. Overall, these studies show that DDs indeed may offer a promising new data structure which, for certain examples, can provide orders of magnitudes of improvement compared to the state-of-the-art, yet also comes with its own, fundamentally different, limitations."}, "cited_paper_content": {"title": "Ising Formulations Of Many Np Problems", "abstract": "We provide Ising formulations for many NP-complete and NP-hard problems, including all of Karp's 21 NP-complete problems. This collects and extends mappings to the Ising model from partitioning, covering and satisfiability. In each case, the required number of spins is at most cubic in the size of the problem. This work may be useful in designing adiabatic quantum optimization algorithms."}, "keywords": ["Hamiltonian simulation", "many computationally hard"], "citation_intent": "background"}
{"citing_id": "2303.01841v1", "cited_id": "1806.07366", "section_title": "Related Work", "citation": "Much more closely aligned to our work, and a natural fit for irregularly sampled data is research that uses differential equations to model continuous-time processes #REFR .", "text_before_citation": ["Time series modelling in machine learning: There is vast literature on the use of machine learning for time series modelling and we highlight some of the ideas that have been explored to adapt diverse kinds of models for irregular time series data.", "Although not naturally well suited to learning representations of such data, there have been modifications proposed to discrete-time models such as recurrent neural networks #OTHEREFR to handle such data.", "Models such as mTANs #OTHEREFR leverage an attention-based approach to interpolate sequences to create discrete-time data from irregularly sampled data.", "Another strategy has been architectural modifications to the recurrence equations e.g.", "CT-GRU #OTHEREFR , GRU-D #OTHEREFR and Unitary RNNs #OTHEREFR ."], "text_after_citation": ["By parameterizing the derivative of a time series using neural networks and integrating the dynamics over unobserved time points, this class of models is well suited to handle irregularly sampled data.", "This includes models such as ODE-RNN #OTHEREFR , ODE-LSTM #OTHEREFR and Neural CDE #OTHEREFR .", "ODE-based approaches require the use of differential equation solvers during training and inference, which can come at the cost of runtime #OTHEREFR .", "PolyODEs lie in this family of models; specifically, this work proposes a new parameterization of the dynamics function and a practical method for learning that enables this model family to accurately forecast the future and reconstruct the past greatly enhancing the scope and utility of the learned embeddings.", "Orthogonal polynomials: PolyODEs are inspired by a rich line of work in orthogonal decomposition of time series data."], "citing_paper_content": {"title": "Anamnesic Neural Differential Equations With Orthogonal Polynomials Projections", "abstract": "Neural ordinary differential equations (Neural ODEs) are an effective framework for learning dynamical systems from irregularly sampled time series data. These models provide a continuous-time latent representation of the underlying dynamical system where new observations at arbitrary time points can be used to update the latent representation of the dynamical system. Existing parameterizations for the dynamics functions of Neural ODEs limit the ability of the model to retain global information about the time series; specifically, a piece-wise integration of the latent process between observations can result in a loss of memory on the dynamic patterns of previously observed data points. We propose PolyODE, a Neural ODE that models the latent continuous-time process as a projection onto a basis of orthogonal polynomials. This formulation enforces long-range memory and preserves a global representation of the underlying dynamical system. Our construction is backed by favourable theoretical guarantees and in a series of experiments, we demonstrate that it outperforms previous works in the reconstruction of past and future data, and in downstream prediction tasks. Our code is available at https://github.com/edebrouwer/polyode."}, "cited_paper_content": {"title": "Neural Ordinary Differential Equations", "abstract": "We introduce a new family of deep neural network models. Instead of specifying a discrete sequence of hidden layers, we parameterize the derivative of the hidden state using a neural network. The output of the network is computed using a black-box differential equation solver. These continuous-depth models have constant memory cost, adapt their evaluation strategy to each input, and can explicitly trade numerical precision for speed. We demonstrate these properties in continuous-depth residual networks and continuous-time latent variable models. We also construct continuous normalizing flows, a generative model that can train by maximum likelihood, without partitioning or ordering the data dimensions. For training, we show how to scalably backpropagate through any ODE solver, without access to its internal operations. This allows end-to-end training of ODEs within larger models."}, "keywords": ["continuous-time processes", "differential equations"], "citation_intent": "background"}
{"citing_id": "2304.11045v1", "cited_id": "1603.09320", "section_title": "Iii. Lightdxml: Architecture", "citation": "Like in Astec, this is achieved in O(log(L)) time per data point using an ANN data structure #REFR .", "text_before_citation": ["Specifically, the goal is now to train label encoder Z l to minimize the mean squared error between feature representations x i (obtained from Step 1) and label embeddingl j if the j th label is positive for the data point i.", "By letting P i be the set of all positive labels for the i th data point, i.e., P = {j :", "EQUATION", "In", "Step 3, LightDXML finds the top k = O(log(L)) most hard negative labels for each data point using the shortlist S k by finding the k nearest label points\u0177 j to the feature representation x i ."], "text_after_citation": ["Using S k , we now have access to both positive and negative labels for a data point.", "In", "Step 4, making use of S k , LightDXML learns the onevs-all classifier Z by jointly learning data feature embedding\u015d Z x , and training on the union of set of positive labels and a set of shortlisted negative labels (instead of training on all labels), i.e., argmin Zx,ZL", "(\u1e90 x , Z) = N i=1 j\u2208P i \u222aS k (x i ) (x i , y ij ;\u1e90 x , Z).", "(3) It should be noted that Problem (3) is more efficient to solve than the original problem (1), due to training on a much reduced set of labels."], "citing_paper_content": {"title": "Light-Weight Deep Extreme Multilabel Classification", "abstract": "Extreme multi-label (XML) classification refers to the task of supervised multi-label learning that involves a large number of labels. Hence, scalability of the classifier with increasing label dimension is an important consideration. In this paper, we develop a method called LightDXML which modifies the recently developed deep learning based XML framework by using label embeddings instead of feature embedding for negative sampling and iterating cyclically through three major phases: (1) proxy training of label embeddings (2) shortlisting of labels for negative sampling and (3) final classifier training using the negative samples. Consequently, LightDXML also removes the requirement of a re-ranker module, thereby, leading to further savings on time and memory requirements. The proposed method achieves the best of both worlds: while the training time, model size and prediction times are on par or better compared to the tree-based methods, it attains much better prediction accuracy that is on par with the deep learning based methods. Moreover, the proposed approach achieves the best tail-label prediction accuracy over most state-of-the-art XML methods on some of the large datasets 1. Code: https://github.com/misterpawan/ LightDXML"}, "cited_paper_content": {"title": "Efficient And Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs", "abstract": "We present a new approach for the approximate K-nearest neighbor search based on navigable small world graphs with controllable hierarchy (Hierarchical NSW, HNSW). The proposed solution is fully graph-based, without any need for additional search structures, which are typically used at the coarse search stage of the most proximity graph techniques. Hierarchical NSW incrementally builds a multi-layer structure consisting from hierarchical set of proximity graphs (layers) for nested subsets of the stored elements. The maximum layer in which an element is present is selected randomly with an exponentially decaying probability distribution. This allows producing graphs similar to the previously studied Navigable Small World (NSW) structures while additionally having the links separated by their characteristic distance scales. Starting search from the upper layer together with utilizing the scale separation boosts the performance compared to NSW and allows a logarithmic complexity scaling. Additional employment of a heuristic for selecting proximity graph neighbors significantly increases performance at high recall and in case of highly clustered data. Performance evaluation has demonstrated that the proposed general metric space search index is able to strongly outperform previous opensource state-of-the-art vector-only approaches. Similarity of the algorithm to the skip list structure allows straightforward balanced distributed implementation."}, "keywords": ["ANN data structure"], "citation_intent": "method"}
{"citing_id": "2304.05642v1", "cited_id": "1810.04805", "section_title": "Introduction", "citation": "Discrete prompt tuning transforms the task into a \"fill-in-the-blank\" format and then utilizes a pre-trained language model to predict the answer, which operates similarly to the masked language model (MLM) #REFR .", "text_before_citation": ["Prompt-based methods can be classified into two categories: discrete prompt tuning (Radford et al., 2018; #OTHEREFR and continuous prompt tuning #OTHEREFR Liu et al., 2021b) ."], "text_after_citation": ["In subsequent research, a recent study #OTHEREFR introduced soft prompts (i.e., prompt embeddings) to replace manual templates, which consist of special tokens with adjustable embeddings.", "We refer to this method as \"prompt tuning\" for simplicity.", "Vanilla prompt tuning, which concatenates prompt embeddings with input tokens in the first layer and updates only the parameters of the prompt embeddings during the training phase, has several limitations #OTHEREFR Su et al., 2021) .", "Firstly, since the effectiveness of prompt embeddings is highly related to the length, it is necessary to use prompt embeddings with hundreds of tokens in length to achieve better downstream task performance, as suggested by #OTHEREFR and (Su et al., 2021) .", "However, longer prompt embeddings also reduce the length space of normal input text."], "citing_paper_content": {"title": "Global Prompt Cell: A Portable Control Module For Effective Prompt Tuning", "abstract": "As a novel approach to tuning pre-trained models, prompt tuning involves freezing the parameters in downstream tasks while inserting trainable embeddings into inputs in the first layer. However, previous methods have mainly focused on the initialization of prompt embeddings. The question of how to train and utilize prompt embeddings in a reasonable way has become a limiting factor in the effectiveness of prompt tuning. To address this issue, we introduce the Global Prompt Cell (GPC), a portable control module for prompt tuning that selectively preserves prompt information across all encoder layers. Our experimental results demonstrate a 5.8% improvement on SuperGLUE datasets compared to vanilla prompt tuning."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["Discrete prompt tuning", "pre-trained language model"], "citation_intent": "method"}
{"citing_id": "2304.12210v1", "cited_id": "1911.05371", "section_title": "Grouping Similar Images Together:", "citation": "Other improvements to deep clustering include using optimal transport methods in feature space to create more informative clusters #REFR .", "text_before_citation": ["One can learn rich features by grouping semantically similar images together.", "K-means clustering is one of the most widely used methods from classical machine learning.", "A number of studies have adapted k-means to perform SSL with neural models.", "Deep clustering alternates between assigning labels to images by performing k-means in the feature space, and updating the model to respect these assigned class labels #OTHEREFR .", "More recent treatments of this approach use mean-shift updates to push features towards their cluster center, and have been shown to complement BYOL, a method based on two networks with the objective to predict pseudo-labels for each sample (discussed in Section 2.3)."], "text_after_citation": [], "citing_paper_content": {"title": "A Cookbook Of Self-Supervised Learning", "abstract": ""}, "cited_paper_content": {"title": "Self-Labelling Via Simultaneous Clustering And Representation Learning", "abstract": "Combining clustering and representation learning is one of the most promising approaches for unsupervised learning of deep neural networks. However, doing so naively leads to ill posed learning problems with degenerate solutions. In this paper, we propose a novel and principled learning formulation that addresses these issues. The method is obtained by maximizing the information between labels and input data indices. We show that this criterion extends standard cross-entropy minimization to an optimal transport problem, which we solve efficiently for millions of input images and thousands of labels using a fast variant of the Sinkhorn-Knopp algorithm. The resulting method is able to self-label visual data so as to train highly competitive image representations without manual labels. Compared to the best previous method in this class, namely DeepCluster, our formulation minimizes a single objective function for both representation learning and clustering; it also significantly outperforms DeepCluster in standard benchmarks."}, "keywords": ["deep clustering"], "citation_intent": "method"}
{"citing_id": "2304.04661v1", "cited_id": "1910.05339", "section_title": "B. Log-Based Rca Problem Definition", "citation": "Knowledge Graph based Methods: Amongst knowledge graph based methods, #REFR diagnoses and triages performance failure issues in an online fashion by continuously building a knowledge base out of rules extracted from a random forest constructed over log data using heuristics and domain knowledge.", "text_before_citation": ["#OTHEREFR , #OTHEREFR uses KNN or its supervised versions to identify loglines that led to a failure.", "Knowledge Mining based Methods: #OTHEREFR , #OTHEREFR takes a different approach of summarizing log events into an entityrelation knowledge graph by extracting custom entities and relationships from log lines and mining temporal and procedural dependencies between them from the overall log dump.", "While this gives a more structured representation of the log summary, it is also an intuitive way of aggregating knowledge from logs, it is also a way to bridge the knowledge gap developer community who creates the log data and the site reliability engineers who typically consume the log data when investigating incidents.", "However, eventually the end goal of constructing this knowledge graph representation of logs is to facilitate RCA.", "While these works do provide use-cases like case-studies on RCA for this vision, but they leave ample scope of research towards a more concrete usage of this kind of knowledge mining in RCA."], "text_after_citation": ["#OTHEREFR constructs a system graph from the combination of KPI metrics and log data.", "Based on the detected anomalies from these data sources, it extracts anomalous subgraphs from it and compares them with the normal system graph to detect the root cause.", "Other works mine normal log patterns #OTHEREFR or time-weighted control flow graphs #OTHEREFR from normal executions and on estimates divergences from them to executions during ongoing failures to suggest root causes.", "#OTHEREFR , #OTHEREFR , #OTHEREFR mines execution sequences or user actions #OTHEREFR either from normal and manually injected failures or from good or bad performing systems, in a knowledge base and utilizes the assumption that similar faults generate similar failures to match and diagnose type of failure.", "Most of these knowledge based approaches incrementally expand their knowledge or rules to cater to newer incident types over time."], "citing_paper_content": {"title": "Ai For It Operations (Aiops) On Cloud Platforms: Reviews, Opportunities And Challenges", "abstract": "Intelligence for IT operations (AIOps) aims to combine the power of AI with the big data generated by IT Operations processes, particularly in cloud infrastructures, to provide actionable insights with the primary goal of maximizing availability. There are a wide variety of problems to address, and multiple use-cases, where AI capabilities can be leveraged to enhance operational efficiency. Here we provide a review of the AIOps vision, trends challenges and opportunities, specifically focusing on the underlying AI techniques. We discuss in depth the key types of data emitted by IT Operations activities, the scale and challenges in analyzing them, and where they can be helpful. We categorize the key AIOps tasks as-incident detection, failure prediction, root cause analysis and automated actions. We discuss the problem formulation for each task, and then present a taxonomy of techniques to solve these problems. We also identify relatively under explored topics, especially those that could significantly benefit from advances in AI literature. We also provide insights into the trends in this field, and what are the key investment opportunities."}, "cited_paper_content": {"title": "Decaf: Diagnosing And Triaging Performance Issues In Large-Scale Cloud Services", "abstract": "Large scale cloud services use Key Performance Indicators (KPIs) for tracking and monitoring performance. They usually have Service Level Objectives (SLOs) baked into the customer agreements which are tied to these KPIs. Dependency failures, code bugs, infrastructure failures, and other problems can cause performance regressions. It is critical to minimize the time and manual effort in diagnosing and triaging such issues to reduce customer impact. Large volumes of logs and mixed type of attributes (categorical, continuous) make any automated or manual diagnosing non-trivial. ::: In this paper, we present the design, implementation and experience from building and deploying DeCaf, a system for automated diagnosis and triaging of KPI issues using service logs. It uses machine learning along with pattern mining to help service owners automatically root cause and triage performance issues. We present the learnings and results from case studies on two large scale cloud services in Microsoft where DeCaf successfully diagnosed 10 known and 31 unknown issues. DeCaf also automatically triages the identified issues by leveraging historical data. Our key insights are that for any such diagnosis tool to be effective in practice, it should a) scale to large volumes of service logs and attributes, b) support different types of KPIs and ranking functions, c) be integrated into the DevOps processes."}, "keywords": ["log data", "performance failure issues"], "citation_intent": "method"}
{"citing_id": "2303.13731v1", "cited_id": "1409.0473", "section_title": "Related Work", "citation": "The attention mechanism #REFR has been used extensively in DL, especially NLP-related tasks, to learn what target tokens the source tokens should \"look at\".", "text_before_citation": ["We refer readers to recent surveys #OTHEREFR , #OTHEREFR for a thorough review of these works.", "Lately, deep transformers demonstrate superior performance than other DL models on 1D sequential data, and multiple visualization works have been introduced for their interpretations #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR .", "The success of transformers has also been extended to 2D images with the seminal work of vision transformers (ViTs) #OTHEREFR .", "However, to the best of our knowledge, no comprehensive visual analyses have been conducted to demystify this type of powerful yet complex models, especially how attention works in the 2D image context. Our work tries to fill this gap.", "Attention Visualization."], "text_after_citation": ["Essentially, attention is a matrix where each cell denotes the attention magnitude that the source token (row) pays to the target (column).", "Popular attention visualization techniques include flow maps #OTHEREFR , #OTHEREFR , parallel coordinates plots (PCPs) #OTHEREFR , and heatmaps #OTHEREFR , #OTHEREFR , #OTHEREFR .", "For example, the flow maps used by Dong et al.", "#OTHEREFR connect the source and target tokens with curves, the widths of which denote the attention strengths.", "Vig #OTHEREFR arranges the source and target tokens along two parallel axes (i.e., a simplified PCP) and connects them with line segments in between to show the attention patterns."], "citing_paper_content": {"title": "How Does Attention Work In Vision Transformers? A Visual Analytics Attempt", "abstract": "Vision transformer (ViT) expands the success of transformer models from sequential data to images. The model decomposes an image into many smaller patches and arranges them into a sequence. Multi-head self-attentions are then applied to the sequence to learn the attention between patches. Despite many successful interpretations of transformers on sequential data, little effort has been devoted to the interpretation of ViTs, and many questions remain unanswered. For example, among the numerous attention heads, which one is more important? How strong are individual patches attending to their spatial neighbors in different heads? What attention patterns have individual heads learned? In this work, we answer these questions through a visual analytics approach. Specifically, we first identify what heads are more important in ViTs by introducing multiple pruning-based metrics. Then, we profile the spatial distribution of attention strengths between patches inside individual heads, as well as the trend of attention strengths across attention layers. Third, using an autoencoder-based learning solution, we summarize all possible attention patterns that individual heads could learn. Examining the attention strengths and patterns of the important heads, we answer why they are important. Through concrete case studies with experienced deep learning experts on multiple ViTs, we validate the effectiveness of our solution that deepens the understanding of ViTs from head importance, head attention strength, and head attention pattern."}, "cited_paper_content": {"title": "Neural Machine Translation By Jointly Learning To Align And Translate", "abstract": "Neural machine translation is a recently proposed approach to machine translation. Unlike the traditional statistical machine translation, the neural machine translation aims at building a single neural network that can be jointly tuned to maximize the translation performance. The models proposed recently for neural machine translation often belong to a family of encoder-decoders and consists of an encoder that encodes a source sentence into a fixed-length vector from which a decoder generates a translation. In this paper, we conjecture that the use of a fixed-length vector is a bottleneck in improving the performance of this basic encoder-decoder architecture, and propose to extend this by allowing a model to automatically (soft-)search for parts of a source sentence that are relevant to predicting a target word, without having to form these parts as a hard segment explicitly. With this new approach, we achieve a translation performance comparable to the existing state-of-the-art phrase-based system on the task of English-to-French translation. Furthermore, qualitative analysis reveals that the (soft-)alignments found by the model agree well with our intuition."}, "keywords": ["attention mechanism", "DL, especially NLP-related"], "citation_intent": "method"}
{"citing_id": "2305.00135v1", "cited_id": "1703.06182", "section_title": "C. Semi-Distributed Multi-Agent Rl For Robust Handover Optimization", "citation": "In contrast, the second learning rate, \u03b7 2 , is much smaller and is utilized to slow down the degradation of Q-values associated with previously positive experiences resulting from successful operations #REFR .", "text_before_citation": ["This is particularly crucial in providing uninterrupted and dependable service for XR users, whose needs may change over time, and guarantees a resilient user experience that can promptly recover from any disruptions.", "Moreover, the semi-distributed learning framework used in this approach ensures guaranteed convergence due to the adoption of asynchronous updates, where agents update their policy and value function independently and asynchronously.", "Although the optimality of the solution may not be guaranteed in highly dynamic and non-stationary environments like THz networks, the quick adaptation of deep hysteretic networks enables the system to maintain robust and resilient performance in response to unpredictable user behavior.", "Hysteretic deep recurrent Q-networks leverage two distinct learning rates to handle the complex dynamics of the learning process.", "The first learning rate, \u03b7 1 , is used when the temporal difference (TD)-error is nonnegative."], "text_after_citation": ["By implementing this approach, hysteresis is introduced, allowing subarrays to be more resilient against negative learning, exploration, and concurrent actions.", "This approach significantly improves the stability and robustness of the learning process, leading to better overall performance and expedited convergence.", "Similarly to #OTHEREFR , the deep hysteretic Q-network adopted is with one input layer, two fully connected hidden layers, one RNN hidden layer, a dueling layer, and an output layer.", "The subarray's local observations and the estimated state-action value Q b.n , respectively define the input and output layer of the hysteretic deep recurrent Q-network.", "Multi-agent RL frameworks are known to face the challenge of shadowed equilibria, a phenomenon where local observations and non-stationarity cause locally optimal actions to become suboptimal at the global level #OTHEREFR ."], "citing_paper_content": {"title": "Joint Sensing, Communication, And Ai: A Trifecta For Resilient Thz User Experiences", "abstract": "In this paper a novel joint sensing, communication, and artificial intelligence (AI) framework is proposed so as to optimize extended reality (XR) experiences over terahertz (THz) wireless systems. The proposed framework consists of three main components. First, a tensor decomposition framework is proposed to extract unique sensing parameters for XR users and their environment by exploiting then THz channel sparsity. Essentially, THz band's quasi-opticality is exploited and the sensing parameters are extracted from the uplink communication signal, thereby allowing for the use of the same waveform, spectrum, and hardware for both communication and sensing functionalities. Then, the Cramer-Rao lower bound is derived to assess the accuracy of the estimated sensing parameters. Second, a non-autoregressive multi-resolution generative artificial intelligence (AI) framework integrated with an adversarial transformer is proposed to predict missing and future sensing information. The proposed framework offers robust and comprehensive historical sensing information and anticipatory forecasts of future environmental changes, which are generalizable to fluctuations in both known and unforeseen user behaviors and environmental conditions. Third, a multi-agent deep recurrent hysteretic Q-neural network is developed to control the handover policy of reconfigurable intelligent surface (RIS) subarrays, leveraging the informative nature of sensing information to minimize handover cost, maximize the individual quality of personal experiences (QoPEs), and improve the robustness and resilience of THz links. Simulation results show a high generalizability of the proposed unsupervised generative AI framework to fluctuations in user behavior and velocity, leading to a 61 % improvement in instantaneous reliability compared to schemes with known channel state information."}, "cited_paper_content": {"title": "Deep Decentralized Multi-Task Multi-Agent Reinforcement Learning Under Partial Observability", "abstract": "Many real-world tasks involve multiple agents with partial observability and limited communication. Learning is challenging in these settings due to local viewpoints of agents, which perceive the world as non-stationary due to concurrently-exploring teammates. Approaches that learn specialized policies for individual tasks face problems when applied to the real world: not only do agents have to learn and store distinct policies for each task, but in practice identities of tasks are often non-observable, making these approaches inapplicable. This paper formalizes and addresses the problem of multi-task multi-agent reinforcement learning under partial observability. We introduce a decentralized single-task learning approach that is robust to concurrent interactions of teammates, and present an approach for distilling single-task policies into a unified policy that performs well across multiple related tasks, without explicit provision of task identity."}, "keywords": ["second learning rate"], "citation_intent": "background"}
{"citing_id": "2303.17358v1", "cited_id": "1207.6083", "section_title": "Dpp-Based Client Selection", "citation": "Note that DPP is a probabilistic model of repulsion, which has been widely adopted for solving subset sampling problems with diversity constraints in machine learning #REFR .", "text_before_citation": ["With the clients' data profiles, a DPP-based CS strategy can be further devised to avoid selecting similar clients in each round of training."], "text_after_citation": ["And the k-DPP is a variant of DPP, with which the size of sampled subsets is fixed at k #OTHEREFR .", "Particularly, a DPP is a probabilistic model over subsets on a finite set, which can be derived from a positive semi-definite similarity kernel matrix #OTHEREFR .", "For a finite set M with M elements, the similarity kernel matrix L can be expressed as L = {lm,n}M\u00d7M , with lm,n representing the similarity between the m-th and n-th elements in M.", "Meanwhile, the DPP assigns a probability to sub-sampling any subset Y of M, which is proportional to the determinant of the sub-matrix LY regarding the subset Y, i.e.,"], "citing_paper_content": {"title": "Dpp-Based Client Selection For Federated Learning With Non-Iid Data", "abstract": "This paper proposes a client selection (CS) method to tackle the communication bottleneck of federated learning (FL) while concurrently coping with FL's data heterogeneity issue. Specifically, we first analyze the effect of CS in FL and show that FL training can be accelerated by adequately choosing participants to diversify the training dataset in each round of training. Based on this, we leverage data profiling and determinantal point process (DPP) sampling techniques to develop an algorithm termed Federated Learning with DPP-based Participant Selection (FL-DP 3 S). This algorithm effectively diversifies the participants' datasets in each round of training while preserving their data privacy. We conduct extensive experiments to examine the efficacy of our proposed method. The results show that our scheme attains a faster convergence rate, as well as a smaller communication overhead than several baselines."}, "cited_paper_content": {"title": "Determinantal Point Processes For Machine Learning", "abstract": "Determinantal point processes (DPPs) are elegant probabilistic models of repulsion that arise in quantum physics and random matrix theory. In contrast to traditional structured models like Markov random fields, which become intractable and hard to approximate in the presence of negative correlations, DPPs offer efficient and exact algorithms for sampling, marginalization, conditioning, and other inference tasks. While they have been studied extensively by mathematicians, giving rise to a deep and beautiful theory, DPPs are relatively new in machine learning. Determinantal Point Processes for Machine Learning provides a comprehensible introduction to DPPs, focusing on the intuitions, algorithms, and extensions that are most relevant to the machine learning community, and shows how DPPs can be applied to real-world applications like finding diverse sets of high-quality search results, building informative summaries by selecting diverse sentences from documents, modeling non-overlapping human poses in images or video, and automatically building timelines of important news stories. It presents the general mathematical background to DPPs along with a range of modeling extensions, efficient algorithms, and theoretical results that aim to enable practical modeling and learning."}, "keywords": ["machine learning", "probabilistic model"], "citation_intent": "method"}
{"citing_id": "2303.04388v1", "cited_id": "1405.0312", "section_title": "Experimental Settings", "citation": "The dataset comprises 33k QA pairs from 28k images sourced from the COCO2014 dataset #REFR .", "text_before_citation": ["Dataset.", "For our experiments, we utilized the VQA-X dataset #OTHEREFR , an extension of the VQA-v2 dataset #OTHEREFR , which included explanations for each answer."], "text_after_citation": ["For the dataset division, we used all the images in the COCO2014 training set with 29k QA pairs as our training set.", "We divided the COCO2014 validation set into our validation set and test set according to the proportion of 3:4, which contains 1.5k QA pairs and 2k QA pairs, respectively.", "Compared to traditional vision models with specific tasks such as image classification and image segmentation, for the vision encoder, we only rely on their primary network function to output simple grid features rather than their time-consuming top-down features.", "Thus, to better adapt to the Vision & Language task, we used the CLIP based on the structure of the vision transformer as the vision encoder.", "The CLIP makes the fusion of vision and language features easier by encoding them in the same hidden space."], "citing_paper_content": {"title": "Interpretable Visual Question Answering Referring To Outside Knowledge", "abstract": "We present a novel multimodal interpretable VQA model that can answer the question more accurately and generate diverse explanations. Although researchers have proposed several methods that can generate human-readable and fine-grained natural language sentences to explain a model's decision, these methods have focused solely on the information in the image. Ideally, the model should refer to various information inside and outside the image to correctly generate explanations, just as we use background knowledge daily. The proposed method incorporates information from outside knowledge and multiple image captions to increase the diversity of information available to the model. The contribution of this paper is to construct an interpretable visual question answering model using multimodal inputs to improve the rationality of generated results. Experimental results show that our model can outperform state-ofthe-art methods regarding answer accuracy and explanation rationality."}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["dataset"], "citation_intent": "method"}
{"citing_id": "2304.06664v1", "cited_id": "1811.10879", "section_title": "Chou, Golovnev, Sudan, And Velusamy [Cgsv21B]", "citation": "Next, in \u00a75.2, we present results due to Chou, Golovnev, Sudan, Velingker, and Velusamy [CGS + 22] in the linear-space streaming setting, which generalize the result we've seen for Max-Cut (discussed in \u00a73.4.3, due to #REFR ).", "text_before_citation": ["Our primary goal is to articulate technical statements of these results, since we'll use them later in Chapters 6 and 7.", "We also give some broad-strokes discussions of the techniques involved, building on our work in the previous two chapters.", "Specifically, in \u00a75.1, we describe the results of Chou, Golovnev, Sudan, and Velusamy [CGSV21a; CGSV21b] on \u221a n-space streaming algorithms, which generalize the results we've already seen for Max-Cut (Theorem 3.1, due to #OTHEREFR ) and Max-DiCut (Theorem 4.1, due to #OTHEREFR ).", "They include a so-called dichotomy theorem, which completely characterizes CSP approximability for \u221a n-space sketching algorithms (see \u00a72.2) and builds on our \"template distribution\" analysis for Max-DiCut in Chapter 4.", "This dichotomy will later be the basis for the [BHP + 22] analysis of Max-BCSP(f ) problems for symmetric f : Z k 2 \u2192 {0, 1}, described in Chapter 7 below."], "text_after_citation": ["We will use these in Chapter 6 to prove linear-space streaming approximation-resistance results for so-called \"ordering constraint satisfaction problems\" from our joint work #OTHEREFR ."], "citing_paper_content": {"title": "On Streaming Approximation Algorithms For Constraint Satisfaction Problems", "abstract": "In this thesis, we explore streaming algorithms for approximating constraint satisfaction problems (CSPs). The setup is roughly the following: A computer has limited memory space, sees a long \"stream\" of local constraints on a set of variables, and tries to estimate how many of the constraints may be simultaneously satisfied. The past ten years have seen a number of works in this area, and this thesis includes both expository material and novel contributions. Throughout, we emphasize connections to the broader theories of CSPs, approximability, and streaming models, and highlight interesting open problems. The first part of our thesis is expository: We present aspects of previous works that completely characterize the approximability of specific CSPs like Max-Cut and Max-DiCut with \u221a n-space streaming algorithm (on n-variable instances), while characterizing the approximability of all CSPs in \u221a n space in the special case of \"composable\" (i.e., sketching) algorithms, and of a particular subclass of CSPs with linear-space streaming algorithms. In the second part of the thesis, we present two of our own joint works. We begin with a work with Madhu Sudan and Santhoshini Velusamy in which we prove linear-space streaming approximation-resistance for all ordering CSPs (OCSPs), which \"CSP-like\" problems maximizing over sets of permutations. Previous works considered the maximum acyclic subgraph problem (MAS), the prototypical OCSP; even for MAS, we improve on both the inapproximability factor and the space bound. Next, we present joint work with Joanna Boyland, Michael Hwang, Tarun Prasad, and Santhoshini Velusamy in which we investigate the \u221a n-space streaming approximability of Boolean CSPs with negations. In particular, we give explicit \u221a n-space sketching approximability ratios for several families of CSPs, including Max-kAND; develop simpler optimal sketching approximation algorithms for threshold predicates; and show that previous iii Thesis advisor: Prof. Madhu Sudan Noah Singer streaming lower bounds are \"incomplete\" in that they fail to characterize the \u221a n-space streaming approximability of Max-3AND."}, "cited_paper_content": {"title": "An Optimal Space Lower Bound For Approximating Max-Cut", "abstract": "We consider the problem of estimating the value of MAX-CUT in a graph in the streaming model of computation. At one extreme, there is a trivial 2-approximation for this problem that uses only O(log n) space, namely, count the number of edges and output half of this value as the estimate for the size of the MAX-CUT. On the other extreme, for any fixed \u0454 > 0, if one allows O(n) space, a (1+\u0454)-approximate solution to the MAX-CUT value can be obtained by storing an O(n)-size sparsifier that essentially preserves MAX-CUT value. Our main result is that any (randomized) single pass streaming algorithm that breaks the 2-approximation barrier requires \u03a9(n)-space, thus resolving the space complexity of any non-trivial approximations of the MAX-CUT value to within polylogarithmic factors in the single pass streaming model. We achieve the result by presenting a tight analysis of the Implicit Hidden Partition Problem introduced by Kapralov et al. [SODA\u201917] for an arbitrarily large number of players. In this problem a number of players receive random matchings of \u03a9(n) size together with random bits on the edges, and their task is to determine whether the bits correspond to parities of some hidden bipartition, or are just uniformly random. Unlike all previous Fourier analytic communication lower bounds, our analysis does not directly use bounds on the l2 norm of Fourier coefficients of a typical message at any given weight level that follow from hypercontractivity. Instead, we use the fact that graphs received by players are sparse (matchings) to obtain strong upper bounds on the l1 norm of the Fourier coefficients of the messages of individual players using their special structure, and then argue, using the convolution theorem, that similar strong bounds on the l1 norm are essentially preserved (up to an exponential loss in the number of players) once messages of different players are combined. We feel that our main technique is likely of independent interest."}, "keywords": ["linear-space streaming setting", "Max-Cut"], "citation_intent": "background"}
{"citing_id": "2304.13013v1", "cited_id": "1804.04235", "section_title": "Preliminaries And Related Work", "citation": "In contrast with Shazeer and Stern #REFR , who only observe instabilities without warmup, we observe instabilities despite a long warmup period.", "text_before_citation": ["Various solutions have been proposed, including freezing the embedding layer #OTHEREFR , adding additional layer normalization #OTHEREFR , or reparametrizing the weights #OTHEREFR .", "In our work we investigate instabilities which arise during CLIP training.", "Unlike the instabilities observed in #OTHEREFR , we find these are not caused by attention entropy collapse.", "Instead, our results indicate that spikes arise when the second moment estimator is out of date for the networks early layers.", "While our analysis and methods build directly on Shazeer and Stern #OTHEREFR (AdaFactor), there are important differences."], "text_after_citation": ["Moreover, in contrast with Shazeer and Stern #OTHEREFR we find that an out-of-date second moment estimator is primarily an issue for the (patch) embedding layer, and measure how well loss spikes are predicted by this event.", "Finally, we note that reserachers have moved away from AdaFactor in its original formulation for large-scale training #OTHEREFR , finding AdaFactor to under-perform AdamW #OTHEREFR .", "We believe this is due to the factored second moment or absence of first moment.", "This is why our focus is AdamW #OTHEREFR which is the de facto standard optimizer for transformers."], "citing_paper_content": {"title": "Stable And Low-Precision Training For Large-Scale Vision-Language Models", "abstract": "We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models. 1) Towards accelerating training, we introduce SwitchBack , a linear layer for int8 quantized training which provides a speed-up of 13-25% while matching the performance of bfloat16 training within 0.1 percentage points for the 1B parameter CLIP ViT-Huge-the largest int8 training to date. Our main focus is int8 as GPU support for float8 is rare, though we also analyze float8 training through simulation. While SwitchBack proves effective for float8, we show that standard techniques are also successful if the network is trained and initialized so that large feature magnitudes are discouraged, which we accomplish via layer-scale initialized with zeros. 2) Towards stable training, we analyze loss spikes and find they consistently occur 1-8 iterations after the squared gradients become underestimated by their AdamW second moment estimator. As a result, we recommend an AdamW-Adafactor hybrid, which we refer to as StableAdamW because it avoids loss spikes when training a CLIP ViT-Huge model and outperforms gradient clipping."}, "cited_paper_content": {"title": "Adafactor: Adaptive Learning Rates With Sublinear Memory Cost", "abstract": "In several recently proposed stochastic optimization methods (e.g. RMSProp, Adam, Adadelta), parameter updates are scaled by the inverse square roots of exponential moving averages of squared past gradients. Maintaining these per-parameter second-moment estimators requires memory equal to the number of parameters. For the case of neural network weight matrices, we propose maintaining only the per-row and per-column sums of these moving averages, and estimating the per-parameter second moments based on these sums. We demonstrate empirically that this method produces similar results to the baseline. Secondly, we show that adaptive methods can produce larger-than-desired updates when the decay rate of the second moment accumulator is too slow. We propose update clipping and a gradually increasing decay rate scheme as remedies. Combining these methods and dropping momentum, we achieve comparable results to the published Adam regime in training the Transformer model on the WMT 2014 English-German machine translation task, while using very little auxiliary storage in the optimizer. Finally, we propose scaling the parameter updates based on the scale of the parameters themselves."}, "keywords": ["contrast", "Stern"], "citation_intent": "result"}
{"citing_id": "2304.02838v1", "cited_id": "1910.00056", "section_title": "Related Work", "citation": "For example, Poirot #REFR threat detection is modeled as an imprecise graph pattern matching problem.", "text_before_citation": ["Provenance graphs have been used extensively in APT detection in recent years due to the superiority of connecting nodes with causal relationships and representing data flow and control flow relationships between system objects.", "Both academics and industry are paying increasing attention to this type of learning method #OTHEREFR .", "APT detection based on provenance graph is mainly divided into three directions: graph matching-based detection, anomaly score-based detection, and tag propagationbased detection.", "The first is graph matching-based detection.", "Due to the substructure in the provenance graph can completely describe malicious behavior, it is a very popular method to use graph matching-based detection method to detect APT attacks."], "text_after_citation": ["A graph matching method is proposed to identify attacks in provenance graphs.", "But it needs to construct the attack graph according to the prior knowledge, and cannot detect the unknown attack method.", "UNICORN #OTHEREFR is the first APT intrusion detection system to analyze the local complete system operation.", "The provenance graph is converted into a sequence of feature vectors and uses an automaton to model the clustering.", "However, the effectiveness of detection is negatively impacted by a large number of automatic opportunities, and the ability to describe the feature sequence is weak."], "citing_paper_content": {"title": "Tbdetector:Transformer-Based Detector For Advanced Persistent Threats With Provenance Graph", "abstract": "APT detection is difficult to detect due to the longterm latency, covert and slow multistage attack patterns of Advanced Persistent Threat (APT). To tackle these issues, we propose TBDetector, a transformer-based advanced persistent threat detection method for APT attack detection. Considering that provenance graphs provide rich historical information and have the powerful attacks historic correlation ability to identify anomalous activities, TBDetector employs provenance analysis for APT detection, which summarizes longrunning system execution with space efficiency and utilizes transformer with self-attention based encoder-decoder to extract long-term contextual features of system states to detect slow-acting attacks. Furthermore, we further introduce anomaly scores to investigate the anomaly of different system states, where each state is calculated with an anomaly score corresponding to its similarity score and isolation score. To evaluate the effectiveness of the proposed method, we have conducted experiments on five public datasets, i.e., streamspot, cadets, shellshock, clearscope, and wget baseline. Experimental results and comparisons with state-of-the-art methods have exhibited better performance of our proposed method."}, "cited_paper_content": {"title": "Poirot: Aligning Attack Behavior With Kernel Audit Records For Cyber Threat Hunting", "abstract": "Cyber threat intelligence (CTI) is being used to search for indicators of attacks that might have compromised an enterprise network for a long time without being discovered. To have a more effective analysis, CTI open standards have incorporated descriptive relationships showing how the indicators or observables are related to each other. However, these relationships are either completely overlooked in information gathering or not used for threat hunting. In this paper, we propose a system, called POIROT, which uses these correlations to uncover the steps of a successful attack campaign. We use kernel audits as a reliable source that covers all causal relations and information flows among system entities and model threat hunting as an inexact graph pattern matching problem. Our technical approach is based on a novel similarity metric which assesses an alignment between a query graph constructed out of CTI correlations and a provenance graph constructed out of kernel audit log records. We evaluate POIROT on publicly released real-world incident reports as well as reports of an adversarial engagement designed by DARPA, including ten distinct attack campaigns against different OS platforms such as Linux, FreeBSD, and Windows. Our evaluation results show that POIROT is capable of searching inside graphs containing millions of nodes and pinpoint the attacks in a few minutes, and the results serve to illustrate that CTI correlations could be used as robust and reliable artifacts for threat hunting."}, "keywords": ["Poirot threat detection"], "citation_intent": "background"}
{"citing_id": "2303.01724v1", "cited_id": "1911.05076", "section_title": "A. Hyperbolic Geometry", "citation": "At each x \u2208 D n c , there is a tangent space T x D n c , which can be viewed as the first-order approximation of the hyperbolic manifold at x #REFR .", "text_before_citation": ["A hyperbolic space is a non-Euclidean space with constant negative curvature.", "There are different but equivalent models to describe the same hyperbolic geometry.", "In this paper, we work with the Poincar\u00e9 ball model, in which all points are inside a ball.", "The hyperbolic space with constant negative curvature c is denoted by (D n c , g c x ).", "It consists of the n-dimensional hyperbolic manifold D n c = {x \u2208 R n : c x < 1} with the Riemannian metric g c x = (\u03bb c x ) 2 g E , where \u03bb c x = 2/(1 \u2212 c x 2 ) and g E = I n is the Euclidean metric."], "text_after_citation": ["The tangent space is then useful to perform Euclidean operations that we are familiar with but are undefined in hyperbolic spaces.", "A hyperbolic space and the tangent space at a point are connected through the exponential map exp c", "x :", "EQUATION", "EQUATION"], "citing_paper_content": {"title": "Node-Specific Space Selection Via Localized Geometric Hyperbolicity In Graph Neural Networks", "abstract": "Many graph neural networks have been developed to learn graph representations in either Euclidean or hyperbolic space, with all nodes' representations embedded in a single space. However, a graph can have hyperbolic and Euclidean geometries at different regions of the graph. Thus, it is suboptimal to indifferently embed an entire graph into a single space. In this paper, we explore and analyze two notions of local hyperbolicity, describing the underlying local geometry: geometric (Gromov) and model-based, to determine the preferred space of embedding for each node. The two hyperbolicities' distributions are aligned using the Wasserstein metric such that the calculated geometric hyperbolicity guides the choice of the learned model hyperbolicity. As such our model Joint Space Graph Neural Network (JSGNN) can leverage both Euclidean and hyperbolic spaces during learning by allowing node-specific geometry space selection. We evaluate our model on both node classification and link prediction tasks and observe promising performance compared to baseline models."}, "cited_paper_content": {"title": "Constant Curvature Graph Convolutional Networks", "abstract": "Interest has been rising lately towards methods representing data in non-Euclidean spaces, e.g. hyperbolic or spherical, that provide specific inductive biases useful for certain real-world data properties, e.g. scale-free, hierarchical or cyclical. However, the popular graph neural networks are currently limited in modeling data only via Euclidean geometry and associated vector space operations. Here, we bridge this gap by proposing mathematically grounded generalizations of graph convolutional networks (GCN) to (products of) constant curvature spaces. We do this by i) introducing a unified formalism that can interpolate smoothly between all geometries of constant curvature, ii) leveraging gyro-barycentric coordinates that generalize the classic Euclidean concept of the center of mass. Our class of models smoothly recover their Euclidean counterparts when the curvature goes to zero from either side. Empirically, we outperform Euclidean GCNs in the tasks of node classification and distortion minimization for symbolic data exhibiting non-Euclidean behavior, according to their discrete curvature."}, "keywords": ["hyperbolic manifold"], "citation_intent": "background"}
{"citing_id": "2303.11620v2", "cited_id": "1011.5553", "section_title": "Introduction", "citation": "An understanding of affine, global (Euclidean), or local (Euclidean) rigidity #REFR of a realization \u0398(S) has importance in several areas such as molecular dynamics [18; 4; 8] or sensor network localization [34; 32] .", "text_before_citation": ["#OTHEREFR Given an alignment S \u2208 O(d) m of the local views, a consensus representation \u0398(S) of the points can be obtained by averaging the local coordinates of the points due to the (rigidly transformed) views containing them.", "In the noiseless setting where the local views are clean measurements of the data (obtained by applying an unknown rigid transformation to a subset of data points), a perfect alignment of views is possible.", "Equivalently, when the views are noiseless, a value of zero for F is attainable, and an S that achieves it is called a \"perfect alignment\".", "Clearly, a perfect alignment is an optimal one, while the converse may not hold.", "To be consistent with previous works [12; 13] , the consensus representation of the points \u0398(S) due to a perfect alignment S of the views is called a realization of the framework."], "text_after_citation": ["Under a mild assumption on the structure of the local views, [3; 31; 12] characterized the affine rigidity of a realization by the rank of a certain matrix derived from the framework \u0398.", "It was shown in #OTHEREFR that deriving a similar characterization of global rigidity is NP-Hard.", "Nevertheless a characterization of the local rigidity is useful from an algorithmic standpoint as we show in this work.", "Furthermore, necessary and sufficient conditions on the overlapping structure of the local views for affine rigidity were derived in #OTHEREFR .", "Similar results in the context of local and global rigidity form our second set of contributions."], "citing_paper_content": {"title": "Non-Degenerate Rigid Alignment In A Patch Framework", "abstract": "Given a set of overlapping local views (patches) of a dataset, we consider the problem of finding a rigid alignment of the views that minimizes a 2-norm based alignment error. In general, the views are noisy and a perfect alignment may not exist. In this work, we characterize the non-degeneracy of an alignment in the noisy setting based on the kernel and positivity of a certain matrix. This leads to a polynomial time algorithm for testing the non-degeneracy of a given alignment. Consequently, we focus on Riemannian gradient descent for minimization of the error and obtain a sufficient condition on an alignment for the algorithm to converge (locally) linearly to it. In the case of noiseless views, a perfect alignment exists, resulting in a realization of the points that respects the geometry of the views. Under a mild condition on the views, we show that the non-degeneracy of a perfect alignment is equivalent to the local rigidity of the resulting realization. By specializing the characterization of a non-degenerate alignment to the noiseless setting, we obtain necessary and sufficient conditions on the overlapping structure of the views for a locally rigid realization. Similar results are also obtained in the context of global rigidity."}, "cited_paper_content": {"title": "On Affine Rigidity", "abstract": "We define the notion of affine rigidity of a hypergraph and prove a variety of fundamental results for this notion. First, we show that affine rigidity can be determined by the rank of a specific matrix which implies that affine rigidity is a generic property of the hypergraph.Then we prove that if a graph is is $(d+1)$-vertex-connected, then it must be \"generically neighborhood affinely rigid\" in $d$-dimensional space. This implies that if a graph is $(d+1)$-vertex-connected then any generic framework of its squared graph must be universally rigid. ::: Our results, and affine rigidity more generally, have natural applications in point registration and localization, as well as connections to manifold learning."}, "keywords": ["local (Euclidean) rigidity"], "citation_intent": "background"}
{"citing_id": "2303.11673v1", "cited_id": "1906.07413", "section_title": "Sampling Techniques", "citation": "It is found that data sampling may generate local models that over-fit (under-fit) for the minority (majority) classes #REFR .", "text_before_citation": ["In general, sampling based techniques are easier to operate compared with other groups of methods (e.g., algorithm-centered techniques) and do not require extensive expertise in ML.", "Thus, these methods have become popular when the models are operated by out-of-domain experts.", "Another strength of sampling based techniques is that they are agnostic to the core classification models #OTHEREFR .", "As such, these techniques can be incorporated into any classification model, and event an ensemble of multiple classification models.", "However, straightforward data sampling is not guaranteed to boost the classification accuracy, and sometimes may even worsen the model performance on one client #OTHEREFR ."], "text_after_citation": ["A possible explanation could be that naive data resampling may prevent clients from learning the \"Special Knowledge\" from local dataset in FL, which eventually leads to low accuracy #OTHEREFR .", "Therefore, sampling strategies need to be carefully designed for FL in order to handle class imbalance effectively.", "Therefore, many attempts have been made to propose sampling strategies that address data imbalance systematically in FL.", "In general, sampling can be done on either the data instance level (i.e., determining if one data instance should be involved in training) or the client level (i.e., determining if the data instances owned by one client should be involved in training).", "In this section, we categorize sampling techniques in FL into three groups and introduce them respectively: (1) data sampling; (2) client sampling; and (3) hybrid data and client sampling."], "citing_paper_content": {"title": "A Survey On Class Imbalance In Federated Learning", "abstract": "Federated learning, which allows multiple client devices in a network to jointly train a machine learning model without direct exposure of clients' data, is an emerging distributed learning technique due to its nature of privacy preservation. However, it has been found that models trained with federated learning usually have worse performance than their counterparts trained in the standard centralized learning mode, especially when the training data is imbalanced. In the context of federated learning, data imbalance may occur either locally one one client device, or globally across many devices. The complexity of different types of data imbalance has posed challenges to the development of federated learning technique, especially considering the need of relieving data imbalance issue and preserving data privacy at the same time. Therefore, in the literature, many attempts have been made to handle class imbalance in federated learning. In this paper, we present a detailed review of recent advancements along this line. We first introduce various types of class imbalance in federated learning, after which we review existing methods for estimating the extent of class imbalance without the need of knowing the actual data to preserve data privacy. After that, we discuss existing methods for handling class imbalance in FL, where the advantages and disadvantages of the these approaches are discussed. We also summarize common evaluation metrics for class imbalanced tasks, and point out potential future directions."}, "cited_paper_content": {"title": "Learning Imbalanced Datasets With Label-Distribution-Aware Margin Loss", "abstract": "Deep learning algorithms can fare poorly when the training dataset suffers from heavy class-imbalance but the testing criterion requires good generalization on less frequent classes. We design two novel methods to improve performance in such scenarios. First, we propose a theoretically-principled label-distribution-aware margin (LDAM) loss motivated by minimizing a margin-based generalization bound. This loss replaces the standard cross-entropy objective during training and can be applied with prior strategies for training with class-imbalance such as re-weighting or re-sampling. Second, we propose a simple, yet effective, training schedule that defers re-weighting until after the initial stage, allowing the model to learn an initial representation while avoiding some of the complications associated with re-weighting or re-sampling. We test our methods on several benchmark vision tasks including the real-world imbalanced dataset iNaturalist 2018. Our experiments show that either of these methods alone can already improve over existing techniques and their combination achieves even better performance gains."}, "keywords": ["minority (majority) classes"], "citation_intent": "background"}
{"citing_id": "2303.17251v1", "cited_id": "2001.10289", "section_title": "Misconception 4", "citation": "In yet some other cases, scientific results about bot activity were later found to largely match with independent platform removals of malicious accounts #REFR Tardelli et al., 2022) , which supports the correctness of the scientific results.", "text_before_citation": ["The previous misconceptions might induce readers to think that social bots research has led to flawed-if not outright useless-results, as claimed by some scholars #OTHEREFR , even in sensationalist terms #OTHEREFR .", "However, we see at least two strong arguments against this thesis.", "Firstly, in spite of the limitations of bot detectors, there have been multiple glaring examples of bot studies that were able to bring to light demonstrably harmful campaigns.", "For example, the detectors developed as part of some scientific endeavors were later deployed on online platforms and used to remove large numbers of malicious accounts #OTHEREFR .", "Similarly, results of some studies on bot activity led platforms to remove the accounts identified as malicious bots (Ferrara, 2022) ."], "text_after_citation": ["These cases represent but some of the success stories of social bots research.", "If researchers had not developed bot detection techniques, we would not have been able to identify pockets of anomalous accounts engaged in well-engineered malicious activities.", "Therefore, even if no universal bot detector still exists and in spite of the many caveats to consider in bot detection tasks, being able to detect some malicious bots puts us in an advantageous position than being able to detect none.", "Secondly, the benefits of social bots research extend beyond the detec-tion of malicious bots.", "For example, research and experimentation on social bots led to the development of neutral bots used for assessing the degree of political polarization on a platform #OTHEREFR , \"news bots\" used for journalistic purposes to curate, aggregate, and disseminate content collected from multiple sources #OTHEREFR , or even bots used for content moderation #OTHEREFR ."], "citing_paper_content": {"title": "Demystifying Misconceptions In Social Bots Research", "abstract": "The science of social bots seeks knowledge and solutions to one of the most debated forms of online misinformation. Yet, social bots research is plagued by widespread biases, hyped results, and misconceptions that set the stage for ambiguities, unrealistic expectations, and seemingly irreconcilable findings. Overcoming such issues is instrumental towards ensuring reliable solutions and reaffirming the validity of the scientific method. In this contribution we revise some recent results in social bots research, highlighting and correcting factual errors as well as methodological and conceptual issues. More importantly, we demystify common misconceptions, addressing fundamental points on how social bots research is discussed. Our analysis surfaces the need to discuss misinformation research in a rigorous, unbiased, and responsible way. This article bolsters such effort by identifying and refuting common fallacious arguments used by both proponents and opponents of social bots research as well as providing indications on the correct methodologies and sound directions for future research in the field."}, "cited_paper_content": {"title": "Charting The Landscape Of Online Cryptocurrency Manipulation", "abstract": "Cryptocurrencies represent one of the most attractive markets for financial speculation. As a consequence, they have attracted unprecedented attention on social media. Besides genuine discussions and legitimate investment initiatives, several deceptive activities have flourished. In this work, we chart the online cryptocurrency landscape across multiple platforms. To reach our goal, we collected a large dataset, composed of more than 50M messages published by almost 7M users on Twitter, Telegram and Discord, over three months. We performed bot detection on Twitter accounts sharing invite links to Telegram and Discord channels, and we discovered that more than 56% of them were bots or suspended accounts. Then, we applied topic modeling techniques to Telegram and Discord messages, unveiling two different deception schemes - \"pump-and-dump\" and \"Ponzi\" - and identifying the channels involved in these frauds. Whereas on Discord we found a negligible level of deception, on Telegram we retrieved 296 channels involved in pump-and-dump and 432 involved in Ponzi schemes, accounting for a striking 20% of the total. Moreover, we observed that 93% of the invite links shared by Twitter bots point to Telegram pump-and-dump channels, shedding light on a little-known social bot activity. Charting the landscape of online cryptocurrency manipulation can inform actionable policies to fight such abuse."}, "keywords": ["bot activity"], "citation_intent": "result"}
{"citing_id": "2305.02704v1", "cited_id": "1403.2079", "section_title": "C. Power Control For Secure Transmission 1) Background:", "citation": "Considering a two-link interference channel with one eavesdropper, the authors of #REFR devise a power control algorithm based on an altruistic and egoistic setting.", "text_before_citation": ["The aim of secure transmission is to send messages to legitimate receivers without any information leakage to eavesdroppers.", "Aside from the information-theoretical study #OTHEREFR - #OTHEREFR , the optimization aspect has attracted extensive research interest as well."], "text_after_citation": ["For a wiretap channel with one legitimate receiver, one eavesdropper, and multiple jammers, #OTHEREFR characterizes the power control as a Srtacjelberg game and thereby proposes an iterative optimization that achieves a Srtacjelberg equilibrium.", "Moreover, #OTHEREFR suggest a second-order-cone approach to power control for secure transmission in a cell-free MIMO network.", "For the UAV network case, #OTHEREFR proposes optimizing powers via block coordinate descent and successive convex approximation.", "The use of machine learning has become popular in this area, e.g., the deep reinforcement learning approach in #OTHEREFR and the Q-learning approach in #OTHEREFR .", "2) Problem Formulation: Consider L base-stations (BSs) each serving a legitimate downlink user terminal."], "citing_paper_content": {"title": "Mixed Max-And-Min Fractional Programming For Wireless Networks", "abstract": "Fractional programming (FP) plays a crucial role in wireless network design because many relevant problems involve maximizing or minimizing ratio terms. Notice that the maximization case and the minimization case of FP cannot be converted to each other in general, so they have to be dealt with separately in most of the previous studies. Thus, an existing FP method for maximizing ratios typically does not work for the minimization case, and vice versa. However, the FP objective can be mixed max-and-min, e.g., one may wish to maximize the signal-tointerference-plus-noise ratio (SINR) of the legitimate receiver while minimizing that of the eavesdropper. We aim to fill the gap between max-FP and min-FP by devising a unified optimization framework. The main results are threefold. First, we extend the existing max-FP technique called quadratic transform to the min-FP, and further develop a full generalization for the mixed case. Second. we provide a minorization-maximization (MM) interpretation of the proposed unified approach, thereby establishing its convergence and also obtaining a matrix extension; another result we obtain is a generalized Lagrangian dual transform which facilitates the solving of the logarithmic FP. Finally, we present three typical applications: the age-of-information (AoI) minimization, the Cram\u00e9r-Rao bound minimization for sensing, and the secure data rate maximization, none of which can be efficiently addressed by the previous FP methods. Index Terms-Multi-ratio mixed max-and-min fractional programming (FP); matrix FP; unified approach; age-of-information (AoI); Cram\u00e9r-Rao bound; sensing; secure transmission."}, "cited_paper_content": {"title": "Joint Power Control In Wiretap Interference Channels", "abstract": "Interference in wireless networks degrades the signal quality at the terminals. However, it can potentially enhance the secrecy rate. This paper investigates the secrecy rate in a two-user interference network where one of the users, namely user 1, requires to establish a confidential connection. User 1 wants to prevent an unintended user of the network to decode its transmission. User 1 has to transmit such that its secrecy rate is maximized while the quality of service at the destination of the other user, user 2, is satisfied, and both user's power limits are taken into account. We consider two scenarios: 1) user 2 changes its power in favor of user 1, an altruistic scenario, 2) user 2 is selfish and only aims to maintain the minimum quality of service at its destination, an egoistic scenario. It is shown that there is a threshold for user 2's transmission power that only below or above which, depending on the channel qualities, user 1 can achieve a positive secrecy rate. Closed-form solutions are obtained in order to perform joint optimal power control. Further, a new metric called secrecy energy efficiency is introduced. We show that in general, the secrecy energy efficiency of user 1 in an interference channel scenario is higher than that of an interference-free channel."}, "keywords": ["two-link interference channel"], "citation_intent": "method"}
{"citing_id": "2305.02528v1", "cited_id": "1906.05332", "section_title": "Results", "citation": "As shown in Table 1 , our self-supervised model achieves comparable performance with supervised HPLFlowNet #REFR on FT3D s dataset.", "text_before_citation": ["From the results, it can be found that our model can outperform all compared self-supervised methods in terms of the four metrics on the FT3D s test data. Especially, our model brings 8.72% gains for metric AS.", "For the KITTI s dataset, our model brings substantial improvements on all metrics.", "To be specific, our model outperforms the second best method RCP #OTHEREFR by 8.68% and 6.58% on metrics AS and AR, respectively.", "Besides, it is worth noting that our model can even achieve an EPE metric of 3.62cm, which is much lower than the EPE (6.19cm ) of recent RigidFlow.", "We also compare our model with some supervised methods, such as FlowNet3D #OTHEREFR and FLOT #OTHEREFR , etc."], "text_after_citation": ["Without any fine-tuning on KITTI s dataset, our model can even outperform the supervised methods listed in Table 1 , which proves that our model has better generalization ability.", "For real scenes, most local regions are with similar flow patterns.", "Thanks to dynamically clustering mechanism, our model clusters points with similar flow pattern into the same clusters and encodes the superpointlevel flow into the GRU for flow refinement, thereby leading to satisfactory performance on real scenes.", "Performance on point clouds with occlusions.", "Following the experimental settings used in Self-Point-Flow #OTHEREFR and RigidFlow #OTHEREFR , we train our model on KITTI r dataset and evaluate our model on both KITTI o and KITTI t datasets."], "citing_paper_content": {"title": "Self-Supervised 3D Scene Flow Estimation Guided By Superpoints", "abstract": "3D scene flow estimation aims to estimate point-wise motions between two consecutive frames of point clouds. Superpoints, i.e., points with similar geometric features, are usually employed to capture similar motions of local regions in 3D scenes for scene flow estimation. However, in existing methods, superpoints are generated with the offline clustering methods, which cannot characterize local regions with similar motions for complex 3D scenes well, leading to inaccurate scene flow estimation. To this end, we propose an iterative end-to-end superpoint based scene flow estimation framework, where the superpoints can be dynamically updated to guide the point-level flow prediction. Specifically, our framework consists of a flow guided superpoint generation module and a superpoint guided flow refinement module. In our superpoint generation module, we utilize the bidirectional flow information at the previous iteration to obtain the matching points of points and superpoint centers for soft point-to-superpoint association construction, in which the superpoints are generated for pairwise point clouds. With the generated superpoints, we first reconstruct the flow for each point by adaptively aggregating the superpoint-level flow, and then encode the consistency between the reconstructed flow of pairwise point clouds. Finally, we feed the consistency encoding along with the reconstructed flow into GRU to refine point-level flow. Extensive experiments on several different datasets show that our method can achieve promising performance. Code is available at https://github. com/supersyq/SPFlowNet."}, "cited_paper_content": {"title": "Hplflownet: Hierarchical Permutohedral Lattice Flownet For Scene Flow Estimation On Large-Scale Point Clouds", "abstract": "We present a novel deep neural network architecture for end-to-end scene flow estimation that directly operates on large-scale 3D point clouds. Inspired by Bilateral Convolutional Layers (BCL), we propose novel DownBCL, UpBCL, and CorrBCL operations that restore structural information from unstructured point clouds, and fuse information from two consecutive point clouds. Operating on discrete and sparse permutohedral lattice points, our architectural design is parsimonious in computational cost. Our model can efficiently process a pair of point cloud frames at once with a maximum of 86K points per frame. Our approach achieves state-of-the-art performance on the FlyingThings3D and KITTI Scene Flow 2015 datasets. Moreover, trained on synthetic data, our approach shows great generalization ability on real-world data and on different point densities without fine-tuning."}, "keywords": ["self-supervised model", "supervised HPLFlowNet"], "citation_intent": "result"}
{"citing_id": "2303.06565v1", "cited_id": "1904.08082", "section_title": "Graph Compressor", "citation": "The graph compressor is inspired by #REFR , and it works by computing the attention scores for all sentence nodes, filtering out nodes with the lowest scores, and then masking the rest using their attention scores.", "text_before_citation": ["Given G D and Q D from the graph encoder, the graph compressor aims to \"compress\" the graph by selecting a subset of salient nodes and edges.", "Here we focus on filtering the sentence nodes, because we want to identify key sentences that help generate the summary.", "After the compression, all selected sentence nodes and their linked word and document nodes represent the compressed graph and their embeddings will be used by the text decoder for summary generation."], "text_after_citation": ["Firstly, attention scores of the sentence nodes are calculated based on the updated node embeddings from our proposed graph encoder MGAT(Q D , G D ):", "EQUATION", "where r is the only trainable parameter of the graph compressor which transforms the updated node embedding into a scalar.", "Then, based on these scores, we select sentence nodes with the highest scores:", "EQUATION"], "citing_paper_content": {"title": "Compressed Heterogeneous Graph For Abstractive Multi-Document Summarization", "abstract": "Multi-document summarization (MDS) aims to generate a summary for a number of related documents. We propose HGSUM-an MDS model that extends an encoder-decoder architecture to incorporate a heterogeneous graph to represent different semantic units (e.g., words and sentences) of the documents. This contrasts with existing MDS models which do not consider different edge types of graphs and as such do not capture the diversity of relationships in the documents. To preserve only key information and relationships of the documents in the heterogeneous graph, HGSUM uses graph pooling to compress the input graph. And to guide HGSUM to learn the compression, we introduce an additional objective that maximizes the similarity between the compressed graph and the graph constructed from the ground-truth summary during training. HGSUM is trained end-to-end with the graph similarity and standard cross-entropy objectives. Experimental results over MULTI-NEWS, WCEP-100, and ARXIV show that HGSUM outperforms state-of-the-art MDS models. The code for our model and experiments is available at: https://github.com/oaimli/HGSum."}, "cited_paper_content": {"title": "Self-Attention Graph Pooling", "abstract": "Advanced methods of applying deep learning to structured data such as graphs have been proposed in recent years. In particular, studies have focused on generalizing convolutional neural networks to graph data, which includes redefining the convolution and the downsampling (pooling) operations for graphs. The method of generalizing the convolution operation to graphs has been proven to improve performance and is widely used. However, the method of applying downsampling to graphs is still difficult to perform and has room for improvement. In this paper, we propose a graph pooling method based on self-attention. Self-attention using graph convolution allows our pooling method to consider both node features and graph topology. To ensure a fair comparison, the same training procedures and model architectures were used for the existing pooling methods and our method. The experimental results demonstrate that our method achieves superior graph classification performance on the benchmark datasets using a reasonable number of parameters."}, "keywords": ["graph compressor"], "citation_intent": "method"}
{"citing_id": "2304.13559v2", "cited_id": "1910.13461", "section_title": "Discussion", "citation": "While text-to-table, for example, also uses pre-trained language models as a basis, it is directly initialized with the pre-trained weights of a language model (i.e., BART #REFR ), which has been pre-trained on plain text only.", "text_before_citation": ["While there exist already approaches that transform textual data to tables like text-to-table #OTHEREFR or STable #OTHEREFR , we think that these approaches are not suitable for implementing multi-modal database operations in an MMDB.", "One key difference of our approach compared to text-to-table and STable is that the underlying models have to be trained from scratch for every new text collection."], "text_after_citation": ["Different from that, we carefully design a new pre-training procedure that allows our MMDB-Model to provide more meaningful representations for table extraction and thus realize MMOps on unseen texts with just a few examples.", "Moreover, another key difference is that text-to-table or STable produce the data for an output table using a transformer-based generative decoder #OTHEREFR .", "The first disadvantage here is that the model can \"make up\" values, that are not actually present in the input text, but that the model picked up during (pre-)training. This phenomenon is called hallucination #OTHEREFR .", "On top of that, transformer-based decoders output tables token-by-token in an autoregressive manner, which requires a pass through the decoder for every token in the output table.", "This results in a computationally expensive decoding process as we show in our evaluation."], "citing_paper_content": {"title": "Towards Multi-Modal Dbmss For Seamless Querying Of Texts And Tables", "abstract": "In this paper, we propose Multi-Modal Databases (MMDBs), which is a new class of database systems that can seamlessly query text and tables using SQL. To enable seamless querying of textual data using SQL in an MMDB, we propose to extend relational databases with so-called multi-modal operators (MMOps) which are based on the advances of recent large language models such as GPT-3. The main idea of MMOps is that they allow text collections to be treated as tables without the need to manually transform the data. As we show in our evaluation, our MMDB prototype can not only outperform state-of-the-art approaches such as text-totable in terms of accuracy and performance but it also requires significantly less training data to fine-tune the model for an unseen text collection."}, "cited_paper_content": {"title": "Bart: Denoising Sequence-To-Sequence Pre-Training For Natural Language Generation, Translation, And Comprehension", "abstract": "We present BART, a denoising autoencoder for pretraining sequence-to-sequence models. BART is trained by (1) corrupting text with an arbitrary noising function, and (2) learning a model to reconstruct the original text. It uses a standard Tranformer-based neural machine translation architecture which, despite its simplicity, can be seen as generalizing BERT (due to the bidirectional encoder), GPT (with the left-to-right decoder), and many other more recent pretraining schemes. We evaluate a number of noising approaches, finding the best performance by both randomly shuffling the order of the original sentences and using a novel in-filling scheme, where spans of text are replaced with a single mask token. BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE. BART also provides a 1.1 BLEU increase over a back-translation system for machine translation, with only target language pretraining. We also report ablation experiments that replicate other pretraining schemes within the BART framework, to better measure which factors most influence end-task performance."}, "keywords": ["language model", "pre-trained language models"], "citation_intent": "method"}
{"citing_id": "2305.02109v1", "cited_id": "1909.11875", "section_title": "B. Open Issues And Challenges", "citation": "Existing FL implementations generally consider static strategies to allocate spectrum (i.e., PRBs) to FLUs for an entire LTR #REFR , leading to underutilization of spectrum, called spectrum fragmentation.", "text_before_citation": ["Due to temporal variation of channels, static resource allocation (e.g., licensed/unlicensed spectrum and transmit power) of current FL implementations #OTHEREFR can lead to two obscure problems.", "(i) Wireless resource over-provisioning (WROP), referring to surpassing FLUs' QoS requirements upon over-provisioning of wireless resources, causing an increase in expenditure, energy consumption, and interference.", "(ii) Wireless resources under provisioning (WRUP), referring to the wireless resource deficiency of FLUs at some time instants, which can cause SLA violations for FLSs.", "WRUP can reduce FLSOs satisfaction, preventing them from running their FLSs in the future.", "To harness WROP/WRUP, we propose two mechanisms in EV-FL to dynamically match resources to FLSOs' de- 3) Spectrum fragmentation: In FL, a local training round (LTR) of FLUs consists of (i) GM downloading, (ii) LM training, and (iii) LM uploading."], "text_after_citation": ["To our knowledge, we are among the first to identify spectrum fragmentation in FL, meaning that FLUs only utilize wireless resources for GM downloading and LM uploading, while these resources are idle during LM training.", "Spectrum fragmentation under DCC. Spectrum fragmentation becomes more severe upon considering DCC.", "In DCC, we have two types of LTRs (see Fig.", "2 ): (i) DPU LTR, consisting of three steps: (i-a) GM downloading, (i-b) LM training, and (i-c) dispersing the trained LM to neighboring CHUs.", "(ii) CHU LTR, consisting of five steps: (ii-a) GM downloading, (ii-b) LM training, (ii-c) waiting for DPUs connected to the CHU to perform DPU LTR, (ii-d) performing local aggregation, and (ii-e) aggregated LM uploading to the BS."], "citing_paper_content": {"title": "Synergies Between Federated Learning And O-Ran: Towards An Elastic Virtualized Architecture For Multiple Distributed Machine Learning Services", "abstract": "Federated learning (FL) is the most popular distributed machine learning technique. However, implementation of FL over modern wireless networks faces key challenges caused by (i) dynamics of the network conditions, (ii) coexistence of multiple FL services/tasks in the system, and (iii) concurrent execution of FL services with other network services, which are not jointly considered in prior works. Motivated by these challenges, we introduce a generic FL paradigm over next-generation (NextG) networks, called dynamic multi-service FL (DMS-FL). We identify three unexplored design considerations in DMS-FL: (i) FL service operator accumulation, (ii) wireless resource fragmentation, and (iii) signal strength fluctuations. We take the first steps towards addressing these design considerations through proposing a novel distributed ML architecture called elastic virtualized FL (EV-FL). EV-FL unleashes the full potential of Open RAN (O-RAN) systems and introduces an elastic resource provisioning methodology to execute FL services. It further constitutes a multi-timescale FL management system that introduces three dimensions into existing FL architectures: (i) virtualization, (ii) scalability, and (iii) elasticity. Through investigating EV-FL, we reveal a series of open research directions for future work. We finally simulate EV-FL to demonstrate its potential to save wireless resources and increase fairness among FL services."}, "cited_paper_content": {"title": "Federated Learning In Mobile Edge Networks: A Comprehensive Survey", "abstract": "In recent years, mobile devices are equipped with increasingly advanced sensing and computing capabilities. Coupled with advancements in Deep Learning (DL), this opens up countless possibilities for meaningful applications. Traditional Machine Learning (ML) approaches require the data to be centralized in a cloud server. However, this results in critical issues related to unacceptable latency and communication inefficiency. To this end, Mobile Edge Computing (MEC) has been proposed to bring intelligence closer to the edge, where data is produced. However, conventional enabling technologies for ML at mobile edge networks still require personal data to be shared with external parties, e.g., edge servers. Recently, in light of increasingly stringent data privacy legislation and growing privacy concerns, the concept of Federated Learning (FL) has been introduced. In FL, end devices use their local data to train an ML model required by the server. The end devices then send the model updates rather than raw data to the server for aggregation. FL can serve as an enabling technology in mobile edge networks since it enables the collaborative training of an ML model and also enables DL for mobile edge network optimization. However, in a large-scale and complex mobile edge network, heterogeneous devices with varying constraints are involved. This raises challenges of communication costs, resource allocation, and privacy and security in the implementation of FL at scale. In this survey, we begin with an introduction to the background and fundamentals of FL. Then, we highlight the aforementioned challenges of FL implementation and review existing solutions. Furthermore, we present the applications of FL for mobile edge network optimization. Finally, we discuss the important challenges, open issues and future research directions in FL."}, "keywords": ["Existing FL implementations", "underutilization"], "citation_intent": "background"}
{"citing_id": "2303.00052v1", "cited_id": "1906.02642", "section_title": "Minimum Spanning Tree Games And Approximation", "citation": "From the more theoretical perspective, one can easily see that for MST games the cost function c( \u2022 ) is subadditive, yet it is not submodular in general; see #REFR for a characterization when it is.", "text_before_citation": ["In this section we address a well known special class of games known as minimum cost spanning tree (MST) games #OTHEREFR , where the cost of the outside option for a set of agents is determined by the cost of a minimum cost spanning tree in the subgraph induced by these agents. MST games are known to have a non-empty core.", "Moreover, it is known that finding some element in the core is computationally easy and can be done by computing a minimum spanning tree #OTHEREFR .", "The optimization problem that we address asks for the maximal amount that can be charged to the agents while no proper subset of agents would prefer the outside option."], "text_after_citation": ["Recalling that the computation of maximum shareable costs can be done in polynomial time when c( \u2022 ) is submodular, it is a natural question to ask if this still holds for subadditive cost functions.", "In that respect, note that the weighted graph games as studied in #OTHEREFR have polynomial time algorithms to decide non-emptiness of the core whenever c( \u2022 ) is subadditive, hence Theorem 4 applied to weighted graphs games does not give an answer to this question.", "We will see that for MST games, maximizing shareable costs cannot be done efficiently unless P=NP.", "In fact, one could define an even more general class of problems in the spirit of cooperative games with restricted coalition formation, by defining a (downward-closed) set system that describes all those subsets of agents that are able to cooperate and hence have access to an outside option, while all other subsets do not have that option.", "The almost core as studied in this paper arises as the special case where the set system is the (n \u2212 1)-uniform matroid."], "citing_paper_content": {"title": "Algorithmic Solutions For Maximizing Shareable Costs", "abstract": "This paper addresses the optimization problem to maximize the total costs that can be shared among a group of agents, while maintaining stability in the sense of the core constraints of a cooperative transferable utility game, or TU game. This means that all subsets of agents have an outside option at a certain cost, and stability requires that the cost shares are defined so that none of the outside options is preferable. When maximizing total shareable costs, the cost shares must satisfy all constraints that define the core of a TU game, except for being budget balanced. The paper gives a fairly complete picture of the computational complexity of this optimization problem, in relation to classical computational problems on the core. We also show that, for games with an empty core, the problem is equivalent to computing minimal core relaxations for several relaxations that have been proposed earlier. As an example for a class of cost sharing games with non-empty core, we address minimum cost spanning tree games. While it is known that cost shares in the core can be found efficiently, we show that the computation of maximal cost shares is NP-hard for minimum cost spanning tree games. We also derive a 2-approximation algorithm. Our work opens several directions for future work."}, "cited_paper_content": {"title": "An Efficient Characterization Of Submodular Spanning Tree Games", "abstract": "Cooperative games are an important class of problems in game theory, where the goal is to distribute a value among a set of players who are allowed to cooperate by forming coalitions. An outcome of the game is given by an allocation vector that assigns a value share to each player. A crucial aspect of such games is submodularity (or convexity). Indeed, convex instances of cooperative games exhibit several nice properties, e.g. regarding the existence and computation of allocations realizing some of the most important solution concepts proposed in the literature. For this reason, a relevant question is whether one can give a polynomial time characterization of submodular instances, for prominent cooperative games that are in general non-convex."}, "keywords": ["cost function", "MST games"], "citation_intent": "background"}
{"citing_id": "2304.07775v1", "cited_id": "1610.09001", "section_title": "Introduction", "citation": "The aforementioned audio-visual distillation methods #REFR highly relate the cross-modal synchronization to semantic shows that we first erase the irrelevant modality noise.", "text_before_citation": ["Knowledge distillation has become an effective approach to ensemble different data sources to enrich the representation ability of target modality #OTHEREFR . Compared with visual-related modalities (e.g.", "RGB and optical flow), sound naturally contains vivid supervisory information from an independent auditory source, which deserves thorough exploration due to its synchronized nature with vision.", "Inspired by this, crossmodal knowledge distillation has been extensively investigated in audio-visual learning #OTHEREFR , leveraging the synchronization between auditory and visual modalities."], "text_after_citation": ["(c) illustrates that we further capture the audio-visual correspondence and rectify the semantic correlation. consistency, and directly transfer knowledge across the audiovisual modalities.", "However, such semantic consistency is hard to guarantee in unconstrained videos, thus, directly transferring knowledge across audio-visual modalities assuming highly-related semantic consistency could harm the distillation performance.", "The similar phenomenon of semantic misalignment is also found in self-supervised multi-modal learning #OTHEREFR but got little attention in cross-modal distillation.", "In this work, we reconsider the conventional semantic consistency assumption and point out that the irrelevant modality noise and differentiated semantic correlation are blamed for the failure of such assumption and harm the crossmodal distillation performance, especially in unconstrained audio-visual scenarios.", "Concretely, in cross-modal distillation, irrelevant modality noise is considered as the signals in one modality that are unrelated to its accompanying target modality."], "citing_paper_content": {"title": "Robust Cross-Modal Knowledge Distillation For Unconstrained Videos", "abstract": "Cross-modal distillation has been widely used to transfer knowledge across different modalities, enriching the representation of the target unimodal one. Recent studies highly relate the temporal synchronization between vision and sound to the semantic consistency for cross-modal distillation. However, such semantic consistency from the synchronization is hard to guarantee in unconstrained videos, due to the irrelevant modality noise and differentiated semantic correlation. To this end, we first propose a Modality Noise Filter (MNF) module to erase the irrelevant noise in teacher modality with cross-modal context. After this purification, we then design a Contrastive Semantic Calibration (CSC) module to adaptively distill useful knowledge for target modality, by referring to the differentiated sample-wise semantic correlation in a contrastive fashion. Extensive experiments show that our method could bring a performance boost compared with other distillation methods in both visual action recognition and video retrieval task. We also extend to the audio tagging task to prove the generalization of our method."}, "cited_paper_content": {"title": "Soundnet: Learning Sound Representations From Unlabeled Video", "abstract": "We learn rich natural sound representations by capitalizing on large amounts of unlabeled sound data collected in the wild. We leverage the natural synchronization between vision and sound to learn an acoustic representation using two-million unlabeled videos. Unlabeled video has the advantage that it can be economically acquired at massive scales, yet contains useful signals about natural sound. We propose a student-teacher training procedure which transfers discriminative visual knowledge from well established visual recognition models into the sound modality using unlabeled video as a bridge. Our sound representation yields significant performance improvements over the state-of-the-art results on standard benchmarks for acoustic scene/object classification. Visualizations suggest some high-level semantics automatically emerge in the sound network, even though it is trained without ground truth labels."}, "keywords": ["aforementioned audio-visual distillation"], "citation_intent": "method"}
{"citing_id": "2303.13809v1", "cited_id": "1904.09675", "section_title": "Meta Evaluation", "citation": "BERTScore #REFR ) is a neural metric that relies on pre-trained models to compute the semantic similarity with the reference.", "text_before_citation": ["We utilize the accuracy of pairwise system-ranking (Kocmi et al., 2021) of three types of Kendall correlation.", "Specifically, these values are computed by flattening the scores into a single vector and calculating the average correlations over systems, or over segments.", "Baseline We compare LLMs with several commonly used baseline metrics for MT evaluation.", "BLEU #OTHEREFR is the most popular metric that compares the n-gram overlap of the translation with human reference, but it has been criticized for not capturing the full semantic meaning of the translation #OTHEREFR ."], "text_after_citation": ["BLEURT #OTHEREFR and COMET #OTHEREFR are supervised neural metrics that leverage human judgments to train. They have shown a high correlation with human judgments."], "citing_paper_content": {"title": "Error Analysis Prompting Enables Human-Like Translation Evaluation In Large Language Models: A Case Study On Chatgpt", "abstract": "Generative large language models (LLMs), e.g., ChatGPT, have demonstrated remarkable proficiency across several NLP tasks such as machine translation, question answering, text summarization, and natural language understanding. Recent research (Kocmi and Federmann, 2023) has shown that utilizing Chat-GPT for assessing the quality of machine translation (MT) achieves state-of-the-art performance at the system level but performs poorly at the segment level. To further improve the performance of LLMs on MT quality assessment, we conducted an investigation into several prompting methods. Our results indicate that by combining Chain-of-Thoughts (Wei et al., 2022) and Error Analysis (Lu et al., 2022), a new prompting method called Error Analysis Prompting, LLMs like ChatGPT can generate human-like MT evaluations at both the system and segment level. Additionally, we discovered some limitations of Chat-GPT as an MT evaluator, such as unstable scoring and biases when provided with multiple translations in a single query. Our findings aim to provide a preliminary experience for appropriately evaluating translation quality on Chat-GPT while offering a variety of tricks in designing prompts for in-context learning. We anticipate that this report will shed new light on advancing the field of translation evaluation with LLMs by enhancing both the accuracy and reliability of metrics."}, "cited_paper_content": {"title": "Bertscore: Evaluating Text Generation With Bert", "abstract": "We propose BERTScore, an automatic evaluation metric for text generation. Analogously to common metrics, BERTScore computes a similarity score for each token in the candidate sentence with each token in the reference sentence. However, instead of exact matches, we compute token similarity using contextual embeddings. We evaluate using the outputs of 363 machine translation and image captioning systems. BERTScore correlates better with human judgments and provides stronger model selection performance than existing metrics. Finally, we use an adversarial paraphrase detection task and show that BERTScore is more robust to challenging examples compared to existing metrics."}, "keywords": ["semantic similarity"], "citation_intent": "method"}
{"citing_id": "2304.11697v1", "cited_id": "1911.10150", "section_title": "B. Multi-Modal Fusion For Object Detection 1) Multi-Modal Object Detection:", "citation": "PointPainting #REFR projects LiDAR points into the output of an image-only semantic segmentation network, and appends the class scores to each point.", "text_before_citation": ["To date, several studies have investigated multi-modal fusion for 2D and 3D object detection.", "Frustum PointNets #OTHEREFR extract the 3D bounding frustum of an object by extruding 2D bounding boxes from image detectors.", "PointFusion #OTHEREFR combines a CNN and a PointNet #OTHEREFR architecture respectively to process images and raw point clouds then predict 3D boxes."], "text_after_citation": ["All these fusion methods of RGB and LiDAR achieve high average precision on the benchmarks, however, the coupling or interrelation of two modalities will cause the whole system to fail easily once part of the sensors break down.", "Besides, the methods above only provide a deterministic predict result, making it risky to carry out in the real application.", "2) Adaptive fusion: Several new studies have proposed self-adaptive techniques in computer vision.", "Therefore, the robustness of those tasks can be improved to some extent.", "Adaptnet #OTHEREFR uses a convoluted mixture of deep experts(CMoDE) fusion techniques to learn features from complementary modalities and spectra."], "citing_paper_content": {"title": "Informative Data Selection With Uncertainty For Multi-Modal Object Detection", "abstract": "Noise has always been nonnegligible trouble in object detection by creating confusion in model reasoning, thereby reducing the informativeness of the data. It can lead to inaccurate recognition due to the shift in the observed pattern, that requires a robust generalization of the models. To implement a general vision model, we need to develop deep learning models that can adaptively select valid information from multi-modal data. This is mainly based on two reasons. Multi-modal learning can break through the inherent defects of single-modal data, and adaptive information selection can reduce chaos in multi-modal data. To tackle this problem, we propose a universal uncertainty-aware multi-modal fusion model. It adopts a multi-pipeline loosely coupled architecture to combine the features and results from point clouds and images. To quantify the correlation in multimodal information, we model the uncertainty, as the inverse of data information, in different modalities and embed it in the bounding box generation. In this way, our model reduces the randomness in fusion and generates reliable output. Moreover, we conducted a completed investigation on the KITTI 2D object detection dataset and its derived dirty data. Our fusion model is proven to resist severe noise interference like Gaussian, motion blur, and frost, with only slight degradation. The experiment results demonstrate the benefits of our adaptive fusion. Our analysis on the robustness of multi-modal fusion will provide further insights for future research."}, "cited_paper_content": {"title": "Pointpainting: Sequential Fusion For 3D Object Detection", "abstract": "Camera and lidar are important sensor modalities for robotics in general and self-driving cars in particular. The sensors provide complementary information offering an opportunity for tight sensor-fusion. Surprisingly, lidar-only methods outperform fusion methods on the main benchmark datasets, suggesting a gap in the literature. In this work, we propose PointPainting: a sequential fusion method to fill this gap. PointPainting works by projecting lidar points into the output of an image-only semantic segmentation network and appending the class scores to each point. The appended (painted) point cloud can then be fed to any lidar-only method. Experiments show large improvements on three different state-of-the art methods, Point-RCNN, VoxelNet and PointPillars on the KITTI and nuScenes datasets. The painted version of PointRCNN represents a new state of the art on the KITTI leaderboard for the bird's-eye view detection task. In ablation, we study how the effects of Painting depends on the quality and format of the semantic segmentation output, and demonstrate how latency can be minimized through pipelining."}, "keywords": ["image-only semantic segmentation", "LiDAR points"], "citation_intent": "method"}
{"citing_id": "2304.00359v1", "cited_id": "1905.05172", "section_title": "Implementation Details", "citation": "For each training subject, we sample (a) 5,000 surface points as G s , which are evenly distributed on the mesh surface; (b) 5,000 occupancy-target points as G o in the 3D space, for which we follow the sampling strategy in #REFR .", "text_before_citation": [], "text_after_citation": ["In particular, 15/16 points in G o are evenly sampled on the mesh surface.", "Gaussian perturbation is then applied to them along the surface normal direction.", "The rest 1/16 points are randomly sampled within the predefined 3D space where the mesh lies in.", "We employ Rembg [1] to segment out the background for real-world images and use Kaolin [2] to calculate the SDF from the SMPL-X model.", "We apply PIXIE #OTHEREFR to estimate the SMPL-X model for the segmented human subject."], "citing_paper_content": {"title": "Sesdf: Self-Evolved Signed Distance Field For Implicit 3D Clothed Human Reconstruction", "abstract": "Figure 1. Self-evolved Signed Distance Field (SeSDF): We propose Self-evolved Signed Distance Field (SeSDF), which can flexibly reconstruct 3D clothed human models from a single-view image (top) or uncalibrated multi-view images (bottom). SeSDF can robustly recover fine geometry details from any type of poses, which allows us to generate clothed human avatars."}, "cited_paper_content": {"title": "Pifu: Pixel-Aligned Implicit Function For High-Resolution Clothed Human Digitization", "abstract": "We introduce Pixel-aligned Implicit Function (PIFu), a highly effective implicit representation that locally aligns pixels of 2D images with the global context of their corresponding 3D object. Using PIFu, we propose an end-to-end deep learning method for digitizing highly detailed clothed humans that can infer both 3D surface and texture from a single image, and optionally, multiple input images. Highly intricate shapes, such as hairstyles, clothing, as well as their variations and deformations can be digitized in a unified way. Compared to existing representations used for 3D deep learning, PIFu can produce high-resolution surfaces including largely unseen regions such as the back of a person. In particular, it is memory efficient unlike the voxel representation, can handle arbitrary topology, and the resulting surface is spatially aligned with the input image. Furthermore, while previous techniques are designed to process either a single image or multiple views, PIFu extends naturally to arbitrary number of views. We demonstrate high-resolution and robust reconstructions on real world images from the DeepFashion dataset, which contains a variety of challenging clothing types. Our method achieves state-of-the-art performance on a public benchmark and outperforms the prior work for clothed human digitization from a single image."}, "keywords": ["3D space", "mesh surface"], "citation_intent": "method"}
{"citing_id": "2303.04544v1", "cited_id": "1809.00549", "section_title": "Social Reasoning Facilitating Communication", "citation": "The obverter mechanism was implemented in modern neural networks by Choi, Lazaridou, and de Freitas (2018) and #REFR , who also observed an increase in compositionality.", "text_before_citation": ["The obverter mechanism makes use of the same signal-meaning mapping for language generation and comprehension.", "This guarantees that agents are internally consistent and use the language they learn in the receiver role when they assume the speaker role.", "It can be interpreted as a rudimentary form of social reasoning: an agent's own cognitive capabilities serve as the model of another agent.", "The symmetry of the roles, together with the selection of speaker-receiver pairs at random from a population of agents, effectively creates a transmission chain similar to the iterated learning setting #OTHEREFR .", "This should constitute a pressure for a systematic communication protocol."], "text_after_citation": ["The more complex mechanism of social reasoning was explored in #OTHEREFR .", "Their research belongs to the domain of artificial intelligence and multi-agent cooperation.", "They introduced the concept of social influence as an auxiliary reward for agents that communicate to solve a sequential social dilemma.", "Social influence was measured using counterfactual reasoning: an agent considered alternative signals it could have sent and assessed their likely effect on the actions of another agent.", "In one variant of their simulations, actions of the other agent were predicted using a specialised internal model trained in a supervised manner."], "citing_paper_content": {"title": "Models Of Symbol Emergence In Communication: A Conceptual Review And A Guide For Avoiding Local Minima", "abstract": "Computational simulations are a popular method for testing hypotheses about the emergence of communication. This kind of research is performed in a variety of traditions including language evolution, developmental psychology, cognitive science, machine learning, robotics, etc. The motivations for the models are different, but the operationalizations and methods used are often similar. We identify the assumptions and explanatory targets of several most representative models and summarise the known results. We claim that some of the assumptions-such as portraying meaning in terms of mapping, focusing on the descriptive function of communication, modelling signals with amodal tokens-may hinder the success of modelling. Relaxing these assumptions and foregrounding the interactions of embodied and situated agents allows one to systematise the multiplicity of pressures under which symbolic systems evolve. In line with this perspective, we sketch the road towards modelling the emergence of meaningful symbolic communication, where symbols are simultaneously grounded in action and perception and form an abstract system."}, "cited_paper_content": {"title": "Emergence Of Communication In An Interactive World With Consistent Speakers", "abstract": "Training agents to communicate with one another given task-based supervision only has attracted considerable attention recently, due to the growing interest in developing models for human-agent interaction. Prior work on the topic focused on simple environments, where training using policy gradient was feasible despite the non-stationarity of the agents during training. In this paper, we present a more challenging environment for testing the emergence of communication from raw pixels, where training using policy gradient fails. We propose a new model and training algorithm, that utilizes the structure of a learned representation space to produce more consistent speakers at the initial phases of training, which stabilizes learning. We empirically show that our algorithm substantially improves performance compared to policy gradient. We also propose a new alignment-based metric for measuring context-independence in emerged communication and find our method increases context-independence compared to policy gradient and other competitive baselines."}, "keywords": ["compositionality", "modern neural networks"], "citation_intent": "method"}
{"citing_id": "2304.02834v2", "cited_id": "1712.02463", "section_title": "C. Corrupted Input Detection", "citation": "CURE-TSR #REFR is a traffic sign recognition dataset that includes real-world and simulated challenging conditions of 12 types and five severity levels.", "text_before_citation": ["In addition to the widely accepted OOD and adversarial detection setup, we consider corrupted inputs as another type of anomaly.", "Deployed in the real world, neural networks are known to suffer from imperfect samples due to the data acquisition process and environmental factors, such as motion blur or weather conditions #OTHEREFR - #OTHEREFR .", "We use image classification datasets designed to benchmark the robustness of neural networks under realistic challenging conditions.", "CIFAR-10-C #OTHEREFR consists of 19 diverse corruption types in four categories, including noise, blur, weather, and digital, at five different severity levels that are applied to test images of CIFAR-10 dataset."], "text_after_citation": ["For each dataset, a ResNet model is trained on corruptionfree images, and the gradients are collected from pristine images in the test sets and their corrupted versions.", "We utilize the Mahalanobis method #OTHEREFR for comparison because it showed the best performance among all the compared methods for OOD detection and adversarial detection.", "In the experiments, we observed that the AUROC scores are highly saturated for both methods in many cases, particularly for CIFAR-10-C dataset, which calls for a more comprehensive comparison.", "To better facilitate the performance comparison, we employ corrected repeated k-fold crossvalidated (CV) paired t-test #OTHEREFR as a measure of statistical significance.", "Paired t-test is a statistical test for comparing two different learning schemes, A and B, based on a number of observations; in this case, predictive accuracy a and b."], "citing_paper_content": {"title": "Probing The Purview Of Neural Networks Via Gradient Analysis", "abstract": "We analyze the data-dependent capacity of neural networks and assess anomalies in inputs from the perspective of networks during inference. The notion of data-dependent capacity allows for analyzing the knowledge base of a model populated by learned features from training data. We define purview as the additional capacity necessary to characterize inference samples that differ from the training data. To probe the purview of a network, we utilize gradients to measure the amount of change required for the model to characterize the given inputs more accurately. To eliminate the dependency on ground-truth labels in generating gradients, we introduce confounding labels that are formulated by combining multiple categorical labels. We demonstrate that our gradient-based approach can effectively differentiate inputs that cannot be accurately represented with learned features. We utilize our approach in applications of detecting anomalous inputs, including out-of-distribution, adversarial, and corrupted samples. Our approach requires no hyperparameter tuning or additional data processing and outperforms state-of-the-art methods by up to 2.7%, 19.8%, and 35.6% of AUROC scores, respectively."}, "cited_paper_content": {"title": "Cure-Tsr: Challenging Unreal And Real Environments For Traffic Sign Recognition", "abstract": "In this paper, we investigate the robustness of traffic sign recognition algorithms under challenging conditions. Existing datasets are limited in terms of their size and challenging condition coverage, which motivated us to generate the Challenging Unreal and Real Environments for Traffic Sign Recognition (CURE-TSR) dataset. It includes more than two million traffic sign images that are based on real-world and simulator data. We benchmark the performance of existing solutions in real-world scenarios and analyze the performance variation with respect to challenging conditions. We show that challenging conditions can decrease the performance of baseline methods significantly, especially if these challenging conditions result in loss or misplacement of spatial information. We also investigate the effect of data augmentation and show that utilization of simulator data along with real-world data enhance the average recognition performance in real-world scenarios. The dataset is publicly available at this https URL"}, "keywords": ["dataset", "traffic sign recognition"], "citation_intent": "background"}
{"citing_id": "2304.02509v1", "cited_id": "0807.3917", "section_title": "The Conjecture", "citation": "Since then, the topic gathered significant attention, and the activity sparked with the emergence of polar codes #REFR ; Ar\u0131kan mentioned this as one of the major open problems in coding theory at ITW Dublin in 2010.", "text_before_citation": ["We refer to Section 9 for the formal definition of BMS channels; for now, it is sufficient to consider the BSC, the main case of interest, as BMS are mixtures of BSCs. Conjecture 1.", "For any BMS channel P, and any rate R below the capacity of P given by C(P) = I(U, P) = (1/2) x\u2208{0,1},y\u2208Y P(y|x) log 2 P(y|x)/(P(y|0)/2 + P(y|1)/2), where 2 Y is the output alphabet of P (with C(P) = 1 \u2212 H( ) when P = BSC( )), a sequence of RM(m i , r i ) codes of rate R i = m i \u2264r i 2 \u2212m i tending to R can be decoded successfully with high probability, i.e., any of the 2 nR i codewords can be decoded with probability 1 \u2212 o m i (1) despite corruptions from P.", "It is hard to track back the first appearance of this belief in the literature, but [KKM + 16a] reports that it was likely already present in the late 60s.", "The claim was mentioned explicitly in a 1993 talk by Shu Lin, entitled ?RM Codes are Not So Bad? #OTHEREFR .", "A 1993 paper by Dumer and Farrell also contains a discussion on the matter #OTHEREFR , as well as the 1997 paper of Costello and Forney on the 'road to channel capacity' #OTHEREFR ."], "text_after_citation": ["Due to the broad relevance 3 of RM codes in computer science, electrical engineering and mathematics, the activity scattered in a wide line of works [Dum04, DS06, Dum06, CG05, HKL05, Ari08, Ar\u013109, Ari10, KLP12, ASW15b, ASW15a, KKM + 16a, KKM + 17, MHU14, SSV17, AY19, YA20, SS20, KKM + 16b, Sam18, Sam20, AHN21, HSS21, LHP20, FFHM21, ASY21, RP21, GEE + 21, #OTHEREFR ; see also #OTHEREFR .", "Relations to polar codes In a breakthrough paper, Ar\u0131kan showed that the explicit class of polar codes achieve Shannon capacity on any BMS channel #OTHEREFR .", "Given the close relationship between polar and RM codes, the belief that RM codes could also be proved to achieve capacity on BMS channels intensified.", "Polar codes are derived from the same square matrix, i.e., the matrix whose rows correspond to evaluations of monomials, which can also be expressed as G n := ( 1 1 0 1 ) \u2297m , but polar codes use a different row selection that is channel dependant.", "One can view the difference as follows: apply G n (over F 2 ) to a vector with i.i.d."], "citing_paper_content": {"title": "A Proof That Reed-Muller Codes Achieve Shannon Capacity On Symmetric Channels", "abstract": "Reed-Muller codes were introduced in 1954, with a simple explicit construction based on polynomial evaluations, and have long been conjectured to achieve Shannon capacity on symmetric channels. Major progress was made towards a proof over the last decades; using combinatorial weight enumerator bounds, a breakthrough on the erasure channel from sharp thresholds, hypercontractivity arguments, and polarization theory. Another major progress recently established that the bit error probability vanishes slowly below capacity. However, when channels allow for errors, the results of Bourgain-Kalai do not apply for converting a vanishing bit to a vanishing block error probability, neither do the known weight enumerator bounds. The conjecture that RM codes achieve Shannon capacity on symmetric channels, with high probability of recovering the codewords, has thus remained open. This paper closes the conjecture's proof. It uses a new recursive boosting framework, which aggregates the decoding of codeword restrictions on 'subspace-sunflowers', handling their dependencies via an L p Boolean Fourier analysis, and using a list-decoding argument with a weight enumerator bound from Sberlo-Shpilka. The proof does not require a vanishing bit error probability for the base case, but only a non-trivial probability, obtained here for general symmetric codes. This gives in particular a shortened and tightened argument for the vanishing bit error probability result of Reeves-Pfister, and with prior works, it implies the strong wire-tap secrecy of RM codes on pure-state classical-quantum channels."}, "cited_paper_content": {"title": "Channel Polarization: A Method For Constructing Capacity-Achieving Codes For Symmetric Binary-Input Memoryless Channels", "abstract": "A method is proposed, called channel polarization, to construct code sequences that achieve the symmetric capacity $I(W)$ of any given binary-input discrete memoryless channel (B-DMC) $W$. The symmetric capacity is the highest rate achievable subject to using the input letters of the channel with equal probability. Channel polarization refers to the fact that it is possible to synthesize, out of $N$ independent copies of a given B-DMC $W$, a second set of $N$ binary-input channels $\\{W_N^{(i)}:1\\le i\\le N\\}$ such that, as $N$ becomes large, the fraction of indices $i$ for which $I(W_N^{(i)})$ is near 1 approaches $I(W)$ and the fraction for which $I(W_N^{(i)})$ is near 0 approaches $1-I(W)$. The polarized channels $\\{W_N^{(i)}\\}$ are well-conditioned for channel coding: one need only send data at rate 1 through those with capacity near 1 and at rate 0 through the remaining. Codes constructed on the basis of this idea are called polar codes. The paper proves that, given any B-DMC $W$ with $I(W)>0$ and any target rate $R<I(W)$, there exists a sequence of polar codes $\\{{\\mathscr C}_n;n\\ge 1\\}$ such that ${\\mathscr C}_n$ has block-length $N=2^n$, rate $\\ge R$, and probability of block error under successive cancellation decoding bounded as $P_{e}(N,R) \\le \\bigoh(N^{-\\frac14})$ independently of the code rate. This performance is achievable by encoders and decoders with complexity $O(N\\log N)$ for each."}, "keywords": ["polar codes"], "citation_intent": "background"}
{"citing_id": "2304.09997v2", "cited_id": "cs/0508003", "section_title": "Contribution 2:", "citation": "A pPDA is PAST iff the equation system from Theorem 7 has a non-negative solution. This is in contrast to e.g. the equation system from #REFR Ex.", "text_before_citation": ["Another important ingredient are weighted PDA over semirings #OTHEREFR .", "Related work: Expected runtimes and, more generally, expected rewards in pPDA are studied in #OTHEREFR - #OTHEREFR , #OTHEREFR .", "However, these works focus mostly on expectations conditioned on reaching the empty stack while being in a certain state.", "In our paper we need such conditional expectations as well; but additionally, we also consider the unconditional expected runtime of pPDA, which is crucial for proving PAST.", "More specifically, our Theorem 7 has -to the best of our knowledge -not appeared in previous works."], "text_after_citation": ["3.3] which has a solution iff the expected runtime conditioned on termination is finite.", "However, this can be the case even if the termination probability is < 1.", "Theorem 7 is actually more in the spirit of #OTHEREFR , #OTHEREFR Ch.", "4] , where 1-exit recursive simple stochastic games with positive rewards are considered, a model that corresponds to a stateless variant of PDA controlled by two players and a probabilistic environment.", "As we consider general pPDA with states, our result does not easily follow from theirs."], "citing_paper_content": {"title": "On Certificates, Expected Runtimes, And Termination In Probabilistic Pushdown Automata", "abstract": "Probabilistic pushdown automata (pPDA) are a natural operational model for a variety of recursive discrete stochastic processes. In this paper, we study certificates-succinct and easily verifiable proofs-for upper and lower bounds on various quantitative properties of a given pPDA. We reveal an intimate, yet surprisingly simple connection between the existence of such certificates and the expected time to termination of the pPDA at hand. This is established by showing that certain intrinsic properties, like the spectral radius of the Jacobian of the pPDA's underlying polynomial equation system, are directly related to expected runtimes. As a consequence, we obtain that there always exist easy-to-check proofs for positive almost-sure termination: does a pPDA terminate in finite expected time?"}, "cited_paper_content": {"title": "Model Checking Probabilistic Pushdown Automata", "abstract": "We consider the model checking problem for probabilistic ::: pushdown automata (pPDA) and properties expressible in various ::: probabilistic logics. We start with properties that can be ::: formulated as instances of a generalized random walk problem. ::: We prove that both qualitative and quantitative model checking ::: for this class of properties and pPDA is decidable. Then we ::: show that model checking for the qualitative fragment of the ::: logic PCTL and pPDA is also decidable. Moreover, we develop an ::: error-tolerant model checking algorithm for PCTL and the ::: subclass of stateless pPDA. Finally, we consider the class of ::: omega-regular properties and show that both qualitative and ::: quantitative model checking for pPDA is decidable."}, "keywords": ["pPDA"], "citation_intent": "result"}
{"citing_id": "2305.01245v1", "cited_id": "1606.03802", "section_title": "B. Image Open-Set Recognition", "citation": "Afterwards, another SVM-based model, i.e., SSVM #REFR , proposes to control the empirical risk to ensure a finite risk of the unknown classes.", "text_before_citation": ["#OTHEREFR in the domain of computer vision, where the task is defined to jointly achieve the known class classification and novel unknown class detection.", "Existing image OSR methods can be grouped into two categories include traditional methods and deep neural networks (DNNs)-based methods.", "1) Traditional Methods: The traditional methods are typically implemented by some classic machine learning models such as SVM, nearest-neighbors, sparse-representation, etc., or statistical models such as extreme value machine.", "Among them, the 1-vs-set machine #OTHEREFR and W-SVM #OTHEREFR are the pioneer works and both of them are derived from SVM.", "The former augments the decision plane of classic SVM from the marginal distances, and the latter integrates useful properties of statistical extreme value theory into SVM to calibrate the classification scores."], "text_after_citation": ["On the other hand, OSNN #OTHEREFR extends the nearest-neighbor classifier to the OSR scenario by fully incorporating the ability to recognize samples belonging to classes that are unknown during training.", "A generalized sparse-representation based classification algorithm is then proposed to make use of the generalized Pareto extreme value distribution for OSR.", "The extreme value theorybased models are a special category derived from the statistical modeling methods for statistical extremes #OTHEREFR .", "Among them, EVM #OTHEREFR is the most representative one, which utilizes datadistribution information with a strong theoretical foundation.", "2) DNNs-based Methods: With the popularity of deep learning techniques, the deep neural networks have became a powerful alternative to implement OSR systems."], "citing_paper_content": {"title": "Mdenet: Multi-Modal Dual-Embedding Networks For Malware Open-Set Recognition", "abstract": "Malware open-set recognition (MOSR) aims at jointly classifying malware samples from known families and detect the ones from novel unknown families, respectively. Existing works mostly rely on a well-trained classifier considering the predicted probabilities of each known family with a thresholdbased detection to achieve the MOSR. However, our observation reveals that the feature distributions of malware samples are extremely similar to each other even between known and unknown families. Thus the obtained classifier may produce overly high probabilities of testing unknown samples toward known families and degrade the model performance. In this paper, we propose the Multi-modal Dual-Embedding Networks, dubbed MDENet, to take advantage of comprehensive malware features from different modalities to enhance the diversity of malware feature space, which is more representative and discriminative for downstream recognition. Concretely, we first generate a malware image for each observed sample based on their numeric features using our proposed numeric encoder with a redesigned multi-scale CNN structure, which can better explore their statistical and spatial correlations. Besides, we propose to organize tokenized malware features into a sentence for each sample considering its behaviors and dynamics, and utilize language models as the textual encoder to transform it into a representable and computable textual vector. Such parallel multi-modal encoders can fuse the above two components to enhance the feature diversity. Last, to further guarantee the open-set recognition, we dually embed the fused multi-modal representation into one primary space and an associated subspace, i.e., discriminative and exclusive spaces, with contrastive sampling and \u03c1-bounded enclosing sphere regularizations, which resort to classification and detection, respectively. Moreover, we also enrich our previously proposed large-scaled malware dataset MAL-100 with multi-modal characteristics and contribute an improved version dubbed MAL-100 +. Experimental results on the widely used malware dataset Mailing and the proposed MAL-100 + demonstrate the effectiveness of our method."}, "cited_paper_content": {"title": "Specialized Support Vector Machines For Open-Set Recognition", "abstract": "Often, when dealing with real-world recognition problems, we do not need, and often cannot have, knowledge of the entire set of possible classes that might appear during operational testing. In such cases, we need to think of robust classification methods able to deal with the\"unknown\"and properly reject samples belonging to classes never seen during training. Notwithstanding, almost all existing classifiers to date were mostly developed for the closed-set scenario, i.e., the classification setup in which it is assumed that all test samples belong to one of the classes with which the classifier was trained. In the open-set scenario, however, a test sample can belong to none of the known classes and the classifier must properly reject it by classifying it as unknown. In this work, we extend upon the well-known Support Vector Machines (SVM) classifier and introduce the Specialized Support Vector Machines (SSVM), which is suitable for recognition in open-set setups. SSVM balances the empirical risk and the risk of the unknown and ensures that the region of the feature space in which a test sample would be classified as known (one of the known classes) is always bounded, ensuring a finite risk of the unknown. In this work, we also highlight the properties of the SVM classifier related to the open-set scenario, and provide necessary and sufficient conditions for an RBF SVM to have bounded open-space risk."}, "keywords": ["another SVM-based model"], "citation_intent": "background"}
{"citing_id": "2305.01261v1", "cited_id": "cs/0610105", "section_title": "B. Block Design Schemes: Locally Differentially Private Schemes Based On Block Designs", "citation": "Moreover, the following theorem says that the above asymptotic optimality implies the non-asymptotic optimality under the unbiased constraint for the mean squared loss (\u2113 #REFR 2 ).", "text_before_citation": ["where", "EQUATION", "and E(0, 1; v) := \u221e.", "Especially, if e \u01eb = E(k, k + 1; v) for some k, then", "K opt (v, \u01eb) = {k, k + 1}. Otherwise, K opt (v, \u01eb) is a singletone."], "text_after_citation": ["Theorem 3.7: For given n, let (Q,P n ) be a block design scheme satisfying k \u2208 K opt (v, \u01eb), and (Q \u2032 ,P \u2032 n ) be any unbiased \u01eb-LDP scheme. Then, we have", "EQUATION", "Proof: For each positive integer m, letP \u2032 nm be the estimator on nm clients given b\u0177", "EQUATION", "Because"], "citing_paper_content": {"title": "Block Design-Based Local Differential Privacy Mechanisms", "abstract": "In this paper, we propose a new class of local differential privacy (LDP) schemes based on combinatorial block designs for a discrete distribution estimation. This class not only recovers many known LDP schemes in a unified framework of combinatorial block design, but also suggests a novel way of finding new schemes achieving the optimal (or near-optimal) privacy-utility trade-off with lower communication costs. Indeed, we find many new LDP schemes that achieve both the optimal privacy-utility trade-off and the minimum communication cost among all the unbiased schemes for a certain set of input data size and LDP constraint. Furthermore, to partially solve the sparse existence issue of block design schemes, we consider a broader class of LDP schemes based on regular and pairwise-balanced designs, called RPBD schemes, which relax one of the symmetry requirements on block designs. By considering this broader class of RPBD schemes, we can find LDP schemes achieving near-optimal privacy-utility trade-off with reasonably low communication costs for a much larger set of input data size and LDP constraint."}, "cited_paper_content": {"title": "How To Break Anonymity Of The Netflix Prize Dataset", "abstract": "We present a new class of statistical de-anonymization attacks against high-dimensional micro-data, such as individual preferences, recommendations, transaction records and so on. Our techniques are robust to perturbation in the data and tolerate some mistakes in the adversary's background knowledge. ::: We apply our de-anonymization methodology to the Netflix Prize dataset, which contains anonymous movie ratings of 500,000 subscribers of Netflix, the world's largest online movie rental service. We demonstrate that an adversary who knows only a little bit about an individual subscriber can easily identify this subscriber's record in the dataset. Using the Internet Movie Database as the source of background knowledge, we successfully identified the Netflix records of known users, uncovering their apparent political preferences and other potentially sensitive information."}, "keywords": ["unbiased constraint"], "citation_intent": "background"}
{"citing_id": "2303.07626v1", "cited_id": "1912.10211", "section_title": "Iv-B. Implementation Details", "citation": "Note that ensemble CAT with other acoustic model #REFR achieves performance comparable to previous work trained on additional modality, e.g.", "text_before_citation": ["It is worth mentioning that following the same ensemble protocol, the CAT achieves performance comparable to SOTA ensemble models.", "By simply ensemble the CAT with PANN #OTHEREFR , we achieve performance comparable to the previous 9-model ensemble #OTHEREFR results.", "This also shows the CAT provides complementary information to ConvNet based model.", "ESC50 The CAT achieves SOTA performance on ESC50 dataset as well (Table I) .", "We observe a similar pattern that the proposed CAT outperforms previous SOTA #OTHEREFR , #OTHEREFR by 0.8% and 0.2% with and without ensemble; demonstrating the effectiveness of the proposed design."], "text_after_citation": ["CAT is 0.7% better than AnT #OTHEREFR which utilizes visualacoustic information.", "It can be anticipated that incorporating multimodality features into CAT may further improve performance."], "citing_paper_content": {"title": "Cat: Causal Audio Transformer For Audio Classification", "abstract": "The attention-based Transformers have been increasingly applied to audio classification because of their global receptive field and ability to handle long-term dependency. However, the existing frameworks which are mainly extended from the Vision Transformers are not perfectly compatible with audio signals. In this paper, we introduce a Causal Audio Transformer (CAT) consisting of a Multi-Resolution Multi-Feature (MRMF) feature extraction with an acoustic attention block for more optimized audio modeling. In addition, we propose a causal module that alleviates over-fitting, helps with knowledge transfer and improves interpretability. CAT obtains higher or comparable state-of-the-art classification performance on ESC50, AudioSet and UrbanSound8K datasets, and can be easily generalized to other Transformerbased models."}, "cited_paper_content": {"title": "Panns: Large-Scale Pretrained Audio Neural Networks For Audio Pattern Recognition", "abstract": "Audio pattern recognition is an important research topic in the machine learning area, and includes several tasks such as audio tagging, acoustic scene classification and sound event detection. Recently neural networks have been applied to solve audio pattern recognition problems. However, previous systems focus on small datasets, which limits the performance of audio pattern recognition systems. Recently in computer vision and natural language processing, systems pretrained on large datasets have generalized well to several tasks. However, there is limited research on pretraining neural networks on large datasets for audio pattern recognition. In this paper, we propose large-scale pretrained audio neural networks (PANNs) trained on AudioSet. We propose to use Wavegram, a feature learned from waveform, and the mel spectrogram as input. We investigate the performance and complexity of a variety of convolutional neural networks. Our proposed AudioSet tagging system achieves a state-of-the-art mean average precision (mAP) of 0.439, outperforming the best previous system of 0.392. We transferred a PANN to six audio pattern recognition tasks and achieve state-of-the-art performance in many tasks. Source code and pretrained models have been released."}, "keywords": ["ensemble CAT", "acoustic model"], "citation_intent": "result"}
{"citing_id": "2303.08225v1", "cited_id": "1706.03762", "section_title": "Related Work", "citation": "The Transformer was first proposed in #REFR for machine translation and has established state-of-the-art results in many natural language processing (NLP) tasks.", "text_before_citation": ["#OTHEREFR tried to generate image layouts and synthesize realistic images from input scene graphs via GCNs. Nauata et al.", "#OTHEREFR proposed a graph-constrained generative adversarial network, whose generator and discriminator are built upon a relational architecture.", "Our innovation is a novel graph Transformer GAN, where the input constraint is encoded into the graph structure of the proposed graph Transformer-based generator and node classificationbased discriminator.", "Experimental results show the effectiveness of GTGAN over all the leading methods.", "Transformers in Computer Vision."], "text_after_citation": ["Recently, the Vision Transformer (ViT) #OTHEREFR equipped with global self-attention has achieved state-of-the-art results on the classification task.", "Since then, Transformer-based approaches have been shown to be efficient in many computer vision tasks including image segmentation #OTHEREFR , object detection #OTHEREFR , depth estimation #OTHEREFR , pose estimation #OTHEREFR , video inpainting #OTHEREFR , vision-and-language navigation #OTHEREFR , video classification #OTHEREFR , human reaction generation #OTHEREFR , 3D pose transfer #OTHEREFR .", "Different from these methods, in this paper, we adopt a Transformer-based network to tackle the graphconstrained house generation task.", "However, integrating graph convolutional networks and ViTs is not trivial.", "To this end, we propose a graph Transformer-based generator to capture both local and global relations across nodes in a graph."], "citing_paper_content": {"title": "Graph Transformer Gans For Graph-Constrained House Generation", "abstract": "We present a novel graph Transformer generative adversarial network (GTGAN) to learn effective graph node relations in an end-to-end fashion for the challenging graphconstrained house generation task. The proposed graph-Transformer-based generator includes a novel graph Transformer encoder that combines graph convolutions and selfattentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node attention (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. Moreover, we propose a new node classification-based discriminator to preserve the high-level semantic and discriminative node features for different house components. Finally, we propose a novel graph-based cycle-consistency loss that aims at maintaining the relative spatial relationships between ground truth and predicted graphs. Experiments on two challenging graph-constrained house generation tasks (i.e., house layout and roof generation) with two public datasets demonstrate the effectiveness of GTGAN in terms of objective quantitative scores and subjective visual realism. New state-of-the-art results are established by large margins on both tasks."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["machine translation"], "citation_intent": "background"}
{"citing_id": "2303.14011v1", "cited_id": "1703.03400", "section_title": "I. Introduction", "citation": "It simulates the train-test process with few examples during training and optimizes the models to reduce the simulated test errors #REFR .", "text_before_citation": ["There are two common paradigms for creating such pretext tasks, i.e., self-supervised learning and intermediate-finetuning.", "Specifically, in self-supervised methods, pseudo annotations are crafted by some heuristics on unannotated data to train models in a supervised learning manner #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR .", "On the other hand, for intermediate finetuning methods, task-related annotated datasets are used to finetune language models before finetuning on the target task #OTHEREFR .", "Although transfer learning methods are usually effective, they are prone to be affected by the data properties or learning objectives in the pretext tasks #OTHEREFR , #OTHEREFR .", "Different from transfer learning, meta learning provides a framework that transfers task-level knowledge instead of data-level one."], "text_after_citation": ["In such schemes, the behaviors of learning with few examples are specifically considered.", "The ability of \"learning to learn\" allows the model to adapt to unseen tasks quickly.", "Nevertheless, two challenges arise in designing such a meta learning framework for abstractive summarization. 1) Expensive computation.", "Meta learning involves the computation of second derivatives, which is expensive for summarization models due to the typically large model size. 2) Meta-dataset construction. Meta learning requires crafted tasks for learning.", "In applications such as image classification #OTHEREFR , tasks can be directly defined as different class combinations in a single dataset, which is not suitable for abstractive summarization."], "citing_paper_content": {"title": "Spec: Summary Preference Decomposition For Low-Resource Abstractive Summarization", "abstract": "Neural abstractive summarization has been widely studied and achieved great success with large-scale corpora. However, the considerable cost of annotating data motivates the need for learning strategies under low-resource settings. In this paper, we investigate the problems of learning summarizers with only few examples and propose corresponding methods for improvements. First, typical transfer learning methods are prone to be affected by data properties and learning objectives in the pretext tasks. Therefore, based on pretrained language models, we further present a meta learning framework to transfer few-shot learning processes from source corpora to the target corpus. Second, previous methods learn from training examples without decomposing the content and preference. The generated summaries could therefore be constrained by the preference bias in the training set, especially under low-resource settings. As such, we propose decomposing the contents and preferences during learning through the parameter modulation, which enables control over preferences during inference. Third, given a target application, specifying required preferences could be non-trivial because the preferences may be difficult to derive through observations. Therefore, we propose a novel decoding method to automatically estimate suitable preferences and generate corresponding summary candidates from the few training examples. Extensive experiments demonstrate that our methods achieve state-of-the-art performance on six diverse corpora with 30.11%/33.95%/27.51% and 26.74%/31.14%/24.48% average improvements on ROUGE-1/2/L under 10-and 100example settings."}, "cited_paper_content": {"title": "Model-Agnostic Meta-Learning For Fast Adaptation Of Deep Networks", "abstract": "We propose an algorithm for meta-learning that is model-agnostic, in the sense that it is compatible with any model trained with gradient descent and applicable to a variety of different learning problems, including classification, regression, and reinforcement learning. The goal of meta-learning is to train a model on a variety of learning tasks, such that it can solve new learning tasks using only a small number of training samples. In our approach, the parameters of the model are explicitly trained such that a small number of gradient steps with a small amount of training data from a new task will produce good generalization performance on that task. In effect, our method trains the model to be easy to fine-tune. We demonstrate that this approach leads to state-of-the-art performance on two few-shot image classification benchmarks, produces good results on few-shot regression, and accelerates fine-tuning for policy gradient reinforcement learning with neural network policies."}, "keywords": ["training"], "citation_intent": "method"}
{"citing_id": "2303.14934v1", "cited_id": "1807.02758", "section_title": "Supervision For Texture Areas", "citation": "In order to further refine the color information, we additionally add several 1\u00d71 convolution blocks with channel attention mechanism #REFR . More details are provided in the supplementary material.", "text_before_citation": ["EQUATION", "wherex 1 andx 2 is the output of BNN and LAN, respectively. sg(\u2022) denotes stop gradient operation #OTHEREFR .", "When training is done, LAN will have a transfer ability to predict the supervision for textured areas.", "The network structure of LAN is simple yet delicate.", "We stack 3\u00d73 convolution layers to create the local receptive field, specifically, k 3\u00d73 layers can make up (2k+1)\u00d7(2k+1) receptive field."], "text_after_citation": [], "citing_paper_content": {"title": "Spatially Adaptive Self-Supervised Learning For Real-World Image Denoising", "abstract": "Significant progress has been made in self-supervised image denoising (SSID) in the recent few years. However, most methods focus on dealing with spatially independent noise, and they have little practicality on real-world sRGB images with spatially correlated noise. Although pixelshuffle downsampling has been suggested for breaking the noise correlation, it breaks the original information of images, which limits the denoising performance. In this paper, we propose a novel perspective to solve this problem, i.e., seeking for spatially adaptive supervision for real-world sRGB image denoising. Specifically, we take into account the respective characteristics of flat and textured regions in noisy images, and construct supervisions for them separately. For flat areas, the supervision can be safely derived from non-adjacent pixels, which are much far from the current pixel for excluding the influence of the noise-correlated ones. And we extend the blind-spot network to a blindneighborhood network (BNN) for providing supervision on flat areas. For textured regions, the supervision has to be closely related to the content of adjacent pixels. And we present a locally aware network (LAN) to meet the requirement, while LAN itself is selectively supervised with the output of BNN. Combining these two supervisions, a denoising network (e.g., U-Net) can be well-trained. Extensive experiments show that our method performs favorably against state-of-the-art SSID methods on real-world sRGB photographs. The code is available at https://github. com/nagejacob/SpatiallyAdaptiveSSID."}, "cited_paper_content": {"title": "Image Super-Resolution Using Very Deep Residual Channel Attention Networks", "abstract": "Convolutional neural network (CNN) depth is of crucial importance for image super-resolution (SR). However, we observe that deeper networks for image SR are more difficult to train. The low-resolution inputs and features contain abundant low-frequency information, which is treated equally across channels, hence hindering the representational ability of CNNs. To solve these problems, we propose the very deep residual channel attention networks (RCAN). Specifically, we propose a residual in residual (RIR) structure to form very deep network, which consists of several residual groups with long skip connections. Each residual group contains some residual blocks with short skip connections. Meanwhile, RIR allows abundant low-frequency information to be bypassed through multiple skip connections, making the main network focus on learning high-frequency information. Furthermore, we propose a channel attention mechanism to adaptively rescale channel-wise features by considering interdependencies among channels. Extensive experiments show that our RCAN achieves better accuracy and visual improvements against state-of-the-art methods."}, "keywords": ["several 1\u00d71 convolution", "channel attention mechanism"], "citation_intent": "method"}
{"citing_id": "2305.00767v1", "cited_id": "2003.14013", "section_title": "Introduction", "citation": "The work in #REFR tried to solve the problem by capturing with stopand-motion manner, which captured each noisy-clean pair with a static scene, and concatenated frames with temporal order to generate paired noisy-clean raw videos, building the CRVD dataset.", "text_before_citation": ["For image denoising, there are mainly two strategies to capture paired noisy-clean images.", "One strategy is capturing noisy-clean image pairs using high and low ISO settings, and the other strategy is averaging a lot of noisy images to get a clean image.", "The two strategies both require the objects to be static and the camera is static.", "Otherwise, it will lead to motion blur or large displacements between noisy and clean images.", "This makes it difficult to capture noisy-clean pairs for videos."], "text_after_citation": ["But #OTHEREFR mainly captured toys which were controlled by people, the motions are monotonous and unnatural, and there are only few scenes (11 scenes) in the dataset.", "The work in #OTHEREFR moved cameras rather than moving the objects to create translation motions, and they only captured sRGB video pairs.", "Neither moving toys nor moving cameras can simulate the various motions in real world.", "Training with the two datasets limits the model's ability in dealing with real-world noisy videos.", "Based on the above observations, we propose to recapture the existing high-resolution videos displayed on a 4K screen to construct our dataset."], "citing_paper_content": {"title": "Rvideformer: Efficient Raw Video Denoising Transformer With A Larger Benchmark Dataset", "abstract": "In recent years, raw video denoising has garnered increased attention due to the consistency with the imaging process and well-studied noise modeling in the raw domain. Despite these advancements, two problems still hinder the denoising performance. Firstly, there is no large dataset with realistic motions for supervised raw video denoising, as capturing noisy and clean frames for real dynamic scenes is difficult. To address this, we propose recapturing existing high-resolution videos displayed on a 4K screen. Specifically, we recapture the screen content with high-low ISO settings to construct noisy-clean paired frames. Afterward, we introduce intensity, spatial, and color correction strategies to make the paired frames well-aligned. Then, the aligned frames are concatenated with temporal order to construct paired videos. In this way, we construct a video denoising dataset (named as ReCRVD) with 120 groups of noisy-clean videos, whose ISO values ranging from 1600 to 25600. Secondly, while non-local temporal-spatial attention is beneficial for denoising, it often leads to heavy computation costs. In this work, we propose an efficient raw video denoising transformer network (RViDeformer) that explores both short and long-distance correlations. Specifically, we introduce Low-Resolution-Window Self-Attention (LWSA), Global-Window Self-Attention (GWSA), and Neighbour-Window Self-Attention (NWSA) to build a multi-branch spatial self-attention for spatial reconstruction. Similarly, Global-Window Temporal Mutual Attention (GTMA) and Neighbour-Window Temporal Mutual Attention (NTMA) are proposed to build multi-branch temporal self-attention for temporal reconstruction. We employ reparameterization to reduce computation costs. Our network is trained in both supervised and unsupervised manners, achieving the best performance compared with state-of-the-art methods. Additionaly, the model trained with our proposed dataset (ReCRVD) outperforms the model trained with previous benchmark dataset (CRVD) when evaluated on the real-world outdoor noisy videos. Our code and dataset will be released after the acceptance of this work. Index Terms-Raw video denoising, ReCRVD dataset, Raw video denoising transformer (RViDeformer), unsupervised learning."}, "cited_paper_content": {"title": "Supervised Raw Video Denoising With A Benchmark Dataset On Dynamic Scenes", "abstract": "In recent years, the supervised learning strategy for real noisy image denoising has been emerging and has achieved promising results. In contrast, realistic noise removal for raw noisy videos is rarely studied due to the lack of noisy-clean pairs for dynamic scenes. Clean video frames for dynamic scenes cannot be captured with a long-exposure shutter or averaging multi-shots as was done for static images. In this paper, we solve this problem by creating motions for controllable objects, such as toys, and capturing each static moment for multiple times to generate clean video frames. In this way, we construct a dataset with 55 groups of noisy-clean videos with ISO values ranging from 1600 to 25600. To our knowledge, this is the first dynamic video dataset with noisy-clean pairs. Correspondingly, we propose a raw video denoising network (RViDeNet) by exploring the temporal, spatial, and channel correlations of video frames. Since the raw video has Bayer patterns, we pack it into four sub-sequences, i.e RGBG sequences, which are denoised by the proposed RViDeNet separately and finally fused into a clean video. In addition, our network not only outputs a raw denoising result, but also the sRGB result by going through an image signal processing (ISP) module, which enables users to generate the sRGB result with their favourite ISPs. Experimental results demonstrate that our method outperforms state-of-the-art video and raw image denoising algorithms on both indoor and outdoor videos."}, "keywords": ["paired noisy-clean raw"], "citation_intent": "method"}
{"citing_id": "2303.06476v1", "cited_id": "1908.00672", "section_title": "Experiments On Synthetic Datasets", "citation": "We follow #REFR to drop the matting refinement stage in DIM for simplicity and fair comparison. The results are illustrated in Tab. 7.", "text_before_citation": ["In contrast, thanks to the large receptive field and tri-token guided design, our proposed model learns image context and object structure effectively, contributing to sophisticated performance in such a challenging region.", "Besides, our proposed model can also accurately estimate the TT-type objects whose background is complex. For example, in the second row of Fig.", "3 , the background is so complicated that the previous methods are hard to describe the details.", "It is worth noting that our TransMattingV2 achieves more high-quality alpha mattes than TransMattingV1, indicating that our tri-token helps to distinguish the foreground and background in deep features.", "To evaluate the universality of our proposed tri-token, we introduce it to three current SOTA methods: DIM #OTHEREFR , GCAMatting #OTHEREFR , and MGMatting #OTHEREFR ."], "text_after_citation": ["As we can see, our proposed tri-token boosts the performance of all three SOTA methods with a big margin.", "Especially for recent proposed GCAMatting and MGMAtting, by replacing the traditional concatenation with our proposed tri-token, the SAD is decreased by 3.98 and 3.18, respectively.", "This proves the effectiveness of the proposed tri-token strategy in existing SOTA methods and indicates a great potential to replace the concatenation strategy with the proposed tritoken in a wide range of image matting methods.", "As illustrated in Tab.", "1, to further verify our method in boosting the performance of TT-type objects, we analyze two types of objects separately."], "citing_paper_content": {"title": "Transmatting: Tri-Token Equipped Transformer Model For Image Matting", "abstract": "Image matting aims to predict alpha values of elaborate uncertainty areas of natural images, like hairs, smoke, and spider web. However, existing methods perform poorly when faced with highly transparent foreground objects due to the large area of uncertainty to predict and the small receptive field of convolutional networks. To address this issue, we propose a Transformer-based network (TransMatting) to model transparent objects with long-range features and collect a high-resolution matting dataset of transparent objects (Transparent-460) for performance evaluation. Specifically, to utilize semantic information in the trimap flexibly and effectively, we also redesign the trimap as three learnable tokens, named tri-token. Both Transformer and convolution matting models could benefit from our proposed tri-token design. By replacing the traditional trimap concatenation strategy with our tri-token, existing matting methods could achieve about 10% improvement in SAD and 20% in MSE. Equipped with the new tri-token design, our proposed TransMatting outperforms current state-of-the-art methods on several popular matting benchmarks and our newly collected Transparent-460."}, "cited_paper_content": {"title": "Indices Matter: Learning To Index For Deep Image Matting", "abstract": "We show that existing upsampling operators can be unified using the notion of the index function. This notion is inspired by an observation in the decoding process of deep image matting where indices-guided unpooling can often recover boundary details considerably better than other upsampling operators such as bilinear interpolation. By viewing the indices as a function of the feature map, we introduce the concept of 'learning to index', and present a novel index-guided encoder-decoder framework where indices are self-learned adaptively from data and are used to guide the pooling and upsampling operators, without extra training supervision. At the core of this framework is a flexible network module, termed IndexNet, which dynamically generates indices conditioned on the feature map. Due to its flexibility, IndexNet can be used as a plug-in applying to almost all off-the-shelf convolutional networks that have coupled downsampling and upsampling stages. We demonstrate the effectiveness of IndexNet on the task of natural image matting where the quality of learned indices can be visually observed from predicted alpha mattes. Results on the Composition-1k matting dataset show that our model built on MobileNetv2 exhibits at least 16.1% improvement over the seminal VGG-16 based deep matting baseline, with less training data and lower model capacity. Code and models have been made available at: https://tinyurl.com/IndexNetV1."}, "keywords": ["matting refinement stage"], "citation_intent": "result"}
{"citing_id": "2303.09946v1", "cited_id": "1108.1855", "section_title": "I. Introduction", "citation": "A neighbor-based tracking control scheme with distributed estimators is applied to advise distributed tracking strategies in #REFR .", "text_before_citation": ["They provide an alternative solution based on relaxing the inter-dependency requirements among the agents, such as the communication range of each agent, the amount of information required about its surrounding agents and the environment, etc.", "A key aspect about any distributed controller is its ability to optimize a set of simultaneously conflicting local and team objectives.", "The underlying control methods are often based on theories of feedback mechanisms and communication graphs #OTHEREFR , #OTHEREFR .", "The behavior of large numbers of interacting agents, such as flocks of flying birds and fish schools, is used in the study the cooperative dynamical systems #OTHEREFR , #OTHEREFR .", "Sliding mode methods are employed to achieve coordination of multi-agent systems in #OTHEREFR , #OTHEREFR ."], "text_after_citation": ["Relevant control approaches relied on the existence of a virtual leader or a moving target in mobile sensor networks #OTHEREFR , #OTHEREFR .", "A common shortcoming to all these techniques is their dependence on the prior knowledge of a system model that is free of unstructured uncertainties.", "A distributed adaptive strategy that is based on a backstepping scheme is adopted to solve a consensus control problem in #OTHEREFR .", "The tracking problem of fractional-order multi-agent systems is addressed in #OTHEREFR using adaptive strategies with neural network approximators.", "Reinforcement Learning (RL) is a machine learning tool that allows agents to learn the best strategies to solve certain problems upon interacting with unknown environments #OTHEREFR , #OTHEREFR ."], "citing_paper_content": {"title": "An Adaptive Fuzzy Reinforcement Learning Cooperative Approach For The Autonomous Control Of Flock Systems", "abstract": "The flock-guidance problem enjoys a challenging structure where multiple optimization objectives are solved simultaneously. This usually necessitates different control approaches to tackle various objectives, such as guidance, collision avoidance, and cohesion. The guidance schemes, in particular, have long suffered from complex tracking-error dynamics. Furthermore, techniques that are based on linear feedback strategies obtained at equilibrium conditions either may not hold or degrade when applied to uncertain dynamic environments. Pre-tuned fuzzy inference architectures lack robustness under such unmodeled conditions. This work introduces an adaptive distributed technique for the autonomous control of flock systems. Its relatively flexible structure is based on online fuzzy reinforcement learning schemes which simultaneously target a number of objectives; namely, following a leader, avoiding collision, and reaching a flock velocity consensus. In addition to its resilience in the face of dynamic disturbances, the algorithm does not require more than the agent position as a feedback signal. The effectiveness of the proposed method is validated with two simulation scenarios and benchmarked against a similar technique from the literature."}, "cited_paper_content": {"title": "Distributed Tracking Control Of Leader-Follower Multi-Agent Systems Under Noisy Measurement", "abstract": "In this paper, a distributed tracking control scheme with distributed estimators has been developed for a leader-follower multi-agent system with measurement noises and directed interconnection topology. It is supposed that each follower can only measure relative positions of its neighbors in a noisy environment, including the relative position of the second-order active leader. A neighbor-based tracking protocol together with distributed estimators is designed based on a novel velocity decomposition technique. It is shown that the closed loop tracking control system is stochastically stable in mean square and the estimation errors converge to zero in mean square as well. A simulation example is finally given to illustrate the performance of the proposed control scheme."}, "keywords": ["distributed tracking strategies", "neighbor-based tracking control"], "citation_intent": "method"}
{"citing_id": "2303.15932v3", "cited_id": "1503.03832", "section_title": "Introduction", "citation": "In implementation, CRA not only learns discriminative visual and textual features based on a set of orthonormal basis and a dual-gate mechanism, but also globally aligns both type of features under the supervision of a triplet contrastive loss #REFR .", "text_before_citation": ["To this end, although there has emerged progresses for improving either global CMA #OTHEREFR or local CMA #OTHEREFR , how to exploit multi-level CMA to enhance radiology report generation is underexplored.", "Considering the above issues, we propose a Unify, Align and Refine (UAR) framework to facilitate multi-level CMA for generating faithful and believable radiology reports.", "Within UAR, we design a Latent Space Unifier (LSU) to tokenize images into discrete visual tokens with a discrete variational autoencoder (dVAE) #OTHEREFR .", "By doing so, LSU unifies vision and text modalities into discrete tokens, making it flexible to design a shared network that seamlessly processes both modalities to learn common knowledge #OTHEREFR .", "Next, we introduce a modality-agnostic Cross-modal Representation Aligner (CRA) to learn global CMA."], "text_after_citation": ["What's more, we improve local CMA by integrating Transformer #OTHEREFR with our proposed Text-to-Image Refiner (TIR).", "Different from the vanilla attention mechanism in Transformer, TIR additionally contains a learnable mask to re-calibrate text-to-image attention activations.", "TIR constrains the learnable mask with an auxiliary loss to focus word prediction on useful visual information.", "Last but not least, we design a two-stage training procedure to make the full model grasp CMA at different levels gradually.", "We conduct experiments on two widely-adopted radiology report generation benchmarks, i.e., IU-Xray #OTHEREFR and MIMIC-CXR #OTHEREFR ."], "citing_paper_content": {"title": "Unify, Align And Refine: Multi-Level Semantic Alignment For Radiology Report Generation", "abstract": "Automatic radiology report generation has attracted enormous research interest due to its practical value in reducing the workload of radiologists. However, simultaneously establishing global correspondences between the image (e.g., Chest X-ray) and its related report and local alignments between image patches and keywords remains challenging. To this end, we propose an Unify, Align and then Refine (UAR) approach to learn multi-level cross-modal alignments and introduce three novel modules: Latent Space Unifier (LSU), Cross-modal Representation Aligner (CRA) and Textto-Image Refiner (TIR). Specifically, LSU unifies multimodal data into discrete tokens, making it flexible to learn common knowledge among modalities with a shared network. The modality-agnostic CRA learns discriminative features via a set of orthonormal basis and a dual-gate mechanism first and then globally aligns visual and textual representations under a triplet contrastive loss. TIR boosts token-level local alignment via calibrating text-to-image attention with a learnable mask. Additionally, we design a two-stage training procedure to make UAR gradually grasp cross-modal alignments at different levels, which imitates radiologists' workflow: writing sentence by sentence first and then checking word by word. Extensive experiments and analyses on IU-Xray and MIMIC-CXR benchmark datasets demonstrate the superiority of our UAR against varied state-of-the-art methods."}, "cited_paper_content": {"title": "Facenet: A Unified Embedding For Face Recognition And Clustering", "abstract": "Despite significant recent advances in the field of face recognition [10, 14, 15, 17], implementing face verification and recognition efficiently at scale presents serious challenges to current approaches. In this paper we present a system, called FaceNet, that directly learns a mapping from face images to a compact Euclidean space where distances directly correspond to a measure of face similarity. Once this space has been produced, tasks such as face recognition, verification and clustering can be easily implemented using standard techniques with FaceNet embeddings as feature vectors."}, "keywords": ["triplet contrastive loss", "discriminative visual"], "citation_intent": "method"}
{"citing_id": "2303.03079v1", "cited_id": "quant-ph/0508070", "section_title": "Introduction", "citation": "Their parameters are written as ((n, K, d)) q , where K is the dimension and d the distance. The most common QECCs are stabilizer codes #REFR .", "text_before_citation": ["Quantum error-correcting codes (QECCs) are an important tool in quantum computation since they allow us to protect quantum information from quantum noise and, in particular, from decoherence.", "The existence of quantum processors that improve the behaviour of classical ones #OTHEREFR explains the importance of obtaining good QECCs.", "A q-ary QECC of length n, q being a power of a prime p, is a subspace of the Hilbert space C q n , where C denotes the complex numbers."], "text_after_citation": ["Many of these codes satisfy K = q k , for some nonnegative integer k, and then their parameters are expressed as [[n, k, d ]] q ; abusing the notation we say that k is the dimension of these codes.", "QECCs were first introduced in the binary case #OTHEREFR and later in the general case #OTHEREFR , where we have cited only some references of a vast literature on the subject.", "QECCs in the nonbinary case are convenient for fault-tolerant quantum computation #OTHEREFR .", "Stabilizer quantum codes are intimately related to self-orthogonal additive codes, where duality is given by a trace-symplectic form #OTHEREFR .", "As particular cases of the above construction, stabilizer quantum codes can be obtained from self-orthogonal linear codes with respect to Hermitian or Euclidean inner products."], "citing_paper_content": {"title": "Steane Enlargement Of Entanglement-Assisted Quantum Error-Correcting Codes", "abstract": "We introduce a Steane-like enlargement procedure for entanglement-assisted quantum error-correcting codes (EAQECCs) obtained by considering Euclidean inner product. We give formulae for the parameters of these enlarged codes and apply our results to explicitly compute the parameters of enlarged EAQECCs coming from some BCH codes."}, "cited_paper_content": {"title": "Nonbinary Stabilizer Codes Over Finite Fields", "abstract": "One formidable difficulty in quantum communication and computation is to protect information-carrying quantum states against undesired interactions with the environment. To address this difficulty, many good quantum error-correcting codes have been derived as binary stabilizer codes. Fault-tolerant quantum computation prompted the study of nonbinary quantum codes, but the theory of such codes is not as advanced as that of binary quantum codes. This paper describes the basic theory of stabilizer codes over finite fields. The relation between stabilizer codes and general quantum codes is clarified by introducing a Galois theory for these objects. A characterization of nonbinary stabilizer codes over Fq in terms of classical codes over Fq 2 is provided that generalizes the well-known notion of additive codes over F4 of the binary case. This paper also derives lower and upper bounds on the minimum distance of stabilizer codes, gives several code constructions, and derives numerous families of stabilizer codes, including quantum Hamming codes, quadratic residue codes, quantum Melas codes, quantum Bose-Chaudhuri-Hocquenghem (BCH) codes, and quantum character codes. The puncturing theory by Rains is generalized to additive codes that are not necessarily pure. Bounds on the maximal length of maximum distance separable stabilizer codes are given. A discussion of open problems concludes this paper"}, "keywords": ["stabilizer codes"], "citation_intent": "background"}
{"citing_id": "2304.01096v1", "cited_id": "1406.2661", "section_title": "Generative Adversarial Optimization", "citation": "One tentative workaround to quantitatively proxying for imitation is proposed by the framework of generative adversarial optimization, recently made popular in the context of gradient optimization of neural networks #REFR .", "text_before_citation": [], "text_after_citation": ["In essence, instead of manually quantifying performance, this framework introduces a second agent optimized to discriminate between real collectable data and artificially produced data.", "Oppositely, the original data-producing agent is now optimized to generate data that the new agent cannot easily discern.", "When optimization conditions are right, generating agents eventually produce synthesized data so similar to real data that discriminator agents can no longer tell the difference. Figure 10 . Metric-Based and Adversarial Optimization for Data Imitation.", "(left) A similarity metric is manually defined from some target dataset for an agent to be optimized based on.", "(right) Instead of a metric, a second agent is put in place and optimized to discriminate between generated and target data."], "citing_paper_content": {"title": "Evolving Artificial Neural Networks To Imitate Human Behaviour In Shinobi Iii : Return Of The Ninja Master Pr\u00e9sent\u00e9 Par", "abstract": "Our society is increasingly fond of computational tools. This phenomenon has greatly increased over the past decade following, among other factors, the emergence of a new Artificial Intelligence paradigm. Specifically, the coupling of two algorithmic techniques, Deep Neural Networks and Stochastic Gradient Descent, thrusted by an exponentially increasing computing capacity, has and is continuing to become a major asset in many modern technologies. However, as progress takes its course, some still wonder whether other methods could similarly or even more greatly benefit from these various hardware advances. In order to further this study, we delve in this thesis into Evolutionary Algorithms and their application to Dynamic Neural Networks, two techniques which despite enjoying many advantageous properties have yet to find their niche in contemporary Artificial Intelligence. We find that by elaborating new methods while exploiting strong computational resources, it becomes possible to develop strongly performing agents on a variety of benchmarks but also some other agents behaving very similarly to human subjects on the video game Shinobi III : Return of The Ninja Master, typical complex tasks previously out of reach for non-gradient-based optimization."}, "cited_paper_content": {"title": "Generative Adversarial Nets", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to \u00bd everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["neural networks", "generative adversarial optimization"], "citation_intent": "method"}
{"citing_id": "2303.10727v1", "cited_id": "1503.02531", "section_title": "Percentage Of Occurrences", "citation": "The proposed efficient knowledge distillation scheme can achieve the same or even better accuracy compared with that of training with vanilla knowledge distillation #REFR yet at a similar training speed as the standard training, as verified in Sec. 4.3.", "text_before_citation": ["EQUATION", "EQUATION", "where S sample and S batch denote the uncertainty score of one sample and a batch of samples, s is the batch size, Yi is the output probability of the model for class i,\u0176i is the prior probability of the labels for class i, and n is the number of classes.", "To simplify the measurement, all our experiments are based on the assumption that different classes have the same prior probability, i.e.,\u0176i =\u0176j for any i and j. As shown in Fig.", "2 , if S batch is larger than the pre-defined threshold, then a query to the teacher model will be made; otherwise, this iteration will proceed with the standard training without knowledge distillation."], "text_after_citation": [], "citing_paper_content": {"title": "Ersam: Neural Architecture Search For Energy-Efficient And Real-Time Social Ambiance Measurement", "abstract": "Social ambiance describes the context in which social interactions happen, and can be measured using speech audio by counting the number of concurrent speakers. This measurement has enabled various mental health tracking and human-centric IoT applications. While on-device Socal Ambiance Measure (SAM) is highly desirable to ensure user privacy and thus facilitate wide adoption of the aforementioned applications, the required computational complexity of state-of-the-art deep neural networks (DNNs) powered SAM solutions stands at odds with the often constrained resources on mobile devices. Furthermore, only limited labeled data is available or practical when it comes to SAM under clinical settings due to various privacy constraints and the required human effort, further challenging the achievable accuracy of on-device SAM solutions. To this end, we propose a dedicated neural architecture search framework for Energy-efficient and Real-time SAM (ERSAM). Specifically, our ERSAM framework can automatically search for DNNs that push forward the achievable accuracy vs. hardware efficiency frontier of mobile SAM solutions. For example, ERSAM-delivered DNNs only consume 40 mW \u2022 12 h energy and 0.05 seconds processing latency for a 5 seconds audio segment on a Pixel 3 phone, while only achieving an error rate of 14.3% on a social ambiance dataset generated by LibriSpeech. We can expect that our ERSAM framework can pave the way for ubiquitous on-device SAM solutions which are in growing demand."}, "cited_paper_content": {"title": "Distilling The Knowledge In A Neural Network", "abstract": "A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel."}, "keywords": ["vanilla knowledge distillation"], "citation_intent": "result"}
{"citing_id": "2303.01041v1", "cited_id": "1804.04159", "section_title": "C. Nt: Network", "citation": "The authors of #REFR also observed that as opposed to the size variations of normal traffic, TCP SYN flood packets are quite small (typically under 100 bytes), largely in order to maximize the number of connection requests per second for victim exhaustion.", "text_before_citation": ["1) INB and OUT: Inbound and Outbound Network Traffic: Among the direct/detectable features, inbound network traffic features describe the traffic whose destination is the IoT device in terms of packet count and size, encryption, etc.", "The outbound network traffic features are similar to the inbound ones, however in this case, the IoT device is the source of the traffic for the features rather than the destination.", "In past research, e.g., #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , these features were found informative for IoT attack detection.", "For instance, most normal IoT packets are sent at regular time intervals #OTHEREFR (e.g., automated network activities).", "In contrast, these intervals are dramatically shorter #OTHEREFR for most DoS attack traffic and thus informative."], "text_after_citation": ["2) SRD: Communication Sources & Destinations: As opposed to PCs and smartphones, IoT devices normally communicate with a limited number of endpoints (IP destination) #OTHEREFR , e.g., for (de)activation from the cloud, retrieving firmware updates, and logging their status; thus, an increased number of destination endpoints might be indicative of attack traffic (e.g., sending messages to the botnet victims), while extraneous source IP addresses might indicate logging attempts into the device, as in Mirai #OTHEREFR .", "Moreover, the set of destination IPs rarely changes over time #OTHEREFR .", "Also, the traffic associated with DDoS attacks has different protocol distributions (roughly equivalent to the number of destination ports), and they also include fewer total protocols #OTHEREFR , thus highly likely to be informative for IoT attack detection."], "citing_paper_content": {"title": "D-Score: An Expert-Based Method For Assessing The Detectability Of Iot-Related Cyber-Attacks", "abstract": "IoT devices are known to be vulnerable to various cyber-attacks, such as data exfiltration and the execution of flooding attacks as part of a DDoS attack. When it comes to detecting such attacks using network traffic analysis, it has been shown that some attack scenarios are not always equally easy to detect if they involve different IoT models. That is, when targeted at some IoT models, a given attack can be detected rather accurately, while when targeted at others the same attack may result in too many false alarms. In this research, we attempt to explain this variability of IoT attack detectability and devise a risk assessment method capable of addressing a key question: how easy is it for an anomaly-based network intrusion detection system to detect a given cyber-attack involving a specific IoT model? In the process of addressing this question we (a) investigate the predictability of IoT network traffic, (b) present a novel taxonomy for IoT attack detection which also encapsulates traffic predictability aspects, (c) propose an expertbased attack detectability estimation method which uses this taxonomy to derive a detectability score (termed 'D-Score') for a given combination of IoT model and attack scenario, and (d) empirically evaluate our method while comparing it with a datadriven method."}, "cited_paper_content": {"title": "Machine Learning Ddos Detection For Consumer Internet Of Things Devices", "abstract": "An increasing number of Internet of Things (IoT) devices are connecting to the Internet, yet many of these devices are fundamentally insecure, exposing the Internet to a variety of attacks. Botnets such as Mirai have used insecure consumer IoT devices to conduct distributed denial of service (DDoS) attacks on critical Internet infrastructure. This motivates the development of new techniques to automatically detect consumer IoT attack traffic. In this paper, we demonstrate that using IoT-specific network behaviors (e.g., limited number of endpoints and regular time intervals between packets) to inform feature selection can result in high accuracy DDoS detection in IoT network traffic with a variety of machine learning algorithms, including neural networks. These results indicate that home gateway routers or other network middleboxes could automatically detect local IoT device sources of DDoS attacks using low-cost machine learning algorithms and traffic data that is flow-based and protocol-agnostic."}, "keywords": ["normal traffic", "TCP SYN flood"], "citation_intent": "background"}
{"citing_id": "2303.15782v1", "cited_id": "2003.08515", "section_title": "Object Set", "citation": "For both experiments, we use 3D models from the PartNet-Mobility #REFR object set to generate a training and test data set.", "text_before_citation": [], "text_after_citation": ["Categories: While PartNet-Mobility provides more than 2000 objects from 46 categories, as done in previous work [5, Table 1 . Overview of Selected Objects.", "We select a subset of the PartNet-Mobility #OTHEREFR object set and report the amount of instances we selected per category for our training and test set.", "Our final object set has 92 objects instances for training and 25 for testing. objects. In Tab.", "1 we list the number of objects per category we selected as well as the context in which they can be used."], "citing_paper_content": {"title": "Carto: Category And Joint Agnostic Reconstruction Of Articulated Objects", "abstract": "Figure 1. Visualization of CARTO on unseen object instances. We first use CARTO to jointly detect all objects in the scene and then articulate them while keeping the predicted shape code constant."}, "cited_paper_content": {"title": "Sapien: A Simulated Part-Based Interactive Environment", "abstract": "Building home assistant robots has long been a pursuit for vision and robotics researchers. To achieve this task, a simulated environment with physically realistic simulation, sufficient articulated objects, and transferability to the real robot is indispensable. Existing environments achieve these requirements for robotics simulation with different levels of simplification and focus. We take one step further in constructing an environment that supports household tasks for training robot learning algorithm. Our work, SAPIEN, is a realistic and physics-rich simulated environment that hosts a large-scale set for articulated objects. Our SAPIEN enables various robotic vision and interaction tasks that require detailed part-level understanding.We evaluate state-of-the-art vision algorithms for part detection and motion attribute recognition as well as demonstrate robotic interaction tasks using heuristic approaches and reinforcement learning algorithms. We hope that our SAPIEN can open a lot of research directions yet to be explored, including learning cognition through interaction, part motion discovery, and construction of robotics-ready simulated game environment."}, "keywords": ["3D models"], "citation_intent": "method"}
{"citing_id": "2303.11828v1", "cited_id": "1902.10903", "section_title": "Comparison With State-Of-The-Art", "citation": "In the single scale setting, compared with the second best CNN-based method BDCN #REFR , we obtain a large performance gain by 2.3%, 2.1% and 4.5% in terms of ODS, OIS and AP.", "text_before_citation": ["In this section, we compare the performance of the proposed UAED with existing excellent edge detectors, including traditional detectors such as Canny #OTHEREFR , CNN-based detectors such as HED #OTHEREFR and RCF #OTHEREFR , and transformerbased detector EDTER #OTHEREFR .", "BSDS results. The results are summarized in Table 1 .", "We can see that our proposed UAED outperforms other previous CNN-based methods."], "text_after_citation": ["We also achieve ODS=0.844, OIS=0.864 and AP=0.905 under the MS-VOC setting, which also surpasses BDCN #OTHEREFR by a large margin (1.6%, 2.0% and 1.5%).", "Compared with transformer-based EDTER #OTHEREFR , we still achieve the best performance and increase the scores by 0.5%, 0.6%, and 1.2% in ODS, OIS and AP metrics when testing on a single scale input.", "The results are only slightly lower than EDTER under multi-scale settings.", "The possible reason is that EDTER is learned in a patch-based manner, so the results are greatly improved when testing under the multi-scale setting.", "To show the results more intuitively, we give the Precision-Recall curve in Figure 3 . The precision-recall curves on BSDS500. Fig. 3 ."], "citing_paper_content": {"title": "The Treasure Beneath Multiple Annotations: An Uncertainty-Aware Edge Detector", "abstract": "Deep learning-based edge detectors heavily rely on pixel-wise labels which are often provided by multiple annotators. Existing methods fuse multiple annotations using a simple voting process, ignoring the inherent ambiguity of edges and labeling bias of annotators. In this paper, we propose a novel uncertainty-aware edge detector (UAED), which employs uncertainty to investigate the subjectivity and ambiguity of diverse annotations. Specifically, we first convert the deterministic label space into a learnable Gaussian distribution, whose variance measures the degree of ambiguity among different annotations. Then we regard the learned variance as the estimated uncertainty of the predicted edge maps, and pixels with higher uncertainty are likely to be hard samples for edge detection. Therefore we design an adaptive weighting loss to emphasize the learning from those pixels with high uncertainty, which helps the network to gradually concentrate on the important pixels. UAED can be combined with various encoder-decoder backbones, and the extensive experiments demonstrate that UAED achieves superior performance consistently across multiple edge detection benchmarks. The source code is available at https://github.com/ZhouCX117/ UAED."}, "cited_paper_content": {"title": "Bi-Directional Cascade Network For Perceptual Edge Detection", "abstract": "Exploiting multi-scale representations is critical to improve edge detection for objects at different scales. To extract edges at dramatically different scales, we propose a Bi-Directional Cascade Network (BDCN) structure, where an individual layer is supervised by labeled edges at its specific scale, rather than directly applying the same supervision to all CNN outputs. Furthermore, to enrich multi-scale representations learned by BDCN, we introduce a Scale Enhancement Module (SEM) which utilizes dilated convolution to generate multi-scale features, instead of using deeper CNNs or explicitly fusing multi-scale edge maps. These new approaches encourage the learning of multi-scale representations in different layers and detect edges that are well delineated by their scales. Learning scale dedicated layers also results in compact network with a fraction of parameters. We evaluate our method on three datasets, i.e., BSDS500, NYUDv2, and Multicue, and achieve ODS Fmeasure of 0.828, 1.3% higher than current state-of-the art on BSDS500."}, "keywords": ["second best CNN-based"], "citation_intent": "result"}
{"citing_id": "2303.06337v1", "cited_id": "1806.09055", "section_title": "2.3.3", "citation": "The session length search part is a Neural Architecture Search (NAS) algorithm based on DARTS #REFR .", "text_before_citation": ["Short-term Interest Module.", "Now we introduce the shortterm interest module for capturing the short-term sequential dependencies in users' historical sequences.", "The short-term interest module follows a similar framework as the long-term interest module, meaning it also contains a sequence-mixer and channel-mixer.", "The key difference is the short-term interest length search section that can automatically select optimal short-term interest length in a data-driven manner. Session Length Search."], "text_after_citation": ["To be specific, we first define some candidate short-term lengths before the last item , representing the possible range of users' short-term interest length, i.e., = { 1 , ..., , ..., }, as shown in Figure 5 . Then we create SRSMLPs, each for individual candidate .", "Next, as shown in Figure 5 , we assign a set of learnable architectural weights = { 1 , ..., , ..., } to the outputs from different SRSMLPs respectively.", "After that, we apply softmax to transform weights as continuous and differentiable approximations:", "\u2026 Short-term Length Search = 2 \u2026 \u2026 \u0391 0.1 0.2 T-2 T-1 T T-2 T-1 T T \u2026 T-2 T-1 SRSMLP Layers SRSMLP Layers SRSMLP Layers SRSMLP Layers 0.2 0.5", "EQUATION"], "citing_paper_content": {"title": "Automlp: Automated Mlp For Sequential Recommendations", "abstract": "Sequential recommender systems aim to predict users' next interested item given their historical interactions. However, a longstanding issue is how to distinguish between users' long/short-term interests, which may be heterogeneous and contribute differently to the next recommendation. Existing approaches usually set predefined short-term interest length by exhaustive search or empirical experience, which is either highly inefficient or yields subpar results. The recent advanced transformer-based models can achieve state-of-the-art performances despite the aforementioned issue, but they have a quadratic computational complexity to the length of the input sequence. To this end, this paper proposes a novel sequential recommender system, AutoMLP, aiming for better modeling users' long/short-term interests from their historical interactions. In addition, we design an automated and adaptive search algorithm for preferable short-term interest length via end-to-end optimization. Through extensive experiments, we show that AutoMLP has competitive performance against state-of-the-art methods, while maintaining linear computational complexity. CCS CONCEPTS \u2022 Information systems \u2192 Recommender systems."}, "cited_paper_content": {"title": "Darts: Differentiable Architecture Search", "abstract": "This paper addresses the scalability challenge of architecture search by formulating the task in a differentiable manner. Unlike conventional approaches of applying evolution or reinforcement learning over a discrete and non-differentiable search space, our method is based on the continuous relaxation of the architecture representation, allowing efficient search of the architecture using gradient descent. Extensive experiments on CIFAR-10, ImageNet, Penn Treebank and WikiText-2 show that our algorithm excels in discovering high-performance convolutional architectures for image classification and recurrent architectures for language modeling, while being orders of magnitude faster than state-of-the-art non-differentiable techniques."}, "keywords": ["session length search", "NAS"], "citation_intent": "method"}
{"citing_id": "2303.14966v1", "cited_id": "1602.05629", "section_title": "I. Introduction", "citation": "The total training will terminate when the accuracy loss of the global model is below the preset threshold #REFR .", "text_before_citation": ["To save computing resources and protect users' privacy, Federated Learning (FL) allows resource-constrained discrete clients to collaboratively learn the global model under the orchestration of a central server while keeping privacy-sensitive data locally #OTHEREFR . * corresponding author.", "The current popular FL algorithms are the Federated Averaging (FedAvg) and its variants #OTHEREFR .", "It usually consists of a central server and a group of distributed clients.", "In each training round, discrete clients first obtain the global model parameters of the previous round from the central server and then uses their own local data to generate the local model parameters.", "Then the central server will aggregate the local model parameters from discrete clients and generate the new global parameters."], "text_after_citation": ["Under this framework, the server effectively utilizes clients' computing power and storage resources while realizing the separation of privacy data #OTHEREFR .", "Although FedAvg performs well on Gboard application #OTHEREFR , recent works #OTHEREFR have pointed out its convergence issues in some settings.", "For example, FedAvg #OTHEREFR will converge slowly and the accuracy will be reduced on the non-independent-andidentically-distributed (Non-IID) client dataset #OTHEREFR . Reddi et al.", "#OTHEREFR show that this is because of client parameter drift and lack of adaptivity to the different models. Moreover, Tan et al.", "#OTHEREFR point out that FedAvg can not learn the personalized FL model for each client."], "citing_paper_content": {"title": "Adaptive Federated Learning Via Entropy Approach", "abstract": "Federated Learning (FL) has recently emerged as a popular framework, which allows resource-constrained discrete clients to cooperatively learn the global model under the orchestration of a central server while storing privacysensitive data locally. However, due to the difference in equipment and data divergence of heterogeneous clients, there will be parameter deviation between local models, resulting in a slow convergence rate and a reduction of the accuracy of the global model. The current FL algorithms use the static client learning strategy pervasively and can not adapt to the dynamic training parameters of different clients. In this paper, by considering the deviation between different local model parameters, we propose an adaptive learning rate scheme for each client based on entropy theory to alleviate the deviation between heterogeneous clients and achieve fast convergence of the global model. It's difficult to design the optimal dynamic learning rate for each client as the local information of other clients is unknown, especially during the local training epochs without communications between local clients and the central server. To enable a decentralized learning rate design for each client, we first introduce meanfield schemes to estimate the terms related to other clients' local model parameters. Then the decentralized adaptive learning rate for each client is obtained in closed form by constructing the Hamilton equation. Moreover, we prove that there exist fixed point solutions for the mean-field estimators, and an algorithm is proposed to obtain them. Finally, extensive experimental results on real datasets show that our algorithm can effectively eliminate the deviation between local model parameters compared to other recent FL algorithms."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["global model", "total training"], "citation_intent": "background"}
{"citing_id": "2304.09975v1", "cited_id": "1611.09940", "section_title": "Ii. Machine Learning Methods For Optimization Problems In Graphs", "citation": "In 2016, researchers presented in #REFR a framework for combinatorial optimization problems using neural networks and reinforcement learning; the work focused on the TSP, which is a graph problem, but the approach was designed to work with any combinatorial optimization problem.", "text_before_citation": ["In the last few years, many machine learning-based approaches that effectively solve several different optimization problems in graphs have been proposed, although none of them designed for solving KEP.", "Graph optimization problems already solved with the help of machine learning include the Set Covering Problem #OTHEREFR , Graph Colouring #OTHEREFR , #OTHEREFR , Minimum Vertex Cover #OTHEREFR , #OTHEREFR , Maximum Cut #OTHEREFR , Graph Partitioning #OTHEREFR , Maximum Independent Set #OTHEREFR , Maximum Common Subgraph #OTHEREFR , and the Travelling Salesperson Problem (also called the travelling salesman problem or TSP), one of the most famous NP-hard problems, often used to represent the class, and some variants of it #OTHEREFR - #OTHEREFR .", "In 2015, the authors of #OTHEREFR presented a new type of neural network called Pointer Networks, designed to learn how to reorder the elements of an input sequence; they validated the method by using it to solve 3 problems, including the TSP."], "text_after_citation": ["In 2017, the authors of #OTHEREFR proposed the utilization of a combination of graph representation learning and reinforcement learning to solve graph optimization problems; they showed that their proposed approach effectively learns to solve at least three of those problems: Minimum Vertex Cover, Maximum Cut, and TSP.", "In 2018, the decision variant of the TSP, called Decision Traveling Salesman Problem (DTSP), which is to decide if a given TSP instance admits a Hamiltonian route with a cost no greater than a given threshold C, and is also NP-Hard, was solved in #OTHEREFR with a GNN.", "In 2019, the authors of #OTHEREFR tried to solve the TSP using a two stage technique that is very similar to the one presented in this study (described at section IV-C2 and illustrated at Figure  1 ): firstly, a GNN processes the input graph and create scores for each edge of the graph; then, a non-learned search heuristic, which in this case was beam search, uses these scores to construct a solution.", "There have been other approaches that use similar techniques: in #OTHEREFR the authors review graph learning methods for solving CO problems, with a focus on two-stage techniques, where the first is based on graph representation learning, which embeds the input graph into low-dimension vectors, and the second uses the embeddings learned in the first stage; #OTHEREFR surveys the use of GNNs as a model of neuralsymbolic computing and their applications, which includes combinatorial optimization problems; #OTHEREFR unifies and refines several of such two stage techniques for neural combinatorial optimization, and test it on the TSP."], "citing_paper_content": {"title": "Solving The Kidney-Exchange Problem Via Graph Neural Networks With No Supervision", "abstract": "This paper introduces a new learning-based approach for approximately solving the Kidney-Exchange Problem (KEP), an NP-hard problem on graphs. The problem consists of, given a pool of kidney donors and patients waiting for kidney donations, optimally selecting a set of donations to optimize the quantity and quality of transplants performed while respecting a set of constraints about the arrangement of these donations. The proposed technique consists of two main steps: the first is a Graph Neural Network (GNN) trained without supervision; the second is a deterministic non-learned search heuristic that uses the output of the GNN to find a valid solution. To allow for comparisons, we also implemented and tested an exact solution method using integer programming, two greedy search heuristics without the machine learning module, and the GNN alone without a heuristic. We analyze and compare the methods and conclude that the learning-based two-stage approach is the best solution quality, outputting approximate solutions on average 1.1 times more valuable than the ones from the deterministic heuristic alone."}, "cited_paper_content": {"title": "Neural Combinatorial Optimization With Reinforcement Learning", "abstract": "This paper presents a framework to tackle combinatorial optimization problems using neural networks and reinforcement learning. We focus on the traveling salesman problem (TSP) and train a recurrent neural network that, given a set of city coordinates, predicts a distribution over different city permutations. Using negative tour length as the reward signal, we optimize the parameters of the recurrent neural network using a policy gradient method. We compare learning the network parameters on a set of training graphs against learning them on individual test graphs. Without much engineering and heuristic designing, Neural Combinatorial Optimization achieves close to optimal results on 2D Euclidean graphs with up to 100 nodes. Applied to the KnapSack, another NP-hard problem, the same method obtains optimal solutions for instances with up to 200 items. These results, albeit still far from state-of-the-art, give insights into how neural networks can be used as a general tool for tackling combinatorial optimization problems."}, "keywords": ["combinatorial optimization problems"], "citation_intent": "background"}
{"citing_id": "2304.07535v1", "cited_id": "1801.06122", "section_title": "Introduction", "citation": "As an example, Shao et al., in #REFR , have highlighted the role of Twitter bots, showing how bots were primarily responsible for the early spread of disinformation, interacting with influential accounts through mentions and replies.", "text_before_citation": ["With the advent of the internet and Online Social Networks (OSNs), production and fruition of information feature less mediated procedures, where content and quality do not always go through a rigorous editorial process #OTHEREFR .", "Thus, although OSNs make our lives easier by giving us immediate access to information and allowing us to exchange opinions about anything, the danger of being exposed to false or misleading news is high #OTHEREFR .", "The promotion of disinformation on OSNs has often been juxtaposed with the existence of automated accounts known as bots."], "text_after_citation": ["The struggle between bots hunters and bots creators has been going on for many years now #OTHEREFR , and the actions of these automated accounts with malicious intent have influenced even the purchase of Twitter itself -just remember the $44 billion deal that went up in smoke precisely because of concerns about the unquantified presence of bots on the platform #OTHEREFR .", "In this study, we provide a novel approach to bot detection, by leveraging the remarkable advancements in the field of image recognition #OTHEREFR .", "To the best of the authors' knowledge, no methodology or tool so far defined for bot detection leverage image recognition.", "In particular, we will exploit the potential of Convolutional Neural Networks (CNNs) to classify Twitter accounts as bots or not.", "Based on the premise that automated accounts are often programmed to carry out spam and/or disinformation campaigns, numerous works in the literature have proposed detection approaches that leverage coordination and synchronism features of accounts, see, e.g. #OTHEREFR ."], "citing_paper_content": {"title": "From Online Behaviours To Images: A Novel Approach To Social Bot Detection", "abstract": "Online Social Networks have revolutionized how we consume and share information, but they have also led to a proliferation of content not always reliable and accurate. One particular type of social accounts is known to promote unreputable content, hyperpartisan, and propagandistic information. They are automated accounts, commonly called bots. Focusing on Twitter accounts, we propose a novel approach to bot detection: we first propose a new algorithm that transforms the sequence of actions that an account performs into an image; then, we leverage the strength of Convolutional Neural Networks to proceed with image classification. We compare our performances with state-of-the-art results for bot detection on genuine accounts / bot accounts datasets well known in the literature. The results confirm the effectiveness of the proposal, because the detection capability is on par with the state of the art, if not better in some cases. This work was partially supported by project SERICS (PE00000014) under the NRRP MUR program funded by the EU-NGEU; by the Integrated Activity Project TOFFEe (TOols for Fighting FakEs) https://toffee.imtlucca.it/; by the IIT-CNR funded Project re-DESIRE (DissEmination of ScIentific REsults 2.0); by 'Prebunking: predicting and mitigating coordinated inauthentic behaviors in social media' project, funded by Sapienza University of Rome."}, "cited_paper_content": {"title": "Anatomy Of An Online Misinformation Network", "abstract": "Massive amounts of fake news and conspiratorial content have spread over social media before and after the 2016 US Presidential Elections despite intense fact-checking efforts. How do the spread of misinformation and fact-checking compete? What are the structural and dynamic characteristics of the core of the misinformation diffusion network, and who are its main purveyors? How to reduce the overall amount of misinformation? To explore these questions we built Hoaxy, an open platform that enables large-scale, systematic studies of how misinformation and fact-checking spread and compete on Twitter. Hoaxy captures public tweets that include links to articles from low-credibility and fact-checking sources. We perform k-core decomposition on a diffusion network obtained from two million retweets produced by several hundred thousand accounts over the six months before the election. As we move from the periphery to the core of the network, fact-checking nearly disappears, while social bots proliferate. The number of users in the main core reaches equilibrium around the time of the election, with limited churn and increasingly dense connections. We conclude by quantifying how effectively the network can be disrupted by penalizing the most central nodes. These findings provide a first look at the anatomy of a massive online misinformation diffusion network."}, "keywords": ["Twitter bots"], "citation_intent": "background"}
{"citing_id": "2303.03395v1", "cited_id": "1507.06527", "section_title": "Case Study #1: A Small Network", "citation": "DRQN #REFR is a DRL method focusing on solving discrete control and POMDPs using historical information.", "text_before_citation": ["Other quantities, including average trip speed and delay, are also important to measure the control effects of the control method.", "The average trip speed is computed as the mean of average speeds for all trips, and the delay is calculated as the ratio of the gaps between free-flow TTT and current TTT over the free-flow TTT.", "Furthermore, we compared the proposed model with the no control, demonstrators (ALINEA or Gating), and other baseline methods, such as DAgger, DQN and DRQN. DAgger #OTHEREFR", "2011) is an imitation learning algorithm that iteratively learns deterministic policies from a prepared generated dataset. DQN #OTHEREFR", "2015 ) is a classic DRL method for discrete control and fully observable problem."], "text_after_citation": [], "citing_paper_content": {"title": "Demonstration-Guided Deep Reinforcement Learning For Coordinated Ramp Metering And Perimeter Control In Large Scale Networks", "abstract": "Effective traffic control methods have great potential in alleviating network congestion. Existing literature generally focuses on a single control approach, while few studies have explored the effectiveness of integrated and coordinated control approaches. This study considers two representative control approaches: ramp metering for freeways and perimeter control for homogeneous urban roads, and we aim to develop a deep reinforcement learning (DRL)-based coordinated control framework for large-scale networks. The main challenges are 1) there is a lack of efficient dynamic models for both freeways and urban roads; 2) the standard DRL method becomes ineffective due to the complex and non-stationary network dynamics. In view of this, we propose a novel meso-macro dynamic network model and first time develop a demonstration-guided DRL method to achieve large-scale coordinated ramp metering and perimeter control. The dynamic network model hybridizes the link and generalized bathtub models to depict the traffic dynamics of freeways and urban roads, respectively. For the DRL method, we incorporate demonstration to guide the DRL method for better convergence by introducing the concept of \"teacher\" and \"student\" models. The teacher models are traditional controllers (e.g., ALINEA, Gating), which provide control demonstrations. The student models are DRL methods, which learn from the teacher and aim to surpass the teacher's performance. To validate the proposed framework, we conduct two case studies in a small-scale network and a real-world large-scale traffic network in Hong Kong. Numerical results show that the proposed DRL method outperforms demonstrators as well as the state-of-the-art DRL methods, and the coordinated control is more effective than just controlling ramps or perimeters by 8.7% and 22.6%, respectively. The research outcome reveals the great potential of combining traditional controllers with DRL for coordinated control in large-scale networks."}, "cited_paper_content": {"title": "Deep Recurrent Q-Learning For Partially Observable Mdps", "abstract": "Deep Reinforcement Learning has yielded proficient controllers for complex tasks. However, these controllers have limited memory and rely on being able to perceive the complete game screen at each decision point. To address these shortcomings, this article investigates the effects of adding recurrency to a Deep Q-Network (DQN) by replacing the first post-convolutional fully-connected layer with a recurrent LSTM. The resulting \\textit{Deep Recurrent Q-Network} (DRQN), although capable of seeing only a single frame at each timestep, successfully integrates information through time and replicates DQN's performance on standard Atari games and partially observed equivalents featuring flickering game screens. Additionally, when trained with partial observations and evaluated with incrementally more complete observations, DRQN's performance scales as a function of observability. Conversely, when trained with full observations and evaluated with partial observations, DRQN's performance degrades less than DQN's. Thus, given the same length of history, recurrency is a viable alternative to stacking a history of frames in the DQN's input layer and while recurrency confers no systematic advantage when learning to play the game, the recurrent net can better adapt at evaluation time if the quality of observations changes."}, "keywords": ["discrete control", "DRQN"], "citation_intent": "method"}
{"citing_id": "2303.05823v1", "cited_id": "1911.06016", "section_title": "Linearly Implicit Methods For Semilinear Evolution Pdes", "citation": "Following the methods introduced in an ODE context in #REFR , we assume we are given an s\u02c6s real or complex matrix D and a real or complex column vector \u0398 \" p\u03b8 1 ,\u00a8\u00a8\u00a8, \u03b8 s q t , and we consider some h \u0105 0 as a time step.", "text_before_citation": ["A b L \"\u00a8a 11 L\u00a8\u00a8\u00a8a 1s L . . . . . . a s1 L\u00a8\u00a8\u00a8a ss L\u2039 \u201a.", "Similarly, we set", "b t b L \" pb 1 L,\u00a8\u00a8\u00a8, b s Lq.", "The notation I will be used for the identity operator in general.", "2.1.2. Linearly implicit methods."], "text_after_citation": ["Since the problem (1) is autonomous, we always consider t \" 0 as the starting time without loss of generality.", "We take u 0 P H \u03c3 (classical Sobolev space of order \u03c3, to be defined in the next section), and we denote by t \u00de \u00d1 uptq the exact solution of (1) with initial datum up0q \" u 0 .", "Of course, for fixed t \u011b 0, the function uptq is itself a function of x P \u2126. We will often omit this dependency in the notations.", "We initialize the linearly implicit method with u 0 P H \u03c3 , close to u 0 .", "We initialize the additional variables p\u03b3\u00b41 ,i q 1\u010fi\u010fs with approximations of pN puppc i\u00b41 qhqqq 1\u010fi\u010fs and we set \u0393\u00b41 \" p\u03b3\u00b41 ,i q t 1\u010fi\u010fs P pH \u03c3 q s ."], "citing_paper_content": {"title": "High Order Linearly Implicit Methods For Semilinear Evolution Pdes", "abstract": "This paper considers the numerical integration of semilinear evolution PDEs using the high order linearly implicit methods developped in [21] in the ODE setting. These methods use a collocation Runge-Kutta method as a basis, and additional variables that are updated explicitly and make the implicit part of the collocation Runge-Kutta method only linearly implicit. In this paper, we introduce several notions of stability for the underlying Runge-Kutta methods as well as for the explicit step on the additional variables necessary to fit the context of evolution PDE. We prove a main theorem about the high order of convergence of these linearly implicit methods in this PDE setting, using the stability hypotheses introduced before. We use nonlinear Schr\u00f6dinger equations and heat equations as main examples but our results extend beyond these two classes of evolution PDEs. We illustrate our main result numerically in dimensions 1 and 2, and we compare the efficiency of the linearly implicit methods with other methods from the litterature. We also illustrate numerically the necessity of the stability conditions of our main result."}, "cited_paper_content": {"title": "High Order Linearly Implicit Methods For Evolution Equations: How To Solve An Ode By Inverting Only Linear Systems", "abstract": "This paper introduces a new class of numerical methods for the time integration of evolution equations set as Cauchy problems of ODEs or PDEs. The systematic design of these methods mixes the Runge-Kutta collocation formalism with collocation techniques, in such a way that the methods are linearly implicit and have high order. The fact that these methods are implicit allows to avoid CFL conditions when the large systems to integrate come from the space discretization of evolution PDEs. Moreover, these methods are expected to be efficient since they only require to solve one linear system of equations at each time step, and efficient techniques from the literature can be used to do so. After the introduction of the methods, we set suitable definitions of consistency and stability for these methods. This allows for a proof that arbitrarily high order linearly implicit methods exist and converge when applied to ODEs. Eventually, we perform numerical experiments on ODEs and PDEs that illustrate our theoretical results for ODEs, and compare our methods with standard methods for several evolution PDEs."}, "keywords": ["ODE context"], "citation_intent": "method"}
{"citing_id": "2303.13401v1", "cited_id": "1912.00049", "section_title": "Numerical Optimization For Max-Loss Form", "citation": "AutoAttack also includes non-PGD methods, e.g., Square Attack #REFR , which is based on gradientfree random search.", "text_before_citation": ["x \u2212 x p \u2264 \u03b5} with p = 1, \u221e, P \u2206(x) takes simple forms.", "For p = 2, sequential projection-first onto the box and then the norm ball, at least finds a feasible solution (see our clarification of these projections in Appendix A.1). Therefore, PGD is viable for these cases.", "For other general p metrics and nonp metrics, analytical projectors mostly cannot be derived and existing PGD methods do not apply.", "Previous work has shown that the solution quality of PGD is sensitive to hyperparameter tuning, e.g., step-size schedule and iteration budget #OTHEREFR .", "PGD variants, such as AGPD methods where 'A' stands for Auto(matic) #OTHEREFR , try to make the tuning process automatic by combining a heuristic adaptive step-size schedule and momentum acceleration under a fixed iteration budget and are built into the popular AutoAttack package #OTHEREFR ."], "text_after_citation": ["Square Attack does not have as effective attack performance as APGDs (see Table 2 in later sections) but is included as #OTHEREFR states that the diversity in the evaluation algorithms is important to achieve reliable numerical RE."], "citing_paper_content": {"title": "Optimization And Optimizers For Adversarial Robustness", "abstract": "Empirical robustness evaluation (RE) of deep learning models against adversarial perturbations entails solving nontrivial constrained optimization problems. Existing numerical algorithms that are commonly used to solve them in practice predominantly rely on projected gradient, and mostly handle perturbations modeled by the 1, 2 and \u221e distances. In this paper, we introduce a novel algorithmic framework that blends a general-purpose constrained-optimization solver PyGRANSO with Constraint-Folding (PWCF), which can add more reliability and generality to the state-of-the-art RE packages, e.g., AutoAttack. Regarding reliability, PWCF provides solutions with stationarity measures and feasibility tests to assess the solution quality. For generality, PWCF can handle perturbation models that are typically inaccessible to the existing projected gradient methods; the main requirement is the distance metric to be almost everywhere differentiable. Taking advantage of PWCF and other existing numerical algorithms, we further explore the distinct patterns in the solutions found for solving these optimization problems using various combinations of losses, perturbation models, and optimization algorithms. We then discuss the implications of these patterns on the current robustness evaluation and adversarial training."}, "cited_paper_content": {"title": "Square Attack: A Query-Efficient Black-Box Adversarial Attack Via Random Search", "abstract": "We propose the Square Attack, a new score-based black-box $l_2$ and $l_\\infty$ adversarial attack that does not rely on local gradient information and thus is not affected by gradient masking. The Square Attack is based on a randomized search scheme where we select localized square-shaped updates at random positions so that the $l_\\infty$- or $l_2$-norm of the perturbation is approximately equal to the maximal budget at each step. Our method is algorithmically transparent, robust to the choice of hyperparameters, and is significantly more query efficient compared to the more complex state-of-the-art methods. In particular, on ImageNet we improve the average query efficiency for various deep networks by a factor of at least $2$ and up to $7$ compared to the recent state-of-the-art $l_\\infty$-attack of Meunier et al. while having a higher success rate. The Square Attack can even be competitive to gradient-based white-box attacks in terms of success rate. Moreover, we show its utility by breaking a recently proposed defense based on randomization. The code of our attack is available at this https URL"}, "keywords": ["gradientfree random search"], "citation_intent": "method"}
{"citing_id": "2303.09827v1", "cited_id": "1907.01669", "section_title": "Method 4.1 Semantic Representation", "citation": "Therefore, we fine-tuned the model with the multi-domain public task-oriented dialogue dataset MultiWOZ 2.2 #REFR .", "text_before_citation": ["With the advance of the pre-trained langue model, leveraging these models for embedding dialogue has exhibited promising results #OTHEREFR .", "Following their success, we utilize the pre-trained SBERT #OTHEREFR ) as a backbone model; SBERT has a siamese network structure and performs well in classification #OTHEREFR , summarization (Zhong et al., 2020) , and intent clustering #OTHEREFR ).", "The SBERT model is pre-trained on a written form text dataset that has a different linguistic pattern with the dialogue utterances.", "This difference could hurt the accuracy on dialogue related tasks #OTHEREFR ."], "text_after_citation": ["This dataset has nine intents types, and we finetuned the model to learn the embedding of utterances according to intent type.", "In this data, intents are spanned multiple turns, and some utterances contain multiple intents in one utterance.", "To clarify the match between the utterance and the intent label, we used only the first utterance of the spanned intent dialogue and excluded utterances that contained multiple intents.", "We analyze the processed fine-tuning dataset in Table 1 .", "As we aim for unsupervised intent induction, domain robust fine-tuning is crucial to identify the intents across the domain."], "citing_paper_content": {"title": "Doric : Domain Robust Fine-Tuning For Open Intent Clustering Through Dependency Parsing", "abstract": "We present our work on Track 2 in the Dialog System Technology Challenges 11 (DSTC11). DSTC11-Track2 aims to provide a benchmark for zero-shot, cross-domain, intent-set induction. In the absence of in-domain training dataset, robust utterance representation that can be used across domains is necessary to induce users' intentions. To achieve this, we leveraged a multi-domain dialogue dataset to fine-tune the language model and proposed extracting Verb-Object pairs to remove the artifacts of unnecessary information. Furthermore, we devised the method that generates each cluster's name for the explainability of clustered results. Our approach achieved 3rd place in the precision score and showed superior accuracy and normalized mutual information (NMI) score than the baseline model on various domain datasets."}, "cited_paper_content": {"title": "Multiwoz 2.1: A Consolidated Multi-Domain Dialogue Dataset With State Corrections And State Tracking Baselines", "abstract": "MultiWOZ 2.0 (Budzianowski et al., 2018) is a recently released multi-domain dialogue dataset spanning 7 distinct domains and containing over 10,000 dialogues. Though immensely useful and one of the largest resources of its kind to-date, MultiWOZ 2.0 has a few shortcomings. Firstly, there is substantial noise in the dialogue state annotations and dialogue utterances which negatively impact the performance of state-tracking models. Secondly, follow-up work (Lee et al., 2019) has augmented the original dataset with user dialogue acts. This leads to multiple co-existent versions of the same dataset with minor modifications. In this work we tackle the aforementioned issues by introducing MultiWOZ 2.1. To fix the noisy state annotations, we use crowdsourced workers to re-annotate state and utterances based on the original utterances in the dataset. This correction process results in changes to over 32% of state annotations across 40% of the dialogue turns. In addition, we fix 146 dialogue utterances by canonicalizing slot values in the utterances to the values in the dataset ontology. To address the second problem, we combined the contributions of the follow-up works into MultiWOZ 2.1. Hence, our dataset also includes user dialogue acts as well as multiple slot descriptions per dialogue state slot. We then benchmark a number of state-of-the-art dialogue state tracking models on the MultiWOZ 2.1 dataset and show the joint state tracking performance on the corrected state annotations. We are publicly releasing MultiWOZ 2.1 to the community, hoping that this dataset resource will allow for more effective models across various dialogue subproblems to be built in the future."}, "keywords": ["dialogue"], "citation_intent": "method"}
{"citing_id": "2304.10677v1", "cited_id": "1409.0575", "section_title": "Transfer Learning", "citation": "ImageNet is a well-known dataset that consists of 14 million images from 20,000 different classes #REFR .", "text_before_citation": ["In various transfer learning applications, it is proposed to utilize a pre-trained neural network with its weights and modify some parts of the model to train for fine-tuning.", "Pre-trained networks decrease the required time and resources for the training since they are already trained with intense resources by deep learning specialists #OTHEREFR .", "Pre-trained CNNs are networks that are previously trained on massive datasets by deep learning specialists.", "The VGG-19, EfficientNet-B0, ResNet50, VGG16, DenseNet121, MobileNet, Inception-v3 and InceptionResNet models are used as base models for the feature extraction during the study.", "Mentioned models are the state-of-the-art pre-trained CNNs that are trained on the ImageNet dataset."], "text_after_citation": ["Researchers have explored the use of pre-trained feature extractors from diverse domains to extract valuable insights from data. Pre-trained feature extractors deliver noteworthy results. As an example, Narin et al.", "built a feedforward neural network to classify COVID-19 patients based on chest X-Ray pictures and utilized pre-trained CNNs to extract features #OTHEREFR .", "In addition, pre-trained CNNs have been applied to extract features from scalograms (combined with SVM classifiers) to detect schizophrenia #OTHEREFR .", "Furthermore, pre-trained CNNs have been utilized to extract features from EEG images to identify neonatal seizures #OTHEREFR .", "Recently, various events recorded by \u03a6-OTDR sensing system were classified using the pre-trained AlexNet architecture followed by an SVM classifier #OTHEREFR ."], "citing_paper_content": {"title": "Deep Reproductive Feature Generation Framework For The Diagnosis Of Covid-19 And Viral Pneumonia Using Chest X-Ray Images", "abstract": "The rapid and accurate detection of COVID-19 cases is critical for timely treatment and preventing the spread of the disease. In this study, a two-stage feature extraction framework using eight state-ofthe-art pre-trained deep Convolutional Neural Networks (CNNs) and an autoencoder is proposed to determine the health conditions of patients (COVID-19, Normal, Viral Pneumonia) based on chest X-rays. The X-ray scans are divided into four equally sized sections and analyzed by deep pre-trained CNNs. Subsequently, an autoencoder with three hidden layers is trained to extract reproductive features from the concatenated ouput of CNNs. To evaluate the performance of the proposed framework, three different classifiers, which are single-layer perceptron (SLP), multi-layer perceptron (MLP), and support vector machine (SVM) are used. Furthermore, the deep CNN architectures are used to create benchmark models and trained on the same dataset for comparision. The proposed framework outperforms other frameworks wih pre-trained feature extractors in binary classification and shows competitive results in three-class classification. The proposed methodology is taskindependent and suitable for addressing various problems. The results show that the discriminative features are a subset of the reproductive features, suggesting that extracting task-independent features is superior to the extraction only task-based features. The flexibility and task-independence of the reproductive features make the conceptive information approach more favorable. The proposed methodology is novel and shows promising results for analyzing medical image data."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["ImageNet"], "citation_intent": "background"}
{"citing_id": "2304.10295v1", "cited_id": "1508.07909", "section_title": "Datasets And Pre-Processing", "citation": "We tokenize and truecase all text via Mose and use BPE #REFR to learn a vocabulary on each translation direction. The vocabulary size is set to 8000.", "text_before_citation": ["We conduct experiments on MuST-C dataset #OTHEREFR to verify the effectiveness of DNKD.", "It contains translation from English (En) to 8 languages: German (De), French (Fr), Russian (Ru), Spanish (Es), Italian (It), Romanian (Ro), Portuguese (Pt), and Dutch (Nl).", "We use dev set for validation and tst-COMMON set for test.", "For speech input, we use the raw 16-bit 16kHz monochannel audio wave."], "text_after_citation": [], "citing_paper_content": {"title": "Decoupled Non-Parametric Knowledge Distillation For End-To-End Speech Translation", "abstract": "Existing techniques often attempt to make knowledge transfer from a powerful machine translation (MT) to speech translation (ST) model with some elaborate techniques, which often requires transcription as extra input during training. However, transcriptions are not always available, and how to improve the ST model performance without transcription, i.e., data efficiency, has rarely been studied in the literature. In this paper, we propose Decoupled Non-parametric Knowledge Distillation (DNKD) from data perspective to improve the data efficiency. Our method follows the knowledge distillation paradigm. However, instead of obtaining the teacher distribution from a sophisticated MT model, we construct it from a non-parametric datastore via k-Nearest-Neighbor (kNN) retrieval, which removes the dependence on transcription and MT model. Then we decouple the classic knowledge distillation loss into target and non-target distillation to enhance the effect of the knowledge among non-target logits, which is the prominent \"dark knowledge\". Experiments on MuST-C corpus show that, the proposed method can achieve consistent improvement over the strong baseline without requiring any transcription."}, "cited_paper_content": {"title": "Neural Machine Translation Of Rare Words With Subword Units", "abstract": "Neural machine translation (NMT) models typically operate with a fixed vocabulary, but translation is an open-vocabulary problem. Previous work addresses the translation of out-of-vocabulary words by backing off to a dictionary. In this paper, we introduce a simpler and more effective approach, making the NMT model capable of open-vocabulary translation by encoding rare and unknown words as sequences of subword units. This is based on the intuition that various word classes are translatable via smaller units than words, for instance names (via character copying or transliteration), compounds (via compositional translation), and cognates and loanwords (via phonological and morphological transformations). We discuss the suitability of different word segmentation techniques, including simple character ngram models and a segmentation based on the byte pair encoding compression algorithm, and empirically show that subword models improve over a back-off dictionary baseline for the WMT 15 translation tasks English!German and English!Russian by up to 1.1 and 1.3 BLEU, respectively."}, "keywords": ["translation direction", "vocabulary size"], "citation_intent": "method"}
{"citing_id": "2303.04453v1", "cited_id": "1410.4084", "section_title": "Graph Parameters", "citation": "The graphs of bounded functionality have at most factorial speed of growth #REFR . However, whether they admit an implicit representation is wide-open.", "text_before_citation": ["More generally, bounded degeneracy in a class provides us with an implicit representation, where the degeneracy of a graph G is the minimum k such that every induced subgraph of G contains a vertex of degree at most k.", "Spinrad showed in #OTHEREFR that bounded clique-width also yields an implicit representation.", "The recently introduced parameter twin-width generalizes clique-width in the sense that bounded clique-width implies bounded twin-width, but not vice versa.", "It was shown in #OTHEREFR that bounded twin-width also implies the existence of an implicit representation.", "The notion of graph functionality, introduced in [2], generalizes both degeneracy and twin-width in the sense that bounded degeneracy or bounded twin-width implies bounded functionality, but not vice versa."], "text_after_citation": ["To approach this question, in Section 3.1 we analyse a parameter intermediate between twin-width and functionality.", "Then in Sections 3.2 and 3.3, we introduce more parameters and report some results concerning them."], "citing_paper_content": {"title": "Graph Parameters, Implicit Representations And Factorial Properties *", "abstract": "How to efficiently represent a graph in computer memory is a fundamental data structuring question. In the present paper, we address this question from a combinatorial point of view. A representation of an n-vertex graph G is called implicit if it assigns to each vertex of G a binary code of length O(log n) so that the adjacency of two vertices is a function of their codes. A necessary condition for a hereditary class X of graphs to admit an implicit representation is that X has at most factorial speed of growth. This condition, however, is not sufficient, as was recently shown in [Hatami & Hatami, FOCS 2022]. Several sufficient conditions for the existence of implicit representations deal with boundedness of some parameters, such as degeneracy or clique-width. In the present paper, we analyze more graph parameters and prove a number of new results related to implicit representation and factorial properties."}, "cited_paper_content": {"title": "Implicit Representations And Factorial Properties Of Graphs", "abstract": "The idea of implicit representation of graphs was introduced in [S. Kannan, M. Naor, S. Rudich, Implicit representation of graphs, SIAM J. Discrete Mathematics, 5 (1992) 596--603] and can be defined as follows. A representation of an $n$-vertex graph $G$ is said to be implicit if it assigns to each vertex of $G$ a binary code of length $O(\\log n)$ so that the adjacency of two vertices is a function of their codes. Since an implicit representation of an $n$-vertex graph uses $O(n\\log n)$ bits, any class of graphs admitting such a representation contains $2^{O(n\\log n)}$ labelled graphs with $n$ vertices. In the terminology of [J. Balogh, B. Bollob\\'{a}s, D. Weinreich, The speed of hereditary properties of graphs, J. Combin. Theory B 79 (2000) 131--156] such classes have at most factorial speed of growth. In this terminology, the implicit graph conjecture can be stated as follows: every class with at most factorial speed of growth which is hereditary admits an implicit representation. The question of deciding whether a given hereditary class has at most factorial speed of growth is far from being trivial. In the present paper, we introduce a number of tools simplifying this question. Some of them can be used to obtain a stronger conclusion on the existence of an implicit representation. We apply our tools to reveal new hereditary classes with the factorial speed of growth. For many of them we show the existence of an implicit representation."}, "keywords": ["graphs", "implicit representation"], "citation_intent": "background"}
{"citing_id": "2303.07713v1", "cited_id": "1706.04048", "section_title": "Properties Of Benamou-Brenier Energy And The Proposed Model", "citation": "Then by #REFR and Proposition 3, we obtain (t \u2192 \u03c1 n t ) \u2208 C w ([0, 1]; M + (\u2126)).", "text_before_citation": ["Therefore, by Proposition 2, it follows that \u03c1 n \u2265 0,", "EQUATION", "By (13), we also have (\u03c1 n , m n ) \u2208 D and in particular \u03c1 n 0 = \u00b5 0 .", "Consequently, it follows from Proposition 1 that \u03c1 n = dt\u2297\u03c1 n t for some \u03c1 n t \u2208 M + (\u2126), t \u2208 [0, 1].", "In particular, m n = dt\u2297(v n t \u03c1 n t )."], "text_after_citation": ["Moreover, by Proposition 1, we have that \u03c1 n t (\u2126) is constant with respect to time t, i.e., \u03c1 n t (\u2126) = \u03c1 n 0 (\u2126), for t \u2208 (0, 1].", "Since \u00b5 0 \u2208 M + (\u2126), there exists a constant C 2 > 0 such that", "EQUATION", "It is then easy to see that there exists a subsequence (not relabeled) \u03c1 n * \u03c1 for some \u03c1 \u2208 M(X). Also, by Lemma 2, we have", "|m n |(X) \u2264 B(\u03c1 n , m n )\u03c1 n (X), It follows that sup n |m n |(X) \u2264 sup n B(\u03c1 n , m n )\u03c1 n (X) < C 1 C 2 ,"], "citing_paper_content": {"title": "An Optimal Transport Regularized Model To Image Reconstruction Problems", "abstract": "Optimal transport problem has gained much attention in image processing field, such as computer vision, image interpolation and medical image registration. In this paper, we incorporate optimal transport into linear inverse problems as a regularization technique. We establish a new variational model based on Benamou-Brenier energy to regularize the evolution path from a template to latent image dynamically. The initial state of the continuity equation can be regarded as a template, which can provide priors for the reconstructed images. Also, we analyze the existence of solutions of such variational problem in Radon measure space. Moreover, the first-order primal-dual algorithm is constructed for solving this general imaging problem in a special grid strategy. Finally, numerical experiments for undersampled MRI reconstruction are presented which show that our proposed model can recover images well with high quality and structure preservation."}, "cited_paper_content": {"title": "Indirect Image Registration With Large Diffeomorphic Deformations", "abstract": "The paper adapts the large deformation diffeomorphic metric mapping framework for image registration to the indirect setting where a template is registered against a target that is given through indirect noisy observations. The registration uses diffeomorphisms that transform the template through a (group) action. These diffeomorphisms are generated by solving a flow equation that is defined by a velocity field with certain regularity. The theoretical analysis includes a proof that indirect image registration has solutions (existence) that are stable and that converge as the data error tends so zero, so it becomes a well-defined regularization method. The paper concludes with examples of indirect image registration in 2D tomography with very sparse and/or highly noisy data."}, "keywords": ["\u2208 C w", "\u2126"], "citation_intent": "background"}
{"citing_id": "2303.10502v1", "cited_id": "1801.05617", "section_title": "A. Cyberbullying Text Detection", "citation": "This automatic detection included discovering cyberbullying signals, such as the various forms of cyberbullying, and identifying posts from bullies, victims, and bystanders #REFR .", "text_before_citation": ["The results of their experiments showed 97% accuracy and 96% for F1-Score.", "According to the researchers, the results of their studies performed very well in detecting cyberbullying in both languages.", "In this way, their model can therefore detect multiple languages in the cyberbullying space #OTHEREFR , #OTHEREFR .", "Automatic detection of cyberbullying in social media text by Van Hee et al.", "#OTHEREFR aimed at exploring cyberbullying automatic detection on social media."], "text_after_citation": ["This study leveraged a binary classification to detect cyberbullying in English and Dutch.", "In addition, the paper highlighted the issue of few publicly available datasets, making it difficult for researchers to conduct constructive research addressing cyberbullying.", "To conclude, the classification revealed that the binary approach is promising to detect cyberbullying automatically in the future.", "However, a dive into using deep learning techniques to detect cyberbullying automatically would improve the ability of the classifier to predict cyberbullying content. In #OTHEREFR , Lu et al.", "proposed a Character-level Convolutional Neural Network with shortcuts (Char-CNNS) for the detection of cyberbullying in a social media text."], "citing_paper_content": {"title": "Cyberbullying In Text Content Detection: An Analytical Review", "abstract": "Technological advancements have resulted in an exponential increase in the use of online social networks (OSNs) worldwide. While online social networks provide a great communication medium, they also increase the user's exposure to life-threatening situations such as suicide, eating disorder, cybercrime, compulsive behavior, anxiety, and depression. To tackle the issue of cyberbullying, most existing literature focuses on developing approaches to identifying factors and understanding the textual factors associated with cyberbullying. While most of these approaches have brought great success in cyberbullying research, data availability needed to develop model detection remains a challenge in the research space. This paper conducts a comprehensive literature review to provide an understanding of cyberbullying detection."}, "cited_paper_content": {"title": "Automatic Detection Of Cyberbullying In Social Media Text", "abstract": "While social media offer great communication opportunities, they also increase the vulnerability of young people to threatening situations online. Recent studies report that cyberbullying constitutes a growing problem among youngsters. Successful prevention depends on the adequate detection of potentially harmful messages and the information overload on the Web requires intelligent systems to identify potential risks automatically. The focus of this paper is on automatic cyberbullying detection in social media text by modelling posts written by bullies, victims, and bystanders of online bullying. We describe the collection and fine-grained annotation of a cyberbullying corpus for English and Dutch and perform a series of binary classification experiments to determine the feasibility of automatic cyberbullying detection. We make use of linear support vector machines exploiting a rich feature set and investigate which information sources contribute the most for the task. Experiments on a hold-out test set reveal promising results for the detection of cyberbullying-related posts. After optimisation of the hyperparameters, the classifier yields an F1 score of 64% and 61% for English and Dutch respectively, and considerably outperforms baseline systems."}, "keywords": ["cyberbullying signals", "cyberbullying"], "citation_intent": "background"}
{"citing_id": "2303.15472v1", "cited_id": "1911.05932", "section_title": "Related Work", "citation": "GIFT #REFR constructs groupequivariant features by rotating or rescaling the images, and then collapses the group dimension using bilinear pooling to obtain invariant local descriptors.", "text_before_citation": ["Several studies #OTHEREFR suggest extracting local descriptors by invariant mapping of the order-based gradient histogram of a patch.", "However, these classical methods for shallow gradient-based feature maps cannot be applied to deep feature maps from standard neural networks, in which rotation induces unpredictable feature variations.", "Therefore, we propose a deep end-to-end pipeline to obtain orientation-normalized local descriptors by utilizing rotation-equivariant CNNs #OTHEREFR with additional losses.", "Learning-based invariant local descriptors.", "A branch of learning-based methods learns to obtain invariant local descriptors in an explicit manner."], "text_after_citation": ["However, their groups are limited to non-cyclic discrete rotations ranging from \u221290 \u2022 to 90", "\u2022 .", "Furthermore, their reliance on data augmentation implies a lower sampling efficiency compared to groupequivariant networks.", "LISRD #OTHEREFR jointly learns meta descriptors with different levels of regional variations and selects the most appropriate level of invariance given the context.", "Another branch of learning methods aims to learn the invariance implicitly using descriptor similarity losses from the image pair using camera pose or homography supervision."], "citing_paper_content": {"title": "Learning Rotation-Equivariant Features For Visual Correspondence", "abstract": "Extracting discriminative local features that are invariant to imaging variations is an integral part of establishing correspondences between images. In this work, we introduce a self-supervised learning framework to extract discriminative rotation-invariant descriptors using groupequivariant CNNs. Thanks to employing group-equivariant CNNs, our method effectively learns to obtain rotationequivariant features and their orientations explicitly, without having to perform sophisticated data augmentations. The resultant features and their orientations are further processed by group aligning, a novel invariant mapping technique that shifts the group-equivariant features by their orientations along the group dimension. Our group aligning technique achieves rotation-invariance without any collapse of the group dimension and thus eschews loss of discriminability. The proposed method is trained end-to-end in a self-supervised manner, where we use an orientation alignment loss for the orientation estimation and a contrastive descriptor loss for robust local descriptors to geometric/photometric variations. Our method demonstrates state-of-the-art matching accuracy among existing rotationinvariant descriptors under varying rotation and also shows competitive results when transferred to the task of keypoint matching and camera pose estimation."}, "cited_paper_content": {"title": "Gift: Learning Transformation-Invariant Dense Visual Descriptors Via Group Cnns", "abstract": "Finding local correspondences between images with different viewpoints requires local descriptors that are robust against geometric transformations. An approach for transformation invariance is to integrate out the transformations by pooling the features extracted from transformed versions of an image. However, the feature pooling may sacrifice the distinctiveness of the resulting descriptors. In this paper, we introduce a novel visual descriptor named Group Invariant Feature Transform (GIFT), which is both discriminative and robust to geometric transformations. The key idea is that the features extracted from the transformed versions of an image can be viewed as a function defined on the group of the transformations. Instead of feature pooling, we use group convolutions to exploit underlying structures of the extracted features on the group, resulting in descriptors that are both discriminative and provably invariant to the group of transformations. Extensive experiments show that GIFT outperforms state-of-the-art methods on several benchmark datasets and practically improves the performance of relative pose estimation."}, "keywords": ["invariant local descriptors"], "citation_intent": "method"}
{"citing_id": "2305.02888v1", "cited_id": "1904.08405", "section_title": "Introduction", "citation": "The fundamental shift in operation grants event cameras higher dynamic range and higher temporal resolution than most conventional shutter cameras #REFR .", "text_before_citation": ["Neuromorphic sensing describes a class of technologies designed to mimic biological sensory and perceptual functions.", "One such device in the field of neuromorphic vision is an event camera.", "Instead of using a conventional shutter-based approach to capture an image, an event camera asynchronously streams 'events' whenever a pixel detects a relative change in brightness.", "Each event has 4 parameters: A timestamp, an x and y coordinate corresponding to the location of the pixel that reported the event, and a polarity to denote whether an increase or decrease in brightness triggered the event.", "Since events are only produced by motion or lighting changes in an otherwise still scene, the output data contains minimal redundant information of components that haven't changed over time."], "text_after_citation": ["Events are logged with microsecond precision and can reach equivalent framerates exceeding 10,000 frames per second.", "These features make event cameras extremely practical sensors for the analysis of motion.", "Drowsy driving is one of the leading causes of motor accidents globally, increasing a driver's risk of an accident by a factor of 5 or more compared to when they are alert #OTHEREFR .", "Moreover, in the context of autonomous driving, drowsiness detection for the driver is important as it provides a key indicator that the DMS should present warnings and reduce the dependence of the vehicular control systems on the driver.", "Many existing drowsiness detection systems tend to focus on analysis of eye-regions and the blinking behaviour of the subject #OTHEREFR ."], "citing_paper_content": {"title": "Neuromorphic Sensing For Yawn Detection In Driver Drowsiness", "abstract": "Driver monitoring systems (DMS) are a key component of vehicular safety and essential for the transition from semiautonomous to fully autonomous driving. A key task for DMS is to ascertain the cognitive state of a driver and to determine their level of tiredness. Neuromorphic vision systems, based on event camera technology, provide advanced sensing of facial characteristics, in particular the behavior of a driver's eyes. This research explores the potential to extend neuromorphic sensing techniques to analyze the entire facial region, detecting yawning behaviors that give a complimentary indicator of tiredness. A neuromorphic dataset is constructed from 952 video clips (481 yawns, 471 not-yawns) captured with an RGB colour camera, with 37 subjects. A total of 95,200 neuromorphic image frames are generated from this video data using a video-to-event converter. From these data 21 subjects were selected to provide a training dataset, 8 subjects were used for validation data, and the remaining 8 subjects were reserved for an 'unseen' test dataset. An additional 12,300 frames were generated from event simulations of a public dataset to test against other methods. A CNN with self-attention and a recurrent head was designed, trained, and tested with these data. Respective precision and recall scores of 95.9% and 94.7% were achieved on our test set, and 89.9% and 91% on the simulated public test set, demonstrating the feasibility to add yawn detection as a sensing component of a neuromorphic DMS."}, "cited_paper_content": {"title": "Event-Based Vision: A Survey", "abstract": "Event cameras are bio-inspired sensors that work radically different from traditional cameras. Instead of capturing images at a fixed rate, they measure per-pixel brightness changes asynchronously. This results in a stream of events, which encode the time, location and sign of the brightness changes. Event cameras posses outstanding properties compared to traditional cameras: very high dynamic range (140 dB vs. 60 dB), high temporal resolution (in the order of microseconds), low power consumption, and do not suffer from motion blur. Hence, event cameras have a large potential for robotics and computer vision in challenging scenarios for traditional cameras, such as high speed and high dynamic range. However, novel methods are required to process the unconventional output of these sensors in order to unlock their potential. This paper provides a comprehensive overview of the emerging field of event-based vision, with a focus on the applications and the algorithms developed to unlock the outstanding properties of event cameras. We present event cameras from their working principle, the actual sensors that are available and the tasks that they have been used for, from low-level vision (feature detection and tracking, optic flow, etc.) to high-level vision (reconstruction, segmentation, recognition). We also discuss the techniques developed to process events, including learning-based techniques, as well as specialized processors for these novel sensors, such as spiking neural networks. Additionally, we highlight the challenges that remain to be tackled and the opportunities that lie ahead in the search for a more efficient, bio-inspired way for machines to perceive and interact with the world."}, "keywords": ["conventional shutter cameras", "higher temporal resolution"], "citation_intent": "background"}
{"citing_id": "2303.12017v1", "cited_id": "1905.04132", "section_title": "Refinement Of Background Scene Flow.", "citation": "Next, we estimate the ego-motion by fitting and decomposing essential matrices from the background flow map using a neural-guided RANSAC #REFR .", "text_before_citation": ["Since most background objects in KITTI are rigid (e.g.", "ground, buildings, etc), we can refine the background scene flow using a rigidity refinement step.", "Specifically, we employ DDRNet-Slim #OTHEREFR , a lightweight 2D semantic segmentation network, to determine the rigid background.", "DDRNet-Slim is pretrained on Cityscapes #OTHEREFR and then fine-tuned on KITTI."], "text_after_citation": ["Finally, the background scene flow is refined using the ego motion and the disparity of the first frame.", "Comparison with State-of-the-art Methods.", "We submit our approach to the website of the KITTI Scene Flow benchmark and report the leaderboard in Tab. 3. A visualized comparison is shown in Fig. 10 .", "Notably, without leveraging any rigid body assumption, CamLiRAFT performs on par with the previous state-of-the-art method RigidMask #OTHEREFR (SF-all: 4.97% vs. 4.89%).", "In contrast, RigidMask leverages more strict rigid-body assumptions by assigning rigid motion to all objects and employs 10 \u00d7 more parameters."], "citing_paper_content": {"title": "Learning Optical Flow And Scene Flow With Bidirectional Camera-Lidar Fusion", "abstract": "In this paper, we study the problem of jointly estimating the optical flow and scene flow from synchronized 2D and 3D data. Previous methods either employ a complex pipeline that splits the joint task into independent stages, or fuse 2D and 3D information in an \"early-fusion\" or \"late-fusion\" manner. Such one-size-fits-all approaches suffer from a dilemma of failing to fully utilize the characteristic of each modality or to maximize the inter-modality complementarity. To address the problem, we propose a novel end-to-end framework, which consists of 2D and 3D branches with multiple bidirectional fusion connections between them in specific layers. Different from previous work, we apply a point-based 3D branch to extract the LiDAR features, as it preserves the geometric structure of point clouds. To fuse dense image features and sparse point features, we propose a learnable operator named bidirectional camera-LiDAR fusion module (Bi-CLFM). We instantiate two types of the bidirectional fusion pipeline, one based on the pyramidal coarse-to-fine architecture (dubbed CamLiPWC), and the other one based on the recurrent all-pairs field transforms (dubbed CamLiRAFT). On FlyingThings3D, both CamLiPWC and CamLiRAFT surpass all existing methods and achieve up to a 47.9% reduction in 3D end-point-error from the best published result. Our best-performing model, CamLiRAFT, achieves an error of 4.26% on the KITTI Scene Flow benchmark, ranking 1st among all submissions with much fewer parameters. Besides, our methods have strong generalization performance and the ability to handle non-rigid motion. Code is available at https://github.com/MCG-NJU/CamLiFlow."}, "cited_paper_content": {"title": "Neural-Guided Ransac: Learning Where To Sample Model Hypotheses", "abstract": "We present Neural-Guided RANSAC (NG-RANSAC), an extension to the classic RANSAC algorithm from robust optimization. NG-RANSAC uses prior information to improve model hypothesis search, increasing the chance of finding outlier-free minimal sets. Previous works use heuristic side-information like hand-crafted descriptor distance to guide hypothesis search. In contrast, we learn hypothesis search in a principled fashion that lets us optimize an arbitrary task loss during training, leading to large improvements on classic computer vision tasks. We present two further extensions to NG-RANSAC. Firstly, using the inlier count itself as training signal allows us to train neural guidance in a self-supervised fashion. Secondly, we combine neural guidance with differentiable RANSAC to build neural networks which focus on certain parts of the input data and make the output predictions as good as possible. We evaluate NG-RANSAC on a wide array of computer vision tasks, namely estimation of epipolar geometry, horizon line estimation and camera re-localization. We achieve superior or competitive results compared to state-of-the-art robust estimators, including very recent, learned ones."}, "keywords": ["background flow map", "neural-guided RANSAC"], "citation_intent": "method"}
{"citing_id": "2303.04611v1", "cited_id": "1901.05573", "section_title": "Using Multi-Objective Metrics", "citation": "This performance surprisingly differs from the performance for OneMax since, as shown in #REFR , the var-ctrl shows competitive results for OneMax.", "text_before_citation": ["For reproducibility, we provide the codes and data in GitHub [20] .", "3.2.1 Results for OneMinMax.", "Table 1 lists the average running time to find the full Pareto set for the GSEMO variants, which use multi-objective metrics to guide self-adaptation.", "Compared to the GSEMO using static mutation rate, we observe that the two-rate GSEMO presents better results for OneMinMax.", "However, the lognormal and var-ctrl use a significantly large number of function evaluations to converge to the full Pareto front."], "text_after_citation": ["Therefore, we plot in Figure 1 the average function evaluations to find each Pareto solution.", "Because all solutions to OneMinMax are located at the Pareto front, this figure can represent the complete convergence procedure.", "We observe that the self-adaptation methods outperform the static setting in the premature optimization stage.", "However, the performance of log-normal and var-ctrl deteriorates later in the convergence process."], "citing_paper_content": {"title": "What Performance Indicators To Use For Self-Adaptation In Multi-Objective Evolutionary Algorithms", "abstract": "Parameter control has succeeded in accelerating the convergence process of evolutionary algorithms. Empirical and theoretical studies for classic pseudo-Boolean problems, such as OneMax, Leadin-gOnes, etc., have explained the impact of parameters and helped us understand the behavior of algorithms for single-objective optimization. In this work, by transmitting the techniques of single-objective optimization, we perform an extensive experimental investigation into the behavior of the self-adaptive GSEMO variants. We test three self-adaptive mutation techniques designed for single-objective optimization for the OneMinMax, COCZ, LOTZ, and OneJumpZeroJump problems. While adopting these techniques for the GSEMO algorithm, we consider different performance metrics based on the current non-dominated solution set. These metrics are used to guide the self-adaption process. Our results indicate the benefits of self-adaptation for the tested benchmark problems. We reveal that the choice of metrics significantly affects the performance of the self-adaptive algorithms. The self-adaptation methods based on the progress in one objective can perform better than the methods using multi-objective metrics such as hypervolume, inverted generational distance, and the number of the obtained Pareto solutions. Moreover, we find that the self-adaptive methods benefit from the large population size for OneMinMax and COCZ."}, "cited_paper_content": {"title": "Interpolating Local And Global Search By Controlling The Variance Of Standard Bit Mutation", "abstract": "A key property underlying the success of evolutionary algorithms (EAs) is their global search behavior, which allows the algorithms to `jump' from a current state to other parts of the search space, thereby avoiding to get stuck in local optima. This property is obtained through a random choice of the radius at which offspring are sampled from previously evaluated solutions. It is well known that, thanks to this global search behavior, the probability that an EA using standard bit mutation finds a global optimum of an arbitrary function $f:\\{0,1\\}^n \\to \\mathbb{R}$ tends to one as the number of function evaluations grows. This advantage over heuristics using a fixed search radius, however, comes at the cost of using non-optimal step sizes also in those regimes in which the optimal rate is stable for a long time. This downside results in significant performance losses for many standard benchmark problems. ::: We introduce in this work a simple way to interpolate between the random global search of EAs and their deterministic counterparts which sample from a fixed radius only. To this end, we introduce \\emph{normalized standard bit mutation}, in which the binomial choice of the search radius is replaced by a normal distribution. Normalized standard bit mutation allows a straightforward way to control its variance, and hence the degree of randomness involved. We experiment with a self-adjusting choice of this variance, and demonstrate its effectiveness for the two classic benchmark problems LeadingOnes and OneMax. Our work thereby also touches a largely ignored question in discrete evolutionary computation: multi-dimensional parameter control."}, "keywords": ["OneMax", "performance"], "citation_intent": "result"}
{"citing_id": "2304.09162v1", "cited_id": "1511.06452", "section_title": "A. Pair-Based Deep Metric Learning", "citation": "Lifted structure loss #REFR make full use of all the samples in mini-batch and mine the features of all pairs.", "text_before_citation": ["It consists of samples of the same class into positive pairs, and different classes of samples into negative pairs.", "It requires positive pairs to be close to each other and negative pairs to be far away from each other.", "Triplet loss #OTHEREFR takes each sample as a query sample, and selects samples from the same class and different classes to form a triplet.", "It requires the distance between negative pairs in the triplet to be greater than that between positive pairs.", "Many subsequent studies are carried out on the basis of triplet, such as N-pair loss #OTHEREFR , Beyond Triplet loss #OTHEREFR , etc."], "text_after_citation": ["It does not distinguish between query samples or positive samples.", "forcing the distance between each sample in the pair and all other negative samples to be greater than the given threshold.", "Multi-similarity loss #OTHEREFR can fully utilize both positive and negative pairs by a more generalized weighting strategy.", "Concretely, Multi-similarity loss mines informative pairs by the triplet criterion for both positive and negative pairs, then respectively assigns different weights for them.", "Circle loss #OTHEREFR use a unified form, which is compatible with both class-level labels and pair-wise labels."], "citing_paper_content": {"title": "Robust Calibrate Proxy Loss For Deep Metric Learning", "abstract": "The mainstream researche in deep metric learning can be divided into two genres: proxy-based and pair-based methods. Proxy-based methods have attracted extensive attention due to the lower training complexity and fast network convergence. However, these methods have limitations as the poxy optimization is done by network, which makes it challenging for the proxy to accurately represent the feature distrubtion of the real class of data. In this paper, we propose a Calibrate Proxy (CP) structure, which uses the real sample information to improve the similarity calculation in proxy-based loss and introduces a calibration loss to constraint the proxy optimization towards the center of the class features. At the same time, we set a small number of proxies for each class to alleviate the impact of intra-class differences on retrieval performance. The effectiveness of our method is evaluated by extensive experiments on three public datasets and multiple synthetic label-noise datasets. The results show that our approach can effectively improve the performance of commonly used proxy-based losses on both regular and noisy datasets."}, "cited_paper_content": {"title": "Deep Metric Learning Via Lifted Structured Feature Embedding", "abstract": "Learning the distance metric between pairs of examples is of great importance for learning and visual recognition. With the remarkable success from the state of the art convolutional neural networks, recent works have shown promising results on discriminatively training the networks to learn semantic feature embeddings where similar examples are mapped close to each other and dissimilar examples are mapped farther apart. In this paper, we describe an algorithm for taking full advantage of the training batches in the neural network training by lifting the vector of pairwise distances within the batch to the matrix of pairwise distances. This step enables the algorithm to learn the state of the art feature embedding by optimizing a novel structured prediction objective on the lifted problem. Additionally, we collected Online Products dataset: 120k images of 23k classes of online products for metric learning. Our experiments on the CUB-200-2011, CARS196, and Online Products datasets demonstrate significant improvement over existing deep feature embedding methods on all experimented embedding sizes with the GoogLeNet network."}, "keywords": ["Lifted structure loss"], "citation_intent": "method"}
{"citing_id": "2303.08691v1", "cited_id": "1512.04433", "section_title": "Signal Recovery Preliminaries", "citation": "Oymak and Recht #REFR present a similar result, stated in terms of the Gaussian width 3 of the signal set instead of the box-counting dimension.", "text_before_citation": ["This result extends Theorem 2 in #OTHEREFR , which holds for k-sparse sets only, to general low-dimensional sets and is included in the Appendix.", "For example, if X is the intersection of L (s + 1)-dimensional subspaces with the unit sphere, Theorem 1 holds with constant 0 = (3 s L) \u2212 1 k\u2212s and k > s.", "This theorem tells us that we can recover sparse signals from binary measurements up to an error of", "O( k m log nm k )", "which is sharp, up to the logarithmic factor #OTHEREFR ."], "text_after_citation": [], "citing_paper_content": {"title": "Learning To Reconstruct Signals From Binary Measurements", "abstract": "Recent advances in unsupervised learning have highlighted the possibility of learning to reconstruct signals from noisy and incomplete linear measurements alone. These methods play a key role in medical and scientific imaging and sensing, where ground truth data is often scarce or difficult to obtain. However, in practice measurements are not only noisy and incomplete but also quantized. Here we explore the extreme case of learning from binary observations and provide necessary and sufficient conditions on the number of measurements required for identifying a set of signals from incomplete binary data. Our results are complementary to existing bounds on signal recovery from binary measurements. Furthermore, we introduce a novel self-supervised learning approach, which we name SSBM, that only requires binary data for training. We demonstrate in a series of experiments with real datasets that SSBM performs on par with supervised learning and outperforms sparse reconstruction methods with a fixed wavelet basis by a large margin."}, "cited_paper_content": {"title": "Near-Optimal Bounds For Binary Embeddings Of Arbitrary Sets", "abstract": "We study embedding a subset $K$ of the unit sphere to the Hamming cube $\\{-1,+1\\}^m$. We characterize the tradeoff between distortion and sample complexity $m$ in terms of the Gaussian width $\\omega(K)$ of the set. For subspaces and several structured sets we show that Gaussian maps provide the optimal tradeoff $m\\sim \\delta^{-2}\\omega^2(K)$, in particular for $\\delta$ distortion one needs $m\\approx\\delta^{-2}{d}$ where $d$ is the subspace dimension. For general sets, we provide sharp characterizations which reduces to $m\\approx{\\delta^{-4}}{\\omega^2(K)}$ after simplification. We provide improved results for local embedding of points that are in close proximity of each other which is related to locality sensitive hashing. We also discuss faster binary embedding where one takes advantage of an initial sketching procedure based on Fast Johnson-Lindenstauss Transform. Finally, we list several numerical observations and discuss open problems."}, "keywords": ["signal", "box-counting dimension"], "citation_intent": "result"}
{"citing_id": "2304.07387v1", "cited_id": "1505.07818", "section_title": "Related Work", "citation": "Domain adversarial neural network (DANN) #REFR employs a domain classifier to distinguish whether the input is from the source or target domains.", "text_before_citation": ["On the one hand, domain discrepancy is directly computed by metrics like maximum mean discrepancy (MMD) #OTHEREFR .", "An adaptation layer with MMD in #OTHEREFR is used to learn domaininvariant deep features.", "#OTHEREFR extends this work by proposing the multi-kernel variant of MMD which selects optimal kernels to improve the transferability of features.", "In #OTHEREFR , weighted MMD introduces class-specific weights for source domain to address the issue of class bias across domains.", "On the other hand, some other works focus on learning domain-invariant features adversarially by using a domain classifier #OTHEREFR ."], "text_after_citation": ["The feature extractor is connected to the domain classifier by a gradient reversal layer which reserves the gradients during back-propagation.", "In #OTHEREFR , the training is divided into two stages, in the first stage, a source model is firstly learnt by only using the source data, then a discriminator along with the pre-trained source model is used for adversarial adaptation to learn the target model.", "Weighted-based adversarial methods are also explored in #OTHEREFR .", "Re-weighted adversarial adaptation network (RAAN) #OTHEREFR proposes to re-weight source domain label distribution to match the target label distribution by minimizing Earth Mover distance to adapt classifier with adversarial learning.", "In #OTHEREFR , an extra domain classifier is used to calculate the weights for source samples."], "citing_paper_content": {"title": "Cross-Domain Food Image-To-Recipe Retrieval By Weighted Adversarial Learning", "abstract": "Food image-to-recipe aims to learn an embedded space linking the rich semantics in recipes with the visual content in food image for cross-modal retrieval. The existing research works carry out the learning of such space by assuming that all the image-recipe training example pairs belong to the same cuisine. As a result, despite the excellent performance reported in the literature, such space is not transferable for retrieving recipes of different cuisine. In this paper, we aim to address this issue by cross-domain food imageto-recipe retrieval, such that by leveraging abundant imagerecipe pairs in source domain (one cuisine), the embedding space is generalizable to a target domain (the other cuisine) that does not have images to pair with recipes for training. With the intuition that the importance of different source samples should vary, this paper proposes two novel mechanisms for cross-domain food image-to-recipe retrieval, i.e., source data selector and weighted cross-modal adversarial learning. The former aims to select source samples similar to the target data and filter out distinctive ones for training. The latter is capable to assign higher weights to the source samples more similar to the target data and lower weights to suppress the distinctive ones for both cross-modal and adversarial learning. The weights are computed from the recipe features extracted from a pre-trained source model. Experiments on three different cuisines (Chuan, Yue and Washoku) demonstrate that the proposed method manages to achieve state-of-the-art performances in all the transfers."}, "cited_paper_content": {"title": "Domain-Adversarial Training Of Neural Networks", "abstract": "We introduce a new representation learning approach for domain adaptation, in which data at training and test time come from similar but different distributions. Our approach is directly inspired by the theory on domain adaptation suggesting that, for effective domain transfer to be achieved, predictions must be made based on features that cannot discriminate between the training (source) and test (target) domains. ::: ::: The approach implements this idea in the context of neural network architectures that are trained on labeled data from the source domain and unlabeled data from the target domain (no labeled target-domain data is necessary). As the training progresses, the approach promotes the emergence of features that are (i) discriminative for the main learning task on the source domain and (ii) indiscriminate with respect to the shift between the domains. We show that this adaptation behaviour can be achieved in almost any feed-forward model by augmenting it with few standard layers and a new gradient reversal layer. The resulting augmented architecture can be trained using standard backpropagation and stochastic gradient descent, and can thus be implemented with little effort using any of the deep learning packages. ::: ::: We demonstrate the success of our approach for two distinct classification problems (document sentiment analysis and image classification), where state-of-the-art domain adaptation performance on standard benchmarks is achieved. We also validate the approach for descriptor learning task in the context of person re-identification application."}, "keywords": ["Domain adversarial neural"], "citation_intent": "method"}
{"citing_id": "2303.13811v1", "cited_id": "2003.09667", "section_title": "Introduction", "citation": "We also continue the work on invariant generation for a sequential circuit we started in #REFR .", "text_before_citation": ["The benefits of property generation by PQE are as follows.", "First, by using PQE one can make property generation efficient.", "Second, PQE facilitates generation of properties covering different parts of the design, which increases the probability of discovering a bug.", "Third, every property generated by PQE specifies a large set of high-quality tests.", "We show how PQE can be used to generate properties for a combinational circuit."], "text_after_citation": ["One can use invariant generation to identify a bug that makes a state of a sequential circuit unreachable (whereas in a correct design this state should be reachable). Such bugs can be easily missed.", "The main body of this paper is structured as follows.", "(A full version with all appendices will be published as a technical report.) In Section 2, we give basic definitions. Section 3 presents property generation for a combinational circuit.", "In Section 4, we describe invariant generation for a sequential circuit.", "Sections 5 and 6 present EG-PQE and EG-PQE + respectively."], "citing_paper_content": {"title": "Partial Quantifier Elimination And Property Generation", "abstract": "We study partial quantifier elimination (PQE) for propositional CNF formulas with existential quantifiers. PQE is a generalization of quantifier elimination where one can limit the set of clauses taken out of the scope of quantifiers to a small subset of clauses. The appeal of PQE is that many verification problems (e.g. equivalence checking and model checking) can be solved in terms of PQE and the latter can be dramatically simpler than full quantifier elimination. We show that PQE can be used for property generation that can be viewed as a generalization of testing. The objective here is to produce an unwanted property of a design implementation thus exposing a bug. We introduce two PQE solvers called EG-PQE and EG-PQE +. EG-PQE is a very simple SAT-based algorithm. EG-PQE + is more sophisticated and robust than EG-PQE. We use these PQE solvers to find an unwanted property (namely, an unwanted invariant) of a buggy FIFO buffer. We also apply them to invariant generation for sequential circuits from a HWMCC benchmark set. Finally, we use these solvers to generate properties of a combinational circuit that mimic symbolic simulation. 2 Basic Definitions In this section, when we say \"formula\" without mentioning quantifiers, we mean \"a quantifier-free formula\". Definition 1. We assume that formulas have only Boolean variables. A literal of a variable v is either v or its negation. A clause is a disjunction of literals. A formula F is in conjunctive normal form (CNF) if F = C 1 \u2227 \u2022 \u2022 \u2022 \u2227 C k where C 1 ,. .. , C k are clauses. We will also view F as a set of clauses {C 1 ,. .. , C k }. We assume that every formula is in CNF unless otherwise stated. Definition 2. Let F be a formula. Then Vars(F) denotes the set of variables of F and Vars(\u2203X[F ]) denotes Vars(F)\\X. Definition 3. Let V be a set of variables. An assignment #\u00bb q to V is a mapping V \u2032 \u2192 {0, 1} where V \u2032 \u2286 V. We will denote the set of variables assigned in #\u00bb q as Vars(q). We will refer to #\u00bb q as a full assignment to V if Vars(q) = V. We will denote as #\u00bb q \u2286 #\u00bb r the fact that a) Vars(q) \u2286 Vars(r) and b) every variable of Vars(q) has the same value in #\u00bb q and #\u00bb r. Definition 4. A literal and a clause are said to be satisfied (respectively falsified) by an assignment #\u00bb q if they evaluate to 1 (respectively 0) under #\u00bb q. Definition 5. Let C be a clause. Let H be a formula that may have quantifiers, and #\u00bb q be an assignment to Vars(H). If C is satisfied by #\u00bb q , then C q \u2261 1. Otherwise, C q is the clause obtained from C by removing all literals falsified by #\u00bb q. Denote by H q the formula obtained from H by removing the clauses satisfied by #\u00bb q and replacing every clause C unsatisfied by #\u00bb q with C q. Definition 6. Given a formula \u2203X[F (X, Y)], a clause C of F is called a quantified clause if Vars(C) \u2229 X = \u2205. If Vars(C) \u2229 X = \u2205, the clause C depends only on free i.e. unquantified variables of F and is called a free clause. Definition 7. Let G, H be formulas that may have existential quantifiers. We say that G, H are equivalent, written G \u2261 H, if G q = H q for all full assignments #\u00bb q to Vars(G) \u222a Vars(H)."}, "cited_paper_content": {"title": "Partial Quantifier Elimination By Certificate Clauses", "abstract": "We study a modification of the Quantifier Elimination (QE) problem called Partial QE (PQE) for propositional CNF formulas. In PQE, only a small subset of target clauses is taken out of the scope of quantifiers. The appeal of PQE is that many verification problems, e.g. equivalence checking and model checking, reduce to PQE and, intuitively, the latter should be much easier than QE. One can perform PQE by adding a set of clauses depending only on free variables that make the target clauses redundant. Proving redundancy of a target clause is done by derivation of a\"certificate\"clause implying the former. This idea is implemented in our PQE algorithm called START. It bears some similarity to a SAT-solver with conflict driven learning. A major difference here is that START backtracks when a target clause is proved redundant in the current subspace (a conflict being just one of backtracking conditions). We experimentally evaluate START on a practical problem. We use this problem to show that PQE can be much easier than QE."}, "keywords": ["invariant generation"], "citation_intent": "background"}
{"citing_id": "2304.07018v1", "cited_id": "1710.01878", "section_title": "Iterative Pruning Strategy", "citation": "Inspired by AGP #REFR , L 2 filter pruning is used in our iterative pruning method for model training.", "text_before_citation": ["Finally, we iteratively pruned the reparameterized student network S: where \u03a6 is the pruning operation, r is the pruning rate, \u03d5 means the finetuning operation, S i p means the network after the i-th pruning."], "text_after_citation": ["We stop pruning until the network cannot make effective predictions, and use the network obtained from the last effective pruning as the final network.", "EQUATION"], "citing_paper_content": {"title": "Dipnet: Efficiency Distillation And Iterative Pruning For Image Super-Resolution", "abstract": "Efficient deep learning-based approaches have achieved remarkable performance in single image super-resolution. However, recent studies on efficient super-resolution have mainly focused on reducing the number of parameters and floating-point operations through various network designs. Although these methods can decrease the number of parameters and floating-point operations, they may not necessarily reduce actual running time. To address this issue, we propose a novel multi-stage lightweight network boosting method, which can enable lightweight networks to achieve outstanding performance. Specifically, we leverage enhanced high-resolution output as additional supervision to improve the learning ability of lightweight student networks. Upon convergence of the student network, we further simplify our network structure to a more lightweight level using reparameterization techniques and iterative network pruning. Meanwhile, we adopt an effective lightweight network training strategy that combines multi-anchor distillation and progressive learning, enabling the lightweight network to achieve outstanding performance. Ultimately, our proposed method achieves the fastest inference time among all participants in the NTIRE 2023 efficient super-resolution challenge while maintaining competitive super-resolution performance. Additionally, extensive experiments are conducted to demonstrate the effectiveness of the proposed components. The results show that our approach achieves comparable performance in representative dataset DIV2K, both qualitatively and quantitatively, with faster inference and fewer number of network parameters."}, "cited_paper_content": {"title": "To Prune, Or Not To Prune: Exploring The Efficacy Of Pruning For Model Compression", "abstract": "Model pruning seeks to induce sparsity in a deep neural network's various connection matrices, thereby reducing the number of nonzero-valued parameters in the model. Recent reports (Han et al., 2015; Narang et al., 2017) prune deep networks at the cost of only a marginal loss in accuracy and achieve a sizable reduction in model size. This hints at the possibility that the baseline models in these experiments are perhaps severely over-parameterized at the outset and a viable alternative for model compression might be to simply reduce the number of hidden units while maintaining the model's dense connection structure, exposing a similar trade-off in model size and accuracy. We investigate these two distinct paths for model compression within the context of energy-efficient inference in resource-constrained environments and propose a new gradual pruning technique that is simple and straightforward to apply across a variety of models/datasets with minimal tuning and can be seamlessly incorporated within the training process. We compare the accuracy of large, but pruned models (large-sparse) and their smaller, but dense (small-dense) counterparts with identical memory footprint. Across a broad range of neural network architectures (deep CNNs, stacked LSTM, and seq2seq LSTM models), we find large-sparse models to consistently outperform small-dense models and achieve up to 10x reduction in number of non-zero parameters with minimal loss in accuracy."}, "keywords": ["model training", "pruning"], "citation_intent": "method"}
{"citing_id": "2305.02757v1", "cited_id": "1802.05694", "section_title": "Overall Framework And Pseudocode", "citation": "In the pseudocode, we also take MAN #REFR as the backbone example for a better explanation.", "text_before_citation": ["Our method serves as a plug-and-play solution for MDL and can be applied to different SP models.", "We provide the pseudocode for our MDAL in Algorithm 1."], "text_after_citation": [], "citing_paper_content": {"title": "Multi-Domain Learning From Insufficient Annotations", "abstract": "Multi-domain learning (MDL) refers to simultaneously constructing a model or a set of models on datasets collected from different domains. Conventional approaches emphasize domain-shared information extraction and domain-private information preservation, following the shared-private framework (SP models), which offers significant advantages over single-domain learning. However, the limited availability of annotated data in each domain considerably hinders the effectiveness of conventional supervised MDL approaches in real-world applications. In this paper, we introduce a novel method called multi-domain contrastive learning (MDCL) to alleviate the impact of insufficient annotations by capturing both semantic and structural information from both labeled and unlabeled data. Specifically, MDCL comprises two modules: inter-domain semantic alignment and intra-domain contrast. The former aims to align annotated instances of the same semantic category from distinct domains within a shared hidden space, while the latter focuses on learning a cluster structure of unlabeled instances in a private hidden space for each domain. MDCL is readily compatible with many SP models, requiring no additional model parameters and allowing for end-to-end training. Experimental results across five textual and image multi-domain datasets demonstrate that MDCL brings noticeable improvement over various SP models. Furthermore, MDCL can further be employed in multi-domain active learning (MDAL) to achieve a superior initialization, eventually leading to better overall performance."}, "cited_paper_content": {"title": "Multinomial Adversarial Networks For Multi-Domain Text Classification", "abstract": "Many text classification tasks are known to be highly domain-dependent. Unfortunately, the availability of training data can vary drastically across domains. Worse still, for some domains there may not be any annotated data at all. In this work, we propose a multinomial adversarial network (MAN) to tackle the text classification problem in this real-world multidomain setting (MDTC). We provide theoretical justifications for the MAN framework, proving that different instances of MANs are essentially minimizers of various f-divergence metrics (Ali and Silvey, 1966) among multiple probability distributions. MANs are thus a theoretically sound generalization of traditional adversarial networks that discriminate over two distributions. More specifically, for the MDTC task, MAN learns features that are invariant across multiple domains by resorting to its ability to reduce the divergence among the feature distributions of each domain. We present experimental results showing that MANs significantly outperform the prior art on the MDTC task. We also show that MANs achieve state-of-the-art performance for domains with no labeled data."}, "keywords": ["pseudocode", "backbone example"], "citation_intent": "method"}
{"citing_id": "2303.05973v1", "cited_id": "2004.03315", "section_title": "B. Prioritized Sampling", "citation": "Theorem 1 states that by using prioritized sampling, the loss function is equivalent to an L2 loss for the non-zero loss portion, while the loss function is an L1 loss when using uniform sampling, as in #REFR . Theorem 2.", "text_before_citation": ["Proof.", "EQUATION", "Then, setting", "EQUATION", "completes the proof."], "text_after_citation": ["Define the absolute prediction error (APE) for the i-th data point in the replay buffer under L1 loss with learning rate \u03b7 \u2208 R + at time t as", "EQUATION", "and under L2 loss with learning rate \u03b7 at time t as", "EQUATION", "Let"], "citing_paper_content": {"title": "Data-Efficient Control Barrier Function Refinement", "abstract": "Control barrier functions (CBFs) have been widely used for synthesizing controllers in safety-critical applications. When used as a safety filter, it provides a simple and computationally efficient way to obtain safe controls from a possibly unsafe performance controller. Despite its conceptual simplicity, constructing a valid CBF is well known to be challenging, especially for high-relative degree systems under nonconvex constraints. Recently, work has been done to learn a valid CBF from data based on a handcrafted CBF (HCBF). Even though the HCBF gives a good initialization point, it still requires a large amount of data to train the CBF network. In this work, we propose a new method to learn more efficiently from the collected data through a novel prioritized data sampling strategy. A priority score is computed from the loss value of each data point. Then, a probability distribution based on the priority score of the data points is used to sample data and update the learned CBF. Using our proposed approach, we can learn a valid CBF that recovers a larger portion of the true safe set using a smaller amount of data. The effectiveness of our method is demonstrated in simulation on a unicycle and a two-link arm."}, "cited_paper_content": {"title": "Learning Control Barrier Functions From Expert Demonstrations", "abstract": "Inspired by the success of imitation and inverse reinforcement learning in replicating expert behavior through optimal control, we propose a learning based approach to safe controller synthesis based on control barrier functions (CBFs). We consider the setting of a known nonlinear control affine dynamical system and assume that we have access to safe trajectories generated by an expert - a practical example of such a setting would be a kinematic model of a self-driving vehicle with safe trajectories (e.g. trajectories that avoid collisions with obstacles in the environment) generated by a human driver. We then propose and analyze an optimization-based approach to learning a CBF that enjoys provable safety guarantees under suitable Lipschitz smoothness assumptions on the underlying dynamical system. A strength of our approach is that it is agnostic to the parameterization used to represent the CBF, assuming only that the Lipschitz constant of such functions can be efficiently bounded. Furthermore, if the CBF parameterization is convex, then under mild assumptions, so is our learning process. We end with extensive numerical evaluations of our results on both planar and realistic examples, using both random feature and deep neural network parameterizations of the CBF. To the best of our knowledge, these are the first results that learn provably safe control barrier functions from data."}, "keywords": ["loss function"], "citation_intent": "background"}
{"citing_id": "2305.01139v1", "cited_id": "1911.11253", "section_title": "Algorithm 1 Consistent Prediction-Based Rejection", "citation": "We note that prior works such as #REFR and Tram\u00e8r (2022) have also proposed the idea of using the consistency check for rejecting inputs to a classifier (equivalent to rejecting inputs based on their distance to the decision boundary).", "text_before_citation": ["The details of our CPR defense are given in Algorithm 1.", "Note that the PGD method in the algorithm is deterministic.", "Thus, the mapping T (x) = x m is deterministic and we can summarize the defense as: if y(T (x)) = y(x), then reject x ; otherwise accept x and output the predicted class y(x).", "The robust base model for our defense h can be trained using methods such as standard adversarial training (AT) #OTHEREFR , TRADES (Zhang et al., 2019) , and MART (Wang et al., 2020) (we will focus on the first two methods).", "An advantage of the proposed defense is that it is agnostic of the rejection loss function used for evaluation, and therefore can generalize to multiple choices of the rejection loss function (see Section 5)."], "text_after_citation": ["However, we are the first to systematically implement and evaluate such a defense, and propose strong adaptive attacks for evaluating it."], "citing_paper_content": {"title": "Stratified Adversarial Robustness With Rejection", "abstract": "Recently, there is an emerging interest in adversarially training a classifier with a rejection option (also known as a selective classifier) for boosting adversarial robustness. While rejection can incur a cost in many applications, existing studies typically associate zero cost with rejecting perturbed inputs, which can result in the rejection of numerous slightly-perturbed inputs that could be correctly classified. In this work, we study adversarially-robust classification with rejection in the stratified rejection setting, where the rejection cost is modeled by rejection loss functions monotonically non-increasing in the perturbation magnitude. We theoretically analyze the stratified rejection setting and propose a novel defense method-Adversarial Training with Consistent Prediction-based Rejection (CPR)-for building a robust selective classifier. Experiments on image datasets demonstrate that the proposed method significantly outperforms existing methods under strong adaptive attacks. For instance, on CIFAR-10, CPR reduces the total robust loss (for different rejection losses) by at least 7.3% under both seen and unseen attacks."}, "cited_paper_content": {"title": "Playing It Safe: Adversarial Robustness With An Abstain Option", "abstract": "We explore adversarial robustness in the setting in which it is acceptable for a classifier to abstain---that is, output no class---on adversarial examples. Adversarial examples are small perturbations of normal inputs to a classifier that cause the classifier to give incorrect output; they present security and safety challenges for machine learning systems. In many safety-critical applications, it is less costly for a classifier to abstain on adversarial examples than to give incorrect output for them. We first introduce a novel objective function for adversarial robustness with an abstain option which characterizes an explicit tradeoff between robustness and accuracy. We then present a simple baseline in which an adversarially-trained classifier abstains on all inputs within a certain distance of the decision boundary, which we theoretically and experimentally evaluate. Finally, we propose Combined Abstention Robustness Learning (CARL), a method for jointly learning a classifier and the region of the input space on which it should abstain. We explore different variations of the PGD and DeepFool adversarial attacks on CARL in the abstain setting. Evaluating against these attacks, we demonstrate that training with CARL results in a more accurate, robust, and efficient classifier than the baseline."}, "keywords": ["classifier"], "citation_intent": "background"}
{"citing_id": "2303.01768v1", "cited_id": "1910.07483", "section_title": "Exploration In Rl", "citation": "MAVEN #REFR uses a hierarchical architecture to generate a shared latent vector for each agent to explore the space cooperatively.", "text_before_citation": ["The Distributional Predict Error (DPE) algorithm #OTHEREFR utilizes Random Network Distillation #OTHEREFR to generate two identical architectures with randomly initialized parameters and use their Wasserstein distance to measure an action's novelty given a state.", "Although these methods are effective in the single-agent domain, na\u00efvely applying them separately and independently to each agent in the MARL setting is bound to result in suboptimal performance.", "This is because the agents mutually influence one another, creating additional uncertainty that has to be taken into account.", "One way of accounting for such uncertainty is to require cooperative behavior between agents.", "Recently, there have been some works on ensuring cooperation between the agents."], "text_after_citation": ["CMAE #OTHEREFR creates an exploration compartment for each agent that is not shared with other agents, drastically reducing the searching space via cooperative behavior.", "However, none of these algorithms account for inherent uncertainty and are applied to other MARL algorithms."], "citing_paper_content": {"title": "Toward Risk-Based Optimistic Exploration For Cooperative Multi-Agent Reinforcement Learning", "abstract": "The multi-agent setting is intricate and unpredictable since the behaviors of multiple agents influence one another. To address this environmental uncertainty, distributional reinforcement learning algorithms that incorporate uncertainty via distributional output have been integrated with multi-agent reinforcement learning (MARL) methods, achieving state-of-the-art performance. However, distributional MARL algorithms still rely on the traditional-greedy, which does not take cooperative strategy into account. In this paper, we present a risk-based exploration that leads to collaboratively optimistic behavior by shifting the sampling region of distribution. Initially, we take expectations from the upper quantiles of state-action values for exploration, which are optimistic actions, and gradually shift the sampling region of quantiles to the full distribution for exploitation. By ensuring that each agent is exposed to the same level of risk, we can force them to take cooperatively optimistic actions. Our method shows remarkable performance in multi-agent settings requiring cooperative exploration based on quantile regression appropriately controlling the level of risk."}, "cited_paper_content": {"title": "Maven: Multi-Agent Variational Exploration", "abstract": "Centralised training with decentralised execution is an important setting for cooperative deep multi-agent reinforcement learning due to communication constraints during execution and computational tractability in training. In this paper, we analyse value-based methods that are known to have superior performance in complex environments. We specifically focus on QMIX, the current state-of-the-art in this domain. We show that the representation constraints on the joint action-values introduced by QMIX and similar methods lead to provably poor exploration and suboptimality. Furthermore, we propose a novel approach called MAVEN that hybridises value and policy-based methods by introducing a latent space for hierarchical control. The value-based agents condition their behaviour on the shared latent variable controlled by a hierarchical policy. This allows MAVEN to achieve committed, temporally extended exploration, which is key to solving complex multi-agent tasks. Our experimental results show that MAVEN achieves significant performance improvements on the challenging SMAC domain."}, "keywords": ["shared latent vector", "hierarchical architecture"], "citation_intent": "method"}
{"citing_id": "2303.02045v2", "cited_id": "1802.10501", "section_title": "Related Work", "citation": "For example, KL-PN #REFR proposes the Prior Networks (PN) trained with two KL divergence terms.", "text_before_citation": ["Dirichlet-based uncertainty models (DBU) predict the parameters of the Dirichlet distribution, which allows the computation of closed-form classical uncertainty metrics such as differential entropy, mutual information, etc.", "These metrics can be used to distinguish among data, model, and distributional uncertainty.", "Different DBU models differ in the parameterization and training strategy of the Dirichlet distribution."], "text_after_citation": ["The first term is used to learn sharp Dirichlet parameters for ID data, while the other learns flat Dirichlet parameters for OOD data.", "Since the forward KL divergence is zero-avoiding, RKL-PN #OTHEREFR introduces the reverse KL divergence to avoid undesired multimodal target distributions.", "Posterior Network (PostN) #OTHEREFR uses Normalizing Flows to predict the posterior distribution of any input sample without training with OOD data.", "Evidential Deep Learning (EDL) #OTHEREFR treats the network's outputs as belief masses based on the Dempster-Shafer Theory of Evidence (DST) #OTHEREFR and derives the loss function using subjective logic #OTHEREFR .", "Moreover, deep evidential regression #OTHEREFR introduces evidential priors over the original Gaussian likelihood function to model the uncertainty of regression networks. #OTHEREFR"], "citing_paper_content": {"title": "Uncertainty Estimation By Fisher Information-Based Evidential Deep Learning", "abstract": "Uncertainty estimation is a key factor that makes deep learning reliable in practical applications. Recently proposed evidential neural networks explicitly account for different uncertainties by treating the network's outputs as evidence to parameterize the Dirichlet distribution, and achieve impressive performance in uncertainty estimation. However, for high data uncertainty samples but annotated with the one-hot label, the evidencelearning process for those mislabeled classes is over-penalized and remains hindered. To address this problem, we propose a novel method, Fisher Information-based Evidential Deep Learning (I-EDL). In particular, we introduce Fisher Information Matrix (FIM) to measure the informativeness of evidence carried by each sample, according to which we can dynamically reweight the objective loss terms to make the network more focus on the representation learning of uncertain classes. The generalization ability of our network is further improved by optimizing the PAC-Bayesian bound. As demonstrated empirically, our proposed method consistently outperforms traditional EDL-related algorithms in multiple uncertainty estimation tasks, especially in the more challenging few-shot classification settings."}, "cited_paper_content": {"title": "Predictive Uncertainty Estimation Via Prior Networks", "abstract": "Estimating uncertainty is important to improving the safety of AI systems. Recently baseline tasks and metrics have been defined and several practical methods for estimating uncertainty developed. However, these approaches attempt to model distributional uncertainty either implicitly through model uncertainty or as data uncertainty. This work proposes a new framework for modeling predictive uncertainty called Prior Networks (PNs) which explicitly models distributional uncertainty. PNs do this by parameterizing a prior distribution over predictive distributions. This work focuses on uncertainty for classification and evaluates PNs on the tasks of identifying out-of-distribution (OOD) samples and detecting misclassification on the MNIST dataset, where they are found to outperform previous methods. Experiments on synthetic and MNIST data show that unlike previous methods PNs are able to distinguish between data and distributional uncertainty."}, "keywords": ["Prior Networks"], "citation_intent": "background"}
{"citing_id": "2304.14281v1", "cited_id": "1703.05175", "section_title": "Class Centroids", "citation": "Following #REFR , we define a class centroid c j \u2208 R d in the embedding space for each class j in the support set S.", "text_before_citation": [], "text_after_citation": ["The centroids are learnable variables but initialized by standard class prototypes #OTHEREFR .", "That is, the centroid c j of class j is initialized by the mean", "EQUATION", "of support embeddings assigned to class j.", "Let C = (c 1 \u2022 \u2022 \u2022 c N )"], "citing_paper_content": {"title": "Adaptive Manifold For Imbalanced Transductive Few-Shot Learning", "abstract": "Transductive few-shot learning algorithms have showed substantially superior performance over their inductive counterparts by leveraging the unlabeled queries. However, the vast majority of such methods are evaluated on perfectly class-balanced benchmarks. It has been shown that they undergo remarkable drop in performance under a more realistic, imbalanced setting. To this end, we propose a novel algorithm to address imbalanced transductive few-shot learning, named Adaptive Manifold. Our method exploits the underlying manifold of the labeled support examples and unlabeled queries by using manifold similarity to predict the class probability distribution per query. It is parameterized by one centroid per class as well as a set of graph-specific parameters that determine the manifold. All parameters are optimized through a loss function that can be tuned towards class-balanced or imbalanced distributions. The manifold similarity shows substantial improvement over Euclidean distance, especially in the 1-shot setting. Our algorithm outperforms or is on par with other state of the art methods in three benchmark datasets, namely miniImageNet, tieredImageNet and CUB, and three different backbones, namely ResNet-18, WideResNet-28-10 and DenseNet-121. In certain cases, our algorithm outperforms the previous state of the art by as much as 4.2%."}, "cited_paper_content": {"title": "Prototypical Networks For Few-Shot Learning", "abstract": "A recent approach to few-shot classification called matching networks has demonstrated the benefits of coupling metric learning with a training procedure that mimics test. This approach relies on a complicated fine-tuning procedure and an attention scheme that forms a distribution over all points in the support set, scaling poorly with its size. We propose a more streamlined approach, prototypical networks, that learns a metric space in which few-shot classification can be performed by computing Euclidean distances to prototype representations of each class, rather than individual points. Our method is competitive with state-of-the-art one-shot classification approaches while being much simpler and more scalable with the size of the support set. We empirically demonstrate the performance of our approach on the Omniglot and mini-ImageNet datasets. We further demonstrate that a similar idea can be used for zero-shot learning, where each class is described by a set of attributes, and achieve state-of-the-art results on the Caltech UCSD bird dataset."}, "keywords": ["embedding space"], "citation_intent": "method"}
{"citing_id": "2304.02699v1", "cited_id": "1810.13306", "section_title": "\u2022 Category 3.3: Specification", "citation": "These artifacts can include a human setting initial goals, providing training data, or providing an initial configurations for components #REFR .", "text_before_citation": ["We consider computational notebooks #OTHEREFR to be artifacts possessing such characteristics, as they are intended to be be shared with other technical data workers or simply with the data scientist themselves.", "Artifacts with distribution (c4.4.2) characteristics are those that aim to have a wide audience.", "These can include presentations or reports, many of which rely on the compilation and analysis of other AutoML artifacts.", "\u2022 Category 4.5: Steering Artifacts can act to orient or intervene within an AutoML process.", "Artifacts with directing (4.5.1) act to initialize or orient the AutoML processes in a specific trajectory."], "text_after_citation": ["They can also include algorithmic processes that define the final configuration of AutoML components.", "In contrast, artifacts with refining (c4.5.2) characteristics are subtler than directing as they largely build off of the existing model structure.", "Fine-tuning operations, example Transfer Learning #OTHEREFR 117] , produce artifacts for refining an existing neural architecture #OTHEREFR , rather than conducting an expensive neural architecture search #OTHEREFR .", "We consider the former to be a refinement, while the latter uses an initial design space configuration to direct the overall neural architecture search process.", "https://osf.io/3nmyj/?view_only=19962103d58b45d289b5c83421f48b36."], "citing_paper_content": {"title": "Tracing And Visualizing Human-Ml/Ai Collaborative Processes Through Artifacts Of Data Work", "abstract": "Artifact Identification and Classification Capture Artifacts Artifact Properties Dependencies Artifact Lineage Model Customization Interaction Version 3 Source files changed Version 2 Source files unchanged Version 1 of artifact Figure 1: We developed an artifact taxonomy that captures both human and ML/AI processes in automated data work. We assess our taxonomy's utility through a collaboration with an enterprise software development team creating an AutoML system. ."}, "cited_paper_content": {"title": "Taking Human Out Of Learning Applications: A Survey On Automated Machine Learning", "abstract": "Machine learning techniques have deeply rooted in our everyday life. However, since it is knowledge- and labor-intensive to pursue good learning performance, human experts are heavily involved in every aspect of machine learning. In order to make machine learning techniques easier to apply and reduce the demand for experienced human experts, automated machine learning (AutoML) has emerged as a hot topic with both industrial and academic interest. In this paper, we provide an up to date survey on AutoML. First, we introduce and define the AutoML problem, with inspiration from both realms of automation and machine learning. Then, we propose a general AutoML framework that not only covers most existing approaches to date but also can guide the design for new methods. Subsequently, we categorize and review the existing works from two aspects, i.e., the problem setup and the employed techniques. Finally, we provide a detailed analysis of AutoML approaches and explain the reasons underneath their successful applications. We hope this survey can serve as not only an insightful guideline for AutoML beginners but also an inspiration for future research."}, "keywords": ["artifacts", "training data"], "citation_intent": "background"}
{"citing_id": "2303.10318v1", "cited_id": "1802.10062", "section_title": "A. Network Architecture", "citation": "To improve the characterization capabilities of the teacher branch further, we adopt a structure similar to CSRNet #REFR by incorporating dilated convolutional layers in the back-end.", "text_before_citation": ["It is important to note that the shared module also serves as the initial layers of the student branch, so it should not have too many parameters.", "In other words, we should avoid introducing too many shallow layers in the teacher branch when designing the shared module.", "2) Teacher branch: In addition to using the first two layers of VGG-16 in the shared shallow module, we employ the remaining layers of the first 10 layers of VGG-16 as the frontend of the teacher branch.", "This allows us to utilize the pretrained model, which has strong transfer learning capabilities and a flexible design.", "The teacher branch is intended to act as an experienced mentor who can quickly learn new information and guide the student branch."], "text_after_citation": ["This type of convolutional layers can directly enhance the network receptive field without increasing its parameters or computational complexity.", "Specifically, we use a light version of CSRNet consisting of 3 dilated convolutional layers.", "3) Student branch: The channel capacity of a network is a critical factor that affects both inference speed and memory consumption.", "In our approach, we limit the number of channels in the teacher branch to approximately one-fourth of its original size to reduce the computational burden.", "However, following the design principle of VGG networks, reducing the number of channels directly to one-fourth may limit the network's ability to express shallow features."], "citing_paper_content": {"title": "Crowd Counting With Online Knowledge Learning", "abstract": "Efficient crowd counting models are urgently required for the applications in scenarios with limited computing resources, such as edge computing and mobile devices. A straightforward method to achieve this is knowledge distillation (KD), which involves using a trained teacher network to guide the training of a student network. However, this traditional twophase training method can be time-consuming, particularly for large datasets, and it is also challenging for the student network to mimic the learning process of the teacher network. To overcome these challenges, we propose an online knowledge learning method for crowd counting. Our method builds an end-to-end training framework that integrates two independent networks into a single architecture, which consists of a shared shallow module, a teacher branch, and a student branch. This approach is more efficient than the two-stage training technique of traditional KD. Moreover, we propose a feature relation distillation method which allows the student branch to more effectively comprehend the evolution of inter-layer features by constructing a new interlayer relationship matrix. It is combined with response distillation and feature internal distillation to enhance the transfer of mutually complementary information from the teacher branch to the student branch. Extensive experiments on four challenging crowd counting datasets demonstrate the effectiveness of our method which achieves comparable performance to state-of-theart methods despite using far fewer parameters."}, "cited_paper_content": {"title": "Csrnet: Dilated Convolutional Neural Networks For Understanding The Highly Congested Scenes", "abstract": "We propose a network for Congested Scene Recognition called CSRNet to provide a data-driven and deep learning method that can understand highly congested scenes and perform accurate count estimation as well as present high-quality density maps. The proposed CSRNet is composed of two major components: a convolutional neural network (CNN) as the front-end for 2D feature extraction and a dilated CNN for the back-end, which uses dilated kernels to deliver larger reception fields and to replace pooling operations. CSRNet is an easy-trained model because of its pure convolutional structure. We demonstrate CSRNet on four datasets (ShanghaiTech dataset, the UCF_CC_50 dataset, the WorldEXPO'10 dataset, and the UCSD dataset) and we deliver the state-of-the-art performance. In the ShanghaiTech Part_B dataset, CSRNet achieves 47.3% lower Mean Absolute Error (MAE) than the previous state-of-the-art method. We extend the targeted applications for counting other objects, such as the vehicle in TRANCOS dataset. Results show that CSRNet significantly improves the output quality with 15.4% lower MAE than the previous state-of-the-art approach."}, "keywords": ["dilated convolutional layers"], "citation_intent": "method"}
{"citing_id": "2303.04289v1", "cited_id": "1803.09047", "section_title": "Related Work", "citation": "In #REFR the authors find that this approach is capable of capturing acoustic correlates of prosody from the reference, even if spoken by an unseen speaker or containing verbal content not related to the target-text.", "text_before_citation": ["Most prosody transfer approaches follow #OTHEREFR and normally comprise a core acoustic model, most commonly Tacotron #OTHEREFR , and a prosody encoder.", "The acoustic model and the prosody encoder are jointly trained.", "The prosody encoder produces a fixed-size latent prosody embedding from the reference mel-spectrogram which is combined with the intermediate acoustic model representation then passed to the decoder which generates the output mel-spectrogram.", "In #OTHEREFR , the joint model is trained to minimize target mel-spectrogram loss, which can be viewed as a reconstruction loss."], "text_after_citation": ["The information capacity of the latent prosody representation has an important effect in the downstream TTS task.", "Representations that are lower-capacity than that in #OTHEREFR , such as #OTHEREFR , for example], are effectively bottle-necked by their capacity and cannot model the same level of detail as #OTHEREFR .", "These models are used for tasks such as speaking-style or emotion modelling.", "Representations of higher capacity than #OTHEREFR , such as #OTHEREFR , capture a considerably higher-detailed representation of the reference and are highly reference-specific.", "Therefore, high-capacity representations are used for tasks such as voice-puppetry where we expect the lexical content of the reference to be similar to the target text."], "citing_paper_content": {"title": "Do Prosody Transfer Models Transfer Prosody?", "abstract": "Some recent models for Text-to-Speech synthesis aim to transfer the prosody of a reference utterance to the generated target synthetic speech. This is done by using a learned embedding of the reference utterance, which is used to condition speech generation. During training, the reference utterance is identical to the target utterance. Yet, during synthesis, these models are often used to transfer prosody from a reference that differs from the text or speaker being synthesized. To address this inconsistency, we propose to use a different, but prosodically-related, utterance during training too. We believe this should encourage the model to learn to transfer only those characteristics that the reference and target have in common. If prosody transfer methods do indeed transfer prosody they should be able to be trained in the way we propose. However, results show that a model trained under these conditions performs significantly worse than one trained using the target utterance as a reference. To explain this, we hypothesize that prosody transfer models do not learn a transferable representation of prosody, but rather a utterance-level representation which is highly dependent on both the reference speaker and reference text."}, "cited_paper_content": {"title": "Towards End-To-End Prosody Transfer For Expressive Speech Synthesis With Tacotron", "abstract": "We present an extension to the Tacotron speech synthesis architecture that learns a latent embedding space of prosody, derived from a reference acoustic representation containing the desired prosody. We show that conditioning Tacotron on this learned embedding space results in synthesized audio that matches the prosody of the reference signal with fine time detail even when the reference and synthesis speakers are different. Additionally, we show that a reference prosody embedding can be used to synthesize text that is different from that of the reference utterance. We define several quantitative and subjective metrics for evaluating prosody transfer, and report results with accompanying audio samples from single-speaker and 44-speaker Tacotron models on a prosody transfer task."}, "keywords": ["prosody"], "citation_intent": "background"}
{"citing_id": "2305.01628v1", "cited_id": "2002.12327", "section_title": "Related Work", "citation": "There have been a number of studies on analyzing the characteristics of different layers of transformer models. #REFR ; Van Aken et al.", "text_before_citation": [], "text_after_citation": ["(2019) used probing to report that in BERT models the lower layers carry the most information about linear word order, syntactic information is most prominent in the middle layers, and the final layers of BERT are the most task-specific. Van Aken et al.", "(2019) also show that similar behavior is observed in other transformer models such as GPT2. #OTHEREFR", "( , 2022 studied the role of feed-forward layers in transformer models.", "They demonstrate that representations across different layers capture meaningful semantic and syntactic patterns, and describe how model predictions are gradually refined as they progress across the different layers.", "Aiming to reduce the computational load of transformers, multiple works have explored earlyexiting, i.e., performing some calculations without passing through all of the model layers."], "citing_paper_content": {"title": "The Benefits Of Bad Advice: Autocontrastive Decoding Across Model Layers", "abstract": "Applying language models to natural language processing tasks typically relies on the representations in the final model layer, as intermediate hidden layer representations are presumed to be less informative. In this work, we argue that due to the gradual improvement across model layers, additional information can be gleaned from the contrast between higher and lower layers during inference. Specifically, in choosing between the probable next token predictions of a generative model, the predictions of lower layers can be used to highlight which candidates are best avoided. We propose a novel approach that utilizes the contrast between layers to improve text generation outputs, and show that it mitigates degenerative behaviors of the model in open-ended generation, significantly improving the quality of generated texts. Furthermore, our results indicate that contrasting between model layers at inference time can yield substantial benefits to certain aspects of general language model capabilities, more effectively extracting knowledge during inference from a given set of model parameters."}, "cited_paper_content": {"title": "A Primer In Bertology: What We Know About How Bert Works", "abstract": "Transformer-based models are now widely used in NLP, but we still do not understand a lot about their inner workings. This paper describes what is known to date about the famous BERT model (Devlin et al. 2019), synthesizing over 40 analysis studies. We also provide an overview of the proposed modifications to the model and its training regime. We then outline the directions for further research."}, "keywords": ["different layers", "transformer models"], "citation_intent": "background"}
{"citing_id": "2304.07574v1", "cited_id": "1805.01677", "section_title": "Results And Analysis", "citation": "Similar observation can be made on TGAN #REFR , i.e., simple fine-tuning based method without explicit knowledge preservation. In contrast, our method (we discuss it in Sec. 5) can address this issue.", "text_before_citation": ["More importantly, the root cause of such incompatible knowledge transfer is the least important filters in G s determined to be irrelevant to the target domain adaptation, and fine-tuning is not adequate for removing the incompatible knowledge after adaptation.", "Specifically, we summarize our observations in Figure 1 and Figure 2 :", "Observation 1: In Figure 1 (c), we visualize the generated images by different methods with fixed noise input.", "Interestingly, features that are incompatible for the target domain are indeed transferred after adaptation with different knowledge preservation criteria, e.g., \"tree on sea\" where \"tree\" is from Church domain, and \"Cat with glasses\" where \"glasses\" is from FFHQ domain.", "All these incompatible source features severely curtail the realism of generated target images."], "text_after_citation": ["Observation 2: In Figure 2 , we dissect and visualize the incompatible features observed in Figure 1 , and find their mostly correlated filters in G s and G t .", "Surprisingly, we find that the filters in G s identified with least importance for target domain are mostly relevant to incompatible features transferred from the source, which is the root cause of degradation of realism of generated images.", "After adaptation, the same filters will still cause the same type of incompatible features, and fine-tuning for knowledge update cannot effectively address this issue.", "When the target domains become distant, this observation is more obvious.", "Our analysis uncovers an important issue of existing SOTA FSIG methods: source domain features that are incompatible to the target domain are transferred to the target generator, with various knowledge preservation criteria."], "citing_paper_content": {"title": "Exploring Incompatible Knowledge Transfer In Few-Shot Image Generation", "abstract": "ProgressiveGAN StyleGAN-V2 Church \u21a6 Sailboat FFHQ \u21a6 AFHQ-Cat Church \u21a6 Sailboat FFHQ \u21a6 AFHQ-Cat Pretraining on large source dataset (e.g., LSUN-Church, FFHQ\u2026) Adaptation on few-shot target samples"}, "cited_paper_content": {"title": "Transferring Gans: Generating Images From Limited Data", "abstract": "Transferring knowledge of pre-trained networks to new domains by means of fine-tuning is a widely used practice for applications based on discriminative models. To the best of our knowledge this practice has not been studied within the context of generative deep networks. Therefore, we study domain adaptation applied to image generation with generative adversarial networks. We evaluate several aspects of domain adaptation, including the impact of target domain size, the relative distance between source and target domain, and the initialization of conditional GANs. Our results show that using knowledge from pre-trained networks can shorten the convergence time and can significantly improve the quality of the generated images, especially when target data is limited. We show that these conclusions can also be drawn for conditional GANs even when the pre-trained model was trained without conditioning. Our results also suggest that density is more important than diversity and a dataset with one or few densely sampled classes is a better source model than more diverse datasets such as ImageNet or Places."}, "keywords": ["TGAN"], "citation_intent": "result"}
{"citing_id": "2305.01481v1", "cited_id": "1912.11370", "section_title": "Performance Evaluation", "citation": "Echoing the previous findings in transfer learning #REFR , we also observe a performance boost of models of larger size pretrained on larger datasets.", "text_before_citation": ["As shown in Table 1 , our method can outperform the baseline methods over different datasets, including large-scale dataset, ImageNet, for both CNN and ViT classifiers.", "Our empirical result confirms the previous findings that ViT can generally perform better than CNN classifiers in accuracy on uncertainty estimation-related tasks #OTHEREFR .", "Effect of Different Foundation Models Given the diversity of foundation models, which can vary in terms of architecture, training data, and optimization losses, it is important to further investigate the impact of different pretrained models on our proposed method.", "We demonstrate the average performance over base models and datasets (excluding ImageNet 1 ) for different pretrained models in Figure 2 , which shows that every pretrained model in our model candidate pool can surpass MSP on average.", "We compare the performance of models pretrained on datasets of increasing size: ImageNet-1K, ImageNet-21K, and CLIP WebImageText dataset (400 million image-text pairs)."], "text_after_citation": ["With the similar model size and pretrained on the same dataset, ViTs generally perform better than CNNs, except for the cases where the dataset is in a smaller scale, e.g., ImageNet-1K. The self-supervised pretrained models, e.g.", "MoCov3 ViT/ResNet, achieve comparable results with models pretrained with supervision signals.", "Similar to the previous findings in #OTHEREFR , which suggests that better performing models transfer better to the downstream tasks, we find the failure detection performance of pretrained models related to these models' accuracy performance in ImagetNet-1K, as shown in Figure  8 .", "Thus, we suggest to select the pretrained model with the highest accuracy on ImageNet-1K for failure detection, without any prior knowledge about the trained models, e.g. the training datasets."], "citing_paper_content": {"title": "Great Models Think Alike: Improving Model Reliability Via Inter-Model Latent Agreement", "abstract": "Reliable application of machine learning is of primary importance to the practical deployment of deep learning methods. A fundamental challenge is that models are often unreliable due to overconfidence (Hendrycks & Gimpel, 2017). In this paper, we estimate a model's reliability by measuring the agreement between its latent space, and the latent space of a foundation model. However, it is challenging to measure the agreement between two different latent spaces due to their incoherence, e.g., arbitrary rotations and different dimensionality. To overcome this incoherence issue, we design a neighborhood agreement measure between latent spaces and find that this agreement is surprisingly well-correlated with the reliability of a model's predictions. Further, we show that fusing neighborhood agreement into a model's predictive confidence in a post-hoc way significantly improves its reliability. Theoretical analysis and extensive experiments on failure detection across various datasets verify the effectiveness of our method on both in-distribution and out-of-distribution settings."}, "cited_paper_content": {"title": "Big Transfer (Bit): General Visual Representation Learning", "abstract": "Transfer of pre-trained representations improves sample efficiency and simplifies hyperparameter tuning when training deep neural networks for vision. We revisit the paradigm of pre-training on large supervised datasets and fine-tuning the model on a target task. We scale up pre-training, and propose a simple recipe that we call Big Transfer (BiT). By combining a few carefully selected components, and transferring using a simple heuristic, we achieve strong performance on over 20 datasets. BiT performs well across a surprisingly wide range of data regimes -- from 1 example per class to 1M total examples. BiT achieves 87.5% top-1 accuracy on ILSVRC-2012, 99.4% on CIFAR-10, and 76.3% on the 19 task Visual Task Adaptation Benchmark (VTAB). On small datasets, BiT attains 76.8% on ILSVRC-2012 with 10 examples per class, and 97.0% on CIFAR-10 with 10 examples per class. We conduct detailed analysis of the main components that lead to high transfer performance."}, "keywords": ["larger datasets"], "citation_intent": "result"}
{"citing_id": "2303.17987v1", "cited_id": "1902.04885", "section_title": "Big Data Of Fl4M", "citation": "In the midst of increased privacy protection and data oversight, the phenomenon of data existing as isolated islands has emerged #REFR .", "text_before_citation": ["Therefore, the acquisition, storage, analysis, and protection of big data is a crucial and challenging task.", "In the metaverse, the constructed digital space is essentially data.", "In other words, the metaverse can be regarded as a fusion mechanism for big data and various information technologies.", "When building a bridge between virtual and reality, real-time data gathering and processing capabilities of sensors and smart devices are the cornerstones for supporting large numbers of users online at once.", "\u2022 Challenges."], "text_after_citation": ["The rich data of multiple business departments in an enterprise is defined by each other independently.", "Data from a single department is similar to an isolated island that is incapable of interacting with other rich data #OTHEREFR .", "In the era of the metaverse, data is undoubtedly one of the most important assets of an enterprise.", "Therefore, it is vital to shatter the isolated island of data in order to actualize information exchange and the hidden value underlying big data #OTHEREFR .", "\u2022 Methods."], "citing_paper_content": {"title": "Federated Learning For Metaverse: A Survey", "abstract": "The metaverse, which is at the stage of innovation and exploration, faces the dilemma of data collection and the problem of private data leakage in the process of development. This can seriously hinder the widespread deployment of the metaverse. Fortunately, federated learning (FL) is a solution to the above problems. FL is a distributed machine learning paradigm with privacy-preserving features designed for a large number of edge devices. Federated learning for metaverse (FL4M) will be a powerful tool. Because FL allows edge devices to participate in training tasks locally using their own data, computational power, and model-building capabilities. Applying FL to the metaverse not only protects the data privacy of participants but also reduces the need for high computing power and high memory on servers. Until now, there have been many studies about FL and the metaverse, respectively. In this paper, we review some of the early advances of FL4M, which will be a research direction with unlimited development potential. We first introduce the concepts of metaverse and FL, respectively. Besides, we discuss the convergence of key metaverse technologies and FL in detail, such as big data, communication technology, the Internet of Things, edge computing, blockchain, and extended reality. Finally, we discuss some key challenges and promising directions of FL4M in detail. In summary, we hope that our up-to-date brief survey can help people better understand FL4M and build a fair, open, and secure metaverse. CCS CONCEPTS \u2022 Computing methodologies \u2192 Federated learning."}, "cited_paper_content": {"title": "Federated Machine Learning: Concept And Applications", "abstract": "Today's AI still faces two major challenges. One is that in most industries, data exists in the form of isolated islands. The other is the strengthening of data privacy and security. We propose a possible solution to these challenges: secure federated learning. Beyond the federated learning framework first proposed by Google in 2016, we introduce a comprehensive secure federated learning framework, which includes horizontal federated learning, vertical federated learning and federated transfer learning. We provide definitions, architectures and applications for the federated learning framework, and provide a comprehensive survey of existing works on this subject. In addition, we propose building data networks among organizations based on federated mechanisms as an effective solution to allow knowledge to be shared without compromising user privacy."}, "keywords": ["increased privacy protection"], "citation_intent": "background"}
{"citing_id": "2303.06636v1", "cited_id": "1205.0858", "section_title": "A T E X I T S H A 1 _ B A S E 6 4 = \" K H E L A V A O A S U R C + R 0 Z R A V F N P T B G 4 = \" > A", "citation": "The problem is known to be hard as it relates to the challenging close-loop controlled sensing problem #REFR .", "text_before_citation": ["v g v T e 8 7 F J v T T I A m O f Z E = \" > A A A C F X i c b V D L S s N A F J 3 U V 4 2 v q k s 3 g 0 V w I S W p g o K b o p s u K 9 o H N L F M p p N m 6 G Q S Z i Z C C f k J N / 6 K G x e K u B X c + T d O 2 i y 0 9 c L A m X P P f R 0 v Z l Q q y / o 2 S k v L K 6 t r 5 X V z Y 3", "N r e 6 e y u 9 e R U S I w a e O I R a L n I U k Y 5 a S t q G K k F w u C Q o + R r j e + z possible values depending on a underlying hypothesis, and the radar receiver aims to guess this hypothesis.", "Sensing performance is measured in terms of exponential decay-rate of the probability of error, either the minimum exponential decay-rate over all hypotheses #OTHEREFR , #OTHEREFR or the set of decayrates that are simultaneously achievable under the different hypotheses #OTHEREFR .", "The work #OTHEREFR also studies a mono-static version of this problem, assuming that Tx coincides with the radar receiver and thus can use the generalized-feedback signals also for communication purposes.", "For this mono-static radar scenario however only a coding scheme but no converse is presented."], "text_after_citation": ["v P d B y I k j f i d m s T E D d G I U 5 9 i p D Q 1 q J w 4 A V K p E y I V Y M T S Z p Y 5 l 6 b j I T y W D M k A w v y b S 2 6 z e y 2 v W j V r G n A R 2 A W o g i J a g 8 q X M 4 x w E h K u s G 4 n + 7 Y V K z d F Q l H M S G Y 6 i S S x n o V G p K 8 h R y G R b j q 9 K o N H m h l C P x L 6 c Q W n 7 O + K F I V S T k J P K / P 1 5 X w u J / / L 9 R P l X 7 g p 5 X G i C M e z Q X 7 C o I p g b h E c U k G w Y h M N E B Z U 7 w p x g A T C S h", "The first information-theoretic work #OTHEREFR on ISAC determined the fundamental limits of the rate-distortion version of the ISAC problem in Figure 1 .", "assuming that the Tx can use the feedback signals also for coding (i.e., under the closeloop coding assumption).", "Extensions to network scenarios and to scenarios with secrecy constraints were subsequently presented in #OTHEREFR - #OTHEREFR .", "In this paper, we consider both the rate-distortion version and the hypothesis testing versions of the model in Figure 1 ."], "citing_paper_content": {"title": "Strong Converses For Memoryless Bi-Static Isac", "abstract": "The paper characterizes the fundamental limits of integrated sensing and communication (ISAC) systems with a bi-static radar, where the radar receiver is located close to the transmitter and estimates or detects the state based on the transmitter's channel inputs and the backscattered signals. Two models are considered. In the first model, the memoryless state sequence is distributed according to a fixed distribution and the goal of the radar receiver is to reconstruct this statesequence with smallest possible distortion. In the second model, the memoryless state is distributed either according to PS or to QS and the radar's goal is to detect this underlying distribution so that the missed-detection error probability has maximum exponential decay-rate (maximum Stein exponent). Similarly to previous results, our fundamental limits show that the tradeoff between sensing and communication solely stems from the empirical statistics of the transmitted codewords which influences both performances. The main technical contribution are two strong converse proofs that hold for all probabilities of communication error and excess-distortion probability or false-alarm probability \u03b4 summing to less than 1, +\u03b4 < 1. These proofs are based on two parallel change-of-measure arguments on the sets of typical sequences, one change-of-measure to obtain the desired bound on the communication rate, and the second to bound the sensing performance. Index Terms-Integrated sensing and communication, strong converse, change of measure arguments."}, "cited_paper_content": {"title": "Controlled Sensing For Multihypothesis Testing", "abstract": "The problem of multiple hypothesis testing with observation control is considered in both fixed sample size and sequential settings. In the fixed sample size setting, for binary hypothesis testing, the optimal exponent for the maximal error probability corresponds to the maximum Chernoff information over the choice of controls, and a pure stationary open-loop control policy is asymptotically optimal within the larger class of all causal control policies. For multihypothesis testing in the fixed sample size setting, lower and upper bounds on the optimal error exponent are derived. It is also shown through an example with three hypotheses that the optimal causal control policy can be strictly better than the optimal open-loop control policy. In the sequential setting, a test based on earlier work by Chernoff for binary hypothesis testing, is shown to be first-order asymptotically optimal for multihypothesis testing in a strong sense, using the notion of decision making risk in place of the overall probability of error. Another test is also designed to meet hard risk constrains while retaining asymptotic optimality. The role of past information and randomization in designing optimal control policies is discussed."}, "keywords": ["sensing problem"], "citation_intent": "background"}
{"citing_id": "2304.08960v1", "cited_id": "1901.05103", "section_title": "Introduction", "citation": "Following the DeepSDF model proposed by #REFR , we represent this SDF function in an implicit neural representation (INR) .", "text_before_citation": ["GANs excel at this second stage owed in part to image-to-image translation methods such as pix2pix #OTHEREFR , which have found widespread application in microscopy image synthesis.", "However, the first stage requires fundamentally different models and, most importantly, shape representations that can deal with the variation in cell shapes. As exemplified in Fig.", "1 , cells exhibit a range of visually observable phenomena including global cell body changes such as cell growth, cell division, deformation in cell tissue, or projections of leading edge during cell motility, as well as localized body changes, such as blebbing (i.e., randomly growing and retracting blobs on the cell surface) or growing filopodia (i.e., highly motile thin protrusions budding off the main cell body).", "In this work, we synthesize living cell shapes in 3D+time.", "We propose to model the cell surface as the zero level-set of a continuous signed distance function (SDF) in 3D space and time."], "text_after_citation": ["Such an INR consists of a multilayer perceptron (MLP) that we jointly optimize with a latent space using a large set of 3D cell shape time-lapse sequences.", "Once trained, owing to the continuous implicit representation, the optimized model can be used to synthesize completely new cell shape sequences at any given spatial or temporal resolution.", "In contrast to voxel-based or mesh-based cell shape representations with discretized sampling rates, our approach alleviates the limitations on spatial and temporal resolution and allows us to model and represent cell shapes at an arbitrary level of detail.", "By parametrizing the cell surface implicitly through differentiable and, therefore, trainable neural networks, the complexity of the resulting model is independent of the spatial and temporal resolution and instead scales with the complexity of the cell surface.", "Hence, in contrast to voxel-based or mesh-based representations, the resulting network has fixed memory requirements."], "citing_paper_content": {"title": "Medical Image Analysis", "abstract": "Data-driven cell tracking and segmentation methods in biomedical imaging require diverse and information-rich training data. In cases where the number of training samples is limited, synthetic computer-generated data sets can be used to improve these methods. This requires the synthesis of cell shapes as well as corresponding microscopy images using generative models. To synthesize realistic living cell shapes, the shape representation used by the generative model should be able to accurately represent fine details and changes in topology, which are common in cells. These requirements are not met by 3D voxel masks, which are restricted in resolution, and polygon meshes, which do not easily model processes like cell growth and mitosis. In this work, we propose to represent living cell shapes as level sets of signed distance functions (SDFs) which are estimated by neural networks. We optimize a fully-connected neural network to provide an implicit representation of the SDF value at any point in a 3D+time domain, conditioned on a learned latent code that is disentangled from the rotation of the cell shape. We demonstrate the effectiveness of this approach on cells that exhibit rapid deformations (Platynereis dumerilii), cells that grow and divide (C. elegans), and cells that have growing and branching filopodial protrusions (A549 human lung carcinoma cells). A quantitative evaluation using shape features, Hausdorff distance, and Dice similarity coefficients of real and synthetic cell shapes shows that our model can generate topologically plausible complex cell shapes in 3D+time with high similarity to real living cell shapes. Finally, we show how microscopy images of living cells that correspond to our generated cell shapes can be synthesized using an image-to-image model."}, "cited_paper_content": {"title": "Deepsdf: Learning Continuous Signed Distance Functions For Shape Representation", "abstract": "Computer graphics, 3D computer vision and robotics communities have produced multiple approaches to representing 3D geometry for rendering and reconstruction. These provide trade-offs across fidelity, efficiency and compression capabilities. In this work, we introduce DeepSDF, a learned continuous Signed Distance Function (SDF) representation of a class of shapes that enables high quality shape representation, interpolation and completion from partial and noisy 3D input data. DeepSDF, like its classical counterpart, represents a shape's surface by a continuous volumetric field: the magnitude of a point in the field represents the distance to the surface boundary and the sign indicates whether the region is inside (-) or outside (+) of the shape, hence our representation implicitly encodes a shape's boundary as the zero-level-set of the learned function while explicitly representing the classification of space as being part of the shapes interior or not. While classical SDF's both in analytical or discretized voxel form typically represent the surface of a single shape, DeepSDF can represent an entire class of shapes. Furthermore, we show state-of-the-art performance for learned 3D shape representation and completion while reducing the model size by an order of magnitude compared with previous work."}, "keywords": ["implicit neural representation"], "citation_intent": "method"}
{"citing_id": "2304.09616v2", "cited_id": "1804.08316", "section_title": "Evaluation Of The Quality Of The Embeddings In Similarity Task", "citation": "In the case of Basque, we used the pure similarity RG dataset and the relatedness dataset WS created by #REFR .", "text_before_citation": ["In this work, we have chosen three types of embeddings, which perform differently in pure semantic similarity #OTHEREFR and relatedness #OTHEREFR measures.", "Broadly, wordnet embeddings are more precise for measuring pure similarity relations, text embeddings are more sensitive to compute relatedness, and hybrid embeddings have been proven to be more robust for capturing semantic relations in general #OTHEREFR .", "Hence in this work, we chose the available similarity and relatedness datasets for evaluating all embeddings in both languages.", "In Spanish, we operated with the pure similarity datasets RG65 (RG) #OTHEREFR and", "SimLex999 (SL) #OTHEREFR , and the relatedness dataset Wordsim353 (WS) #OTHEREFR ."], "text_after_citation": ["In both Spanish and Basque, F T txt , F T kb and F T hyb representations have been compared with the baseline, using publicly available text-based represen- \u2022 F T txt embeddings: F T txt results perform similarly to the baseline ones in both languages, with two exceptions.", "One, the Basque F T txt result in the WS dataset is higher than the baseline, likely due to the greater corpus size Euscrawl.", "Second, the Spanish F T txt in the SL dataset is lower than the baseline.", "The only plausible explanation for this result in the Spanish SL may lie in the differences in the pre-processing of the corpus; the baseline text corpus was tokenized with Europarl pre-processing tools 14 , whereas the NLTK tokenizer has been used in the present work.", "\u2022 F T kb embeddings: being this type of embeddings more suited for pure similarity datasets (RG and SL) rather than for relatedness ones (WS), F T kb results are higher than the baseline in the Basque RG and the Spanish SL, but not in the Spanish RG."], "citing_paper_content": {"title": "Bridging Natural Language Processing And Psycholinguistics: Computationally Grounded Semantic Similarity Datasets For Basque And Spanish", "abstract": "We present a computationally-grounded word similarity dataset based on two well-known Natural Language Processing resources; text corpora and knowledge bases. This dataset aims to fulfil a gap in psycholinguistic research by providing a variety of quantifications of semantic similarity in an extensive set of noun pairs controlled by variables that play a significant role in lexical processing. The dataset creation has consisted in three steps, 1) computing four key psycholinguistic features for each noun; concreteness, frequency, semantic and phonological neighbourhood density; 2) pairing nouns across these four variables; 3) for each noun pair, assigning three types of word similarity measurements, computed out of text, Wordnet and hybrid embeddings. The present dataset includes noun pairs' information in Basque and European Spanish, but further work intends to extend it to more languages."}, "cited_paper_content": {"title": "Bilingual Embeddings With Random Walks Over Multilingual Wordnets", "abstract": "Bilingual word embeddings represent words of two languages in the same space, and allow to transfer knowledge from one language to the other without machine translation. The main approach is to train monolingual embeddings first and then map them using bilingual dictionaries. In this work, we present a novel method to learn bilingual embeddings based on multilingual knowledge bases (KB) such as WordNet. Our method extracts bilingual information from multilingual wordnets via random walks and learns a joint embedding space in one go. We further reinforce cross-lingual equivalence adding bilingual constraints in the loss function of the popular Skip-gram model. Our experiments on twelve cross-lingual word similarity and relatedness datasets in six language pairs covering four languages show that: 1) our method outperforms the state-of-the-art mapping method using dictionaries; 2) multilingual wordnets on their own improve over text-based systems in similarity datasets; 3) the combination of wordnet-generated information and text is key for good results. Our method can be applied to richer KBs like DBpedia or BabelNet, and can be easily extended to multilingual embeddings. All our software and resources are open source."}, "keywords": ["relatedness dataset WS"], "citation_intent": "method"}
{"citing_id": "2305.01214v1", "cited_id": "1511.04629", "section_title": "B. Error-Rate Performance", "citation": "As a baseline, we also plot the O(n \u22122 ) approximation of the finite blocklength Polyanskyi-Poor-Verd\u00fa (PPV) meta-converse bound #REFR .", "text_before_citation": ["We evaluate the performance of rate-compatible polar codes with 64 \u2264 N \u2264 1024 for the additive white Gaussian noise (AWGN) channel with binary phase shift keying (BPSK) mapping.", "We compare the proposed s-symmetric \u03b2 -expansionbased codes under automorphism ensemble successive cancellation (AE-SC) decoding with ensemble size M = 8 to the 5G standardized codes under CRC-aided successive cancellation list (CA-SCL) decoding with list size L = 8.", "The 5G codes use a 6-bit CRC code for K < 20 and a 11-bit CRC for K \u2265 20 #OTHEREFR .", "Note that AED uses random automorphisms from the full affine automorphism group of each code, which may be larger than the specified minimum symmetry in the code design. Fig.", "2 shows the required signal-to-noise-ratio (SNR) to reach a BLER of 10 \u22123 ."], "text_after_citation": ["As we can see, across all values of N and rates considered, the proposed codes using AE-SC perform similarly to the standardized codes using CA-SCL, and outperform it for shorter blocklengths.", "However, for longer blocklengths (N \u2265 512), AED falls behind SCL, but still remains within 1 dB of the 5G polar code."], "citing_paper_content": {"title": "Rate-Compatible Polar Codes For Automorphism Ensemble Decoding", "abstract": "Recently, automorphism ensemble decoding (AED) has drawn research interest as a more computationally efficient alternative to successive cancellation list (SCL) decoding of polar codes. Although AED has demonstrated superior performance for specific code parameters, a flexible code design that can accommodate varying code rates does not yet exist. This work proposes a theoretical framework for constructing ratecompatible polar codes with a prescribed automorphism group, which is a key requirement for AED. We first prove that a one-bit granular sequence with useful automorphisms cannot exist. However, by allowing larger steps in the code dimension, flexible code sequences can be constructed. An explicit synthetic channel ranking based on the \u03b2-expansion is then proposed to ensure that all constructed codes possess the desired symmetries. Simulation results, covering a broad range of code dimensions and blocklengths, show a performance comparable to that of 5G polar codes under cyclic redundancy check (CRC)-aided SCL decoding, however, with lower complexity."}, "cited_paper_content": {"title": "Coding In The Finite-Blocklength Regime: Bounds Based On Laplace Integrals And Their Asymptotic Approximations", "abstract": "In this paper, we provide new compact integral expressions and associated simple asymptotic approximations for converse and achievability bounds in the finite blocklength regime. The chosen converse and random coding union bounds were taken from the recent work of Polyanskyi\u2013Poor\u2013Verdu, and are investigated under parallel AWGN channels, the AWGN channel, the BI-AWGN channel, and the BSC. The technique we use, which is a generalization of some recent results available from the literature, is to map the probabilities of interest into a Laplace integral, and then solve (or approximate) the integral by use of a steepest descent technique. The proposed results are particularly useful for short packet lengths, where the normal approximation may provide unreliable results."}, "keywords": ["finite blocklength Polyanskyi-Poor-Verd\u00fa"], "citation_intent": "method"}
{"citing_id": "2303.01400v1", "cited_id": "1904.12334", "section_title": "General Overview Of The Methods", "citation": "Moreover, algorithms with improved approximation guarantees can be obtained that is FPT in k and #REFR .", "text_before_citation": ["Coreset construction was popularized by an initial set of works that obtained small-sized coresets in low-dimensional Euclidean spaces #OTHEREFR .", "Chen #OTHEREFR obtained the first coreset for Euclidean spaces with polynomial dependence on the dimension and the first coreset in general metrics, where the size is O(k 2 \u22122 log 2 n) for k-median.", "Subsequently, the dependence on the dimension has been further improved #OTHEREFR . Finally, such dependence was removed in #OTHEREFR .", "See also #OTHEREFR for recent improvements.", "Both k-median and k-means admit polynomial-time O(1)-approximations in general metrics #OTHEREFR ."], "text_after_citation": ["Naturally, the problems have also been studied in specialized metrics.", "Polynomial-time approximation schemes (PTASes) are known for Euclidean k-median #OTHEREFR and kmeans #OTHEREFR . See #OTHEREFR for other improvements.", "Similar to geometric clustering, clustering in graphic setting is also widely studied. PTASes are known for excluded-minor graphs #OTHEREFR .", "Also, FPT approximation schemes are known for graphs of bounded-treewidth #OTHEREFR and graphs of bounded highway dimension #OTHEREFR ."], "citing_paper_content": {"title": "Coresets For Clustering In Geometric Intersection Graphs", "abstract": "Designing coresets-small-space sketches of the data preserving cost of the solutions within (1 \u00b1)-approximate factor-is an important research direction in the study of center-based kclustering problems, such as k-means or k-median. Feldman and Langberg [STOC'11] have shown that for k-clustering of n points in general metrics, it is possible to obtain coresets whose size depends logarithmically in n. Moreover, such a dependency in n is inevitable in general metrics. A significant amount of recent work in the area is devoted to obtaining coresests whose sizes are independent of n for special metrics, like d-dimensional Euclidean space [Huang,"}, "cited_paper_content": {"title": "Tight Fpt Approximations For $K$-Median And $K$-Means", "abstract": "We investigate the fine-grained complexity of approximating the classical k-Median/k-Means clustering problems in general metric spaces. We show how to improve the approximation factors to (1+2/e+epsilon) and (1+8/e+epsilon) respectively, using algorithms that run in fixed-parameter time. Moreover, we show that we cannot do better in FPT time, modulo recent complexity-theoretic conjectures."}, "keywords": ["improved approximation guarantees"], "citation_intent": "background"}
{"citing_id": "2305.02507v1", "cited_id": "2003.11142", "section_title": "One-Shot Nas", "citation": "Following this work, BigNAS #REFR introduces several technologies to train a high-quality single-stage model, whose child models can be directly deployed without extra retraining or post-processing steps.", "text_before_citation": ["One-shot NAS is an important branch of neural architecture search (NAS) #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR that aims to drastically reduce the training and search time for the whole pipeline.", "Along this direction, #OTHEREFR trains a once-for-all (OFA) network with progressive shrinking and knowledge distillation to support all kinds of architectural settings."], "text_after_citation": ["Both OFA and BigNAS aim at simultaneously training and searching various networks with different resolutions, depths, widths and operations.", "Differently, the proposed method aims at improving a given residual network and can be seamlessly applied to the searched model of NAS.", "In addition, as OFA and BigNAS are not designed to solve the loafing problem and improve the main network results, their subnet input designs, sampling rules, and supervision signals are all different from the proposed method."], "citing_paper_content": {"title": "Stimulative Training++: Go Beyond The Performance Limits Of Residual Networks", "abstract": "Residual networks have shown great success and become indispensable in recent deep neural network models. In this work, we aim to re-investigate the training process of residual networks from a novel social psychology perspective of loafing, and further propose a new training scheme as well as three improved strategies for boosting residual networks beyond their performance limits. Previous research has suggested that residual networks can be considered as ensembles of shallow networks, which implies that the final performance of a residual network is influenced by a group of subnetworks. We identify a previously overlooked problem that is analogous to social loafing, where subnetworks within a residual network are prone to exert less effort when working as part of a group compared to working alone. We define this problem as network loafing. Similar to the decreased individual productivity and overall performance as demonstrated in society, network loafing inevitably causes sub-par performance. Inspired by solutions from social psychology, we first propose a novel training scheme called stimulative training, which randomly samples a residual subnetwork and calculates the KL divergence loss between the sampled subnetwork and the given residual network for extra supervision. In order to unleash the potential of stimulative training, we further propose three simple-yet-effective strategies, including a novel KL-loss that only aligns the network logits direction, random smaller inputs for subnetworks, and inter-stage sampling rules. Comprehensive experiments and analysis verify the effectiveness of stimulative training as well as its three improved strategies. For example, the proposed method can boost the performance of ResNet50 on ImageNet to 80.5% Top1 accuracy without using any extra data, model, trick, or changing the structure. With only uniform augment, the performance can be further improved to 81.0% Top1 accuracy, better than the best training recipes provided by Timm library and PyTorch official version. We also verify its superiority on various typical models, datasets, and tasks and give some theoretical analysis. As such, we advocate utilizing the proposed method as a general and next-generation technology to train residual networks. The code is available at https://github.com/Sunshine-Ye/NIPS22-ST."}, "cited_paper_content": {"title": "Bignas: Scaling Up Neural Architecture Search With Big Single-Stage Models", "abstract": "Neural architecture search (NAS) has shown promising results discovering models that are both accurate and fast. For NAS, training a one-shot model has become a popular strategy to rank the relative quality of different architectures (child models) using a single set of shared weights. However, while one-shot model weights can effectively rank different network architectures, the absolute accuracies from these shared weights are typically far below those obtained from stand-alone training. To compensate, existing methods assume that the weights must be retrained, finetuned, or otherwise post-processed after the search is completed. These steps significantly increase the compute requirements and complexity of the architecture search and model deployment. In this work, we propose BigNAS, an approach that challenges the conventional wisdom that post-processing of the weights is necessary to get good prediction accuracies. Without extra retraining or post-processing steps, we are able to train a single set of shared weights on ImageNet and use these weights to obtain child models whose sizes range from 200 to 1000 MFLOPs. Our discovered model family, BigNASModels, achieve top-1 accuracies ranging from 76.5% to 80.9%, surpassing state-of-the-art models in this range including EfficientNets and Once-for-All networks without extra retraining or post-processing. We present ablative study and analysis to further understand the proposed BigNASModels."}, "keywords": ["extra retraining", "BigNAS"], "citation_intent": "method"}
{"citing_id": "2303.14011v1", "cited_id": "1910.13461", "section_title": "A. Base Model", "citation": "We use BART #REFR , a sequence-to-sequence pretrained Transformer, to initialize both the encoder and decoder.", "text_before_citation": ["The self-attention and cross-attention layers can then be written as:", "EQUATION", "where h enc represents the final encoder hidden states.", "Based on the self-attention and cross-attention layers, the encoder (ENC) and decoder (DEC) layers can be expressed as:", "EQUATION"], "text_after_citation": [], "citing_paper_content": {"title": "Spec: Summary Preference Decomposition For Low-Resource Abstractive Summarization", "abstract": "Neural abstractive summarization has been widely studied and achieved great success with large-scale corpora. However, the considerable cost of annotating data motivates the need for learning strategies under low-resource settings. In this paper, we investigate the problems of learning summarizers with only few examples and propose corresponding methods for improvements. First, typical transfer learning methods are prone to be affected by data properties and learning objectives in the pretext tasks. Therefore, based on pretrained language models, we further present a meta learning framework to transfer few-shot learning processes from source corpora to the target corpus. Second, previous methods learn from training examples without decomposing the content and preference. The generated summaries could therefore be constrained by the preference bias in the training set, especially under low-resource settings. As such, we propose decomposing the contents and preferences during learning through the parameter modulation, which enables control over preferences during inference. Third, given a target application, specifying required preferences could be non-trivial because the preferences may be difficult to derive through observations. Therefore, we propose a novel decoding method to automatically estimate suitable preferences and generate corresponding summary candidates from the few training examples. Extensive experiments demonstrate that our methods achieve state-of-the-art performance on six diverse corpora with 30.11%/33.95%/27.51% and 26.74%/31.14%/24.48% average improvements on ROUGE-1/2/L under 10-and 100example settings."}, "cited_paper_content": {"title": "Bart: Denoising Sequence-To-Sequence Pre-Training For Natural Language Generation, Translation, And Comprehension", "abstract": "We present BART, a denoising autoencoder for pretraining sequence-to-sequence models. BART is trained by (1) corrupting text with an arbitrary noising function, and (2) learning a model to reconstruct the original text. It uses a standard Tranformer-based neural machine translation architecture which, despite its simplicity, can be seen as generalizing BERT (due to the bidirectional encoder), GPT (with the left-to-right decoder), and many other more recent pretraining schemes. We evaluate a number of noising approaches, finding the best performance by both randomly shuffling the order of the original sentences and using a novel in-filling scheme, where spans of text are replaced with a single mask token. BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE. BART also provides a 1.1 BLEU increase over a back-translation system for machine translation, with only target language pretraining. We also report ablation experiments that replicate other pretraining schemes within the BART framework, to better measure which factors most influence end-task performance."}, "keywords": ["decoder"], "citation_intent": "method"}
{"citing_id": "2303.01894v3", "cited_id": "1711.10398", "section_title": "Trr360D Annotation Format", "citation": "According to the DOTA #REFR dataset annotation format, a line in the text file corresponding to the annotation of a rotated table instance is shown 2.", "text_before_citation": [], "text_after_citation": ["Point A represents the top-left point of the table, and points ABCD are arranged clockwise.", "Point D represents the detection difficulty of the sample, which is uniformly defined as 0 in TRR360D.", "EQUATION", "In this format, x1, y1, x2, y2, x3, y3, x4, and y4 represent the x and y coordinates of the four points A, B, C, and D, respectively.", "The last column represents the detection difficulty, which is set to 0 for all samples in TRR360D."], "citing_paper_content": {"title": "Trr360D: A Dataset For 360 Degree Rotated Rectangular Box Table Detection", "abstract": "To address the problem of scarcity and high annotation costs of rotated image table detection datasets, this paper proposes a method for building a rotated image table detection dataset. Based on the ICDAR2019MTD modern table detection dataset, we refer to the annotation format of the DOTA dataset to create the TRR360D rotated table detection dataset. The training set contains 600 rotated images and 977 annotated instances, and the test set contains 240 rotated images and 499 annotated instances. The AP50(T<90) evaluation metric is defined, and this dataset is available for future researchers to study rotated table detection algorithms and promote the development of table detection technology. The TRR360D rotated table detection dataset was created by constraining the starting point and annotation direction, and is publicly available at https://github.com/vansin/ TRR360D."}, "cited_paper_content": {"title": "Dota: A Large-Scale Dataset For Object Detection In Aerial Images", "abstract": "Object detection is an important and challenging problem in computer vision. Although the past decade has witnessed major advances in object detection in natural scenes, such successes have been slow to aerial imagery, not only because of the huge variation in the scale, orientation and shape of the object instances on the earth's surface, but also due to the scarcity of well-annotated datasets of objects in aerial scenes. To advance object detection research in Earth Vision, also known as Earth Observation and Remote Sensing, we introduce a large-scale Dataset for Object deTection in Aerial images (DOTA). To this end, we collect $2806$ aerial images from different sensors and platforms. Each image is of the size about 4000-by-4000 pixels and contains objects exhibiting a wide variety of scales, orientations, and shapes. These DOTA images are then annotated by experts in aerial image interpretation using $15$ common object categories. The fully annotated DOTA images contains $188,282$ instances, each of which is labeled by an arbitrary (8 d.o.f.) quadrilateral To build a baseline for object detection in Earth Vision, we evaluate state-of-the-art object detection algorithms on DOTA. Experiments demonstrate that DOTA well represents real Earth Vision applications and are quite challenging."}, "keywords": ["rotated table instance", "DOTA dataset annotation"], "citation_intent": "method"}
{"citing_id": "2303.11045v1", "cited_id": "2002.00160", "section_title": "D. Independent Groups", "citation": "As all transactions are globally shared in GeoBFT #REFR , each cluster has access to all transactions, making cross-cluster transactions cheap.", "text_before_citation": ["There are two broad approaches to this: either with the actual transactions #OTHEREFR , #OTHEREFR - #OTHEREFR , or with \"prepackaged\" blocks of transactions #OTHEREFR , #OTHEREFR arranged in the upper layer.", "For example, in GeoBFT #OTHEREFR or DP-Hybrid #OTHEREFR , every transaction is shared through the upper layer with the other groups.", "In C-PBFT #OTHEREFR , on the other hand, the lower layer already constructs blocks of transactions, which are then confirmed by adding the block header to the upper-layer blockchain.", "As with sharding, it can happen that some transaction requires data from multiple groups.", "For this, affected groups can coordinate with their least common ancestor group #OTHEREFR ."], "text_after_citation": ["As global sharing is performed before execution, this leads to higher latencies.", "Different layers can have different failure modes #OTHEREFR , #OTHEREFR , e. g., when different applications use a shared blockchain. This can be used to optimize the system, e.", "g., if the applications require different security or fault tolerance levels.", "Then it might be possible to use different consensus protocols in separate lower layers, such as BFT or crash fault-tolerant protocols.", "Further, there are protocols that measure the behavior of nodes and reward good behavior, e."], "citing_paper_content": {"title": "Sok: Scalability Techniques For Bft Consensus", "abstract": "With the advancement of blockchain systems, many recent research works have proposed distributed ledger technology (DLT) that employs Byzantine fault-tolerant (BFT) consensus protocols to decide which block to append next to the ledger. Notably, BFT consensus can offer high performance, energy efficiency, and provable correctness properties, and it is thus considered a promising building block for creating highly resilient and performant blockchain infrastructures. Yet, a major ongoing challenge is to make BFT consensus applicable to large-scale environments. A large body of recent work addresses this challenge by developing novel ideas to improve the scalability of BFT consensus, thus opening the path for a new generation of BFT protocols tailored to the needs of blockchain. In this survey, we create a systematization of knowledge about the novel scalability-enhancing techniques that state-of-the-art BFT consensus protocols use. For our comparison, we closely analyze the efforts, assumptions, and trade-offs these protocols make."}, "cited_paper_content": {"title": "Resilientdb: Global Scale Resilient Blockchain Fabric", "abstract": "Recent developments in blockchain technology have inspired innovative new designs in distributed resilient database systems. At their core, these database systems typically use Byzantine fault-tolerant consensus protocols to maintain a common state across all replicas, even if some replicas are faulty or malicious. Unfortunately, existing consensus protocols are not designed to deal with geo-scale deployments in which many replicas spread across a geographically large area participate in consensus. To address this, we present the Geo-Scale Byzantine Fault-Tolerant consensus protocol (GeoBFT). GeoBFT is designed for excellent scalability by using a topological-aware grouping of replicas in local clusters, by introducing parallelization of consensus at the local level, and by minimizing communication between clusters. To validate our vision of high-performance geo-scale resilient database systems using GeoBFT, we implement GeoBFT in our efficient ResilientDB blockchain and database fabric. We show that GeoBFT is not only sound and provides great scalability, but also outperforms state-of-the-art consensus protocols by a factor of six in geo-scale deployments."}, "keywords": ["cross-cluster transactions"], "citation_intent": "background"}
{"citing_id": "2303.10752v4", "cited_id": "1505.04597", "section_title": "Daif-Net", "citation": "We modify the encoder of the U-Net #REFR , and design our DAIF-Net, as shown in Figure 4 .", "text_before_citation": ["Taking advantage that the defocus images in a focal stack are sharing the same depth and AIF image.", "Thus, we can train our model on a focal stack to make this task solvable, and our DAIF-Net has the following expression:", "EQUATION", "where that the accurate depth map and AIF image is the only solution to this function.", "Consequently, from the aspect of model architecture, the DAIF-Net is supposed to take a focal stack with an arbitrary number of images and output the underlying AIF image and depth map."], "text_after_citation": ["In order to support a flexible number of input branches, every image will be passed through the same encoder and bottleneck.", "In order to learn the depth and AIF image from a focal stack simultaneously, we observe that sharpness is the connection between them.", "Intuitively, a sharper point indicates a closer distance between the depth and the focus distance.", "Meanwhile, sharper points better preserve the corresponding colors of the AIF images.", "Therefore, training a model that is sensitive to the sharpest points in the focal stacks is beneficial to predicting both AIF images and depth maps."], "citing_paper_content": {"title": "Fully Self-Supervised Depth Estimation From Defocus Clue", "abstract": "Depth-from-defocus (DFD), modeling the relationship between depth and defocus pattern in images, has demonstrated promising performance in depth estimation. Recently, several self-supervised works try to overcome the difficulties in acquiring accurate depth ground-truth. However, they depend on the all-in-focus (AIF) images, which cannot be captured in real-world scenarios. Such limitation discourages the applications of DFD methods. To tackle this issue, we propose a completely self-supervised framework that estimates depth purely from a sparse focal stack. We show that our framework circumvents the needs for the depth and AIF image ground-truth, and receives superior predictions, thus closing the gap between the theoretical success of DFD works and their applications in the real world. In particular, we propose (i) a more realistic setting for DFD tasks, where no depth or AIF image ground-truth is available; (ii) a novel selfsupervision framework that provides reliable predictions of depth and AIF image under the challenging setting. The proposed framework uses a neural model to predict the depth and AIF image, and utilizes an optical model to validate and refine the prediction. We verify our framework on three benchmark datasets with rendered focal stacks and real focal stacks. Qualitative and quantitative evaluations show that our method provides a strong baseline for selfsupervised DFD tasks. The source code is publicly available at https://github.com/Ehzoahis/DEReD."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["DAIF-Net"], "citation_intent": "method"}
{"citing_id": "2304.02014v1", "cited_id": "2002.08155", "section_title": "Background And Related Work 2.1 Large Language Model", "citation": "Recently, researchers have begun adopting LLMs for code-related tasks, e.g., via further fine-tuning LLMs on code snippets from open-source repositories #REFR .", "text_before_citation": ["Large Language Models (LLMs) have demonstrated impressive performance across a wide range of NLP tasks by training on large corpora of text data scraped from the Internet #OTHEREFR ."], "text_after_citation": ["LLMs can be classified based on variations of the popular Transformer architecture #OTHEREFR into: Encoder-only, Decoder-only and Encoder-Decoder models.", "Decoder-only LLMs (e.g., GPT #OTHEREFR 48] , Codex [11] and CodeGen #OTHEREFR ) focus on autoregressive completion tasks by learning to predict the probability of the next token given previously generated tokens.", "Encoder-only (e.g., CodeBERT #OTHEREFR and GraphCodeBERT #OTHEREFR ) and Encoder-Decoder (e.g., CodeT5 #OTHEREFR and PLBART #OTHEREFR ) models on the other hand are designed for infilling tasks, where the pre-training objective is to recover masked-out tokens or token spans in the training data by using bi-directional context.", "In order to apply LLMs for down-stream applications, there are two main paradigms: Fine-tuning #OTHEREFR and In-context learning #OTHEREFR .", "Fine-tuning is the process of updating the model weights by learning the desired output from the given input of a specific downstream task dataset."], "citing_paper_content": {"title": "Large Language Models Are Edge-Case Fuzzers: Testing Deep Learning Libraries Via Fuzzgpt", "abstract": "Bugs in Deep Learning (DL) libraries may affect almost all downstream DL applications, and it is crucial to ensure the quality of such systems. It is challenging to generate valid input programs for fuzzing DL libraries, since the input programs need to satisfy both the syntax/semantics of the supported languages (e.g., Python) and the tensor/operator constraints for constructing valid computational graphs. Recently, the TitanFuzz work demonstrates that modern Large Language Models (LLMs) can be directly leveraged to implicitly learn all the language and DL computation constraints to generate valid programs for fuzzing DL libraries. However, LLMs tend to generate ordinary programs following similar patterns/tokens with typical programs seen in their massive training corpora (e.g., GitHub), while fuzzing favors unusual inputs that cover edge cases or are unlikely to be manually produced. To fill this gap, this paper proposes FuzzGPT, the first technique to prime LLMs to synthesize unusual programs for fuzzing. FuzzGPT is built on the well-known hypothesis that historical bug-triggering programs may include rare/valuable code ingredients important for bug finding. Meanwhile, while traditional techniques leveraging such historical information require intensive human efforts to both design dedicated generators and ensure the syntactic/semantic validity of generated programs, FuzzGPT demonstrates that this process can be fully automated via the intrinsic capabilities of LLMs (including fine-tuning and in-context learning), while being generalizable and applicable to challenging domains. While FuzzGPT can be applied with different LLMs, this paper focuses on the powerful GPT-style models: Codex and CodeGen. Moreover, FuzzGPT also shows the potential of directly leveraging the instruct-following capability of the recent ChatGPT for effective fuzzing. The experimental study on two popular DL libraries (PyTorch and TensorFlow) shows that FuzzGPT can substantially outperform TitanFuzz, detecting 76 bugs, with 49 already confirmed as previously unknown bugs, including 11 high-priority bugs or security vulnerabilities."}, "cited_paper_content": {"title": "Codebert: A Pre-Trained Model For Programming And Natural Languages", "abstract": "We present CodeBERT, a bimodal pre-trained model for programming language (PL) and nat-ural language (NL). CodeBERT learns general-purpose representations that support downstream NL-PL applications such as natural language codesearch, code documentation generation, etc. We develop CodeBERT with Transformer-based neural architecture, and train it with a hybrid objective function that incorporates the pre-training task of replaced token detection, which is to detect plausible alternatives sampled from generators. This enables us to utilize both bimodal data of NL-PL pairs and unimodal data, where the former provides input tokens for model training while the latter helps to learn better generators. We evaluate CodeBERT on two NL-PL applications by fine-tuning model parameters. Results show that CodeBERT achieves state-of-the-art performance on both natural language code search and code documentation generation tasks. Furthermore, to investigate what type of knowledge is learned in CodeBERT, we construct a dataset for NL-PL probing, and evaluate in a zero-shot setting where parameters of pre-trained models are fixed. Results show that CodeBERT performs better than previous pre-trained models on NL-PL probing."}, "keywords": ["code snippets"], "citation_intent": "background"}
{"citing_id": "2304.04410v1", "cited_id": "1911.12834", "section_title": "Existing Results", "citation": "The work of #REFR follows a similar paradigm, but randomly selects one non-empty dimension from s dimensions.", "text_before_citation": ["It remains an open question whether it holds the same bounds as the general case or has tighter bounds.", "Recently, for a broad family of -LDP estimation problems that can be cast as mean estimation problems, #OTHEREFR studies sample complexity lower bounds under certain error tolerance \u03b1, but their sample complexity results for s-sparse numerical vectors exhibit at least a 1/\u03b1 gap compared to our minimax optimal sample complexity results.", "In practice, numerous -LDP mechanisms have been proposed for statistical estimation, such as multinomial distribution estimation on categorical data #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR and one-dimensional mean estimation on numerical values #OTHEREFR , #OTHEREFR .", "For -LDP numerical vector or key-value data aggregation, existing approaches handle both dense numerical vectors (e.g., in #OTHEREFR , #OTHEREFR , #OTHEREFR ) and sparse numerical vectors (e.g., in #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR ).", "Specifically, #OTHEREFR , #OTHEREFR uniformly and randomly select one dimension from [d] and transform the multi-dimensional estimation problem to a one-dimensional numerical/categorical problem."], "text_after_citation": ["However, as we will show in Section 3, these mechanisms are sub-optimal.", "To mitigate the high noise needed for LDP, #OTHEREFR , #OTHEREFR introduce a (semi-trusted) shuffler to hide private views in the crowd.", "The seminal work #OTHEREFR shows that n shuffled -LDP views can preserve (O( log(1/\u03b4)/n), \u03b4)-differential privacy.", "The work #OTHEREFR derives a similar conclusion specifically for binary randomized response messages.", "A later work #OTHEREFR considers private views from other users as a \"privacy blanket\" and derives tighter privacy amplification bounds (O(min{ 0 , 1}e 0 log(1/\u03b4)/n), \u03b4)."], "citing_paper_content": {"title": "Differentially Private Numerical Vector Analyses In The Local And Shuffle Model", "abstract": "Numerical vector aggregation plays a crucial role in privacy-sensitive applications, such as distributed gradient estimation in federated learning and statistical analysis of key-value data. In the context of local differential privacy, this study provides a tight minimax error bound of O(ds n 2), where d represents the dimension of the numerical vector and s denotes the number of non-zero entries. By converting the conditional/unconditional numerical mean estimation problem into a frequency estimation problem, we develop an optimal and efficient mechanism called Collision. In contrast, existing methods exhibit sub-optimal error rates of O(d 2 n 2) or O(ds 2 n 2). Specifically, for unconditional mean estimation, we leverage the negative correlation between two frequencies in each dimension and propose the CoCo mechanism, which further reduces estimation errors for mean values compared to Collision. Moreover, to surpass the error barrier in local privacy, we examine privacy amplification in the shuffle model for the proposed mechanisms and derive precisely tight amplification bounds. Our experiments validate and compare our mechanisms with existing approaches, demonstrating significant error reductions for frequency estimation and mean estimation on numerical vectors."}, "cited_paper_content": {"title": "Pckv: Locally Differentially Private Correlated Key-Value Data Collection With Optimized Utility", "abstract": "Data collection under local differential privacy (LDP) has been mostly studied for homogeneous data. Real-world applications often involve a mixture of different data types such as key-value pairs, where the frequency of keys and mean of values under each key must be estimated simultaneously. For key-value data collection with LDP, it is challenging to achieve a good utility-privacy tradeoff since the data contains two dimensions and a user may possess multiple key-value pairs. There is also an inherent correlation between key and values which if not harnessed, will lead to poor utility. In this paper, we propose a locally differentially private key-value data collection framework that utilizes correlated perturbations to enhance utility. We instantiate our framework by two protocols PCKV-UE (based on Unary Encoding) and PCKV-GRR (based on Generalized Randomized Response), where we design an advanced Padding-and-Sampling mechanism and an improved mean estimator which is non-interactive. Due to our correlated key and value perturbation mechanisms, the composed privacy budget is shown to be less than that of independent perturbation of key and value, which enables us to further optimize the perturbation parameters via budget allocation. Experimental results on both synthetic and real-world datasets show that our proposed protocols achieve better utility for both frequency and mean estimations under the same LDP guarantees than state-of-the-art mechanisms."}, "keywords": ["one non-empty dimension"], "citation_intent": "background"}
{"citing_id": "2303.07107v1", "cited_id": "1201.0490", "section_title": "Experiment Settings", "citation": "For the validation method, we opted for a stratified 10-fold cross-validation with an MCC score metric since k-fold cross-validation segments the training data set in folds, each containing approximately the same distribution of class labels as in the training data set #REFR .", "text_before_citation": ["The wallclock-time represents the time in seconds between starting the SMAC optimisation and terminating it.", "In order to find the wallclock-time SMAC needs to find near-optimal hyperparameter values, we proceeded as follows: First, we split the entire data set into 67% training and 33% test set, and then, with [\"linear\", \"poly\", \"rbf\", \"sigmoid\"]", "Categorical increment-steps of 25 seconds, we performed 15 independent optimisation runs per wallclock-time on the pipeline configuration yielding the most hyperparameters (i.e., noise removal with an RF model).", "This method resulted in a wallclock-time of 500 seconds at which the highest performance reflected in the MCC score was reached. This optimal value is adopted in the bootstrapping protocol.", "A single pipeline optimization within the bootstrapping protocol #OTHEREFR consists of the following steps: First, we split the entire data set again into a 67% training and 33% test set after which we exclusively utilise the training set to perform the optimisation with SMAC."], "text_after_citation": ["Given that SMAC targets HPO as a minimisation problem #OTHEREFR , we aggregated over the 10 resulting performances and subtracted the outcome from 1 as an indicator of the respective pipeline configuration performance.", "In this respect, we let SMAC optimise a single pipeline configuration 15 times on the training data set.", "Next, from the 15 generated ones, we randomly sample 5 incumbents from which we select the one yielding the lowest performance score.", "We repeated this process 50 times for every pipeline configuration.", "Subsequently, we retrained 50 models according to the 50 best-found incumbents on the training data set."], "citing_paper_content": {"title": "Automated Classification Of Pre-Defined Movement Patterns: A Comparison Between Gnss And Uwb Technology", "abstract": "Advanced real-time location systems (RTLS) allow for collecting spatio-temporal data from human movement behaviours. Tracking individuals in small areas such as schoolyards or nursing homes might impose difficulties for RTLS in terms of positioning accuracy. However, to date, few studies have investigated the performance of different localisation systems regarding the classification of human movement patterns in small areas. The current study aims to design and evaluate an automated framework to classify human movement trajectories obtained from two different RTLS: Global Navigation Satellite System (GNSS) and Ultra-wideband (UWB), in areas of approximately 100 square meters. Specifically, we designed a versatile framework which takes GNSS or UWB data as input, extracts features from these data and classifies them according to the annotated spatial patterns. The automated framework contains three choices for applying noise removal: (i) no noise removal, (ii) Savitzky Golay filter on the raw location data or (iii) Savitzky Golay filter on the extracted features, as well as three choices regarding the classification algorithm: Decision Tree (DT), Random Forest (RF) or Support Vector Machine (SVM). We integrated different stages within the framework with the Sequential Model-Based Algorithm Configuration (SMAC) to perform automated hyperparameter optimisation. The best performance is achieved with a pipeline consisting of noise removal applied to the raw location data with an RF model for the GNSS and no noise removal with an SVM model for the UWB. We further demonstrate through statistical analysis that the UWB achieves significantly higher results than the GNSS in classifying movement patterns."}, "cited_paper_content": {"title": "Scikit-Learn: Machine Learning In Python", "abstract": "Scikit-learn is a Python module integrating a wide range of state-of-the-art machine learning algorithms for medium-scale supervised and unsupervised problems. This package focuses on bringing machine learning to non-specialists using a general-purpose high-level language. Emphasis is put on ease of use, performance, documentation, and API consistency. It has minimal dependencies and is distributed under the simplified BSD license, encouraging its use in both academic and commercial settings. Source code, binaries, and documentation can be downloaded from http://scikit-learn.sourceforge.net."}, "keywords": ["training data"], "citation_intent": "method"}
{"citing_id": "2304.05857v1", "cited_id": "1910.03882", "section_title": "Scordelis-Lo Roof With Four Holes Under Single Load", "citation": "The deformation plot has its maximum and minimum at u z,min \u2248 \u22123.3 \u2022 10 \u22121 and u z,max \u2248 3.7 \u2022 10 \u22121 , respectively. These results are also obtained in #REFR .", "text_before_citation": ["25 with the trimming curves shown in red and the point load applied in the center in green.", "The results of the model can be compared to deformation plots presented in #OTHEREFR Figure 11 (a) ], which proposes an IGA shell with adaptive refinement. Fig.", "18 shows the deformed structure with the displacement in z-direction for the SB-IGA approach of p = 3, r = 1 and h = 0.125.", "u z Figure 18 : Deformed Scordelis-Lo roof with four holes with coloring of the z-displacement obtained for p = 3, h = 1/6, and r = 1.", "The results point out, that the approach derived herein is capable of computing the deformation for trimmed domains of shells."], "text_after_citation": ["Good agreement is visible on the whole domain including the area of the load application and the free edges."], "citing_paper_content": {"title": "Smooth Multi-Patch Scaled Boundary Isogeometric Analysis For Kirchhoff-Love Shells", "abstract": "In this work, a linear Kirchhoff-Love shell formulation in the framework of scaled boundary isogeometric analysis is presented that aims to provide a simple approach to trimming for NURBS-based shell analysis. To obtain a global C 1-regular test function space for the shell discretization, an inter-patch coupling is applied with adjusted basis functions in the vicinity of the scaling center to ensure the approximation ability. Doing so, the scaled boundary geometries are related to the concept of analysis-suitable G 1 parametrizations.This yields a coupling of patch boundaries in a strong sense that is restricted to G 1-smooth surfaces. The proposed approach is advantageous to trimmed geometries due to the incorporation of the trimming curve in the boundary representation that provides an exact representation in the planar domain. The potential of the approach is demonstrated by several problems of untrimmed and trimmed geometries of Kirchhoff-Love shell analysis evaluated against error norms and displacements. Lastly, the applicability is highlighted in the analysis of a violin structure including arbitrarily shaped patches."}, "cited_paper_content": {"title": "Adaptive Isogeometric Analysis On Two-Dimensional Trimmed Domains Based On A Hierarchical Approach", "abstract": "The focus of this work is on the development of an error-driven isogeometric framework, capable of automatically performing an adaptive simulation in the context of second- and fourth-order, elliptic partial differential equations defined on two-dimensional trimmed domains. The method is steered by an a posteriori error estimator, which is computed with the aid of an auxiliary residual-like problem formulated onto a space spanned by splines with single element support. The local refinement of the basis is achieved thanks to the use of truncated hierarchical B-splines. We prove numerically the applicability of the proposed estimator to various engineering-relevant problems, namely the Poisson problem, linear elasticity and Kirchhoff-Love shells, formulated on trimmed geometries. In particular, we study several benchmark problems which exhibit both smooth and singular solutions, where we recover optimal asymptotic rates of convergence for the error measured in the energy norm and we observe a substantial increase in accuracy per-degree-of-freedom compared to uniform refinement. Lastly, we show the applicability of our framework to the adaptive shell analysis of an industrial-like trimmed geometry modeled in the commercial software Rhinoceros, which represents the B-pillar of a car."}, "keywords": ["deformation plot"], "citation_intent": "result"}
{"citing_id": "2304.05543v1", "cited_id": "0803.0811", "section_title": "Candidate Generation Using Gpsp", "citation": "However, due to the 0 -norm constraint, exactly solving the non-convex and non-differentiable problem (9) is NP-hard #REFR .", "text_before_citation": ["where A and y are obtained from A and y by normalizing each column, i.e., the column norms are 1, and", "EQUATION", "represents the number of groups with non-zero coefficients.", "The constraint enforces group sparsity by explicitly specifying that only k groups of features have nonzero coefficients.", "The solution of (9) corresponds to a PDE model with exactly k features that best fits the given data."], "text_after_citation": ["We propose Group Projected Subspace Pursuit (GPSP) to find a group k-sparse vector c * (k) for k = 1, 2, . . . , K max .", "Given a fixed k, the proposed GPSP iteratively searches for k groups highly correlated to the residuals in a greedy manner (See Section 3).", "For each sparsity level k, we denote the index set corresponding to the active group features by T (k) \u2286 {1, . . . , G}.", "We obtain K max candidate PDEs whose active features are indexed by T (k) for each sparsity level k = 1, . . . , K max respectively."], "citing_paper_content": {"title": "Group Projected Subspace Pursuit For Identification Of Variable Coefficient Differential Equations (Gp-Ident)", "abstract": "We propose an effective and robust algorithm for identifying partial differential equations (PDEs) with space-time varying coefficients from a single trajectory of noisy observations. Identifying unknown differential equations from noisy observations is a difficult task, and it is even more challenging with space and time varying coefficients in the PDE. The proposed algorithm, GP-IDENT, has three ingredients: (i) we use B-spline bases to express the unknown space and time varying coefficients, (ii) we propose Group Projected Subspace Pursuit (GPSP) to find a sequence of candidate PDEs with various levels of complexity, and (iii) we propose a new criterion for model selection using the Reduction in Residual (RR) to choose an optimal one among the pool of candidates. The new GPSP considers group projected subspaces which is more robust than existing methods in distinguishing correlated group features. We test GP-IDENT on a variety of PDEs and PDE systems, and compare it with the state-of-the-art parametric PDE identification algorithms under different settings to illustrate its outstanding performance. Our experiments show that GP-IDENT is effective in identifying the correct terms from a large dictionary and the model selection scheme is robust to noise."}, "cited_paper_content": {"title": "Subspace Pursuit For Compressive Sensing Signal Reconstruction", "abstract": "We propose a new method for reconstruction of sparse signals with and without noisy perturbations, termed the subspace pursuit algorithm. The algorithm has two important characteristics: low computational complexity, comparable to that of orthogonal matching pursuit techniques when applied to very sparse signals, and reconstruction accuracy of the same order as that of linear programming (LP) optimization methods. The presented analysis shows that in the noiseless setting, the proposed algorithm can exactly reconstruct arbitrary sparse signals provided that the sensing matrix satisfies the restricted isometry property with a constant parameter. In the noisy setting and in the case that the signal is not exactly sparse, it can be shown that the mean-squared error of the reconstruction is upper-bounded by constant multiples of the measurement and signal perturbation energies."}, "keywords": ["0 -norm constraint"], "citation_intent": "background"}
{"citing_id": "2303.06532v1", "cited_id": "1401.3492", "section_title": "Local Search.", "citation": "Global perturbation means reinitializing the currently designed algorithm with a certain probability to escape from local optima #REFR .", "text_before_citation": ["The iterative local search #OTHEREFR was utilized as the design strategy in #OTHEREFR .", "It is often used to search the design space with algorithmic operators. Its key steps are local improvement and global perturbation.", "Here, local improvement refers to changing only one dimension of the representation of the currently designed algorithm at each iteration.", "A local optima is reached once the change in this dimension cannot bring algorithm performance improvement."], "text_after_citation": ["Apart from the iterative local search, SA #OTHEREFR was also employed as the design strategy #OTHEREFR .", "The workflow of local search for metaheuristic algorithm design is shown in Fig. 5 . 4.1.2 Evolutionary Search.", "The evolutionary search #OTHEREFR serves as the design strategy in the GP-based algorithm design literature.", "It uses crossover and mutation to globally search the design space.", "The sub-tree crossover and mutation are common for the GP-based design with tree representation #OTHEREFR ."], "citing_paper_content": {"title": "A Survey On Automated Design Of Metaheuristic Algorithms", "abstract": "Metaheuristic algorithms have attracted wide attention from academia and industry due to their capability of conducting search independent of problem structures and problem domains. Often, human experts are requested to manually tailor algorithms to fit for solving a targeted problem. The manual tailoring process may be laborious, error-prone, and require intensive specialized knowledge. This gives rise to increasing interests and demands for automated design of metaheuristic algorithms with less human intervention. The automated design could make high-performance algorithms accessible to a much broader range of researchers and practitioners; and by leveraging computing power to fully explore the potential design choices, automated design could reach or even surpass human-level design. This paper presents a broad picture of the formalization, methodologies, challenges, and research trends of automated design of metaheuristic algorithms, by conducting a survey on the common grounds and representative techniques in this field. In the survey, we first present the concept of automated design of metaheuristic algorithms and provide a taxonomy by abstracting the automated design process into four parts, i.e., design space, design strategies, performance evaluation strategies, and targeted problems. Then, we overview the techniques concerning the four parts of the taxonomy and discuss their strengths, weaknesses, challenges, and usability, respectively. Finally, we present research trends in this field. CCS Concepts: \u2022 Computing methodologies \u2192 Search methodologies."}, "cited_paper_content": {"title": "Paramils: An Automatic Algorithm Configuration Framework", "abstract": "The identification of performance-optimizing parameter settings is an important part of the development and application of algorithms. We describe an automatic framework for this algorithm configuration problem. More formally, we provide methods for optimizing a target algorithm's performance on a given class of problem instances by varying a set of ordinal and/or categorical parameters. We review a family of local-search-based algorithm configuration procedures and present novel techniques for accelerating them by adaptively limiting the time spent for evaluating individual configurations. We describe the results of a comprehensive experimental evaluation of our methods, based on the configuration of prominent complete and incomplete algorithms for SAT. We also present what is, to our knowledge, the first published work on automatically configuring the CPLEX mixed integer programming solver. All the algorithms we considered had default parameter settings that were manually identified with considerable effort. Nevertheless, using our automated algorithm configuration procedures, we achieved substantial and consistent performance improvements."}, "keywords": ["currently designed algorithm"], "citation_intent": "background"}
{"citing_id": "2304.04991v1", "cited_id": "1909.11942", "section_title": "C. Ablation Studies", "citation": "Analysis of Group Quantities: The cross-layer weight sharing technique used in ALBert #REFR simply shares the weights of all layers in the model.", "text_before_citation": ["Moreover, the proposed multiplexing technique has also been applied in baseline Transformer model to verify its impact on the encoder and decoder, respectively. The experimental results are shown in TABLE VI.", "When only applying the multiplexing technique in the encoder, the CER can be improved by 0.1% over the baseline Transformer on the Test dataset with a model size of 17.21M and a 43% compression.", "On the other hand, 0.1% CER improvement over the baseline Transformer on the Dev dataset with a model size of 22.47M and a 26% compression can be obtained if only the decoder has been applied the multiplexing technique.", "Implementation of the weight reuse technique in both the encoder and decoder can compress the model size to 9.32M and improve 0.1% CER on the Dev dataset.", "Therefore, it is logic to say that the model can be compressed with no sacrifice of the model performance regardless of the implementation of the multiplexing technique in either encoder or decoder."], "text_after_citation": ["A similar technique is also adopted in Universal Transformer #OTHEREFR and MiniVit #OTHEREFR .", "Different from these two approaches, the Sim-T proposed in this study divides all encoder layers into multiple groups, and the proposed multiplexing technique is used in each group.", "For It should be highlighted that the model performance degrades with the decreasing number of encoder groups.", "Also, the model size gradually decreases from 9.32M to 7.99M with the increasing number of encoder groups.", "From the table, one can note that the model performance is undesirable when the encoder group is one, although a greater degree of model compression can be obtained."], "citing_paper_content": {"title": "Sim-T: Simplify The Transformer Network By Multiplexing Technique For Speech Recognition", "abstract": "In recent years, a great deal of attention has been paid to the Transformer network for speech recognition tasks due to its excellent model performance. However, the Transformer network always involves heavy computation and large number of parameters, causing serious deployment problems in devices with limited computation sources or storage memory. In this paper, a new lightweight model called Sim-T has been proposed to expand the generality of the Transformer model. Under the help of the newly developed multiplexing technique, the Sim-T can efficiently compress the model with negligible sacrifice on its performance. To be more precise, the proposed technique includes two parts, that are, module weight multiplexing and attention score multiplexing. Moreover, a novel decoder structure has been proposed to facilitate the attention score multiplexing. Extensive experiments have been conducted to validate the effectiveness of Sim-T. In Aishell-1 dataset, when the proposed Sim-T is 48% parameter less than the baseline Transformer, 0.4% CER improvement can be obtained. Alternatively, 69% parameter reduction can be achieved if the Sim-T gives the same performance as the baseline Transformer. With regard to the HKUST and WSJ eval92 datasets, CER and WER will be improved by 0.3% and 0.2%, respectively, when parameters in Sim-T are 40% less than the baseline Transformer."}, "cited_paper_content": {"title": "Albert: A Lite Bert For Self-Supervised Learning Of Language Representations", "abstract": "Increasing model size when pretraining natural language representations often results in improved performance on downstream tasks. However, at some point further model increases become harder due to GPU/TPU memory limitations, longer training times, and unexpected model degradation. To address these problems, we present two parameter-reduction techniques to lower memory consumption and increase the training speed of BERT. Comprehensive empirical evidence shows that our proposed methods lead to models that scale much better compared to the original BERT. We also use a self-supervised loss that focuses on modeling inter-sentence coherence, and show it consistently helps downstream tasks with multi-sentence inputs. As a result, our best model establishes new state-of-the-art results on the GLUE, RACE, and SQuAD benchmarks while having fewer parameters compared to BERT-large."}, "keywords": ["cross-layer weight sharing"], "citation_intent": "method"}
{"citing_id": "2304.00446v1", "cited_id": "1905.11286", "section_title": "B. Model Architecture", "citation": "NovoGrad #REFR optimizer is employed for training across a maximum of 15000 iterations on a batch of 64 randomly sampled channel realizations with early stopping.", "text_before_citation": ["Our proposed feedforward UWMMSE architecture is composed of 3 unfolded-WMMSE layers with a 2-layered CV-GCN -shared by the unfolded layers -modeling the function \u03a8 in #OTHEREFR .", "We set the hidden layer of CV-GNN as F = 32 and that of CV-MLP as G = 16. The model consists of 3302 trainable parameters."], "text_after_citation": ["The initial learning rate is set to 1 \u00d7 10 \u22122 .", "An interesting observation was that while 3 unfolded-layers offered the best performance trade-off in terms of sum-rate and time at inference, the best training performance was achieved with just 1 unfolded layer.", "Recalling that our model is flexible insofar as to have different numbers of unfolded-layers at training and inference on account of parameter-sharing, we established a model setup wherein UWMMSE is trained with a single layer and then 2 more layers are appended at inference, which share the learned components.", "Experimental results presented in this section demonstrate the effectiveness of this setup.", "At inference, we average the achieved sum-rate over 10000 channel realizations for all experiments."], "citing_paper_content": {"title": "Deep Graph Unfolding For Beamforming In Mu-Mimo Interference Networks", "abstract": "We develop an efficient and near-optimal solution for beamforming in multiuser multiple-input-multiple-output single-hop wireless ad-hoc interference networks. Inspired by the weighted minimum mean squared error (WMMSE) method, a classical approach to solving this problem, and the principle of algorithm unfolding, we present unfolded WMMSE (UWMMSE) for MU-MIMO. This method learns a parameterized functional transformation of key WMMSE parameters using graph neural networks (GNNs), where the channel and interference components of a wireless network constitute the underlying graph. These GNNs are trained through gradient descent on a network utility metric using multiple instances of the beamforming problem. Comprehensive experimental analyses illustrate the superiority of UWMMSE over the classical WMMSE and stateof-the-art learning-based methods in terms of performance, generalizability, and robustness."}, "cited_paper_content": {"title": "Stochastic Gradient Methods With Layer-Wise Adaptive Moments For Training Of Deep Networks", "abstract": "We propose NovoGrad, an adaptive stochastic gradient descent method with layer-wise gradient normalization and decoupled weight decay. In our experiments on neural networks for image classification, speech recognition, machine translation, and language modeling, it performs on par or better than well tuned SGD with momentum and Adam or AdamW. Additionally, NovoGrad (1) is robust to the choice of learning rate and weight initialization, (2) works well in a large batch setting, and (3) has two times smaller memory footprint than Adam."}, "keywords": ["channel realizations", "NovoGrad optimizer"], "citation_intent": "method"}
{"citing_id": "2304.01729v1", "cited_id": "1301.1493", "section_title": "Computational Experiments", "citation": "For finding orbits, we used nauty 2.7r3, a software library for computing automorphism groups of graphs by McKay and Piperno #REFR .", "text_before_citation": ["This explains why the only solved instances are for m = Z(d) values.", "When m = Z(d), the resulting graph is d-regular or almost d-regular, which has an equivalent edge count to the bound we discussed.", "From an inspection of the results, we can conclude that our initial approaches are not good enough for decreasing the upper bound found by CPLEX, but they are promising for finding feasible solutions. This might be due to symmetry.", "In our next experiment, we test the efficacy of using Orbital Branching with Basic Formulation (Subsection 3.1).", "We implemented Orbital Branching using the callback mechanism of CPLEX."], "text_after_citation": ["At each branch-and-cut node, we find the variables set to 0 and 1 from CPLEX and update the constraints accordingly.", "Then, we construct the updated bipartite graph using binding constraints. The resulting graph (and coloring) is sent to nauty.", "For calculating orbits, nauty calculates the automorphism groups utilizing graph isomorphism problem, which is a GI \u2212complete problem #OTHEREFR .", "This implies that the problem of calculating orbits is neither known to be NP-complete nor known to be polynomial-time solvable.", "For large instances, run time can be as high as 150 seconds, while it is around 0.1 seconds for small instances."], "citing_paper_content": {"title": "Constructing Extremal Triangle-Free Graphs Using Integer Programming", "abstract": "The maximum number of edges in a graph with matching number m and maximum degree d has been determined in [1] and [2], where some extremal graphs have also been provided. Then, a new question has emerged: how the maximum edge count is affected by forbidding some subgraphs occurring in these extremal graphs? In [3], the problem is solved in triangle-free graphs for d \u2265 m, and for d < m with either Z(d) \u2264 m < 2d or d \u2264 6, where Z(d) is approximately 5d/4. The authors derived structural properties of triangle-free extremal graphs, which allows us to focus on constructing small extremal components to form an extremal graph. Based on these findings, in this paper, we develop an integer programming formulation for constructing extremal graphs. Since our formulation is highly symmetric, we use our own implementation of Orbital Branching to reduce symmetry. We also implement our integer programming formulation so that the feasible region is restricted iteratively. Using a combination of the two approaches, we expand the solution into d \u2264 10 instead of d \u2264 6 for m > d. Our results endorse the formula for the number of edges in all extremal triangle-free graphs conjectured in [3]."}, "cited_paper_content": {"title": "Practical Graph Isomorphism, Ii", "abstract": "We report the current state of the graph isomorphism problem from the practical point of view. After describing the general principles of the refinement-individualization paradigm and proving its validity, we explain how it is implemented in several of the key programs. In particular, we bring the description of the best known program nauty up to date and describe an innovative approach called Traces that outperforms the competitors for many difficult graph classes. Detailed comparisons against saucy, Bliss and conauto are presented."}, "keywords": ["graphs"], "citation_intent": "method"}
{"citing_id": "2303.02717v1", "cited_id": "1704.00390", "section_title": "Method Seen Unseen", "citation": "We used an initial learning rate of \u03bb = 10 \u22124 and a weight decay of 10 \u22124 and trained for 30 (600) epochs for indoor (outdoor) localization. Our loss is initialized as in #REFR .", "text_before_citation": ["We follow the training procedure of #OTHEREFR and apply the augmentation method suggested by #OTHEREFR .", "Specifically, during training, images are first rescaled such that their smaller edge is resized to 256 pixels.", "We extract random 224 \u00d7 224 crops and randomly jitter the brightness, saturation, and contrast.", "At test time, images are rescaled, and the center crop is used for inference, with no further augmentations.", "Our model is optimized using Adam, with \u03b2 1 = 0.9, \u03b2 2 = 0.999 and = 10 \u221210 and a batch size of 8."], "text_after_citation": ["The training pairs for the 7Scenes dataset are generated as in #OTHEREFR .", "We further extend this dataset with pairs from #OTHEREFR .", "Pairs from the CambridgeLandmarks dataset are computed by encoding images with a pre-trained NetVLAD model #OTHEREFR , and fetching the nearest 40 neighbors based on cosine similarity.", "At train time, we sample from the pool of neighbors to generate different pairs.", "We provide the training pairs, code, configuration files, and pre-trained models in order to support the reproducibility of our results."], "citing_paper_content": {"title": "Learning To Localize In Unseen Scenes With Relative Pose Regressors", "abstract": "Relative pose regressors (RPRs) localize a camera by estimating its relative translation and rotation to a poselabelled reference. Unlike scene coordinate regression and absolute pose regression methods, which learn absolute scene parameters, RPRs can (theoretically) localize in unseen environments, since they only learn the residual pose between camera pairs. In practice, however, the performance of RPRs is significantly degraded in unseen scenes. In this work, we propose to aggregate paired feature maps into latent codes, instead of operating on global image descriptors, in order to improve the generalization of RPRs. We implement aggregation with concatenation, projection, and attention operations (Transformer Encoders) and learn to regress the relative pose parameters from the resulting latent codes. We further make use of a recently proposed continuous representation of rotation matrices, which alleviates the limitations of the commonly used quaternions. Compared to stateof-the-art RPRs, our model is shown to localize significantly better in unseen environments, across both indoor and outdoor benchmarks, while maintaining competitive performance in seen scenes. We validate our findings and architecture design through multiple ablations. Our code and pretrained models is publicly available 1 ."}, "cited_paper_content": {"title": "Geometric Loss Functions For Camera Pose Regression With Deep Learning", "abstract": "Deep learning has shown to be effective for robust and real-time monocular image relocalisation. In particular, PoseNet [22] is a deep convolutional neural network which learns to regress the 6-DOF camera pose from a single image. It learns to localize using high level features and is robust to difficult lighting, motion blur and unknown camera intrinsics, where point based SIFT registration fails. However, it was trained using a naive loss function, with hyper-parameters which require expensive tuning. In this paper, we give the problem a more fundamental theoretical treatment. We explore a number of novel loss functions for learning camera pose which are based on geometry and scene reprojection error. Additionally we show how to automatically learn an optimal weighting to simultaneously regress position and orientation. By leveraging geometry, we demonstrate that our technique significantly improves PoseNets performance across datasets ranging from indoor rooms to a small city."}, "keywords": ["indoor (outdoor) localization"], "citation_intent": "method"}
{"citing_id": "2303.10351v1", "cited_id": "1912.10211", "section_title": "The Teacher Model And Baslines", "citation": "As shown in Table 1 , our model achieves 29.6 mAP on the AudioSet evaluation set, which outperforms the CNN10 model (when also trained on balanced set) as shown in #REFR .", "text_before_citation": ["The teacher model ft uses the CNN10 architecture designed in #OTHEREFR .", "In order to achieve good performance, we train the teacher model using data augmentation techniques (e.g., mixup #OTHEREFR and specaugment #OTHEREFR ) up to 400K iterations using batch size of 64 on AudioSet balanced train set."], "text_after_citation": ["We compare OFA training with two baseline approaches: Baseline: Training each model from scratch using BCE loss and specaugment, which is a standard way to train an AEC model. Fig. 3 . Comparison of different search spaces for the OFA supernet.", "We selectively apply ED on certain blocks, indicated with \"X\".", "Knowledge Distillation (KD): Training each model with knowledge distillation and specaugment, which is a stronger baseline.", "For the architecture of individual models, based on the supernet search space, we set each model with a fixed width ratio and depth across all the dynamic blocks.", "For example, \"width ratio = 0.4, depth = 2\" for all three dynamic CNN blocks."], "citing_paper_content": {"title": "Weight-Sharing Supernet For Searching Specialized Acoustic Event Classification Networks Across Device Constraints", "abstract": "Acoustic Event Classification (AEC) has been widely used in devices such as smart speakers and mobile phones for home safety or accessibility support [1]. As AEC models run on more and more devices with diverse computation resource constraints, it became increasingly expensive to develop models that are tuned to achieve optimal accuracy/computation trade-off for each given computation resource constraint. In this paper, we introduce a Once-For-All (OFA) Neural Architecture Search (NAS) framework for AEC. Specifically, we first train a weight-sharing supernet that supports different model architectures, followed by automatically searching for a model given specific computational resource constraints. Our experimental results showed that by just training once, the resulting model from NAS significantly outperforms both models trained individually from scratch and knowledge distillation (25.4% and 7.3% relative improvement). We also found that the benefit of weightsharing supernet training of ultra-small models comes not only from searching but from optimization."}, "cited_paper_content": {"title": "Panns: Large-Scale Pretrained Audio Neural Networks For Audio Pattern Recognition", "abstract": "Audio pattern recognition is an important research topic in the machine learning area, and includes several tasks such as audio tagging, acoustic scene classification and sound event detection. Recently neural networks have been applied to solve audio pattern recognition problems. However, previous systems focus on small datasets, which limits the performance of audio pattern recognition systems. Recently in computer vision and natural language processing, systems pretrained on large datasets have generalized well to several tasks. However, there is limited research on pretraining neural networks on large datasets for audio pattern recognition. In this paper, we propose large-scale pretrained audio neural networks (PANNs) trained on AudioSet. We propose to use Wavegram, a feature learned from waveform, and the mel spectrogram as input. We investigate the performance and complexity of a variety of convolutional neural networks. Our proposed AudioSet tagging system achieves a state-of-the-art mean average precision (mAP) of 0.439, outperforming the best previous system of 0.392. We transferred a PANN to six audio pattern recognition tasks and achieve state-of-the-art performance in many tasks. Source code and pretrained models have been released."}, "keywords": ["AudioSet evaluation set"], "citation_intent": "result"}
{"citing_id": "2303.04735v1", "cited_id": "1611.08554", "section_title": "Introduction", "citation": "The work on MSC is extended in #REFR , which proves that while MSC corresponds to synchronized automata, the \u00b5-fragment of the modal \u00b5calculus similarly captures asynchronous distributed automata.", "text_before_citation": ["Related work.", "As already mentioned, descriptive complexity of distributed computing has been largely initiated in #OTHEREFR , which characterizes a range of related complexity classes via modal logics.", "It is shown, for example, that graded modal logic captures the class MB(1) containing problems solvable in constant time by algorithms whose recognition capacity is sufficient all the way up to distinguishing between multisets of incoming messages, but no further.", "In the paper, the link to logic helps also in separating some of the studied classes.", "The constant-time limitation is lifted in #OTHEREFR , which shows that finite distributed message passing automata (FMPAs) correspond to modal substitution calculus MSC, which is the logic studied also in the current paper."], "text_after_citation": ["Distributed computing with identifiers has been studied from the point of view of logic earlier in #OTHEREFR .", "The paper #OTHEREFR approaches identifiers via a uniform logical characterization of a certain class of algorithms using IDs, while our work is based on the circuit-style paradigm with formulas and circuits being given based on model size.", "Thus, the two approaches are not comparable in any uniquely obvious way.", "Nevertheless, one simple difference between our work and #OTHEREFR is that we treat IDs bit by bit as concrete bit strings.", "Thus, we can express, e.g., that the current ID has a bit 1 without implying that the current node cannot have the smallest ID in the system."], "citing_paper_content": {"title": "Descriptive Complexity For Distributed Computing With Circuits", "abstract": "We consider distributed algorithms in the realistic scenario where distributed message passing is operated via circuits. We show that within this setting, modal substitution calculus MSC captures the expressive power of circuits. The translations between circuits and MSC-programs are linear in both directions. Furthermore, we show that the colouring algorithm based on Cole-Vishkin can be specified via logarithmic size programs."}, "cited_paper_content": {"title": "Asynchronous Distributed Automata: A Characterization Of The Modal Mu-Fragment", "abstract": "We establish the equivalence between a class of asynchronous distributed automata and a small fragment of least fixpoint logic, when restricted to finite directed graphs. More specifically, the logic we consider is (a variant of) the fragment of the modal $\\mu$-calculus that allows least fixpoints but forbids greatest fixpoints. The corresponding automaton model uses a network of identical finite-state machines that communicate in an asynchronous manner and whose state diagram must be acyclic except for self-loops. Exploiting the connection with logic, we also prove that the expressive power of those machines is independent of whether or not messages can be lost."}, "keywords": ["asynchronous distributed automata"], "citation_intent": "background"}
{"citing_id": "2304.11900v1", "cited_id": "1902.09212", "section_title": "Hybrid Feature Extraction", "citation": "For multi-view RGB frames, we directly filter them using HRNet #REFR to acquire 2D feature maps F i .", "text_before_citation": ["Specifically, we follow #OTHEREFR to unproject depth image into pointcloud, voxelize and filter using 3D convolutional as coarse feature volume.", "For a query point X, we acquire its local 3D feature F D by trilinear interpolation based on point coordinates and then share it with both visibility and occupancy MLPs.", "Intuitively, the 3D feature originates from voxelized point cloud, which can be viewed as noisy samples of the underlying surface (decision boundary of the occupancy field).", "By applying 3D convolution, the network reasons the coarse geometric surface feature over a sufficiently large receptive field, consequently aiding visibility inference.", "3D feature is necessary for accurate visibility prediction as ablated in Fig. 8 ."], "text_after_citation": ["The local 2D features are then extracted in the pixel-aligned fashion as in #OTHEREFR , by projecting the point coordinate X onto each view as the image coordinate \u03c0(X), then bilinearly interpolating the corresponding feature maps F i (\u03c0(X)).", "Compared to feature volume, 2D feature maps have much higher resolution and thus grant better details.", "Follow Eq.", "(4), we infer V \u03c6|\u03c9 (n) by providing 3D feature F D and averaged 2D feature F avg :", "EQUATION"], "citing_paper_content": {"title": "Learning Visibility Field For Detailed 3D Human Reconstruction And Relighting", "abstract": "Figure 1. Given sparse-view RGBD input (left), our method reconstructs detailed geometry (middle left), albedo (middle right), and visibility for realistic 3D human relighting under different illuminations (right)."}, "cited_paper_content": {"title": "Deep High-Resolution Representation Learning For Human Pose Estimation", "abstract": "This is an official pytorch implementation of Deep High-Resolution Representation Learning for Human Pose Estimation. In this work, we are interested in the human pose estimation problem with a focus on learning reliable high-resolution representations. Most existing methods recover high-resolution representations from low-resolution representations produced by a high-to-low resolution network. Instead, our proposed network maintains high-resolution representations through the whole process. We start from a high-resolution subnetwork as the first stage, gradually add high-to-low resolution subnetworks one by one to form more stages, and connect the mutli-resolution subnetworks in parallel. We conduct repeated multi-scale fusions such that each of the high-to-low resolution representations receives information from other parallel representations over and over, leading to rich high-resolution representations. As a result, the predicted keypoint heatmap is potentially more accurate and spatially more precise. We empirically demonstrate the effectiveness of our network through the superior pose estimation results over two benchmark datasets: the COCO keypoint detection dataset and the MPII Human Pose dataset. The code and models have been publicly available at \\url{this https URL}."}, "keywords": ["multi-view RGB frames", "2D feature maps"], "citation_intent": "method"}
{"citing_id": "2304.12847v1", "cited_id": "1810.04805", "section_title": "Module 2: Ensemble", "citation": "For the individual classifiers, we examined different PLMs such as BERT #REFR , RoBERTa , and DeBERTa.", "text_before_citation": ["Similar to the full pipeline, Module 2 can also be broken down into its individual constituents, which are (i) the pre-processing module, (ii) the individual classifiers, and (iii) the ensembling method to combine the decision of these classifiers.", "Firstly, a pre-processing step is needed, as the data for the tasks was collected from noisy resources (Reddit, and Gab).", "For this, we used the same common techniques for all models.", "In particular, we converted all uppercase characters to lowercase, removed repetitive patterns like \"heeeey\" and additional spaces, eliminated special characters like emojis and hashtags (#), and deleted numbers."], "text_after_citation": ["each of these models was initially fine-tuned using the entire dataset.", "Lastly, we employed an ensemble of the three best-performing classifiers from the previous step for the final submission, namely BERTweet-large, DeBERTa-v3-large, and RoBERTa-large.", "Ensembling multiple models can potentially prevent egregious mistakes made by a single model #OTHEREFR .", "We used two ensemble approaches: majority voting, a hard voting method where the prediction of each classifier is treated as a vote, and the class with the most votes is ultimately selected as the predicted class, and soft average ensemble, in which the output of each model is averaged as shown in Equation 1.", "EQUATION"], "citing_paper_content": {"title": "Nlp-Ltu At Semeval-2023 Task 10: The Impact Of Data Augmentation And Semi-Supervised Learning Techniques On Text Classification Performance On An Imbalanced Dataset", "abstract": "In this paper, we propose a methodology for task 10 of SemEval23, focusing on detecting and classifying online sexism in social media posts. The task is tackling a serious issue, as detecting harmful content on social media platforms is crucial for mitigating the harm of these posts on users. Our solution for this task is based on an ensemble of fine-tuned transformer-based models (BERTweet, RoBERTa, and DeBERTa). To alleviate problems related to class imbalance, and to improve the generalization capability of our model, we also experiment with data augmentation and semi-supervised learning. In particular, for data augmentation, we use backtranslation, either on all classes, or on the underrepresented classes only. We analyze the impact of these strategies on the overall performance of the pipeline through extensive experiments. while for semi-supervised learning, we found that with a substantial amount of unlabelled, in-domain data available, semi-supervised learning can enhance the performance of certain models. Our proposed method (for which the source code is available on Github 12) attains an F 1-score of 0.8613 for sub-taskA, which ranked us 10th in the competition."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["individual classifiers", "BERT"], "citation_intent": "method"}
{"citing_id": "2305.00281v1", "cited_id": "1904.00078", "section_title": "Ii. Related Work", "citation": "In #REFR it is presented a tethered UAV-UGV team that navigates through unstructured or confined spaces.", "text_before_citation": ["However, the tether is assumed to be taut and the planner generates trajectories with sweeping patterns only for the UAV system.", "A tether state estimation technique is presented in #OTHEREFR , which describes a relative localization system for a marsupial tethered UGV-UAV pair.", "The estimation of the tether angles, as measured by the cable's provider SDK, and a catenary model #OTHEREFR are used to estimate the relative position of the UAV with respect to the UGV.", "The same authors propose in #OTHEREFR a PRM planner in which reachable space is constrained by the tether in two ways: avoiding any contacts of the tether with obstacles, or allowing up to two different contact points to extend the effective range.", "Nevertheless, again the planner assumes a tensed tether at any time, enabling collision checking based on ray-tracing."], "text_after_citation": ["The UAV navigates autonomously and performs as visual assistant for UGV, while UGV is teleoperated. This approach also assumes a taut tether.", "Regarding path planning, #OTHEREFR presents a method for autonomous exploration of unknown cavities in three dimensions (3D) that focuses on minimizing the distance traveled and the length of the unwound tether.", "The method uses a hierarchical framework composed of a global level of exploration path planning that solves a Traveling Salesman Problem (TSP) and a local level of path planning whose parameters make it possible to adjust the trade-off between the tether unwinding and the path cost.", "This approach also assumes the tether to be always in a taut condition.", "In #OTHEREFR , we proposed a trajectory planning method for an UAV tied to a fixed element."], "citing_paper_content": {"title": "Path And Trajectory Planning Of A Tethered Uav-Ugv Marsupial Robotic System*", "abstract": "This letter addresses the problem of trajectory planning in a marsupial robotic system consisting of an unmanned aerial vehicle (UAV) linked to an unmanned ground vehicle (UGV) through a non-taut tether with controllable length. To the best of our knowledge, this is the first method that addresses the trajectory planning of a marsupial UGV-UAV with a non-taut tether. The objective is to determine a synchronized collision-free trajectory for the three marsupial system agents: UAV, UGV, and tether. First, we present a path planning solution based on optimal Rapidly-exploring Random Trees (RRT*) with novel sampling and steering techniques to speed-up the computation. This algorithm is able to obtain collision-free paths for the UAV and the UGV, taking into account the 3D environment and the tether. Then, the letter presents a trajectory planner based on non-linear least squares. The optimizer takes into account aspects not considered in the path planning, like temporal constraints of the motion imposed by limits on the velocities and accelerations of the robots, or raising the tether's clearance. Simulated and field test results demonstrate that the approach generates obstacle-free, smooth, and feasible trajectories for the marsupial system."}, "cited_paper_content": {"title": "Autonomous Visual Assistance For Robot Operations Using A Tethered Uav", "abstract": "This paper develops an autonomous tethered aerial visual assistant for robot operations in unstructured or confined environments. Robotic tele-operation in remote environments is difficult due to lack of sufficient situational awareness, mostly caused by the stationary and limited field-of-view and lack of depth perception from the robot's onboard camera. The emerging state of the practice is to use two robots, a primary and a secondary that acts as a visual assistant to overcome the perceptual limitations of the onboard sensors by providing an external viewpoint. However, problems exist when using a tele-operated visual assistant: extra manpower, manually chosen suboptimal viewpoint, and extra teamwork demand between primary and secondary operators. In this work, we use an autonomous tethered aerial visual assistant to replace the secondary robot and operator, reducing human robot ratio from 2:2 to 1:2. This visual assistant is able to autonomously navigate through unstructured or confined spaces in a risk-aware manner, while continuously maintaining good viewpoint quality to increase the primary operator's situational awareness. With the proposed co-robots team, tele-operation missions in nuclear operations, bomb squad, disaster robots, and other domains with novel tasks or highly occluded environments could benefit from reduced manpower and teamwork demand, along with improved visual assistance quality based on trustworthy risk-aware motion in cluttered environments."}, "keywords": ["tethered UAV-UGV team"], "citation_intent": "background"}
{"citing_id": "2304.00022v1", "cited_id": "1801.07829", "section_title": "Influence Of Backbone Architecture On Fsl", "citation": "One can conclude that the graph-based network DGCNN #REFR achieves higher classification accuracy than other networks on these two datasets.", "text_before_citation": ["\u2022 Pointwise-based: PointNet #OTHEREFR and PointNet++ #OTHEREFR .", "\u2022 Convolution-based: PointCNN [15] , RSCNN [18] and DensePoint [17] .", "\u2022 Graph-based: DGCNN #OTHEREFR .", "For fair comparisons, we remove the last fully connected layers and train the networks from scratch with the same training strategy as suggested by these methods.", "After that, we feed the embedding features generated by these networks into ProtoNet #OTHEREFR for few-shot classification and the comparison results are shown in Table 2 ."], "text_after_citation": ["The reason may be that graph-based methods dynamically update the point-wise connection graph in feature space and extracts edge features with EdgeConv layers hence can learn more discriminative features.", "Therefore, we will use ProtoNet #OTHEREFR with DGCNN #OTHEREFR as the strong baseline for benchmarking with 3D FSL."], "citing_paper_content": {"title": "What Makes For Effective Few-Shot Point Cloud Classification?", "abstract": "Due to the emergence of powerful computing resources and large-scale annotated datasets, deep learning has seen wide applications in our daily life. However, most current methods require extensive data collection and retraining when dealing with novel classes never seen before. On the other hand, we humans can quickly recognize new classes by looking at a few samples, which motivates the recent popularity of few-shot learning (FSL) in machine learning communities. Most current FSL approaches work on 2D image domain, however, its implication in 3D perception is relatively under-explored. Not only needs to recognize the unseen examples as in 2D domain, 3D few-shot learning is more challenging with unordered structures, high intraclass variances, and subtle inter-class differences. Moreover, different architectures and learning algorithms make it difficult to study the effectiveness of existing 2D methods when migrating to the 3D domain. In this work, for the first time, we perform systematic and extensive studies of recent 2D FSL and 3D backbone networks for benchmarking few-shot point cloud classification, and we suggest a strong baseline and learning architectures for 3D FSL. Then, we propose a novel plug-and-play component called Cross-Instance Adaptation (CIA) module, to address the high intra-class variances and subtle inter-class differences issues, which can be easily inserted into current baselines with significant performance improvement. Extensive experiments on two newly introduced benchmark datasets, ModelNet40-FS and ShapeNet70-FS, demonstrate the superiority of our proposed network for 3D FSL."}, "cited_paper_content": {"title": "Dynamic Graph Cnn For Learning On Point Clouds", "abstract": "Point clouds provide a flexible geometric representation suitable for countless applications in computer graphics; they also comprise the raw output of most 3D data acquisition devices. While hand-designed features on point clouds have long been proposed in graphics and vision, however, the recent overwhelming success of convolutional neural networks (CNNs) for image analysis suggests the value of adapting insight from CNN to the point cloud world. Point clouds inherently lack topological information, so designing a model to recover topology can enrich the representation power of point clouds. To this end, we propose a new neural network module dubbed EdgeConv suitable for CNN-based high-level tasks on point clouds, including classification and segmentation. EdgeConv acts on graphs dynamically computed in each layer of the network. It is differentiable and can be plugged into existing architectures. Compared to existing modules operating in extrinsic space or treating each point independently, EdgeConv has several appealing properties: It incorporates local neighborhood information; it can be stacked applied to learn global shape properties; and in multi-layer systems affinity in feature space captures semantic characteristics over potentially long distances in the original embedding. We show the performance of our model on standard benchmarks, including ModelNet40, ShapeNetPart, and S3DIS."}, "keywords": ["graph-based network DGCNN"], "citation_intent": "result"}
{"citing_id": "2304.03094v1", "cited_id": "1810.06773", "section_title": "A.9 Image Segmentation Results", "citation": "Rather than re-implementing ESGD from scratch, we replicate the same image recognition experiment used by #REFR and compare PAPA to ESGD in this setting.", "text_before_citation": ["We also see that PAPA is additionally improved by soups (an advantage not shared by training a single model for p times more epochs).", "This demonstrates that PAPA can bring more performance gain than training models for more epochs.", "This suggests that PAPA could provide a better and more efficient way of training large models on large data by parallelizing the training length over multiple PAPA networks trained for less time.", "#OTHEREFR ] is a method that leverages a population of networks through a complicated genetic algorithm. In this section, we compare PAPA to ESGD.", "Since ESGD did not release their code as open source, direct comparison to PAPA is difficult."], "text_after_citation": ["From the paper, the experiments were done with CIFAR-10, using the same data processing as our main experiments using a different schedule (160 epochs, learning starts at 0.1 and is decayed by a scaling of 0.1 at 81 epochs and again at 122 epochs).", "The authors do not mention what batch size they used, but having tried 32, 64, and 128, 128 gives us numbers closest to the one from their paper; thus, we use a batch size of 128.", "We do not report soups since we can only compare the results with respect to the mean, min, and max since these were the metrics reported in #OTHEREFR .", "For the PAPA data augmentations and regularizations, we use a random draw from Mixup \u03b1 \u2208 {0, 0.5, 1} and Label smoothing \u03b1 \u2208 {0, 0.05, 0.1}", "The results are shown in the table below."], "citing_paper_content": {"title": "Population Parameter Averaging (Papa)", "abstract": "Ensemble methods combine the predictions of multiple models to improve performance, but they require significantly higher computation costs at inference time. To avoid these costs, multiple neural networks can be combined into one by averaging their weights (model soups). However, this usually performs significantly worse than ensembling. Weight averaging is only beneficial when weights are similar enough (in weight or feature space) to average well but different enough to benefit from combining them. Based on this idea, we propose PopulAtion Parameter Averaging (PAPA): a method that combines the generality of ensembling with the efficiency of weight averaging. PAPA leverages a population of diverse models (trained on different data orders, augmentations, and regularizations) while occasionally (not too often, not too rarely) replacing the weights of the networks with the population average of the weights. PAPA reduces the performance gap between averaging and ensembling, increasing the average accuracy of a population of models by up to 1.1% on CIFAR-10, 2.4% on CIFAR-100, and 1.9% on ImageNet when compared to training independent (non-averaged) models."}, "cited_paper_content": {"title": "Evolutionary Stochastic Gradient Descent For Optimization Of Deep Neural Networks", "abstract": "We propose a population-based Evolutionary Stochastic Gradient Descent (ESGD) framework for optimizing deep neural networks. ESGD combines SGD and gradient-free evolutionary algorithms as complementary algorithms in one framework in which the optimization alternates between the SGD step and evolution step to improve the average fitness of the population. With a back-off strategy in the SGD step and an elitist strategy in the evolution step, it guarantees that the best fitness in the population will never degrade. In addition, various SGD-based optimizers with distinct hyper-parameters are considered in the SGD step as competing species in a coevolution setting such that their complementarity is also taken into account. The effectiveness of ESGD is demonstrated across multiple applications including speech recognition, image recognition and language modeling using networks with a variety of deep architectures."}, "keywords": ["ESGD"], "citation_intent": "method"}
{"citing_id": "2304.06523v1", "cited_id": "1710.10964", "section_title": "Introduction", "citation": "In #REFR Kempa and Prezza also show that computing the smallest k-attractor is NP -complete for any k \u2265 3, by giving a reduction from k-set cover, and extend this proof to non-constant k, especially for k = n.", "text_before_citation": ["For a string of length n, an attractor is a set of positions \u0393 \u2286 [1, n] covering all distinct substrings, that is, every distinct substring has an occurrence crossing at least one of the selected positions.", "If only distinct substrings up to a certain length k need to be covered, we speak of the k-attractor.", "For the string T = abbcabccac, the set of markings \u0393 = {2, 7, 9} is a 2-attractor for T .", "It is not a valid 3-attractor and therefore not an attractor because the substrings bca and cab are not covered.", "Adding either 4 or 5 to \u0393 results in an attractor, and there is no smaller attractor for T ."], "text_after_citation": ["Thus, the problem is solvable in polynomial time for k = 1, and the complexity for k = 2 was left as an open problem.", "Variants of the problem have been introduced, such as the sharp k-attractor in #OTHEREFR , which only considers distinct substrings of length exactly k, and the circular attractor in #OTHEREFR , which also requires to cover circular substrings.", "Interestingly, the sharp k-attractor problem can also be reduced from and to k-set cover and is therefore NP -complete for k \u2265 3.", "However, the sharp 2-attractor problem reduces to 2-set cover, which is equivalent to edge cover and therefore solvable in polynomial time.", "Notably, the sharp variant does not exhibit the same gap as the k-attractor problem."], "citing_paper_content": {"title": "The 2-Attractor Problem Is Np-Complete", "abstract": "A k-attractor is a combinatorial object unifying dictionary-based compression. It allows to compare the repetitiveness measures of different dictionary compressors such as Lempel-Ziv 77, the Burrows-Wheeler transform, straight line programs and macro schemes. For a string T \u2208 \u03a3 n , the k-attractor is defined as a set of positions \u0393 \u2286 [1, n], such that every distinct substring of length at most k is covered by at least one of the selected positions. Thus, if a substring occurs multiple times in T, one position suffices to cover it. A 1-attractor is easily computed in linear time, while Kempa and Prezza [STOC 2018] have shown that for k \u2265 3, it is NP-complete to compute the smallest k-attractor by a reduction from k-set cover. The main result of this paper answers the open question for the complexity of the 2-attractor problem, showing that the problem remains NP-complete. Kempa and Prezza's proof for k \u2265 3 also reduces the 2-attractor problem to the 2-set cover problem, which is equivalent to edge cover, but that does not fully capture the complexity of the 2-attractor problem. For this reason, we extend edge cover by a color function on the edges, yielding the colorful edge cover problem. Any edge cover must then satisfy the additional constraint that each color is represented. This extension raises the complexity such that colorful edge cover becomes NP-complete while also more precisely modeling the 2-attractor problem. We obtain a reduction showing k-attractor to be NP-complete for any k \u2265 2."}, "cited_paper_content": {"title": "At The Roots Of Dictionary Compression: String Attractors", "abstract": "A well-known fact in the field of lossless text compression is that high-order entropy is a weak model when the input contains long repetitions. Motivated by this fact, decades of research have generated myriads of so-called dictionary compressors: algorithms able to reduce the text\u2019s size by exploiting its repetitiveness. Lempel-Ziv 77 is one of the most successful and well-known tools of this kind, followed by straight-line programs, run-length Burrows-Wheeler transform, macro schemes, collage systems, and the compact directed acyclic word graph. In this paper, we show that these techniques are different solutions to the same, elegant, combinatorial problem: to find a small set of positions capturing all distinct text\u2019s substrings. We call such a set a string attractor. We first show reductions between dictionary compressors and string attractors. This gives the approximation ratios of dictionary compressors with respect to the smallest string attractor and allows us to uncover new asymptotic relations between the output sizes of different dictionary compressors. We then show that the k-attractor problem \u2014 deciding whether a text has a size-t set of positions capturing all substrings of length at most k \u2014 is NP-complete for k\u2265 3. This, in particular, includes the full string attractor problem. We provide several approximation techniques for the smallest k-attractor, show that the problem is APX-complete for constant k, and give strong inapproximability results. To conclude, we provide matching lower and upper bounds for the random access problem on string attractors. The upper bound is proved by showing a data structure supporting queries in optimal time. Our data structure is universal: by our reductions to string attractors, it supports random access on any dictionary-compression scheme. In particular, it matches the lower bound also on LZ77, straight-line programs, collage systems, and macro schemes, and therefore essentially closes (at once) the random access problem for all these compressors."}, "keywords": ["k-set cover"], "citation_intent": "background"}
{"citing_id": "2303.08665v1", "cited_id": "1512.03385", "section_title": "Introduction", "citation": "It is adapted from ResNet #REFR by replacing the pooling and stride-convolution layers with a lowpass filter based on Discrete Wavelet Transform (DWT).", "text_before_citation": ["Authors in #OTHEREFR developed efficient low-resolution face recognition models with very low computational cost by distilling the most informative facial features from teacher to student stream.", "More recently, #OTHEREFR In this paper, a cross-resolution knowledge distillation framework is first employed, where the targeted student network is trained with multi-scale LR data and optimized with both face recognition and distillation losses.", "Despite the guidance of the prior knowledge extracted from HR face images, the huge resolution disparity between HR and LR images makes it difficult for the student network to capture informative features.", "Frequency analysis in #OTHEREFR has shown that the high-frequency information in a facial image, such as edge and noise, is eliminated during the resolution reduction, while the low-frequency subbands still preserve the most discriminative features.", "This work proposes an identity-preserving network, WaveResNet, to capture the discriminative information stored in low-frequency components of the LR images."], "text_after_citation": ["The high-frequency subbands of the intermediate feature maps are filtered out to remove ambiguous and noisy information and enforce the network to focus on the more discriminative lowfrequency information.", "In addition, a wavelet similarity loss is designed as an auxiliary distillation loss in order to further enhance attention in low-frequency subbands.", "Moreover, a degradation model is designed to simulate real-world LR training data and develop a more robust recognition system.", "The proposed method has been evaluated on four datasets in a variety of resolutions and it outperforms the baseline model and some other related work."], "citing_paper_content": {"title": "Identity-Preserving Knowledge Distillation For Low-Resolution Face Recognition", "abstract": "Low-resolution face recognition (LRFR) has become a challenging problem for modern deep face recognition systems. Existing methods mainly leverage prior information from high-resolution (HR) images by either reconstructing facial details with super-resolution techniques or learning a unified feature space. To address this issue, this paper proposes a novel approach which enforces the network to focus on the discriminative information stored in the lowfrequency components of a low-resolution (LR) image. A cross-resolution knowledge distillation paradigm is first employed as the learning framework. An identity-preserving network, WaveResNet, and a wavelet similarity loss are then designed to capture low-frequency details and boost performance. Finally, an image degradation model is conceived to simulate more realistic LR training data. Consequently, extensive experimental results show that the proposed method consistently outperforms the baseline model and other stateof-the-art methods across a variety of image resolutions."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["Discrete Wavelet Transform", "stride-convolution layers"], "citation_intent": "method"}
{"citing_id": "2303.08011v1", "cited_id": "1605.07678", "section_title": "I. Introduction", "citation": "Our long-term forecasting experiments require \u223c 10 18 floating point operations for training, model selection and hyperparameter tuning, a figure comparable to the scale of other recent large-scale machine learning benchmarks #REFR .", "text_before_citation": ["Traditional methods include standard linear regression, autoregressive moving averages (ARIMA), exponential smoothing, Kalman filtering, Fourier mode extrapolation, boosted random forest models #OTHEREFR , and newly-introduced linear models that account for trends and distribution shift #OTHEREFR .", "Current state-of-the-art models for general time series forecasting are based on deep neural networks: the transformer model #OTHEREFR , long-short-termmemory networks (LSTM), block recurrent neural networks (RNN), temporal convolutional neural networks #OTHEREFR , and neural basis expansion/neural hierarchical interpolation (NBEATS/NHiTS) #OTHEREFR .", "The latter methods generates forecasts hierarchically by aggregating separate forecasts at distinct timescales (NBEATS), and can explicitly coarse-grain the time series to further reduce computational costs (NHiTS).", "We also consider hybrid physics-motivated methods such as neural ordinary differential equations #OTHEREFR , which approximate the continuous-time differential equation underlying time series, as well as echo-state networks and their generalization, nonlinear vector autoregressive models (nVAR), which use a trainable linear recurrent structure with fixed \"reservoir\" of random nonlinearities #OTHEREFR .", "Our main forecasting results are summarized in Figure  2 ."], "text_after_citation": ["For brevity, we report here only results for the top 14 forecasting models in terms of the sMAPE error, and defer the full tabular results and accuracy metrics to the supplementary material and our open-source repository.", "We observe that large, domain-agnostic forecasting methods successfully forecast diverse chaotic systems, with the strongest methods succeeding consistently across diverse systems and forecasting horizons.", "The strong relative performance of machine learning models is highlighted in Fig 2C, where NBEATS successfully forecasts the Mackey-Glass equation for \u223c22 Lyapunov times without losing track of the global phase.", "This and other strongly-performing methods, such as the transformer and LSTM, comprise large models originally designed for generic sequential datasets, and which do not make any assumptions regarding whether observed time series data arises from a dynamical system.", "This observation suggests that more flexible, generic architectures may prove preferable for problems where some physical structure is present (e.g."], "citing_paper_content": {"title": "Large Statistical Learning Models Effectively Forecast Diverse Chaotic Systems", "abstract": "Chaos and unpredictability are traditionally synonymous, yet recent advances in statistical forecasting suggest that large machine learning models can derive unexpected insight from extended observation of complex systems. Here, we study the forecasting of chaos at scale, by performing a large-scale comparison of 24 representative state-of-the-art multivariate forecasting methods on a crowdsourced database of 135 distinct low-dimensional chaotic systems. We find that large, domain-agnostic time series forecasting methods based on artificial neural networks consistently exhibit strong forecasting performance, in some cases producing accurate predictions lasting for dozens of Lyapunov times. Best-in-class results for forecasting chaos are achieved by recently-introduced hierarchical neural basis function models, though even generic transformers and recurrent neural networks perform strongly. However, physics-inspired hybrid methods like neural ordinary equations and reservoir computers contain inductive biases conferring greater data efficiency and lower training times in data-limited settings. We observe consistent correlation across all methods despite their widely-varying architectures, as well as universal structure in how predictions decay over long time intervals. Our results suggest that a key advantage of modern forecasting methods stems not from their architectural details, but rather from their capacity to learn the large-scale structure of chaotic attractors."}, "cited_paper_content": {"title": "An Analysis Of Deep Neural Network Models For Practical Applications", "abstract": "Since the emergence of Deep Neural Networks (DNNs) as a prominent technique in the field of computer vision, the ImageNet classification challenge has played a major role in advancing the state-of-the-art. While accuracy figures have steadily increased, the resource utilisation of winning models has not been properly taken into account. In this work, we present a comprehensive analysis of important metrics in practical applications: accuracy, memory footprint, parameters, operations count, inference time and power consumption. Key findings are: (1) power consumption is independent of batch size and architecture; (2) accuracy and inference time are in a hyperbolic relationship; (3) energy constraint are an upper bound on the maximum achievable accuracy and model complexity; (4) the number of operations is a reliable estimate of the inference time. We believe our analysis provides a compelling set of information that helps design and engineer efficient DNNs."}, "keywords": ["long-term forecasting experiments", "learning benchmarks"], "citation_intent": "result"}
{"citing_id": "2303.13773v1", "cited_id": "1906.01629", "section_title": "Satgnn", "citation": "The message-passing is split into two steps, one for each set of nodes, similar to the approach of #REFR .", "text_before_citation": ["This allows for more complex features to be extracted at each model layer.", "Algorithm 1: Application of an M P operator with K convolutions to update the features of a node v through message-passing.", "Data:", "Target node v, node features h v , neighbors' features H N (v) , edge weights w u,v , \u2200u \u2208 N (v). Result: Updated node features h * v .", "h (0) \u2190 h v for k \u2190 1 to K do h (k) \u2190 U (k) h (k\u22121) , Aggregation (k) M (k) (h u , w u,v ) : u \u2208 N (v) end h * v \u2190 h (K)"], "text_after_citation": ["At each layer, the messages are propagated first from the variable nodes to the constraint nodes and then from the constraint nodes to the variable nodes, exploiting the bipartite nature of the graph. Algorithm 2 describes this process with further detail.", "Finally, the output of model is generated from the last hidden feature vectors of the variable nodes, generating an output in the same shape as the problem's variable vector.", "The hidden features are fed to an MLP with two hidden layers and ReLU activations NN out , which maps each d-dimensional vector into a single output, i.e., Figure 2 shows an overview of the architecture. Data:", "y v = NN out (h (L) v ), \u2200v \u2208 V var .", "Bipartite graph (V var , V con , E), node features H (0) \u2208 R (n+m)\u00d7d , edge weights w e \u2208 R, \u2200e \u2208 E, number of message passing iterations L. Result: Node embeddings H (L) ."], "citing_paper_content": {"title": "A Graph Neural Network Approach To Nanosatellite Task Scheduling: Insights Into Learning Mixed-Integer Models", "abstract": "This study investigates how to schedule nanosatellite tasks more efficiently using Graph Neural Networks (GNN). In the Offline Nanosatellite Task Scheduling (ONTS) problem, the goal is to find the optimal schedule for tasks to be carried out in orbit while taking into account Quality-of-Service (QoS) considerations such as priority, minimum and maximum activation events, execution time-frames, periods, and execution windows, as well as constraints on the satellite's power resources and the complexity of energy harvesting and management. The ONTS problem has been approached using conventional mathematical formulations and precise methods, but their applicability to challenging cases of the problem is limited. This study examines the use of GNNs in this context, which has been effectively applied to many optimization problems, including traveling salesman problems, scheduling problems, and facility placement problems. Here, we fully represent MILP instances of the ONTS problem in bipartite graphs. We apply a feature aggregation and message-passing methodology allied to a ReLU activation function to learn using a classic deep learning model, obtaining an optimal set of parameters. Furthermore, we apply Explainable AI (XAI), another emerging field of research, to determine which features-nodes, constraints-had the most significant impact on learning performance, shedding light on the inner workings and decision process of such models. We also explored an early fixing approach, obtaining an accuracy above 80% both in predicting the feasibility of a solution and the probability of a decision variable value being in the optimal solution. Our results point to GNNs as a potentially effective method for scheduling nanosatellite tasks and shed light on the advantages of explainable machine learning models for challenging combinatorial optimization problems."}, "cited_paper_content": {"title": "Exact Combinatorial Optimization With Graph Convolutional Neural Networks", "abstract": "Combinatorial optimization problems are typically tackled by the branch-and-bound paradigm. We propose a new graph convolutional neural network model for learning branch-and-bound variable selection policies, which leverages the natural variable-constraint bipartite graph representation of mixed-integer linear programs. We train our model via imitation learning from the strong branching expert rule, and demonstrate on a series of hard problems that our approach produces policies that improve upon state-of-the-art machine-learning methods for branching and generalize to instances significantly larger than seen during training. Moreover, we improve for the first time over expert-designed branching rules implemented in a state-of-the-art solver on large problems. Code for reproducing all the experiments can be found at this https URL."}, "keywords": ["nodes"], "citation_intent": "method"}
{"citing_id": "2303.14265v1", "cited_id": "1705.10528", "section_title": "Iv. Methodology A. Adapting The Safe Set Algorithm", "citation": "Thus we decide to consider the current positions, estimate future positions of all approaching obstacles, and modify the direction of safe control by tuning parameters in the QP problem #REFR .", "text_before_citation": ["Vanilla SSA would output safe control that drives the system to the currently safest direction, which may not be an efficient direction in the long run, see fig. 1a .", "Besides, in multi-obstacle environment, it's not safe to directly add constraints for every dangerous obstacle whose \u03c6 is positive.", "In detail, since vanilla SSA only considers these dangerous obstacles, it may push the robot to the direction that is safe now but risky in the next step if there are unconsidered obstacles (\u03c6 values are negative) in that direction, see fig. 1b ."], "text_after_citation": ["After adapting the projection direction, we expect to generate control signal that is safe to all approaching obstacles, even these \u03c6 values are negative, and efficient for the longer time horizon.", "EQUATION", "In detail, we define the approaching obstacles as those whose distances to the robot are smaller than a threshold value.", "With the current position of the robot (x 0 , y 0 ) and the current positions of approaching obstacles (x i , y i ), i = 1, 2, ..., n, we first predict the next k steps positions of each obstacle (x j i , y j i ), j = 1, 2, ..., k using constant velocity model.", "Then we solve the line l \u03b8 : \u2212 sin(\u03b8)x + cos(\u03b8)y = 0 that maximizes the distance to all approaching obstacles"], "citing_paper_content": {"title": "Safe And Sample-Efficient Reinforcement Learning For Clustered Dynamic Environments", "abstract": "This study proposes a safe and sample-efficient reinforcement learning (RL) framework to address two major challenges in developing applicable RL algorithms: satisfying safety constraints and efficiently learning with limited samples. To guarantee safety in real-world complex environments, we use the safe set algorithm (SSA) to monitor and modify the nominal controls, and evaluate SSA+RL in a clustered dynamic environment which is challenging to be solved by existing RL algorithms. However, the SSA+RL framework is usually not sample-efficient especially in reward-sparse environments, which has not been addressed in previous safe RL works. To improve the learning efficiency, we propose three techniques: (1) avoiding behaving overly conservative by adapting the SSA; (2) encouraging safe exploration using random network distillation with safety constraints; (3) improving policy convergence by treating SSA as expert demonstrations and directly learn from that. The experimental results show that our framework can achieve better safety performance compare to other safe RL methods during training and solve the task with substantially fewer episodes. Project website: https://hychennaza.github.io/projects/Safe RL/."}, "cited_paper_content": {"title": "Constrained Policy Optimization", "abstract": "For many applications of reinforcement learning it can be more convenient to specify both a reward function and constraints, rather than trying to design behavior through the reward function. For example, systems that physically interact with or around humans should satisfy safety constraints. Recent advances in policy search algorithms (Mnih et al., 2016, Schulman et al., 2015, Lillicrap et al., 2016, Levine et al., 2016) have enabled new capabilities in high-dimensional control, but do not consider the constrained setting. ::: We propose Constrained Policy Optimization (CPO), the first general-purpose policy search algorithm for constrained reinforcement learning with guarantees for near-constraint satisfaction at each iteration. Our method allows us to train neural network policies for high-dimensional control while making guarantees about policy behavior all throughout training. Our guarantees are based on a new theoretical result, which is of independent interest: we prove a bound relating the expected returns of two policies to an average divergence between them. We demonstrate the effectiveness of our approach on simulated robot locomotion tasks where the agent must satisfy constraints motivated by safety."}, "keywords": ["safe control"], "citation_intent": "method"}
{"citing_id": "2303.11191v1", "cited_id": "1911.06854", "section_title": "Hyper-Parameter Selection", "citation": "In #REFR , a study is performed evaluating thirty-three different OPE methods by measuring the distance between the estimated policy value and the true policy value.", "text_before_citation": ["Generally, it is difficult to perform policy rollouts in real-world settings, due to costs or safety concerns.", "This is especially true for evaluating policies that are still being trained.", "Performing evaluation before the training of the policy to determine the ideal value of hyper-parameters, would allow for changing the values of hyper-parameters and safe time and resources, instead of training a complete policy.", "Hence, this is another reason restricting demonstration learning to simulators.", "Off-policy methods aim at evaluating a policy from previous data, without requiring interaction with the environment."], "text_after_citation": ["The study concluded that evaluation with the state-action function outperforms the remaining methods.", "To help accelerate the development of OPE methods, #OTHEREFR propose the Deep Off-Policy Evaluation (DOPE) benchmark.", "The available options are using inaccurate OPE methods or training the model for a fixed number of steps and adjusting the values based on the early results.", "The challenge of the selection of hyper-parameter values is common to other machine-learning paradigms.", "However, demonstration learning needs to overcome this problem if it truly wishes to reduce the requirement of expert programming."], "citing_paper_content": {"title": "A Survey Of Demonstration Learning", "abstract": "With the fast improvement of machine learning, reinforcement learning (RL) has been used to automate human tasks in different areas. However, training such agents is difficult and restricted to expert users. Moreover, it is mostly limited to simulation environments due to the high cost and safety concerns of interactions in the real world. Demonstration Learning is a paradigm in which an agent learns to perform a task by imitating the behavior of an expert shown in demonstrations. It is a relatively recent area in machine learning, but it is gaining significant traction due to having tremendous potential for learning complex behaviors from demonstrations. Learning from demonstration accelerates the learning process by improving sample efficiency, while also reducing the effort of the programmer. Due to learning without interacting with the environment, demonstration learning would allow the automation of a wide range of real world applications such as robotics and healthcare. This paper provides a survey of demonstration learning, where we formally introduce the demonstration problem along with its main challenges and provide a comprehensive overview of the process of learning from demonstrations from the creation of the demonstration data set, to learning methods from demonstrations, and optimization by combining demonstration learning with different machine learning methods. We also review the existing benchmarks and identify their strengths and limitations. Additionally, we discuss the advantages and disadvantages of the paradigm as well as its main applications. Lastly, we discuss our perspective on open problems and research directions for this rapidly growing field."}, "cited_paper_content": {"title": "Empirical Study Of Off-Policy Policy Evaluation For Reinforcement Learning", "abstract": "Off-policy policy evaluation (OPE) is the problem of estimating the online performance of a policy using only pre-collected historical data generated by another policy. Given the increasing interest in deploying learning-based methods for safety-critical applications, many recent OPE methods have recently been proposed. Due to disparate experimental conditions from recent literature, the relative performance of current OPE methods is not well understood. In this work, we present the first comprehensive empirical analysis of a broad suite of OPE methods. Based on thousands of experiments and detailed empirical analyses, we offer a summarized set of guidelines for effectively using OPE in practice, and suggest directions for future research."}, "keywords": ["estimated policy value"], "citation_intent": "method"}
{"citing_id": "2303.02660v1", "cited_id": "1907.04047", "section_title": "Base Presentation Attack Detectors", "citation": "PixBis #REFR employs a binary supervisory strategy at pixel-level to simplify the problem and obviate the need for a computationally intensive synthesis of depth maps.", "text_before_citation": ["ResNet #OTHEREFR is one of the most popular backbone architectures used in face PAD algorithm design #OTHEREFR .", "We report the results of a model trained from scratch based on the ResNet-18 model architecture.", "A cross-entropy loss function is used in the training phase and formulated as follows:", "EQUATION", "where y is the ground truth (1 for bona fide and 0 for attack in our case) and p is the predicted score."], "text_after_citation": ["Two dense blocks of DenseNet121 #OTHEREFR are utilized as the model backbone and a combination of two binary cross-entropy loss functions is used to train the model for both pixel-wise and binary output.", "The combined loss equation for the training of all models is formed as:", "EQUATION", "where L pixel\u2212wise CE refers to the loss based on the pixel-wise output and L binary CE refers to the loss based on the binary output."], "citing_paper_content": {"title": "Synthaspoof: Developing Face Presentation Attack Detection Based On Privacy-Friendly Synthetic Data", "abstract": "Recently, significant progress has been made in face presentation attack detection (PAD), which aims to secure face recognition systems against presentation attacks, owing to the availability of several face PAD datasets. However, all available datasets are based on privacy and legallysensitive authentic biometric data with a limited number of subjects. To target these legal and technical challenges, this work presents the first synthetic-based face PAD dataset, named SynthASpoof, as a large-scale PAD development dataset. The bona fide samples in SynthASpoof are synthetically generated and the attack samples are collected by presenting such synthetic data to capture systems in a real attack scenario. The experimental results demonstrate the feasibility of using SynthASpoof for the development of face PAD. Moreover, we boost the performance of such a solution by incorporating the domain generalization tool MixStyle into the PAD solutions. Additionally, we showed the viability of using synthetic data as a supplement to enrich the diversity of limited authentic training data and consistently enhance PAD performances. The SynthASpoof dataset, containing 25,000 bona fide and 78,800 attack samples, the implementation, and the pre-trained weights are made publicly available 1 ."}, "cited_paper_content": {"title": "Deep Pixel-Wise Binary Supervision For Face Presentation Attack Detection", "abstract": "Face recognition has evolved as a prominent biometric authentication modality. However, vulnerability to presentation attacks curtails its reliable deployment. Automatic detection of presentation attacks is essential for secure use of face recognition technology in unattended scenarios. In this work, we introduce a Convolutional Neural Network (CNN) based framework for presentation attack detection, with deep pixel-wise supervision. The framework uses only frame level information making it suitable for deployment in smart devices with minimal computational and time overhead. We demonstrate the effectiveness of the proposed approach in public datasets for both intra as well as cross-dataset experiments. The proposed approach achieves an HTER of 0% in Replay Mobile dataset and an ACER of 0.42% in Protocol-1 of OULU dataset outperforming state of the art methods."}, "keywords": ["computationally intensive synthesis", "pixel-level"], "citation_intent": "method"}
{"citing_id": "2303.00521v2", "cited_id": "1512.03385", "section_title": "Implementation Details", "citation": "Specifically, we use a ResNet50 #REFR with an extra two-layer MLP as the encoder and train it on ImageNet.", "text_before_citation": ["Our experiments are performed using PyTorch #OTHEREFR , and are all conducted on 8 Nvidia V100 GPUs.", "Following the classical pre-training and fine-tuning paradigm, we run the whole experiments in two stages:", "Pre-training.", "We inherit most settings from MoCoV2 #OTHEREFR , while modify the pretext task and degradation process for the quality-aware purpose."], "text_after_citation": ["We use SGD optimizer with weight decay of 0.0001, momentum of 0.9 and a mini-batch size of 256.", "The initial learning rate is 0.03 and multiplied by 0.1 at 120 and 160 epoch. The hyper-parameters, \u03b2 and \u03c4 in Eq. 4 Fine-tuning.", "Following the scheme from #OTHEREFR , we resize the smaller edge of images to 340 while maintaining aspect ratio, and randomly sample and horizontally flip patches with size 320x320 pixels.", "We use AdamW optimizer with weight decay of 0.01 and mini-batch size of 64.", "The Learning rate is initialized with 0.0002, and decayed by the cosine annealing strategy."], "citing_paper_content": {"title": "Quality-Aware Pre-Trained Models For Blind Image Quality Assessment", "abstract": "Blind image quality assessment (BIQA) aims to automatically evaluate the perceived quality of a single image, whose performance has been improved by deep learningbased methods in recent years. However, the paucity of labeled data somewhat restrains deep learning-based BIQA methods from unleashing their full potential. In this paper, we propose to solve the problem by a pretext task customized for BIQA in a self-supervised learning manner, which enables learning representations from orders of magnitude more data. To constrain the learning process, we propose a quality-aware contrastive loss based on a simple assumption: the quality of patches from a distorted image should be similar, but vary from patches from the same image with different degradations and patches from different images. Further, we improve the existing degradation process and form a degradation space with the size of roughly 2 \u00d7 10 7. After pre-trained on ImageNet using our method, models are more sensitive to image quality and perform significantly better on downstream BIQA tasks. Experimental results show that our method obtains remarkable improvements on popular BIQA datasets."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["ImageNet"], "citation_intent": "method"}
{"citing_id": "2303.00235v1", "cited_id": "1603.00762", "section_title": "Self-Dual Consta-Dihedral Codes", "citation": "If q \u2261 1 (mod 4), then the self-dual quasi-cyclic codes of index 2 over F and the LCD quasi-cyclic codes of index 2 over F of rate #REFR 2 are both asymptotically good.", "text_before_citation": ["Assume that q is even or 4 | (q \u2212 1) (i.e. q \u2261 3 (mod 4)).", "Then the self-dual consta-dihedral codes over F are asymptotically good.", "In particular, self-dual quasi-cyclic codes of index 2 over F are asymptotically good.", "It is known that if q is odd then LCD dihedral codes of rate 1 2 are asymptotically good, see #OTHEREFR Theorem 1.2] . We get the following consequence.", "Corollary 5.13."], "text_after_citation": [], "citing_paper_content": {"title": "Consta-Dihedral Codes Over Finite Fields", "abstract": "It is proved in a reference (Fan, Lin, IEEE TIT, vol.67, pp.5016-5025) that the self-dual (LCD respectively) dihedral codes over a finite field F with |F | = q are asymptotically good if q is even (odd respectively). In this paper, we investigate the algebraic property and the asymptotic property of conta-dihedral codes over F , and show that: if q is even or 4 | (q \u2212 1), then the self-dual consta-dihedral codes are asymptotically good; otherwise, the LCD consta-dihedral codes are asymptotically good. And, with the help of a technique developed in this paper, some errors in the reference mentioned above are corrected."}, "cited_paper_content": {"title": "On Self-Dual Double Circulant Codes", "abstract": "Self-dual double circulant codes of odd dimension are shown to be dihedral in even characteristic and consta-dihedral in odd characteristic. Exact counting formulae are derived for them, generalizing some old results of MacWilliams on the enumeration of circulant orthogonal matrices. These formulae, in turn, are instrumental in deriving a Varshamov\u2013Gilbert bound on the relative minimum distance of this family of codes."}, "keywords": ["self-dual quasi-cyclic codes"], "citation_intent": "background"}
{"citing_id": "2304.13689v1", "cited_id": "1804.00320", "section_title": "Related Work", "citation": "Models employing the pipeline architecture have highlighted the effects of errors arising from the ASR component #REFR .", "text_before_citation": ["Spoken Question Answering Spoken Question Answering (SQA) #OTHEREFR involves a set of approaches aimed at appropriately answering a spoken question.", "There are two types of SQA systems: in context SQA systems #OTHEREFR and open domain SQA systems.", "While open domain SQA systems are built with the aim to answer questions without context, in context SQA systems typically take the form of Machine Reading Comprehension (MRC) #OTHEREFR where the question is accompanied by a passage."], "text_after_citation": ["#OTHEREFR explored the use of sub-word strategy while #OTHEREFR utilized BERT based model to learn joint audio-text features with the aim of mitigating the errors from ASR.", "However, such techniques do not explicitly consider the mapping between ASR transcript and the corresponding human transcription.", "In our work, we present experiments with models explicitly aimed at exploiting this mapping as well, by attempting to directly learn the said mapping in a translation fashion.", "In #OTHEREFR authors present a method to learn representations using contrastive learning and show improved performance on three SQA datasets.", "While the augmentation strategies presented may implicitly contribute to improvements in representations learned, it is not clear how they explicitly improve the performance of an SQA system."], "citing_paper_content": {"title": "Heysquad: A Spoken Question Answering Dataset", "abstract": "Human-spoken questions are critical to evaluating the performance of spoken question answering (SQA) systems that serve several realworld use cases including digital assistants. We present a new large-scale communityshared SQA dataset, HeySQuAD that consists of 76k human-spoken questions and 97k machine-generated questions and corresponding textual answers derived from the SQuAD QA dataset. The goal of HeySQuAD is to measure the ability of machines to understand noisy spoken questions and answer the questions accurately. To this end, we run extensive benchmarks on the human-spoken and machine-generated questions to quantify the differences in noise from both sources and its subsequent impact on the model and answering accuracy. Importantly, for the task of SQA, where we want to answer human-spoken questions, we observe that training using the transcribed human-spoken and original SQuAD questions leads to significant improvements (12.51%) over training using only the original SQuAD textual questions."}, "cited_paper_content": {"title": "Spoken Squad: A Study Of Mitigating The Impact Of Speech Recognition Errors On Listening Comprehension", "abstract": "Reading comprehension has been widely studied. One of the most representative reading comprehension tasks is Stanford Question Answering Dataset (SQuAD), on which machine is already comparable with human. On the other hand, accessing large collections of multimedia or spoken content is much more difficult and time-consuming than plain text content for humans. It's therefore highly attractive to develop machines which can automatically understand spoken content. In this paper, we propose a new listening comprehension task - Spoken SQuAD. On the new task, we found that speech recognition errors have catastrophic impact on machine comprehension, and several approaches are proposed to mitigate the impact."}, "keywords": ["errors"], "citation_intent": "background"}
{"citing_id": "2303.01363v1", "cited_id": "1906.05909", "section_title": "Spatial Attention", "citation": "In this work, we rather consider the use of local self-attention layers, and more specifically the stand-alone self-attention layers proposed by #REFR .", "text_before_citation": ["Several strategies have been proposed, including training a subnetwork to identify the important regions #OTHEREFR , or increasing the receptive field of CNNs.", "The methods based on the latest strategy use self-attention mechanisms, which were introduced in computer vision tasks by #OTHEREFR .", "They lead to impressive results compared to the performance achieved so far using CNNs, especially when it comes to the use of Vision Transformers (ViT) for various visual tasks #OTHEREFR .", "However, this process is computationally very expensive, and it also requires a lot of training data.", "In addition, for small object detection, the spatial dependencies are mainly local."], "text_after_citation": [], "citing_paper_content": {"title": "Deep-Nfa: A Deep A Contrario Framework For Small Object Detection", "abstract": "The detection of small objects is a challenging task in computer vision. Conventional object detection methods have difficulty in finding the balance between high detection and low false alarm rates. In the literature, some methods have addressed this issue by enhancing the feature map responses, but without guaranteeing robustness with respect to the number of false alarms induced by background elements. To tackle this problem, we introduce an a contrario decision criterion into the learning process to take into account the unexpectedness of small objects. This statistic criterion enhances the feature map responses while controlling the number of false alarms (NFA) and can be integrated into any semantic segmentation neural network. Our add-on NFA module not only allows us to obtain competitive results for small target and crack detection tasks respectively, but also leads to more robust and interpretable results."}, "cited_paper_content": {"title": "Stand-Alone Self-Attention In Vision Models", "abstract": "Convolutions are a fundamental building block of modern computer vision systems. Recent approaches have argued for going beyond convolutions in order to capture long-range dependencies. These efforts focus on augmenting convolutional models with content-based interactions, such as self-attention and non-local means, to achieve gains on a number of vision tasks. The natural question that arises is whether attention can be a stand-alone primitive for vision models instead of serving as just an augmentation on top of convolutions. In developing and testing a pure self-attention vision model, we verify that self-attention can indeed be an effective stand-alone layer. A simple procedure of replacing all instances of spatial convolutions with a form of self-attention to ResNet-50 produces a fully self-attentional model that outperforms the baseline on ImageNet classification with 12% fewer FLOPS and 29% fewer parameters. On COCO object detection, a fully self-attention model matches the mAP of a baseline RetinaNet while having 39% fewer FLOPS and 34% fewer parameters. Detailed ablation studies demonstrate that self-attention is especially impactful when used in later layers. These results establish that stand-alone self-attention is an important addition to the vision practitioner's toolbox."}, "keywords": ["local self-attention layers", "stand-alone self-attention layers"], "citation_intent": "method"}
{"citing_id": "2303.13778v1", "cited_id": "1903.03283", "section_title": "Iv. Insufficiently Separated Markov Chains", "citation": "This Lemma 3 bound seems strikingly similar to the bounding result of Lemma 2 in #REFR with relative entropy rate between chains replacing the role of relative entropy between measurement densities in that result.", "text_before_citation": ["EQUATION", "For sufficiently large k, we can now expand the second term of (5) as", "EQUATION", "where the first line follows from the definition of expectation operation, the 2nd line follows from application of (6), the second last line follows from the definition of relative entropy rate between two chains and the last line follows becaus\u00ea M b k\u22121 \u2264 1 and relative entropy rates such as R(A a |A b ) are non-negative.", "The Lemma statement then follows from (5) and (7) by noting that for any \u03b4 < 0 there is a h \u03b4 > 0 such that for allM k\u22121 < h \u03b4 we have E \u03c0 \u03b3 k M b k\u22121 < \u03b4 as log is monotonically increasing and the elements A i,j b and A i,j a are bounded in size (no greater than 1)."], "text_after_citation": ["The lemma's assumptions on p(X 0 ) being the stationary distribution of A b and the requirement for sufficiently largely k have been included to simplify analysis rather than being fundamental to the bounding mechanism.", "We would expect similar bounded behaviour under relaxation of these two assumptions.", "The lemma's assumption of the existence of stationary distributions is ensured under this paper's standing assumption of aperiodic and irreducible chains #OTHEREFR Ch. 4] .", "Whilst such ergodic chains are a large and useful class, the lemma does exclude non-ergodic chains such as those with transient states, absorbing states or chains that exhibit periodic behaviours.", "We now investigate a phenomenon that occurs when the Markov chains A b and A a are too close in the sense of having small relative entropy rate R(A b |A a ) and are unable to overcome the change event's geometric prior information."], "citing_paper_content": {"title": "Exactly Optimal Quickest Change Detection Of Markov Chains", "abstract": "This paper establishes that an exactly optimal rule for Bayesian Quickest Change Detection (QCD) of Markov chains is a threshold test on the no change posterior. We also provide a computationally efficient scalar filter for the no change posterior whose effort is independent of the dimension of the chains. We establish that an (undesirable) weak practical super-martingale phenomenon can be exhibited by the no change posterior when the before and after chains are too close in a relative entropy rate sense. The proposed detector is examined in simulation studies."}, "cited_paper_content": {"title": "On The Informativeness Of Measurements In Shiryaev'S Bayesian Quickest Change Detection", "abstract": "This paper provides the first description of a weak practical super-martingale phenomenon that can emerge in the test statistic in Shiryaev's Bayesian quickest change detection (QCD) problem. We establish that this super-martingale phenomenon can emerge under a condition on the relative entropy between pre and post change densities when the measurements are insufficiently informative to overcome the change time's geometric prior. We illustrate this super-martingale phenomenon in a simple Bayesian QCD problem which highlights the unsuitability of Shiryaev's test statistic for detecting subtle change events."}, "keywords": ["relative entropy rate"], "citation_intent": "result"}
{"citing_id": "2303.07811v1", "cited_id": "1806.10574", "section_title": "Icicle", "citation": "Random initialization of prototypes, proposed in #REFR , fails in the incremental learning (see Table 5 ).", "text_before_citation": ["L IR = H i=0 W j=0 |sim(p t\u22121 , z t i,j ) \u2212 sim(p t , z t i,j )| \u2022 S i,j (1)", "where", "sim(p t\u22121 , z t i,j )", "is computed for the model stored before training task t, and S is a binary mask of size H \u00d7 W , indicating the representation pixels with the highest similarity (\u03b3 quantile of those pixels).", "Such similarity distillation gives higher plasticity when learning a new task but, at the same time, reduces the interpretability drift. Proximity-based prototype initialization."], "text_after_citation": ["Most probably because new prototypes can be created far from the old ones, which are only slightly different (e.g. wing prototypes of various bird species).", "Therefore, we introduce initialization that sets new prototypes close to the old ones (see Figure 4) .", "We start by passing training samples of task t We introduce a new proximity-based prototype initialization.", "It starts by passing training samples of task t through the network (green dots) and choosing representations closest to existing prototypes (violet diamonds).", "This results in many points, which we cluster (purple circles) to obtain the initial locations of task t prototypical parts (yellow diamonds)."], "citing_paper_content": {"title": "Icicle: Interpretable Class Incremental Continual Learning", "abstract": "Continual learning enables incremental learning of new tasks without forgetting those previously learned, resulting in positive knowledge transfer that can enhance performance on both new and old tasks. However, continual learning poses new challenges for interpretability, as the rationale behind model predictions may change over time, leading to interpretability concept drift. We address this problem by proposing Interpretable Class-InCremental LEarning (ICICLE), an exemplar-free approach that adopts a prototypical part-based approach. It consists of three crucial novelties: interpretability regularization that distills previously learned concepts while preserving user-friendly positive reasoning; proximity-based prototype initialization strategy dedicated to the fine-grained setting; and taskrecency bias compensation devoted to prototypical parts. Our experimental results demonstrate that ICICLE reduces the interpretability concept drift and outperforms the existing exemplar-free methods of common class-incremental learning when applied to concept-based models. We make the code available."}, "cited_paper_content": {"title": "This Looks Like That: Deep Learning For Interpretable Image Recognition", "abstract": "When we are faced with challenging image classification tasks, we often explain our reasoning by dissecting the image, and pointing out prototypical aspects of one class or another. The mounting evidence for each of the classes helps us make our final decision. In this work, we introduce a deep network architecture -- prototypical part network (ProtoPNet), that reasons in a similar way: the network dissects the image by finding prototypical parts, and combines evidence from the prototypes to make a final classification. The model thus reasons in a way that is qualitatively similar to the way ornithologists, physicians, and others would explain to people on how to solve challenging image classification tasks. The network uses only image-level labels for training without any annotations for parts of images. We demonstrate our method on the CUB-200-2011 dataset and the Stanford Cars dataset. Our experiments show that ProtoPNet can achieve comparable accuracy with its analogous non-interpretable counterpart, and when several ProtoPNets are combined into a larger network, it can achieve an accuracy that is on par with some of the best-performing deep models. Moreover, ProtoPNet provides a level of interpretability that is absent in other interpretable deep models."}, "keywords": ["incremental learning"], "citation_intent": "background"}
{"citing_id": "2304.01143v1", "cited_id": "1710.05381", "section_title": "Properties Of Long-Tail Benchmarks", "citation": "In #REFR , experiments were performed with class counts that decay linearly or decay with a step-function.", "text_before_citation": ["Established benchmarks for long-tail image recognition #OTHEREFR have shaped the progress of long-tail methods.", "These followed earlier efforts that investigated the desired data distribution characteristics for long-tail benchmarks."], "text_after_citation": ["They noted that a larger imbalance between majority (now known as 'head') and minority (i.e.", "'tail') classes increases difficulty and that a longer tail negatively affects classifier performance for both linear and step class count decays.", "Interestingly, imbalance was shown to affect higher complexity tasks (e.g. CIFAR) significantly more than lower complexity tasks (e.g. MNIST).", "Step and exponential class count decays were also investigated in #OTHEREFR , with similar conclusions.", "In #OTHEREFR , multiple long-tail versions of CIFAR #OTHEREFR were curated by changing the minimum class size."], "citing_paper_content": {"title": "Use Your Head: Improving Long-Tail Video Recognition", "abstract": "This paper presents an investigation into long-tail video recognition. We demonstrate that, unlike naturallycollected video datasets and existing long-tail image benchmarks, current video benchmarks fall short on multiple long-tailed properties. Most critically, they lack few-shot classes in their tails. In response, we propose new video benchmarks that better assess long-tail recognition, by sampling subsets from two datasets: SSv2 and VideoLT. We then propose a method, Long-Tail Mixed Reconstruction (LMR), which reduces overfitting to instances from few-shot classes by reconstructing them as weighted combinations of samples from head classes. LMR then employs label mixing to learn robust decision boundaries. It achieves state-of-the-art average class accuracy on EPIC-KITCHENS and the proposed SSv2-LT and VideoLT-LT."}, "cited_paper_content": {"title": "A Systematic Study Of The Class Imbalance Problem In Convolutional Neural Networks", "abstract": "In this study, we systematically investigate the impact of class imbalance on classification performance of convolutional neural networks (CNNs) and compare frequently used methods to address the issue. Class imbalance is a common problem that has been comprehensively studied in classical machine learning, yet very limited systematic research is available in the context of deep learning. In our study, we use three benchmark datasets of increasing complexity, MNIST, CIFAR-10 and ImageNet, to investigate the effects of imbalance on classification and perform an extensive comparison of several methods to address the issue: oversampling, undersampling, two-phase training, and thresholding that compensates for prior class probabilities. Our main evaluation metric is area under the receiver operating characteristic curve (ROC AUC) adjusted to multi-class tasks since overall accuracy metric is associated with notable difficulties in the context of imbalanced data. Based on results from our experiments we conclude that (i) the effect of class imbalance on classification performance is detrimental; (ii) the method of addressing class imbalance that emerged as dominant in almost all analyzed scenarios was oversampling; (iii) oversampling should be applied to the level that completely eliminates the imbalance, whereas the optimal undersampling ratio depends on the extent of imbalance; (iv) as opposed to some classical machine learning models, oversampling does not cause overfitting of CNNs; (v) thresholding should be applied to compensate for prior class probabilities when overall number of properly classified cases is of interest."}, "keywords": ["class counts"], "citation_intent": "method"}
{"citing_id": "2303.14011v1", "cited_id": "2002.02095", "section_title": "C. Meta-Datasets", "citation": "Next, we conduct experiments to investigate the relation between performance and source corpora choice #REFR . According to the results, we have the following observations.", "text_before_citation": ["Meanwhile, each source corpus should possess as many identities to the target corpus as possible.", "This idea could be achieved with properly designed similarity criteria, where the performance of models should show a monotonous change with corpora chosen along the similarity rankings.", "Based on this idea, we consider the following hypotheses that may help in source corpora choice.", "1) Semantics: the document embedding similarity, 2) Word overlapping: token cosine similarity, 3) Information coverage: ROUGE recall, 4) Information density: ROUGE precision, 5) Length: absolute difference in token length between articles.", "Given a target corpus, we rank the source corpora and create meta-datasets in different ranking sections for each criterion."], "text_after_citation": ["First, the statistical similarity is more important than semantic similarity.", "Second, corpora from the same domain may not always provide better knowledge.", "Third, rather than choosing source corpora that can cover all used words in the target corpus, source corpora should have a high information density.", "Fourth, source corpora with similar average article lengths are better than source corpora with different average article lengths.", "With the above observations, we use the average ranking of the following criteria: 1) cosine similarity, 2) ROUGE precision, and 3) article length to choose our source corpora for the meta-dataset."], "citing_paper_content": {"title": "Spec: Summary Preference Decomposition For Low-Resource Abstractive Summarization", "abstract": "Neural abstractive summarization has been widely studied and achieved great success with large-scale corpora. However, the considerable cost of annotating data motivates the need for learning strategies under low-resource settings. In this paper, we investigate the problems of learning summarizers with only few examples and propose corresponding methods for improvements. First, typical transfer learning methods are prone to be affected by data properties and learning objectives in the pretext tasks. Therefore, based on pretrained language models, we further present a meta learning framework to transfer few-shot learning processes from source corpora to the target corpus. Second, previous methods learn from training examples without decomposing the content and preference. The generated summaries could therefore be constrained by the preference bias in the training set, especially under low-resource settings. As such, we propose decomposing the contents and preferences during learning through the parameter modulation, which enables control over preferences during inference. Third, given a target application, specifying required preferences could be non-trivial because the preferences may be difficult to derive through observations. Therefore, we propose a novel decoding method to automatically estimate suitable preferences and generate corresponding summary candidates from the few training examples. Extensive experiments demonstrate that our methods achieve state-of-the-art performance on six diverse corpora with 30.11%/33.95%/27.51% and 26.74%/31.14%/24.48% average improvements on ROUGE-1/2/L under 10-and 100example settings."}, "cited_paper_content": {"title": "Attractive Or Faithful? Popularity-Reinforced Learning For Inspired Headline Generation", "abstract": "With the rapid proliferation of online media sources and published news, headlines have become increasingly important for attracting readers to news articles, since users may be overwhelmed with the massive information. In this paper, we generate inspired headlines that preserve the nature of news articles and catch the eye of the reader simultaneously. The task of inspired headline generation can be viewed as a specific form of Headline Generation (HG) task, with the emphasis on creating an attractive headline from a given news article. To generate inspired headlines, we propose a novel framework called POpularity-Reinforced Learning for inspired Headline Generation (PORL-HG). PORL-HG exploits the extractive-abstractive architecture with 1) Popular Topic Attention (PTA) for guiding the extractor to select the attractive sentence from the article and 2) a popularity predictor for guiding the abstractor to rewrite the attractive sentence. Moreover, since the sentence selection of the extractor is not differentiable, techniques of reinforcement learning (RL) are utilized to bridge the gap with rewards obtained from a popularity score predictor. Through quantitative and qualitative experiments, we show that the proposed PORL-HG significantly outperforms the state-of-the-art headline generation models in terms of attractiveness evaluated by both human (71.03%) and the predictor (at least 27.60%), while the faithfulness of PORL-HG is also comparable to the state-of-the-art generation model."}, "keywords": ["source corpora choice"], "citation_intent": "result"}
{"citing_id": "2304.06212v1", "cited_id": "1906.00817", "section_title": "Related Works", "citation": "The purpose of zero-shot semantic segmentation is to enable the model to predict new categories exclusively from auxiliary semantic information without any extra pixel-level annotations of new categories #REFR , further relaxing the annotation requirements for new categories.", "text_before_citation": ["Generalized semantic segmentation.", "The semantic segmentation task is to make pixel-by-pixel category predictions for an image on a fixed category label #OTHEREFR .", "Few-shot semantic segmentation is to relax the restrictions of fixed categories #OTHEREFR .", "By providing several images of new categories and pixel-level annotations (support set), the dense prediction ability of the network can be generalized to new categories.", "However, few-shot semantic segmentation needs to provide at least one annotation of new categories."], "text_after_citation": ["Therefore, the model must establish the correlation between the image and auxiliary semantic information from seen categories.", "The rise of CLIP provides a fresh perspective on the interaction between auxiliary semantic information and image information.", "Prompt Learning.", "Prompt learning was first widely used in NLP #OTHEREFR .", "The initial approach was to design different hard prompts to bridge the gap between the downstream task and the pre-trained model, guiding the model to complete the corresponding downstream task #OTHEREFR ."], "citing_paper_content": {"title": "Cls] Token Is All You Need For Zero-Shot Semantic Segmentation", "abstract": "In this paper, we propose an embarrassingly simple yet highly effective zero-shot semantic segmentation (ZS3) method, based on the pre-trained vision-language model CLIP. First, our study provides a couple of key discoveries: (i) the global tokens (a.k.a [CLS] tokens in Transformer) of the text branch in CLIP provide a powerful representation of semantic information and (ii) these text-side [CLS] tokens can be regarded as category priors to guide CLIP visual encoder pay more attention on the corresponding region of interest. Based on that, we build upon the CLIP model as a backbone which we extend with a One-Way [CLS] token navigation from text to the visual branch that enables zero-shot dense prediction, dubbed ClsCLIP. Specifically, we use the [CLS] token output from the text branch, as an auxiliary semantic prompt, to replace the [CLS] token in shallow layers of the ViT-based visual encoder. This one-way navigation embeds such global category prior earlier and thus promotes semantic segmentation. Furthermore, to better segment tiny objects in ZS3, we further enhance ClsCLIP with a local zoom-in strategy, which employs a region proposal pre-processing and we get ClsCLIP+. Extensive experiments demonstrate that our proposed ZS3 method achieves a SOTA performance, and it is even comparable with those few-shot semantic segmentation methods."}, "cited_paper_content": {"title": "Zero-Shot Semantic Segmentation", "abstract": "Semantic segmentation models are limited in their ability to scale to large numbers of object classes. In this paper, we introduce the new task of zero-shot semantic segmentation: learning pixel-wise classifiers for never-seen object categories with zero training examples. To this end, we present a novel architecture, ZS3Net, combining a deep visual segmentation model with an approach to generate visual representations from semantic word embeddings. By this way, ZS3Net addresses pixel classification tasks where both seen and unseen categories are faced at test time (so called generalized zero-shot classification). Performance is further improved by a self-training step that relies on automatic pseudo-labeling of pixels from unseen classes. On the two standard segmentation datasets, Pascal-VOC and Pascal-Context, we propose zero-shot benchmarks and set competitive baselines. For complex scenes as ones in the Pascal-Context dataset, we extend our approach by using a graph-context encoding to fully leverage spatial context priors coming from class-wise segmentation maps."}, "keywords": ["zero-shot semantic segmentation"], "citation_intent": "background"}
{"citing_id": "2303.13852v1", "cited_id": "1801.01075", "section_title": "A.2. Additional Experiments A.2.1 Evaluation Of Light-Net", "citation": "To evaluate the performance of Light-Net, we randomly sampled a testing set of 200 images from LIME #REFR .", "text_before_citation": [], "text_after_citation": ["It is a synthetic dataset of Bigbird #OTHEREFR and ShapeNet #OTHEREFR objects with ground truth normal, albedo, and shading.", "We use lighting coefficients predicted by Light-Net to render shading with ground truth normal maps.", "By comparing the rendered shading and ground truths, we can evaluate the accuracy of the estimated lighting coefficients. For quantitative"], "citing_paper_content": {"title": "Weakly-Supervised Single-View Image Relighting", "abstract": "We present a learning-based approach to relight a single image of Lambertian and low-frequency specular objects. Our method enables inserting objects from photographs into new scenes and relighting them under the new environment lighting, which is essential for AR applications. To relight the object, we solve both inverse rendering and re-rendering. To resolve the ill-posed inverse rendering, we propose a weakly-supervised method by a low-rank constraint. To facilitate the weakly-supervised training, we contribute Relit, a large-scale (750K images) dataset of videos with aligned objects under changing illuminations. For re-rendering, we propose a differentiable specular rendering layer to render low-frequency non-Lambertian materials under various illuminations of spherical harmonics. The whole pipeline is end-to-end and efficient, allowing for a mobile app implementation of AR object insertion. Extensive evaluations demonstrate that our method achieves state-of-the-art performance. Project"}, "cited_paper_content": {"title": "Lime: Live Intrinsic Material Estimation", "abstract": "We present the first end to end approach for real time material estimation for general object shapes with uniform material that only requires a single color image as input. In addition to Lambertian surface properties, our approach fully automatically computes the specular albedo, material shininess, and a foreground segmentation. We tackle this challenging and ill posed inverse rendering problem using recent advances in image to image translation techniques based on deep convolutional encoder decoder architectures. The underlying core representations of our approach are specular shading, diffuse shading and mirror images, which allow to learn the effective and accurate separation of diffuse and specular albedo. In addition, we propose a novel highly efficient perceptual rendering loss that mimics real world image formation and obtains intermediate results even during run time. The estimation of material parameters at real time frame rates enables exciting mixed reality applications, such as seamless illumination consistent integration of virtual objects into real world scenes, and virtual material cloning. We demonstrate our approach in a live setup, compare it to the state of the art, and demonstrate its effectiveness through quantitative and qualitative evaluation."}, "keywords": ["Light-Net"], "citation_intent": "method"}
{"citing_id": "2303.14501v1", "cited_id": "1706.02216", "section_title": "Message-Passing Layers", "citation": "GraphSAGE's message-passing layer #REFR does not directly aggregate central node features with incoming messages.", "text_before_citation": ["GNNs utilize the concept of message-passing to encode semantically rich features within network-structured data.", "Over time, multiple variations of message-passing layers have been proposed [40, #OTHEREFR .", "For instance, GCN's message-passing layer #OTHEREFR weighs each incoming message with a fixed coefficient, the node degree, before aggregation.", "In contrast, GAT's message-passing layer #OTHEREFR learns aggregation weights dynamically based on attention scores."], "text_after_citation": ["Instead, it distinguishes these two kinds of features and learns two different transformations, one on the central node and another on incoming messages.", "EdgeConv #OTHEREFR aggregates the feature difference between the central node and its neighbors combined with the central node's features.", "Thus, EdgeConv draws parallels to aggregating spatial vectors if the nodes embed spatial positions.", "However, our proposed GAV layer differs significantly from EdgeConv, as our method explicitly constrains the update of vector embeddings to imitate the simplified dynamics of physical flow in spatial networks.", "Importantly, only a few works tried to adapt the message-passing paradigm to spatial networks #OTHEREFR ."], "citing_paper_content": {"title": "Link Prediction For Flow-Driven Spatial Networks", "abstract": "Link prediction algorithms predict the existence of connections between nodes in network-structured data and are typically applied to refine the connectivity among nodes by proposing meaningful new links. In this work, we focus on link prediction for flow-driven spatial networks, which are embedded in a Euclidean space and relate to physical exchange and transportation processes (e.g., blood flow in vessels or traffic flow in road networks). To this end, we propose the Graph Attentive Vectors (GAV) link prediction framework. GAV models simplified dynamics of physical flow in spatial networks via an attentive, neighborhoodaware message-passing paradigm, updating vector embeddings in a constrained manner. We evaluate GAV on eight flow-driven spatial networks given by whole-brain vessel graphs and road networks. GAV demonstrates superior performances across all datasets and metrics and outperforms the current state-of-the-art on the ogbl-vessel benchmark by more than 18% (98.38 vs. 83.07 AUC)."}, "cited_paper_content": {"title": "Inductive Representation Learning On Large Graphs", "abstract": "Low-dimensional embeddings of nodes in large graphs have proved extremely useful in a variety of prediction tasks, from content recommendation to identifying protein functions. However, most existing approaches require that all nodes in the graph are present during training of the embeddings; these previous approaches are inherently transductive and do not naturally generalize to unseen nodes. Here we present GraphSAGE, a general, inductive framework that leverages node feature information (e.g., text attributes) to efficiently generate node embeddings for previously unseen data. Instead of training individual embeddings for each node, we learn a function that generates embeddings by sampling and aggregating features from a node's local neighborhood. Our algorithm outperforms strong baselines on three inductive node-classification benchmarks: we classify the category of unseen nodes in evolving information graphs based on citation and Reddit post data, and we show that our algorithm generalizes to completely unseen graphs using a multi-graph dataset of protein-protein interactions."}, "keywords": ["central node features", "GraphSAGE's message-passing layer"], "citation_intent": "background"}
{"citing_id": "2304.14232v1", "cited_id": "1910.00497", "section_title": "A. Microstrip Patch Type Cell", "citation": "The metasurface cell in #REFR consists of two layers of dielec- tric substrate, as shown in Fig.15 (a).", "text_before_citation": ["To isolate the DC feed port from the microwave signal, three 30 nH inductors are used.", "In this cell design, the phase difference between two adjacent states is (90", "\u2022 \u2212 15 \u2022 , 90 \u2022 + 15", "\u2022 ) in the range of frequency around 3.2 GHz.", "Thus, this cell can be viewed as a 2-bit digital coded metasurface cell at all digital states around 3.2 GHz, and can achieve reflectance of more than 85%."], "text_after_citation": ["The square patch at the top is responsible for reflecting the incident electromagnetic wave, and integrates a PIN diode, which is connected to the ground plane through a via.", "The chip inductance L = 33 nH is used to suppress the coupling between AC and ground.", "Simulation results and experimental results are compared in Fig.15 Fig.", "16 (a) demonstrated, the cell structure in #OTHEREFR consists of two pairs of rectangular patches with a slotted copper ground floor on the back of the dielectric plate.", "Each pair of metal patches is connected by a varactor diode, and the junction capacitance is strongly affected by the external bias voltage."], "citing_paper_content": {"title": "Design Of Reconfigurable Intelligent Surfaces For Wireless Communication: A Review", "abstract": "Existing literature reviews predominantly focus on the theoretical aspects of reconfigurable intelligent surfaces (RISs), such as algorithms and models, while neglecting a thorough examination of the associated hardware components. To bridge this gap, this research paper presents a comprehensive overview of the hardware structure of RISs. The paper provides a classification of RIS cell designs and prototype systems, offering insights into the diverse configurations and functionalities. Moreover, the study explores potential future directions for RIS development. Notably, a novel RIS prototype design is introduced, which integrates seamlessly with a communication system for performance evaluation through signal gain and image formation experiments. The results demonstrate the significant potential of RISs in enhancing communication quality within signal blind zones and facilitating effective radio wave imaging."}, "cited_paper_content": {"title": "Intelligent Metasurface Imager And Recognizer", "abstract": "It is ever-increasingly demanded to remotely monitor people in daily life using radio-frequency probing signals. However, conventional systems can hardly be deployed in real-world settings since they typically require objects to either deliberately cooperate or carry a wireless active device or identification tag. To accomplish the complicated successive tasks using a single device in real time, we propose a smart metasurface imager and recognizer simultaneously, empowered by a network of artificial neural networks (ANNs) for adaptively controlling data flow. Here, three ANNs are employed in an integrated hierarchy: transforming measured microwave data into images of whole human body; classifying the specifically designated spots (hand and chest) within the whole image; and recognizing human hand signs instantly at Wi-Fi frequency of 2.4 GHz. Instantaneous in-situ imaging of full scene and adaptive recognition of hand signs and vital signs of multiple non-cooperative people have been experimentally demonstrated. We also show that the proposed intelligent metasurface system work well even when it is passively excited by stray Wi-Fi signals that ubiquitously exist in our daily lives. The reported strategy could open a new avenue for future smart cities, smart homes, human-device interactive interfaces, healthy monitoring, and safety screening free of visual privacy issues."}, "keywords": ["metasurface cell"], "citation_intent": "background"}
{"citing_id": "2305.02596v1", "cited_id": "1409.1259", "section_title": "1) Actor Network", "citation": "Herein, GRU is a type of recurrent neural networks (RNNs) with a gating mechanism #REFR .", "text_before_citation": ["As shown in Fig.", "4 , the actor network (proxy model) is established as a GRU-equipped DNN, which is designed to output proper actions (Var compensation) for inverters based on the historical trajectory (actor network's input)."], "text_after_citation": ["Such refinement of RNN includes an update gate and a reset gate, which determine what information is allowed through to the output, and it can be trained to retain information over time.", "With this simulated ability to remember information, GRU is used in this paper to process the time-series coupled variable-length historical trajectory as:", "EQUATION", "where denotes the GRU mapping rule, and represents its network parameter.", "represents the GRU hidden state at time instant , which will be updated at each time step."], "citing_paper_content": {"title": "A Soft Coordination Method Of Heterogeneous Devices In Distribution System Voltage Control", "abstract": "With the continuous increase of photovoltaic (PV) penetration, the voltage control interactions between newly installed PV inverters and previously deployed on-load tapchanger (OLTC) transformers become ever more significant. To achieve coordinated voltage regulation, current methods often rely on a decision-making algorithm to fully take over the control of all devices, requiring OLTC to give up its existing tap switching logic and execute corresponding upgrades. Aiming at bridging this gap, a soft coordination framework is proposed in this paper. Specifically, the decision-making commands are only applied on inverters, and OLTC that retains its own operation rule will be indirectly controlled by the changed system voltage, which is a result of appropriately adjusting inverters' Var output. The proposed method achieves the soft coordination by establishing a modified actor-critic algorithm to train a proxy model of inverters. The well-trained proxy model can properly adjust inverters' Var output to \"softly\" coordinate OLTC's tap operations, which finally attains coordinated voltage regulation and line loss minimization. Simulation results verify the superiority of our proposed method over traditional ones in coordinating heterogeneous devices for voltage control."}, "cited_paper_content": {"title": "On The Properties Of Neural Machine Translation: Encoder-Decoder Approaches", "abstract": "Neural machine translation is a relatively new approach to statistical machine translation based purely on neural networks. The neural machine translation models often consist of an encoder and a decoder. The encoder extracts a fixed-length representation from a variable-length input sentence, and the decoder generates a correct translation from this representation. In this paper, we focus on analyzing the properties of the neural machine translation using two models; RNN Encoder--Decoder and a newly proposed gated recursive convolutional neural network. We show that the neural machine translation performs relatively well on short sentences without unknown words, but its performance degrades rapidly as the length of the sentence and the number of unknown words increase. Furthermore, we find that the proposed gated recursive convolutional network learns a grammatical structure of a sentence automatically."}, "keywords": ["recurrent neural networks"], "citation_intent": "background"}
{"citing_id": "2303.08253v1", "cited_id": "2002.07686", "section_title": "Weight Distribution", "citation": "Contrary to that, the weights of the model trained with range-regularization follows a more uniform distribution as expected in #REFR .", "text_before_citation": ["To understand behavior of range-regularization, we analyse weight distribution change during the pre-training stage of ResNet-18 with and without range-regularization.", "Figure 2 shows the weight distribution change of the first layer (conv1), first convolution of second residual block (block2.conv1), first downsampling layer in second residual block and the final linear layer of ResNet-18.", "For the no range regularization case (original) we observe that all the demonstrated weights follow a normal distribution with outliers distributed wide apart."], "text_after_citation": ["We observe the same trend as shown in Table 6 .", "The weight range for each layer with range-regularization is less than the model trained without range-regularization while both models achieve similar accuracy, 70.15% and 69.76% with and without range-regularization respectively, in full-precision evaluation.", "Therefore, we demonstrate that range-regularization effectively removes the outliers in weights without notice- Table 6 it can also be observed that even the standard deviation of the weights for most of the layers are the same with and without range-regularization.", "This indicates that the proposed approach doesn't cause a major shift in most of the weights, therefore the ensuring minimal regression to the accuracy of the pre-trained full precision model.", "In addition, for ResNet-18, range regularization shows some effect in addressing over-fitting as seen from the pre-train full precision metrics in Table 1 ."], "citing_paper_content": {"title": "R 2 : Range Regularization For Model Compression And Quantization", "abstract": "Model parameter regularization is a widely used technique to improve generalization, but also can be used to shape the weight distributions for various purposes. In this work, we shed light on how weight regularization can assist model quantization and compression techniques, and then propose range regularization (R 2) to further boost the quality of model optimization by focusing on the outlier prevention. By effectively regulating the minimum and maximum weight values from a distribution, we mold the overall distribution into a tight shape so that model compression and quantization techniques can better utilize their limited numeric representation powers. We introduce L \u221e regularization, its extension margin regularization and a new soft-min-max regularization to be used as a regularization loss during full-precision model training. Coupled with state-of-the-art quantization and compression techniques, models trained with R 2 perform better on an average, specifically at lower bit weights with 16x compression ratio. We also demonstrate that R 2 helps parameter constrained models like MobileNetV1 achieve significant improvement of around 8% for 2 bit quantization and 7% for 1 bit compression."}, "cited_paper_content": {"title": "Robust Quantization: One Model To Rule Them All", "abstract": "Neural network quantization methods often involve simulating the quantization process during training. This makes the trained model highly dependent on the precise way quantization is performed. Since low-precision accelerators differ in their quantization policies and their supported mix of data-types, a model trained for one accelerator may not be suitable for another. To address this issue, we propose KURE, a method that provides intrinsic robustness to the model against a broad range of quantization implementations. We show that KURE yields a generic model that may be deployed on numerous inference accelerators without a significant loss in accuracy."}, "keywords": ["range-regularization"], "citation_intent": "result"}
{"citing_id": "2304.00959v1", "cited_id": "2002.04920", "section_title": "Ii. Related Work", "citation": "In #REFR , it is shown a real-world application of the chance-constrained MPC for dynamic obstacle avoidance.", "text_before_citation": ["MPC has been used for perching on power lines #OTHEREFR and agile flights #OTHEREFR .", "The first work introducing perception awareness in MPC is #OTHEREFR where the authors propose to include a perception objective in the MPC to keep a point of interest in the camera field of view. Our MPC controller is inspired by #OTHEREFR .", "However, we deal with two different perception objectives, line tracking, and collision avoidance, which conflict with each other when the drone approaches the power masts.", "To enable collision avidance capabilities, in #OTHEREFR , the authors utilized a chanceconstrained MPC formulation.", "This probabilistic collision constraint allows to account for the perceptual uncertainty and consequently enhances the obstacle avoidance robustness."], "text_after_citation": ["In #OTHEREFR , an MPC-based reactive planner for visual target tracking and obstacle avoidance is presented.", "Different from our method, this MPC does not directly generate control commands but planned trajectories, which are tracked by another low-level controller."], "citing_paper_content": {"title": "Autonomous Power Line Inspection With Drones Via Perception-Aware Mpc", "abstract": "The original rough reference trajectory is depicted in yellow and the drone deviates from it while avoiding obstacles (the power masts) and keeping the power line visible in the field of view of the onboard camera."}, "cited_paper_content": {"title": "Robust Vision-Based Obstacle Avoidance For Micro Aerial Vehicles In Dynamic Environments", "abstract": "In this paper, we present an on-board vision-based approach for avoidance of moving obstacles in dynamic environments. Our approach relies on an efficient obstacle detection and tracking algorithm based on depth image pairs, which provides the estimated position, velocity and size of the obstacles. Robust collision avoidance is achieved by formulating a chance-constrained model predictive controller (CC-MPC) to ensure that the collision probability between the micro aerial vehicle (MAV) and each moving obstacle is below a specified threshold. The method takes into account MAV dynamics, state estimation and obstacle sensing uncertainties. The proposed approach is implemented on a quadrotor equipped with a stereo camera and is tested in a variety of environments, showing effective on-line collision avoidance of moving obstacles."}, "keywords": ["dynamic obstacle avoidance"], "citation_intent": "method"}
{"citing_id": "2304.02628v2", "cited_id": "1801.01450", "section_title": "Data Augmentations", "citation": "We complement the finding of #REFR by showing that random crops increase not only translation invariance but also translation equivariance in the intermediate layers.", "text_before_citation": ["We train for 200 epochs using SGD with a batch size of 128 and a learning rate of 0.1 and a momentum of 0.9, L2 regularization at 0.0001 and weight decay at epochs 100 and 150 with a factor of 10.0.", "In principle we expect the in-and equivariance to be the same since translation equivariance should be provided by the convolution.", "However, we include equivariance in our experiments since there are works showing that the information about location can be encoded in different channels #OTHEREFR .", "In Figure 6a we show learned translation equivariance for the tested data augmentations.", "Random crops and Ran-dAugment increase the equivariance of learned features in the Middle, Late and Final parts, while the other data augmentation methods do not have any significant effect, with CutMix even having less equivariance than the baseline in the Middle part."], "text_after_citation": ["Also, we do not see any difference between invariance and equivariance for any data augmentation, which means that any equivariance learned is just invariance."], "citing_paper_content": {"title": "What Affects Learned Equivariance In Deep Image Recognition Models?", "abstract": "Equivariance w.r.t. geometric transformations in neural networks improves data efficiency, parameter efficiency and robustness to out-of-domain perspective shifts. When equivariance is not designed into a neural network, the network can still learn equivariant functions from the data. We quantify this learned equivariance, by proposing an improved measure for equivariance. We find evidence for a correlation between learned translation equivariance and validation accuracy on ImageNet. We therefore investigate what can increase the learned equivariance in neural networks, and find that data augmentation, reduced model capacity and inductive bias in the form of convolutions induce higher learned equivariance in neural networks."}, "cited_paper_content": {"title": "Quantifying Translation-Invariance In Convolutional Neural Networks", "abstract": "A fundamental problem in object recognition is the development of image representations that are invariant to common transformations such as translation, rotation, and small deformations. There are multiple hypotheses regarding the source of translation invariance in CNNs. One idea is that translation invariance is due to the increasing receptive field size of neurons in successive convolution layers. Another possibility is that invariance is due to the pooling operation. We develop a simple a tool, the translation-sensitivity map, which we use to visualize and quantify the translation-invariance of various architectures. We obtain the surprising result that architectural choices such as the number of pooling layers and the convolution filter size have only a secondary effect on the translation-invariance of a network. Our analysis identifies training data augmentation as the most important factor in obtaining translation-invariant representations of images using convolutional neural networks."}, "keywords": ["equivariance", "translation invariance"], "citation_intent": "result"}
{"citing_id": "2304.13616v1", "cited_id": "1906.11300", "section_title": "Introduction", "citation": "On the contrary, despite tracing insufficient generalization to overfitting to the training data, occurrences of benign overfitting have been shown and characterized #REFR .", "text_before_citation": ["Whilst important for supervised-and unsupervised-learning tasks, said generalization is especially important for the safe application * Accepted for publication at IJCAI 2023 of reinforcement learning (RL), where unexpected observations may cause unintended and potentially unsafe behavior #OTHEREFR .", "For instance, a shift might be present when transferring a robotics model from the training simulation to the real-world .", "Generally, increased generalization from a few examples, also referred to as frew-shot learning, is induced by altering the training data to increase its diversity and impede overfitting .", "Those data augmentation techniques have been successfully applied to both supervised and reinforcement learning by increasing the information used for training the model [Laskin et al., 2020] .", "However, while the enlargement of the data may prevent overfitting and thus improve robustness to OOD samples, the increased complexity also increases training difficulty and hinders optimal convergence."], "text_after_citation": ["Also, the states observed by the agent might not resemble an optimal representation, an \"issues that are so often critical to successful applications\" of reinforcement learning #OTHEREFR .", "Extending on these insights, we propose to apply Compact Reshaped Observation Processing (CROP) to reduce the observation such that presumably irrelevant details are removed and the remaining information is sufficient for learning a robust policy that generalizes to unseen shifted observations. Overall, we provide the following contribution:", "\u2022 We formulate three concrete CROP methods applicable to fully observable state and action spaces, reducing the information with regards to the agent's position within the environment, the agent's action space and affect, and the objects within the environment (Figure 1 ).", "\u2022 We provide proof-of-concept and show the strengths of each CROP over the full observation in a safety gridworld training environment and a shifted test environment.", "\u2022 We evaluate the proposed CROPs and compare their effect in zero-shot generalization to the state-of the art of data augmentation in randomly generated mazes."], "citing_paper_content": {"title": "Crop: Towards Distributional-Shift Robust Reinforcement Learning Using Compact Reshaped Observation Processing *", "abstract": "The safe application of reinforcement learning (RL) requires generalization from limited training data to unseen scenarios. Yet, fulfilling tasks under changing circumstances is a key challenge in RL. Current state-of-the-art approaches for generalization apply data augmentation techniques to increase the diversity of training data. Even though this prevents overfitting to the training environment(s), it hinders policy optimization. Crafting a suitable observation, only containing crucial information, has been shown to be a challenging task itself. To improve data efficiency and generalization capabilities, we propose Compact Reshaped Observation Processing (CROP) to reduce the state information used for policy optimization. By providing only relevant information, overfitting to a specific training layout is precluded and generalization to unseen environments is improved. We formulate three CROPs that can be applied to fully observable observation-and action-spaces and provide methodical foundation. We empirically show the improvements of CROP in a distributionally shifted safety gridworld. We furthermore provide benchmark comparisons to full observability and data-augmentation in two different-sized procedurally generated mazes."}, "cited_paper_content": {"title": "Benign Overfitting In Linear Regression", "abstract": "The phenomenon of benign overfitting is one of the key mysteries uncovered by deep learning methodology: deep neural networks seem to predict well, even with a perfect fit to noisy training data. Motivated by this phenomenon, we consider when a perfect fit to training data in linear regression is compatible with accurate prediction. We give a characterization of gaussian linear regression problems for which the minimum norm interpolating prediction rule has near-optimal prediction accuracy. The characterization is in terms of two notions of the effective rank of the data covariance. It shows that overparameterization is essential for benign overfitting in this setting: the number of directions in parameter space that are unimportant for prediction must significantly exceed the sample size."}, "keywords": ["training data", "benign overfitting"], "citation_intent": "background"}
{"citing_id": "2304.14874v1", "cited_id": "1909.03252", "section_title": "C. Proposal Relations Modeling", "citation": "In the field of action localization, P-GCN #REFR constructs a graph of proposals by establishing the edges and then applying GCNs to do message aggregation among proposals. For 3D object detection, Feng et al.", "text_before_citation": ["Modeling proposal relations has emerged in many fields, including 2D #OTHEREFR and 3D object detection #OTHEREFR ."], "text_after_citation": ["#OTHEREFR extract uniform appearance features for each 3D object proposal and construct a relation graph that exploits the 3D object-object relationships.", "In order to explore and utilize object-to-object relations, RDN #OTHEREFR assembles and propagates object relations to enhance object features for video object detection.", "Despite the rapid development of modeling the proposal relations of 2D or 3D bounding boxes, the relations between lane proposals have not been well exploited yet.", "In our work, we aim to mine the location and score relations between proposals to generate diverse and discriminative proposals."], "citing_paper_content": {"title": "Dense Hybrid Proposal Modulation For Lane Detection", "abstract": "In this paper, we present a dense hybrid proposal modulation (DHPM) method for lane detection. Most existing methods perform sparse supervision on a subset of high-scoring proposals, while other proposals fail to obtain effective shape and location guidance, resulting in poor overall quality. To address this, we densely modulate all proposals to generate topologically and spatially high-quality lane predictions with discriminative representations. Specifically, we first ensure that lane proposals are physically meaningful by applying single-lane shape and location constraints. Benefitting from the proposed proposal-to-label matching algorithm, we assign each proposal a target ground truth lane to efficiently learn from spatial layout priors. To enhance the generalization and model the inter-proposal relations, we diversify the shape difference of proposals matching the same ground-truth lane. In addition to the shape and location constraints, we design a quality-aware classification loss to adaptively supervise each positive proposal so that the discriminative power can be further boosted. Our DHPM achieves very competitive performances on four popular benchmark datasets. Moreover, we consistently outperform the baseline model on most metrics without introducing new parameters and reducing inference speed. The codes of our method are available at https://github.com/wuyuej/DHPM."}, "cited_paper_content": {"title": "Graph Convolutional Networks For Temporal Action Localization", "abstract": "Most state-of-the-art action localization systems process each action proposal individually, without explicitly exploiting their relations during learning. However, the relations between proposals actually play an important role in action localization, since a meaningful action always consists of multiple proposals in a video. In this paper, we propose to exploit the proposal-proposal relations using Graph Convolutional Networks (GCNs). First, we construct an action proposal graph, where each proposal is represented as a node and their relations between two proposals as an edge. Here, we use two types of relations, one for capturing the context information for each proposal and the other one for characterizing the correlations between distinct actions. Then we apply the GCNs over the graph to model the relations among different proposals and learn powerful representations for the action classification and localization. Experimental results show that our approach significantly outperforms the state-of-the-art on THUMOS14 (49.1% versus 42.8%). Moreover, augmentation experiments on ActivityNet also verify the efficacy of modeling action proposal relationships. Codes are available at this https URL."}, "keywords": ["3D object detection", "action localization"], "citation_intent": "method"}
{"citing_id": "2303.10070v1", "cited_id": "1902.00751", "section_title": "Related Works", "citation": "As an efficient alternative to full fine-tuning, Adapter-Tuning #REFR was first proposed to transfer large pre-trained Language models to downstream tasks.", "text_before_citation": ["Parameter-Efficient Tuning."], "text_after_citation": ["Inspired by textual prompting, Prompt-Tuning #OTHEREFR and Prefix-Tuning #OTHEREFR insert learnable tokens to adapt to the new task.", "More advanced methods #OTHEREFR achieved comparable or superior performance to full finetuning, and keep the same inference cost by merging the additional learnable parameters to the original pre-trained model.", "Following the step of successful Vision Transformers #OTHEREFR , VPT #OTHEREFR and AdapterFormer #OTHEREFR have been proposed to solve visual transfer learning problems.", "Prompt-Tuning and Prefix-Tuning depend on the transformer architecture because they modify input or hidden tokens.", "Adapter and its variants are network architecture generalizable since they are new modules that can be implemented in forms compatible with pre-trained models."], "citing_paper_content": {"title": "A Unified Continual Learning Framework With General Parameter-Efficient Tuning", "abstract": "The \"pre-training \u2192 downstream adaptation\" presents both new opportunities and challenges for Continual Learning (CL). Although the recent state-of-the-art in CL is achieved through Parameter-Efficient-Tuning (PET) adaptation paradigm, only prompt has been explored, limiting its application to Transformers only. In this paper, we position prompting as one instantiation of PET, and propose a unified CL framework with general PET, dubbed as Learning-Accumulation-Ensemble (LAE). PET, e.g., using Adapter, LoRA, or Prefix, can adapt a pre-trained model to downstream tasks with fewer parameters and resources. Given a PET method, our LAE framework incorporates it for CL with three novel designs. 1) Learning: the pre-trained model adapts to the new task by tuning an online PET module, along with our adaptation speed calibration to align different PET modules, 2) Accumulation: the task-specific knowledge learned by the online PET module is accumulated into an offline PET module through momentum update, 3) Ensemble: During inference, we respectively construct two experts with online/offline PET modules (which are favored by the novel/historical tasks) for prediction ensemble. We show that LAE is compatible with a battery of PET methods and gains strong CL capability. For example, LAE with Adaptor PET surpasses the prior state-ofthe-art by 1.3% and 3.6% in last-incremental accuracy on CIFAR100 and ImageNet-R datasets, respectively."}, "cited_paper_content": {"title": "Parameter-Efficient Transfer Learning For Nlp", "abstract": "Fine-tuning large pre-trained models is an effective transfer mechanism in NLP. However, in the presence of many downstream tasks, fine-tuning is parameter inefficient: an entire new model is required for every task. As an alternative, we propose transfer with adapter modules. Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones. The parameters of the original network remain fixed, yielding a high degree of parameter sharing. To demonstrate adapter's effectiveness, we transfer the recently proposed BERT Transformer model to 26 diverse text classification tasks, including the GLUE benchmark. Adapters attain near state-of-the-art performance, whilst adding only a few parameters per task. On GLUE, we attain within 0.4% of the performance of full fine-tuning, adding only 3.6% parameters per task. By contrast, fine-tuning trains 100% of the parameters per task."}, "keywords": ["large pre-trained Language"], "citation_intent": "method"}
{"citing_id": "2303.15072v1", "cited_id": "1811.01157", "section_title": "Related Works", "citation": "Besides, #REFR proposed to directly manipulate the neuron values to ameliorate the model performance, and similar strategy has been adopted by us to verify the functionality of neurons.", "text_before_citation": ["One of the most commonly used strategy utilizes visualization to observe the dynamics of hidden states for each time-step #OTHEREFR .", "In #OTHEREFR , it shows that counting behavior can be achieved much easier in LSTM, while in other variants such as GRU, the dynamics of most neuron values seem to be irregular.", "Our findings can be considered as offering an explanation how GRU can perform counting even though each individual neuron's value does not seem to have strong correlation with the counts.", "There have been some works utilizing diagnostic classifier #OTHEREFR , Pearson Correlation #OTHEREFR , and ablation test to probe the functions of neurons, which shares similar spirits to the first strategy we used (the only difference is that we further select a subset of dominant features).", "Nevertheless, as we have pointed out, simply performing diagnostic classification or correlation analysis cannot directly confirm the hypothesis, and the third stage of our solution (verification by manipulating neuron values) is the key factor to verify the functions of neurons."], "text_after_citation": ["Nevertheless, we believe the proposed 3-stage strategy can be a general mechanism that researchers can adopt to confirm the hypothesis of functions of neurons."], "citing_paper_content": {"title": "Exposing The Functionalities Of Neurons For Gated Recurrent Unit Based Sequence-To-Sequence Model", "abstract": "The goal of this paper is to report certain scientific discoveries about a Seq2Seq model. It is known that analyzing the behavior of RNN-based models at the neuron level is considered a more challenging task than analyzing a DNN or CNN models due to their recursive mechanism in nature. This paper aims to provide neuron-level analysis to explain why a vanilla GRU-based Seq2Seq model without attention can achieve token-positioning. We found four different types of neurons: storing, counting, triggering, and outputting and further uncover the mechanism for these neurons to work together in order to produce the right token in the right position. Preprint. Under review."}, "cited_paper_content": {"title": "Identifying And Controlling Important Neurons In Neural Machine Translation", "abstract": "Neural machine translation (NMT) models learn representations containing substantial linguistic information. However, it is not clear if such information is fully distributed or if some of it can be attributed to individual neurons. We develop unsupervised methods for discovering important neurons in NMT models. Our methods rely on the intuition that different models learn similar properties, and do not require any costly external supervision. We show experimentally that translation quality depends on the discovered neurons, and find that many of them capture common linguistic phenomena. Finally, we show how to control NMT translations in predictable ways, by modifying activations of individual neurons."}, "keywords": ["neurons"], "citation_intent": "method"}
{"citing_id": "2303.03290v1", "cited_id": "1606.05250", "section_title": "Introduction", "citation": "Factoid extractive QA datasets like SQuAD #REFR , come up with a challenge to measure a QA model competency in identifying the span of an answer from a context for factoid questions.", "text_before_citation": ["In the latter approach, question-answer pairs are formulated using language generation models, machine translation, or manual/learned templates.", "The main challenge in automatic generation is gold answer extraction. Mostly accomplished using existing QA models.", "But getting a dependable model, as perfect as a human, that can produce a correct answer is challenging.", "So, to minimize the generation of trivial and un-grammatical question-1 Amharic is written using Ge'ez script known as \u134a\u12f0\u120d (Fidel) answer pairs, aside from improving the performance of the generation models, experts paraphrase the generated question-answer pairs #OTHEREFR .", "The distinction between the existing datasets lies in the question types (factoid vs non-factoid) and answer formulation sub-task (extractive vs abstractive)."], "text_after_citation": ["Factoid questions like 'What is the capital city of Ethiopia?' (Answer: Addis Ababa) seeks a factual answer that appears as a named entity such as date, location, proper noun, other short noun phrases, or short sentence.", "Unlike that, abstractive QA datasets contain questions whose answer is a comprehension of a context, not a direct copy #OTHEREFR .", "Recently, the QA field of study is getting too many datasets in mono, cross, and multi-lingual settings #OTHEREFR .", "However, Amharic 1 is not included yet in the map of the QA datasets.", "Specific to Amharic there are attempts to develop datasets for other Natural Language Processing (NLP) tasks like sentiment analysis #OTHEREFR , morphologically annotated corpus #OTHEREFR , contemporary Amharic corpus #OTHEREFR , and parallel corpora for machine translation #OTHEREFR ."], "citing_paper_content": {"title": "Amqa: Amharic Question Answering Dataset", "abstract": "Question Answering (QA) returns concise answers or answer lists from natural language text given a context document. To advance robust models' development, large amounts of resources go into curating QA datasets. There is a surge of QA datasets for languages like English, however, this is not the case for Amharic. Amharic, the official language of Ethiopia, is the second most spoken Semitic language in the world. There is no published or publicly available Amharic QA dataset. Hence, to foster the research in Amharic QA, we present the first Amharic QA (AmQA) dataset. We crowdsourced 2628 question-answer pairs over 378 Wikipedia articles. Additionally, we run an XLMRLarge-based baseline model to spark opendomain QA research interest. The best-performing baseline achieves an F-score of 69.58 and 71.74 in reader-retriever QA and reading comprehension settings respectively."}, "cited_paper_content": {"title": "Squad: 100,000+ Questions For Machine Comprehension Of Text", "abstract": "We present the Stanford Question Answering Dataset (SQuAD), a new reading comprehension dataset consisting of 100,000+ questions posed by crowdworkers on a set of Wikipedia articles, where the answer to each question is a segment of text from the corresponding reading passage. We analyze the dataset to understand the types of reasoning required to answer the questions, leaning heavily on dependency and constituency trees. We build a strong logistic regression model, which achieves an F1 score of 51.0%, a significant improvement over a simple baseline (20%). However, human performance (86.8%) is much higher, indicating that the dataset presents a good challenge problem for future research. ::: The dataset is freely available at this https URL"}, "keywords": ["Factoid extractive QA"], "citation_intent": "background"}
{"citing_id": "2303.06779v1", "cited_id": "1908.03119", "section_title": "I. Introduction", "citation": "Since the information of all APs and users are required in cell-free massive MIMO, it is suggested in literature to serve each user by a subset of APs to limit data sharing and computational complexity via two networkcentric and user-centric approaches #REFR .", "text_before_citation": ["In #OTHEREFR , three methods for user scheduling has been presented to optimize the ergodic sum-rate.", "A joint user scheduling and transceiver design scheme for cross-link interference suppression that works on per-frame basis in interfering massive MIMO multicell scenarios has been reported in #OTHEREFR .", "In #OTHEREFR , a userselection method based on simple zero-forcing beamforming with selection (ZFS) has been proposed which attains a sig-nificant fraction of sum capacity and throughput of optimal method.", "Unlike multicell massive MIMO systems where users in each cell are served by a BS, a network with the same area of cells including randomly located single-antenna access points (APs) serving all users in the same time-frequency resource was introduced in #OTHEREFR , #OTHEREFR , #OTHEREFR and called cell-free massive MIMO which presents higher throughput compared with small cells.", "Cell-free massive MIMO is compared with cellular massive MIMO systems in #OTHEREFR , and it is shown that cell-free systems can provide significantly higher spectral efficiency for all users."], "text_after_citation": ["The first approach divides the APs into several clusters, where each includes different APs from other clusters.", "The latter, considers the AP subset providing the best channel conditions for each user which could be dynamic.", "In this paper, we investigate the sum-rate performance of the multicell and cell-free massive MIMO systems using greedy user scheduling based on ZFS algorithm in the downlink #OTHEREFR .", "We consider the use of a network-centric clustering approach to reduce the computational complexity of the techniques applied to the cell-free systems #OTHEREFR .", "We then develop a greedy algorithm that considers multiple candidates for the subset of users to be scheduled and that approaches the performance of the optimal exhaustive search."], "citing_paper_content": {"title": "Study Of Multiuser Scheduling With Enhanced Greedy Techniques For Multicell And Cell-Free Massive Mimo Networks", "abstract": "In this work, we investigate the sum-rate performance of multicell and cell-free massive MIMO systems using linear precoding and multiuser scheduling algorithms. We consider the use of a network-centric clustering approach to reduce the computational complexity of the techniques applied to the cellfree system. We then develop a greedy algorithm that considers multiple candidates for the subset of users to be scheduled and that approaches the performance of the optimal exhaustive search. We assess the proposed and existing scheduling algorithms in both multicell and cell-free networks with the same coverage area. Numerical results illustrate the sum-rate performance of the proposed scheduling algorithm against existing approaches."}, "cited_paper_content": {"title": "Scalable Cell-Free Massive Mimo Systems", "abstract": "Imagine a coverage area with many wireless access points that cooperate to jointly serve the users, instead of creating autonomous cells. Such a cell-free network operation can potentially resolve many of the interference issues that appear in current cellular networks. This ambition was previously called Network MIMO (multiple-input multiple-output) and has recently reappeared under the name Cell-Free Massive MIMO. The main challenge is to achieve the benefits of cell-free operation in a practically feasible way, with computational complexity and fronthaul requirements that are scalable to large networks with many users. We propose a new framework for scalable Cell-Free Massive MIMO systems by exploiting the dynamic cooperation cluster concept from the Network MIMO literature. We provide algorithms for initial access, pilot assignment, cluster formation, precoding, and combining that are proved to be scalable. Interestingly, the proposed scalable precoding and combining outperform conventional maximum ratio processing and also performs closely to the best unscalable alternatives."}, "keywords": ["cell-free massive MIMO"], "citation_intent": "background"}
{"citing_id": "2303.00735v1", "cited_id": "1912.01703", "section_title": "Scalability And Real-World Traffic", "citation": "In our implementation, the Pytorch #REFR auto-differentiation package is used to calculate the gradients.", "text_before_citation": ["A key insight is that with real-world traffic, one may expect certain patterns in this mapping; for example, if two histories of traffic conditions are very similar, their corresponding optimal TE configurations should also be similar. However, measuring and explicitly quantifying such similarities is nontrivial.", "Our approach is to exploit deep neural networks, which have demonstrated remarkable success in identifying complex patterns in high dimensional data, for this task.", "DOTE", "L(D i t , \u03c0 \u03b8 (D i t\u22121 , . . . , D i t\u2212H )", ") can be calculated in closed form via backpropagation."], "text_after_citation": [], "citing_paper_content": {"title": "Dote: Rethinking (Predictive) Wan Traffic Engineering", "abstract": "We explore a new design point for traffic engineering on wide-area networks (WANs): directly optimizing traffic flow on the WAN using only historical data about traffic demands. Doing so obviates the need to explicitly estimate, or predict, future demands. Our method, which utilizes stochastic optimization, provably converges to the global optimum in well-studied theoretical models. We employ deep learning to scale to large WANs and real-world traffic. Our extensive empirical evaluation on real-world traffic and network topologies establishes that our approach's TE quality almost matches that of an (infeasible) omniscient oracle, outperforming previously proposed approaches, and also substantially lowers runtimes."}, "cited_paper_content": {"title": "Pytorch: An Imperative Style, High-Performance Deep Learning Library", "abstract": "Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. ::: In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. ::: We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks."}, "keywords": ["Pytorch auto-differentiation package"], "citation_intent": "method"}
{"citing_id": "2304.12486v2", "cited_id": "1902.06705", "section_title": "Models And Defenses", "citation": "We conducted our experiments on two visual model architectures that currently perform best on RVL-CDIP with less than a hundred million of parameters, as far as we know #REFR : an EfficientNetB0 and a ResNet50. According to the method of Ferrando et al.", "text_before_citation": ["Model backbones."], "text_after_citation": ["#OTHEREFR , who observed the best accuracy of both models on our classification task, we initialized these models with weights that have been pre-trained on the ImageNet classification task.", "Reactive defenses.", "We implemented preprocessing steps, that precede the model backbone, with the aim of improving their robustness.", "The first of these reactive defenses we implemented is the JPEG preprocessing.", "We compress the document image into the JPEG format, then decompress each image that is given as input of the model backbone (JPEG defense). This transformation has been identified by Dsiugaite et al. #OTHEREFR as a factor of robustness against adversarial attacks."], "citing_paper_content": {"title": "Evaluating Adversarial Robustness On Document Image Classification", "abstract": "Adversarial attacks and defenses have gained increasing interest on computer vision systems in recent years, but as of today, most investigations are limited to natural images. However, many artificial intelligence models actually handle documentary data, which is very different from real world images. Hence, in this work, we try to apply the adversarial attack philosophy on documentary data and to protect models against such attacks. Our methodology is to implement untargeted gradient-based, transfer-based and score-based attacks and evaluate the impact of defenses such as adversarial training, JPEG input compression and grey-scale input transformation on the robustness of ResNet50 and EfficientNetB0 model architectures. To the best of our knowledge, no such work has been conducted by the community in order to study the impact of these attacks on the document image classification task."}, "cited_paper_content": {"title": "On Evaluating Adversarial Robustness", "abstract": "Correctly evaluating defenses against adversarial examples has proven to be extremely difficult. Despite the significant amount of recent work attempting to design defenses that withstand adaptive attacks, few have succeeded; most papers that propose defenses are quickly shown to be incorrect. ::: We believe a large contributing factor is the difficulty of performing security evaluations. In this paper, we discuss the methodological foundations, review commonly accepted best practices, and suggest new methods for evaluating defenses to adversarial examples. We hope that both researchers developing defenses as well as readers and reviewers who wish to understand the completeness of an evaluation consider our advice in order to avoid common pitfalls."}, "keywords": ["EfficientNetB0", "experiments"], "citation_intent": "method"}
{"citing_id": "2303.11522v1", "cited_id": "1909.05207", "section_title": "Fixed Cost Functions And Time-Varying Demand", "citation": "Finally, we note that the regret guarantee obtained in Theorem 3 compares favorably to classical O( \u221a T ) regret guarantees in the OCO or MAB literature #REFR .", "text_before_citation": ["Finally, choosing \u03b3 = 1 \u221a T achieves an optimal balance (up to logarithmic terms) between the two quantities above, i.e., O(\u03b3 \u22121 log log T ) and O(\u03b3T ), which establishes the O( \u221a T log log T ) regret bound.", "For a complete proof of Theorem 3, see Appendix B.", "We reiterate that Theorem 3 applies to strongly convex cost functions as with Theorem 1 and that extending this result to general convex cost functions, e.g., linear functions, is, in general, not possible (see Example 1 in Section 4.2).", "Furthermore, compared to the regret guarantee obtained in Theorem 1, Theorem 3 establishes that the time-varying nature of the customer demand incurs an additional factor of O( \u221a T ) in the regret guarantee as compared to the setting with fixed demands.", "However, we do note that if the set of demand realizations D is known a priori to be o( \u221a T ), then the regret guarantee in Theorem 2 can be improved to O(|D| log log T ) by running the direct extension of Algorithm 1, wherein a feasible price interval is maintained for each realized demand."], "text_after_citation": [], "citing_paper_content": {"title": "Online Learning For Equilibrium Pricing In Markets Under Incomplete Information", "abstract": "The study of market equilibria is central to economic theory, particularly in efficiently allocating scarce resources. However, the computation of equilibrium prices at which the supply of goods matches their demand typically relies on having access to complete information on private attributes of agents, e.g., suppliers' cost functions, which are often unavailable in practice. Motivated by this practical consideration, we consider the problem of setting equilibrium prices in the incomplete information setting wherein a market operator seeks to satisfy the customer demand for a commodity by purchasing the required amount from competing suppliers with privately known cost functions unknown to the market operator. In this incomplete information setting, we consider the online learning problem of learning equilibrium prices over time while jointly optimizing three performance metricsunmet demand, cost regret, and payment regret-pertinent in the context of equilibrium pricing over a horizon of T periods. We first consider the setting when suppliers' cost functions are fixed and develop algorithms that achieve a regret of O(log log T) when the customer demand is constant over time, or O(\u221a T log log T) when the demand is variable over time. Next, we consider the setting when the suppliers' cost functions can vary over time and illustrate that no online algorithm can achieve sublinear regret on all three metrics when the market operator has no information about how the cost functions change over time. Thus, we consider an augmented setting wherein the operator has access to hints/contexts that, without revealing the complete specification of the cost functions, reflect the variation in the cost functions over time and propose an algorithm with sublinear regret in this augmented setting."}, "cited_paper_content": {"title": "Introduction To Online Convex Optimization", "abstract": "This monograph portrays optimization as a process. In many practical applications the environment is so complex that it is infeasible to lay out a comprehensive theoretical model and use classical algorithmic theory and mathematical optimization. It is necessary as well as beneficial to take a robust approach, by applying an optimization method that learns as one goes along, learning from experience as more aspects of the problem are observed. This view of optimization as a process has become prominent in varied fields and has led to some spectacular success in modeling and systems that are now part of our daily lives."}, "keywords": ["regret guarantees"], "citation_intent": "result"}
{"citing_id": "2304.09373v1", "cited_id": "1703.06868", "section_title": "A. Coarse-Fusion Network", "citation": "We choose adaptive instance normalization #REFR to build the AIN module due to its efficiency and compact representation. Fig. 3 shows the architecture of the AIN module.", "text_before_citation": ["For a given input hyperspectral image, the proposed model first downsamples the input image into 1/2 and 1/4 scales by using Gaussian kernels.", "Shallow features are extracted by multiple parallel convolutions, as illustrated in Fig. 2 (the initial layer).", "Next, the coarse-fusion network extracts deep features and fuses the multiscale information through several parallel adaptive instance (AIN) modules.", "The motivations for designing the coarse-fusion network are twofold: 1) The multiscale structure presents a solution to increase the receptive field to capture more content.", "2) The AIN module can transfer the basic structures from the low-resolution feature maps to the high-resolution ones."], "text_after_citation": ["The adaptive instance normalization affine transforms the normalized feature map h \u2208 R H\u00d7W \u00d7C by taking an input", "h \u2208 R H 2 \u00d7 W 2 \u00d72C", ".", "Here H and W denote the height and width of the feature map, respectively. C is the number of channels.", "It should be noted that h is the feature from the downscale."], "citing_paper_content": {"title": "Multi-Scale Adaptive Fusion Network For Hyperspectral Image Denoising", "abstract": "Removing the noise and improving the visual quality of hyperspectral images (HSIs) is challenging in academia and industry. Great efforts have been made to leverage local, global or spectral context information for HSI denoising. However, existing methods still have limitations in feature interaction exploitation among multiple scales and rich spectral structure preservation. In view of this, we propose a novel solution to investigate the HSI denoising using a Multi-scale Adaptive Fusion Network (MAFNet), which can learn the complex nonlinear mapping between clean and noisy HSI. Two key components contribute to improving the hyperspectral image denoising: A progressively multiscale information aggregation network and a co-attention fusion module. Specifically, we first generate a set of multiscale images and feed them into a coarse-fusion network to exploit the contextual texture correlation. Thereafter, a fine fusion network is followed to exchange the information across the parallel multiscale subnetworks. Furthermore, we design a co-attention fusion module to adaptively emphasize informative features from different scales, and thereby enhance the discriminative learning capability for denoising. Extensive experiments on synthetic and real HSI datasets demonstrate that the proposed MAFNet has achieved better denoising performance than other state-of-the-art techniques. Our codes are available at https://github.com/summitgao/MAFNet."}, "cited_paper_content": {"title": "Arbitrary Style Transfer In Real-Time With Adaptive Instance Normalization", "abstract": "Gatys et al. recently introduced a neural algorithm that renders a content image in the style of another image, achieving so-called style transfer. However, their framework requires a slow iterative optimization process, which limits its practical application. Fast approximations with feed-forward neural networks have been proposed to speed up neural style transfer. Unfortunately, the speed improvement comes at a cost: the network is usually tied to a fixed set of styles and cannot adapt to arbitrary new styles. In this paper, we present a simple yet effective approach that for the first time enables arbitrary style transfer in real-time. At the heart of our method is a novel adaptive instance normalization (AdaIN) layer that aligns the mean and variance of the content features with those of the style features. Our method achieves speed comparable to the fastest existing approach, without the restriction to a pre-defined set of styles. In addition, our approach allows flexible user controls such as content-style trade-off, style interpolation, color & spatial controls, all using a single feed-forward neural network."}, "keywords": ["adaptive instance normalization"], "citation_intent": "method"}
{"citing_id": "2303.02081v1", "cited_id": "1710.09412", "section_title": "Related Work", "citation": "MixUp #REFR is a mixing approach that linearly combines two data samples as well as the labels.", "text_before_citation": ["Gaussian patch augmentation provides better robustness to corruptions while keeping the accuracy unaffected.", "SaliencyMix #OTHEREFR utilizes salient information to select the dominant patch in a data sample.", "The patch is then used to replace a patch in a different random data sample.", "ResizeMix #OTHEREFR is a similar approach to CutMix, however, instead of replacing the source image patch with a random patch, it uses a whole random, resized data sample.", "The authors make the argument that salient information used by some augmentations for guidance is redundant and outperformed by their re-scaling -pasting into a random position algorithm."], "text_after_citation": ["The ratio between mixed samples is controlled by the \u03bb parameter sampled from Beta distribution.", "Mixup reduces generalization error, memorization of corrupted labels, and improves GAN training #OTHEREFR .", "Puzzle Mix #OTHEREFR combines two images and their labels to create a new data sample.", "The method preserves a local image statistic and at the same time maximizes the amount of saliency information exposed in the resulting data sample.", "Super-Mix #OTHEREFR generalizes mixing up to k images using supervised learning to preserve saliency information."], "citing_paper_content": {"title": "Unproportional Mosaicing", "abstract": "Data shift is a gap between data distribution used for training and data distribution encountered in the real-world. Data augmentations help narrow the gap by generating new data samples, increasing data variability, and data space coverage. We present a new data augmentation: Unproportional mosaicing (Unprop). Our augmentation randomly splits an image into various-sized blocks and swaps its content (pixels) while maintaining block sizes. Our method achieves a lower error rate when combined with other state-of-the-art augmentations. Keywords image augmentation \u2022 convolutional neural network \u2022 image processing"}, "cited_paper_content": {"title": "Mixup: Beyond Empirical Risk Minimization", "abstract": "Large deep neural networks are powerful, but exhibit undesirable behaviors such as memorization and sensitivity to adversarial examples. In this work, we propose mixup, a simple learning principle to alleviate these issues. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularizes the neural network to favor simple linear behavior in-between training examples. Our experiments on the ImageNet-2012, CIFAR-10, CIFAR-100, Google commands and UCI datasets show that mixup improves the generalization of state-of-the-art neural network architectures. We also find that mixup reduces the memorization of corrupt labels, increases the robustness to adversarial examples, and stabilizes the training of generative adversarial networks."}, "keywords": ["mixing approach"], "citation_intent": "background"}
{"citing_id": "2304.02569v1", "cited_id": "1910.14453", "section_title": "Related Work", "citation": "LiDAR-Flow #REFR fuses the LiDAR with a stereo camera to boost the cost computation in the matching process for both stereo and optical flow.", "text_before_citation": ["However, #OTHEREFR heavily rely on a rigid motion prior for moving vehicles, which does not apply to debris flows. #OTHEREFR suffer from overfitting to specific camera intrinsics.", "Their networks are trained on KITTI #OTHEREFR with known camera intrinsics, but encounter problems when applied to other datasets, e.g., nuScenes #OTHEREFR or our debris-flow data.", "Camera-LiDAR fusion.", "An intuitive way to disambiguate depth scaling is to incorporate direct 3D measurements into image-based pipelines.", "Many works have thus focused on sensor fusion for scene flow estimation."], "text_after_citation": ["Subsequent work like DeepLiDARFlow #OTHEREFR encapsulates the solution into an end-to-end deep neural network that regresses scene flow from consecutive LiDAR scans and monocular images.", "CamLiFlow #OTHEREFR mitigates the sen-sor fusion problem with a multi-stage fusion framework and estimates point-wise motion for sparse LiDAR point clouds.", "Multi-frame flow processing.", "Two-frame methods #OTHEREFR do not account for the temporal continuity across multiple consecutive frames.", "In contrast, #OTHEREFR enforces consistency for both spatial and temporal neighbors in a sliding temporal window, #OTHEREFR leverages multi-frame consistency to detect moving object regions, and #OTHEREFR introduces a convolutional LSTM on top of #OTHEREFR to propagate the hidden states via forward warping, but requires stereo images as supervision. Huang et al."], "citing_paper_content": {"title": "Deflow: Self-Supervised 3D Motion Estimation Of Debris Flow", "abstract": "Existing work on scene flow estimation focuses on autonomous driving and mobile robotics, while automated solutions are lacking for motion in nature, such as that exhibited by debris flows. We propose DEFLOW, a model for 3D motion estimation of debris flows, together with a newly captured dataset. We adopt a novel multi-level sensor fusion architecture and self-supervision to incorporate the inductive biases of the scene. We further adopt a multi-frame temporal processing module to enable flow speed estimation over time. Our model achieves state-of-the-art optical flow and depth estimation on our dataset, and fully automates the motion estimation for debris flows. Source code and dataset are available at project page."}, "cited_paper_content": {"title": "Lidar-Flow: Dense Scene Flow Estimation From Sparse Lidar And Stereo Images", "abstract": "We propose a new approach called LiDAR-Flow to robustly estimate a dense scene flow by fusing a sparse LiDAR with stereo images. We take the advantage of the high accuracy of LiDAR to resolve the lack of information in some regions of stereo images due to textureless objects, shadows, ill-conditioned light environment and many more. Additionally, this fusion can overcome the difficulty of matching unstructured 3D points between LiDAR-only scans. Our LiDAR-Flow approach consists of three main steps; each of them exploits LiDAR measurements. First, we build strong seeds from LiDAR to enhance the robustness of matches between stereo images. The imagery part seeks the motion matches and increases the density of scene flow estimation. Then, a consistency check employs LiDAR seeds to remove the possible mismatches. Finally, LiDAR measurements constraint the edge-preserving interpolation method to fill the remaining gaps. In our evaluation we investigate the individual processing steps of our LiDAR-Flow approach and demonstrate the superior performance compared to image-only approach."}, "keywords": ["LiDAR-Flow"], "citation_intent": "method"}
{"citing_id": "2304.01893v1", "cited_id": "1707.06347", "section_title": "Physics-Based Pedestrian Animation", "citation": "We utilize Proximal Policy Optimization (PPO) #REFR to find the optimal control policy \u03c0 PACER .", "text_before_citation": ["To enable full-body pedestrian simulation, we design the Pedestrian Animation ControllER (PACER) to execute the 2D trajectories generated by TRACE in a physics simulator. Background: Goal-Conditioned RL. Our framework ( Fig.", "3) follows the general goal-conditioned reinforcement learning framework, where a goal-conditioned policy \u03c0 PACER is trained to follow 2D target trajectories specified by \u03c4 s .", "The task is formulated as a Markov Decision Process (MDP) defined by a tuple M = S, A, T , R, \u03b3 of states, actions, transition dynamics, reward function, and discount factor.", "The state S, transition dynamics T , and reward R are calculated by the environment based on the current simulation and goal, while the action A is computed by the policy \u03c0 PACER .", "The policy's objective is to maximize the discounted return E T t=1 \u03b3 t\u22121 r t where r t is the reward per timestep."], "text_after_citation": ["Terrain, Social, and Body Awareness.", "To create a controller that can simulate crowds in realistic 3D scenes (e.g. scans, neural reconstructions, or artist-created meshes (Fig.", "1) ), our humanoid must be terrain aware, socially aware of other agents, and support diverse body types.", "We use a humanoid model that conforms to the kinematic structure of SMPL #OTHEREFR , and is automatically generated using a procedure similar to #OTHEREFR .", "Our control policy \u03c0 PACER (a t |h t , o t , \u03b2, \u03c4 s ) is conditioned on the state of the simulated character h t , environmental features o t , body type \u03b2, and goal trajectory \u03c4 s ."], "citing_paper_content": {"title": "Trace And Pace: Controllable Pedestrian Animation Via Guided Trajectory Diffusion", "abstract": "Figure 1. (Left) We propose TRACE, a trajectory diffusion model that enables user control through test-time guidance. (Right) Generated trajectories are passed to a novel physics-based humanoid controller (PACER), forming a closed-loop pedestrian animation system."}, "cited_paper_content": {"title": "Proximal Policy Optimization Algorithms", "abstract": "We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a \"surrogate\" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time."}, "keywords": ["optimal control policy", "Proximal Policy Optimization"], "citation_intent": "method"}
{"citing_id": "2303.02472v1", "cited_id": "1810.04805", "section_title": "Datasets And Models", "citation": "Specifically, we used Bert-base #REFR and Roberta-base (Liu et al.) , for SNLI and ANLI, respectively.", "text_before_citation": ["The data consists of 162,865/3,200/3,200 sentence pairs for train/val/test set respectively.", "The max length of the input is set to 128.", "For the Image Classification datasets, we used Convolutional Neural Networks (CNNs).", "Specifically, we used LeNet5 #OTHEREFR , ResNet50, ResNet34, and ResNet18 #OTHEREFR , for MNIST, CIFAR10, CIFAR100, and ImageNet100, respectively.", "For the NLI datasets, we finetuned transformer based Pre-trained Language Models (PLMs)."], "text_after_citation": [], "citing_paper_content": {"title": "Esd: Expected Squared Difference As A Tuning-Free Trainable Calibration Measure", "abstract": "Studies have shown that modern neural networks tend to be poorly calibrated due to over-confident predictions. Traditionally, post-processing methods have been used to calibrate the model after training. In recent years, various trainable calibration measures have been proposed to incorporate them directly into the training process. However, these methods all incorporate internal hyperparameters, and the performance of these calibration objectives relies on tuning these hyperparameters, incurring more computational costs as the size of neural networks and datasets become larger. As such, we present Expected Squared Difference (ESD), a tuning-free (i.e., hyperparameter-free) trainable calibration objective loss, where we view the calibration error from the perspective of the squared difference between the two expectations. With extensive experiments on several architectures (CNNs, Transformers) and datasets, we demonstrate that (1) incorporating ESD into the training improves model calibration in various batch size settings without the need for internal hyperparameter tuning, (2) ESD yields the best-calibrated results compared with previous approaches, and (3) ESD drastically improves the computational costs required for calibration during training due to the absence of internal hyperparameter. The code is publicly accessible at https://github.com/heesuk-yoon/ESD."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["Bert-base"], "citation_intent": "method"}
{"citing_id": "2304.03801v1", "cited_id": "1703.00056", "section_title": "Group Fairness", "citation": "Calibration #REFR : A classifier is said to satisfy calibration if both protected and unprotected groups have almost similar positive predictive values (PPV).", "text_before_citation": ["for all k \u2208 Y and X m , X m \u2208 X .", "If the difference SP m,k \u2212 SP m ,k is greater than 0, then the protected group is benefited.", "On the other hand, if the difference is less than 0, the unprotected group is benefited.", "Note that the statistical parity of any system can be directly measured from the system's outcome labels, without any need for an alternative classifier.", "However, this is not the case with other statistical fairness metrics."], "text_after_citation": ["The PPV represents the probability of an individual with a positive prediction actually experiencing a positive outcome.", "Formally, if C m,k = P(z = k | y = k, x \u2208 X m ) denote the positive predictive rate for the group X m , the calibration of the system g with respect to the classifier f is computed as", "EQUATION", "for all k \u2208 Y and X m , X m \u2208 X .", "Accuracy Equality #OTHEREFR : This statistical measure computes the probability that both the classifiers g and f yielding the same label."], "citing_paper_content": {"title": "Towards Inclusive Fairness Evaluation Via Eliciting Disagreement Feedback From Non-Expert Stakeholders", "abstract": "Traditional algorithmic fairness notions rely on label feedback, which can only be elicited from expert critics. However, in most practical applications, several non-expert stakeholders also play a major role in the system and can have distinctive opinions about the decision making philosophy. For example, in kidney placement programs, transplant surgeons are very wary about accepting kidney offers for black patients due to genetic reasons. However, non-expert stakeholders in kidney placement programs (e.g. patients, donors and their family members) may misinterpret such decisions from the perspective of social discrimination. This paper evaluates group fairness notions from the viewpoint of non-expert stakeholders, who can only provide binary agreement/disagreement feedback regarding the decision in context. Specifically, two types of group fairness notions have been identified: (i) definite notions (e.g. calibration), which can be evaluated exactly using disagreement feedback, and (ii) indefinite notions (e.g. equal opportunity) which suffer from uncertainty due to lack of label feedback. In the case of indefinite notions, bounds are presented based on disagreement rates, and an estimate is constructed based on established bounds. The efficacy of all our findings are validated empirically on real human feedback dataset. This work is supported by an NSF grant..."}, "cited_paper_content": {"title": "Fair Prediction With Disparate Impact: A Study Of Bias In Recidivism Prediction Instruments", "abstract": "Recidivism prediction instruments (RPI's) provide decision makers with an assessment of the likelihood that a criminal defendant will reoffend at a future point in time. While such instruments are gaining increasing popularity across the country, their use is attracting tremendous controversy. Much of the controversy concerns potential discriminatory bias in the risk assessments that are produced. This paper discusses several fairness criteria that have recently been applied to assess the fairness of recidivism prediction instruments. We demonstrate that the criteria cannot all be simultaneously satisfied when recidivism prevalence differs across groups. We then show how disparate impact can arise when a recidivism prediction instrument fails to satisfy the criterion of error rate balance."}, "keywords": ["classifier", "predictive values"], "citation_intent": "background"}
{"citing_id": "2303.03295v1", "cited_id": "2003.04031", "section_title": "B. Semi-Decentralized Equilibrium Seeking", "citation": "In particular, we adopt the Inertial Forward-Reflected-Backward (I-FoRB) algorithm #REFR , for its convergence speed and low computational complexity.", "text_before_citation": ["To solve the game in (16), we focus on the computation of a variational GNE (v-GNE) #OTHEREFR Def.", "3.10] , that is, the subset of GNEs which satisfy the KKT conditions", "\u03c9 \u03bb \u2208 zer \u00d7 i\u2208I N \u2126i (\u03c9 i ) + F (\u03c9) + col(A i \u03bb) i N R |E| \u22650 (\u03bb) \u2212 A\u03c9 + b (23) where \u03bb \u2208 R |E| \u22650", "is the dual variable associated to the shared constraints in #OTHEREFR .", "The v-GNEs have desirable characteristics of fairness between agents and there exist several efficient algorithms for their computation."], "text_after_citation": ["The I-FoRB algorithm guarantees convergence in the general class of (non-strictly) monotone games.", "On the contrary, the algorithm proposed in #OTHEREFR converges only if the game is strongly Algorithm 1 I-FoRB GNE seeking for traffic routing (k) and computes:", "Initialization.", "Set (\u03c9 (0) i ) i\u2208I , (\u03c9 (1) i ) i\u2208I \u2208 \u2126 j , \u03bb (1) \u2208 R |E| \u22650 . Set \u03c3 (1) = avg(\u03c9 (1) ).", "Set \u03c9 (1) i = \u03c9 (1) i + \u03b8(\u03c9 (1) i \u2212 \u03c9 (0) i ) for all i \u2208 I, \u03bb (1) = \u03bb (1) + \u03b8(\u03bb (1) \u2212 \u03bb (0) ) ."], "citing_paper_content": {"title": "Probabilistic Game-Theoretic Traffic Routing", "abstract": "We examine the routing problem for self-interested vehicles using stochastic decision strategies. By approximating the road latency functions and a non-linear variable transformation, we frame the problem as an aggregative game. We characterize the approximation error and we derive a new monotonicity condition for a broad category of games that encompasses the problem under consideration. Next, we propose a semidecentralized algorithm to calculate the routing as a variational generalized Nash equilibrium and demonstrate the solution's benefits with numerical simulations. We also explore a recursive receding-horizon formulation of the routing problem for potential games, showing asymptotic convergence to destinations and analysing closed-loop performance dependence on horizon length through numerical simulations."}, "cited_paper_content": {"title": "Semi-Decentralized Generalized Nash Equilibrium Seeking In Monotone Aggregative Games", "abstract": "We address the generalized Nash equilibrium seeking problem for a population of agents playing aggregative games with affine coupling constraints. We focus on semi-decentralized communication architectures, where there is a central coordinator able to gather and broadcast signals of aggregative nature to the agents. By exploiting the framework of monotone operator theory and operator splitting, we first critically review the most relevant available algorithms and then design two novel schemes: (i) a single-layer, fixed-step algorithm with convergence guarantee for general (non cocoercive, non-strictly) monotone aggregative games and (ii) a single-layer proximal-type algorithm for a class of monotone aggregative games with linearly coupled cost functions. We also design novel accelerated variants of two algorithms via (alternating) inertial steps. Finally, we show via numerical simulations that the proposed algorithms outperform those in the literature in terms of convergence speed."}, "keywords": ["low computational complexity", "convergence speed"], "citation_intent": "method"}
{"citing_id": "2304.08586v1", "cited_id": "2004.03315", "section_title": "I. Introduction", "citation": "Given this difficulty, work has been done in learning CBFs from expert data of safe and unsafe interactions #REFR .", "text_before_citation": ["Model predictive control (MPC) #OTHEREFR and trajectory optimization (TO) based methods #OTHEREFR have been widely used for safetycritical robot applications, e.g., obstacle avoidance.", "However, the computation time of MPC and TO based methods has made them difficult to be directly deployed on systems that require fast response time.", "Recently, a new control paradigm, CBF-based control #OTHEREFR , has become increasingly popular for safe robotic control, since it provides a simple yet computationally efficient way for safe control synthesis.", "One of the main hurdles to applying CBF-based methods to robotic systems is the construction of a valid CBF.", "Although work has been done in synthesizing CBFs for robots, the constructed CBF is often not intuitive #OTHEREFR ."], "text_after_citation": ["Also, CBFs for robotics systems can be learned online using onboard sensor data #OTHEREFR .", "Additionally, work has been done in learning CBFs starting from a handcrafted CBF #OTHEREFR .", "Although the learning-based methods for CBF synthesis are appealing, Fig. 1 .", "The structure of safe robotic control using differentiableoptimization-based CBFs.", "acquiring a dataset may be costly in many cases, e.g., selfdriving cars."], "citing_paper_content": {"title": "Safe Navigation And Obstacle Avoidance Using Differentiable Optimization Based Control Barrier Functions", "abstract": "Control barrier functions (CBFs) have been widely applied to safety-critical robotic applications. However, the construction of control barrier functions for robotic systems remains a challenging task. Recently, collision detection using differentiable optimization has provided a way to compute the minimum uniform scaling factor that results in an intersection between two convex shapes and to also compute the Jacobian of the scaling factor. In this paper, we propose a framework that uses this scaling factor, with an offset, to systematically define a CBF for obstacle avoidance tasks. We provide a theoretical analysis that proves the continuity of the proposed CBF. Empirically, we show that the proposed CBF is continuously differentiable, and the resulting optimal control problem is computationally efficient, which makes it applicable for realtime robotic control. We validate our approach, first using a 2D mobile robot example, then on the Franka-Emika Research 3 (FR3) robot manipulator both in simulation and experiment."}, "cited_paper_content": {"title": "Learning Control Barrier Functions From Expert Demonstrations", "abstract": "Inspired by the success of imitation and inverse reinforcement learning in replicating expert behavior through optimal control, we propose a learning based approach to safe controller synthesis based on control barrier functions (CBFs). We consider the setting of a known nonlinear control affine dynamical system and assume that we have access to safe trajectories generated by an expert - a practical example of such a setting would be a kinematic model of a self-driving vehicle with safe trajectories (e.g. trajectories that avoid collisions with obstacles in the environment) generated by a human driver. We then propose and analyze an optimization-based approach to learning a CBF that enjoys provable safety guarantees under suitable Lipschitz smoothness assumptions on the underlying dynamical system. A strength of our approach is that it is agnostic to the parameterization used to represent the CBF, assuming only that the Lipschitz constant of such functions can be efficiently bounded. Furthermore, if the CBF parameterization is convex, then under mild assumptions, so is our learning process. We end with extensive numerical evaluations of our results on both planar and realistic examples, using both random feature and deep neural network parameterizations of the CBF. To the best of our knowledge, these are the first results that learn provably safe control barrier functions from data."}, "keywords": ["unsafe interactions", "expert data"], "citation_intent": "background"}
{"citing_id": "2303.08692v1", "cited_id": "1706.05587", "section_title": "Spatial-Aware Recursive Meshing", "citation": "SRM module is composed of a modified ASPP block #REFR and three spatial-aware feature refinement blocks, as shown in Figure 5 .", "text_before_citation": ["Semantic segmentation is a pixel-level scene understanding task, which relies on fine-grained semantic features for pixel-wise classification.", "However, detailed information loss caused by downsampling is inevitable during encoding.", "To compensate information loss and refine the finegrained features, we propose a SRM module that leverages the fused multimodal features in a recursive manner with spatial awareness."], "text_after_citation": ["For RGB modality, the encoded feature map f rgb4 is first embedded with more global features via atrous spatial pyramid pooling.", "We use three dilation rates (d = 2, 4, 8), and the number of channels is reduced to 256 after this block.", "To refine the fine-grained semantic features, the features of different receptive fields are recursively introduced via skip-connection.", "Instead of simple concatenation of features, we perform spatial-aware feature refinement to proactively mesh upsampled features with multimodal features f mi which contain rich detailed semantic information from both RGB-T signals.", "Considering the complexity, channel reduction is applied to f mi via convolutional operation."], "citing_paper_content": {"title": "Spidermesh: Spatial-Aware Demand-Guided Recursive Meshing For Rgb-T Semantic Segmentation", "abstract": "For semantic segmentation in urban scene understanding, RGB cameras alone often fail to capture a clear holistic topology, especially in challenging lighting conditions. Thermal signal is an informative additional channel that can bring to light the contour and fine-grained texture of blurred regions in low-quality RGB image. Aiming at RGB-T (thermal) segmentation, existing methods either use simple passive channel/spatial-wise fusion for cross-modal interaction, or rely on heavy labeling of ambiguous boundaries for fine-grained supervision. We propose a Spatialaware Demand-guided Recursive Meshing (SpiderMesh) framework that: 1) proactively compensates inadequate contextual semantics in optically-impaired regions via a demand-guided target masking algorithm; 2) refines multimodal semantic features with recursive meshing to improve pixel-level semantic analysis performance. We further introduce an asymmetric data augmentation technique M-CutOut, and enable semi-supervised learning to fully utilize RGB-T labels only sparsely available in practical use. Extensive experiments on MFNet and PST900 datasets demonstrate that SpiderMesh achieves state-of-the-art performance on standard RGB-T segmentation benchmarks."}, "cited_paper_content": {"title": "Rethinking Atrous Convolution For Semantic Image Segmentation", "abstract": "In this work, we revisit atrous convolution, a powerful tool to explicitly adjust filter's field-of-view as well as control the resolution of feature responses computed by Deep Convolutional Neural Networks, in the application of semantic image segmentation. To handle the problem of segmenting objects at multiple scales, we design modules which employ atrous convolution in cascade or in parallel to capture multi-scale context by adopting multiple atrous rates. Furthermore, we propose to augment our previously proposed Atrous Spatial Pyramid Pooling module, which probes convolutional features at multiple scales, with image-level features encoding global context and further boost performance. We also elaborate on implementation details and share our experience on training our system. The proposed `DeepLabv3' system significantly improves over our previous DeepLab versions without DenseCRF post-processing and attains comparable performance with other state-of-art models on the PASCAL VOC 2012 semantic image segmentation benchmark."}, "keywords": ["three spatial-aware feature"], "citation_intent": "method"}
{"citing_id": "2304.00485v1", "cited_id": "2002.12307", "section_title": "Task Graph Type Graph Approach Task Level Paper", "citation": "The aggregation and burst phenomenon observed by #REFR suggested that a large number of fraudsters logging the same device in a short time tend to be high-risk.", "text_before_citation": ["Financial fraud Transactions Attributed (statistical feature) Subgraph #OTHEREFR Node #OTHEREFR Att+Dee+Hom (graph embedding) #OTHEREFR Att+Dee+Het (graph embedding) #OTHEREFR Account-device #OTHEREFR Underground market", "Underground forum Str+Sha+Het (graph embedding) #OTHEREFR Att+Dee+Het (graph embedding)", "Edge #OTHEREFR [41]", "Node #OTHEREFR Word co-occurrence Att+Sha+Hom (graph embedding) Node/graph #OTHEREFR Transaction Structural (statistical feature) Node #OTHEREFR Host Attributed (statistical feature) Subgraph #OTHEREFR systems (e.g., Alipay 2 ) #OTHEREFR .", "#OTHEREFR extracted high-risk dense subgraphs and obtain the risk scores based on homogeneous graph embeddings for further screening."], "text_after_citation": ["It initiated user node vectors by temporal activity statistics and utilized GNNs to aggregate local neighbors.", "However, normal accounts may also log in to the same device so that the above GNN-based models raise a high false positive rate. #OTHEREFR solved this problem by introducing adaptive receptive paths.", "Besides these general solutions, graph mining techniques have also been successfully used in many specific tasks of financial risk management recently. We introduce these tasks and corresponding solutions as follows.", "Money Laundering.", "Money laundering is a financial industry behavior, which makes illegal income legalization."], "citing_paper_content": {"title": "Graph Mining For Cybersecurity: A Survey", "abstract": "The explosive growth of cyber attacks nowadays, such as malware, spam, and intrusions, caused severe consequences on society. Securing cyberspace has become an utmost concern for organizations and governments. Traditional Machine Learning (ML) based methods are extensively used in detecting cyber threats, but they hardly model the correlations between real-world cyber entities. In recent years, with the proliferation of graph mining techniques, many researchers investigated these techniques for capturing correlations between cyber entities and achieving high performance. It is imperative to summarize existing graph-based cybersecurity solutions to provide a guide for future studies. Therefore, as a key contribution of this paper, we provide a comprehensive review of graph mining for cybersecurity, including an overview of cybersecurity tasks, the typical graph mining techniques, and the general process of applying them to cybersecurity, as well as various solutions for different cybersecurity tasks. For each task, we probe into relevant methods and highlight the graph types, graph approaches, and task levels in their modeling. Furthermore, we collect open datasets and toolkits for graph-based cybersecurity. Finally, we outlook the potential directions of this field for future research."}, "cited_paper_content": {"title": "Heterogeneous Graph Neural Networks For Malicious Account Detection", "abstract": "We present, GEM, the first heterogeneous graph neural network approach for detecting malicious accounts at Alipay, one of the world's leading mobile cashless payment platform. Our approach, inspired from a connected subgraph approach, adaptively learns discriminative embeddings from heterogeneous account-device graphs based on two fundamental weaknesses of attackers, i.e. device aggregation and activity aggregation. For the heterogeneous graph consists of various types of nodes, we propose an attention mechanism to learn the importance of different types of nodes, while using the sum operator for modeling the aggregation patterns of nodes in each type. Experiments show that our approaches consistently perform promising results compared with competitive methods over time."}, "keywords": ["aggregation"], "citation_intent": "background"}
{"citing_id": "2303.01123v1", "cited_id": "1909.01193", "section_title": "Ii. Related Work", "citation": "We provide an extension of their work, by introducing a learning method to estimate the bias model parameters using raw sensory measurements. Sterzentsenko et al. #REFR presented self-supervised depth denoising method for RGBD sensors.", "text_before_citation": ["A laser beam width is typically modeled as a Gaussian function of the distance along the beam #OTHEREFR . Using non-zero-mean Gaussian beam model, Laconte et al.", "#OTHEREFR were able to measure the bias of several commercial lidar sensors as a function of depth and incidence angle.", "The authors additionally present an experimental methodology on estimation of the bias model parameters for a lidar sensor."], "text_after_citation": ["Similar to our approach, the authors support the idea of utilizing spatial consistency and surface smoothness as optimization objectives.", "We use the lidar depth correction model, introduced in #OTHEREFR .", "It estimates the bias as a polynomial function of laser beam incidence angle, Section III-A.", "The authors utilize the model for lidars and cameras calibration scenario.", "However, the application of the model in localization and mapping scenarios was not studied in their paper."], "citing_paper_content": {"title": "Self-Supervised Depth Correction Of Lidar Measurements From Map Consistency Loss", "abstract": "Depth perception is considered an invaluable source of information in the context of 3D mapping and various robotics applications. However, point cloud maps acquired using consumer-level light detection and ranging sensors (lidars) still suffer from bias related to local surface properties such as measuring beam-to-surface incidence angle, distance, texture, reflectance, or illumination conditions. This fact has recently motivated researchers to exploit traditional filters, as well as the deep learning paradigm, in order to suppress the aforementioned depth sensors error while preserving geometric and map consistency details. Despite the effort, depth correction of lidar measurements is still an open challenge mainly due to the lack of clean 3D data that could be used as ground truth. In this paper, we introduce two novel point cloud map consistency losses, which facilitate self-supervised learning on real data of lidar depth correction models. Specifically, the models exploit multiple point cloud measurements of the same scene from different viewpoints in order to learn to reduce the bias based on the constructed map consistency signal. Complementary to the removal of the bias from the measurements, we demonstrate that the depth correction models help to reduce localization drift. Additionally, we release a data set that contains point cloud data captured in an indoor corridor environment with precise localization and ground truth mapping information 1 ."}, "cited_paper_content": {"title": "Self-Supervised Deep Depth Denoising", "abstract": "Depth perception is considered an invaluable source of information for various vision tasks. However, depth maps acquired using consumer-level sensors still suffer from non-negligible noise. This fact has recently motivated researchers to exploit traditional filters, as well as the deep learning paradigm, in order to suppress the aforementioned non-uniform noise, while preserving geometric details. Despite the effort, deep depth denoising is still an open challenge mainly due to the lack of clean data that could be used as ground truth. In this paper, we propose a fully convolutional deep autoencoder that learns to denoise depth maps, surpassing the lack of ground truth data. Specifically, the proposed autoencoder exploits multiple views of the same scene from different points of view in order to learn to suppress noise in a self-supervised end-to-end manner using depth and color information during training, yet only depth during inference. To enforce selfsupervision, we leverage a differentiable rendering technique to exploit photometric supervision, which is further regularized using geometric and surface priors. As the proposed approach relies on raw data acquisition, a large RGB-D corpus is collected using Intel RealSense sensors. Complementary to a quantitative evaluation, we demonstrate the effectiveness of the proposed self-supervised denoising approach on established 3D reconstruction applications. Code is avalable at this https URL"}, "keywords": ["self-supervised depth denoising"], "citation_intent": "method"}
{"citing_id": "2303.08225v1", "cited_id": "2003.06988", "section_title": "Ablation Study", "citation": "GTGAN has 11 baselines, as shown in Table 3 : (1) B1 is our baseline combining HouseGAN G and HouseGAN D (i.e., the original HouseGAN #REFR ).", "text_before_citation": ["We conduct extensive ablation studies on house layout generation (\"10-12\" subset) to evaluate the effectiveness of each component of the proposed GTGAN. Baselines Models."], "text_after_citation": ["(2) B2 adopts the combination of GTGAN G and HouseGAN D. (3) B3 combines HouseGAN G and GTGAN D. (4) B4 employs both GTGAN G and GTGAN D. (5) B5 is our baseline without using NNA. (6) B6 is our baseline without using CNA. (7) B7 is our baseline without using GMB.", "(8) B8 is our baseline using Transformer layers instead of Conv-MPN layers. (9) B9 is the variation using Eq. (3) instead of Eq. (2). (10) B10 is the variation using Eq. (4) instead of Eq. (2).", "(11) B11 is our full model, using the proposed graph-based cycle-consistency loss upon B4. Ablation Analysis.", "The results of the ablation study, shown in Table 3 , prove that our graph Transformer generator G and node classification-based discriminator D improve the generation performance over the baseline models, validating the effectiveness of the proposed framework.", "Specifically, when using our generator, B2 yields further improvements over B1, meaning that our generator learns local and global relations across connected and non-connected nodes more effectively, confirming our design motivation."], "citing_paper_content": {"title": "Graph Transformer Gans For Graph-Constrained House Generation", "abstract": "We present a novel graph Transformer generative adversarial network (GTGAN) to learn effective graph node relations in an end-to-end fashion for the challenging graphconstrained house generation task. The proposed graph-Transformer-based generator includes a novel graph Transformer encoder that combines graph convolutions and selfattentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node attention (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. Moreover, we propose a new node classification-based discriminator to preserve the high-level semantic and discriminative node features for different house components. Finally, we propose a novel graph-based cycle-consistency loss that aims at maintaining the relative spatial relationships between ground truth and predicted graphs. Experiments on two challenging graph-constrained house generation tasks (i.e., house layout and roof generation) with two public datasets demonstrate the effectiveness of GTGAN in terms of objective quantitative scores and subjective visual realism. New state-of-the-art results are established by large margins on both tasks."}, "cited_paper_content": {"title": "House-Gan: Relational Generative Adversarial Networks For Graph-Constrained House Layout Generation", "abstract": "This paper proposes a novel graph-constrained generative adversarial network, whose generator and discriminator are built upon relational architecture. The main idea is to encode the constraint into the graph structure of its relational networks. We have demonstrated the proposed architecture for a new house layout generation problem, whose task is to take an architectural constraint as a graph (i.e., the number and types of rooms with their spatial adjacency) and produce a set of axis-aligned bounding boxes of rooms. We measure the quality of generated house layouts with the three metrics: the realism, the diversity, and the compatibility with the input graph constraint. Our qualitative and quantitative evaluations over 117,000 real floorplan images demonstrate that the proposed approach outperforms existing methods and baselines. We will publicly share all our code and data."}, "keywords": ["original HouseGAN", "HouseGAN"], "citation_intent": "method"}
{"citing_id": "2303.15532v1", "cited_id": "1902.07243", "section_title": "Collaborative Filtering", "citation": "Alternatively, attention mechanisms have been also proposed to capture the importance of different relationships between users and items #REFR .", "text_before_citation": ["proposed a Neural Graph Collaborative Filtering (NGCF) framework that captured collaborative signals in high-hop neighbors and integrates them into the embedding learning process #OTHEREFR .", "However, further studies showed that NGCF demonstrates higher training loss and worse generalization performance with nonlinear activation and feature transformation #OTHEREFR .", "As a result, the authors proposed a simplified model named Light Graph Convolution Network (LightGCN).", "Other works have leveraged LightGCN by aggregating information from different aspect-level graphs #OTHEREFR (e.g., adding a user-director graph on a user-movie recommendation to guide the embedding learning process).", "These RS models typically aggregate information by averaging data from neighbors."], "text_after_citation": ["In our experiments, we also test various aggregation strategies (see Section 5.4)"], "citing_paper_content": {"title": "Stance Inference In Twitter Through Graph Convolutional Collaborative Filtering Networks With Minimal Supervision", "abstract": "Social Media (SM) has become a stage for people to share thoughts, emotions, opinions, and almost every other aspect of their daily lives. This abundance of human interaction makes SM particularly attractive for social sensing. Especially during polarizing events such as political elections or referendums, users post information and encourage others to support their side, using symbols such as hashtags to represent their attitudes. However, many users choose not to attach hashtags to their messages, use a different language, or show their position only indirectly. Thus, automatically identifying their opinions becomes a more challenging task. To uncover these implicit perspectives, we propose a collaborative filtering model based on Graph Convolutional Networks that exploits the textual content in messages and the rich connections between users and topics. Moreover, our approach only requires a small annotation effort compared to state-of-the-art solutions. Nevertheless, the proposed model achieves competitive performance in predicting individuals' stances. We analyze users' attitudes ahead of two constitutional referendums in Chile in 2020 and 2022. Using two large Twitter datasets, our model achieves improvements of 3.4% in recall and 3.6% in accuracy over the baselines."}, "cited_paper_content": {"title": "Graph Neural Networks For Social Recommendation", "abstract": "In recent years, Graph Neural Networks (GNNs), which can naturally integrate node information and topological structure, have been demonstrated to be powerful in learning on graph data. These advantages of GNNs provide great potential to advance social recommendation since data in social recommender systems can be represented as user-user social graph and user-item graph; and learning latent factors of users and items is the key. However, building social recommender systems based on GNNs faces challenges. For example, the user-item graph encodes both interactions and their associated opinions; social relations have heterogeneous strengths; users involve in two graphs (e.g., the user-user social graph and the user-item graph). To address the three aforementioned challenges simultaneously, in this paper, we present a novel graph neural network framework (GraphRec) for social recommendations. In particular, we provide a principled approach to jointly capture interactions and opinions in the user-item graph and propose the framework GraphRec, which coherently models two graphs and heterogeneous strengths. Extensive experiments on two real-world datasets demonstrate the effectiveness of the proposed framework GraphRec. Our code is available at \\url{this https URL}"}, "keywords": ["users", "different relationships"], "citation_intent": "background"}
{"citing_id": "2304.04308v1", "cited_id": "1411.1607", "section_title": "Software And Computational Resources", "citation": "We wrote all code in Julia 1.6 #REFR , using the JuMP package (Dunning et al., 2017) to write optimization functions and Gurobi (Gurobi Optimization, LLC, 2022) as the solver.", "text_before_citation": [], "text_after_citation": ["We performed each individual experiment reported in this paper using 4 Intel Xeon Platinum 8260 CPU cores from the Supercloud cluster #OTHEREFR ."], "citing_paper_content": {"title": "Ensemble Modeling For Time Series Forecasting: An Adaptive Robust Optimization Approach", "abstract": "Accurate time series forecasting is critical for a wide range of problems with temporal data. Ensemble modeling is a well-established technique for leveraging multiple predictive models to increase accuracy and robustness, as the performance of a single predictor can be highly variable due to shifts in the underlying data distribution. This paper proposes a new methodology for building robust ensembles of time series forecasting models. Our approach utilizes Adaptive Robust Optimization (ARO) to construct a linear regression ensemble in which the models' weights can adapt over time. We demonstrate the effectiveness of our method through a series of synthetic experiments and real-world applications, including air pollution management, energy consumption forecasting, and tropical cyclone intensity forecasting. Our results show that our adaptive ensembles outperform the best ensemble member in hindsight by 16-26% in root mean square error and 14-28% in conditional value at risk and improve over competitive ensemble techniques."}, "cited_paper_content": {"title": "Julia: A Fresh Approach To Numerical Computing", "abstract": "The Julia programming language is gaining enormous popularity. Julia was designed to be easy and fast. Most importantly, Julia shatters deeply established notions widely held in the applied community: $\\begin{array}{l} \\mbox{1. High-level, dynamic code has to be slow by some sort of law of nature. } \\\\ \\mbox{2. It is sensible to prototype in one language and then recode in another language.} \\\\ \\mbox {3. There are parts of a system for the programmer, and other parts best left untouched as they are built by the experts.} \\end{array}$ Julia began with a deep understanding of the needs of the scientific programmer and the needs of the computer in mind. Bridging cultures that have often been distant, Julia combines expertise from computer science and computational science creating a new approach to scientific computing. This note introduces the programmer to the language and the underlying design theory. It invites the reader to rethink the fundamental foundations of numerical computing systems. In particular, there is the fascinating dance between specialization and abstraction. Specialization allows for custom treatment. We can pick just the right algorithm for the right circumstance and this can happen at runtime based on argument types (code selection via multiple dispatch). Abstraction recognizes what remains the same after differences are stripped away and ignored as irrelevant. The recognition of abstraction allows for code reuse (generic programming). A simple idea that yields incredible power. The Julia design facilitates this interplay in many explicit and subtle ways for machine performance and, most importantly, human convenience."}, "keywords": ["optimization functions", "Julia"], "citation_intent": "method"}
{"citing_id": "2303.13506v1", "cited_id": "2001.08361", "section_title": "The Distribution Over Per-Token Losses", "citation": "The parameter scaling exponent \u03b1 N is measured to be \u2248 0.083 from the first six points along the curve (the seventh model appears to break the trend), roughly similar to the parameter scaling exponent measured in #REFR .", "text_before_citation": ["This is roughly in line with the parameter scaling exponent of 0.076 measured in Kaplan et al. 2020 1 .", "The 6.4b model does not fit the scaling curve well, so we excluded its loss when measuring the scaling exponent.", "Next, we plot the probability distribution over per-token losses p(L).", "We find that losses close to zero are by far the most common, and that scaling increases the portion of approximately-zero losses.", "We also plot Figure 3 : Top left: Scaling of mean test loss w.r.t. non-embedding parameters for the Eleuther Pythia models."], "text_after_citation": ["Top center: the distribution p(L) over losses on individual tokens for models of different size.", "Token losses \u2248 0 are by far the most common, and larger models achieve \u2248 0 loss on an increasing fraction of tokens.", "Top right: the expected loss integrand L \u2022 p(L) for models of different sizes.", "Despite their very high prevalence, low-loss tokens contribute minimal mass to the mean loss, which is instead dominated by tokens with much higher loss of 5-10 bits (depending on scale). Bottom left: Training curves (scaling w.r.t. steps S) of mean test loss for Pythia models. We measure exponents \u03b1 S between 0.037 and 0.06. Bottom center: the distribution p(L) over time.", "Over time, models achieve \u2248 0 loss on an increasing fraction of tokens, similar to scaling in model size. Bottom right:"], "citing_paper_content": {"title": "The Quantization Model Of Neural Scaling", "abstract": "We propose the Quantization Model of neural scaling laws, explaining both the observed power law dropoff of loss with model and data size, and also the sudden emergence of new capabilities with scale. We derive this model from what we call the Quantization Hypothesis, where learned network capabilities are quantized into discrete chunks (quanta). We show that when quanta are learned in order of decreasing use frequency, then a power law in use frequencies explains observed power law scaling of loss. We validate this prediction on toy datasets, then study how scaling curves decompose for large language models. Using language model internals, we auto-discover diverse model capabilities (quanta) and find tentative evidence that the distribution over corresponding subproblems in the prediction of natural text is compatible with the power law predicted from the neural scaling exponent as predicted from our theory."}, "cited_paper_content": {"title": "Scaling Laws For Neural Language Models", "abstract": "We study empirical scaling laws for language model performance on the cross-entropy loss. The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude. Other architectural details such as network width or depth have minimal effects within a wide range. Simple equations govern the dependence of overfitting on model/dataset size and the dependence of training speed on model size. These relationships allow us to determine the optimal allocation of a fixed compute budget. Larger models are significantly more sample-efficient, such that optimally compute-efficient training involves training very large models on a relatively modest amount of data and stopping significantly before convergence."}, "keywords": ["parameter scaling exponent"], "citation_intent": "result"}
{"citing_id": "2304.01768v1", "cited_id": "1310.7991", "section_title": "Attainable Accuracy:", "citation": "Comparison to existing results for MOD: Finally, we can compare our result with that in #REFR for MOD with an \u2113 1 -minimisation based sparse approximation routine.", "text_before_citation": ["The main reason is that (13) is a very light condition which does not exclude the existence and rare selection of sparse supports, leading to ill-conditioned or even rank-deficient matrices \u03a8 I .", "Such supports cannot be recovered by thresholding or any other sparse approximation algorithm.", "This failure probability stops the algorithm from attaining arbitrarily small precision.", "If we exchange (13) by the more restrictive assumption 2S\u00b5(\u03a6) < 1 then all supports of size S are well enough conditioned to be identified by a sparse algorithm like OMP using \u03a6 (or some very small perturbation of it).", "This means that \u03a6 is a fixed point of the algorithm and that for a close enough initialisation we have convergence to \u03a6 in expectation."], "text_after_citation": ["Assuming a signal model with uniformly selected random supports and a non-zero coefficient distribution, which is bounded from above, it is shown that for an approximately tight dictionary with coherence \u00b5(\u03a6) 1/ \u221a d and a sparsity level S d 1/6 MOD will converge from any initialisation satisfying \u03b5(\u03a8, \u03a6) 1/S 2 , given enough training signals.", "In this special case our assumptions on the generating dictionary -\u00b5(\u03a6) 1/ log K and S d/ log(K) -are more relaxed.", "As already seen, this comes at the price of a theoretical limit on the achievable accuracy, which in practice, however, is determined by the number of available signals.", "Further, we also have lighter conditions on the distance between the initialisation and the generating dictionary.", "To see this, note that for a tight dictionary our conditions can be written as"], "citing_paper_content": {"title": "Convergence Of Alternating Minimisation Algorithms For Dictionary Learning", "abstract": "In this paper we derive sufficient conditions for the convergence of two popular alternating minimisation algorithms for dictionary learning-the Method of Optimal Directions (MOD) and Online Dictionary Learning (ODL), which can also be thought of as approximative K-SVD. We show that given a well-behaved initialisation that is either within distance at most 1/ log(K) to the generating dictionary or has a special structure ensuring that each element of the initialisation only points to one generating element, both algorithms will converge with geometric convergence rate to the generating dictionary. This is done even for data models with non-uniform distributions on the supports of the sparse coefficients. These allow the appearance frequency of the dictionary elements to vary heavily and thus model real data more closely."}, "cited_paper_content": {"title": "Learning Sparsely Used Overcomplete Dictionaries Via Alternating Minimization", "abstract": "We consider the problem of sparse coding, where each sample consists of a sparse linear combination of a set of dictionary atoms, and the task is to learn both the dictionary elements and the mixing coefficients. Alternating minimization is a popular heuristic for sparse coding, where the dictionary and the coefficients are estimated in alternate steps, keeping the other fixed. Typically, the coefficients are estimated via $\\ell_1$ minimization, keeping the dictionary fixed, and the dictionary is estimated through least squares, keeping the coefficients fixed. In this paper, we establish local linear convergence for this variant of alternating minimization and establish that the basin of attraction for the global optimum (corresponding to the true dictionary and the coefficients) is $\\mathcal{O}(1/s^{2})$, where $s$ is the sparsity level in each sample and the dictionary satisfies restricted isometry property. Combined with the recent results of approximate dictionary estimation, this yields provable guar..."}, "keywords": ["based sparse approximation"], "citation_intent": "result"}
{"citing_id": "2305.01657v1", "cited_id": "1607.00133", "section_title": "Privacy Aspects", "citation": "For example, differential privacy (DP) can be implemented, in a similar concept to DP in training ML models #REFR .", "text_before_citation": ["Sharing the deep features across clients can potentially increase the risk for membership inference attacks, meaning attacks that determine if a data point (i.e., a medical image) is part of the training process.", "However, membership inference attacks may not be a significant concern in such a healthcare setting, as two institutions would need the same or a very similar data point of a patient, and the maximum information gained would be that a patient visited both institutions.", "More concerning would be reconstruction attacks, in which SL clients could reconstruct training data points (e.g., medical images) from the data.", "Such reconstruction attacks are difficult here, since only the last output layer of the neural network in an inference with training data points is shared. Contemporary attacks require access to more layers #OTHEREFR .", "Still, there could be ways for additional protection and guarantees against such hypothetically possible attack scenarios."], "text_after_citation": ["Thereby, before clients submit the vectors of the last output layer of each training data point to the servers, they can add Gaussian noise to the vectors.", "The standard deviation of the noise could depend on the standard deviation of each vector dimension, and a scaling factor which controls the level of privacy protection. The mean of the noise is 0. Additionally, a clipping can be implemented. We leave this implementation to future research.", "An interesting question for future research would be the impact of DP on the SV estimation accuracy, and how it affects SVs (e.g., SVs become more uniform with more noise)."], "citing_paper_content": {"title": "Scalable Data Point Valuation In Decentralized Learning", "abstract": "Existing research on data valuation in federated and swarm learning focuses on valuing client contributions and works best when data across clients is independent and identically distributed (IID). In practice, data is rarely distributed IID. We develop an approach called DDVal for decentralized data valuation, capable of valuing individual data points in federated and swarm learning. DDVal is based on sharing deep features and approximating Shapley values through a k-nearest neighbor approximation method. This allows for novel applications, for example, to simultaneously reward institutions and individuals for providing data to a decentralized machine learning task. The valuation of data points through DDVal allows to also draw hierarchical conclusions on the contribution of institutions, and we empirically show that the accuracy of DDVal in estimating institutional contributions is higher than existing Shapley value approximation methods for federated learning. Specifically, it reaches a cosine similarity in approximating Shapley values of 99.969 % in both, IID and non-IID data distributions across institutions, compared with 99.301 % and 97.250 % for the best state of the art methods. DDVal scales with the number of data points instead of the number of clients, and has a loglinear complexity. This scales more favorably than existing approaches with an exponential complexity. We show that DDVal is especially efficient in data distribution scenarios with many clients that have few data points-for example, more than 16 clients with 8,000 data points each. By integrating DDVal into a decentralized system, we show that it is not only suitable for centralized federated learning, but also decentralized swarm learning, which aligns well with the research on emerging internet technologies such as web3 to reward users for providing data to algorithms."}, "cited_paper_content": {"title": "Deep Learning With Differential Privacy", "abstract": "Machine learning techniques based on neural networks are achieving remarkable results in a wide variety of domains. Often, the training of models requires large, representative datasets, which may be crowdsourced and contain sensitive information. The models should not expose private information in these datasets. Addressing this goal, we develop new algorithmic techniques for learning and a refined analysis of privacy costs within the framework of differential privacy. Our implementation and experiments demonstrate that we can train deep neural networks with non-convex objectives, under a modest privacy budget, and at a manageable cost in software complexity, training efficiency, and model quality."}, "keywords": ["differential privacy"], "citation_intent": "background"}
{"citing_id": "2304.09471v1", "cited_id": "1902.03604", "section_title": "Monocular Multi-Object Tracking", "citation": "Meanwhile, #REFR thought that bounding box level tracking is saturating and introduced pixel-level tracking as a way to improve performance.", "text_before_citation": ["Another work #OTHEREFR focused on improving tracking performance with simple motion prediction using a Kalman filter to predict object location in the next frame.", "While this approach is fast and effective in simple environments, it has limitations in complex scenarios with a high number of objects to track.", "Other studies, such as #OTHEREFR , focused on using appearance information to match objects, by extracting feature vectors from detected objects and comparing them across frames.", "Many studies #OTHEREFR have proposed various approaches to enhance Multi-Object Tracking (MOT) performance with low computational cost.", "#OTHEREFR addressed the issue of the lack of a tracking dataset compared to the detection dataset."], "text_after_citation": ["Other studies, such as #OTHEREFR , focused on learning instance embeddings using 2D point cloud representations to avoid background features.", "#OTHEREFR proposed a real-time Joint Detection and Embedding (JDE) system for MOT, which outperformed the Separate Detection and Embedding (SDE) system used in previous studies.", "#OTHEREFR treated MOT as a multi-task learning problem of object detection and asso- 1Single-Camera Tracking is first conducted with a standard tracking-by-detection scheme, where the extracted frames are fed into a detector and a feature extractor to get the detections and re-id features for the preliminary tracking.", "(2) Then Anchor-Guided Clustering will assign a global ID as well as fixing the ID switches for each trajectory.", "(3) Lastly, Spatio-Temporal Consistency ID Reassignment will utilize the 2D human pose with camera self-calibration to reproject on the map for final post-processing."], "citing_paper_content": {"title": "Enhancing Multi-Camera People Tracking With Anchor-Guided Clustering And Spatio-Temporal Consistency Id Re-Assignment", "abstract": "Multi-camera multiple people tracking has become an increasingly important area of research due to the growing demand for accurate and efficient indoor people tracking systems, particularly in settings such as retail, healthcare centers, and transit hubs. We proposed a novel multi-camera multiple people tracking method that uses anchor-guided clustering for cross-camera re-identification and spatio-temporal consistency for geometry-based crosscamera ID reassigning. Our approach aims to improve the accuracy of tracking by identifying key features that are unique to every individual and utilizing the overlap of views between cameras to predict accurate trajectories without needing the actual camera parameters. The method has demonstrated robustness and effectiveness in handling both synthetic and real-world data. The proposed method is evaluated on CVPR AI City Challenge 2023 dataset, achieving IDF1 of 95.36% with the first-place ranking in the challenge. The code is available at: https://github. com/ipl-uw/AIC23_Track1_UWIPL_ETRI. * indicates equal contributions."}, "cited_paper_content": {"title": "Mots: Multi-Object Tracking And Segmentation", "abstract": "This paper extends the popular task of multi-object tracking to multi-object tracking and segmentation (MOTS). Towards this goal, we create dense pixel-level annotations for two existing tracking datasets using a semi-automatic annotation procedure. Our new annotations comprise 65,213 pixel masks for 977 distinct objects (cars and pedestrians) in 10,870 video frames. For evaluation, we extend existing multi-object tracking metrics to this new task. Moreover, we propose a new baseline method which jointly addresses detection, tracking, and segmentation with a single convolutional network. We demonstrate the value of our datasets by achieving improvements in performance when training on MOTS annotations. We believe that our datasets, metrics and baseline will become a valuable resource towards developing multi-object tracking approaches that go beyond 2D bounding boxes. We make our annotations, code, and models available at this https URL."}, "keywords": ["tracking"], "citation_intent": "background"}
{"citing_id": "2304.08589v1", "cited_id": "2002.11005", "section_title": "Iii. Strategy And Main Results", "citation": "Knowing the previous switching time t \u03c4 \u22121 and the parameters for the next stage, i.e., k \u03c4 +1 and \u03b2 \u03c4 +1 , we follow the steps of #REFR to determine the optimal switching times in theory, and by this to analyze possible gains of the proposed scheme.", "text_before_citation": ["To optimize the speed of convergence, we group iterations into multiple stages \u03c4 \u2208 N, where each stage is characterized by a certain number k \u03c4 of workers to wait for and a certain scale parameter \u03b2 \u03c4 for the batch size.", "We declare a new stage upon changing k and/or \u03b2.", "The time when switching from stage \u03c4 to \u03c4 +1 is denoted by t \u03c4 , where t 0 := 0.", "Consequently, the duration of a stage is t \u03c4 \u2212 t \u03c4 \u22121 ."], "text_after_citation": ["In contrast to #OTHEREFR , the order statistics and the error floors now also depend on \u03b2. The result is given in Theorem 2.", "For ease of notation, we omit the dependencies on k \u03c4 and \u03b2 \u03c4 and refer to the order statistics \u00b5 k\u03c4 :n (\u03b2 \u03c4 ) as \u00b5 \u03c4 .", "We define \u03c6 \u03c4 := k \u03c4 \u03b2 \u03c4 to reflect the batch size in stage \u03c4 . Theorem 2.", "Given t \u03c4 \u22121 , the switching time t \u03c4 is given by", "t \u03c4 = t \u03c4 \u22121 + \u00b5 \u03c4 \u03b1 \u2212 log(\u00b5 \u03c4 \u03b7L\u03c3 2 grad (\u03c6 \u03c4 +1 \u2212 \u03c6 \u03c4 )) + log(\u00b5 \u03c4 +1 \u2212\u00b5 \u03c4 ) + log(\u03c6 \u03c4 +1 (2c\u03c6 \u03c4 se(t \u03c4 \u22121 )\u2212\u03b7L\u03c3 2 grad )) ,"], "citing_paper_content": {"title": "Fast And Straggler-Tolerant Distributed Sgd With Reduced Computation Load", "abstract": "In distributed machine learning, a central node outsources computationally expensive calculations to external worker nodes. The properties of optimization procedures like stochastic gradient descent (SGD) can be leveraged to mitigate the effect of unresponsive or slow workers called stragglers, that otherwise degrade the benefit of outsourcing the computation. This can be done by only waiting for a subset of the workers to finish their computation at each iteration of the algorithm. Previous works proposed to adapt the number of workers to wait for as the algorithm evolves to optimize the speed of convergence. In contrast, we model the communication and computation times using independent random variables. Considering this model, we construct a novel scheme that adapts both the number of workers and the computation load throughout the run-time of the algorithm. Consequently, we improve the convergence speed of distributed SGD while significantly reducing the computation load, at the expense of a slight increase in communication load. This project has received funding from the German Research Foundation (DFG) under Grant Agreement Nos. BI 2492/1-1 and WA 3907/7-1. 1 This holds true if the workers' data is generated from the same distribution. 2 Readers who are not familiar with GD algorithms are referred to Section II."}, "cited_paper_content": {"title": "Adaptive Distributed Stochastic Gradient Descent For Minimizing Delay In The Presence Of Stragglers", "abstract": "We consider the setting where a master wants to run a distributed stochastic gradient descent (SGD) algorithm on $n$ workers each having a subset of the data. Distributed SGD may suffer from the effect of stragglers, i.e., slow or unresponsive workers who cause delays. One solution studied in the literature is to wait at each iteration for the responses of the fastest $k<n$ workers before updating the model, where $k$ is a fixed parameter. The choice of the value of $k$ presents a trade-off between the runtime (i.e., convergence rate) of SGD and the error of the model. Towards optimizing the error-runtime trade-off, we investigate distributed SGD with adaptive $k$. We first design an adaptive policy for varying $k$ that optimizes this trade-off based on an upper bound on the error as a function of the wall-clock time which we derive. Then, we propose an algorithm for adaptive distributed SGD that is based on a statistical heuristic. We implement our algorithm and provide numerical simulations which confirm our intuition and theoretical analysis."}, "keywords": ["optimal switching times"], "citation_intent": "method"}
{"citing_id": "2303.04887v1", "cited_id": "1603.05027", "section_title": "Experimental Setups", "citation": "Using Pre-Activation of ResNet-20 (PreResNet-20) #REFR as an example, we show the relation of the memory cost of each block between width-wise and depth-wise training in Table 1 .", "text_before_citation": ["To simulate the non-IID setting with class imbalance, we follow #OTHEREFR to distribute each class to clients using the Dirichlet distribution with \u03b1(\u03bb), where \u03bb = {0.3, 1.0}.", "Besides, we adopt pathological non-IID data partition \u03b2(\u039b) that is used by the selected baselines -Het-eroFL and SplitMix #OTHEREFR , where each device has unique \u039b labels with \u039b = {2, 5} for CIFAR-10 while \u039b = {10, 30} for CIFAR-100.", "We note that the balanced data partition is applied by default, which makes each client holds the same number of examples.", "The unbalanced \u03b1u(\u03bb) non-IID, where clients may have a different amount of samples with different feature distribution skew, is also used to evaluate the stability of FEDEPTH.", "We consider 100 clients in all experiments, and in Figure 7 2. Memory budgets."], "text_after_citation": ["We can see that if clients afford to train #OTHEREFR 6width PreResNet-20, they can train the full-size neural network via depth-wise training. B4 \u2192 B5,6 \u2192 B7,8,9 }.", "Inspired by this example, we simulate three memory budget scenarios.", "\u2022 (Fair).", "The memory budgets depend on the hidden channel shrinkage ratio, r = { 1 6 , 1 3 , 1 2 , 1}, and are uniformly distributed into clients.", "It means 1/4 of clients can train a PreResNet-20 model within a maximal width of #OTHEREFR 6 -, 1 3 -, 1 2and full width, respectively."], "citing_paper_content": {"title": "Memory-Adaptive Depth-Wise Heterogenous Federated Learning", "abstract": "Federated learning is a promising paradigm that allows multiple clients to collaboratively train a model without sharing the local data. However, the presence of heterogeneous devices in federated learning, such as mobile phones and IoT devices with varying memory capabilities, would limit the scale and hence the performance of the model could be trained. The mainstream approaches to address memory limitations focus on width-slimming techniques, where different clients train subnetworks with reduced widths locally and then the server aggregates the subnetworks. The global model produced from these methods suffers from performance degradation due to the negative impact of the actions taken to handle the varying subnetwork widths in the aggregation phase. In this paper, we introduce a memory-adaptive depth-wise learning solution in FL called FEDEPTH, which adaptively decomposes the full model into blocks according to the memory budgets of each client and trains blocks sequentially to obtain a full inference model. Our method outperforms state-of-the-art approaches, achieving 5% and more than 10% improvements in top-1 accuracy on CIFAR-10 and CIFAR-100, respectively. We also demonstrate the effectiveness of depth-wise fine-tuning on ViT."}, "cited_paper_content": {"title": "Identity Mappings In Deep Residual Networks", "abstract": "Deep residual networks have emerged as a family of extremely deep architectures showing compelling accuracy and nice convergence behaviors. In this paper, we analyze the propagation formulations behind the residual building blocks, which suggest that the forward and backward signals can be directly propagated from one block to any other block, when using identity mappings as the skip connections and after-addition activation. A series of ablation experiments support the importance of these identity mappings. This motivates us to propose a new residual unit, which makes training easier and improves generalization. We report improved results using a 1001-layer ResNet on CIFAR-10 (4.62 % error) and CIFAR-100, and a 200-layer ResNet on ImageNet. Code is available at: https://github.com/KaimingHe/resnet-1k-layers."}, "keywords": ["depth-wise training"], "citation_intent": "method"}
{"citing_id": "2304.05459v1", "cited_id": "1911.07750", "section_title": "Results", "citation": "Regarding ProbLog2, its performance in LUBM010 turned out to be too low to be further considered, as also observed by #REFR .", "text_before_citation": ["For LUBM, we present a comparison between LTGs and all the other engines.", "For DBpedia, Claros, Smokers, YAGO, and WN18RR, the comparison does not consider ProbLog2 and Scallop."], "text_after_citation": ["Regarding Scallop, either it did not support some rules in the benchmarks or the data was too large to be loaded (Scallop's authors' highlight scalability as a direction for future work #OTHEREFR ).", "For VQAR, we show a comparison only between LTGs w/ and Scallop: neither LTGs w/o nor vProbLog were able to compute the least parameterized model due to the combinatorial explosion of the derivations.", "LTGs is the only technique that can compute the full least parameterized model despite this explosion.", "Notice that computing the full lineage does not necessarily mean that we can always compute its exact probability, as the problem is #P-hard #OTHEREFR .", "To deal with such cases, approximations can be employed either upfront, by reducing the size of the lineage (Scallop) or after the full lineage has been collected like #OTHEREFR ."], "citing_paper_content": {"title": "Probabilistic Reasoning At Scale: Trigger Graphs To The Rescue", "abstract": "The role of uncertainty in data management has become more prominent than ever before, especially because of the growing importance of machine learning-driven applications that produce large uncertain databases. A well-known approach to querying such databases is to blend rule-based reasoning with uncertainty. However, techniques proposed so far struggle with large databases. In this paper, we address this problem by presenting a new technique for probabilistic reasoning that exploits Trigger Graphs (TGs)-a notion recently introduced for the non-probabilistic setting. The intuition is that TGs can effectively store a probabilistic model by avoiding an explicit materialization of the lineage and by grouping together similar derivations of the same fact. Firstly, we show how TGs can be adapted to support the possible world semantics. Then, we describe techniques for efficiently computing a probabilistic model, and formally establish the correctness of our approach. We also present an extensive empirical evaluation using a prototype called LTGs. Our comparison against other leading engines shows that LTGs is not only faster, even against approximate reasoning techniques, but can also reason over probabilistic databases that existing engines cannot scale to."}, "cited_paper_content": {"title": "Beyond The Grounding Bottleneck: Datalog Techniques For Inference In Probabilistic Logic Programs (Technical Report)", "abstract": "State-of-the-art inference approaches in probabilistic logic programming typically start by computing the relevant ground program with respect to the queries of interest, and then use this program for probabilistic inference using knowledge compilation and weighted model counting. We propose an alternative approach that uses efficient Datalog techniques to integrate knowledge compilation with forward reasoning with a non-ground program. This effectively eliminates the grounding bottleneck that so far has prohibited the application of probabilistic logic programming in query answering scenarios over knowledge graphs, while also providing fast approximations on classical benchmarks in the field."}, "keywords": ["ProbLog2"], "citation_intent": "result"}
{"citing_id": "2304.07963v1", "cited_id": "1605.09763", "section_title": "Summary", "citation": "Second, there are many other forms of the convection term as shown in #REFR , whose theoretical and computational properties can be studied.", "text_before_citation": ["The numerical simulations show that the fully implicit scheme is more stable and accurate and efficient than the semi-implicit form.", "Note it is the limit scheme, the one achieved when the iterative projections are convergent in each time step, that endows this method the properties such as stability, accuracy, and robustness to Reynolds numbers.", "Thus, there are many ways this work can be improved and extended.", "First, the scheme proposed in this work does not use any stabilization techniques mentioned in #OTHEREFR .", "A simple and appealing stabilization is the grad-div method as mentioned in #OTHEREFR , also used in the augmented Lagrangian method #OTHEREFR , which has been shown to achieve robust error estimates for non-div-free FEMs."], "text_after_citation": ["Third, the convergence rate of the projection iterations for (4) and (5) is first order, and its comparison in efficiency with Newton's method with GMRES approach of solving the nonlinear system as in #OTHEREFR will be an interesting future work.", "Finally, this iterative projection method is a completely new and efficient numerical method for the saddle point problem from the discretization of the time dependent Navier-Stokes equations.", "It is an extension of the Uzawa method by adding only one Poisson equation in the pressure space, which is from the Hodge decomposition of the intermediate velocity.", "The same as the Uzawa method, the method has a linear convergence rate.", "However, normal mode analysis and numerical experiments reveal that with proper parameters, the iterative projection method is much faster than the Uzawa method when the viscosity is small or the Reynolds number is large."], "citing_paper_content": {"title": "Iterative Projection Method For Unsteady Navier-Stokes Equations With High Reynolds Numbers", "abstract": "A new approach, iteration projection method, is proposed to solve the saddle point problem obtained after the full discretization of the unsteady Navier-Stokes equations. It is motivated by the Uzawa method, an iterative method for saddle point problems, and the convectional projection method for the Navier-Stokes equations that projects the intermediate velocity to the divergence free space only once per time step. The proposed method iterates projections in each time step with a proper convection form. We prove that the projection iterations converge with a certain parameter regime. Optimal iteration convergence can be achieved with the modulation of parameter values. This new method has several significant improvements over the Uzawa method and the projection method both theoretically and practically. First, when the iterative projections are fully convergent in each time step, the numerical velocity is weakly divergence free (pointwise divergence free in divergence free finite element methods), and the stability and error estimate are rigorously proven. With proper parameters, this method converges much faster than the Uzawa method. Second, numerical simulations show that with rather relaxed stopping criteria which require only a few iterations each time step, the numerical solution preserves stability and accuracy for high Reynolds numbers, where the convectional projection method would fail. Furthermore, this method retains the efficiency of the traditional projection method by decoupling the velocity and pressure fields, which splits the saddle point system into small elliptic problems. Three dimensional simulations with Taylor-Hood P2/P1 finite elements are presented to demonstrate the performance and efficiency of this method. More importantly, this method is a generic approach and thus there are many potential improvements and extensions of the iterative projection method, including utilization of various convection forms, association with stabilization techniques for high Reynolds numbers, applications to other saddle point problems."}, "cited_paper_content": {"title": "On Conservation Laws Of Navier-Stokes Galerkin Discretizations", "abstract": "We study conservation properties of Galerkin methods for the incompressible Navier\u2013Stokes equations, without the divergence constraint strongly enforced. In typical discretizations such as the mixed finite element method, the conservation of mass is enforced only weakly, and this leads to discrete solutions which may not conserve energy, momentum, angular momentum, helicity, or vorticity, even though the physics of the Navier\u2013Stokes equations dictate that they should. We aim in this work to construct discrete formulations that conserve as many physical laws as possible without utilizing a strong enforcement of the divergence constraint, and doing so leads us to a new formulation that conserves each of energy, momentum, angular momentum, enstrophy in 2D, helicity and vorticity (for reference, the usual convective formulation does not conserve most of these quantities). Several numerical experiments are performed, which verify the theory and test the new formulation."}, "keywords": ["convection term"], "citation_intent": "background"}
{"citing_id": "2304.08275v1", "cited_id": "1802.01029", "section_title": "G. Fairness Vs. Transparency", "citation": "This is particularly important where end users are accountable for the decisions made using an AI/ML system #REFR .", "text_before_citation": ["Through increasing transparency of an AI/ML system it may become apparent to the designers or the end users that the system is unfair #OTHEREFR , thereby possibly leading to internal or external pressure to increase fairness.", "Furthermore, if the designers of an AI/ML system are required to demonstrate that the system is fair, an effective means for doing so is to increase transparency (and possibly explainability).", "The appropriate definition of fairness for an AI/ML system will depend on the context of its use #OTHEREFR .", "Disclosing how designers have understood and implemented fairness will assist the end users of an AI/ML system in evaluating whether it is likely to produce fair outcomes."], "text_after_citation": ["A possible tension between fairness and transparency is the potential for malicious or mischievous actors to exploit an AI/ML system to their own advantage based on the information disclosed about the system [34] , #OTHEREFR .", "In such cases, transparency may need to be limited to protect the fairness of the system.", "The designers should consider to whom the AI/ML system should be transparent, and what information should be provided #OTHEREFR .", "A similar tension is created if aspects of an AI/ML system are protected as trade secrets #OTHEREFR .", "To maintain accountability where public transparency is impossible, the AI/ML system may be made transparent to a trusted auditor for review #OTHEREFR (cf. Section III-J)."], "citing_paper_content": {"title": "Implementing Responsible Ai: Tensions And Trade-Offs Between Ethics Aspects", "abstract": "Many sets of ethics principles for responsible AI have been proposed to allay concerns about misuse and abuse of AI/ML systems. The underlying aspects of such sets of principles include privacy, accuracy, fairness, robustness, explainability, and transparency. However, there are potential tensions between these aspects that pose difficulties for AI/ML developers seeking to follow these principles. For example, increasing the accuracy of an AI/ML system may reduce its explainability. As part of the ongoing effort to operationalise the principles into practice, in this work we compile and discuss a catalogue of 10 notable tensions, trade-offs and other interactions between the underlying aspects. We primarily focus on two-sided interactions, drawing on support spread across a diverse literature. This catalogue can be helpful in raising awareness of the possible interactions between aspects of ethics principles, as well as facilitating well-supported judgements by the designers and developers of AI/ML systems."}, "cited_paper_content": {"title": "Fairness And Accountability Design Needs For Algorithmic Support In High-Stakes Public Sector Decision-Making", "abstract": "Calls for heightened consideration of fairness and accountability in algorithmically-informed public decisions-like taxation, justice, and child protection-are now commonplace. How might designers support such human values? We interviewed 27 public sector machine learning practitioners across 5 OECD countries regarding challenges understanding and imbuing public values into their work. The results suggest a disconnect between organisational and institutional realities, constraints and needs, and those addressed by current research into usable, transparent and 'discrimination-aware' machine learning-absences likely to undermine practical initiatives unless addressed. We see design opportunities in this disconnect, such as in supporting the tracking of concept drift in secondary data sources, and in building usable transparency tools to identify risks and incorporate domain knowledge, aimed both at managers and at the 'street-level bureaucrats' on the frontlines of public service. We conclude by outlining ethical challenges and future directions for collaboration in these high-stakes applications."}, "keywords": ["AI/ML system"], "citation_intent": "method"}
{"citing_id": "2304.06966v1", "cited_id": "1806.01260", "section_title": "The Final Loss Is Calculated Using A Linear Combination Of Lp And Ls As Follows:", "citation": "Looking at the results, we were able to replicate the performance metrics as mentioned by the author #REFR with an error threshold of 0.5% . A sample prediction is displayed in Fig. 6 .", "text_before_citation": ["Losses are then back propagated to the trainable parameters which are a combination of ResNet-50 encoder, Depth Decoder and Pose Decoder models. The optimizer used is Adam with a StepLR scheduler.", "We don't use losses from depth maps to train the network.", "The depth metrics -rms, squared relative, log_rms, a1, a2, a3 are as proposed #OTHEREFR .", "where, Next, coming to the hyperparameters, we used a batch size of 12, lr=1e-4, a lr scheduler with steps every 15 batches for a total of 20 epochs. The training completed in approximately 8 hrs."], "text_after_citation": [], "citing_paper_content": {"title": "Self-Supervised Learning Based Depth Estimation From Monocular Images", "abstract": "Depth Estimation has wide reaching applications in the field of Computer vision such as target tracking, augmented reality, and self-driving cars. The goal of Monocular Depth Estimation is to predict the depth map, given a 2D monocular RGB image as input. The traditional depth estimation methods are based on depth cues and used concepts like epipolar geometry. With the evolution of Convolutional Neural Networks, depth estimation has undergone tremendous strides. In this project, our aim is to explore possible extensions to existing SoTA Deep Learning based Depth Estimation Models and to see whether performance metrics could be further improved. In a broader sense, we are looking at the possibility of implementing Pose Estimation, Efficient Sub-Pixel Convolution Interpolation, Semantic Segmentation Estimation techniques to further enhance our proposed architecture and to provide fine-grained and more globally coherent depth map predictions. We also plan to do away with camera intrinsic parameters during training and apply weather augmentations to further generalize our model."}, "cited_paper_content": {"title": "Digging Into Self-Supervised Monocular Depth Estimation", "abstract": "Per-pixel ground-truth depth data is challenging to acquire at scale. To overcome this limitation, self-supervised learning has emerged as a promising alternative for training models to perform monocular depth estimation. In this paper, we propose a set of improvements, which together result in both quantitatively and qualitatively improved depth maps compared to competing self-supervised methods. Research on self-supervised monocular training usually explores increasingly complex architectures, loss functions, and image formation models, all of which have recently helped to close the gap with fully-supervised methods. We show that a surprisingly simple model, and associated design choices, lead to superior predictions. In particular, we propose (i) a minimum reprojection loss, designed to robustly handle occlusions, (ii) a full-resolution multi-scale sampling method that reduces visual artifacts, and (iii) an auto-masking loss to ignore training pixels that violate camera motion assumptions. We demonstrate the effectiveness of each component in isolation, and show high quality, state-of-the-art results on the KITTI benchmark."}, "keywords": ["sample prediction"], "citation_intent": "result"}
{"citing_id": "2303.10971v1", "cited_id": "1612.08012", "section_title": "Multimodal Medical Shape Data", "citation": "To this end, we use the real-world LUNA16 dataset #REFR , which provides chest CTscans with corresponding lung segmentation mask.", "text_before_citation": ["To demonstrate the potential for real-world applications, we conduct an experiment for multimodal matching in the context of medical image analysis."], "text_after_citation": ["Based on the provided segmentation masks, we reconstruct 3D lung shapes represented as triangle meshes and simulate partial point clouds using a subset of slices of the volumetric images.", "Since there are no ground-truth correspondences available, we restrict ourselves to qualitative results, which are shown in Fig.", "7 , where it can be seen that reliable correspondences between different lung shapes can be obtained from our method."], "citing_paper_content": {"title": "Self-Supervised Learning For Multimodal Non-Rigid 3D Shape Matching", "abstract": "The matching of 3D shapes has been extensively studied for shapes represented as surface meshes, as well as for shapes represented as point clouds. While point clouds are a common representation of raw real-world 3D data (e.g. from laser scanners), meshes encode rich and expressive topological information, but their creation typically requires some form of (often manual) curation. In turn, methods that purely rely on point clouds are unable to meet the matching quality of mesh-based methods that utilise the additional topological structure. In this work we close this gap by introducing a self-supervised multimodal learning strategy that combines mesh-based functional map regularisation with a contrastive loss that couples mesh and point cloud data. Our shape matching approach allows to obtain intramodal correspondences for triangle meshes, complete point clouds, and partially observed point clouds, as well as correspondences across these data modalities. We demonstrate that our method achieves state-of-the-art results on several challenging benchmark datasets even in comparison to recent supervised methods, and that our method reaches previously unseen cross-dataset generalisation ability. Our code is available at https://github.com/ dongliangcao/Self-Supervised-Multimodal-Shape-Matching."}, "cited_paper_content": {"title": "Validation, Comparison, And Combination Of Algorithms For Automatic Detection Of Pulmonary Nodules In Computed Tomography Images: The Luna16 Challenge", "abstract": "Automatic detection of pulmonary nodules in thoracic computed tomography (CT) scans has been an active area of research for the last two decades. However, there have only been few studies that provide a comparative performance evaluation of different systems on a common database. We have therefore set up the LUNA16 challenge, an objective evaluation framework for automatic nodule detection algorithms using the largest publicly available reference database of chest CT scans, the LIDC-IDRI data set. In LUNA16, participants develop their algorithm and upload their predictions on 888 CT scans in one of the two tracks: 1) the complete nodule detection track where a complete CAD system should be developed, or 2) the false positive reduction track where a provided set of nodule candidates should be classified. This paper describes the setup of LUNA16 and presents the results of the challenge so far. Moreover, the impact of combining individual systems on the detection performance was also investigated. It was observed that the leading solutions employed convolutional networks and used the provided set of nodule candidates. The combination of these solutions achieved an excellent sensitivity of over 95% at fewer than 1.0 false positives per scan. This highlights the potential of combining algorithms to improve the detection performance. Our observer study with four expert readers has shown that the best system detects nodules that were missed by expert readers who originally annotated the LIDC-IDRI data. We released this set of additional nodules for further development of CAD systems."}, "keywords": ["corresponding lung segmentation", "chest CTscans"], "citation_intent": "method"}
{"citing_id": "2303.04226v1", "cited_id": "1909.03193", "section_title": "Text Graph Generation.", "citation": "KG-BERT #REFR utilizes the power of pre-trained language models to capture more contextualized information during KGC.", "text_before_citation": ["There are emerging works bridging highly structured molecule graph with language descriptions, which facilitates human understanding of profound molecular knowledge and novel molecule exploration.", "In the following, we briefly overview some representative works in these four topics.", "Text To Knowledge Graph Generation. Li et al.", "#OTHEREFR treat text-to-KG construction as a process of knowledge graph completion (KGC), where missing terms are progressively covered by inference.", "A bilinear model and another DNN-based model are adopted to embed terms and compute score of arbitrary tuples for additive operation."], "text_after_citation": ["The idea is to represent triplets as textual sequences and models graph completion as a sequence classification problem by finetuned BERT model. Malaviya et al.", "#OTHEREFR propose an approach incorporating graph convolutional network (GCN) for to extract more structural and semantic context.", "It also tackles graph sparsity and scalability issues by introducing graph augmentation and progressive masking strategies.", "Alternatively, another line of works #OTHEREFR directly query pre-trained language models to obtain a semantic knowledge network.", "Specifically, language models are repeatedly prompted to predict the masked terms in cloze sentence to acquire relational knowledge."], "citing_paper_content": {"title": "A Comprehensive Survey Of Ai-Generated Content (Aigc): A History Of Generative Ai From Gan To Chatgpt", "abstract": "Recently, ChatGPT, along with DALL-E-2 [1] and Codex [2],has been gaining significant attention from society. As a result, many individuals have become interested in related resources and are seeking to uncover the background and secrets behind its impressive performance. In fact, ChatGPT and other Generative AI (GAI) techniques belong to the category of Artificial Intelligence Generated Content (AIGC), which involves the creation of digital content, such as images, music, and natural language, through AI models. The goal of AIGC is to make the content creation process more efficient and accessible, allowing for the production of high-quality content at a faster pace. AIGC is achieved by extracting and understanding intent information from instructions provided by human, and generating the content according to its knowledge and the intent information. In recent years, large-scale models have become increasingly important in AIGC as they provide better intent extraction and thus, improved generation results. With the growth of data and the size of the models, the distribution that the model can learn becomes more comprehensive and closer to reality, leading to more realistic and high-quality content generation. This survey provides a comprehensive review on the history of generative models, and basic components, recent advances in AIGC from unimodal interaction and multimodal interaction. From the perspective of unimodality, we introduce the generation tasks and relative models of text and image. From the perspective of multimodality, we introduce the cross-application between the modalities mentioned above. Finally, we discuss the existing open problems and future challenges in AIGC."}, "cited_paper_content": {"title": "Kg-Bert: Bert For Knowledge Graph Completion", "abstract": "Knowledge graphs are important resources for many artificial intelligence tasks but often suffer from incompleteness. In this work, we propose to use pre-trained language models for knowledge graph completion. We treat triples in knowledge graphs as textual sequences and propose a novel framework named Knowledge Graph Bidirectional Encoder Representations from Transformer (KG-BERT) to model these triples. Our method takes entity and relation descriptions of a triple as input and computes scoring function of the triple with the KG-BERT language model. Experimental results on multiple benchmark knowledge graphs show that our method can achieve state-of-the-art performance in triple classification, link prediction and relation prediction tasks."}, "keywords": ["contextualized information", "pre-trained language models"], "citation_intent": "method"}
{"citing_id": "2304.06371v2", "cited_id": "1808.06226", "section_title": "Id Text Preprocessing Rbleu", "citation": "In our study, we utilize the SentencePiece #REFR tokenizer and experiment with dictionary sizes of 1000, 4000, and 7000 sub-words to evaluate their respective impacts on NMT performance.", "text_before_citation": ["(1) no 0.62 #OTHEREFR yes 0.98 Table 7 . Validation scores during the exploration of the model architecture.", "As previously discussed in Section 3.2, the selection of appropriate vocabulary size is a trade-off between enhancing the representation of rare words or producing shorter sequences."], "text_after_citation": ["Results in Table 6 show that larger dictionary improves results, in our experiments it increases 0.14 points of reduced BLEU score.", "Thus, we always use a Sentencepiece tokenizer with a vocabulary of 7000 sub-words.", "A current observation in Transformers is that increasing the number of parameters will improve the performance.", "However, in low-resource languages, increasing the number of model parameters can hinder performance #OTHEREFR .", "We study the effect of using a deeper and shallower Transformer by changing the number of layers in the encoder and decoder, the number of attention heads, the feed-forward layer dimension, and embedding dimensions."], "citing_paper_content": {"title": "Sign Language Translation From Instructional Videos", "abstract": "The advances in automatic sign language translation (SLT) to spoken languages have been mostly benchmarked with datasets of limited size and restricted domains. Our work advances the state of the art by providing the first baseline results on How2Sign, a large and broad dataset. We train a Transformer over I3D video features, using the reduced BLEU as a reference metric for validation, instead of the widely used BLEU score. We report a result of 8.03 on the BLEU score, and publish the first open-source implementation of its kind to promote further advances."}, "cited_paper_content": {"title": "Sentencepiece: A Simple And Language Independent Subword Tokenizer And Detokenizer For Neural Text Processing", "abstract": "This paper describes SentencePiece, a language-independent subword tokenizer and detokenizer designed for Neural-based text processing, including Neural Machine Translation. It provides open-source C++ and Python implementations for subword units. While existing subword segmentation tools assume that the input is pre-tokenized into word sequences, SentencePiece can train subword models directly from raw sentences, which allows us to make a purely end-to-end and language independent system. We perform a validation experiment of NMT on English-Japanese machine translation, and find that it is possible to achieve comparable accuracy to direct subword training from raw sentences. We also compare the performance of subword training and segmentation with various configurations. SentencePiece is available under the Apache 2 license at https://github.com/google/sentencepiece."}, "keywords": ["words", "SentencePiece tokenizer"], "citation_intent": "method"}
{"citing_id": "2303.12733v1", "cited_id": "1702.08734", "section_title": "Multi-Modal Search Results", "citation": "Figure 2 also indicates that indices similar to those in #REFR , using a PCA as a first compression and IVFPQ as a quantizer, have some drop off in performance, as with the CLIP only indices.", "text_before_citation": ["Figure 2 presents the retrieval benchmarks for the L400M and L2B datasets.", "We assessed each method across a range of benchmarks and created Pareto fronts by varying the number of chunks for product quantization (i.e., M = 2, 4, 8) and the n probe parameter (\u03c4 = 1, 2, 4).", "Notably, MSE and SNIP performed similarly on image similarity search, while CLIP-only loss networks consistently performed worse.", "However, as mentioned, the CLIP networks (both CLIP and SNIP) do perform better on multi-modal tasks.", "Thus, SNIP can be seen as a best of both worlds descriptor for these types of indices."], "text_after_citation": [], "citing_paper_content": {"title": "On The De-Duplication Of Laion-2B", "abstract": "Generative models, such as DALL-E, Midjourney, and Stable Diffusion, have societal implications that extend beyond the field of computer science. These models require large image databases like LAION-2B, which contain two billion images. At this scale, manual inspection is difficult and automated analysis is challenging. In addition, recent studies show that duplicated images pose copyright problems for models trained on LAION2B, which hinders its usability. This paper proposes an algorithmic chain that runs with modest compute, that compresses CLIP features to enable efficient duplicate detection, even for vast image volumes. Our approach demonstrates that roughly 700 million, or about a third of LAION-2B's images, are duplicates. Our method also provides the histograms of duplication on this dataset, which we use to reveal more examples of verbatim copies by Stable Diffusion and further justify the approach. The current version of the de-duplicated set can be found at github.com/ryanwebster90/snip-dedup."}, "cited_paper_content": {"title": "Billion-Scale Similarity Search With Gpus", "abstract": "Similarity search finds application in specialized database systems handling complex data such as images or videos, which are typically represented by high-dimensional features and require specific indexing structures. This paper tackles the problem of better utilizing GPUs for this task. While GPUs excel at data-parallel tasks, prior approaches are bottlenecked by algorithms that expose less parallelism, such as k-min selection, or make poor use of the memory hierarchy. ::: We propose a design for k-selection that operates at up to 55% of theoretical peak performance, enabling a nearest neighbor implementation that is 8.5x faster than prior GPU state of the art. We apply it in different similarity search scenarios, by proposing optimized design for brute-force, approximate and compressed-domain search based on product quantization. In all these setups, we outperform the state of the art by large margins. Our implementation enables the construction of a high accuracy k-NN graph on 95 million images from the Yfcc100M dataset in 35 minutes, and of a graph connecting 1 billion vectors in less than 12 hours on 4 Maxwell Titan X GPUs. We have open-sourced our approach for the sake of comparison and reproducibility."}, "keywords": ["quantizer"], "citation_intent": "result"}
{"citing_id": "2304.12821v1", "cited_id": "1712.05889", "section_title": "A. Environmental Setup", "citation": "We use Ray #REFR to implement parallel sampling and use 16 workers with different seeds to sample experiences in parallel across four scenarios. Each scenario is assigned 4 workers.", "text_before_citation": ["The maximum steps for one episode are T and when an agent survives T steps in the environment, we call it \"timeout\".", "An agent is marked as a success only when it passes through the interaction zone 1 before the timeout occurs.", "Training Framework: We use PyTorch 1.9.1 2 to implement policy architectures and use Adam #OTHEREFR to optimize policies. All models are trained with one Nvidia RTX 3090.", "Detailed parameters of Soft-Actor-Critic (SAC) are shown in Table I .", "We use precisely the same parameters for both singleand multi-agent policy learning."], "text_after_citation": [], "citing_paper_content": {"title": "Zero-Shot Transfer Learning Of Driving Policy Via Socially Adversarial Traffic Flow", "abstract": "Acquiring driving policies that can transfer to unseen environments is challenging when driving in dense traffic flows. The design of traffic flow is essential and previous studies are unable to balance interaction and safety-criticism. To tackle this problem, we propose a socially adversarial traffic flow. We propose a Contextual Partially-Observable Stochastic Game to model traffic flow and assign Social Value Orientation (SVO) as context. We then adopt a two-stage framework. In Stage 1, each agent in our socially-aware traffic flow is driven by a hierarchical policy where upper-level policy communicates genuine SVOs of all agents, which the lower-level policy takes as input. In Stage 2, each agent in the socially adversarial traffic flow is driven by the hierarchical policy where upper-level communicates mistaken SVOs, taken by the lower-level policy trained in Stage 1. Driving policy is adversarially trained through a zero-sum game formulation with upper-level policies, resulting in a policy with enhanced zero-shot transfer capability to unseen traffic flows. Comprehensive experiments on cross-validation verify the superior zero-shot transfer performance of our method."}, "cited_paper_content": {"title": "Ray: A Distributed Framework For Emerging Ai Applications", "abstract": "The next generation of AI applications will continuously interact with the environment and learn from these interactions. These applications impose new and demanding systems requirements, both in terms of performance and flexibility. In this paper, we consider these requirements and present Ray--a distributed system to address them. Ray implements a unified interface that can express both task-parallel and actor-based computations, supported by a single dynamic execution engine. To meet the performance requirements, Ray employs a distributed scheduler and a distributed and fault-tolerant store to manage the system's control state. In our experiments, we demonstrate scaling beyond 1.8 million tasks per second and better performance than existing specialized systems for several challenging reinforcement learning applications."}, "keywords": ["scenario", "parallel"], "citation_intent": "method"}
{"citing_id": "2304.10611v1", "cited_id": "1910.13461", "section_title": "Experiments", "citation": "For baseline, we use pretrained and finetuned until convergence BART model #REFR with cross entropy loss.", "text_before_citation": ["First, we evaluate the performance of SentenceBERT-based post-processing method. Then, we compared different aspects of unlikelihood training objectives.", "Lastly, we show the efficacy of our novel block loss term in the setting of sequence-level unlikelihood objective."], "text_after_citation": ["Token-level unlikelihood objective is applied during fine-tuning of pretrained BART model.", "Sequence-level unlikelihood objective is applied 50% of the time during fine-tuning of pretrained BART model, with the other 50% of the time using basic likelihood loss.", "Teacher forcing #OTHEREFR is applied during training to ensure a faster and more stable convergence.", "We use beam search as the decoding strategy with beam size 5 across all experiments.", "All training is done on an Azure ML Compute instance with 4 Nvidia V100 GPUs."], "citing_paper_content": {"title": "Multi-Aspect Repetition Suppression And Content Moderation Of Large Language Models Authors", "abstract": "Natural language generation is one of the most impactful fields in NLP, and recent years have witnessed its evolution brought about by large language models (LLMs). As the key instrument for writing assistance applications, they are generally prone to replicating or extending offensive content provided in the input. In low-resource data regime, they can also lead to repetitive outputs (Holtzman et al., 2019) [1]. Usually, offensive content and repetitions are mitigated with post-hoc methods, including n-gram level blocklists, top-k and nucleus sampling. In this paper, we introduce a combination of exact and nonexact repetition suppression using token and sequence level unlikelihood loss, repetition penalty during training, inference, and post-processing respectively. We further explore multi-level unlikelihood loss to the extent that it endows the model with abilities to avoid generating offensive words and phrases from the beginning. Finally, with comprehensive experiments, we demonstrate that our proposed methods work exceptionally in controlling the repetition and content quality of LLM outputs."}, "cited_paper_content": {"title": "Bart: Denoising Sequence-To-Sequence Pre-Training For Natural Language Generation, Translation, And Comprehension", "abstract": "We present BART, a denoising autoencoder for pretraining sequence-to-sequence models. BART is trained by (1) corrupting text with an arbitrary noising function, and (2) learning a model to reconstruct the original text. It uses a standard Tranformer-based neural machine translation architecture which, despite its simplicity, can be seen as generalizing BERT (due to the bidirectional encoder), GPT (with the left-to-right decoder), and many other more recent pretraining schemes. We evaluate a number of noising approaches, finding the best performance by both randomly shuffling the order of the original sentences and using a novel in-filling scheme, where spans of text are replaced with a single mask token. BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE. BART also provides a 1.1 BLEU increase over a back-translation system for machine translation, with only target language pretraining. We also report ablation experiments that replicate other pretraining schemes within the BART framework, to better measure which factors most influence end-task performance."}, "keywords": ["cross entropy loss", "baseline"], "citation_intent": "method"}
{"citing_id": "2304.13130v1", "cited_id": "1910.10683", "section_title": "Implementation", "citation": "For our learned hypernymization approach, we use the T5-small #REFR language model, which we fine-tune for 5 epochs on one Google Cloud TPU with learning rate of 0.0001, batch size of 12, and gradient accumulation step of 8.", "text_before_citation": ["To perform named entity recognition, we use the off-the-shelf Flair tagger #OTHEREFR , available from Hugging Face #OTHEREFR ."], "text_after_citation": ["Given the affinity of caption reconstruction with summarization, we add the \"summarize:\" prompt to the beginning of each input caption.", "To create the artificial, NE-enriched captions, we set = 0.7, \u2212 = = 0.3 (Section 3.2).", "For pre-training, we adapt the code from #OTHEREFR with a ResNet50 #OTHEREFR as the visual encoder and a frozen BERT-base model #OTHEREFR as the text encoder.", "We set the learning rate to 0.001 (decreased by a factor of 10 at 50% and 80% of training).", "We set the batch size to 9 and we train using 3 Nvidia Titan X GPUs."], "citing_paper_content": {"title": "Hypernymization Of Named Entity-Rich Captions For Grounding-Based Multi-Modal Pretraining", "abstract": "Figure 1: Our key ideas: an object mentioned using a named entity cannot be well grounded with an image. We thus introduce two methods to carry out hypernymization on the caption and show better grounding performance between the image and the hypernymized captions."}, "cited_paper_content": {"title": "Exploring The Limits Of Transfer Learning With A Unified Text-To-Text Transformer", "abstract": "Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new \"Colossal Clean Crawled Corpus\", we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code."}, "keywords": ["learned hypernymization approach", "T5-small language model"], "citation_intent": "method"}
{"citing_id": "2303.00154v1", "cited_id": "1312.6034", "section_title": "Inference Of Yarn Characteristics From Input Images", "citation": "We considered the saliency maps #REFR of networks trained to predict the set of raw yarn's parameters and the set of the flyaway parameters with two independent models.", "text_before_citation": ["We model the prediction of the parameters for our procedural yarn model from images as a regression problem.", "Our training and validation dataset consists of annotated synthetic yarn images that we use to train a model that allows inferring yarn parameters from novel images of yarns not seen during training or validation.", "Before providing details of our respective approach (see Section 4.1), we motivate our choice of a suitable network architecture that is capable of handling the challenging nature of the underlying problem."], "text_after_citation": ["An entry m i, j in a saliency map for a model f that has been trained on a subset of P parameters is the maximum derivative of the average value of the predicted parameters with respect to a pixel x i, j,c in the input image over the color channels c, i.e.", "EQUATION", "In contrast to saliency maps that have been proposed within the context of classification networks, we consider the derivative of the mean of the predicted parameters because we need to investigate the effect of a pixel on the entire subset on which the network was trained.", "The saliency maps (Figure 7 ) for the network trained to predict the raw yarn parameters illustrate a higher susceptibility to changes in the yarn center region of the image.", "On the other hand, the saliency maps for the network that predicts the flyaway parameters indicate that such a network exhibits a higher sensitivity towards the border regions within the input images."], "citing_paper_content": {"title": "Neural Inverse Procedural Modeling Of Knitting Yarns From Images", "abstract": "We investigate the capabilities of neural inverse procedural modeling to infer high-quality procedural yarn models with fiber-level details from single images of depicted yarn samples. While directly inferring all parameters of the underlying yarn model based on a single neural network may seem an intuitive choice, we show that the complexity of yarn structures in terms of twisting and migration characteristics of the involved fibers can be better encountered in terms of ensembles of networks that focus on individual characteristics. We analyze the effect of different loss functions including a parameter loss to penalize the deviation of inferred parameters to ground truth annotations, a reconstruction loss to enforce similar statistics of the image generated for the estimated parameters in comparison to training images as well as an additional regularization term to explicitly penalize deviations between latent codes of synthetic images and the average latent code of real images in the encoder's latent space. We demonstrate that the combination of a carefully designed parametric, procedural yarn model with respective network ensembles as well as loss functions even allows robust parameter inference when solely trained on synthetic data. Since our approach relies on the availability of a yarn database with parameter annotations and we are not aware of such a respectively available dataset, we additionally provide, to the best of our knowledge, the first dataset of yarn images with annotations regarding the respective yarn parameters. For this purpose, we use a novel yarn generator that improves the realism of the produced results over previous approaches."}, "cited_paper_content": {"title": "Deep Inside Convolutional Networks: Visualising Image Classification Models And Saliency Maps", "abstract": "This paper addresses the visualisation of image classification models, learnt using deep Convolutional Networks (ConvNets). We consider two visualisation techniques, based on computing the gradient of the class score with respect to the input image. The first one generates an image, which maximises the class score [5], thus visualising the notion of the class, captured by a ConvNet. The second technique computes a class saliency map, specific to a given image and class. We show that such maps can be employed for weakly supervised object segmentation using classification ConvNets. Finally, we establish the connection between the gradient-based ConvNet visualisation methods and deconvolutional networks [13]."}, "keywords": ["raw yarn's parameters", "saliency maps"], "citation_intent": "method"}
{"citing_id": "2303.10411v1", "cited_id": "1712.00726", "section_title": "Introduction", "citation": "It was followed by a series of improvements, such as Cascade R-CNN #REFR , etc.", "text_before_citation": ["\u2022 We show the impact of MSIL on regression and classification by using relevant information through visualization.", "And experiments on two datasets verify that MSIL can leverage the relevant information between semantics to improves the performance of different methods.", "#OTHEREFR further improves the efficiency by simplifying the pooling operation with RoI pool #OTHEREFR , and integrating the regressor into the neural network.", "Faster R-CNN #OTHEREFR is the paradigm of the R-CNN series, which proposed the Region Proposal Network (RPN) based on subsequent improvements.", "Discarding the previous selective search #OTHEREFR algorithm, RPN uses neural networks to generate candidate boxes."], "text_after_citation": ["It is worth mentioning that the R-CNN series has always used one branch to predict position and class.", "Despite various adjustments made by subsequent improvements, regression and classification have always been coupled together. Later, Double-Head #OTHEREFR discussed in this regard.", "YoloV1 #OTHEREFR is the first single-stage detector, which divides the image into individual lattices, and then predicts all lattices at once.", "It is worth mentioning that YoloV1 only uses points near the center of the object to predict edges.", "Because these points are thought to produce higher quality detections. However, this strategy leads to low recall of YoloV1."], "citing_paper_content": {"title": "Multi-Semantic Interactive Learning For Object Detection", "abstract": "Single-branch object detection methods use shared features for localization and classification, yet the shared features are not fit for the two different tasks simultaneously. Multi-branch object detection methods usually use different features for localization and classification separately, ignoring the relevance between different tasks. Therefore, we propose multi-semantic interactive learning (MSIL) to mine the semantic relevance between different branches and extract multi-semantic enhanced features of objects. MSIL first performs semantic alignment of regression and classification branches, then merges the features of different branches by semantic fusion, finally extracts relevant information by semantic separation and passes it back to the regression and classification branches respectively. More importantly, MSIL can be integrated into existing object detection nets as a plug-and-play component. Experiments on the MS COCO, and Pascal VOC datasets show that the integration of MSIL with existing algorithms can utilize the relevant information between semantics of different tasks and achieve better performance."}, "cited_paper_content": {"title": "Cascade R-Cnn: Delving Into High Quality Object Detection", "abstract": "In object detection, an intersection over union (IoU) threshold is required to define positives and negatives. An object detector, trained with low IoU threshold, e.g. 0.5, usually produces noisy detections. However, detection performance tends to degrade with increasing the IoU thresholds. Two main factors are responsible for this: 1) overfitting during training, due to exponentially vanishing positive samples, and 2) inference-time mismatch between the IoUs for which the detector is optimal and those of the input hypotheses. A multi-stage object detection architecture, the Cascade R-CNN, is proposed to address these problems. It consists of a sequence of detectors trained with increasing IoU thresholds, to be sequentially more selective against close false positives. The detectors are trained stage by stage, leveraging the observation that the output of a detector is a good distribution for training the next higher quality detector. The resampling of progressively improved hypotheses guarantees that all detectors have a positive set of examples of equivalent size, reducing the overfitting problem. The same cascade procedure is applied at inference, enabling a closer match between the hypotheses and the detector quality of each stage. A simple implementation of the Cascade R-CNN is shown to surpass all single-model object detectors on the challenging COCO dataset. Experiments also show that the Cascade R-CNN is widely applicable across detector architectures, achieving consistent gains independently of the baseline detector strength. The code will be made available at this https URL."}, "keywords": ["Cascade R-CNN"], "citation_intent": "method"}
{"citing_id": "2303.13929v1", "cited_id": "1912.01703", "section_title": "A. Experiment Setup", "citation": "The PPO agent and base controllers, i.e., H \u221e and PIDs, are implemented based on Pytorch #REFR to facilitate vectorized computing.", "text_before_citation": ["We perform all experiments on a single computer (AMD Ryzen Threadripper 3960X, 24x 3.8GHz, NVIDIA GeForce RTX2080 Ti, 11GB)."], "text_after_citation": ["Both TurtleSim and the blimp simulator are implemented based on the ROS and the latter is integrated with the Gazebo SITL simulation.", "We compare our robust H \u221e -RL controller to the previous PID-RL baseline in the TurtleSim (Sec. III-B.1) and blimp simulator (Sec. III-B.2).", "1) H \u221e -PPO agent: our proposed approach 2) PID-PPO agent: previous approach #OTHEREFR .", "We reimplemented with a randomized mixing factor q and a servo control so that it has the chance to challenge our more difficult waypoint following task.", "Note when the mixing factor is q = 1 we recover the base control, and q = 0 corresponds to the pure PPO agent."], "citing_paper_content": {"title": "Autonomous Blimp Control Via H \u221e Robust Deep Residual Reinforcement Learning", "abstract": "Due to their superior energy efficiency, blimps may replace quadcopters for long-duration aerial tasks. However, designing a controller for blimps to handle complex dynamics, modeling errors, and disturbances remains an unsolved challenge. One recent work combines reinforcement learning (RL) and a PID controller to address this challenge and demonstrates its effectiveness in real-world experiments. In the current work, we build on that using an H\u221e robust controller to expand the stability margin and improve the RL agent's performance. Empirical analysis of different mixing methods reveals that the resulting H\u221e-RL controller outperforms the prior PID-RL combination and can handle more complex tasks involving intensive thrust vectoring. We provide our code as open-source at https://github.com/robot-perception-group/ robust_deep_residual_blimp."}, "cited_paper_content": {"title": "Pytorch: An Imperative Style, High-Performance Deep Learning Library", "abstract": "Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. ::: In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. ::: We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks."}, "keywords": ["base controllers", "Pytorch"], "citation_intent": "method"}
{"citing_id": "2305.02507v1", "cited_id": "2003.08237", "section_title": "Effectiveness Of Different Components", "citation": "When we further adjust the validation resize size and train crop size, it can not only get a slightly higher performance gain (3.34% vs 3.13%) but also speed up the training process, similar to the conclusions in #REFR .", "text_before_citation": ["To verify the effectiveness of the proposed three strategies (i.e., KL-+Snet6, SITrans, and InterSST), we progressively add these components to the original ST and compare their main network performance. All experiments are performed with ResNet50 on ImageNet.", "As shown in Table 3 , original ST can already improve the Top-1 classification results of CT by 0.86%.", "When adding KL-+Snet6 and SITrans to ST respectively, the Top-1 performance improvement increases to 1.83% and 1.86% respectively.", "When adding both KL-+Snet6 and SITrans to ST, the Top-1 performance improvement is up to 3.13%."], "text_after_citation": ["Then, when we add the above strategies and InterSSR to ST, the Top-1 performance gain is boosted to 3.80%.", "In addition, the proposed stimulative training and three simple-but-effective strategies are complementary to other training ingredients in essence.", "To verify this, on the basis of the above strategies, when we further add a data augmentation strategy called uniform augment #OTHEREFR and extend the training epoch to 400, the Top-1 performance gain continues to increase to 4.27%.", "In summary, the proposed stimulative training and its associated three strategies, although simple, can remarkably improve the performance of residual networks and serve as a strong booster for the current training scheme."], "citing_paper_content": {"title": "Stimulative Training++: Go Beyond The Performance Limits Of Residual Networks", "abstract": "Residual networks have shown great success and become indispensable in recent deep neural network models. In this work, we aim to re-investigate the training process of residual networks from a novel social psychology perspective of loafing, and further propose a new training scheme as well as three improved strategies for boosting residual networks beyond their performance limits. Previous research has suggested that residual networks can be considered as ensembles of shallow networks, which implies that the final performance of a residual network is influenced by a group of subnetworks. We identify a previously overlooked problem that is analogous to social loafing, where subnetworks within a residual network are prone to exert less effort when working as part of a group compared to working alone. We define this problem as network loafing. Similar to the decreased individual productivity and overall performance as demonstrated in society, network loafing inevitably causes sub-par performance. Inspired by solutions from social psychology, we first propose a novel training scheme called stimulative training, which randomly samples a residual subnetwork and calculates the KL divergence loss between the sampled subnetwork and the given residual network for extra supervision. In order to unleash the potential of stimulative training, we further propose three simple-yet-effective strategies, including a novel KL-loss that only aligns the network logits direction, random smaller inputs for subnetworks, and inter-stage sampling rules. Comprehensive experiments and analysis verify the effectiveness of stimulative training as well as its three improved strategies. For example, the proposed method can boost the performance of ResNet50 on ImageNet to 80.5% Top1 accuracy without using any extra data, model, trick, or changing the structure. With only uniform augment, the performance can be further improved to 81.0% Top1 accuracy, better than the best training recipes provided by Timm library and PyTorch official version. We also verify its superiority on various typical models, datasets, and tasks and give some theoretical analysis. As such, we advocate utilizing the proposed method as a general and next-generation technology to train residual networks. The code is available at https://github.com/Sunshine-Ye/NIPS22-ST."}, "cited_paper_content": {"title": "Fixing The Train-Test Resolution Discrepancy: Fixefficientnet", "abstract": "This note complements the paper\"Fixing the train-test resolution discrepancy\"that introduced the FixRes method. First, we show that this strategy is advantageously combined with recent training recipes from the literature. Most importantly, we provide new results for the EfficientNet architecture. The resulting network, called FixEfficientNet, significantly outperforms the initial architecture with the same number of parameters. For instance, our FixEfficientNet-B0 trained without additional training data achieves 79.3% top-1 accuracy on ImageNet with 5.3M parameters. This is a +0.5% absolute improvement over the Noisy student EfficientNet-B0 trained with 300M unlabeled images and +1.7% compared to the EfficientNet-B0 trained with adversarial examples. An EfficientNet-L2 pre-trained with weak supervision on 300M unlabeled images and further optimized with FixRes achieves 88.5% top-1 accuracy (top-5: 98.7%), which establishes the new state of the art for ImageNet with a single crop."}, "keywords": ["training process", "train crop size"], "citation_intent": "result"}
{"citing_id": "2304.14291v1", "cited_id": "1801.00868", "section_title": "Introduction", "citation": "Panoptic segmentation #REFR of images is a core computer vision task that jointly solves two related problems -semantic segmentation and instance segmentation.", "text_before_citation": [], "text_after_citation": ["With the rise of robotics and emerging autonomous driving markets, efficient visual perception stacks are in high demand.", "However, large-scale supervised learning of panoptic segmentation #OTHEREFR is prohibitively expensive as it requires dense annotations for both semantics and instances, which require time-consuming manual labeling.", "ground truth CVRN PQ=32.1 EDAPS PQ=41.2 Figure 1 : EDAPS is an architecture and a collection of recipes, designed specifically for Domain-Adaptive Panoptic Segmentation.", "It demonstrates a significant improvement over the prior art on the challenging synthetic-to-real benchmarks.", "As shown above, it is better than CVRN #OTHEREFR on SYNTHIA \u2192 Cityscapes by 9 mPQ."], "citing_paper_content": {"title": "Edaps: Enhanced Domain-Adaptive Panoptic Segmentation", "abstract": "With autonomous industries on the rise, domain adaptation of the visual perception stack is an important research direction due to the cost savings promise. Much prior art was dedicated to domain-adaptive semantic segmentation in the synthetic-to-real context. Despite being a crucial output of the perception stack, panoptic segmentation has been largely overlooked by the domain adaptation community. Therefore, we revisit well-performing domain adaptation strategies from other fields, adapt them to panoptic segmentation, and show that they can effectively enhance panoptic domain adaptation. Further, we study the panoptic network design and propose a novel architecture (EDAPS) designed explicitly for domain-adaptive panoptic segmentation. It uses a shared, domain-robust transformer encoder to facilitate the joint adaptation of semantic and instance features, but task-specific decoders tailored for the specific requirements of both domain-adaptive semantic and instance segmentation. As a result, the performance gap seen in challenging panoptic benchmarks is substantially narrowed. EDAPS significantly improves the state-of-the-art performance for panoptic segmentation UDA by a large margin of 25% on SYNTHIA-to-Cityscapes and even 72% on the more challenging SYNTHIA-to-Mapillary Vistas. The implementation is available at https://github.com/susaha/edaps."}, "cited_paper_content": {"title": "Panoptic Segmentation", "abstract": "We propose and study a task we name panoptic segmentation (PS). Panoptic segmentation unifies the typically distinct tasks of semantic segmentation (assign a class label to each pixel) and instance segmentation (detect and segment each object instance). The proposed task requires generating a coherent scene segmentation that is rich and complete, an important step toward real-world vision systems. While early work in computer vision addressed related image/scene parsing tasks, these are not currently popular, possibly due to lack of appropriate metrics or associated recognition challenges. To address this, we propose a novel panoptic quality (PQ) metric that captures performance for all classes (stuff and things) in an interpretable and unified manner. Using the proposed metric, we perform a rigorous study of both human and machine performance for PS on three existing datasets, revealing interesting insights about the task. The aim of our work is to revive the interest of the community in a more unified view of image segmentation. For more analysis and up-to-date results, please check the arXiv version of the paper: {\\small\\url{https://arxiv.org/abs/1801.00868}}."}, "keywords": ["Panoptic segmentation"], "citation_intent": "background"}
{"citing_id": "2303.17755v1", "cited_id": "1808.02639", "section_title": "Seeking Better Weights", "citation": "On the one hand they appear in the component-by-component (CBC) construction for determining the lattice generating vector z gen , which in turn determines the interpolation points through #REFR .", "text_before_citation": ["In implementations of the lattice-based kernel interpolant of #OTHEREFR the weights \u03b3 u play a dual role."], "text_after_citation": ["On the other hand they appear in the formula (10) for the kernel.", "In both roles only special forms of weights are feasible, given that there are 2 s subsets u \u2286 {1 : s}.", "The SPOD weights described above are feasible (and were used in the computations in #OTHEREFR ), but encounter two computational bottlenecks:", "1.", "The CBC construction used to obtain the generating vector z gen using SPOD weights has cost O(s n log(n) + s 3 \u03b1 2 n), see Table 1 and #OTHEREFR . 2."], "citing_paper_content": {"title": "Lattice-Based Kernel Approximation And Serendipitous Weights For Parametric Pdes In Very High Dimensions", "abstract": "We describe a fast method for solving elliptic partial differential equations (PDEs) with uncertain coefficients using kernel interpolation at a lattice point set. By representing the input random field of the system using the model proposed by Kaarnioja, Kuo, and Sloan (SIAM J. Numer. Anal. 2020), in which a countable number of independent random variables enter the random field as periodic functions, it was shown by Kaarnioja, Kazashi, Kuo, Nobile, and Sloan (Numer. Math. 2022) that the lattice-based kernel interpolant can be constructed for the PDE solution as a function of the stochastic variables in a highly efficient manner using fast Fourier transform (FFT). In this work, we discuss the connection between our model and the popular \"affine and uniform model\" studied widely in the literature of uncertainty quantification for PDEs with uncertain coefficients. We also propose a new class of weights entering the construction of the kernel interpolant-serendipitous weights-which dramatically improve the computational performance of the kernel interpolant for PDE problems with uncertain coefficients, and allow us to tackle function approximation problems up to very high dimensionalities. Numerical experiments are presented to showcase the performance of the serendipitous weights."}, "cited_paper_content": {"title": "Analysis Of Quasi-Monte Carlo Methods For Elliptic Eigenvalue Problems With Stochastic Coefficients", "abstract": "We consider the forward problem of uncertainty quantification for the generalised Dirichlet eigenvalue problem for a coercive second order partial differential operator with random coefficients, motivated by problems in structural mechanics, photonic crystals and neutron diffusion. The PDE coefficients are assumed to be uniformly bounded random fields, represented as infinite series parametrised by uniformly distributed i.i.d. random variables. The expectation of the fundamental eigenvalue of this problem is computed by (a) truncating the infinite series which define the coefficients; (b) approximating the resulting truncated problem using lowest order conforming finite elements and a sparse matrix eigenvalue solver; and (c) approximating the resulting finite (but high dimensional) integral by a randomly shifted quasi-Monte Carlo lattice rule, with specially chosen generating vector. We prove error estimates for the combined error, which depend on the truncation dimension s, the finite element mesh diameter h, and the number of quasi-Monte Carlo samples N. Under suitable regularity assumptions, our bounds are of the particular form \\({\\mathcal {O}}(h^2 + N^{-1 + \\delta })\\), where \\(\\delta > 0\\) is arbitrary and the hidden constant is independent of the truncation dimension, which needs to grow as \\(h\\rightarrow 0\\) and \\(N \\rightarrow \\infty \\). As for the analogous PDE source problem, the conditions under which our error bounds hold depend on a parameter \\(p \\in (0, 1)\\) representing the summability of the terms in the series expansions of the coefficients. Although the eigenvalue problem is nonlinear, which means it is generally considered harder than the source problem, in almost all cases (\\(p \\ne 1\\)) we obtain error bounds that converge at the same rate as the corresponding rate for the source problem. The proof involves a detailed study of the regularity of the fundamental eigenvalue as a function of the random parameters. As a key intermediate result in the analysis, we prove that the spectral gap (between the fundamental and the second eigenvalues) is uniformly positive over all realisations of the random problem."}, "keywords": ["interpolation", "lattice"], "citation_intent": "background"}
{"citing_id": "2303.15149v1", "cited_id": "1904.03451", "section_title": "Experiments", "citation": "For category-level object detection, we train on category-level sketch/photo pairs in QuickDraw-Extended #REFR having 330k sketches and 204k photos from 110 categories.", "text_before_citation": ["Dataset We train our object detector using existing crosscategory FG-SBIR dataset -Sketchy #OTHEREFR that contains 125 categories, each with 100 photos.", "Every photo in #OTHEREFR has at least 5 instance-level paired sketches.", "To evaluate finegrained object detection, we use SketchyCOCO #OTHEREFR comprising of natural images in MS-COCO #OTHEREFR with instancelevel paired sketches. Following Liu et al.", "#OTHEREFR , we select 1, 225 sketch/photo pairs from SketchyCOCO #OTHEREFR with at least one foreground sketched object.", "We filter the overlapping categories of in SketchyCOCO #OTHEREFR from Sketchy #OTHEREFR to measure true zero-shot performance."], "text_after_citation": ["Following #OTHEREFR , we evaluate on a subset of standard object detection datasets like PASCAL-VOC #OTHEREFR and MS-COCO #OTHEREFR that have 20 and 56 overlapping categories in QuickDraw #OTHEREFR .", "Implementation Details Our model is implemented in PyTorch on a 11GB Nvidia RTX 2080-Ti GPU.", "First, we train a generalised cross-category FG-SBIR with image size (224\u00d7224) by adapting CLIP with ViT #OTHEREFR backbone (ViT-B/32 weights) using prompt learning #OTHEREFR .", "The prompts (P = 3) are trained with triplet loss #OTHEREFR , margin \u00b5 = 0.3, Adam optimiser with learning rate 1e \u2212 4 for 60 epochs, and batch size 64.", "Our object detection pipeline is build using Detectron2 [93] ."], "citing_paper_content": {"title": "What Can Human Sketches Do For Object Detection?", "abstract": "Sketches are highly expressive, inherently capturing subjective and fine-grained visual cues. The exploration of such innate properties of human sketches has, however, been limited to that of image retrieval. In this paper, for the first time, we cultivate the expressiveness of sketches but for the fundamental vision task of object detection. The end result is a sketch-enabled object detection framework that detects based on what you sketch-that \"zebra\" (e.g., one that is eating the grass) in a herd of zebras (instance-aware detection), and only the part (e.g., \"head\" of a \"zebra\") that you desire (part-aware detection). We further dictate that our model works without (i) knowing which category to expect at testing (zero-shot) and (ii) not requiring additional bounding boxes (as per fully supervised) and class labels (as per weakly supervised). Instead of devising a model from the ground up, we show an intuitive synergy between foundation models (e.g., CLIP) and existing sketch models build for sketch-based image retrieval (SBIR), which can already elegantly solve the task-CLIP to provide model generalisation, and SBIR to bridge the (sketch\u2192photo) gap. In particular, we first perform independent prompting on both sketch and photo branches of an SBIR model to build highly generalisable sketch and photo encoders on the back of the generalisation ability of CLIP. We then devise a training paradigm to adapt the learned encoders for object detection, such that the region embeddings of detected boxes are aligned with the sketch and photo embeddings from SBIR. Evaluating our framework on standard object detection datasets like PASCAL-VOC and MS-COCO outperforms both supervised (SOD) and weakly-supervised object detectors (WSOD) on zero-shot setups. Project"}, "cited_paper_content": {"title": "Doodle To Search: Practical Zero-Shot Sketch-Based Image Retrieval", "abstract": "In this paper, we investigate the problem of zero-shot sketch-based image retrieval (ZS-SBIR), where human sketches are used as queries to conduct retrieval of photos from unseen categories. We importantly advance prior arts by proposing a novel ZS-SBIR scenario that represents a firm step forward in its practical application. The new setting uniquely recognizes two important yet often neglected challenges of practical ZS-SBIR, (i) the large domain gap between amateur sketch and photo, and (ii) the necessity for moving towards large-scale retrieval. We first contribute to the community a novel ZS-SBIR dataset, QuickDraw-Extended, that consists of 330,000 sketches and 204,000 photos spanning across 110 categories. Highly abstract amateur human sketches are purposefully sourced to maximize the domain gap, instead of ones included in existing datasets that can often be semi-photorealistic. We then formulate a ZS-SBIR framework to jointly model sketches and photos into a common embedding space. A novel strategy to mine the mutual information among domains is specifically engineered to alleviate the domain gap. External semantic knowledge is further embedded to aid semantic transfer. We show that, rather surprisingly, retrieval performance significantly outperforms that of state-of-the-art on existing datasets that can already be achieved using a reduced version of our model. We further demonstrate the superior performance of our full model by comparing with a number of alternatives on the newly proposed dataset. The new dataset, plus all training and testing code of our model, will be publicly released to facilitate future research"}, "keywords": ["category-level sketch/photo pairs"], "citation_intent": "method"}
{"citing_id": "2303.17017v1", "cited_id": "1904.04637", "section_title": "Conclusions And Future Work", "citation": "We showed that this problem is coNP-complete, which is the same complexity we obtained in the relational case #REFR .", "text_before_citation": ["We have presented two algorithms that decide the definability problem by quantifier-free first-order formulas over a purely functional language.", "These algorithms rely on the semantic characterization of quantifier-free definability given in #OTHEREFR ."], "text_after_citation": ["Prior to this we developed an algorithm that gives a characterization of the isomorphism type of a tuple\u0101 in the given structure (in particular, the subuniverse it generates is obtained).", "Our first definability-decision procedure (Merging Algorithm) partitions the set of all ktuples from the universe with k \u2208 spec(R), and has to exhaust the search space to give a positive answer.", "Therefore, the execution time of our algorithm depends exponentially on the parameter k.", "The empirical tests confirm the exponential dependence on k, and give us an idea of the impact of the cardinality of the input structure and the family of algebras on which we performed the tests.", "They also confirm the conjecture that this strategy, which is based on the existence of symmetries, would perform better on models with a large amount of subisomorphisms, as we The Splitting algorithm starts with a single block of tuples, and partitions blocks as the determination of the isomorphism types of their tuples progresses."], "citing_paper_content": {"title": "Two Algorithms To Decide Quantifier-Free Definability In Finite Algebraic Structures", "abstract": "This work deals with the definability problem by quantifier-free first-order formulas over a finite algebraic structure. We show the problem to be coNP-complete and present two decision algorithms based on a semantical characterization of definable relations as those preserved by isomorphisms of substructures, the second one also providing a formula in the positive case. Our approach also includes the design of an algorithm that computes the isomorphism type of a tuple in a finite algebraic structure. Proofs of soundness and completeness of the algorithms are presented, as well as empirical tests assessing their performances."}, "cited_paper_content": {"title": "The Complexity Of Definability By Open First-Order Formulas", "abstract": "In this article we formally define and investigate the computational complexity of the Definability Problem for open first-order formulas (i.e., quantifier free first-order formulas) with equality. Given a logic $\\mathbf{\\mathcal{L}}$, the $\\mathbf{\\mathcal{L}}$-Definability Problem for finite structures takes as input a finite structure $\\mathbf{A}$ and a target relation $T$ over the domain of $\\mathbf{A}$, and determines whether there is a formula of $\\mathbf{\\mathcal{L}}$ whose interpretation in $\\mathbf{A}$ coincides with $T$. We show that the complexity of this problem for open first-order formulas (open definability, for short) is coNP-complete. We also investigate the parametric complexity of the problem, and prove that if the size and the arity of the target relation $T$ are taken as parameters then open definability is $\\mathrm{coW}[1]$-complete for every vocabulary $\\tau$ with at least one, at least binary, relation."}, "keywords": ["complexity"], "citation_intent": "result"}
{"citing_id": "2303.03283v1", "cited_id": "1706.03762", "section_title": "Generative Models And Personalization", "citation": "Generative models often implement a Transformer architecture that includes an encoder and a decoder, utilizing an attention mechanism that allows back-references to prior content #REFR .", "text_before_citation": ["In recent years, various generative models have emerged for tasks such as text-to-text #OTHEREFR and text-to-image #OTHEREFR generation, music composition based on prompts #OTHEREFR , and text simplification #OTHEREFR .", "Other models generate images #OTHEREFR , poetry #OTHEREFR , rap lyrics #OTHEREFR , or music #OTHEREFR in the style of previous works."], "text_after_citation": ["These models are typically trained to predict probable next tokens from previous text. Related tasks include infilling masked tokens.", "Thus, they learn connections between subsequent words and are well-suited for text generation.", "In this work, we focus on text generation with GPT-3.", "Released in 2020, GPT-3 currently is a particularly large and powerful language model #OTHEREFR .", "It uses deep learning methods such as attention mechanisms and the Transformer architecture with autoregressive pretraining #OTHEREFR ."], "citing_paper_content": {"title": "The Ai Ghostwriter Effect: Users Do Not Perceive Ownership Of Ai-Generated Text But Self-Declare As Authors", "abstract": "Human-AI interaction in text production increases complexity in authorship. In two empirical studies (n1 = 30 & n2 = 96), we investigate authorship and ownership in human-AI collaboration for personalized language generation models. We show an AI Ghostwriter Effect: Users do not consider themselves the owners and authors of AI-generated text but refrain from publicly declaring AI authorship. The degree of personalization did not impact the AI Ghostwriter Effect, and control over the model increased participants' sense of ownership. We also found that the discrepancy between the sense of ownership and the authorship declaration is stronger in interactions with a human ghostwriter and that people use similar rationalizations for authorship in AI ghostwriters and human ghostwriters. We discuss how our findings relate to psychological ownership and human-AI interaction to lay the foundations for adapting authorship frameworks and user interfaces in AI in text-generation tasks. CCS Concepts: \u2022 Computing methodologies \u2192 Natural language generation; \u2022 Applied computing \u2192 Publishing; Text editing; \u2022 Social and professional topics \u2192 Intellectual property."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["Generative models", "decoder"], "citation_intent": "background"}
{"citing_id": "2303.12733v1", "cited_id": "1702.08734", "section_title": "Image Similarity Search", "citation": "We note that the approach in #REFR also used a similar dimension for the compressed space before quantization.", "text_before_citation": ["We then compared the index retrieval results with this ground truth as follows:", "R \u2229 k := 1 N q\u2208Q |k-NN(q i ) \u2229 k-ANN(q i )| k (R \u2229 k)", "The approximate k-nearest neighbors (k-ANN) were computed for the index under review, while the exhaustive search on raw features (k-NN) served as the ground truth.", "It's worth noting that the approach used to construct the ground truth for the (non-multimodal) Deep1B dataset #OTHEREFR is similar to the one described in #OTHEREFR .", "We did ablations over the latent space dimensions and found that encoding to a relatively high dimension of 128 first (from 1024 for ViT-H-14) was more effective, than for example encoding to a low dimension directly and giving more memory to the quantizer."], "text_after_citation": ["Also like in #OTHEREFR , we used the IVFPQ index for every index we built atop of our descriptors and finally did several comparisons to a vanilla IVFPQ on raw features.", "For the IVF, we chose 2 #OTHEREFR centroids, and for the PQ we constructed collections of indices with different values of M (the number of dimension chunks) with 2 8 centroids each.", "Finally, we highlight the popular open source tool AutoFaiss, which will automatically construct and efficient PQ index given a memory budget.", "All of our AutoFaiss indices use multiple PQ steps, with an initial optimized product quantization (OPQ) step for dimensionality reduction (see #OTHEREFR ), as well as a hierarchical navigable world search (HSNW) #OTHEREFR .", "For instance, when providing the tool with \"6G\" total memory for the index, it produced a faiss construction pipeline of \"OPQ16 112,IVF65536 HNSW32,PQ16x8\" (Please see the faiss library for more details #OTHEREFR )."], "citing_paper_content": {"title": "On The De-Duplication Of Laion-2B", "abstract": "Generative models, such as DALL-E, Midjourney, and Stable Diffusion, have societal implications that extend beyond the field of computer science. These models require large image databases like LAION-2B, which contain two billion images. At this scale, manual inspection is difficult and automated analysis is challenging. In addition, recent studies show that duplicated images pose copyright problems for models trained on LAION2B, which hinders its usability. This paper proposes an algorithmic chain that runs with modest compute, that compresses CLIP features to enable efficient duplicate detection, even for vast image volumes. Our approach demonstrates that roughly 700 million, or about a third of LAION-2B's images, are duplicates. Our method also provides the histograms of duplication on this dataset, which we use to reveal more examples of verbatim copies by Stable Diffusion and further justify the approach. The current version of the de-duplicated set can be found at github.com/ryanwebster90/snip-dedup."}, "cited_paper_content": {"title": "Billion-Scale Similarity Search With Gpus", "abstract": "Similarity search finds application in specialized database systems handling complex data such as images or videos, which are typically represented by high-dimensional features and require specific indexing structures. This paper tackles the problem of better utilizing GPUs for this task. While GPUs excel at data-parallel tasks, prior approaches are bottlenecked by algorithms that expose less parallelism, such as k-min selection, or make poor use of the memory hierarchy. ::: We propose a design for k-selection that operates at up to 55% of theoretical peak performance, enabling a nearest neighbor implementation that is 8.5x faster than prior GPU state of the art. We apply it in different similarity search scenarios, by proposing optimized design for brute-force, approximate and compressed-domain search based on product quantization. In all these setups, we outperform the state of the art by large margins. Our implementation enables the construction of a high accuracy k-NN graph on 95 million images from the Yfcc100M dataset in 35 minutes, and of a graph connecting 1 billion vectors in less than 12 hours on 4 Maxwell Titan X GPUs. We have open-sourced our approach for the sake of comparison and reproducibility."}, "keywords": ["quantization", "similar dimension"], "citation_intent": "method"}
{"citing_id": "2303.09012v1", "cited_id": "1607.08022", "section_title": "Pix2Pix", "citation": "In Pix2Pix, using instance normalization, which involves setting the batch size to 1, is a better choice than batch normalization for training generative models #REFR .", "text_before_citation": ["One benefit of the PatchGAN discriminator is its ability to apply a fixed-size patch discriminator on images of any size, unlike typical discriminators that only score the full image.", "Additionally, a smaller PatchGAN has fewer parameters, runs faster, and can be applied to images of any size.", "Training and optimization: GAN models are notoriously challenging to train and converge due to the simultaneous training of the generator and discriminator, with improvements to one model often coming at the expense of the other.", "The differing learning rates of the two models can also make convergence difficult.", "As suggested in the original GAN paper #OTHEREFR , instead of training the generator G to minimize log(1 \u2212 D(x, G(x, z))), it is more effective to train G to maximize log(D(x, G(x, z))) to improve convergence."], "text_after_citation": [], "citing_paper_content": {"title": "Exploring The Power Of Generative Deep Learning For Image-To-Image Translation And Mri Reconstruction: A Cross-Domain Review", "abstract": "Deep learning has become a prominent computational modeling tool in the areas of computer vision and image processing in recent years. This research comprehensively analyzes the different deep-learning methods used for image-to-image translation and reconstruction in the natural and medical imaging domains. We examine the famous deep learning frameworks, such as convolutional neural networks and generative adversarial networks, and their variants, delving into the fundamental principles and difficulties of each. In the field of natural computer vision, we investigate the development and extension of various deeplearning generative models. In comparison, we investigate the possible applications of deep learning to generative medical imaging problems, including medical image translation, MRI reconstruction, and multi-contrast MRI synthesis. This thorough review provides scholars and practitioners in the areas of generative computer vision and medical imaging with useful insights for summarizing past works and getting insight into future research paths."}, "cited_paper_content": {"title": "Instance Normalization: The Missing Ingredient For Fast Stylization", "abstract": "It this paper we revisit the fast stylization method introduced in Ulyanov et. al. (2016). We show how a small change in the stylization architecture results in a significant qualitative improvement in the generated images. The change is limited to swapping batch normalization with instance normalization, and to apply the latter both at training and testing times. The resulting method can be used to train high-performance architectures for real-time image generation. The code will is made available on github at this https URL. Full paper can be found at arXiv:1701.02096."}, "keywords": ["generative models", "instance normalization"], "citation_intent": "method"}
{"citing_id": "2305.02330v1", "cited_id": "1907.06772", "section_title": "Hotspot Identification Via Fish Detection", "citation": "Next, we also investigate using a fully out-of-sample detector, which we call the MegaFishDetector (MFD), similar to #REFR , but trained using several publicly available fish detection datasets using the same YOLOv5m at 1280p network.", "text_before_citation": ["However, for these detectors to be useful and drastically reduce annotation effort, they must detect fish with similar accuracy on other reefs, or at different times in various conditions.", "We thus provide an initial investigation by deploying the same detector used for the in-sample study, and apply it without modification to a survey on another coral We see that MFD alone is insufficient, and JS alone is still not robust even to small changes in site context.", "JSMF D is slightly more robust to out of sample data, but still more work needs to be done.", "reef.", "We minimize changes in the environment by selecting a reef that is nearby to the original, and survey at a similar time of day and weather conditions, only one day apart."], "text_after_citation": ["This dataset consists of roughly 210K training images and 36K validation images, assembled together from the following datasets: AIMS OzFish #OTHEREFR , FathomNet (only the gnathostomata class) #OTHEREFR , VIAME FishTrack22 (subsampled by 50%) #OTHEREFR , NOAA Puget Sound Nearshore Fish #OTHEREFR , NOAA Labeled Fishes In the Wild #OTHEREFR , and DeepFish (Segmentation only) #OTHEREFR .", "Finally, we compare the ability of using MFD alone or as a pre-trained network that can be used to fine-tune the in-sample network.", "MFD code and models are available at https:// github.com/warplab/megafishdetector."], "citing_paper_content": {"title": "Biological Hotspot Mapping In Coral Reefs With Robotic Visual Surveys", "abstract": "Figure 1. Hotspot detection at two unique coral reef sites in the US Virgin Islands-Joel's Shoal (JS) and Tektite (TK). From visual surveys collected by an AUV, we generate hotspot maps of fish counts along the robot trajectory and plotted with a log-colorscale (a/d). Our fish detector is fine-tuned on labelled images from JS, but not from TK. We can see distinct hotspots at both reefs, correlated with distinct reef structure visible in our 3D reconstruction, an overhanging structure and a pillar coral (b), as well as with rugosity (c), which is a measure of topographic complexity used by the biological community as a proxy for biodiversity and abundance. Our system enables rapid deployment of an AUV over a predefined but unknown area to identify biologically interesting hotspots."}, "cited_paper_content": {"title": "Efficient Pipeline For Camera Trap Image Review", "abstract": "Biologists all over the world use camera traps to monitor biodiversity and wildlife population density. The computer vision community has been making strides towards automating the species classification challenge in camera traps, but it has proven difficult to to apply models trained in one region to images collected in different geographic areas. In some cases, accuracy falls off catastrophically in new region, due to both changes in background and the presence of previously-unseen species. We propose a pipeline that takes advantage of a pre-trained general animal detector and a smaller set of labeled images to train a classification model that can efficiently achieve accurate results in a new region."}, "keywords": ["fish detection datasets"], "citation_intent": "method"}
{"citing_id": "2303.06670v1", "cited_id": "1512.03385", "section_title": "Dino With Different Backbones", "citation": "A deep residual learning framework (ResNet) is proposed to overcome the degradation issue by #REFR .", "text_before_citation": ["In this work, there are four different networks applied as the backbone of DINO, DINO-TP, and DINO-MC, including ViT, Swin Transformer, ResNet, and WRN.", "ViT preserves the Transformer structure used in NLP as much as possible and performs very well in computer vision tasks #OTHEREFR .", "The input of transformer encoder is the embedding formed by adding the patch embedding with the position encoding.", "The transformer encoder of ViT consists mainly of layer normalization (LN) which is applied before each block, multi-head attention, and multi-layer perceptron block (MLP).", "Inspired by ViT, Swin Transformer is proposed with a hierarchical transformer, which computes representation with both regular and shifted windows to capture features from different levels and resolutions."], "text_after_citation": ["They add shortcut connections to transmit the information of shallow layers directly to the deeper layers of the neural network and require no additional computation.", "With this mechanism, the residual net is deepened to 152 layers and achieves stateof-the-art results in multiple computer vision tasks.", "Furthermore, with the development of self-supervised representation learning, ResNet also becomes an effective back-bone widely used in multiple self-supervised architectures #OTHEREFR .", "WRN [49] is proposed to improve the performance of ResNet by increase the width of the residual block, i.e., widening the convolutional layers.", "When the same number of parameters are utilized, wide residual block can get superior outcomes with less training time compared to the original residual block."], "citing_paper_content": {"title": "Dino-Mc: Self-Supervised Contrastive Learning For Remote Sensing Imagery With Multi-Sized Local Crops", "abstract": "Due to the costly nature of remote sensing image labeling and the large volume of available unlabeled imagery, self-supervised methods that can learn feature representations without manual annotation have received great attention. While prior works have explored self-supervised learning in remote sensing tasks, pretext tasks based on local-global view alignment remain underexplored. Inspired by DINO [6], which employs an effective representation learning structure with knowledge distillation based on global-local view alignment, we formulate two pretext tasks for use in self-supervised learning on remote sensing imagery (SSLRS). Using these tasks, we explore the effectiveness of positive temporal contrast as well as multi-sized views on SSLRS. Moreover, we extend DINO and propose DINO-MC which uses local views of various sized crops instead of a single fixed size. Our experiments demonstrate that even when pre-trained on only 10% of the dataset, DINO-MC performs on par or better than existing state of the art SSLRS methods on multiple remote sensing tasks, while using less computational resources. All codes, models and results are available at https://github.com/ WennyXY/DINO-MC."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["deep residual learning"], "citation_intent": "method"}
{"citing_id": "2304.09355v2", "cited_id": "1810.05728", "section_title": "Information In Deterministic Networks", "citation": "Moreover, this discretization depends on the chosen bin size and does not track the mutual information across varying bin sizes #REFR ; Ross (2014).", "text_before_citation": ["Information-theoretic methods have made a significant impact on deep learning #OTHEREFR .", "However, a key challenge is addressing the source of randomness in deterministic DNNs.", "The mutual information between the input and representation is infinite, leading to ill-posed optimization problems or piecewise constant outcomes #OTHEREFR . To tackle this issue, researchers have proposed various solutions.", "One common approach is to discretize the input distribution and real-valued hidden representations by binning, which facilitates non-trivial measurements and prevents the mutual information from always taking the maximum value of the log of the dataset size, thus avoiding ill-posed optimization problems #OTHEREFR .", "However, binning and discretization are essentially equivalent to geometrical compression and serve as clustering measures #OTHEREFR ."], "text_after_citation": ["To address these limitations, researchers have proposed alternative approaches such as interpreting binned information as a weight decay penalty #OTHEREFR", "(2019b) , estimating mutual information based on lower bounds assuming a continuous input distribution without making assumptions about the network's output distribution properties #OTHEREFR , injecting additive noise, and considering data augmentation as the source of noise (Dubois et al., 2021; #OTHEREFR ."], "citing_paper_content": {"title": "To Compress Or Not To Compress -Self-Supervised Learning And Information Theory: A Review", "abstract": "Deep neural networks have demonstrated remarkable performance in supervised learning tasks but require large amounts of labeled data. Self-supervised learning offers an alternative paradigm, enabling the model to learn from data without explicit labels. Information theory has been instrumental in understanding and optimizing deep neural networks. Specifically, the information bottleneck principle has been applied to optimize the trade-off between compression and relevant information preservation in supervised settings. However, the optimal information objective in self-supervised learning remains unclear. In this paper, we review various approaches to self-supervised learning from an information-theoretic standpoint and present a unified framework that formalizes the self-supervised informationtheoretic learning problem. We integrate existing research into a coherent framework, examine recent self-supervised methods, and identify research opportunities and challenges. Moreover, we discuss empirical measurement of information-theoretic quantities and their estimators. This paper offers a comprehensive review of the intersection between information theory, self-supervised learning, and deep neural networks."}, "cited_paper_content": {"title": "Estimating Information Flow In Neural Networks", "abstract": "We study the flow of information and the evolution of internal representations during deep neural network (DNN) training, aiming to demystify the compression aspect of the information bottleneck theory. The theory suggests that DNN training comprises a rapid fitting phase followed by a slower compression phase, in which the mutual information $I(X;T)$ between the input $X$ and internal representations $T$ decreases. Several papers observe compression of estimated mutual information on different DNN models, but the true $I(X;T)$ over these networks is provably either constant (discrete $X$) or infinite (continuous $X$). This work explains the discrepancy between theory and experiments, and clarifies what was actually measured by these past works. To this end, we introduce an auxiliary (noisy) DNN framework for which $I(X;T)$ is a meaningful quantity that depends on the network's parameters. This noisy framework is shown to be a good proxy for the original (deterministic) DNN both in terms of performance and the learned representations. We then develop a rigorous estimator for $I(X;T)$ in noisy DNNs and observe compression in various models. By relating $I(X;T)$ in the noisy DNN to an information-theoretic communication problem, we show that compression is driven by the progressive clustering of hidden representations of inputs from the same class. Several methods to directly monitor clustering of hidden representations, both in noisy and deterministic DNNs, are used to show that meaningful clusters form in the $T$ space. Finally, we return to the estimator of $I(X;T)$ employed in past works, and demonstrate that while it fails to capture the true (vacuous) mutual information, it does serve as a measure for clustering. This clarifies the past observations of compression and isolates the geometric clustering of hidden representations as the true phenomenon of interest."}, "keywords": ["mutual information"], "citation_intent": "background"}
{"citing_id": "2303.13637v1", "cited_id": "1908.01755", "section_title": "Introduction", "citation": "These results corroborate the \"Rashomon\" theory #REFR that, for some problems, there exist simple models with good accuracy and meet special requirements, such as the limited memory size of embedded devices in this case.", "text_before_citation": ["Note that, 300-second HRV monitoring can be used for caring for chronic renal failure and diabetes #OTHEREFR , showing the healthcare potential of PPG-based HRV monitoring.", "The evaluation results also show that our method is significantly more accurate than the signal-processing-only and ML-only methods.", "Moreover, the model exploration showed that DT and MLP models are usually smaller with good accuracy, making them more suitable for small embedded devices.", "DT models have an average error of 13.0% and are usually less than 10KB with inference time less than 10 s.", "MLP models have an average error of 9.1% and are less than 469KB with inference time less than 1ms."], "text_after_citation": ["The contributions of this paper include,", "\u2022 The compound and direct HRV inference methodology combines signal processing and ML to directly infer RMSSD/SDNN to achieve high accuracy with small and fast ML models. Figure 1 .", "An example of ECG waveform showcasing R peaks (red dots) and RR intervals.", "RR interval R R R R R", "\u2022 A systematic exploration of different ML algorithms to study their impact on the accuracy, model size, and time on HRV inference."], "citing_paper_content": {"title": "Efficient And Direct Inference Of Heart Rate Variability Using Both Signal Processing And Machine Learning", "abstract": "Heart Rate Variability (HRV) measures the variation of the time between consecutive heartbeats and is a major indicator of physical and mental health. Recent research has demonstrated that photoplethysmography (PPG) sensors can be used to infer HRV. However, many prior studies had high errors because they only employed signal processing or machine learning (ML), or because they indirectly inferred HRV, or because there lacks large training datasets. Many prior studies may also require large ML models. The low accuracy and large model sizes limit their applications to small embedded devices and potential future use in healthcare. To address the above issues, we first collected a large dataset of PPG signals and HRV ground truth. With this dataset, we developed HRV models that combine signal processing and ML to directly infer HRV. Evaluation results show that our method had errors between 3.5% to 25.7% and outperformed signal-processing-only and ML-only methods. We also explored different ML models, which showed that Decision Trees and Multi-level Perceptrons have 13.0% and 9.1% errors on average with models at most hundreds of KB and inference time less than 1ms. Hence, they are more suitable for small embedded devices and potentially enable the future use of PPG-based HRV monitoring in healthcare."}, "cited_paper_content": {"title": "A Study In Rashomon Curves And Volumes: A New Perspective On Generalization And Model Simplicity In Machine Learning", "abstract": "The Rashomon effect occurs when many different explanations exist for the same phenomenon. In machine learning, Leo Breiman used this term to characterize problems where many accurate-but-different models exist to describe the same data. In this work, we study how the Rashomon effect can be useful for understanding the relationship between training and test performance, and the possibility that simple-yet-accurate models exist for many problems. We consider the Rashomon set - the set of almost-equally-accurate models for a given problem - and study its properties and the types of models it could contain. We present the Rashomon ratio as a new measure related to simplicity of model classes, which is the ratio of the volume of the set of accurate models to the volume of the hypothesis space; the Rashomon ratio is different from standard complexity measures from statistical learning theory. For a hierarchy of hypothesis spaces, the Rashomon ratio can help modelers to navigate the trade-off between simplicity and accuracy. In particular, we find empirically that a plot of empirical risk vs. Rashomon ratio forms a characteristic $\\Gamma$-shaped Rashomon curve, whose elbow seems to be a reliable model selection criterion. When the Rashomon set is large, models that are accurate - but that also have various other useful properties - can often be obtained. These models might obey various constraints such as interpretability, fairness, or monotonicity."}, "keywords": ["embedded devices", "simple models"], "citation_intent": "result"}
{"citing_id": "2304.00698v1", "cited_id": "1903.07293", "section_title": "Discussion", "citation": "For example, if we apply HGPF on HAN #REFR , the running time of HGPF on ACM dataset is about 5 minutes with a single GPU device of GeForce RTX 3090.", "text_before_citation": ["4.4.1 Computational Complexity.", "The time and space complexity of our framework is linear to the scale of an HIN, i.e., the number of nodes, edges, features and meta-path instances.", "In fact, the training of HGPF is very time efficient."], "text_after_citation": [], "citing_paper_content": {"title": "A Post-Training Framework For Improving Heterogeneous Graph Neural Networks", "abstract": "Recent years have witnessed the success of heterogeneous graph neural networks (HGNNs) in modeling heterogeneous information networks (HINs). In this paper, we focus on the benchmark task of HGNNs, i.e., node classification, and empirically find that typical HGNNs are not good at predicting the label of a test node whose receptive field (1) has few training nodes from the same category or (2) has multiple training nodes from different categories. A possible explanation is that their message passing mechanisms may involve noises from different categories, and cannot fully explore task-specific knowledge such as the label dependency between distant nodes. Therefore, instead of introducing a new HGNN model, we propose a general post-training framework that can be applied on any pretrained HGNNs to further inject task-specific knowledge and enhance their prediction performance. Specifically, we first design an auxiliary system that estimates node labels based on (1) a global inference module of multi-channel label propagation and (2) a local inference module of network schema-aware prediction. The mechanism of our auxiliary system can complement the pretrained HGNNs by providing extra task-specific knowledge. During the post-training process, we will strengthen both system-level and module-level consistencies to encourage the cooperation between a pretrained HGNN and our auxiliary system. In this way, both systems can learn from each other for better performance. In experiments, we apply our framework to four typical HGNNs. Experimental results on three benchmark datasets show that compared with pretrained HGNNs, our post-training framework can enhance Micro-F1 by a relative improvement of 3.9% on average. Code, data"}, "cited_paper_content": {"title": "Heterogeneous Graph Attention Network", "abstract": "Graph neural network, as a powerful graph representation technique based on deep learning, has shown superior performance and attracted considerable research interest. However, it has not been fully considered in graph neural network for heterogeneous graph which contains different types of nodes and links. The heterogeneity and rich semantic information bring great challenges for designing a graph neural network for heterogeneous graph. Recently, one of the most exciting advancements in deep learning is the attention mechanism, whose great potential has been well demonstrated in various areas. In this paper, we first propose a novel heterogeneous graph neural network based on the hierarchical attention, including node-level and semantic-level attentions. Specifically, the node-level attention aims to learn the importance between a node and its metapath based neighbors, while the semantic-level attention is able to learn the importance of different meta-paths. With the learned importance from both node-level and semantic-level attention, the importance of node and meta-path can be fully considered. Then the proposed model can generate node embedding by aggregating features from meta-path based neighbors in a hierarchical manner. Extensive experimental results on three real-world heterogeneous graphs not only show the superior performance of our proposed model over the state-of-the-arts, but also demonstrate its potentially good interpretability for graph analysis."}, "keywords": ["HAN"], "citation_intent": "method"}
{"citing_id": "2304.13017v1", "cited_id": "1810.04805", "section_title": "Related Work", "citation": "SSL with Transformers has driven recent advances in NLP and computer vision (e.g. #REFR ; Dosovitskiy et al. (2021)), and has clear potential to advance other fields.", "text_before_citation": ["(2022) uses Transformers instead in an alternating manner, performing local attention operations across spatial and channel dimensions.", "However, these computer vision models create an internal channel dimension without a semantic relation to an input dimension, whereas we design our input embedding and Transformer layers to preserve the relationship of the event dimension to input events. In dense time-series forecasting, TSMixer #OTHEREFR", "(2023) has shown the promise of mixing across dimensions with MLP models.", "Recently, Crossformer Zhang and Yan (2023) has introduced a modified attention function to mix across channels in a dense time-series forecasting Transformer model, but does not apply full quadratic attention or feedforward processing across the channel dimension as our model does.", "SSL has become an important framework to enable learning useful representations from data without relying on labels."], "text_after_citation": ["Different approaches for SSL with numeric EHR data are explored in #OTHEREFR", "(2021) , but these are applied to a relatively basic GRU model and do not claim to achieve state-of-the-art results.", "mTAN uses a similar approach to SSL based on reconstructing inputs, incorporating a variational loss into their training procedure, but without a distinct pre-training stage.", "SSL is used with Transformers in Hi-BEHRT, which applies BYOL #OTHEREFR to augmentations of EHR time series data.", "STraTs uses masked value prediction for SSL with Transformers, while RAPT additionally uses a reasonability check to identify corrupted sequences and a contrastive patient similarity task."], "citing_paper_content": {"title": "Duett: Dual Event Time Transformer For Electronic Health Records", "abstract": "Electronic health records (EHRs) recorded in hospital settings typically contain a wide range of numeric time series data that is characterized by high sparsity and irregular observations. Effective modelling for such data must exploit its time series nature, the semantic relationship between different types of observations, and information in the sparsity structure of the data. Self-supervised Transformers have shown outstanding performance in a variety of structured tasks in NLP and computer vision. But multivariate time series data contains structured relationships over two dimensions: time and recorded event type, and straightforward applications of Transformers to time series data do not leverage this distinct structure. The quadratic scaling of self-attention layers can also significantly limit the input sequence length without appropriate input engineering. We introduce the DuETT architecture, an extension of Transformers designed to attend over both time and event type dimensions, yielding robust representations from EHR data. DuETT uses an aggregated input where sparse time series are transformed into a regular sequence with fixed length; this lowers the computational complexity relative to previous EHR Transformer models and, more importantly, enables the use of larger and deeper neural networks. When trained with self-supervised prediction tasks, that provide rich and informative signals for model pre-training, our model outperforms state-of-the-art deep learning models on multiple downstream tasks from the MIMIC-IV and PhysioNet-2012 EHR datasets."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["NLP"], "citation_intent": "background"}
{"citing_id": "2303.03127v1", "cited_id": "1712.07487", "section_title": "Evaluation On Icfhr 2016 Competition'S Datasets", "citation": "In summary, we note that the ST-KeyS semi-supervised model outperforms the fully supervised PHOC-based methods #REFR that were conducted in the literature on the BOT dataset and yields moderate results with these methods on the AK dataset.", "text_before_citation": ["From Table 3 , we notice that the proposed ST-KeyS model yields good results on the BOT dataset reaching a performance of 59.10% which outperforms TPP-PHOCNet(CPS), PHOCNet(CPS) and Triplet CNN architectures #OTHEREFR .", "However, we obtained moderate results on the AK dataset comparing to TPP-PHOCNet(CPS) and PHOCNet(CPS) architectures #OTHEREFR .", "Moreover, we observe that the ST-KeyS leads to better results than learningfree based methods #OTHEREFR on both the BOT and AK datasets although they represent different calligraphic styles compared to the IAM database utilized in the first pretraining stage and also despite the difference in the written language script.", "Additionally, our ST-KeyS model gives better results compared to the Graph Edit Distance method proposed in #OTHEREFR thanks to the transformer's capacity to focus on meaningful information included in the word image.", "This capacity is owed to the attention mechanism allowing to extract discriminant features from word images and which are used later in the matching process."], "text_after_citation": ["Generally, these methods rely on synthetically generated data to learn the PHOC contextual representations.", "Thus, such models are able to treat large vocabulary allowing to alleviate to the out of vocabulary words issue #OTHEREFR .", "In the remainder of our work, we limit the PHOC learning stage to the data provided in the competition image databases.", "In other words, we are based on a closed vocabulary since our goal is to build a robust model allowing to retrieve handwritten words independently of the out of vocabulary word problem.", "Therefore, it is unfair to compare our results to PHOC-based methods that were conducted in the literature."], "citing_paper_content": {"title": "St-Keys : Self-Supervised Transformer For Keyword Spotting In Historical Handwritten Documents", "abstract": "Keyword spotting (KWS) in historical documents is an important tool for the initial exploration of digitized collections. Nowadays, the most efficient KWS methods are relying on machine learning techniques that require a large amount of annotated training data. However, in the case of historical manuscripts, there is a lack of annotated corpus for training. To handle the data scarcity issue, we investigate the merits of the self-supervised learning to extract useful representations of the input data without relying on human annotations and then using these representations in the downstream task. We propose ST-KeyS, a masked auto-encoder model based on vision transformers where the pretraining stage is based on the mask-and-predict paradigm, without the need of labeled data. In the fine-tuning stage, the pre-trained encoder is integrated into a siamese neural network model that is fine-tuned to improve feature embedding from the input images. We further improve the image representation using pyramidal histogram of characters (PHOC) embedding to create and exploit an intermediate representation of images based on text attributes. In an exhaustive experimental evaluation on three widely used benchmark datasets (Botany, Alvermann Konzilsprotokolle and George Washington), the proposed approach outperforms state-of-the-art methods trained on the same datasets."}, "cited_paper_content": {"title": "Attribute Cnns For Word Spotting In Handwritten Documents", "abstract": "Word spotting has become a field of strong research interest in document image analysis over the last years. Recently, AttributeSVMs were proposed which predict a binary attribute representation. At their time, this influential method defined the state-of-the-art in segmentation-based word spotting. In this work, we present an approach for learning attribute representations with Convolutional Neural Networks (CNNs). By taking a probabilistic perspective on training CNNs, we derive two different loss functions for binary and real-valued word string embeddings. In addition, we propose two different CNN architectures, specifically designed for word spotting. These architectures are able to be trained in an end-to-end fashion. In a number of experiments, we investigate the influence of different word string embeddings and optimization strategies. We show our Attribute CNNs to achieve state-of-the-art results for segmentation-based word spotting on a large variety of data sets."}, "keywords": ["fully supervised PHOC-based", "AK dataset"], "citation_intent": "result"}
{"citing_id": "2303.14877v1", "cited_id": "1812.01041", "section_title": "Analytically Exact Simulation", "citation": "It is worth noting that stability is of great importance for optimizing over the QAOA parameter landscape, as the great number of local minima requires generically exponential independent optimization trials to reach a global minimum #REFR .", "text_before_citation": ["2 shows the results for a collection of different optimizers, including Adam #OTHEREFR , COBYLA #OTHEREFR , and DARBO.", "(For benchmark with more choices of common optimizers, see the Supplemental Material.) The results are collected from five different n = 16 w3R graph instances, and for each graph and each optimization method, the best record over 20 independent optimization trials is reported.", "The results show the superior efficiency of DARBO over other common optimizers.", "For all different p values, the approximation gaps 1 \u2212 r of Adam and COBYLA are 1.02 \u223c 2.08 times and 1.28 \u223c 3.47 times larger than those of DARBO, respectively.", "In addition to the efficiency and accuracy, the DARBO performance is also more stable in terms of different problem graph instances."], "text_after_citation": ["As a result, with the increasing depth p of QAOA, the optimization problem becomes harder, and we see that Adam and COBYLA do not even exhibit a monotonic growth of the accuracy as p increases, despite the fact that the cut size given by the optimal |\u03c8(\u03b3, \u03b2) (over the choice of 2p parameters (\u03b3, \u03b2)) is clearly non-decreasing with p.", "The reason of the performance drop is due to a lack of training stability and the sensitivity on initial parameter choices for conventional local optimization methods.", "DARBO, on the contrary, does give better results with large depth p.", "In other words, one important optimization advantage brought by DARBO is its capability of finding the near optimal parameters with a small number of independent optimization trials, which are not enough to locate the global optimal parameters for conventional optimizers.", "Some advanced optimization methods such as FOURIER heuristics reported in #OTHEREFR can be good at locating global minimum but depend on progressive optimization on lower-depth QAOA, leading to a much larger total amount of required circuit evaluations."], "citing_paper_content": {"title": "Error-Mitigated Quantum Approximate Optimization Via Learning-Based Adaptive Optimization", "abstract": "Combinatorial optimization problems are ubiquitous and computationally hard to solve in general. Quantum computing is envisioned as a powerful tool offering potential computational advantages for solving some of these problems. Quantum approximate optimization algorithm (QAOA), one of the most representative quantum-classical hybrid algorithms, is designed to solve certain combinatorial optimization problems by transforming a discrete optimization problem into a classical optimization problem over a continuous circuit parameter domain. QAOA objective landscape over the parameter variables is notorious for pervasive local minima and barren plateaus, and its viability in training significantly relies on the efficacy of the classical optimization algorithm. To enhance the performance of QAOA, we design double adaptive-region Bayesian optimization (DARBO), an adaptive classical optimizer for QAOA. Our experimental results demonstrate that the algorithm greatly outperforms conventional gradient-based and gradient-free optimizers in terms of speed, accuracy, and stability. We also address the issues of measurement efficiency and the suppression of quantum noise by successfully conducting the full optimization loop on the superconducting quantum processor. This work helps to unlock the full power of QAOA and paves the way toward achieving quantum advantage in practical classical tasks."}, "cited_paper_content": {"title": "Quantum Approximate Optimization Algorithm: Performance, Mechanism, And Implementation On Near-Term Devices", "abstract": "The Quantum Approximate Optimization Algorithm (QAOA) is a hybrid quantum-classical variational algorithm designed to tackle combinatorial optimization problems. Despite its promise for near-term quantum applications, not much is currently understood about QAOA's performance beyond its lowest-depth variant. An essential but missing ingredient for understanding and deploying QAOA is a constructive approach to carry out the outer-loop classical optimization. We provide an in-depth study of the performance of QAOA on MaxCut problems by developing an efficient parameter-optimization procedure and revealing its ability to exploit non-adiabatic operations. Building on observed patterns in optimal parameters, we propose heuristic strategies for initializing optimizations to find quasi-optimal $p$-level QAOA parameters in $O(\\text{poly}(p))$ time, whereas the standard strategy of random initialization requires $2^{O(p)}$ optimization runs to achieve similar performance. We then benchmark QAOA and compare it with quantum annealing, especially on difficult instances where adiabatic quantum annealing fails due to small spectral gaps. The comparison reveals that QAOA can learn via optimization to utilize non-adiabatic mechanisms to circumvent the challenges associated with vanishing spectral gaps. Finally, we provide a realistic resource analysis on the experimental implementation of QAOA. When quantum fluctuations in measurements are accounted for, we illustrate that optimization will be important only for problem sizes beyond numerical simulations, but accessible on near-term devices. We propose a feasible implementation of large MaxCut problems with a few hundred vertices in a system of 2D neutral atoms, reaching the regime to challenge the best classical algorithms."}, "keywords": ["optimization trials"], "citation_intent": "background"}
{"citing_id": "2303.09841v1", "cited_id": "1810.04805", "section_title": "Architecture And Loss Objective", "citation": "This is realized by a BERT #REFR encoder, a composition of layers for input embedding, positional encoding and multi-head self-attention blocks (cf.", "text_before_citation": ["Differently to the GAD characterization and aggregation function (cf. Eq. 3 and Eq.", "4) of the deep generative models of #OTHEREFR our deep bidirectional GAD-Former \u03a8 models the characterization and aggregation functions as follows with", "EQUATION", "The characterization function f \u03b8 of GADFormer maps the bidirectional relationships between group members o i of a group G m (representations of trajectory point p or segment s of trajectory \u039b m ) to an attention-based feature map b Gm , representing the behavior of an individual group (the individual trajectory path)."], "text_after_citation": ["Figure 1 ) using group member embeddings pe as input tokens.", "EQUATION", "The aggregation function g \u03a6 of GADFormer approximates instead of a distribution for normal group representations G (ref ) with distance measure d a probability p m for abnormal group behavior (abnormal trajectory path).", "This is realized by non-linear custom task layers of the output block, which maps the group behavior characteristics b \u039bm to a task specific feature map representation z \u039bm .", "EQUATION"], "citing_paper_content": {"title": "Gadformer: An Attention-Based Model For Group Anomaly Detection On Trajectories", "abstract": "Group Anomaly Detection (GAD) reveals anomalous behavior among groups consisting of multiple member instances, which are, individually considered, not necessarily anomalous. This task is of major importance across multiple disciplines, in which also sequences like trajectories can be considered as a group. However, with increasing amount and heterogenity of group members, actual abnormal groups get harder to detect, especially in an unsupervised or semi-supervised setting. Recurrent Neural Networks are well established deep sequence models, but recent works have shown that their performance can decrease with increasing sequence lengths. Hence, we introduce with this paper GADFormer, a GAD specific BERT architecture, capable to perform attention-based Group Anomaly Detection on trajectories in an unsupervised and semi-supervised setting. We show formally and experimentally how trajectory outlier detection can be realized as an attention-based Group Anomaly Detection problem. Furthermore, we introduce a Block Attention-anomaly Score (BAS) to improve the interpretability of transformer encoder blocks for GAD. In addition to that, synthetic trajectory generation allows us to optimize the training for domain-specific GAD. In extensive experiments we investigate our approach versus GRU in their robustness for trajectory noise and novelties on synthetic and real world datasets."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["BERT encoder"], "citation_intent": "method"}
{"citing_id": "2303.08132v2", "cited_id": "1506.04214", "section_title": "Problem Formulation And Motivation", "citation": "A naive solution would be to use recurrent neural networks (RNNs) such as Conv-LSTM #REFR to extract the dynamics from the input frames and predict the next, but this solution yields poor results.", "text_before_citation": ["Since we focus on segmentation tasks, we directly use the binary mask of an instance to represent the shape and position and predict motion upon it.", "Let I t \u2208 R w\u00d7h\u00d73 denotes the t-th frame of the input video and m k t \u2208 R w\u00d7h denotes the binary mask of instance k in this frame.", "The goal of this module is to learn the motion and deformation from m k t\u2212n to m k t\u22121 , and then predict the shape and position, denoted by m k t .", "In our problem, we only require our module to predict one frame forward, but we wish it can take various number of frames as input, i.e.", "m k t = F (m k t\u2212n:t\u22121 ), \u2200n \u2265 2 where F denotes our motion module."], "text_after_citation": ["They tend to make pixel-wise predictions for the entire image, while we want to learn instance-level motion and thus all pixels should be considered as a unit.", "Following the concept of conditional variational autoencode (CVAE) that models the dynamics with a Gaussian and samples latent variables z t from this distribution for prediction #OTHEREFR , we also use an encoder during training to learn the motion pattern represented by z t .", "However, we maintain a memory network M to store representative motion patterns.", "The stored motion patterns are then used during inference to help refine the incomplete pattern that is extracted only from previous frames m k t\u2212n:t\u22121 , and the refined pattern, which contains the instance-level motion information, is used to assist the prediction of the next frames m k t . Fig.", "2 shows the overall framework of our motion prediction plugin."], "citing_paper_content": {"title": "Instmove: Instance Motion For Object-Centric Video Segmentation", "abstract": "Despite significant efforts, cutting-edge video segmentation methods still remain sensitive to occlusion and rapid movement, due to their reliance on the appearance of objects in the form of object embeddings, which are vulnerable to these disturbances. A common solution is to use optical flow to provide motion information, but essentially it only considers pixel-level motion, which still relies on appearance similarity and hence is often inaccurate under occlusion and fast movement. In this work, we study the instance-level motion and present InstMove, which stands for Instance Motion for Object-centric Video Segmentation. In comparison to pixel-wise motion, Inst-Move mainly relies on instance-level motion information that is free from image feature embeddings, and features physical interpretations, making it more accurate and robust toward occlusion and fast-moving objects. To better fit in with the video segmentation tasks, InstMove uses instance masks to model the physical presence of an object and learns the dynamic model through a memory network to predict its position and shape in the next frame. With only a few lines of code, InstMove can be integrated into current SOTA methods for three different video segmentation tasks and boost their performance. Specifically, we improve the previous arts by 1.5 AP on OVIS dataset, which features heavy occlusions, and 4.9 AP on YouTubeVIS-Long dataset, which mainly contains fast moving objects. These results suggest that instance-level motion is robust and accurate, and hence serving as a powerful solution in complex scenarios for object-centric video segmentation."}, "cited_paper_content": {"title": "Convolutional Lstm Network: A Machine Learning Approach For Precipitation Nowcasting", "abstract": "The goal of precipitation nowcasting is to predict the future rainfall intensity in a local region over a relatively short period of time. Very few previous studies have examined this crucial and challenging weather forecasting problem from the machine learning perspective. In this paper, we formulate precipitation nowcasting as a spatiotemporal sequence forecasting problem in which both the input and the prediction target are spatiotemporal sequences. By extending the fully connected LSTM (FC-LSTM) to have convolutional structures in both the input-to-state and state-to-state transitions, we propose the convolutional LSTM (ConvLSTM) and use it to build an end-to-end trainable model for the precipitation nowcasting problem. Experiments show that our ConvLSTM network captures spatiotemporal correlations better and consistently outperforms FC-LSTM and the state-of-the-art operational ROVER algorithm for precipitation nowcasting."}, "keywords": ["Conv-LSTM", "recurrent neural networks"], "citation_intent": "method"}
{"citing_id": "2304.00935v1", "cited_id": "1611.01144", "section_title": "A. Unsupervised Learning Task", "citation": "The SDLGM is compared with state-of-the-art methods, including Gumbel-Softmax, ST Gumbel-Softmax, Score-Function (SF), DARN, MuProp, Straight-Through (ST), and Slope-Annealed ST #REFR .", "text_before_citation": ["We start with the representation learning task."], "text_after_citation": ["The performance of the methods is measured with negative variational lower bound of the log-likelihood (lower is better), which is a commonly used criterion in the literature.", "In the experiments, we follow the similar settings as #OTHEREFR . The popular image dataset MNIST is used.", "The encoder is defined as We first quantitatively demonstrate the performance of the proposed method with the commonly used measurement: the negative variational lower bound (lower is better).", "We compare the SDLGM with multiple baselines, shown as Table I .", "For a fair comparison, we directly use the results reported in #OTHEREFR . More details of the baselines refer to #OTHEREFR ."], "citing_paper_content": {"title": "Learning Sparsity Of Representations With Discrete Latent Variables", "abstract": "Deep latent generative models have attracted increasing attention due to the capacity of combining the strengths of deep learning and probabilistic models in an elegant way. The data representations learned with the models are often continuous and dense. However in many applications, sparse representations are expected, such as learning sparse high dimensional embedding of data in an unsupervised setting, and learning multi-labels from thousands of candidate tags in a supervised setting. In some scenarios, there could be further restriction on degree of sparsity: the number of non-zero features of a representation cannot be larger than a pre-defined threshold L0. In this paper we propose a sparse deep latent generative model SDLGM to explicitly model degree of sparsity and thus enable to learn the sparse structure of the data with the quantified sparsity constraint. The resulting sparsity of a representation is not fixed, but fits to the observation itself under the pre-defined restriction. In particular, we introduce to each observation i an auxiliary random variable Li, which models the sparsity of its representation. The sparse representations are then generated with a two-step sampling process via two Gumbel-Softmax distributions. For inference and learning, we develop an amortized variational method based on MC gradient estimator. The resulting sparse representations are differentiable with backpropagation. The experimental evaluation on multiple datasets for unsupervised and supervised learning problems shows the benefits of the proposed method."}, "cited_paper_content": {"title": "Categorical Reparameterization With Gumbel-Softmax", "abstract": "Categorical variables are a natural choice for representing discrete structure in the world. However, stochastic neural networks rarely use categorical latent variables due to the inability to backpropagate through samples. In this work, we present an efficient gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample from a novel Gumbel-Softmax distribution. This distribution has the essential property that it can be smoothly annealed into a categorical distribution. We show that our Gumbel-Softmax estimator outperforms state-of-the-art gradient estimators on structured output prediction and unsupervised generative modeling tasks with categorical latent variables, and enables large speedups on semi-supervised classification."}, "keywords": ["ST Gumbel-Softmax"], "citation_intent": "method"}
{"citing_id": "2305.00673v1", "cited_id": "1710.09412", "section_title": "Ablation Studies", "citation": "Interpolation Strategies We compare BCP with other two interpolation strategies: Mixup #REFR and Fine-Grained CutMix (FG-CutMix).", "text_before_citation": ["Inward and outward copy-paste (In and Out in the table) mean using", "X l M + X u (1 \u2212 M) or X u M + X l", "(1 \u2212 M) respectively to train the network.", "We also conduct within-set copy-paste (CP in the table), i.e., copy-paste labeled data on another labeled data and copy-paste unlabeled data on another unlabeled data.", "We can see that all these variants get inferior performances compared with our BCP, since they either lack consistent manner for training labeled and unlabeled data, or lack common semantics transfer between labeled and unlabeled data."], "text_after_citation": ["For Mixup, we superimpose labeled and unlabeled data to generate new training images, imitating the framework of GuidedMix-Net #OTHEREFR .", "For FG-CutMix, we crop training images into 4\u00d74 patches and combine them in batch to generate new images.", "The network predictions of new images are re-combined and then supervised by ground-truth or pseudo-labels.", "The results of LA dataset are shown in Table 5 .", "Due to similar spatial structures of medical images, Mixup brings more influential noise in medical images. FG-CutMix maintains less structure information after CutMix than BCP. More details will be discussed in supplementary material. Fig."], "citing_paper_content": {"title": "Bidirectional Copy-Paste For Semi-Supervised Medical Image Segmentation", "abstract": "In semi-supervised medical image segmentation, there exist empirical mismatch problems between labeled and unlabeled data distribution. The knowledge learned from the labeled data may be largely discarded if treating labeled and unlabeled data separately or in an inconsistent manner. We propose a straightforward method for alleviating the problem \u2212 copy-pasting labeled and unlabeled data bidirectionally, in a simple Mean Teacher architecture. The method encourages unlabeled data to learn comprehensive common semantics from the labeled data in both inward and outward directions. More importantly, the consistent learning procedure for labeled and unlabeled data can largely reduce the empirical distribution gap. In detail, we copypaste a random crop from a labeled image (foreground) onto an unlabeled image (background) and an unlabeled image (foreground) onto a labeled image (background), respectively. The two mixed images are fed into a Student network and supervised by the mixed supervisory signals of pseudo-labels and ground-truth. We reveal that the simple mechanism of copy-pasting bidirectionally between labeled and unlabeled data is good enough and the experiments show solid gains (e.g., over 21% Dice improvement on ACDC dataset with 5% labeled data) compared with other state-of-the-arts on various semi-supervised medical image segmentation datasets. Code is available at https: //github.com/DeepMed-Lab-ECNU/BCP."}, "cited_paper_content": {"title": "Mixup: Beyond Empirical Risk Minimization", "abstract": "Large deep neural networks are powerful, but exhibit undesirable behaviors such as memorization and sensitivity to adversarial examples. In this work, we propose mixup, a simple learning principle to alleviate these issues. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularizes the neural network to favor simple linear behavior in-between training examples. Our experiments on the ImageNet-2012, CIFAR-10, CIFAR-100, Google commands and UCI datasets show that mixup improves the generalization of state-of-the-art neural network architectures. We also find that mixup reduces the memorization of corrupt labels, increases the robustness to adversarial examples, and stabilizes the training of generative adversarial networks."}, "keywords": ["Fine-Grained CutMix", "Mixup"], "citation_intent": "method"}
{"citing_id": "2304.13911v1", "cited_id": "1602.05629", "section_title": "Introduction", "citation": "Recent progressives in federated learning (FL) #REFR have proved that utilizing distributed data sources can both preserve data privacy and enhance model training.", "text_before_citation": ["#OTHEREFR How to improve the question answering accuracy has become a serious challenge which decides whether LLMs can be accepted as a robust and reliable part in realistic applications.", "One commonsense is that can we crowd-source many questions and aggregate those questions to better understand some common questions.", "A common question might be asked frequently as its variants in the concrete parameters or rephrased formulations.", "For example, the Chickens-and-rabbits questions can be asked with different number of heads and feet.", "Now we want to ask Can we fully utilize those similar questions to improve the question answering of the LLMs without tuning the model parameters or infringing user privacy?"], "text_after_citation": ["In the FL paradigm, each client trains a local learning model with own data, while a central server regularly communicates with all agents to generate a better global model through the aggregation of the local models.", "In this study, we consider improving the reasoning capacity of LLMs by better understanding crowd-sourced similar questions, from which we can explore the contextual information and improve the LLM answers substantially.", "Inspired by FL, we propose two typical scenarios when a user sends a QA request to the LLM and the LLM tries to answer with a collected question database.", "-Synonymous Questions with Same Parameters (SP-questions).", "The cloud-deployed system retrieves from the database and finds several synonymous but rephrased questions with exactly the same parameters."], "citing_paper_content": {"title": "Federated Prompting And Chain-Of-Thought Reasoning For Improving Llms Answering", "abstract": "We investigate how to enhance answer precision in frequently asked questions posed by distributed users using cloud-based Large Language Models (LLMs). Our study focuses on a typical situations where users ask similar queries that involve identical mathematical reasoning steps and problem-solving procedures. Due to the unsatisfactory accuracy of LLMs' zero-shot prompting with standalone questions, we propose to improve the distributed synonymous questions using Self-Consistency (SC) and Chain-of-Thought (CoT) techniques. Specifically, we first retrieve synonymous questions from a crowd-sourced database and create a federated question pool. We call these federated synonymous questions with the same or different parameters SP-questions or DP-questions, respectively. We refer to our methods as Fed-SP-SC and Fed-DP-CoT, which can generate significantly more accurate answers for all user queries without requiring sophisticated model-tuning. Through extensive experiments, we demonstrate that our proposed methods can significantly enhance question accuracy by fully exploring the synonymous nature of the questions and the consistency of the answers."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["distributed data sources", "federated learning"], "citation_intent": "background"}
{"citing_id": "2303.12419v1", "cited_id": "1904.11238", "section_title": "Learning With Noisy Labels", "citation": "Furthermore, Dynamic bootstrapping #REFR use a Beta-Mixture-Model to combine label dynamically for every sample.", "text_before_citation": ["Therefore, various methods were proposed to improve robustness against noisy labels.", "The typical algorithms for combating noisy labels including adding regularization #OTHEREFR , designing robust loss functions #OTHEREFR , selecting possible clean samples #OTHEREFR and correcting the labels #OTHEREFR .", "We mainly introduce the label correction approaches which are most related to this work.", "Specifically, label correction algorithms aim to identify and correct suspicious labels in an iterative framework.", "Bootstrapping #OTHEREFR combines the noisy label and the DNN prediction as a soft pseudo-label to replace the original label."], "text_after_citation": ["Differently, Joint Optimization #OTHEREFR trained their model with a large learning rate for several epochs firstly, and then generated pseudo-labels by averaging the predictions of that epochs.", "Recently, SELFIE #OTHEREFR and Ada-Corr #OTHEREFR selectively combine the labels of noisy data for corresponding to true labels with a high probability.", "However, almost all existing noisy label studies, focusing on the classification task, cannot be adapted to cross-modal matching directly in which the noisy label refers to the alignment errors between data pairs rather than the categorical labeling errors.", "The label of cross-modal matching can be consider as y i \u2208 {0, 1}, where data pair is positively correlated (y i = 1) or not (y i = 0).", "Therefore, label correction of noisy multi-modal data pairs can be translated into generating a continuous value between 0 and 1, called soft correspondence labels in this paper, to measure the true degree of correspondence."], "citing_paper_content": {"title": "Bicro: Noisy Correspondence Rectification For Multi-Modality Data Via Bi-Directional Cross-Modal Similarity Consistency", "abstract": "As one of the most fundamental techniques in multimodal learning, cross-modal matching aims to project various sensory modalities into a shared feature space. To achieve this, massive and correctly aligned data pairs are required for model training. However, unlike unimodal datasets, multimodal datasets are extremely harder to collect and annotate precisely. As an alternative, the co-occurred data pairs (e.g., image-text pairs) collected from the Internet have been widely exploited in the area. Unfortunately, the cheaply collected dataset unavoidably contains many mismatched data pairs, which have been proven to be harmful to the model's performance. To address this, we propose a general framework called BiCro (Bidirectional Cross-modal similarity consistency), which can be easily integrated into existing cross-modal matching models and improve their robustness against noisy data. Specifically, BiCro aims to estimate soft labels for noisy data pairs to reflect their true correspondence degree. The basic idea of BiCro is motivated by that-taking image-text matching as an example-similar images should have similar textual descriptions and vice versa. Then the consistency of these two similarities can be recast as the estimated soft labels to train the matching model. The experiments on three popular cross-modal matching datasets demonstrate that our method significantly improves the noise-robustness of various matching models, and surpass the state-of-the-art by a clear margin.The code is available at https://github.com/xu5zhao/BiCro."}, "cited_paper_content": {"title": "Unsupervised Label Noise Modeling And Loss Correction", "abstract": "Despite being robust to small amounts of label noise, convolutional neural networks trained with stochastic gradient methods have been shown to easily fit random labels. When there are a mixture of correct and mislabelled targets, networks ::: tend to fit the former before the latter. This suggests using a suitable two-component mixture model as an unsupervised generative model of sample loss values during training to allow online estimation of the probability that a sample is mislabelled. Specifically, we propose a beta mixture to estimate this probability and correct the loss by relying on the network prediction (the so-called bootstrapping loss). We further adapt mixup augmentation to drive our approach a step further. Experiments on CIFAR-10/100 and TinyImageNet demonstrate a robustness to label noise that substantially outperforms recent state-of-the-art. Source code is available at https://git.io/fjsvE and Appendix at https://arxiv.org/abs/1904.11238."}, "keywords": ["label"], "citation_intent": "method"}
{"citing_id": "2304.14130v1", "cited_id": "1811.07901", "section_title": "Performance", "citation": "On the other hand, participants in the Prediction condition failed to surpass the model's performance, suggesting that predictions alone are not as effective in improving the joint performance, supporting the findings in #REFR .", "text_before_citation": ["Figure 2a , depicts participants' unassisted performance per condition, along with a 95% confidence interval. Details about all CIs are presented in Appendix A.", "Figure 2a shows that 75% belongs to all CIs, so participants and model showed comparable performance in all conditions, thus recreating the setting in #OTHEREFR .", "Then, we compared participants' performance after exposure to the model's assistance (Assisted Performance) to the model's accuracy.", "Figure 2a , shows the assisted performance, along with the corresponding 95% CIs.", "Participants' assisted performance was significantly better than 75% in all but the Prediction condition, suggesting that even the simple strategy of pairing model predictions with confidence, as in the Local condition, is beneficial to participants' performance, in line with the findings in #OTHEREFR ."], "text_after_citation": ["Having established that the model's assistance helped the ensemble surpass the individual model accuracy, we continue by examining whether it surpassed participant's individual accuracy as well.", "Figure 2b , shows the 95% CIs of the difference between participants' assisted and unassisted performance, per condition.", "Participants' assisted performance was significantly better than their unassisted performance in the Prediction and Local conditions.", "On the contrary, the same comparison did not yield statistically significant results in the Combined and Explanations conditions, even though the point estimates were positive.", "This pattern can be explained, at least in part, by the fact that participants in the Combined and Explanations conditions already had better performance in their unassisted predictions compared to participants in the Prediction and Local conditions, leaving less room for improvement for them."], "citing_paper_content": {"title": "Why Not Both? Complementing Explanations With Uncertainty, And The Role Of Self-Confidence In Human-Ai Collaboration", "abstract": "AI and ML models have already found many applications in critical domains, such as healthcare and criminal justice. However, fully automating such high-stakes applications can raise ethical or fairness concerns. Instead, in such cases, humans should be assisted by automated systems so that the two parties reach a joint decision, stemming out of their interaction. In this work we conduct an empirical study to identify how uncertainty estimates and model explanations affect users' reliance, understanding, and trust towards a model, looking for potential benefits of bringing the two together. Moreover, we seek to assess how users' behaviour is affected by their own self-confidence in their abilities to perform a certain task, while we also discuss how the latter may distort the outcome of an analysis based on agreement and switching percentages. Preprint. Under review."}, "cited_paper_content": {"title": "On Human Predictions With Explanations And Predictions Of Machine Learning Models: A Case Study On Deception Detection", "abstract": "Humans are the final decision makers in critical tasks that involve ethical and legal concerns, ranging from recidivism prediction, to medical diagnosis, to fighting against fake news. Although machine learning models can sometimes achieve impressive performance in these tasks, these tasks are not amenable to full automation. To realize the potential of machine learning for improving human decisions, it is important to understand how assistance from machine learning models affects human performance and human agency. In this paper, we use deception detection as a testbed and investigate how we can harness explanations and predictions of machine learning models to improve human performance while retaining human agency. We propose a spectrum between full human agency and full automation, and develop varying levels of machine assistance along the spectrum that gradually increase the influence of machine predictions. We find that without showing predicted labels, explanations alone slightly improve human performance in the end task. In comparison, human performance is greatly improved by showing predicted labels (>20% relative improvement) and can be further improved by explicitly suggesting strong machine performance. Interestingly, when predicted labels are shown, explanations of machine predictions induce a similar level of accuracy as an explicit statement of strong machine performance. Our results demonstrate a tradeoff between human performance and human agency and show that explanations of machine predictions can moderate this tradeoff."}, "keywords": ["predictions"], "citation_intent": "result"}
{"citing_id": "2304.07327v1", "cited_id": "1711.09050", "section_title": "Safety And Ethical Implications", "citation": "LLMs can also produce toxic or hateful content and fail to follow provided user-constraints #REFR .", "text_before_citation": ["We presented the OpenAssistant Conversations dataset, an outcome of a crowd-sourcing initiative aimed at promoting research in the area of alignment in LLMs.", "We recognize that sufficiently powerful language models can have a significant impact on society #OTHEREFR , and therefore we believe it is essential to promote transparency and ethical considerations in their development and deployment.", "These models are often prone to generating inaccurate information about people, places, or facts, a phenomenon commonly known as 'hallucinations' #OTHEREFR ."], "text_after_citation": ["Additionally, these models tend to incorporate biases present in their training data, leading to unfair and discriminatory outputs #OTHEREFR .", "While methods such as RLHF can mitigate some of these shortcomings, they may exacerbate others [26, 20] .", "We hope that alignment can fix some of these issues #OTHEREFR , but it is important to acknowledge that achieving alignment is a complex and ongoing challenge.", "Our team has put in significant effort to ensure that the community has access to an open-source high-quality dataset free of unethical or harmful responses.", "We believe that creating a safe and respectful environment for our users is paramount, and we encourage them to generate prompts and replies that are not only polite, but also creative and detailed."], "citing_paper_content": {"title": "Openassistant Conversations -Democratizing Large Language Model Alignment", "abstract": "Aligning large language models (LLMs) with human preferences has proven to drastically improve usability and has driven rapid adoption as demonstrated by ChatGPT. Alignment techniques such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) greatly reduce the required skill and domain knowledge to effectively harness the capabilities of LLMs, increasing their accessibility and utility across various domains. However, state-of-the-art alignment techniques like RLHF rely on high-quality human feedback data, which is expensive to create and often remains proprietary. In an effort to democratize research on large-scale alignment, we release OpenAssistant Conversations, a human-generated, human-annotated assistant-style conversation corpus consisting of 161,443 messages distributed across 66,497 conversation trees, in 35 different languages, annotated with 461,292 quality ratings. The corpus is a product of a worldwide crowd-sourcing effort involving over 13,500 volunteers. 2 To demonstrate the OpenAssistant Conversations dataset's effectiveness, we present OpenAssistant, the first fully open-source large-scale instruction-tuned model to be trained on human data. A preference study revealed that OpenAssistant replies are comparably preferred to GPT-3.5-turbo (ChatGPT) with a relative winrate of 48.3% vs. 51.7% respectively. We release our code 3 and data 4 under fully permissive licenses."}, "cited_paper_content": {"title": "Ethical Challenges In Data-Driven Dialogue Systems", "abstract": "The use of dialogue systems as a medium for human-machine interaction is an increasingly prevalent paradigm. A growing number of dialogue systems use conversation strategies that are learned from large datasets. There are well documented instances where interactions with these system have resulted in biased or even offensive conversations due to the data-driven training process. Here, we highlight potential ethical issues that arise in dialogue systems research, including: implicit biases in data-driven systems, the rise of adversarial examples, potential sources of privacy violations, safety concerns, special considerations for reinforcement learning systems, and reproducibility concerns. We also suggest areas stemming from these issues that deserve further investigation. Through this initial survey, we hope to spur research leading to robust, safe, and ethically sound dialogue systems."}, "keywords": ["LLMs", "hateful content"], "citation_intent": "background"}
{"citing_id": "2305.01118v1", "cited_id": "1512.00567", "section_title": "Fine-Grained Species Recognition", "citation": "Only (ImageNet) #REFR 5 d on the model performance, we conduct an additional ablation study of d on the iNat2018 dataset with different \u03bb and report the results in Table 3 .", "text_before_citation": ["To understand the effectiveness of each loss component in CSP (see Figure 3 and Equation 4), we conduct an ablation study on the iNat2018 dataset with different \u03bb and report the results in Table 2 .", "We can see that each component contributes to the final model performance. Deleting any of them will lead to performance drops.", "To understand the effect of location embedding dimension 4 https://github.com/visipedia/inat_comp/ tree/master/2018 Table 1 : The Top1 accuracy of different models and training strategies on the iNat2018 validation dataset for the species finegrain recognition task with different training data ratios, where \u03bb% = 100% indicates the fully supervised setting.", "We run each model 5 times and report the standard deviation in \"()\".", "Ratio \u03bb% 5% 10% 20% 100% Img."], "text_after_citation": ["We can see that at the few-shot setting \u03bb% = 5%, 10%, 20%, models with d = 256 achieve the best performance.", "In the fully supervised setting, the model with d = 1024 leads to the best performance.", "Last but not least, we also explore whether our CSP is effective on different image encoders.", "We conduct an ablation study of different F() on the iNat2018 dataset with \u03bb% = 5%. Table 4 summarizes the results.", "We can see that no matter which F() we use, Inception V3 or ViT, our CSP-MC-BLD consistently yields the best results, and ViT improves the model performance a lot."], "citing_paper_content": {"title": "Csp: Self-Supervised Contrastive Spatial Pre-Training For Geospatial-Visual Representations", "abstract": "Geo-tagged images are publicly available in large quantities, whereas labels such as object classes are rather scarce and expensive to collect. Meanwhile, contrastive learning has achieved tremendous success in various natural image and language tasks with limited labeled data. However, existing methods fail to fully leverage geospatial information, which can be paramount to distinguishing objects that are visually similar. To directly leverage the abundant geospatial information associated with images in pre-training, fine-tuning, and inference stages, we present Contrastive Spatial Pre-Training (CSP), a selfsupervised learning framework for geo-tagged images. We use a dual-encoder to separately encode the images and their corresponding geo-locations, and use contrastive objectives to learn effective location representations from images, which can be transferred to downstream supervised tasks such as image classification. Experiments show that CSP can improve model performance on both iNat2018 and fMoW dataset. Especially, on iNat2018, CSP significantly boosts the model performance with 10-34% relative improvement with various labeled training data sampling ratios 1 ."}, "cited_paper_content": {"title": "Rethinking The Inception Architecture For Computer Vision", "abstract": "Convolutional networks are at the core of most state of-the-art computer vision solutions for a wide variety of tasks. Since 2014 very deep convolutional networks started to become mainstream, yielding substantial gains in various benchmarks. Although increased model size and computational cost tend to translate to immediate quality gains for most tasks (as long as enough labeled data is provided for training), computational efficiency and low parameter count are still enabling factors for various use cases such as mobile vision and big-data scenarios. Here we are exploring ways to scale up networks in ways that aim at utilizing the added computation as efficiently as possible by suitably factorized convolutions and aggressive regularization. We benchmark our methods on the ILSVRC 2012 classification challenge validation set demonstrate substantial gains over the state of the art: 21:2% top-1 and 5:6% top-5 error for single frame evaluation using a network with a computational cost of 5 billion multiply-adds per inference and with using less than 25 million parameters. With an ensemble of 4 models and multi-crop evaluation, we report 3:5% top-5 error and 17:3% top-1 error on the validation set and 3:6% top-5 error on the official test set."}, "keywords": ["iNat2018", "model performance"], "citation_intent": "result"}
{"citing_id": "2303.05463v1", "cited_id": "1908.04321", "section_title": "Related Work", "citation": "IITB-Corridor #REFR has one of the most extensive collections of frames but focuses on a single location which may not be generalized to other settings.", "text_before_citation": ["However, contemporary videobased methods #OTHEREFR discussed below were found to be more accurate, utilizing deep learning techniques.", "Numerous methods have been proposed for detecting anomalous behaviors using video-based anomaly datasets.", "These datasets are videos of normal and abnormal behaviors in the training and validation set.", "One of the most recent datasets, Charlotte Anomaly Dataset (CHAD) [25] , provides high-resolution multi-camera anomaly detection with bounding box, identity, and pose annotations.", "ShanghaiTech Campus dataset #OTHEREFR is a well-known dataset that provides a collection of 317K normal and anomalous frames across several scenes."], "text_after_citation": ["In contrast with previously mentioned datasets which collect videos from continuous streaming cameras, UCF Crime #OTHEREFR is based on the videos collected from YouTube, covering many scenes and thus changing the context with each scene.", "This may not be ideal for anomaly detection as it is context-specific.", "Another new dataset is the NOLA dataset #OTHEREFR , which consists of a single scene covered by a moving camera.", "One of the characteristics of NOLA is that they use a moving camera, making them unique compared to their peers.", "Unsupervised learning approaches for vision-based anomaly detection are well-established and widely used."], "citing_paper_content": {"title": "Understanding The Challenges And Opportunities Of Pose-Based Anomaly Detection", "abstract": "Pose-based anomaly detection is a video-analysis technique for detecting anomalous events or behaviors by examining human pose extracted from the video frames. Utilizing pose data alleviates privacy and ethical issues. Also, computation-wise, the complexity of pose-based models is lower than pixel-based approaches. However, it introduces more challenges, such as noisy skeleton data, losing important pixel information, and not having enriched enough features. These problems are exacerbated by a lack of anomaly detection datasets that are good enough representatives of real-world scenarios. In this work, we analyze and quantify the characteristics of two well-known video anomaly datasets to better understand the difficulties of pose-based anomaly detection. We take a step forward, exploring the discriminating power of pose and trajectory for video anomaly detection and their effectiveness based on context. We believe these experiments are beneficial for a better comprehension of pose-based anomaly detection and the datasets currently available. This will aid researchers in tackling the task of anomaly detection with a more lucid perspective, accelerating the development of robust models with better performance."}, "cited_paper_content": {"title": "Multi-Timescale Trajectory Prediction For Abnormal Human Activity Detection", "abstract": "A classical approach to abnormal activity detection is to learn a representation for normal activities from the training data and then use this learned representation to detect abnormal activities while testing. Typically, the methods based on this approach operate at a fixed timescale - either a single time-instant (eg. frame-based) or a constant time duration (eg. video-clip based). But human abnormal activities can take place at different timescales. For example, jumping is a short term anomaly and loitering is a long term anomaly in a surveillance scenario. A single and pre-defined timescale is not enough to capture the wide range of anomalies occurring with different time duration. In this paper, we propose a multi-timescale model to capture the temporal dynamics at different timescales. In particular, the proposed model makes future and past predictions at different timescales for a given input pose trajectory. The model is multi-layered where intermediate layers are responsible to generate predictions corresponding to different timescales. These predictions are combined to detect abnormal activities. In addition, we also introduce an abnormal activity data-set for research use that contains 4,83,566 annotated frames. Data-set will be made available at this https URL Our experiments show that the proposed model can capture the anomalies of different time duration and outperforms existing methods."}, "keywords": ["frames"], "citation_intent": "background"}
{"citing_id": "2305.02024v1", "cited_id": "1902.09630", "section_title": "Limitation Of Proxy Loss Functions", "citation": "The use of these proxy losses does not have a strong correlation with IoU #REFR . As noted by Yu et al.", "text_before_citation": ["Thus, the proxy loss function, in this case, is harsher on the model than the evaluation metric.", "The mismatch between the per-character cross-entropy loss and the edit distance is shown in Figure 1 .1 through an example.", "\u2022 Scene Text Detection.", "Given a natural scene image, the goal is to pre-cisely localize all instances of text at a word level. The ground truth consists of rotated bounding boxes.", "Popular approaches use smooth-L 1 or L 2 distance for regressing bounding box coordinates."], "text_after_citation": ["#OTHEREFR , IoU accounts for a bounding box as a whole, whereas regressing using an L n proxy loss treats each point independently.", "\u2022 Image compression.", "The objective of compression approaches is to minimize the storage cost of images.", "In a lossy setting, this is achieved by removing information that is least noticeable to humans.", "Learning-based compression approaches follow a rate-distortion objective, which tries to minimize the storage requirements while keeping the distortion as low as possible."], "citing_paper_content": {"title": "", "abstract": "I hereby declare I have written this doctoral dissertation independently and quoted all the sources of information used in accordance with methodological instructions on ethical principles for writing an academic dissertation. Moreover, I state that this dissertation has neither been submitted nor accepted for any other degree. The results presented in this dissertation were were achieved during my own research in cooperation with my dissertation supervisor Jiri Matas published in [1]-[8]. During the Ph.D., I also collaborated with several researchers on multiple projects, with R. Manmatha, Srikar"}, "cited_paper_content": {"title": "Generalized Intersection Over Union: A Metric And A Loss For Bounding Box Regression", "abstract": "Intersection over Union (IoU) is the most popular evaluation metric used in the object detection benchmarks. However, there is a gap between optimizing the commonly used distance losses for regressing the parameters of a bounding box and maximizing this metric value. The optimal objective for a metric is the metric itself. In the case of axis-aligned 2D bounding boxes, it can be shown that IoU can be directly used as a regression loss. However, IoU has a plateau making it infeasible to optimize in the case of non-overlapping bounding boxes. In this paper, we address the this weakness by introducing a generalized version of IoU as both a new loss and a new metric. By incorporating this generalized IoU ( GIoU) as a loss into the state-of-the art object detection frameworks, we show a consistent improvement on their performance using both the standard, IoU based, and new, GIoU based, performance measures on popular object detection benchmarks such as PASCAL VOC and MS COCO."}, "keywords": ["proxy losses", "IoU"], "citation_intent": "background"}
{"citing_id": "2305.02474v1", "cited_id": "1901.09451", "section_title": "Data Pre-Processing", "citation": "In some specific problems, removing subgroup indicators such as the sex of a job candidate from their application has shown to have minimal influence on classifier accuracy while improving the fairness #REFR .", "text_before_citation": ["#OTHEREFR provide a detailed overview of data anonymization models and techniques in healthcare such as kanonymity, k-map, l-diversity, t-closeness, \u03b4-disclosure privacy, \u03b2-likeness, \u03b4-presence, and ( , \u03b4)-differential privacy.", "To avoid data leakage, many tools for data anonymization and its evaluation tools #OTHEREFR such as Sec-Graph #OTHEREFR , ARX-tool for anonymizing biomedical data #OTHEREFR , Amnesia 29 #OTHEREFR , PySyft #OTHEREFR , Synthea #OTHEREFR and Anonimatron 30 (open-source data anonymization tool written in Java) can be incorporated in the MLHOps pipeline.", "3. Removing subgroups indicators.", "Changing the race of the patients can have a dramatic impact on the outcome of an algorithm that is designed to fill a prompt #OTHEREFR .", "Therefore, the existence of race attributes in the text can decrease the fairness of the model dramatically."], "text_after_citation": ["This method is applicable mostly in text-based data where sensitive attributes are easily removable.", "As a preprocessing step, one can estimate the effect of keeping or removing such sensitive attributes on the overall accuracy and fairness of a developed model.", "At the same time, it is not always possible to remove the sensitive attributes from the data.", "For example, AI models can predict patient race from medical images, but it is not yet clear how they can do it #OTHEREFR .", "In one study #OTHEREFR , researchers did not provide the patient race during model training, but they also could not find a particular patch or region in the data for which AI failed to detect race by removing that part."], "citing_paper_content": {"title": "Mlhops: Machine Learning For Healthcare Operations", "abstract": "Machine Learning Health Operations (MLHOps) is the combination of processes for reliable, efficient, usable, and ethical deployment and maintenance of machine learning models in healthcare settings. This paper provides both a survey of work in this area and guidelines for developers and clinicians to deploy and maintain their own models in clinical practice. We cover the foundational concepts of general machine learning operations, describe the initial setup of MLHOps pipelines (including data sources, preparation, engineering, and tools). We then describe long-term monitoring and updating (including data distribution shifts and model updating) and ethical considerations (including bias, fairness, interpretability, and privacy). This work therefore provides guidance across the full pipeline of MLHOps from conception to initial and ongoing deployment."}, "cited_paper_content": {"title": "Bias In Bios: A Case Study Of Semantic Representation Bias In A High-Stakes Setting", "abstract": "We present a large-scale study of gender bias in occupation classification, a task where the use of machine learning may lead to negative outcomes on peoples' lives. We analyze the potential allocation harms that can result from semantic representation bias. To do so, we study the impact on occupation classification of including explicit gender indicators---such as first names and pronouns---in different semantic representations of online biographies. Additionally, we quantify the bias that remains when these indicators are \"scrubbed,\" and describe proxy behavior that occurs in the absence of explicit gender indicators. As we demonstrate, differences in true positive rates between genders are correlated with existing gender imbalances in occupations, which may compound these imbalances."}, "keywords": ["classifier accuracy", "job candidate"], "citation_intent": "background"}
{"citing_id": "2303.01864v1", "cited_id": "1608.01953", "section_title": "B. Results", "citation": "A similar behavior is observed for Mix+Incons_hardMag, which outperforms its special case \u03c3 = 0 #REFR .", "text_before_citation": ["First, let us investigate the influence of the consistency wiehgt \u03c3 onto performance. From the results displayed in Fig.", "1a-1c , we remark that Mag+Incons_hardMix exhibits a stable performance with respect to \u03c3.", "Conversely, for the other algorithms, the SDR peaks at specific values of \u03c3.", "In particular, a properly tuned Mix+Incons outperforms its special cases corresponding to \u03c3 = 0 and +\u221e, which were use in #OTHEREFR ."], "text_after_citation": ["This demonstrates the interest of our framework, which allows for deriving general algorithms that outperform their special cases found in the literature. Besides, Fig.", "1d reveals that the algorithms exhibit a different behavior over iterations.", "Indeed, while for most methods, the SDR steadily increases over iterations, MISI reaches its peak performance after very few iterations, and then its performance drops severely.", "This was notably observed in #OTHEREFR ; nonetheless, the modified MISI algorithm introduced in this paper (which is similar to Mix+Incons_hardMag) was compared to MISI after 200 iterations, which is somewhat unfair to MISI.", "Here, we select the optimal weight and number of iterations for each algorithm in order to run them on the test set in a more fair setup."], "citing_paper_content": {"title": "Spectrogram Inversion For Audio Source Separation Via Consistency, Mixing, And Magnitude Constraints", "abstract": "Audio source separation is often achieved by estimating the magnitude spectrogram of each source, and then applying a phase recovery (or spectrogram inversion) algorithm to retrieve time-domain signals. Typically, spectrogram inversion is treated as an optimization problem involving one or several terms in order to promote estimates that comply with a consistency property, a mixing constraint, and/or a target magnitude objective. Nonetheless, it is still unclear which set of constraints and problem formulation is the most appropriate in practice. In this paper, we design a general framework for deriving spectrogram inversion algorithm, which is based on formulating optimization problems by combining these objectives either as soft penalties or hard constraints. We solve these by means of algorithms that perform alternating projections on the subsets corresponding to each objective/constraint. Our framework encompasses existing techniques from the literature as well as novel algorithms. We investigate the potential of these approaches for a speech enhancement task. In particular, one of our novel algorithms outperforms other approaches in a realistic setting where the magnitudes are estimated beforehand using a neural network."}, "cited_paper_content": {"title": "Model-Based Stft Phase Recovery For Audio Source Separation", "abstract": "For audio source separation applications, it is common to estimate the magnitude of the short-time Fourier transform (STFT) of each source. In order to further synthesizing time-domain signals, it is necessary to recover the phase of the corresponding complex-valued STFT. Most authors in this field choose a Wiener-like filtering approach which boils down to using the phase of the original mixture. In this paper, a different standpoint is adopted. Many music events are partially composed of slowly varying sinusoids and the STFT phase increment over time of those frequency components takes a specific form. This allows phase recovery by an unwrapping technique once a short-term frequency estimate has been obtained. Herein, a novel iterative source separation procedure is proposed which builds upon these results. It consists in minimizing the mixing error by means of the auxiliary function method. This procedure is initialized by exploiting the unwrapping technique in order to generate estimates that benefit from a temporal continuity property. Experiments conducted on realistic music pieces show that, given accurate magnitude estimates, this procedure outperforms the state-of-the-art consistent Wiener filter."}, "keywords": ["Mix+Incons_hardMag"], "citation_intent": "result"}
{"citing_id": "2304.12429v1", "cited_id": "1908.00780", "section_title": "Comparing Sparsifier With Related Works", "citation": "We explore the synthetic and KDDCUP'99 datasets since these are what were tested in the private ADMM paper #REFR .", "text_before_citation": ["All reported \u01eb values indicate the privacy of the output weights, \u01eb 1 + \u01eb 2 . Since we were unable to Fig. 1 .", "Comparing the number of correctly identified zeros of Sparsifier with twostage, private ADMM, and private IGHT. Sparsifier and private IGHT were run for 1000 iterations.", "A black line at 92 coefficients has been drawn indicating the number of true zero coefficients.", "Higher numbers of correct zero coefficients indicate that an algorithm is able to maintain sparsity.", "replicate the sparse outputs of the private ADMM algorithm, we directly use the results provided in that paper."], "text_after_citation": ["Synthetic Data Here, we test Sparsifier on a synthetic dataset to demonstrate that it outperforms prior solutions for sparse private logistic regression.", "We compare our solution to the the two-stage procedure developed by Kifer et al., private ADMM algorithm developed by Wang & Zhang, and the private IGHT method developed by Wang & Gu #OTHEREFR .", "For this comparison, we generated synthetic data according to the details provided by Wang & Zhang #OTHEREFR .", "Specifically, we generated 10, 000 datapoints where each x i \u2208 R 100 was drawn from N (0, \u03a3) with \u03a3 ij = 0.5 |i\u2212j| and scaled each feature so its maximum absolute value in the dataset was 1.", "The true weight vector w * had its first 8 components set to 10 9 8 7 6 5 4 0.5 and its last 92 components set to 0."], "citing_paper_content": {"title": "Sparse Private Lasso Logistic Regression", "abstract": "LASSO regularized logistic regression is particularly useful for its built-in feature selection, allowing coefficients to be removed from deployment and producing sparse solutions. Differentially private versions of LASSO logistic regression have been developed, but generally produce dense solutions, reducing the intrinsic utility of the LASSO penalty. In this paper, we present a differentially private method for sparse logistic regression that maintains hard zeros. Our key insight is to first train a non-private LASSO logistic regression model to determine an appropriate privatized number of non-zero coefficients to use in final model selection. To demonstrate our method's performance, we run experiments on synthetic and real-world datasets."}, "cited_paper_content": {"title": "Differential Privacy For Sparse Classification Learning", "abstract": "In this paper, we present a differential privacy version of convex and nonconvex sparse classification approach. Based on alternating direction method of multiplier (ADMM) algorithm, we transform the solving of sparse problem into the multistep iteration process. Then we add exponential noise to stable steps to achieve privacy protection. By the property of the post-processing holding of differential privacy, the proposed approach satisfies the \u03f5 \u2212 differential privacy even when the original problem is unstable. Furthermore, we present the theoretical privacy bound of the differential privacy classification algorithm. Specifically, the privacy bound of our algorithm is controlled by the algorithm iteration number, the privacy parameter, the parameter of loss function, ADMM pre-selected parameter, and the data size. Finally we apply our framework to logistic regression with L1 regularizer and logistic regression with L1/2 regularizer. Numerical studies demonstrate that our method is both effective and efficient which performs well in sensitive data analysis."}, "keywords": ["private ADMM paper"], "citation_intent": "method"}
{"citing_id": "2303.12818v1", "cited_id": "1502.03167", "section_title": "Analysis Of Weights And Gradients", "citation": "Since AffineLayer performs significantly worse than BatchNorm, these results cast doubt on the claim that the success of BatchNorm is due to it stabilizing the gradients #REFR .", "text_before_citation": ["Here, we extract the weights and gradients for each of input-early, input-late, final-early, and final-late, and we plot various histograms of these distributions.", "In Figure 3(a) , we see that the weight distribution changes from input-early to final-late are fairly modest for BatchNorm, while Figure 3 (b) yields the same conclusion for AffineLayer.", "In Figure 3 (c), we have overlayed the final-late weights of BatchNorm and AffineLayer, while Figure 3(d) gives the analogous result for the gradients.", "We see that the weights and gradients behave similarly, and hence whatever is causing BatchNorm to outperform AffineLayer does not appear to be distinguishable through this weight and gradient distribution comparison.", "The bottom line here is that AffineLayer has a similar effect on the gradients as BatchNorm, at least in the all-important first epoch."], "text_after_citation": ["The results in Figure 3 show minimal differences between BatchNorm and AffineLayer.", "A significant difference between AffineLayer and BatchNorm can be observed by comparing the input-early and input-late updates.", "Specifically, in Figure 4(a) , we compare the input-early gradient with the input-late gradient for BatchNorm, while Figure 4 (b) provides the analogous comparison for AffineLayer.", "We observe from Figure 4 (a) that for BatchNorm, the gradient starts spread out and become more tightly focused, whereas Figure 4(b) shows the opposite behavior for the AffineLayer gradient.", "Figure 4 provides additional evidence that it is the weight normalization at the beginning of each epoch that is causing the observed performance differences between BatchNorm and AffineLayer, at least in the case of ResNet18. We discuss this further in the next section."], "citing_paper_content": {"title": "An Empirical Analysis Of The Shift And Scale Parameters In Batchnorm", "abstract": "Batch Normalization (BatchNorm) is a technique that improves the training of deep neural networks, especially Convolutional Neural Networks (CNN). It has been empirically demonstrated that BatchNorm increases performance, stability, and accuracy, although the reasons for such improvements are unclear. BatchNorm includes a normalization step as well as trainable shift and scale parameters. In this paper, we empirically examine the relative contribution to the success of BatchNorm of the normalization step, as compared to the re-parameterization via shifting and scaling. To conduct our experiments, we implement two new optimizers in PyTorch, namely, a version of BatchNorm that we refer to as AffineLayer, which includes the re-parameterization step without normalization, and a version with just the normalization step, that we call BatchNorm-minus. We compare the performance of our AffineLayer and BatchNorm-minus implementations to standard BatchNorm, and we also compare these to the case where no batch normalization is used. We experiment with four ResNet architectures (ResNet18, ResNet34, ResNet50, and ResNet101) over a standard image dataset and multiple batch sizes. Among other findings, we provide empirical evidence that the success of BatchNorm may derive primarily from improved weight initialization."}, "cited_paper_content": {"title": "Batch Normalization: Accelerating Deep Network Training By Reducing Internal Covariate Shift", "abstract": "Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization, and in some cases eliminates the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.82% top-5 test error, exceeding the accuracy of human raters."}, "keywords": ["BatchNorm"], "citation_intent": "result"}
{"citing_id": "2304.14633v1", "cited_id": "2003.10432", "section_title": "Method", "citation": "To verify our proposed novel RCCV is a general 3D geometric feature representation and not limited to a specific framework, we additionally apply it to Atlas #REFR to replace its simple back-projected feature. The results meet our expectations.", "text_before_citation": ["The lower the better The higher the better Accuracy Completeness Chamfer Precision Recall F1 Score Atlas #OTHEREFR Table 4 . Atlas with Our RCCV."], "text_after_citation": ["nism.", "3 sets of experiments are designed as follows: #OTHEREFR We simply concatenate the keyframe 2D image feature to the Ray-compensated Cost Volume.", "All 3D mesh metrics are improved by this concatenation, confirming the effectiveness of the contextual feature.", "(2)A 2D convolutional layer is employed to better fuse the contextual feature with the cost volume, the improvements are relatively minor.", "3We use a group convolution layer to limit the perceptive field of the depth planes in the cost volume to be within itself."], "citing_paper_content": {"title": "Cvrecon: Rethinking 3D Geometric Feature Learning For Neural Reconstruction", "abstract": "Recent advances in neural reconstruction using posed image sequences have made remarkable progress. However, due to the lack of depth information, existing volumetric-based techniques simply duplicate 2D image features of the object surface along the entire camera ray. We contend this duplication introduces noise in empty and occluded spaces, posing challenges for producing highquality 3D geometry. Drawing inspiration from traditional multi-view stereo methods, we propose an end-to-end 3D neural reconstruction framework CVRecon, designed to exploit the rich geometric embedding in the cost volumes to facilitate 3D geometric feature learning. Furthermore, we present Ray-contextual Compensated Cost Volume (RCCV), a novel 3D geometric feature representation that encodes view-dependent information with improved integrity and robustness. Through comprehensive experiments, we demonstrate that our approach significantly improves the reconstruction quality in various metrics and recovers clear fine details of the 3D geometries. Our extensive ablation studies provide insights into the development of effective 3D geometric feature learning schemes. Project page: https://cvrecon.ziyue.cool/."}, "cited_paper_content": {"title": "Atlas: End-To-End 3D Scene Reconstruction From Posed Images", "abstract": "We present an end-to-end 3D reconstruction method for a scene by directly regressing a truncated signed distance function (TSDF) from a set of posed RGB images. Traditional approaches to 3D reconstruction rely on an intermediate representation of depth maps prior to estimating a full 3D model of a scene. We hypothesize that a direct regression to 3D is more effective. A 2D CNN extracts features from each image independently which are then back-projected and accumulated into a voxel volume using the camera intrinsics and extrinsics. After accumulation, a 3D CNN refines the accumulated features and predicts the TSDF values. Additionally, semantic segmentation of the 3D model is obtained without significant computation. This approach is evaluated on the Scannet dataset where we significantly outperform state-of-the-art baselines (deep multiview stereo followed by traditional TSDF fusion) both quantitatively and qualitatively. We compare our 3D semantic segmentation to prior methods that use a depth sensor since no previous work attempts the problem with only RGB input."}, "keywords": ["general 3D geometric", "simple back-projected feature"], "citation_intent": "result"}
{"citing_id": "2304.01768v1", "cited_id": "1310.7991", "section_title": "Introduction", "citation": "For MOD a convergence radius of order 1/S 2 was derived in #REFR , however, our results show that even under milder assumptions the basin of convergence is several orders of magnitude larger.", "text_before_citation": ["In particular, when assuming that the data follows a random sparse model with an overcomplete generating dictionary \u03a6, there are no global recovery guarantees, in the sense that one of the algorithms above or a variant will recover \u03a6 with high probability.", "Our Contribution In this paper we will cement the theoretical foundations of dictionary learning via alternating minimisation by characterising the convergence basin for both MOD and ODL (aK-SVD) in combination with thresholding as sparse approximation algorithm as summarised in Algorithm 1.1.", "In particular we will show that under mild conditions on the generating dictionary, any initialisation within atom-wise \u2113 2 -norm and scaled operator norm distance of order 1/ log K will converge exponentially close to the generating dictionary given enough training signals.", "Moreover, we provide further conditions on an initialisation that ensure convergence even for distances close to the maximal distance \u221a 2.", "In case of ODL (aK-SVD) this is the first result of this kind."], "text_after_citation": ["Finally, our most important contribution is that we provide the first theoretical results which are valid for data generating models where the non-zero entries of the generating coefficients are not essentially uniformly distributed.", "This means that some dictionary elements can be more likely to appear in the data representation than others.", "Such a non-homogeneous use of dictionary elements can usually be observed in dictionaries learned on image or audio data.", "Our results, proving the stability and convergence of alternating minimisation methods for non-uniform sparse support models, might be an explanation for their practical advantages over graph clustering or element-wise dictionary learning methods.", "Organisation The remainder of the paper is organised as follows."], "citing_paper_content": {"title": "Convergence Of Alternating Minimisation Algorithms For Dictionary Learning", "abstract": "In this paper we derive sufficient conditions for the convergence of two popular alternating minimisation algorithms for dictionary learning-the Method of Optimal Directions (MOD) and Online Dictionary Learning (ODL), which can also be thought of as approximative K-SVD. We show that given a well-behaved initialisation that is either within distance at most 1/ log(K) to the generating dictionary or has a special structure ensuring that each element of the initialisation only points to one generating element, both algorithms will converge with geometric convergence rate to the generating dictionary. This is done even for data models with non-uniform distributions on the supports of the sparse coefficients. These allow the appearance frequency of the dictionary elements to vary heavily and thus model real data more closely."}, "cited_paper_content": {"title": "Learning Sparsely Used Overcomplete Dictionaries Via Alternating Minimization", "abstract": "We consider the problem of sparse coding, where each sample consists of a sparse linear combination of a set of dictionary atoms, and the task is to learn both the dictionary elements and the mixing coefficients. Alternating minimization is a popular heuristic for sparse coding, where the dictionary and the coefficients are estimated in alternate steps, keeping the other fixed. Typically, the coefficients are estimated via $\\ell_1$ minimization, keeping the dictionary fixed, and the dictionary is estimated through least squares, keeping the coefficients fixed. In this paper, we establish local linear convergence for this variant of alternating minimization and establish that the basin of attraction for the global optimum (corresponding to the true dictionary and the coefficients) is $\\mathcal{O}(1/s^{2})$, where $s$ is the sparsity level in each sample and the dictionary satisfies restricted isometry property. Combined with the recent results of approximate dictionary estimation, this yields provable guar..."}, "keywords": ["convergence"], "citation_intent": "result"}
{"citing_id": "2304.05310v1", "cited_id": "1904.01681", "section_title": "Related Works", "citation": "As emphasized in #REFR , the flow of the NODEs cannot represent some functions omnipresently emergent in applications.", "text_before_citation": ["EQUATION", "Ref.", "#OTHEREFR calculated the gradients by calling an ODE solver with extended ODEs (i.e., concatenating the original state, the adjoint, and the other partial derivatives for the parameters at each time point into a single vector).", "Notably, for the regression task of the time series, the loss function probably depends on the state at multiple observational times, such as the form of L(h(t 0 ), h(t 1 ), ..., h(t n )).", "Under such a case, we must update the adjoint state instantly by adding the partial derivative of the loss at each observational time point."], "text_after_citation": ["Typical examples include the following two-valued function with one argument: g 1-D (1) = \u22121 and g 1-D (\u22121) = 1.", "Our framework desires to conquer the representation limitation observed in applying the NODEs.", "Optimal control As mentioned above, a closed connection between deep neural networks and dynamical systems has been emphasized in the literature and, correspondingly, theories, methods, and tools of dynamical systems have been employed, e.g.", "the theory of optimal control #OTHEREFR - #OTHEREFR , #OTHEREFR - #OTHEREFR .", "Generally, we model a typical task using a deep neural network and then train the network parameters such that the given loss function can be reduced by some learning algorithm."], "citing_paper_content": {"title": "Neural Delay Differential Equations: System Reconstruction And Image Classification", "abstract": "Neural Ordinary Differential Equations (NODEs), a framework of continuous-depth neural networks, have been widely applied, showing exceptional efficacy in coping with representative datasets. Recently, an augmented framework has been developed to overcome some limitations that emerged in the application of the original framework. In this paper, we propose a new class of continuous-depth neural networks with delay, named Neural Delay Differential Equations (NDDEs). To compute the corresponding gradients, we use the adjoint sensitivity method to obtain the delayed dynamics of the adjoint. Differential equations with delays are typically seen as dynamical systems of infinite dimensions that possess more fruitful dynamics. Compared to NODEs, NDDEs have a stronger capacity for nonlinear representations. We use several illustrative examples to demonstrate this outstanding capacity. Firstly, we successfully model the delayed dynamics where the trajectories in the lower-dimensional phase space could be mutually intersected and even chaotic in a model-free or model-based manner. Traditional NODEs, without any argumentation, are not directly applicable to such modeling. Secondly, we achieve lower loss and higher accuracy not only for the data produced synthetically by complex models but also for the CIFAR10, a well-known image dataset. Our results on the NDDEs demonstrate that appropriately articulating the elements of dynamical systems into the network design is truly beneficial in promoting network performance."}, "cited_paper_content": {"title": "Augmented Neural Odes", "abstract": "We show that Neural Ordinary Differential Equations (ODEs) learn representations that preserve the topology of the input space and prove that this implies the existence of functions Neural ODEs cannot represent. To address these limitations, we introduce Augmented Neural ODEs which, in addition to being more expressive models, are empirically more stable, generalize better and have a lower computational cost than Neural ODEs."}, "keywords": ["NODEs", "functions"], "citation_intent": "background"}
{"citing_id": "2304.13135v1", "cited_id": "1705.01809", "section_title": "Data Preprocessing", "citation": "Second, to accommodate the pixel-value representation necessary for image processing, we normalize the range of pixel values from (0, 255) to (0, 1) to train the deep learning model correctly #REFR .", "text_before_citation": ["To begin, a random 60%, 20%, 10%, and 10% split is made between the training, validation, testing A, and testing B sets from the specified dataset."], "text_after_citation": ["It can be characterized as follows: In this paper, we use pre-trained feature extractors for easy modeling because it may be difficult to train a deep learning model for recognizing COVID-19 infection from scratch due to a lack of CT scans of COVID-19 subjects.", "The transfer learning method and several different pretrained models address this problem #OTHEREFR .", "One of transfer learning's key benefits is that it allows data to be trained with fewer samples and in less time #OTHEREFR .", "And it allows a newly trained model to benefit from the experience of a previously trained model #OTHEREFR .", "Table 3 presents the pre-trained part and the novelty portion of our MEDNC. DenseNet169."], "citing_paper_content": {"title": "Mednc: Multi-Ensemble Deep Neural Network For Covid-19 Diagnosis", "abstract": "Coronavirus disease 2019 (COVID-19) has spread all over the world for three years, but medical facilities in many areas still aren't adequate. There is a need for rapid COVID-19 diagnosis to identify highrisk patients and maximize the use of limited medical resources. Motivated by this fact, we proposed the deep learning framework MEDNC for automatic prediction and diagnosis of COVID-19 using computed tomography (CT) images. Our model was trained using two publicly available sets of COVID-19 data. And it was built with the inspiration of transfer learning. Results indicated that the MEDNC greatly enhanced the detection of COVID-19 infections, reaching an accuracy of 98.79% and 99.82% respectively. We tested MEDNC on a brain tumor and a blood cell dataset to show that our model applies to a wide range of problems. The outcomes demonstrated that our proposed models attained an accuracy of 99.39% and 99.28%, respectively. This COVID-19 recognition tool could help optimize healthcare resources and reduce clinicians' workload when screening for the virus."}, "cited_paper_content": {"title": "Pixel Normalization From Numeric Data As Input To Neural Networks: For Machine Learning And Image Processing", "abstract": "Text to image transformation for input to neural networks requires intermediate steps. This paper attempts to present a new approach to pixel normalization so as to convert textual data into image, suitable as input for neural networks. This method can be further improved by its Graphics Processing Unit (GPU) implementation to provide significant speedup in computational time."}, "keywords": ["deep learning model", "image processing"], "citation_intent": "method"}
{"citing_id": "2303.12768v1", "cited_id": "1505.05599", "section_title": "Additive Spanners", "citation": "In what follows we will mainly highlight the difference between the construction by #REFR and ours.", "text_before_citation": ["We now provide an overview for our additive spanner construction."], "text_after_citation": ["First, we apply the clustering algorithm from #OTHEREFR with radius R to G and obtain a set B of balls.", "Similar to the construction of sublinear and pairwise additive spanners, while processing unsettled pairs of vertices, we will add demand pairs to the corresponding balls that partition their shortest paths.", "It has been shown that O( \u221a n) demand pairs can be preserved exactly with O(n)", "edges #OTHEREFR , and the bound O( \u221a n) is cannot be improved, so we need to ensure that each ball B collects at most O(|B| 1/2 ) demand pairs.", "On the one hand, the small balls B with |B| \u2264 R 4/3 can be handled using previous work on subset distance preservers #OTHEREFR ."], "citing_paper_content": {"title": "Almost-Optimal Sublinear Additive Spanners", "abstract": "Given an undirected unweighted graph G = (V, E) on n vertices and m edges, a subgraph H \u2286 G is a spanner of G with stretch function f : R + \u2192 R + , iff for every pair s, t of vertices in V , dist H (s, t) \u2264 f (dist G (s, t)). When f (d) = d + o(d), H is called a sublinear additive spanner ; when f (d) = d + o(n), H is called an additive spanner, and f (d) \u2212 d is usually called the additive stretch of H. As our primary result, we show that for any constant \u03b4 > 0 and constant integer k \u2265 2, every graph on n vertices has a sublinear additive spanner with stretch function f (d) = d + O(d 1\u22121/k) and O n 1+ 1+\u03b4 2 k+1 \u22121 edges. When k = 2, this improves upon the previous spanner construction with stretch function f (d) = d + O(d 1/2) and\u00d5(n 1+3/17) edges [Chechik, 2013]; for any constant integer k \u2265 3, this improves upon the previous spanner construction with stretch function f (d) = d + O(d 1\u22121/k) and O n"}, "cited_paper_content": {"title": "Better Distance Preservers And Additive Spanners", "abstract": "We make improvements to the upper bounds on several popular types of distance preserving graph sketches. These sketches are all various restrictions of the {\\em additive pairwise spanner} problem, in which one is given an undirected unweighted graph $G$, a set of node pairs $P$, and an error allowance $+\\beta$, and one must construct a sparse subgraph $H$ satisfying $\\delta_H(u, v) \\le \\delta_G(u, v) + \\beta$ for all $(u, v) \\in P$. The first part of our paper concerns {\\em pairwise distance preservers}, which make the restriction $\\beta=0$ (i.e. distances must be preserved {\\em exactly}). Our main result here is an upper bound of $|H| = O(n^{2/3}|P|^{2/3} + n|P|^{1/3})$ when $G$ is undirected and unweighted. This improves on existing bounds whenever $|P| = \\omega(n^{3/4})$, and it is the first such improvement in the last ten years. We then devise a new application of distance preservers to graph clustering algorithms, and we apply this algorithm to {\\em subset spanners}, which require $P = S \\times S$ for some node subset $S$, and {\\em (standard) spanners}, which require $P = V \\times V$. For both of these objects, our construction generalizes the best known bounds when the error allowance is constant, and we obtain the strongest polynomial error/sparsity tradeoff that has yet been reported (in fact, for subset spanners, ours is the {\\em first} nontrivial construction that enjoys improved sparsity from a polynomial error allowance). We leave open a conjecture that $O(n^{2/3}|P|^{2/3} + n)$ pairwise distance preservers are possible for undirected unweighted graphs. Resolving this conjecture in the affirmative would improve and simplify our upper bounds for all the graph sketches mentioned above."}, "keywords": ["construction", "difference"], "citation_intent": "background"}
{"citing_id": "2303.03315v1", "cited_id": "1512.03012", "section_title": "Exploration Of Large 3D Scenes", "citation": "The other methods were trained on ShapeNet #REFR with 3D supervision since their learning process cannot scale to unknown, large environments.", "text_before_citation": ["To compare the different methods, we follow #OTHEREFR and compute the area under the curve of the evolution of the total surface coverage during exploration, after 100 NBV iterations, as presented in Table 1 .", "The surface coverage is computed using the ground truth meshes.", "AUCs are averaged on multiple trajectories in each scene: We use the same starting camera poses and same sets of candidate camera poses C t for each method for fair comparison.", "During this experiment, all weights of MACARONS are frozen and we only perform inference computation.", "The model was trained on a set of previous scenes: The 8 scenes above the bar in Table 1 were seen during online training (but with different, random starting camera poses and trajectories), and the 4 scenes below the bar were not."], "text_after_citation": ["Even if it only uses RGB images, our model is able to outperform the baselines in large environments since, contrary to other methods, its self-supervised online training strategy allows it to scale its learning process to any kind of environment.", "Figures 4 and 5 show examples of trajectories as well as of surface reconstructions computed with MACARONS."], "citing_paper_content": {"title": "Macarons: Mapping And Coverage Anticipation With Rgb Online Self-Supervision", "abstract": "Figure 1. Mapping and Coverage Anticipation with RGB Online Self-Supervision. (a) NBV methods such as [27] rely on a depth sensor to perform path planning (bottom) and scan the environment (top). They need to be trained with explicit 3D supervision, generally on small-scale meshes. (b) Our approach MACARONS instead simultaneously learns to efficiently explore the scene and to reconstruct it (top) using an RGB sensor only. Its self-supervised, online learning process scales to large-scale and complex scenes."}, "cited_paper_content": {"title": "Shapenet: An Information-Rich 3D Model Repository", "abstract": "We present ShapeNet: a richly-annotated, large-scale repository of shapes represented by 3D CAD models of objects. ShapeNet contains 3D models from a multitude of semantic categories and organizes them under the WordNet taxonomy. It is a collection of datasets providing many semantic annotations for each 3D model such as consistent rigid alignments, parts and bilateral symmetry planes, physical sizes, keywords, as well as other planned annotations. Annotations are made available through a public web-based interface to enable data visualization of object attributes, promote data-driven geometric analysis, and provide a large-scale quantitative benchmark for research in computer graphics and vision. At the time of this technical report, ShapeNet has indexed more than 3,000,000 models, 220,000 models out of which are classified into 3,135 categories (WordNet synsets). In this report we describe the ShapeNet effort as a whole, provide details for all currently available datasets, and summarize future plans."}, "keywords": ["3D supervision", "ShapeNet"], "citation_intent": "method"}
{"citing_id": "2304.03428v1", "cited_id": "1804.02767", "section_title": "Comparison With Sota Lightweight Generic Detectors", "citation": "TinyDet-L achieves the same 18.3 AP s with YOLOv3 #REFR , but it only has 1/30 computation cost of YOLOv3. Some visualized results are shown in Fig. 7 .", "text_before_citation": ["Among the compared methods, ThunderNet #OTHEREFR and Efficient-Det #OTHEREFR are regarded as the most recent SOTA lightweight detectors.", "The results show that our models obviously outperform them with fewer computation costs: TinyDet-S surpasses ThunderNet-SNet146 by 2.4% AP with similar FLOPs; TinyDet-M surpasses ThunderNet-SNet535 by 2.3% AP with 76% FLOPs; and TinyDet-L surpasses EfficientDet-D0 by 3.1% AP with similar FLOPs.", "Besides, TinyDet has better performance-computation tradeoffs than the automatically searched lightweight detec-tor (i.e., NAS-FPNLite #OTHEREFR ) and the popular YOLO series #OTHEREFR .", "Our TinyDet models obtain extraordinary performance on detecting small objects.", "With similar computation cost (i.e., FLOPs), TinyDet-M achieves 13.5 AP s , which is over 100% improvement over ThunderNet-SNet535 with 6.5 AP s ; and TinyDet-S achieves 9.6 AP s , which is also over 100% improvement over ThunderNet-SNet146 with 4.6 AP s ."], "text_after_citation": [], "citing_paper_content": {"title": "Tinydet: Accurate Small Object Detection In Lightweight Generic Detectors", "abstract": "Small object detection requires the detection head to scan a large number of positions on image feature maps, which is extremely hard for computation-and energyefficient lightweight generic detectors. To accurately detect small objects with limited computation, we propose a two-stage lightweight detection framework with extremely low computation complexity, termed as TinyDet. It enables high-resolution feature maps for dense anchoring to better cover small objects, proposes a sparsely-connected convolution for computation reduction, enhances the early stage features in the backbone, and addresses the feature misalignment problem for accurate small object detection. On the COCO benchmark, our TinyDet-M achieves 30.3 AP and 13.5 AP s with only 991 MFLOPs, which is the first detector that has an AP over 30 with less than 1 GFLOPs; besides, TinyDet-S and TinyDet-L achieve promising performance under different computation limitation."}, "cited_paper_content": {"title": "Yolov3: An Incremental Improvement", "abstract": "We present some updates to YOLO! We made a bunch of little design changes to make it better. We also trained this new network that's pretty swell. It's a little bigger than last time but more accurate. It's still fast though, don't worry. At 320x320 YOLOv3 runs in 22 ms at 28.2 mAP, as accurate as SSD but three times faster. When we look at the old .5 IOU mAP detection metric YOLOv3 is quite good. It achieves 57.9 mAP@50 in 51 ms on a Titan X, compared to 57.5 mAP@50 in 198 ms by RetinaNet, similar performance but 3.8x faster. As always, all the code is online at this https URL"}, "keywords": ["TinyDet-L", "YOLOv3"], "citation_intent": "result"}
{"citing_id": "2304.10593v1", "cited_id": "1411.1792", "section_title": "C.3. Explanation And Intuition For Capacity-Criticality-Tradeoff", "citation": "Further, we note that the aforementioned observation resonates with the findings made in #REFR -neurons in the middle layers of a network (i.e., Stage2 and Stage3) exhibit fragile co-adaption, which is difficult to re-learn.", "text_before_citation": ["To gain a better understanding of why having a higher fraction of Stage1 ReLUs is preferable for achieving superior performance at lower ReLU counts, we examined the ReLU dropping strategies employed in the prevalent ReLU optimization techniques.", "Specifically, we examined the strategies employed by DeepReDuce #OTHEREFR , SNL #OTHEREFR , and SENet #OTHEREFR , and found that they consistently demonstrate that Stage1 ReLUs are the least critical, and as such, all Stage1 ReLUs are dropped first to achieve very low ReLU counts.", "Consequently, networks with a greater proportion of non-critical ReLUs will drop a lower fraction of their critical ReLUs when aiming for very low ReLU counts, as opposed to networks with a lower proportion of non-critical ReLUs. This can be observed in Figure 17 .", "In particular, regardless of the total ReLU count of the network, WRN22x8 (ResNet18(m=16)-4x4x4x) drops a higher fraction of their critical ReLUs compared to ResNet18 (ResNet34) in order to attain a ReLU count of 25K.", "Thus, dropping a higher fraction of critical ReLUs lead to a significant loss in accuracy and results in inferior performance."], "text_after_citation": ["Consequently, dropping more ReLUs from these stages in a network with lower fraction of non-critical (Stage1) ReLUs would disrupt the fragile co-adaption and significantly reduce performance."], "citing_paper_content": {"title": "Deepreshape: Redesigning Neural Networks For Efficient Private Inference", "abstract": "The increasing demand for privacy and security has driven the advancement of private inference (PI), a cryptographic method enabling inferences directly on encrypted data. However, the computational and storage burdens of non-linear operators (e.g., ReLUs) render it impractical. Despite these limitations, prior ReLU optimization methods consistently relied on classical networks, that are not optimized for PI. Moreover, the selection of baseline networks in these ReLU optimization methods remains enigmatic and fails to provide insights into network attributes contributing to PI efficiency. In this paper, we investigate the desirable network architecture for efficient PI, and key finding is wider networks are superior at higher ReLU counts, while networks with a greater proportion of least-critical ReLUs excel at lower ReLU counts. Leveraging these findings, we develop a novel network redesign technique (DeepReShape) with a complexity of O(1), and synthesize specialized architectures(HybReNet). Compared to the state-of-the-art (SNL on CIFAR-100), we achieve a 2.35% accuracy gain at 180K ReLUs, and for ResNet50 on TinyImageNet our method saves 4.2\u00d7 ReLUs at iso-accuracy."}, "cited_paper_content": {"title": "How Transferable Are Features In Deep Neural Networks?", "abstract": "Many deep neural networks trained on natural images exhibit a curious phenomenon in common: on the first layer they learn features similar to Gabor filters and color blobs. Such first-layer features appear not to be specific to a particular dataset or task, but general in that they are applicable to many datasets and tasks. Features must eventually transition from general to specific by the last layer of the network, but this transition has not been studied extensively. In this paper we experimentally quantify the generality versus specificity of neurons in each layer of a deep convolutional neural network and report a few surprising results. Transferability is negatively affected by two distinct issues: (1) the specialization of higher layer neurons to their original task at the expense of performance on the target task, which was expected, and (2) optimization difficulties related to splitting networks between co-adapted neurons, which was not expected. In an example network trained on ImageNet, we demonstrate that either of these two issues may dominate, depending on whether features are transferred from the bottom, middle, or top of the network. We also document that the transferability of features decreases as the distance between the base task and target task increases, but that transferring features even from distant tasks can be better than using random features. A final surprising result is that initializing a network with transferred features from almost any number of layers can produce a boost to generalization that lingers even after fine-tuning to the target dataset."}, "keywords": ["network", "-neurons"], "citation_intent": "result"}
{"citing_id": "2304.05729v1", "cited_id": "1511.05298", "section_title": "Temporal Edge Modeling", "citation": "An equivalent static graph G = {V , E S , E T } is obtained by modelling temporal edges #REFR .", "text_before_citation": ["Since applying convolution on a static graph is generally easier than encoding across multiple snapshots, the DG encoding problem is frequently transformed into encoding a static graph where each node is connected to itself in the adjacent snapshot #OTHEREFR .", "This approach can also be interpreted as constructing a time-expanded graph #OTHEREFR or node-oriented ESG (see section 2.2.4) and is widely used to encode Trans-fix V,E cases, i.e. STGs.", "In more complex configurations, nodes are also connected with their k-hop neighbours in the adjacent snapshot(s) #OTHEREFR .", "A simple example of such a strategy is shown in fig. 8 ."], "text_after_citation": ["G has Once defined the connection rule for temporal edges, the traditional convolution for static graphs is applicable on ESGs. A state-of-the-art example is the ST-GCN module #OTHEREFR .", "To update the hidden states h to h in an ST-GCN layer (see Eqn.", "6), a typical spatial GNN structure aggregates the neighbour features with the msg(\u2022) function, computes their weights by the w(\u2022) function, and then sums them after normalisation with the norm(\u2022) function.", "Note that, unlike the neighbourhood definition in static graphs, the neighbourhood set N of v t i is defined by: (1) spatially, the shortest path distance with neighbour v j obeys d(v j , v i ) \u2264 K and (2) temporally, the time difference between timestamps q and t, i.e. |q \u2212 t| is not greater than \u0393/2 .", "#OTHEREFR , i.e., in a stacked way #OTHEREFR , or incorporates graph convolution when encoding each snapshot across time #OTHEREFR , i.e., in an integrated way #OTHEREFR ."], "citing_paper_content": {"title": "Dynamic Graph Representation Learning With Neural Networks: A Survey", "abstract": "In recent years, Dynamic Graph (DG) representations have been increasingly used for modeling dynamic systems due to their ability to integrate both topological and temporal information in a compact representation. Dynamic graphs allow to efficiently handle applications such as social network prediction, recommender systems, traffic forecasting or electroencephalography analysis, that can not be adressed using standard numeric representations. As a direct consequence of the emergence of dynamic graph representations, dynamic graph learning has emerged as a new machine learning problem, combining challenges from both sequential/temporal data processing and static graph learning. In this research area, Dynamic Graph Neural Network (DGNN) has became the state of the art approach and plethora of models have been proposed in the very recent years. This paper aims at providing a review of problems and models related to dynamic graph learning. The various dynamic graph supervised learning settings are analysed and discussed. We identify the similarities and differences between existing models with respect to the way time information is modeled. Finally, general guidelines for a DGNN designer when faced with a dynamic graph learning problem are provided."}, "cited_paper_content": {"title": "Structural-Rnn: Deep Learning On Spatio-Temporal Graphs", "abstract": "Deep Recurrent Neural Network architectures, though remarkably capable at modeling sequences, lack an intuitive high-level spatio-temporal structure. That is while many problems in computer vision inherently have an underlying high-level structure and can benefit from it. Spatio-temporal graphs are a popular tool for imposing such high-level intuitions in the formulation of real world problems. In this paper, we propose an approach for combining the power of high-level spatio-temporal graphs and sequence learning success of Recurrent Neural Networks~(RNNs). We develop a scalable method for casting an arbitrary spatio-temporal graph as a rich RNN mixture that is feedforward, fully differentiable, and jointly trainable. The proposed method is generic and principled as it can be used for transforming any spatio-temporal graph through employing a certain set of well defined steps. The evaluations of the proposed approach on a diverse set of problems, ranging from modeling human motion to object interactions, shows improvement over the state-of-the-art with a large margin. We expect this method to empower new approaches to problem formulation through high-level spatio-temporal graphs and Recurrent Neural Networks."}, "keywords": ["temporal edges"], "citation_intent": "method"}
{"citing_id": "2303.14592v1", "cited_id": "1703.03864", "section_title": "Results", "citation": "There is no evidence that modifying the Gaussian noise standard deviation at onset meaningfully changes the result, which is in line with #REFR who find that varying this parameter at runtime gives no difference.", "text_before_citation": ["A sample agent might be denoted as 1-0-1-0-0-0-0-0-0-0 if it won the first and third levels and no others.", "1024 agents are possible for most trials, 2 20 = 1, 048, 576 when the key win is treated separately.", "\u2022 Most Levels Solved refers to the highest count of levels solved by any individual agent. For 1-0-1-0-0-0-0-0-0-0, this number is 2.", "Tables 1 through 4 show evaluation for Vanilla Map Elites.", "Although longrunning evaluations did give rise to 10 agents, no agent came close to solving all levels."], "text_after_citation": ["Switching to a tile map view and including an explore reward did not change the highest count of levels solved.", "However, slightly relaxing the steps per level and adding partial map elites cells for gaining a key allowed the agent to solve levels far sooner (at 80K vs 400K iterations).", "Including key cells did have a meaningful difference in the number of elites in the map, including 6 elites which acquired the key in a level other than the one solved.", "Using prior knowledge to identify intermediate points in level solving is highly beneficial in solving levels more quickly and in increasing the diversity of solutions tracked in the map elites map."], "citing_paper_content": {"title": "Exploring Novel Quality Diversity Methods For Generalization In Reinforcement Learning", "abstract": "The Reinforcement Learning field is strong on achievements and weak on reapplication; a computer playing GO at a superhuman level is still terrible at Tic-Tac-Toe. This paper asks whether the method of training networks improves their generalization. Specifically we explore core quality diversity algorithms, compare against two recent algorithms, and propose a new algorithm to deal with shortcomings in existing methods. Although results of these methods are well below the performance hoped for, our work raises important points about the choice of behavior criterion in quality diversity, the interaction of differential and evolutionary training methods, and the role of offline reinforcement learning and randomized learning in evolutionary search."}, "cited_paper_content": {"title": "Evolution Strategies As A Scalable Alternative To Reinforcement Learning", "abstract": "We explore the use of Evolution Strategies (ES), a class of black box optimization algorithms, as an alternative to popular MDP-based RL techniques such as Q-learning and Policy Gradients. Experiments on MuJoCo and Atari show that ES is a viable solution strategy that scales extremely well with the number of CPUs available: By using a novel communication strategy based on common random numbers, our ES implementation only needs to communicate scalars, making it possible to scale to over a thousand parallel workers. This allows us to solve 3D humanoid walking in 10 minutes and obtain competitive results on most Atari games after one hour of training. In addition, we highlight several advantages of ES as a black box optimization technique: it is invariant to action frequency and delayed rewards, tolerant of extremely long horizons, and does not need temporal discounting or value function approximation."}, "keywords": ["deviation", "runtime"], "citation_intent": "result"}
{"citing_id": "2303.16206v1", "cited_id": "1611.05507", "section_title": "Experiments And Discussion", "citation": "These color shifts disappear when a larger image quality factor \u03bb or a more diverse dataset is used (a similar observation was made by #REFR for image transformations).", "text_before_citation": ["Essentially LISO's ability to learn from past optimizations and to stay on the image manifold, allows it to achieve low error rates, high image quality, and initializations (for further L-BFGS optimization) near a good minimum.", "Figure 4 shows the steganographic images produced by LISO under different payloads.", "Qualitatively we observe that the image quality is quite good even when a large number of bits per pixel are hidden in the cover images.", "The CelebA example shows a rare example where the hue changes slightly while the image remains un-pixelated and natural; in contrast, the low-quality steganographic images produced by FNNS tend to be pixelated.", "Different from previous image steganography methods that produce unnatural Gaussian noise, LISO's perturbation is smooth and natural because LISO implicitly learns the manifold of natural images and produces steganographic images that stay close it."], "text_after_citation": ["We can also explicitly control the trade-off between error rate and image quality by changing the image quality factors \u03bb and \u00b5 on the loss terms if desired.", "Additional qualitative images obtained by varying these are provided in Appendix J. Optimization with LISO.", "In Figure 3 , we compare optimizing with the LISO encoder against popular numerical optimization algorithms including projected gradient descent #OTHEREFR and L-BFGS #OTHEREFR ) on the steganographic encoding task.", "Note that L-BFGS-Random is equivalent to FNNS-R and L-BFGS-Pretrained is equivalent to FNNS-D (Kishore et al., 2021).", "LISO converges faster than any other optimization algorithm and this validates our hypothesis that LISO can learn a descent direction that is better than the one found by common gradientbased optimization methods; LISO converges to 100\u00d7 lower loss and 80\u00d7 lower error rate."], "citing_paper_content": {"title": "Learning Iterative Neural Optimizers For Im-Age Steganography", "abstract": "Image steganography is the process of concealing secret information in images through imperceptible changes. Recent work has formulated this task as a classic constrained optimization problem. In this paper, we argue that image steganography is inherently performed on the (elusive) manifold of natural images, and propose an iterative neural network trained to perform the optimization steps. In contrast to classical optimization methods like L-BFGS or projected gradient descent, we train the neural network to also stay close to the manifold of natural images throughout the optimization. We show that our learned neural optimization is faster and more reliable than classical optimization approaches. In comparison to previous state-of-the-art encoder-decoder based steganography methods, it reduces the recovery error rate by multiple orders of magnitude and achieves zero error up to 3 bits per pixel (bpp) without the need for error-correcting codes."}, "cited_paper_content": {"title": "Deep Feature Interpolation For Image Content Changes", "abstract": "We propose Deep Feature Interpolation (DFI), a new data-driven baseline for automatic high-resolution image transformation. As the name suggests, it relies only on simple linear interpolation of deep convolutional features from pre-trained convnets. We show that despite its simplicity, DFI can perform high-level semantic transformations like \"make older/younger\", \"make bespectacled\", \"add smile\", among others, surprisingly well - sometimes even matching or outperforming the state-of-the-art. This is particularly unexpected as DFI requires no specialized network architecture or even any deep network to be trained for these tasks. DFI therefore can be used as a new baseline to evaluate more complex algorithms and provides a practical answer to the question of which image transformation tasks are still challenging in the rise of deep learning."}, "keywords": ["image transformations"], "citation_intent": "result"}
{"citing_id": "2304.03378v1", "cited_id": "2002.05709", "section_title": "Introduction", "citation": "We inject self-supervision into video similarity learning by adopting the concept of instancediscrimination #REFR , where each video forms its own class, and any transformation of it preserves the class label.", "text_before_citation": ["In this work, three cases are considered: i) video copies #OTHEREFR , i.e., edited versions of the same source video, ii) videos of the same incident #OTHEREFR , i.e., videos capturing the same spatio-temporal span, and iii) videos of the same event #OTHEREFR , i.e., videos capturing the same spatial or temporal span. In this work, we target both retrieval and detection.", "In the former, only ranking per query matters; therefore, the distribution of similarities can vary among queries.", "While in the latter, the ability to apply a similarity threshold and detect relevant videos matters.", "Supervised learning of specialised models per task is very demanding in terms of training data collection, especially in the video domain.", "Instead, in the proposed method, we are learning a single model via SSL to perform all retrieval and detection tasks (see Figure 1 ) without further fine-tuning."], "text_after_citation": ["In this work, we adopt the ViSiL #OTHEREFR architecture for video similarity, which needs labeled video datasets for its development in prior works #OTHEREFR , but we train it in a self-supervised way and argue that instance-discrimination through augmentations is well suited for all the aforementioned tasks.", "To pronounce the synergy, we develop an appropriate composition of video augmentations and propose a model-tailored loss combined with a standard SSL loss.", "By eliminating the need for video annotations, we are able to train on large video datasets and achieve state-of-the-art results on all target retrieval and detection tasks.", "Evaluation is performed on three standard benchmarks, namely, VCDB #OTHEREFR , FIVR #OTHEREFR , and EVVE #OTHEREFR .", "In summary, our contributions include the following:"], "citing_paper_content": {"title": "Self-Supervised Video Similarity Learning", "abstract": "We introduce S 2 VS, a video similarity learning approach with self-supervision. Self-Supervised Learning (SSL) is typically used to train deep models on a proxy task so as to have strong transferability on target tasks after fine-tuning. Here, in contrast to prior work, SSL is used to perform video similarity learning and address multiple retrieval and detection tasks at once with no use of labeled data. This is achieved by learning via instance-discrimination with tasktailored augmentations and the widely used InfoNCE loss together with an additional loss operating jointly on selfsimilarity and hard-negative similarity. We benchmark our method on tasks where video relevance is defined with varying granularity, ranging from video copies to videos depicting the same incident or event. We learn a single universal model that achieves state-of-the-art performance on all tasks, surpassing previously proposed methods that use labeled data. The code and pretrained models are publicly"}, "cited_paper_content": {"title": "A Simple Framework For Contrastive Learning Of Visual Representations", "abstract": "This paper presents SimCLR: a simple framework for contrastive learning of visual representations. We simplify recently proposed contrastive self-supervised learning algorithms without requiring specialized architectures or a memory bank. In order to understand what enables the contrastive prediction tasks to learn useful representations, we systematically study the major components of our framework. We show that (1) composition of data augmentations plays a critical role in defining effective predictive tasks, (2) introducing a learnable nonlinear transformation between the representation and the contrastive loss substantially improves the quality of the learned representations, and (3) contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. By combining these findings, we are able to considerably outperform previous methods for self-supervised and semi-supervised learning on ImageNet. A linear classifier trained on self-supervised representations learned by SimCLR achieves 76.5% top-1 accuracy, which is a 7% relative improvement over previous state-of-the-art, matching the performance of a supervised ResNet-50. When fine-tuned on only 1% of the labels, we achieve 85.8% top-5 accuracy, outperforming AlexNet with 100X fewer labels."}, "keywords": ["video similarity", "self-supervision"], "citation_intent": "method"}
{"citing_id": "2303.17683v1", "cited_id": "1810.04805", "section_title": "Introduction", "citation": "Indeed, several monolingual and multilingual BERT models are available that encompass a number of languages #REFR .", "text_before_citation": ["Contemporary NLP methods such as BERT #OTHEREFR , with the large amount of knowledge contained within their parameters, paired with the relatively low computational power required to finetune them for a downstream task, have taken over many NLP applications."], "text_after_citation": ["However, the strength of these models is tied to the availability of data, and the large amounts of data required to pre-train such models exclude some languages for which it is difficult to collect large amounts of written text.", "The scarcity of data becomes more severe with dialects and language varieties.", "In fact, the very nature of dialects as an evolving form of the language, often spoken rather than written, with various social and cultural nuances, can make it difficult to develop systems tailored to specific dialects.", "In many applications, users may span a continuum of idiolects, some falling into established dialects and others not.", "It may therefore be impossible to train a system on even a small amount of data in every idiolect."], "citing_paper_content": {"title": "Fine-Tuning Bert With Character-Level Noise For Zero-Shot Transfer To Dialects And Closely-Related Languages", "abstract": "In this work, we induce character-level noise in various forms when fine-tuning BERT to enable zero-shot cross-lingual transfer to unseen dialects and languages. We fine-tune BERT on three sentence-level classification tasks and evaluate our approach on an assortment of unseen dialects and languages. We find that character-level noise can be an extremely effective agent of cross-lingual transfer under certain conditions, while it is not as helpful in others. Specifically, we explore these differences in terms of the nature of the task and the relationships between source and target languages, finding that introduction of characterlevel noise during fine-tuning is particularly helpful when a task draws on surface level cues and the source-target cross-lingual pair has a relatively high lexical overlap with shorter (i.e., less meaningful) unseen tokens on average."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["multilingual BERT models"], "citation_intent": "background"}
{"citing_id": "2303.06135v1", "cited_id": "1912.02164", "section_title": "Related Work", "citation": "As an alternative to reinforcement learning where model parameters are updated by PPO, the reward model can also be used at inference time by, for instance, rejecting the worst responses generated by the model #REFR .", "text_before_citation": ["With the emergence of pre-trained large language models (Zhu, 2022), social chatbots have recently been dominated by transformer-based #OTHEREFR designs #OTHEREFR .", "Typically, transformerbased chatbots are fine-tuned on conversational data in a specific domain, e.g. GPT-2 finetuned on conversational reddit data #OTHEREFR .", "Over the last few years, larger and better transformerbased models have been trained on conversational data for the development of more sophisticated social chatbots #OTHEREFR .", "Learning using Human Feedback: Instead of purely fine-tuning large language models on conversational data, incorporating human feedback in chatbot development has been shown to be an effective method to obtain more human-aligned responses #OTHEREFR Askell et al., 2021; #OTHEREFR .", "Popularized by the training of In-structGPT #OTHEREFR , reinforcement learning from human feedback (RLHF; #OTHEREFR ) trains a reward model on user ranks of system responses, followed by reinforcement learning as an agent interacting with the environment."], "text_after_citation": ["Human feedback has also been used in several other ways including constrained optimization (Achiam et al., 2017) , control codes #OTHEREFR and expert iteration #OTHEREFR .", "Chatbot Evaluation: The traditional method of assessing chatbots is through direct human assessment.", "This can be done through rating individual responses #OTHEREFR , scoring aggregated qualities across the multi-turn exchange, or through ranking overall conversational experiences #OTHEREFR .", "Alternatively, text overlap based metrics such as BLEU, METEOR and ROUGE #OTHEREFR and F-score metrics #OTHEREFR have also been proposed to assess chatbots.", "Other assessment methods #OTHEREFR include sen-tence perplexity #OTHEREFR John et al., 2017; #OTHEREFR , entities per exchange #OTHEREFR , number of questions raised, specificity #OTHEREFR , turns per conversation #OTHEREFR , inconsistency detection, and relevance to history."], "citing_paper_content": {"title": "Rewarding Chatbots For Real-World Engagement With Millions Of Users", "abstract": "The emergence of pretrained large language models has led to the deployment of a range of social chatbots for chitchat. Although these chatbots demonstrate language ability and fluency, they are not guaranteed to be engaging and can struggle to retain users. This work investigates the development of social chatbots that prioritize user engagement to enhance retention, specifically examining the use of human feedback to efficiently develop highly engaging chatbots. The proposed approach uses automatic pseudo-labels collected from user interactions to train a reward model that can be used to reject low-scoring sample responses generated by the chatbot model at inference time. Intuitive evaluation metrics, such as mean conversation length (MCL), are introduced as proxies to measure the level of engagement of deployed chatbots. A/B testing on groups of 10,000 new daily chatbot users on the Chai Research platform shows that this approach increases the MCL by up to 70%, which translates to a more than 30% increase in user retention for a GPT-J 6B model. Future work aims to use the reward model to realise a data flywheel , where the latest user conversations can be used to alternately fine-tune the language model and the reward model 1 ."}, "cited_paper_content": {"title": "Plug And Play Language Models: A Simple Approach To Controlled Text Generation", "abstract": "Large transformer-based language models (LMs) trained on huge text corpora have shown unparalleled generation capabilities. However, controlling attributes of the generated language (e.g. switching topic or sentiment) is difficult without modifying the model architecture or fine-tuning on attribute-specific data and entailing the significant cost of retraining. We propose a simple alternative: the Plug and Play Language Model (PPLM) for controllable language generation, which combines a pretrained LM with one or more simple attribute classifiers that guide text generation without any further training of the LM. In the canonical scenario we present, the attribute models are simple classifiers consisting of a user-specified bag of words or a single learned layer with 100,000 times fewer parameters than the LM. Sampling entails a forward and backward pass in which gradients from the attribute model push the LM's hidden activations and thus guide the generation. Model samples demonstrate control over a range of topics and sentiment styles, and extensive automated and human annotated evaluations show attribute alignment and fluency. PPLMs are flexible in that any combination of differentiable attribute models may be used to steer text generation, which will allow for diverse and creative applications beyond the examples given in this paper."}, "keywords": ["reward model", "model"], "citation_intent": "method"}
{"citing_id": "2303.07123v1", "cited_id": "1710.03077", "section_title": "Single-Dg On Image Recognition", "citation": "Setup and Implementation Details: We validate the proposed method on two image datasets: PACS #REFR , a widelyused benchmark for domain generalization with four domains: Photo (P), Art Painting (A), Cartoon (C), and Sketch (S).", "text_before_citation": [], "text_after_citation": ["VLCS #OTHEREFR , another commonly adopted benchmark for domain generalization with four different domains: VOC2007 (V), LabelMe (L), Caltech101 (C), SUN09 (S).", "In our implementation, we adopt the ResNet-18 #OTHEREFR pretrained on ImageNet #OTHEREFR as backbone. We apply the SGD optimizer with momentum 0.9. The batch size is set to 64. We set the learning rate to 2e-3/1e-3 for PACS/VLCS.", "Experiments are conducted on a Tesla P40 GPU with PyTorch-1.5.", "Following #OTHEREFR , we split the training domain into training and validation subsets, and select the best-performing model on validation set to report the OOD performances.", "Experiment Results: We first conduct experiments on PACS, shown in Table 1 ."], "citing_paper_content": {"title": "Modality-Agnostic Debiasing For Single Domain Generalization", "abstract": "Deep neural networks (DNNs) usually fail to generalize well to outside of distribution (OOD) data, especially in the extreme case of single domain generalization (single-DG) that transfers DNNs from single domain to multiple unseen domains. Existing single-DG techniques commonly devise various data-augmentation algorithms, and remould the multi-source domain generalization methodology to learn domain-generalized (semantic) features. Nevertheless, these methods are typically modality-specific, thereby being only applicable to one single modality (e.g., image). In contrast, we target a versatile Modality-Agnostic Debiasing (MAD) framework for single-DG, that enables generalization for different modalities. Technically, MAD introduces a novel two-branch classifier: a biased-branch encourages the classifier to identify the domain-specific (superficial) features, and a general-branch captures domaingeneralized features based on the knowledge from biasedbranch. Our MAD is appealing in view that it is pluggable to most single-DG models. We validate the superiority of our MAD in a variety of single-DG scenarios with different modalities, including recognition on 1D texts, 2D images, 3D point clouds, and semantic segmentation on 2D images. More remarkably, for recognition on 3D point clouds and semantic segmentation on 2D images, MAD improves DSU by 2.82% and 1.5% in accuracy and mIOU."}, "cited_paper_content": {"title": "Deeper, Broader And Artier Domain Generalization", "abstract": "The problem of domain generalization is to learn from multiple training domains, and extract a domain-agnostic model that can then be applied to an unseen domain. Domain generalization (DG) has a clear motivation in contexts where there are target domains with distinct characteristics, yet sparse data for training. For example recognition in sketch images, which are distinctly more abstract and rarer than photos. Nevertheless, DG methods have primarily been evaluated on photo-only benchmarks focusing on alleviating the dataset bias where both problems of domain distinctiveness and data sparsity can be minimal. We argue that these benchmarks are overly straightforward, and show that simple deep learning baselines perform surprisingly well on them. In this paper, we make two main contributions: Firstly, we build upon the favorable domain shift-robust properties of deep learning methods, and develop a low-rank parameterized CNN model for end-to-end DG learning. Secondly, we develop a DG benchmark dataset covering photo, sketch, cartoon and painting domains. This is both more practically relevant, and harder (bigger domain shift) than existing benchmarks. The results show that our method outperforms existing DG alternatives, and our dataset provides a more significant DG challenge to drive future research."}, "keywords": ["domain generalization"], "citation_intent": "method"}
{"citing_id": "2305.01618v1", "cited_id": "2003.08515", "section_title": "Contactart Dataset", "citation": "We deploy the teleoperation system introduced in [58] to manipulate articulated objects in the Sapien #REFR simulation.", "text_before_citation": ["To the best of our knowledge, there is only one largescale dataset #OTHEREFR including 3D hand-articulated object interaction. However, it still holds the following limitations.", "(i) HOI4D dataset is only captured in the egocentric view and can not generalize to the third view.", "(ii) The annotation of the HOI4D dataset is not accurate enough to provide contact information.", "Therefore we build a large handarticulated object interaction dataset with no annotation effort and more accurate pose and contact information."], "text_after_citation": ["This simulation system allows us to get accurate pose annotation and contact information using only a mobile phone and a laptop ( Fig. 2 (a) ).", "Instead of using professional equipments like Kinect or Real-Sense cameras, the utilization of a mobile phone makes the data collection more accessible and easier to scale up.", "Since the teleoperation is in the simulation, the annotations can be automatically recorded, which will make it easy to scale up the size of the dataset. We train our models with this collected dataset.", "While the use of the teleoperation system itself is not new, its application in collecting 3D interaction prior data for pose estimation is novel, compared to previous literature on collecting human demonstrations. Dataset collection.", "We use the front camera of a mobile phone to stream the RGB-D video at 15 fps. The setup and collection interface are shown in Fig. 2 (a) ."], "citing_paper_content": {"title": "Contactart: Learning 3D Interaction Priors For Category-Level Articulated Object And Hand Poses Estimation", "abstract": "Figure 1. Overview. We collect a dataset named ContactArt, which is created by human interacting with the articulated objects in a simulator, using teleoperation. Two interaction priors are learned from ContactArt: (i) a contact prior predicted by a diffusion model to improve 3D hand pose estimation; (ii) an articulation prior with a discriminator to improve category-level articulated object pose estimation. We visualize the pose estimation results in real-world data, leveraging the learned priors."}, "cited_paper_content": {"title": "Sapien: A Simulated Part-Based Interactive Environment", "abstract": "Building home assistant robots has long been a pursuit for vision and robotics researchers. To achieve this task, a simulated environment with physically realistic simulation, sufficient articulated objects, and transferability to the real robot is indispensable. Existing environments achieve these requirements for robotics simulation with different levels of simplification and focus. We take one step further in constructing an environment that supports household tasks for training robot learning algorithm. Our work, SAPIEN, is a realistic and physics-rich simulated environment that hosts a large-scale set for articulated objects. Our SAPIEN enables various robotic vision and interaction tasks that require detailed part-level understanding.We evaluate state-of-the-art vision algorithms for part detection and motion attribute recognition as well as demonstrate robotic interaction tasks using heuristic approaches and reinforcement learning algorithms. We hope that our SAPIEN can open a lot of research directions yet to be explored, including learning cognition through interaction, part motion discovery, and construction of robotics-ready simulated game environment."}, "keywords": ["articulated objects"], "citation_intent": "method"}
{"citing_id": "2303.03440v1", "cited_id": "1411.4380", "section_title": "I. Fixpoint Operators For 1-Categories", "citation": "As M fin and M fin M fin have bifree algebras, we recover the fixpoint operator in the relational model #REFR and obtain that it is unique.", "text_before_citation": ["Consider the lifting comonad (\u2212) on Cppo : this comonad freely adjoins a bottom element and its co-Kleisli category is isomorphic to Cppo.", "Since the endofunctors (\u2212) and (\u2212) \u25cb (\u2212) are Cppo-enriched, the category Cppo has a unique dinatural fixpoint operator uniform with respect to the free functor Cppo \u2192 Cppo and it is given by the standard formula", "f * = \u22c1 n\u2208\u03c9 f n ( ) for f \u2236 A \u2192 A in Cppo.", "We can also consider the finite multiset comonad M fin on the category Rel.", "It leads to a quantitative model of linear logic where multisets allow to count multiplicities of the inputs for a term #OTHEREFR ."], "text_after_citation": ["Lastly, the category Lin (also called the linear Scott category ScottL) can be equipped with the \u2228-semi-lattice comonad yielding a qualitative model of linear logic where substitution allows for duplication and erasure.", "This comonad also verifies the necessary enrichment conditions to obtain a unique uniform dinatural fixpoint operator."], "citing_paper_content": {"title": "Fixpoint Operators For 2-Categorical Structures", "abstract": "Fixpoint operators are tools to reason on recursive programs and data types obtained by induction (e.g. lists, trees) or coinduction (e.g. streams). They were given a categorical treatment with the notion of categories with fixpoints. A theorem by Plotkin and Simpson characterizes existence and uniqueness of fixpoint operators for categories satisfying some conditions on bifree algebras and recovers the standard examples of the category Cppo (\u03c9-complete pointed partial orders and continuous functions) in domain theory and the relational model in linear logic. We present a categorification of this result and develop the theory of 2-categorical fixpoint operators where the 2dimensional framework allows to model the execution steps for languages with (co)recursive principles. We recover the standard categorical constructions of initial algebras and final coalgebras for endofunctors as well as fixpoints of generalized species and polynomial functors."}, "cited_paper_content": {"title": "An Infinitary Model Of Linear Logic", "abstract": "In this paper, we construct an infinitary variant of the relational model of linear logic, where the exponential modality is interpreted as the set of finite or countable multisets. We explain how to interpret in this model the fixpoint operator Y as a Conway operator alternatively defined in an inductive or a coinductive way. We then extend the relational semantics with a notion of color or priority in the sense of parity games. This extension enables us to define a new fixpoint operator Y combining both inductive and coinductive policies. We conclude the paper by mentionning a connection between the resulting model of \u03bb-calculus with recursion and higher-order model-checking."}, "keywords": ["fixpoint operator", "relational model"], "citation_intent": "background"}
{"citing_id": "2304.14996v1", "cited_id": "1403.3286", "section_title": "Introduction", "citation": "Several approaches that abstract to finite Markov decision processes have been proposed: In #REFR abstractions for uncountable-state discrete-time stochastic processes are proposed for SHA, where all nondeterminism is resolved probabilistically.", "text_before_citation": ["The application of model checking methods for probabilistic HA was enabled by CEGAR-style abstractions #OTHEREFR .", "Extending decidable subclasses of HA by discrete probability distributions on jumps ( #OTHEREFR , #OTHEREFR ) preserves decidability of reachability.", "An extension of probabilistic timed automata is presented in #OTHEREFR , where continuously distributed resets are used and randomized schedulers resolve discrete nondeterminism for a discretized state-space.", "Further approaches for (networks of) stochastic timed automata ( #OTHEREFR ) maintain the probabilistic approach of resolving nondeterminism.", "Approaches for more general classes either rely on stochastic approximation (also for nondeterminism) #OTHEREFR or on a combination of discretization and randomness #OTHEREFR ."], "text_after_citation": ["In #OTHEREFR an abstraction to interval Markov decision processes is proposed.", "Both approaches feature a stochastic kernel, and can hence not be compared to our work.", "The analysis proposed in #OTHEREFR resolves discrete nondeterminism prophetically and continuous nondeterminism via a safe overapproximation (c.f. #OTHEREFR ).", "The same approach has been specified for stochastic timed automata in #OTHEREFR .", "In #OTHEREFR , discrete nondeterminism is resolved (non-)prophetically for stochastic automata, where all continuous variables are random clocks."], "citing_paper_content": {"title": "Maximizing Reachability Probabilities In Rectangular Automata With Random Clocks", "abstract": "This paper proposes an algorithm to maximize reachability probabilities for rectangular automata with random clocks via a history-dependent prophetic scheduler. This model class incorporates time-induced nondeterminism on discrete behavior and nondeterminism in the dynamic behavior. After computing reachable state sets via a forward flowpipe construction, we use backward refinement to compute maximum reachability probabilities. The feasibility of the presented approach is illustrated on a scalable model."}, "cited_paper_content": {"title": "Faust$^2$: Formal Abstractions Of Uncountable-State Stochastic Processes", "abstract": "FAUST $^{\\mathsf 2}$ is a software tool that generates formal abstractions of possibly non-deterministic discrete-time Markov processes dtMP defined over uncountable continuous state spaces. A dtMP model is specified in MATLAB and abstracted as a finite-state Markov chain or a Markov decision process. The abstraction procedure runs in MATLAB and employs parallel computations and fast manipulations based on vector calculus, which allows scaling beyond state-of-the-art alternatives. The abstract model is formally put in relationship with the concrete dtMP via a user-defined maximum threshold on the approximation error introduced by the abstraction procedure. FAUST $^{\\mathsf 2}$ allows exporting the abstract model to well-known probabilistic model checkers, such as PRISM or MRMC. Alternatively, it can handle internally the computation of PCTL properties e.g. safety or reach-avoid over the abstract model. FAUST $^{\\mathsf 2}$ allows refining the outcomes of the verification procedures over the concrete dtMP in view of the quantified and tunable error, which depends on the dtMP dynamics and on the given formula. The toolbox is available at http://sourceforge.net/projects/faust2/"}, "keywords": ["uncountable-state discrete-time stochastic"], "citation_intent": "method"}
{"citing_id": "2304.14169v1", "cited_id": "1312.1135", "section_title": "Tractability For The Fourier System", "citation": "It is interesting to note that the polynomial order d 2 \u03b5 \u22122 that we obtained for L 2 -approximation is the same as the one obtained in #REFR for numerical integration.", "text_before_citation": ["per . By the triangle inequality,", "f \u2212 P [\u2212m,m] d f \u221e \u2264 d i=1 S x 1 m \u2022 \u2022 \u2022 S x i\u22121 m (f \u2212 S x i m f ) \u221e \u2264 d i=1 (c 1 log(m)) i\u22121 \u2022 c 2 log(m)m \u2212\u03b1 \u2264 (c 3 log(m)) d m \u03b1 .", "This term is smaller than \u03b5 for m \u2265 exp(c 4 d log", "2 (d) log(\u03b5 \u22121 )), thus log N \u221e B d (\u03b5, F 3 d ) \u2264 c 5 d 2 log 2 (d) log(\u03b5 \u22121 )", ". Now the bound is obtained from Proposition 8."], "text_after_citation": ["It is surely an interesting open problem to improve the complexity bounds in Corollaries 11, 12 and 13 and/or provide matching lower bounds."], "citing_paper_content": {"title": "Tractability Of Sampling Recovery On Unweighted Function Classes", "abstract": "It is well-known that the problem of sampling recovery in the L 2-norm on unweighted Korobov spaces (Sobolev spaces with mixed smoothness) as well as classical smoothness classes such as H\u00f6lder classes suffers from the curse of dimensionality. We show that the problem is tractable for those classes if they are intersected with the Wiener algebra of functions with summable Fourier coefficients. In fact, this is a relatively simple implication of powerful results by Rauhut and Ward [Appl. Comput. Harmon. Anal. 40 (2016), pp. 321-351]. Tractability is achieved by the use of nonlinear algorithms, while linear algorithms cannot do the job."}, "cited_paper_content": {"title": "Numerical Integration Of H\\\"Older Continuous, Absolutely Convergent Fourier-, Fourier Cosine-, And Walsh Series", "abstract": "We introduce quasi-Monte Carlo rules for the numerical integration of functions f defined on [ 0 , 1 ] s , s \u2265 1 , which satisfy the following properties: the Fourier, Fourier cosine or Walsh coefficients of f are absolutely summable and f satisfies a Holder condition of order \u03b1 , for some 0 \u03b1 \u2264 1 . We show a convergence rate of the integration error of order max ( ( s \u2212 1 ) N \u2212 1 / 2 , s \u03b1 / 2 N \u2212 \u03b1 ) . The construction of the quadrature points is explicit and is based on Weil sums."}, "keywords": ["L 2 -approximation", "numerical integration"], "citation_intent": "result"}
{"citing_id": "2303.15986v1", "cited_id": "1807.00459", "section_title": "B. Clustered Federated Learning Process For Heterogeneous Devices", "citation": "Model poisoning attacks against FL are outside the scope of this paper, even if already considered in some academic works #REFR .", "text_before_citation": ["For the FL deployment, we consider a typical cross-device setting #OTHEREFR with a large number of devices.", "However, due to availability guarantees required by many IoT devices, especially in industrial settings, we expect most devices to participate in the FL training process.", "This allows the server to maintain persistent state for each client to perform the clustering step.", "In this work, we assume that no IoT device is infected prior to the FL model training and that none of them behave in an adversarial manner."], "text_after_citation": ["Our clustered FL builds upon the generalized FL setting proposed by Reddi et al. #OTHEREFR .", "This generalized FL setting, described in Algorithm 1, improves over standard FL optimization methods such as the popular Federated Averaging (FedAvg) #OTHEREFR by including adaptive optimization methods.", "The proposed clustered FL is described in Algorithm 2.", "First, the aggregation server initializes the model weights W 0 and selects the training hyperparameters.", "Then, the server sends those values to all the participating clients."], "citing_paper_content": {"title": "Clustered Federated Learning Architecture For Network Anomaly Detection In Large Scale Heterogeneous Iot Networks", "abstract": "There is a growing trend of cyberattacks against Internet of Things (IoT) devices; moreover, the sophistication and motivation of those attacks is increasing. The vast scale of IoT, diverse hardware and software, and being typically placed in uncontrolled environments make traditional IT security mechanisms such as signature-based intrusion detection and prevention systems challenging to integrate. They also struggle to cope with the rapidly evolving IoT threat landscape due to long delays between the analysis and publication of the detection rules. Machine learning methods have shown faster response to emerging threats; however, model training architectures like cloud or edge computing face multiple drawbacks in IoT settings, including network overhead and data isolation arising from the large scale and heterogeneity that characterizes these networks. This work presents an architecture for training unsupervised models for network intrusion detection in large, distributed IoT and Industrial IoT (IIoT) deployments. We leverage Federated Learning (FL) to collaboratively train between peers and reduce isolation and network overhead problems. We build upon it to include an unsupervised device clustering algorithm fully integrated into the FL pipeline to address the heterogeneity issues that arise in FL settings. The architecture is implemented and evaluated using a testbed that includes various emulated IoT/IIoT devices and attackers interacting in a complex network topology comprising 100 emulated devices, 30 switches and 10 routers. The anomaly detection models are evaluated on real attacks performed by the testbed's threat actors, including the entire Mirai malware lifecycle, an additional botnet based on the Merlin command and control server and other red-teaming tools performing scanning activities and multiple attacks targeting the emulated devices."}, "cited_paper_content": {"title": "How To Backdoor Federated Learning", "abstract": "Federated learning enables thousands of participants to construct a deep learning model without sharing their private training data with each other. For example, multiple smartphones can jointly train a next-word predictor for keyboards without revealing what individual users type. We demonstrate that any participant in federated learning can introduce hidden backdoor functionality into the joint global model, e.g., to ensure that an image classifier assigns an attacker-chosen label to images with certain features, or that a word predictor completes certain sentences with an attacker-chosen word. We design and evaluate a new model-poisoning methodology based on model replacement. An attacker selected in a single round of federated learning can cause the global model to immediately reach 100% accuracy on the backdoor task. We evaluate the attack under different assumptions for the standard federated-learning tasks and show that it greatly outperforms data poisoning. Our generic constrain-and-scale technique also evades anomaly detection-based defenses by incorporating the evasion into the attacker's loss function during training."}, "keywords": ["Model poisoning attacks"], "citation_intent": "background"}
{"citing_id": "2304.09802v1", "cited_id": "1602.05916", "section_title": "Estimation Error Bounds: Local Properties", "citation": "For instance, if |h(y) \u2212 x| \u2264 1 for any h \u2208 H, y \u2208 R ny and x \u2208 R nx , then the condition is satisfied with C = 1 #REFR .", "text_before_citation": ["Based on these restricted class of functions, we present the following assumption and theorem (the proof is provided in the supplementary material). Assumption 1.", "There exists a constant C \u2265 1 such that for every probability distribution D, and", "estimator h \u2208 H, E D nx j=1 (h j \u2212 h * j ) 2 \u2264 CE D nx j=1 (h j ) \u2212 (h * j )", ", where h j and h * j denote the jth coordinate of the estimators.", "As pointed out in #OTHEREFR , this condition usually follows from a uniform convexity condition on the loss function ."], "text_after_citation": ["Theorem 5 (Estimation error bound of ISTA, ADMM, and ReLU networks).", "Consider the class of functions represented by depth-L ISTA networks H L I as detailed in Section 2.1, m i.i.d training samples, and a per-coordinate loss satisfying Assumption 1 with a constant C. Let", "||W l \u2212 W l, * || \u221e \u2264 \u03b1 \u221a r for some \u03b1 > 0. Moreover, B \u2265 max{\u03b1 \u221a r, 1}.", "Then there exists T in the interval T \u2208 0, min \u221a mB0B L\u22121 2 L \u03bb\u03b7 , m , where \u03b7 = LB L\u22121 (B\u22121)\u2212B L +1 (B\u22121) 2", ", such that for any s > 0 with probability at least 1 \u2212 e \u2212s ,"], "citing_paper_content": {"title": "Generalization And Estimation Error Bounds For Model-Based Neural Networks", "abstract": "Model-based neural networks provide unparalleled performance for various tasks, such as sparse coding and compressed sensing problems. Due to the strong connection with the sensing model, these networks are interpretable and inherit prior structure of the problem. In practice, model-based neural networks exhibit higher generalization capability compared to ReLU neural networks. However, this phenomenon was not addressed theoretically. Here, we leverage complexity measures including the global and local Rademacher complexities, in order to provide upper bounds on the generalization and estimation errors of model-based networks. We show that the generalization abilities of model-based networks for sparse recovery outperform those of regular ReLU networks, and derive practical design rules that allow to construct model-based networks with guaranteed high generalization. We demonstrate through a series of experiments that our theoretical insights shed light on a few behaviours experienced in practice, including the fact that ISTA and ADMM networks exhibit higher generalization abilities (especially for small number of training samples), compared to ReLU networks."}, "cited_paper_content": {"title": "Local Rademacher Complexity-Based Learning Guarantees For Multi-Task Learning", "abstract": "We show a Talagrand-type concentration inequality for Multi-Task Learning (MTL), using which we establish sharp excess risk bounds for MTL in terms of distribution- and data-dependent versions of the Local Rademacher Complexity (LRC). We also give a new bound on the LRC for norm regularized as well as strongly convex hypothesis classes, which applies not only to MTL but also to the standard i.i.d. setting. Combining both results, one can now easily derive fast-rate bounds on the excess risk for many prominent MTL methods, including---as we demonstrate---Schatten-norm, group-norm, and graph-regularized MTL. The derived bounds reflect a relationship akeen to a conservation law of asymptotic convergence rates. This very relationship allows for trading off slower rates w.r.t. the number of tasks for faster rates with respect to the number of available samples per task, when compared to the rates obtained via a traditional, global Rademacher analysis."}, "keywords": ["\u2264"], "citation_intent": "background"}
{"citing_id": "2304.07354v1", "cited_id": "1702.05464", "section_title": "Nev-Ncd", "citation": "We adopt adversarial training #REFR to enforce the encoder to learn view-invariant discriminative action features by alternatively updating a discriminator and encoder towards an adversarial objective.", "text_before_citation": ["We explicitly apply this constraint as regularization 8 for the empirical distribution of the sharpened (sharpening hyper-parameter Sr) unlabeled prediction head\u1ef9 i = \u03c3 k (\u0177 i /Sr) for relatively large batch size.", "L var = E xi\u2208D u |var(\u1ef9 i )\u2212 k\u22121 k 2 | 2 2 (8)", "Joint Optimization: We jointly optimize the learning objective 9 of a weighted (\u03bb) sum of the discussed losses and regularization by utilizing both D l and D u in a single stage with supervised contrastive architecture framework.", "L = \u03bb ce L ce +\u03bb cl L cl +\u03bb nl L nl +\u03bb H L H +\u03bb var L var (9) View-invariance constraint:", "We impose optional additional constraints with the joint optimization to learn viewinvariant features given the availability of the viewpoint specifications."], "text_after_citation": ["In our setting, we place a multi-class discriminator f d to retrieve the source viewpoint information [minimizing the loss of equation 10], whereas the encoder network f \u03b8 aims to learn viewpoint information free action embedding [minimizing the loss of 11] to make the f d task no better than a random guess than maximize the discriminator loss of 10.", "EQUATION", "Where, e j denotes one-hot vector for j \u2212th position.", "EQUATION", "Moreover, we enforce contrastive learning objectives to collapse the distributions from different viewpoints to learn a view-invariant representation that only depends on the underlying action class."], "citing_paper_content": {"title": "Nev-Ncd: Negative Learning, Entropy, And Variance Regularization Based Novel Action Categories Discovery", "abstract": "Novel Categories Discovery (NCD) facilitates learning from a partially annotated label space and enables deep learning (DL) models to operate in an open-world setting by identifying and differentiating instances of novel classes based on the labeled data notions. One of the primary assumptions of NCD is that the novel label space is perfectly disjoint and can be equipartitioned, but it is rarely realized by most NCD approaches in practice. To better align with this assumption, we propose a novel singlestage joint optimization-based NCD method, Negative learning, Entropy, and Variance regularization NCD (NEV-NCD). We demonstrate the efficacy of NEV-NCD in previously unexplored NCD applications of video action recognition (VAR) with the public UCF101 dataset and a curated in-house partial action-space annotated multi-view video dataset. We perform a thorough ablation study by varying the composition of final joint loss and associated hyper-parameters. During our experiments with UCF101 and multi-view action dataset, NEV-NCD achieves \u2248 83% classification accuracy in test instances of labeled data. NEV-NCD achieves \u2248 70% clustering accuracy over unlabeled data outperforming both naive baselines (by \u2248 40%) and state-of-the-art pseudo-labeling-based approaches (by \u2248 3.5%) over both datasets. Further, we propose to incorporate optional view-invariant feature learning with the multiview dataset to identify novel categories from novel viewpoints. Our additional view-invariance constraint improves the discriminative accuracy for both known and unknown categories by \u2248 10% for novel viewpoints."}, "cited_paper_content": {"title": "Adversarial Discriminative Domain Adaptation", "abstract": "Adversarial learning methods are a promising approach to training robust deep networks, and can generate complex samples across diverse domains. They can also improve recognition despite the presence of domain shift or dataset bias: recent adversarial approaches to unsupervised domain adaptation reduce the difference between the training and test domain distributions and thus improve generalization performance. However, while generative adversarial networks (GANs) show compelling visualizations, they are not optimal on discriminative tasks and can be limited to smaller shifts. On the other hand, discriminative approaches can handle larger domain shifts, but impose tied weights on the model and do not exploit a GAN-based loss. In this work, we first outline a novel generalized framework for adversarial adaptation, which subsumes recent state-of-the-art approaches as special cases, and use this generalized view to better relate prior approaches. We then propose a previously unexplored instance of our general framework which combines discriminative modeling, untied weight sharing, and a GAN loss, which we call Adversarial Discriminative Domain Adaptation (ADDA). We show that ADDA is more effective yet considerably simpler than competing domain-adversarial methods, and demonstrate the promise of our approach by exceeding state-of-the-art unsupervised adaptation results on standard domain adaptation tasks as well as a difficult cross-modality object classification task."}, "keywords": ["view-invariant discriminative action", "adversarial training"], "citation_intent": "method"}
{"citing_id": "2304.03763v1", "cited_id": "1604.07379", "section_title": "Rgb Image Inpainting", "citation": "Since the seminal work #REFR that introduced adversarial training and encoder-decoder structure into this task, various techniques have been proposed.", "text_before_citation": ["Traditional single-view methods #OTHEREFR such as patch-based or diffusion-based work often achieve realistic results but lack high-level image understanding."], "text_after_citation": ["These include multi-stage generation #OTHEREFR , gated convolution #OTHEREFR , recurrent methods #OTHEREFR , etc.", "More recent work begins to focus on more challenging settings such as pluralistic inpainting #OTHEREFR and large hole filling #OTHEREFR .", "Specifically, LaMa #OTHEREFR uses fast Fourier Convolutions and extremely large masks for training.", "It achieves high-fidelity results with relatively low training and inference cost.", "Single image 3D photography aims to synthesize novel views using image-based rendering techniques #OTHEREFR ."], "citing_paper_content": {"title": "Clutter Detection And Removal In 3D Scenes With View-Consistent Inpainting", "abstract": "Removing clutter from scenes is essential in many applications, ranging from privacy-concerned content filtering to data augmentation. In this work, we present an automatic system that removes clutter from 3D scenes and inpaints with coherent geometry and texture. We propose techniques for its two key components: 3D segmentation based on shared properties and 3D inpainting, both of which are important problems. The definition of 3D scene clutter (frequently-moving objects) is not well captured by commonly-studied object categories in computer vision. To tackle the lack of well-defined clutter annotations, we group noisy fine-grained labels, leverage virtual rendering, and impose an instance-level area-sensitive loss. Once clutter is removed, we inpaint geometry and texture in the resulting holes by merging inpainted RGB-D images. This requires novel voting and pruning strategies that guarantee multi-view consistency across individually inpainted images for mesh reconstruction. Experiments on ScanNet and Matterport3D dataset show that our method outperforms baselines for clutter segmentation and 3D inpainting, both visually and quantitatively. Project Page: https://weify627.github.io/clutter/."}, "cited_paper_content": {"title": "Context Encoders: Feature Learning By Inpainting", "abstract": "We present an unsupervised visual feature learning algorithm driven by context-based pixel prediction. By analogy with auto-encoders, we propose Context Encoders -- a convolutional neural network trained to generate the contents of an arbitrary image region conditioned on its surroundings. In order to succeed at this task, context encoders need to both understand the content of the entire image, as well as produce a plausible hypothesis for the missing part(s). When training context encoders, we have experimented with both a standard pixel-wise reconstruction loss, as well as a reconstruction plus an adversarial loss. The latter produces much sharper results because it can better handle multiple modes in the output. We found that a context encoder learns a representation that captures not just appearance but also the semantics of visual structures. We quantitatively demonstrate the effectiveness of our learned features for CNN pre-training on classification, detection, and segmentation tasks. Furthermore, context encoders can be used for semantic inpainting tasks, either stand-alone or as initialization for non-parametric methods."}, "keywords": ["adversarial training"], "citation_intent": "method"}
{"citing_id": "2303.04887v1", "cited_id": "1905.00414", "section_title": "Handle Extrem Memory Constraints With Partial Tranining", "citation": "This mechanism would not incur significant performance degradation because the input-side layers learn similar representations on different datasets #REFR , and clients with sufficient memory will provide model parameters of these input-side layers in the FL aggregation phase. Figure 6 emeprically validates such a strategy.", "text_before_citation": ["According to the memory consumption analysis in Section 3.2, the memory bottleneck of training a neural network is related to the block with the largest activations, which Figure 6 : Correspondences between layers of different local neural networks trained from private datasets in FL under non-IID distribution.", "We observe that early layers, but not later layers, learn similar representations. some devices may still not afford.", "These \"large\" blocks are usually the layers close to the input side.", "To tackle this issue, we borrow the idea of partial training in FEDEPTH, where we skip the first few blocks that are too large to fit even after the finnest blocks decomposition."], "text_after_citation": ["We train a customized 14-layer ResNet (13 convolution layers and one classifier) on MNIST with 20 clients under non-IID distribution, respectively and measure the similarity of neural network representations, using both canonical correlation analysis (CCA) and centered kernel alignment (CKA) #OTHEREFR ."], "citing_paper_content": {"title": "Memory-Adaptive Depth-Wise Heterogenous Federated Learning", "abstract": "Federated learning is a promising paradigm that allows multiple clients to collaboratively train a model without sharing the local data. However, the presence of heterogeneous devices in federated learning, such as mobile phones and IoT devices with varying memory capabilities, would limit the scale and hence the performance of the model could be trained. The mainstream approaches to address memory limitations focus on width-slimming techniques, where different clients train subnetworks with reduced widths locally and then the server aggregates the subnetworks. The global model produced from these methods suffers from performance degradation due to the negative impact of the actions taken to handle the varying subnetwork widths in the aggregation phase. In this paper, we introduce a memory-adaptive depth-wise learning solution in FL called FEDEPTH, which adaptively decomposes the full model into blocks according to the memory budgets of each client and trains blocks sequentially to obtain a full inference model. Our method outperforms state-of-the-art approaches, achieving 5% and more than 10% improvements in top-1 accuracy on CIFAR-10 and CIFAR-100, respectively. We also demonstrate the effectiveness of depth-wise fine-tuning on ViT."}, "cited_paper_content": {"title": "Similarity Of Neural Network Representations Revisited", "abstract": "Recent work has sought to understand the behavior of neural networks by comparing representations between layers and between different trained models. We examine methods for comparing neural network representations based on canonical correlation analysis (CCA). We show that CCA belongs to a family of statistics for measuring multivariate similarity, but that neither CCA nor any other statistic that is invariant to invertible linear transformation can measure meaningful similarities between representations of higher dimension than the number of data points. We introduce a similarity index that measures the relationship between representational similarity matrices and does not suffer from this limitation. This similarity index is equivalent to centered kernel alignment (CKA) and is also closely connected to CCA. Unlike CCA, CKA can reliably identify correspondences between representations in networks trained from different initializations."}, "keywords": ["input-side layers", "similar representations"], "citation_intent": "background"}
{"citing_id": "2304.01371v1", "cited_id": "1312.4511", "section_title": "Limitations", "citation": "Second, we overlooked YT channels shared on Twitter to safeguard our analysis from Twitter users who just advertise their own (or others) YouTube channel #REFR .", "text_before_citation": ["There is a number of limitations in our study.", "First, neither Twitter nor YouTube provides any supplementary information about account suspension and video moderation, as well as we do not know when these interventions occurred.", "As a result, there is no guarantee that YT videos were still online when re-shared through retweets on Twitter, but we can confidently assume they were not moderated yet when shared in an original tweet.", "In addition, we acknowledge our analyses, like several previous works #OTHEREFR , might be biased towards moderated YT content which includes not only videos violating YouTube policies but also those removed by their publishers for any reason."], "text_after_citation": ["However, this choice might prevent us to consider another potential source of harmful YT content on Twitter.", "Third, the partition strategy to define the two groups of mobilizers is quite conservative as we considered Twitter users who never share moderated videos (NMYT mobilizers) and those who mostly share moderated videos (MYT mobilizers).", "The investigation of the users sharing a limited number of moderated videos is left for future research."], "citing_paper_content": {"title": "The Interconnected Nature Of Online Harm And Moderation: Investigating The Cross-Platform Spread Of Harmful Content Between Youtube And Twitter", "abstract": "The proliferation of harmful content shared online poses a threat to online information integrity and the integrity of discussion across platforms. Despite various moderation interventions adopted by social media platforms, researchers and policymakers are calling for holistic solutions. This study explores how a target platform could leverage content that has been deemed harmful on a source platform by investigating the behavior and characteristics of Twitter users responsible for sharing moderated YouTube videos. Using a large-scale dataset of 600M tweets related to the 2020 U.S. election, we find that moderated Youtube videos are extensively shared on Twitter and that users who share these videos also endorse extreme and conspiratorial ideologies. A fraction of these users are eventually suspended by Twitter, but they do not appear to be involved in state-backed information operations. The findings of this study highlight the complex and interconnected nature of harmful cross-platform information diffusion, raising the need for cross-platform moderation strategies. CCS Concepts: \u2022 Information systems \u2192 Social networks; \u2022 Human-centered computing \u2192 Empirical studies in collaborative and social computing."}, "cited_paper_content": {"title": "Who Watches (And Shares) What On Youtube? And When? Using Twitter To Understand Youtube Viewership", "abstract": "By combining multiple social media datasets, it is possible to gain insight into each dataset that goes beyond what could be obtained with either individually. In this paper we combine user-centric data from Twitter with video-centric data from YouTube to build a rich picture of who watches and shares what on YouTube. We study 87K Twitter users, 5.6 million YouTube videos and 15 million video sharing events from user-, video- and sharing-event-centric perspectives. We show that features of Twitter users correlate with YouTube features and sharing-related features. For example, urban users are quicker to share than rural users. We find a superlinear relationship between initial Twitter shares and the final amounts of views. We discover that Twitter activity metrics play more role in video popularity than mere amount of followers. We also reveal the existence of correlated behavior concerning the time between video creation and sharing within certain timescales, showing the time onset for a coherent response, and the time limit after which collective responses are extremely unlikely. Response times depend on the category of the video, suggesting Twitter video sharing is highly dependent on the video content. To the best of our knowledge, this is the first large-scale study combining YouTube and Twitter data, and it reveals novel, detailed insights into who watches (and shares) what on YouTube, and when."}, "keywords": ["Twitter users"], "citation_intent": "background"}
{"citing_id": "2303.02266v1", "cited_id": "1909.07972", "section_title": "Iii. Trajectory Optimization", "citation": "In comparison with #REFR , this theorem shows the separate effects of the sensor noise levels on the accuracy of the trained NN, allowing noise-aware optimization of the drone trajectory.", "text_before_citation": ["Iterative application of (1) over a series of T aggregation rounds, gives a bound on the expected convergence speed and final performance of the trained NN.", "Assuming {\u03a6t} T t=1 < 1, the speed of convergence is determined only by {\u03a6t} T t=1 values, while the expected final performance of the trained NN is determined by the interplay of the collective effects from Jt, Kt, and \u03a6t terms over T aggregation rounds.", "In an ideal scenario without packet errors and sensor noises, i.e.", "assuming ei,t = 0 and \u03c3 2 i = 0, the expected training loss would asymptotically converge to the optimum loss value, i.e.", "limT \u2192\u221e E(F (\u03c9T +1)) = F (\u03c9 * ), while the packet errors and sensor noises have a destructive effect on the convergence speed and final performance of the trained NN."], "text_after_citation": ["The packet errors increase \u03a6t values, thereby slowing down the convergence.", "The sensor noises seem to have no effect on the convergence speed, but degrade the final performance of the trained NN by introducing the additional error terms Kt values.", "To speed up convergence and improve the final performance of the NN, the drone should position itself closer to devices with larger local datasets and less noisy data.", "In the following, we study the optimum placement and trajectory of the drone, based on the above insights."], "citing_paper_content": {"title": "Collaborative Learning With A Drone Orchestrator", "abstract": "In this paper, the problem of drone-assisted collaborative learning is considered. In this scenario, swarm of intelligent wireless devices train a shared neural network (NN) model with the help of a drone. Using its sensors, each device records samples from its environment to gather a local dataset for training. The training data is severely heterogeneous as various devices have different amount of data and sensor noise level. The intelligent devices iteratively train the NN on their local datasets and exchange the model parameters with the drone for aggregation. For this system, the convergence rate of collaborative learning is derived while considering data heterogeneity, sensor noise levels, and communication errors, then, the drone trajectory that maximizes the final accuracy of the trained NN is obtained. The proposed trajectory optimization approach is aware of both the devices data characteristics (i.e., local dataset size and noise level) and their wireless channel conditions, and significantly improves the convergence rate and final accuracy in comparison with baselines that only consider data characteristics or channel conditions. Compared to state-of-the-art baselines, the proposed approach achieves an average 3.85% and 3.54% improvement in the final accuracy of the trained NN on benchmark datasets for image recognition and semantic segmentation tasks, respectively. Moreover, the proposed framework achieves a significant speedup in training, leading to an average 24% and 87% saving in the drone's hovering time, communication overhead, and battery usage, respectively for these tasks."}, "cited_paper_content": {"title": "A Joint Learning And Communications Framework For Federated Learning Over Wireless Networks", "abstract": "In this paper, the problem of training federated learning (FL) algorithms over a realistic wireless network is studied. In particular, in the considered model, wireless users execute an FL algorithm while training their local FL models using their own data and transmitting the trained local FL models to a base station (BS) that will generate a global FL model and send it back to the users. Since all training parameters are transmitted over wireless links, the quality of the training will be affected by wireless factors such as packet errors and the availability of wireless resources. Meanwhile, due to the limited wireless bandwidth, the BS must select an appropriate subset of users to execute the FL algorithm so as to build a global FL model accurately. This joint learning, wireless resource allocation, and user selection problem is formulated as an optimization problem whose goal is to minimize an FL loss function that captures the performance of the FL algorithm. To address this problem, a closed-form expression for the expected convergence rate of the FL algorithm is first derived to quantify the impact of wireless factors on FL. Then, based on the expected convergence rate of the FL algorithm, the optimal transmit power for each user is derived, under a given user selection and uplink resource block (RB) allocation scheme. Finally, the user selection and uplink RB allocation is optimized so as to minimize the FL loss function. Simulation results show that the proposed joint federated learning and communication framework can reduce the FL loss function value by up to 10% and 16%, respectively, compared to: 1) An optimal user selection algorithm with random resource allocation and 2) a standard FL algorithm with random user selection and resource allocation."}, "keywords": ["drone trajectory", "trained NN"], "citation_intent": "result"}
{"citing_id": "2303.07929v1", "cited_id": "1904.05059", "section_title": "Experiments Results And Analysis", "citation": "C3AE #REFR is a lightweight network with a multi-scale image input, aiming to show that the designed DAA is also a lightweight operation.", "text_before_citation": ["The DAA operation we designed can be transplanted to different feature extraction networks for age estimation. Our methods are expressed as Resnet18+DAA and C3AE+DAA."], "text_after_citation": ["Following the design in C3AE #OTHEREFR , we also give the \"plain\" configuration, applying single-scale input.", "Comparisons on Morph dataset.", "As is shown in Table  1 , the upper part of the table is the method with large parameters, while the lower part is the lightweight network.", "Compared with Ranking-CNN, AP, MV, and PML, our proposed method achieved the best MAEs value of 2.06 with the pre-trained operation.", "At the same time, the parameters of our approach are significantly reduced compared with other methods."], "citing_paper_content": {"title": "Daa: A Delta Age Adain Operation For Age Estimation Via Binary Code Transformer", "abstract": "Naked eye recognition of age is usually based on comparison with the age of others. However, this idea is ignored by computer tasks because it is difficult to obtain representative contrast images of each age. Inspired by the transfer learning, we designed the Delta Age AdaIN (DAA) operation to obtain the feature difference with each age, which obtains the style map of each age through the learned values representing the mean and standard deviation. We let the input of transfer learning as the binary code of age natural number to obtain continuous age feature information. The learned two groups of values in Binary code mapping are corresponding to the mean and standard deviation of the comparison ages. In summary, our method consists of four parts: FaceEncoder, DAA operation, Binary code mapping, and AgeDecoder modules. After getting the delta age via AgeDecoder, we take the average value of all comparison ages and delta ages as the predicted age. Compared with state-of-the-art methods, our method achieves better performance with fewer parameters on multiple facial age datasets."}, "cited_paper_content": {"title": "C3Ae: Exploring The Limits Of Compact Model For Age Estimation", "abstract": "Age estimation is a classic learning problem in computer vision. Many larger and deeper CNNs have been proposed with promising performance, such as AlexNet, VggNet, GoogLeNet and ResNet. However, these models are not practical for the embedded/mobile devices. Recently, MobileNets and ShuffleNets have been proposed to reduce the number of parameters, yielding lightweight models. However, their representation has been weakened because of the adoption of depth-wise separable convolution. In this work, we investigate the limits of compact model for small-scale image and propose an extremely Compact yet efficient Cascade Context-based Age Estimation model(C3AE). This model possesses only 1/9 and 1/2000 parameters compared with MobileNets/ShuffleNets and VggNet, while achieves competitive performance. In particular, we re-define age estimation problem by two-points representation, which is implemented by a cascade model. Moreover, to fully utilize the facial context information, multi-branch CNN network is proposed to aggregate multi-scale context. Experiments are carried out on three age estimation datasets. The state-of-the-art performance on compact model has been achieved with a relatively large margin."}, "keywords": ["multi-scale image input"], "citation_intent": "method"}
{"citing_id": "2304.08648v1", "cited_id": "1408.4156", "section_title": "Upper Bound On The Competitive Ratio Of Move To Front", "citation": "This significantly improves the result of Kamali and L\u00f3pez-Ortiz #REFR , who showed that Move To Front has an asymptotic competitive ratio of 6\u00b5 + 7, i.e., for any item list R, they showed cost(MF, R) \u2264 (6\u00b5 + 7) \u2022 OPT(R) + 3(\u00b5 + 1).", "text_before_citation": ["In this section we prove the first main result of our paper.", "For d = 1, our result implies that Move To Front has a competitive ratio of at most 2\u00b5 + 2."], "text_after_citation": ["Our result also nearly settles their conjecture of the CR being 2\u00b5 + 1.", "Their analysis decomposes the active span into segments of length (\u00b5 + 1) and compares the cost of OPT with the cost of Move To Front in each such interval. It turns out that this decomposition is sub-optimal.", "Instead, we directly use the nature of the Move To Front algorithm and develop a novel decomposition of the usage periods of each bin B into intervals based on whether or not in the interval B is the most recently used bin. We now prove Theorem 2.", "Suppose Move To Front uses m bins B 1 , B 2 , . . . , B m on an input sequence R.", "As mentioned earlier, we can assume that \u222a r\u2208R I(r) = [0, span(R)) and that the usage period of each bin is We say a bin is a leader at time t if it is in the front of the list L at time t."], "citing_paper_content": {"title": "Dynamic Vector Bin Packing For Online Resource Allocation In The Cloud", "abstract": "Several cloud-based applications, such as cloud gaming, rent servers to execute jobs which arrive in an online fashion. Each job has a resource demand, such as GPU requirement, and must be dispatched to a cloud server which has enough resources to execute the job, which departs after its completion. Under the \"pay-as-you-go\" billing model, the server rental cost is proportional to the total time that servers are actively running jobs. The problem of efficiently allocating a sequence of online jobs to servers without exceeding the resource capacity of any server while minimizing total server usage time can be modelled as a variant of the dynamic bin packing problem (DBP), called MinUsageTime DBP [21]. In this work, we initiate the study of the problem with multi-dimensional resource demands (e.g. CPU/GPU usage, memory requirement, bandwidth usage, etc.), called MinUsageTime Dynamic Vector Bin Packing (DVBP). We study the competitive ratio (CR) of Any Fit packing algorithms for this problem. We show almost-tight bounds on the CR of three specific Any Fit packing algorithms, namely First Fit, Next Fit, and Move To Front. We prove that the CR of Move To Front is at most (2\u00b5 + 1)d + 1, where \u00b5 is the ratio of the max/min item durations. For d = 1, this implies a significant improvement over the previously known upper bound of 6\u00b5 + 7 [18]. We then prove the CR of First Fit and Next Fit are bounded by (\u00b5 + 2)d + 1 and 2\u00b5d + 1, respectively. Next, we prove a lower bound of (\u00b5 + 1)d on the CR of any Any Fit packing algorithm, an improved lower bound of 2\u00b5d for Next Fit, and a lower bound of 2\u00b5 for Move To Front in the 1-D case. All our bounds improve or match the best-known bounds for the 1-D case. Finally, we experimentally study the average-case performance of these algorithms on randomly generated synthetic data, and observe that Move To Front outperforms other Any Fit packing algorithms."}, "cited_paper_content": {"title": "Efficient Online Strategies For Renting Servers In The Cloud", "abstract": "When scheduling jobs for systems in the cloud, we often deal with jobs that arrive and depart in an online manner. Each job should be assigned to a server upon arrival. Jobs are annotated with sizes which define the amount of resources that they need. Servers have uniform capacity and, at all times, the total size of jobs assigned to a server should not exceed this capacity. This setting is closely related to the classic bin packing problem. The difference is that, in bin packing, the objective is to minimize the total number of used servers. In the cloud systems, however, the cost for each server is proportional to the length of the time interval it is rented for, and the goal is to minimize the cost involved in renting all used servers. Recently, certain bin packing strategies were considered for renting servers in the cloud [Li et al. SPAA\u201914]. There, it is proved that all Any-Fit strategies have a competitive ratio of at least \u03bc, where \u03bc is the max/min interval length ratio of jobs. It is also proved that First Fit has a competitive ratio of 2\u03bc + 13, while Best Fit is not competitive at all. We observe that the lower bound of \u03bc extends to all online algorithms. We also prove that, surprisingly, Next Fit algorithm has a competitive ratio of at most 2 \u03bc + 1. We also show that a variant of Next Fit achieves a competitive ratio of K \u00d7 max {1,\u03bc/(K \u2212 1)} + 1, where K is a parameter of the algorithm. In particular, if the value of \u03bc is known, the algorithm has a competitive ratio of \u03bc + 2; this improves upon the existing upper bound of \u03bc + 8. Finally, we introduce a simple algorithm called Move To Front (Mtf) which has a competitive ratio of at most 6\u03bc + 8. We experimentally study the average-case performance of different algorithms and observe that the typical behaviour of Mtf is better than other algorithms."}, "keywords": ["asymptotic competitive ratio", "cost(MF"], "citation_intent": "result"}
{"citing_id": "2304.11970v1", "cited_id": "1905.05172", "section_title": "Overview Of Gsdf", "citation": "To alleviate the limitation, inspired by #REFR , we apply the geometry transformation to extract aligned local visual features.", "text_before_citation": ["To improve the hand pose estimation, we propose to first predict sparse 3D joint locations j h from the image and then use inverse kinematics to derive pose transformations \u03b8 h from the predicted joints.", "In this way, we are able to obtain kinematic features e h and e o for the hand and the object respectively.", "The visual feature encodes the visual appearance for the point x to provide more shape details.", "Prior works #OTHEREFR use the same global visual feature for all the points, e.g., averaging the feature map of a SDF feature encoder on the spatial dimension.", "Such global visual features suffers from imprecise geometry alignment between a point and its visual appearance."], "text_after_citation": ["Moreover, to address hard cases with occlusions and motion blur in a single image I t , we propose to enhance the local visual feature with its temporal contexts from videos using a spatio-temporal transformer.", "We denote the local visual feature of a point as e v .", "Finally, we concatenate the kinematic feature and local visual feature to predict the signed distance for x:", "EQUATION", "where f h and f o are the hand SDF decoder and the object SDF decoder respectively."], "citing_paper_content": {"title": "Gsdf: Geometry-Driven Signed Distance Functions For 3D Hand-Object Reconstruction", "abstract": "Signed distance functions (SDFs) is an attractive framework that has recently shown promising results for 3D shape reconstruction from images. SDFs seamlessly generalize to different shape resolutions and topologies but lack explicit modelling of the underlying 3D geometry. In this work, we exploit the hand structure and use it as guidance for SDF-based shape reconstruction. In particular, we address reconstruction of hands and manipulated objects from monocular RGB images. To this end, we estimate poses of hands and objects and use them to guide 3D reconstruction. More specifically, we predict kinematic chains of pose transformations and align SDFs with highly-articulated hand poses. We improve the visual features of 3D points with geometry alignment and further leverage temporal information to enhance the robustness to occlusion and motion blurs. We conduct extensive experiments on the challenging ObMan and DexYCB benchmarks and demonstrate significant improvements of the proposed method over the state of the art."}, "cited_paper_content": {"title": "Pifu: Pixel-Aligned Implicit Function For High-Resolution Clothed Human Digitization", "abstract": "We introduce Pixel-aligned Implicit Function (PIFu), a highly effective implicit representation that locally aligns pixels of 2D images with the global context of their corresponding 3D object. Using PIFu, we propose an end-to-end deep learning method for digitizing highly detailed clothed humans that can infer both 3D surface and texture from a single image, and optionally, multiple input images. Highly intricate shapes, such as hairstyles, clothing, as well as their variations and deformations can be digitized in a unified way. Compared to existing representations used for 3D deep learning, PIFu can produce high-resolution surfaces including largely unseen regions such as the back of a person. In particular, it is memory efficient unlike the voxel representation, can handle arbitrary topology, and the resulting surface is spatially aligned with the input image. Furthermore, while previous techniques are designed to process either a single image or multiple views, PIFu extends naturally to arbitrary number of views. We demonstrate high-resolution and robust reconstructions on real world images from the DeepFashion dataset, which contains a variety of challenging clothing types. Our method achieves state-of-the-art performance on a public benchmark and outperforms the prior work for clothed human digitization from a single image."}, "keywords": ["aligned local visual"], "citation_intent": "method"}
{"citing_id": "2304.12541v1", "cited_id": "1808.04730", "section_title": "Physics-Informed Invertible Neural Networks", "citation": "Consequently, z is assumed to be independent of c and follow a tractable probability distribution, where the uniqueness of the mapping between \u03bb and [c, z] #REFR is constructed.", "text_before_citation": ["Similar to DeepONets #OTHEREFR , the separate representation of the solutions is employed by", "EQUATION", "where c i represents the expansion coefficient, and \u03c6 i (x) is the basis function defined on the spatial domain.", "The INN is designed to learn the bijective mapping between the parametric input \u03bb and the unique pairs [c, z] of expansion coefficients and random latent variables.", "The random latent variable z is utilized to capture the inherent information loss in the output layer, which is not contained in c."], "text_after_citation": ["In this work, the standard Gaussian distribution for z is adopted.", "Furthermore, the INN architecture consists of alternating layers and affine coupling layers.", "The partitioned input of the coupling layer remains unchanged for one subset, while the other subset undergoes a coupling operation.", "Then, the two parts of output are exchanged in alternating layers, which ensure that every dimension of the input is updated #OTHEREFR .", "Then, a fully-connected neural network named NB-Net is employed to obtain the spatial basis functions, and the outputs of the NB-Net can be orthogonalized."], "citing_paper_content": {"title": "Efficient Bayesian Inference Using Physics-Informed Invertible Neural Networks For Inverse Problems", "abstract": "In the paper, we propose a novel approach for solving Bayesian inverse problems with physics-informed invertible neural networks (PI-INN). The architecture of PI-INN consists of two sub-networks: an invertible neural network (INN) and a neural basis network (NB-Net). The invertible map between the parametric input and the INN output with the aid of NB-Net is constructed to provide a tractable estimation of the posterior distribution, which enables efficient sampling and accurate density evaluation. Furthermore, the loss function of PI-INN includes two components: a residual-based physics-informed loss term and a new independence loss term. The presented independence loss term can Gaussianize the random latent variables and ensure statistical independence between two parts of INN output by effectively utilizing the estimated density function. Several numerical experiments are presented to demonstrate the efficiency and accuracy of the proposed PI-INN, including inverse kinematics, inverse problems of the 1-d and 2-d diffusion equations, and seismic traveltime tomography."}, "cited_paper_content": {"title": "Analyzing Inverse Problems With Invertible Neural Networks", "abstract": "In many tasks, in particular in natural science, the goal is to determine hidden system parameters from a set of measurements. Often, the forward process from parameter- to measurement-space is a well-defined function, whereas the inverse problem is ambiguous: one measurement may map to multiple different sets of parameters. In this setting, the posterior parameter distribution, conditioned on an input measurement, has to be determined. We argue that a particular class of neural networks is well suited for this task -- so-called Invertible Neural Networks (INNs). Although INNs are not new, they have, so far, received little attention in literature. While classical neural networks attempt to solve the ambiguous inverse problem directly, INNs are able to learn it jointly with the well-defined forward process, using additional latent output variables to capture the information otherwise lost. Given a specific measurement and sampled latent variables, the inverse pass of the INN provides a full distribution over parameter space. We verify experimentally, on artificial data and real-world problems from astrophysics and medicine, that INNs are a powerful analysis tool to find multi-modalities in parameter space, to uncover parameter correlations, and to identify unrecoverable parameters."}, "keywords": ["mapping", "tractable probability distribution"], "citation_intent": "background"}
{"citing_id": "2305.02024v1", "cited_id": "2003.08505", "section_title": "Discussion.", "citation": "The setup of #REFR is not directly usable for experiments with the RS@k loss, as large batch sizes are crucial to estimate recall@k accurately. Furthermore, their setup does not allow mixup.", "text_before_citation": ["The methods in the literature use different embedding sizes, d, therefore, the models for the RS@k loss are trained with two embedding sizes of d = 128 and d = 512 for image retrieval benchmarks #OTHEREFR - #OTHEREFR , and d = 2048 for ROxford/RParis #OTHEREFR , to allow a fair comparison.", "In the standard split, the image retrieval benchmarks #OTHEREFR - #OTHEREFR do not contain an explicit validation set; as a result, image retrieval methods often tune the hyper-parameters on the test set, leading to the issue of training with test set feedback.", "This issue has been studied in #OTHEREFR , which proposes to train different methods with identical hyper-parameters."], "text_after_citation": ["Therefore, instead of following #OTHEREFR , the issue is eliminated by using a part of the training set for validation as described above."], "citing_paper_content": {"title": "", "abstract": "I hereby declare I have written this doctoral dissertation independently and quoted all the sources of information used in accordance with methodological instructions on ethical principles for writing an academic dissertation. Moreover, I state that this dissertation has neither been submitted nor accepted for any other degree. The results presented in this dissertation were were achieved during my own research in cooperation with my dissertation supervisor Jiri Matas published in [1]-[8]. During the Ph.D., I also collaborated with several researchers on multiple projects, with R. Manmatha, Srikar"}, "cited_paper_content": {"title": "A Metric Learning Reality Check", "abstract": "Deep metric learning papers from the past four years have consistently claimed great advances in accuracy, often more than doubling the performance of decade-old methods. In this paper, we take a closer look at the field to see if this is actually true. We find flaws in the experimental setup of these papers, and propose a new way to evaluate metric learning algorithms. Finally, we present experimental results that show that the improvements over time have been marginal at best."}, "keywords": ["RS@k loss"], "citation_intent": "background"}
{"citing_id": "2304.02595v1", "cited_id": "1412.6980", "section_title": "Results", "citation": "We use the Sunspot time series 5 data and Abalone #REFR datasets for regression problems.", "text_before_citation": [], "text_after_citation": ["The Abalone dataset provides the ring age for Abalone based on eight features that represent physical properties such as length, width, and weight and associated target feature, i.e., the ring age.", "Determining the age of Abalone is difficult as requires cutting the shell and counting the number of rings using a microscope.", "However, other physical measurements can be used to predict the age and a model can be developed to use the physical features to determine the ring age.", "Sunspots are regions of reduced surface temperature in the Sun's photosphere caused by concentrations of magnetic field flux, and appear as spots darker than the surrounding areas.", "The Sunspot cycles are about every eleven years and over the solar cycle, the number of occurrence of Sunspot change more rapidly."], "citing_paper_content": {"title": "Bayesian Neural Networks Via Mcmc: A Python-Based Tutorial", "abstract": "Bayesian inference provides a methodology for parameter estimation and uncertainty quantification in machine learning and deep learning methods. Variational inference and Markov Chain Monte-Carlo (MCMC) sampling techniques are used to implement Bayesian inference. In the past three decades, MCMC methods have faced a number of challenges in being adapted to larger models (such as in deep learning) and big data problems. Advanced proposals that incorporate gradients, such as a Langevin proposal distribution, provide a means to address some of the limitations of MCMC sampling for Bayesian neural networks. Furthermore, MCMC methods have typically been constrained to use by statisticians and are still not prominent among deep learning researchers. We present a tutorial for MCMC methods that covers simple Bayesian linear and logistic models, and Bayesian neural networks. The aim of this tutorial is to bridge the gap between theory and implementation via coding, given a general sparsity of libraries and tutorials to this end. This tutorial provides code in Python with data and instructions that enable their use and extension. We provide results for some benchmark problems showing the strengths and weaknesses of implementing the respective Bayesian models via MCMC. We highlight the challenges in sampling multi-modal posterior distributions in particular for the case of Bayesian neural networks, and the need for further improvement of convergence diagnosis."}, "cited_paper_content": {"title": "Adam: A Method For Stochastic Optimization", "abstract": "We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm."}, "keywords": ["Abalone datasets"], "citation_intent": "method"}
{"citing_id": "2303.08115v1", "cited_id": "1612.00796", "section_title": "Related Work", "citation": "Studies on continual reinforcement learning seek to determine how learning one or more tasks can accelerate learning of other tasks, as well as avoid ruinous forgetting #REFR .", "text_before_citation": ["In policy transfer, behaviors are transferred to a new agent (student) by a pre-trained agent (teacher).", "The student is trained to minimize the KullbackLeibler divergence to the teacher using policy distillation, for example [36] .", "Other methods, re-use policies from source tasks directly to create a student policy #OTHEREFR .", "On the other hand, a pre-learned state representation used in feature transfer encourages the agent to explore when it is presented with tasks #OTHEREFR .", "In exploration transfer, in a task-agnostic and environment-centric scenario, the agent is first attempted by defining an intrinsic reward to visit the states as more as possible especially interesting states in multiple environments instead of learning the main task, and then this prior knowledge is transferred as a bias to the main environment in which the main task is to be trained #OTHEREFR ."], "text_after_citation": ["Meta reinforcement learning and curriculum learning, on the other hand, focus on exploiting underlying similarities between tasks to speed up the learning process of new tasks #OTHEREFR .", "Many studies have been done under the Lifelong RL framework, where the tasks change following a specific distribution #OTHEREFR .", "In #OTHEREFR , by spending a lot of computational costs, knowledge transfer is done through transferring initializers.", "Most are only compatible with ( , )-PAC algorithms #OTHEREFR In the mentioned incrementally learning task methods, it is assumed that there are several tasks/environments that, with the help of the experience gained on one or multiple tasks/environments, we seek to accelerate the learning of the same task in the new environment or the new task in the same environment.", "As a result, the most important hypothesis is that tasks/environments are very similar to each other, which is not always consistent with reality and deprives us of the possibility of using these methods on any task and any environment."], "citing_paper_content": {"title": "Human-Inspired Framework To Accelerate Reinforcement Learning", "abstract": "While deep reinforcement learning (RL) is becoming an integral part of good decision-making in data science, it is still plagued with sample inefficiency. This can be challenging when applying deep-RL in real-world environments where physical interactions are expensive and can risk system safety. To improve the sample efficiency of RL algorithms, this paper proposes a novel humaninspired framework that facilitates fast exploration and learning for difficult RL tasks. The main idea is to first provide the learning agent with simpler but similar tasks that gradually grow in difficulty and progress toward the main task. The proposed method requires no pre-training phase. Specifically, the learning of simpler tasks is only done for one iteration. The generated knowledge could be used by any transfer learning, including value transfer and policy transfer, to reduce the sample complexity while not adding to the computational complexity. So, it can be applied to any goal, environment, and reinforcement learning algorithm-both value-based methods and policy-based methods and both tabular methods and deep-RL methods. We have evaluated our proposed framework on both a simple Random Walk for illustration purposes and on more challenging optimal control problems with constraint. The experiments show the good performance of our proposed framework in improving the sample efficiency of RL-learning algorithms, especially when the main task is difficult."}, "cited_paper_content": {"title": "Overcoming Catastrophic Forgetting In Neural Networks", "abstract": "The ability to learn tasks in a sequential fashion is crucial to the development of artificial intelligence. Until now neural networks have not been capable of this and it has been widely thought that catastrophic forgetting is an inevitable feature of connectionist models. We show that it is possible to overcome this limitation and train networks that can maintain expertise on tasks that they have not experienced for a long time. Our approach remembers old tasks by selectively slowing down learning on the weights important for those tasks. We demonstrate our approach is scalable and effective by solving a set of classification tasks based on a hand-written digit dataset and by learning several Atari 2600 games sequentially."}, "keywords": ["continual reinforcement learning"], "citation_intent": "background"}
{"citing_id": "2303.07463v1", "cited_id": "1910.04630", "section_title": "The Landau-Lifshitz-Gilbert-Equation. The Strong Form Of The Llg Equations Is To", "citation": "Note that in case one of the weak solutions is also a strong solution, we have uniqueness #REFR .", "text_before_citation": ["T(m(t)) = \u03d5 \u2208 H 1 (\u2126) 3 | m(t) \u2022 \u03d5 = 0 .", "For ease of notation we will omit the argument t in the subsequel and arrive formally at the following weak equation for m", "\u03b1(\u2202 t m, \u03d5) \u2126 + (m \u00d7 \u2202 t m, \u03d5) \u2126 + C e (\u2207m, \u2207\u03d5) \u2126 = (H ext , \u03d5) \u2126 for all \u03d5 \u2208 T(m) .", "Existence of weak solutions m \u2208 H 1 (\u2126) 3 is well-known (see, e.g., [27] ) and follows with a standard Galerkin-Fadeo approximation and compactness arguments.", "Uniqueness of weak solutions is open in general and there are non-uniqueness results for the related harmonic map heat flow #OTHEREFR ."], "text_after_citation": ["In terms of regularity, it is suspected that even smooth initial conditions can lead to blow up of \u2207m in L \u221e (\u2126) #OTHEREFR in finite time, however, no proof is known.", "For LLG on smooth bounded domains with Neumann boundary conditions, smooth initial data close to constants lead to arbitrarily smooth solutions #OTHEREFR ."], "citing_paper_content": {"title": "Adaptive Mesh Refinement For The Landau-Lifshitz-Gilbert Equation", "abstract": "We propose a new adaptive algorithm for the approximation of the Landau-Lifshitz-Gilbert equation via a higher-order tangent plane scheme. We show that the adaptive approximation satisfies an energy inequality and demonstrate numerically, that the adaptive algorithm outperforms uniform approaches."}, "cited_paper_content": {"title": "Weak-Strong Uniqueness For The Landau-Lifshitz-Gilbert Equation In Micromagnetics", "abstract": "We consider the time-dependent Landau-Lifshitz-Gilbert equation. We prove that each weak solution coincides with the (unique) strong solution, as long as the latter exists in time. Unlike available results in the literature, our analysis also includes the physically relevant lower-order terms like Zeeman contribution, anisotropy, stray field, and the Dzyaloshinskii-Moriya interaction (which accounts for the emergence of magnetic Skyrmions). Moreover, our proof gives a template on how to approach weak-strong uniqueness for even more complicated problems, where LLG is (nonlinearly) coupled to other (nonlinear) PDE systems."}, "keywords": ["weak solutions"], "citation_intent": "background"}
{"citing_id": "2304.09507v1", "cited_id": "1904.03485", "section_title": "Overview", "citation": "The pixel-shuffle downsampling (PD) #REFR also loosens the spatial correlation of the noise, but it generates severe checkerboard artifacts.", "text_before_citation": ["We propose a novel loss function that can be directly optimized on the input image without loss of information.", "It consists of self-supervised loss and downsampled invariance loss that controls the extent of the blindness.", "Our main idea of the downsampled invariance loss is to make a blind-spot network serve as regularization of the same network while preserving network parameters.", "To this end, we propose a conditional blind-spot network (denoted C-BSN in the figure) to selectively mask the center pixel in the receptive field.", "In addition, we introduce Random Subsampler (RS) to decorrelate noise spatially."], "text_after_citation": ["On the contrary, since our RS draws a pixel randomly from each grid, it does not produce such artifacts.", "We denote the noisy input image as x and the corresponding clean image as y.", "For brevity, the channel dimension is omitted, and spatial dimensions are vectorized, i.e., x, y \u2208 R m ."], "citing_paper_content": {"title": "Self-Supervised Image Denoising With Downsampled Invariance Loss And Conditional Blind-Spot Network", "abstract": "There have been many image denoisers using deep neural networks, which outperform conventional model-based methods by large margins. Recently, self-supervised methods have attracted attention because constructing a large real noise dataset for supervised training is an enormous burden. The most representative self-supervised denoisers are based on blind-spot networks, which exclude the receptive field's center pixel. However, excluding any input pixel is abandoning some information, especially when the input pixel at the corresponding output position is excluded. In addition, a standard blind-spot network fails to reduce real camera noise due to the pixel-wise correlation of noise, though it successfully removes independently distributed synthetic noise. Hence, to realize a more practical denoiser, we propose a novel self-supervised training framework that can remove real noise. For this, we derive the theoretic upper bound of a supervised loss where the network is guided by the downsampled blinded output. Also, we design a conditional blind-spot network (C-BSN), which selectively controls the blindness of the network to use the center pixel information. Furthermore, we exploit a random subsampler to decorrelate noise spatially, making the C-BSN free of visual artifacts that were often seen in downsample-based methods. Extensive experiments show that the proposed C-BSN achieves state-of-the-art performance on real-world datasets as a self-supervised denoiser and shows qualitatively pleasing results without any postprocessing or refinement."}, "cited_paper_content": {"title": "When Awgn-Based Denoiser Meets Real Noises", "abstract": "Discriminative learning based image denoisers have achieved promising performance on synthetic noise such as the additive Gaussian noise. However, their performance on images with real noise is often not satisfactory. The main reason is that real noises are mostly spatially/channel-correlated and spatial/channel-variant. In contrast, the synthetic Additive White Gaussian Noise (AWGN) adopted in most previous work is pixel-independent. In this paper, we propose a novel approach to boost the performance of a real image denoiser which is trained only with synthetic pixel-independent noise data. First, we train a deep model that consists of a noise estimator and a denoiser with mixed AWGN and Random Value Impulse Noise (RVIN). We then investigate Pixel-shuffle Down-sampling (PD) strategy to adapt the trained model to real noises. Extensive experiments demonstrate the effectiveness and generalization ability of the proposed approach. Notably, our method achieves state-of-the-art performance on real sRGB images in the DND benchmark. Codes are available at this https URL."}, "keywords": ["pixel-shuffle downsampling"], "citation_intent": "background"}
{"citing_id": "2305.01656v1", "cited_id": "1510.07898", "section_title": "Our Previous Work", "citation": "Our initial analysis of AppTracker #REFR used a different inferred admixture model, a restricted set of generic temporal properties, and segmented data sets. There was no inductive coding nor menu redesign.", "text_before_citation": ["We developed our analytics over a number of years: refining the Markovian models, the temporal properties, and segmentation of data sets.", "Our first study #OTHEREFR involved the iOS multi-player game Hungry Yoshi #OTHEREFR , one data set, a simple inferred model, simple temporal logic properties, and K = 2.", "The app was different to AppTracker in that it had clear (user) goal states and (in addition to user interactions) external events when the device picked up scanned Wi-Fi access points.", "We uncovered two interaction styles representing different strategies for playing the game, but could not implement a redesign because Apple's iOS changes meant we could no longer scan for Wi-Fi access points. The app is no longer under development."], "text_after_citation": ["The GPAM model was defined in #OTHEREFR , where we also introduced the possibility of logic formulae over the latent variables, though we did not consider the probability to transition between activity patterns.", "Significantly, none of our earlier work involved specific design recommendations, the implementation and deployment of a new design, nor analysis of interaction styles in the new design."], "citing_paper_content": {"title": "Probabilistic Formal Modelling To Uncover And Interpret Interaction Styles", "abstract": "We present a study using new computational methods, based on a novel combination of machine learning for inferring admixture hidden Markov models and probabilistic model checking, to uncover interaction styles in a mobile app. These styles are then used to inform a redesign, which is implemented, deployed, and then analysed using the same methods. The data sets are logged user traces, collected over two six-month deployments of each version, involving thousands of users and segmented into different time intervals. The methods do not assume tasks or absolute metrics such as measures of engagement, but uncover the styles through unsupervised inference of clusters and analysis with probabilistic temporal logic. For both versions there was a clear distinction between the styles adopted by users during the first day/week/month of usage, and during the second and third months, a result we had not anticipated."}, "cited_paper_content": {"title": "Probabilistic Formal Analysis Of App Usage To Inform Redesign", "abstract": "Evaluation and redesign of user-intensive mobile applications is challenging because users are often heterogeneous, adopting different patterns of activity, at different times. We set out a process of integrating statistical, longitudinal analysis of actual logged behaviours, formal, probabilistic discrete state models of activity patterns, and hypotheses over those models expressed as probabilistic temporal logic properties to inform redesign. We employ formal methods not to the design of the mobile application, but to characterise the different probabilistic patterns of actual use over various time cuts within a population of users. We define the whole process from identifying questions that give us insight into application usage, to event logging, data abstraction from logs, model inference, temporal logic property formulation, visualisation of results, and interpretation in the context of redesign. We illustrate the process through a real-life case study, which results in a new and principled way for selecting content for an extension to the mobile application."}, "keywords": ["AppTracker", "generic temporal properties"], "citation_intent": "method"}
{"citing_id": "2303.09395v1", "cited_id": "1512.03385", "section_title": "Quantitative Results", "citation": "Then, we use the corresponding ECGs to train a multi-label classifier with ResNet-34 #REFR architecture and 1-D convolutions.", "text_before_citation": ["We conduct various evaluations to compare our model against other baseline models: Diffusion model, Tacotron 2 + Griffin-Lim, and Tacotron 2 + Hifi-GAN.", "The first evaluation is conducted to test how well the generated ECGs can capture representative cardiac diagnoses when they are given as textual input.", "We first collect most frequent textual prompts with each diagnosis in our dataset, which is determined with a Word N-gram of N between 4 and 10."], "text_after_citation": ["Finally, we generate 128 ECG samples for each diagnosis with the collected textual prompts.", "The classification results of different models are shown in Table 1 .", "The second evaluation is conducted to assess text-ECG alignment performance, and is based on Contrastive Language-Image Pre-training (CLIP) #OTHEREFR method.", "Instead of training an image encoder, we train an ECG encoder jointly with a text encoder to predict correct pairings of ECG and text.", "We then generate ECGs based on full text reports in the test set with each model."], "citing_paper_content": {"title": "Text-To-Ecg: 12-Lead Electrocardiogram Synthesis Conditioned On Clinical Text Reports", "abstract": "Electrocardiogram (ECG) synthesis is the area of research focused on generating realistic synthetic ECG signals for medical use without concerns over annotation costs or clinical data privacy restrictions. Traditional ECG generation models consider a single ECG lead and utilize GAN-based generative models. These models can only generate single lead samples and require separate training for each diagnosis class. The diagnosis classes of ECGs are insufficient to capture the intricate differences between ECGs depending on various features (e.g. patient demographic details, co-existing diagnosis classes, etc.). To alleviate these challenges, we present a text-to-ECG task, in which textual inputs are used to produce ECG outputs. Then we propose Auto-TTE, an autoregressive generative model conditioned on clinical text reports to synthesize 12-lead ECGs, for the first time to our knowledge. We compare the performance of our model with other representative models in text-to-speech and text-to-image. Experimental results show the superiority of our model in various quantitative evaluations and qualitative analysis. Finally, we conduct a user study with three board-certified cardiologists to confirm the fidelity and semantic alignment of generated samples. our code will be available at https:"}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["corresponding ECGs", "1-D convolutions"], "citation_intent": "method"}
{"citing_id": "2303.05342v1", "cited_id": "1904.11622", "section_title": "Experimental Settings", "citation": "This benchmark is proposed by #REFR which removes hypernyms and redundant synonyms in the 50-way benchmark, giving a cleaner relationship schema.", "text_before_citation": ["This benchmark contains the 50 most frequent relationships and 150 object classes in Visual Genome, and is widely used in many supervised VRD works #OTHEREFR .", "(2) 25-way.", "This benchmark is proposed in #OTHEREFR by removing the 25 most frequent relationships from the 50-way benchmark.", "Many existing few-shot VRD works #OTHEREFR evaluated their approach on this benchmark and use samples of removed relationships to pre-train their models.", "(3) 20-way."], "text_after_citation": ["We compare our method extensively with other baselines on all three benchmarks and conduct ablation studies on the 25-way benchmark with 5 training examples.", "We refer readers to the appendix for the full list of relationships used in each benchmark and other statistics.", "Evaluation Metrics.", "We evaluate the overall VRD performance of our method using Recall@k (R@k) and mean-Recall@k (mR@k), which measure how many labeled visual relations are recalled.", "To support multi-dimensional and more fine-grained evaluation on the methods' generalization capability, we devise novel Seen/Unseen Recall metrics."], "citing_paper_content": {"title": "Knowledge-Augmented Few-Shot Visual Relation Detection", "abstract": "Visual Relation Detection (VRD) aims to detect relationships between objects for image understanding. Most existing VRD methods rely on thousands of training samples of each relationship to achieve satisfactory performance. Some recent papers tackle this problem by few-shot learning with elaborately designed pipelines and pre-trained word vectors. However, the performance of existing few-shot VRD models is severely hampered by the poor generalization capability, as they struggle to handle the vast semantic diversity of visual relationships. Nonetheless, humans have the ability to learn new relationships with just few examples based on their knowledge. Inspired by this, we devise a knowledge-augmented, few-shot VRD framework leveraging both textual knowledge and visual relation knowledge to improve the generalization ability of few-shot VRD. The textual knowledge and visual relation knowledge are acquired from a pre-trained language model and an automatically constructed visual relation knowledge graph, respectively. We extensively validate the effectiveness of our framework. Experiments conducted on three benchmarks from the commonly used Visual Genome dataset show that our performance surpasses existing state-of-the-art models with a large improvement."}, "cited_paper_content": {"title": "Scene Graph Prediction With Limited Labels", "abstract": "Visual knowledge bases such as Visual Genome power numerous applications in computer vision, including visual question answering and captioning, but suffer from sparse, incomplete relationships. All scene graph models to date are limited to training on a small set of visual relationships that have thousands of training labels each. Hiring human annotators is expensive, and using textual knowledge base completion methods are incompatible with visual data. In this paper, we introduce a semi-supervised method that assigns probabilistic relationship labels to a large number of unlabeled images using few labeled examples. We analyze visual relationships to suggest two types of image-agnostic features that are used to generate noisy heuristics, whose outputs are aggregated using a factor graph-based generative model. With as few as 10 labeled examples per relationship, the generative model creates enough training data to train any existing state-of-the-art scene graph model. We demonstrate that our method outperforms all baseline approaches on scene graph prediction by5.16 recall@100 for PREDCLS. In our limited label setting, we define a complexity metric for relationships that serves as an indicator (R^2 = 0.778) for conditions under which our method succeeds over transfer learning, the de-facto approach for training with limited labels."}, "keywords": ["cleaner relationship schema"], "citation_intent": "method"}
{"citing_id": "2303.09397v1", "cited_id": "1908.10063", "section_title": "Sentiment Prediction", "citation": "FinBert is a language model based on BERT, built to tackle NLP problems in the financial domain #REFR .", "text_before_citation": ["The sentiment prediction in this paper is done using a fine-tuned FinBERT-based model."], "text_after_citation": ["On various NLP tasks, pre-trained generic language models #OTHEREFR - #OTHEREFR have produced excellent results.", "Such models can be used for practically any task after being unsupervised and trained on massive quantities of text.", "One of the most effective language models now in use is BERT #OTHEREFR . The transformer encoder is the foundation of this concept.", "The Transformer is a sequence-tosequence architecture that only relies on decoder and encoder attention methods.", "Since the BERT design is not a sequence-to-sequence model, it ignores the decoder network and uses simply a transformer encoder (although it can be used in such tasks)."], "citing_paper_content": {"title": "Cryptocurrency Price Prediction Using Twitter Sentiment Analysis", "abstract": "The cryptocurrency ecosystem has been the centre of discussion on many social media platforms, following its noted volatility and varied opinions. Twitter is rapidly being utilised as a news source and a medium for bitcoin discussion. Our algorithm seeks to use historical prices and sentiment of tweets to forecast the price of Bitcoin. In this study, we develop an end-to-end model that can forecast the sentiment of a set of tweets (using a Bidirectional Encoder Representations from Transformers-based Neural Network Model) and forecast the price of Bitcoin (using Gated Recurrent Unit) using the predicted sentiment and other metrics like historical cryptocurrency price data, tweet volume, a user's following, and whether or not a user is verified. The sentiment prediction gave a Mean Absolute Percentage Error of 9.45%, an average of real-time data, and test data. The mean absolute percent error for the price prediction was 3.6%."}, "cited_paper_content": {"title": "Finbert: Financial Sentiment Analysis With Pre-Trained Language Models", "abstract": "Financial sentiment analysis is a challenging task due to the specialized language and lack of labeled data in that domain. General-purpose models are not effective enough because of the specialized language used in a financial context. We hypothesize that pre-trained language models can help with this problem because they require fewer labeled examples and they can be further trained on domain-specific corpora. We introduce FinBERT, a language model based on BERT, to tackle NLP tasks in the financial domain. Our results show improvement in every measured metric on current state-of-the-art results for two financial sentiment analysis datasets. We find that even with a smaller training set and fine-tuning only a part of the model, FinBERT outperforms state-of-the-art machine learning methods."}, "keywords": ["financial domain", "NLP problems"], "citation_intent": "method"}
{"citing_id": "2304.07305v1", "cited_id": "1412.6980", "section_title": "C. Training Procedure", "citation": "Each batch consists of 32 samples and the weights of the network are optimized using the Adam #REFR optimization algorithm with a weight decay factor of 5 \u00d7 10 \u22125 .", "text_before_citation": ["Hence, 20% of the data are reserved for testing, while the remaining 80% are used for training the model.", "Again, an 80/20 split is performed on the remaining four folds to generate the training and the validation set.", "Each split is performed in a such a manner that the number of examples for each class is identical in each partition.", "When the model is trained for either one of the operating conditions listed in Tab.", "II, 50,000 samples are available and thus the number of samples in the test, the training and the validation set correspond to (N test , N train , N val ) = (10000, 32000, 8000), which corresponds to 2000, 6400 and 1600 data samples per class label."], "text_after_citation": ["The weight decay factor is a regularization parameter that is multiplied with the L2 norm of the weight matrix.", "Weight decay encourages the optimizer to keep the weights small, which can prevent overfitting and improve generalization ability #OTHEREFR .", "The initial learning rate is set to \u03bb = 0.001 but is decreased by a factor of 0.8, whenever the accuracy on the validation set has not increased by at least 0.01% over the last 5 epochs.", "To even further mitigate the risk of overfitting, an early stopping scheme is implemented.", "The model is trained for a minimum number of epochs, i.e., 65 epochs, but then training is stopped when the validation accuracy has not improved over the most recent 25 epochs. The maximum number of epochs is 120."], "citing_paper_content": {"title": "-D Residual Convolutional Neural Network Coupled With Data Augmentation And Regularization Techniques For The Icphm 2023 Data Challenge", "abstract": "In this article, we present our contribution to the International Conference on Prognostics and Health Management (ICPHM) 2023 Data Challenge on Industrial Systems' Health Monitoring using Vibration Analysis. For the task of classifying sun gear faults in a gearbox, we propose a residual Convolutive Neural Network (CNN) that operates on raw three-channel time-domain vibration signals. In conjunction with data augmentation and regularization techniques, the proposed model yields very good results in a multi-class classification scenario with realworld data despite its relatively small size, i.e., with less than 30,000 trainable parameters. Even when presented with data obtained from multiple operating conditions, the network is still capable to accurately predict the condition of the gearbox under inspection."}, "cited_paper_content": {"title": "Adam: A Method For Stochastic Optimization", "abstract": "We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm."}, "keywords": ["network", "Adam optimization algorithm"], "citation_intent": "method"}
{"citing_id": "2304.11249v1", "cited_id": "1808.00897", "section_title": "Embedded-Compute-Ready Obstacle Detection Network Ewasr", "citation": "The channel refinement module (CRM) applies the cARM #REFR module to enable a global cross-channel information flow.", "text_before_citation": ["However, the transformer applies token cross-attention #OTHEREFR , which still adds a computationally prohibitive overhead.", "We propose a more efficient feature mixer that draws on findings of Yu et al.", "#OTHEREFR that computationally-costly token cross-attentions in transformers can be replaced by alternative operations, as long as they implement cross-token information flow.", "We thus propose a Light-weight Scale-Aware Semantic Extractor (LSSE) for the eWaSR feature mixer, which is composed of two metaformer refinement modules CRM and SRM (Figure 3) .", "Both modules follow the metaformer #OTHEREFR structure and differ in the type of information flow implemented by the token mixer."], "text_after_citation": ["This is followed by a spatial refinement module (SRM), which applies sARM #OTHEREFR to enable cross-channel spatial information flow.", "To make the LSSE suitable for our target hardware, we replace the commonlyused GeLU and layer normalization blocks of metaformer by ReLU and batch normalization.", "The proposed LSSE is much more computationally efficent than SSE #OTHEREFR .", "For example, with a ResNet-18 encoder, the SSE would contain 66.4M parameters (requires 3.2GFlops), while the LSSE contains 47.9M parameters (requires 2.1GFlops).", "We also simplify the WaSR decoder following the TopFormer #OTHEREFR semantic-enrichment routines to avoid the computationally-costly FFM and ASPP modules."], "citing_paper_content": {"title": "Ewasr -An Embedded-Compute-Ready Maritime Obstacle Detection Network", "abstract": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY"}, "cited_paper_content": {"title": "Bisenet: Bilateral Segmentation Network For Real-Time Semantic Segmentation", "abstract": "Semantic segmentation requires both rich spatial information and sizeable receptive field. However, modern approaches usually compromise spatial resolution to achieve real-time inference speed, which leads to poor performance. In this paper, we address this dilemma with a novel Bilateral Segmentation Network (BiSeNet). We first design a Spatial Path with a small stride to preserve the spatial information and generate high-resolution features. Meanwhile, a Context Path with a fast downsampling strategy is employed to obtain sufficient receptive field. On top of the two paths, we introduce a new Feature Fusion Module to combine features efficiently. The proposed architecture makes a right balance between the speed and segmentation performance on Cityscapes, CamVid, and COCO-Stuff datasets. Specifically, for a 2048 \\(\\times \\) 1024 input, we achieve 68.4% Mean IOU on the Cityscapes test dataset with speed of 105 FPS on one NVIDIA Titan XP card, which is significantly faster than the existing methods with comparable performance."}, "keywords": ["global cross-channel information", "flow"], "citation_intent": "method"}
{"citing_id": "2304.10482v1", "cited_id": "1904.05866", "section_title": "B. Sgnify Objective", "citation": "For more details on these terms, please refer to the original paper of SMPLify-X #REFR .", "text_before_citation": ["For the hands, SMPL-X uses a low-dimensional PCA pose space such that", "\u03b8 h = |m h | n=1 m hn M,", "where M are principal components capturing the finger pose variations and m hn are the corresponding PCA coefficients.", "Thus, E m h (m h ) is an L2 prior on the coefficients m h .", "E J represents the joint re-projection loss, and E \u03b1 (\u03b8 b ) is a prior penalizing extreme bending only for elbows and knees."], "text_after_citation": ["E O is a bone-orientation term, which factors out the residual of the parent joint from the residual of the child joint.", "For more details about this term, please refer to the original paper of RICH #OTHEREFR .", "E P and E A are used to prevent self-interpenetration.", "When self-contact occurs, the E P term pushes vertices that are inside the mesh to the surface, and E A aligns the surface normals of the vertices in contact.", "For more details, please refer to the original paper of TUCH #OTHEREFR ."], "citing_paper_content": {"title": "Reconstructing Signing Avatars From Video Using Linguistic Priors", "abstract": "Figure 1. Given a monocular, in-the-wild video of a sign-language sign, SGNify automatically reconstructs a 3D body with accurate hand pose, facial motion, and body pose. Note that motion blur obscures the finger articulations in several video frames; this is a common problem. Our novel linguistic priors enable accurate 3D reconstruction despite such image degradation."}, "cited_paper_content": {"title": "Expressive Body Capture: 3D Hands, Face, And Body From A Single Image", "abstract": "To facilitate the analysis of human actions, interactions and emotions, we compute a 3D model of human body pose, hand pose, and facial expression from a single monocular image. To achieve this, we use thousands of 3D scans to train a new, unified, 3D model of the human body, SMPL-X, that extends SMPL with fully articulated hands and an expressive face. Learning to regress the parameters of SMPL-X directly from images is challenging without paired images and 3D ground truth. Consequently, we follow the approach of SMPLify, which estimates 2D features and then optimizes model parameters to fit the features. We improve on SMPLify in several significant ways: (1) we detect 2D features corresponding to the face, hands, and feet and fit the full SMPL-X model to these; (2) we train a new neural network pose prior using a large MoCap dataset; (3) we define a new interpenetration penalty that is both fast and accurate; (4) we automatically detect gender and the appropriate body models (male, female, or neutral); (5) our PyTorch implementation achieves a speedup of more than 8x over Chumpy. We use the new method, SMPLify-X, to fit SMPL-X to both controlled images and images in the wild. We evaluate 3D accuracy on a new curated dataset comprising 100 images with pseudo ground-truth. This is a step towards automatic expressive human capture from monocular RGB data. The models, code, and data are available for research purposes at this https URL."}, "keywords": ["SMPLify-X"], "citation_intent": "background"}
{"citing_id": "2304.00902v1", "cited_id": "1909.03276", "section_title": "Evaluation Metric", "citation": "We note that through optimized FuxiCTR implementation and sufficient hyper-parameter tuning, we obtain much better model performance than what was reported in #REFR 3 .", "text_before_citation": ["Implementation We reuse the baseline models and implement our models based on FuxiCTR #OTHEREFR , an open-source CTR prediction library 2 .", "Our evaluation follows the same experimental settings with AFN #OTHEREFR , by setting the embedding dimension to 10, batch size to 4096, and the default MLP size to [400, 400, 400].", "For DualMLP and FinalMLP, we tune the two MLPs in 1\u223c3 layers to enhance stream diversity.", "We set the learning rate to 1e \u2212 3 or 5e \u2212 4.", "We tune all the other hyper-parameters (e.g., embedding regularization and dropout rate) of all the studied models via extensive grid search (about 30 runs per model on average)."], "text_after_citation": ["Thus, we report our own experimental results instead of reusing theirs to make a fair comparison.", "To promote reproducible research, we open sourced the code and running logs of FinalMLP and all the baselines used."], "citing_paper_content": {"title": "Finalmlp: An Enhanced Two-Stream Mlp Model For Ctr Prediction", "abstract": "Click-through rate (CTR) prediction is one of the fundamental tasks for online advertising and recommendation. While multi-layer perceptron (MLP) serves as a core component in many deep CTR prediction models, it has been widely recognized that applying a vanilla MLP network alone is inefficient in learning multiplicative feature interactions. As such, many two-stream interaction models (e.g., DeepFM and DCN) have been proposed by integrating an MLP network with another dedicated network for enhanced CTR prediction. As the MLP stream learns feature interactions implicitly, existing research focuses mainly on enhancing explicit feature interactions in the complementary stream. In contrast, our empirical study shows that a well-tuned two-stream MLP model that simply combines two MLPs can even achieve surprisingly good performance, which has never been reported before by existing work. Based on this observation, we further propose feature selection and interaction aggregation layers that can be easily plugged to make an enhanced two-stream MLP model, FinalMLP. In this way, it not only enables differentiated feature inputs but also effectively fuses stream-level interactions across two streams. Our evaluation results on four open benchmark datasets as well as an online A/B test in our industrial system show that FinalMLP achieves better performance than many sophisticated two-stream CTR models. Our source code will be available at MindSpore/models and FuxiCTR/model zoo."}, "cited_paper_content": {"title": "Adaptive Factorization Network: Learning Adaptive-Order Feature Interactions", "abstract": "Various factorization-based methods have been proposed to leverage second-order, or higher-order cross features for boosting the performance of predictive models. They generally enumerate all the cross features under a predefined maximum order, and then identify useful feature interactions through model training, which suffer from two drawbacks. First, they have to make a trade-off between the expressiveness of higher-order cross features and the computational cost, resulting in suboptimal predictions. Second, enumerating all the cross features, including irrelevant ones, may introduce noisy feature combinations that degrade model performance. In this work, we propose the Adaptive Factorization Network (AFN), a new model that learns arbitrary-order cross features adaptively from data. The core of AFN is a logarithmic transformation layer to convert the power of each feature in a feature combination into the coefficient to be learned. The experimental results on four real datasets demonstrate the superior predictive performance of AFN against the start-of-the-arts."}, "keywords": ["sufficient hyper-parameter tuning"], "citation_intent": "result"}
{"citing_id": "2303.06747v1", "cited_id": "1909.06581", "section_title": "Experiments", "citation": "Following the same training strategy and hyperparameters in IRN baseline, our models were trained on the DIV2K #REFR dataset, which includes 800 HR training images.", "text_before_citation": [], "text_after_citation": ["IRN-M and IRN-A were trained with 500,000 and 250,000 iterations respectively.", "Both models were evaluated across five benchmark datasets: Set5 #OTHEREFR , Set14 #OTHEREFR , BSD100 #OTHEREFR , Ur-ban100 #OTHEREFR and the validation set of DIV2K.", "The upscaled images quality across different models were assessed via the peak noise-signal ratio (PSNR) and SSIM on the Y channel of the YCbCr color space.", "Following previous works #OTHEREFR , as it is not beneficial to add randomness in restoring HR images, we set\u1e91 as 0 during the inverse up-scaling process for both training and validation steps in all experiments."], "citing_paper_content": {"title": "Raising The Limit Of Image Rescaling Using Auxiliary Encoding", "abstract": "Normalizing flow models using invertible neural networks (INN) have been widely investigated for successful generative image super-resolution (SR) by learning the transformation between the normal distribution of latent variable z and the conditional distribution of high-resolution (HR) images gave a low-resolution (LR) input. Recently, image rescaling models like IRN utilize the bidirectional nature of INN to push the performance limit of image upscaling by optimizing the downscaling and upscaling steps jointly. While the random sampling of latent variable z is useful in generating diverse photo-realistic images, it is not desirable for image rescaling when accurate restoration of the HR image is more important. Hence, in places of random sampling of z, we propose auxiliary encoding modules to further push the limit of image rescaling performance. Two options to store the encoded latent variables in downscaled LR images, both readily supported in existing image file format, are proposed. One is saved as the alpha-channel, the other is saved as meta-data in the image header, and the corresponding modules are denoted as suffixes-A and-M respectively. Optimal network architectural changes are investigated for both options to demonstrate their effectiveness in raising the rescaling performance limit on different baseline models including IRN and DLV-IRN."}, "cited_paper_content": {"title": "Blind Super-Resolution Kernel Estimation Using An Internal-Gan", "abstract": "Super resolution (SR) methods typically assume that the low-resolution (LR) image was downscaled from the unknown high-resolution (HR) image by a fixed `ideal\u2019 downscaling kernel (e.g. Bicubic downscaling). However, this is rarely the case in real LR images, in contrast to synthetically generated SR datasets. When the assumed downscaling kernel deviates from the true one, the performance of SR methods significantly deteriorates. This gave rise to Blind-SR - namely, SR when the downscaling kernel (``SR-kernel\u2019\u2019) is unknown. It was further shown that the true SR-kernel is the one that maximizes the recurrence of patches across scales of the LR image. In this paper we show how this powerful cross-scale recurrence property can be realized using Deep Internal Learning. We introduce ``KernelGAN\u2019\u2019, an image-specific Internal-GAN, which trains solely on the LR test image at test time, and learns its internal distribution of patches. Its Generator is trained to produce a downscaled version of the LR test image, such that its Discriminator cannot distinguish between the patch distribution of the downscaled image, and the patch distribution of the original LR image. The Generator, once trained, constitutes the downscaling operation with the correct image-specific SR-kernel. KernelGAN is fully unsupervised, requires no training data other than the input image itself, and leads to state-of-the-art results in Blind-SR when plugged into existing SR algorithms."}, "keywords": ["DIV2K dataset"], "citation_intent": "method"}
{"citing_id": "2303.07564v2", "cited_id": "2004.01905", "section_title": "Experiment Setup", "citation": "We seek the real foggy dataset collected by DenseFogFlow #REFR , namely DenseFog, of which 2346 images for training and 100 images for testing.", "text_before_citation": ["We select 8400 images of Fog-KITTI2015 dataset for training and 400 images for testing.", "\u2022 Fog-GOF.", "GOF #OTHEREFR is a dataset containing four different scenes with synchronized gyro readings, such as regular scenes and adverse weather scenes.", "We choose foggy images of GOF to compose a new foggy dataset, namely Fog-GOF, of which 1000 images for training and 105 images for testing.", "\u2022 DenseFog."], "text_after_citation": ["\u2022 Real-Fog World.", "We collect degraded videos under real foggy scenes from Youtube, with 1200 and 240 images for training and testing, respectively. Comparison Methods.", "We choose three competing methods GyroFlow #OTHEREFR , DenseFogFlow #OTHEREFR and RobustFlow #OTHEREFR which are designed for adverse weather optical flow.", "Moreover, we select several state-of-the-art supervised (GMA #OTHEREFR ) and unsupervised (SMURF #OTHEREFR , UFlow #OTHEREFR and Selflow #OTHEREFR ) optical flow approaches designed for clean scenes.", "The unsupervised methods are first trained on KITTI2015 for initialization and re-trained on the target degraded dataset."], "citing_paper_content": {"title": "Unsupervised Cumulative Domain Adaptation For Foggy Scene Optical Flow", "abstract": "Optical flow has achieved great success under clean scenes, but suffers from restricted performance under foggy scenes. To bridge the clean-to-foggy domain gap, the existing methods typically adopt the domain adaptation to transfer the motion knowledge from clean to synthetic foggy domain. However, these methods unexpectedly neglect the synthetic-to-real domain gap, and thus are erroneous when applied to real-world scenes. To handle the practical optical flow under real foggy scenes, in this work, we propose a novel unsupervised cumulative domain adaptation optical flow (UCDA-Flow) framework: depth-association motion adaptation and correlation-alignment motion adaptation. Specifically, we discover that depth is a key ingredient to influence the optical flow: the deeper depth, the inferior optical flow, which motivates us to design a depth-association motion adaptation module to bridge the clean-to-foggy domain gap. Moreover, we figure out that the cost volume correlation shares similar distribution of the synthetic and real foggy images, which enlightens us to devise a correlation-alignment motion adaptation module to distill motion knowledge of the synthetic foggy domain to the real foggy domain. Note that synthetic fog is designed as the intermediate domain. Under this unified framework, the proposed cumulative adaptation progressively transfers knowledge from clean scenes to real foggy scenes. Extensive experiments have been performed to verify the superiority of the proposed method."}, "cited_paper_content": {"title": "Optical Flow In Dense Foggy Scenes Using Semi-Supervised Learning", "abstract": "In dense foggy scenes, existing optical flow methods are erroneous. This is due to the degradation caused by dense fog particles that break the optical flow basic assumptions such as brightness and gradient constancy. To address the problem, we introduce a semi-supervised deep learning technique that employs real fog images without optical flow ground-truths in the training process. Our network integrates the domain transformation and optical flow networks in one framework. Initially, given a pair of synthetic fog images, its corresponding clean images and optical flow ground-truths, in one training batch we train our network in a supervised manner. Subsequently, given a pair of real fog images and a pair of clean images that are not corresponding to each other (unpaired), in the next training batch, we train our network in an unsupervised manner. We then alternate the training of synthetic and real data iteratively. We use real data without ground-truths, since to have ground-truths in such conditions is intractable, and also to avoid the overfitting problem of synthetic data training, where the knowledge learned on synthetic data cannot be generalized to real data testing. Together with the network architecture design, we propose a new training strategy that combines supervised synthetic-data training and unsupervised real-data training. Experimental results show that our method is effective and outperforms the state-of-the-art methods in estimating optical flow in dense foggy scenes."}, "keywords": ["real foggy dataset"], "citation_intent": "method"}
{"citing_id": "2304.01351v1", "cited_id": "1712.02862", "section_title": "Experiments And Results", "citation": "While MOL-LR 3D performs recovery in 3D, MoDL #REFR could do it only in 2D due to lack of memory for multiple unrolls. MOL-LR 3D outperforms MoDL 2D as shown in Fig. 2 .", "text_before_citation": ["By contrast, the experiments in the second row of Fig.", "1(a) show that MOL-SN and MOL-LR are relatively more robust to adversarial noise, compared to MoDL, UNET, and ADMM-Net.", "Both MOL-SN and MOL-LR are associated with guarantees on robustness. The plots in Fig.", "1(b) shows drop in PSNR with respect to percentage of Adversarial/Gaussian noise (0% to 20% for = [0.0, 0.2]) and shows similar trend.", "We demonstrate the benefit of reduced memory demand in MOL over unrolled networks through experiments on 2D+time cardiac data, run on a 16GB GPU."], "text_after_citation": ["MoDL 2D's poor performance can be attributed to lack of temporal information which is well utilized by MOL-LR 3D."], "citing_paper_content": {"title": "Accelerated Parallel Mri Using Memory Efficient And Robust Monotone Operator Learning (Mol)", "abstract": "Model-based deep learning methods that combine imaging physics with learned regularization priors have been emerging as powerful tools for parallel MRI acceleration. The main focus of this paper is to determine the utility of the monotone operator learning (MOL) framework in the parallel MRI setting. The MOL algorithm alternates between a gradient descent step using a monotone convolutional neural network (CNN) and a conjugate gradient algorithm to encourage data consistency. The benefits of this approach include similar guarantees as compressive sensing algorithms including uniqueness, convergence, and stability, while being significantly more memory efficient than unrolled methods. We validate the proposed scheme by comparing it with different unrolled algorithms in the context of accelerated parallel MRI for static and dynamic settings."}, "cited_paper_content": {"title": "Modl: Model Based Deep Learning Architecture For Inverse Problems", "abstract": "We introduce a model-based image reconstruction framework with a convolution neural network (CNN)-based regularization prior. The proposed formulation provides a systematic approach for deriving deep architectures for inverse problems with the arbitrary structure. Since the forward model is explicitly accounted for, a smaller network with fewer parameters is sufficient to capture the image information compared to direct inversion approaches. Thus, reducing the demand for training data and training time. Since we rely on end-to-end training with weight sharing across iterations, the CNN weights are customized to the forward model, thus offering improved performance over approaches that rely on pre-trained denoisers. Our experiments show that the decoupling of the number of iterations from the network complexity offered by this approach provides benefits, including lower demand for training data, reduced risk of overfitting, and implementations with significantly reduced memory footprint. We propose to enforce data-consistency by using numerical optimization blocks, such as conjugate gradients algorithm within the network. This approach offers faster convergence per iteration, compared to methods that rely on proximal gradients steps to enforce data consistency. Our experiments show that the faster convergence translates to improved performance, primarily when the available GPU memory restricts the number of iterations."}, "keywords": ["MOL-LR 3D"], "citation_intent": "background"}
{"citing_id": "2305.00599v1", "cited_id": "1801.03924", "section_title": "Conditional Generation", "citation": "We substitute PTI's locality regularization with our codebook-perseverance regularization: we push randomly sampled codes to keep their syntheses unchanged via an LPIPS #REFR loss.", "text_before_citation": ["Adding real images in the codebook.", "We extend the Pivotal Tuning Inversion #OTHEREFR to project real images into our codebook in Fig. 3 .", "We start by concurrently optimizing a set of vectors in the underlying continuous space to produce the images we aim to invert.", "Then, we find the indices of the nearest-neighbor variant for each gene in the codebook.", "We train both the generator and the codebook to recreate the images, based only on these indices."], "text_after_citation": ["We find this step important to retain the perceptual quality of the genome."], "citing_paper_content": {"title": "Stylegenes: Discrete And Efficient Latent Distributions For Gans", "abstract": "We propose a discrete latent distribution for Generative Adversarial Networks (GANs). Instead of drawing latent vectors from a continuous prior, we sample from a finite set of learnable latents. However, a direct parametrization of such a distribution leads to an intractable linear increase in memory in order to ensure sufficient sample diversity. We address this key issue by taking inspiration from the encoding of information in biological organisms. Instead of learning a separate latent vector for each sample, we split the latent space into a set of genes. For each gene, we train a small bank of gene variants. Thus, by independently sampling a variant for each gene and combining them into the final latent vector, our approach can represent a vast number of unique latent samples from a compact set of learnable parameters. Interestingly, our gene-inspired latent encoding allows for new and intuitive approaches to latent-space exploration, enabling conditional sampling from our unconditionally trained model. Moreover, our approach preserves state-of-the-art photo-realism while achieving better disentanglement than the widely-used StyleMapping network."}, "cited_paper_content": {"title": "The Unreasonable Effectiveness Of Deep Features As A Perceptual Metric", "abstract": "While it is nearly effortless for humans to quickly assess the perceptual similarity between two images, the underlying processes are thought to be quite complex. Despite this, the most widely used perceptual metrics today, such as PSNR and SSIM, are simple, shallow functions, and fail to account for many nuances of human perception. Recently, the deep learning community has found that features of the VGG network trained on the ImageNet classification task has been remarkably useful as a training loss for image synthesis. But how perceptual are these so-called\"perceptual losses\"? What elements are critical for their success? To answer these questions, we introduce a new Full Reference Image Quality Assessment (FR-IQA) dataset of perceptual human judgments, orders of magnitude larger than previous datasets. We systematically evaluate deep features across different architectures and tasks and compare them with classic metrics. We find that deep features outperform all previous metrics by huge margins. More surprisingly, this result is not restricted to ImageNet-trained VGG features, but holds across different deep architectures and levels of supervision (supervised, self-supervised, or even unsupervised). Our results suggest that perceptual similarity is an emergent property shared across deep visual representations."}, "keywords": ["randomly sampled codes", "PTI's locality regularization"], "citation_intent": "method"}
{"citing_id": "2303.05118v1", "cited_id": "1905.13260", "section_title": "Experimental Results", "citation": "As for fine-grained continual learning benchmarks in Table 2 , SLCA achieves the strongest performance for supervised pre-training, while for self-supervised pre-training the performance is comparable or slightly inferior to the SL version of BiC #REFR , which additionally employs some old training samples.", "text_before_citation": ["Overall Performance: All baseline approaches in Table 1, 2 are trained with our Slow Learner for a fair comparison, except the prompt-based approaches that keep the representation layer fixed.", "For continual learning of relatively coarse-grained classification tasks, such as Split CIFAR-100 and Split ImageNet-R in Table 1 (also shown in Fig.", "2 ), the SL can substantially enhance the performance of continual learning.", "With the help of Classifier Alignment (CA) and its Logit Normalization (LN), our approach clearly outperforms L2P #OTHEREFR and DualPrompt #OTHEREFR , and almost reach the joint training upper bound (the performance gap is less than 2% for supervised pre-training and 4% for self-supervised pre-training)."], "text_after_citation": ["Please note that the presented implementation of SLCA can be seen as adding CA+LN together with SL to the simplest sequential fine-tuning (Seq FT) baseline.", "In fact, the proposed SLCA can be naturally plug-and-play with other continual learning approaches. We leave it as a further work.", "It is worth noting that different replay-based approaches (w/ SL) behave differently in CLPM.", "In general, BiC #OTHEREFR , which updated the entire model with the old training samples in a similar way to learning the new ones and added a bias correction layer to mitigate imbalance between old and new classes, receives the most substantial improvements.", "While GDumb #OTHEREFR , which simply used the old training samples to train a new model from scratch at test time, has difficulty to adapt the pre-trained model with limited training samples and thus performs the worst."], "citing_paper_content": {"title": "Slca: Slow Learner With Classifier Alignment For Continual Learning On A Pre-Trained Model", "abstract": "The goal of continual learning is to improve the performance of recognition models in learning sequentially arrived data. Although most existing works are established on the premise of learning from scratch, growing efforts have been devoted to incorporating the benefits of pre-training. However, how to adaptively exploit the pre-trained knowledge for each incremental task while maintaining its generalizability remains an open question. In this work, we present an extensive analysis for continual learning on a pre-trained model (CLPM), and attribute the key challenge to a progressive overfitting problem. Observing that selectively reducing the learning rate can almost resolve this issue in the representation layer, we propose a simple but extremely effective approach named Slow Learner with Classifier Alignment (SLCA), which further improves the classification layer by modeling the class-wise distributions and aligning the classification layers in a post-hoc fashion. Across a variety of scenarios, our proposal provides substantial improvements for CLPM (e.g., up to 49.76%, 50.05%, 44.69% and 40.16% on Split CIFAR-100, Split ImageNet-R, Split CUB-200 and Split Cars-196, respectively), and thus outperforms state-of-the-art approaches by a large margin. Based on such a strong baseline, critical factors and promising directions are analyzed in-depth to facilitate subsequent research."}, "cited_paper_content": {"title": "Large Scale Incremental Learning", "abstract": "Modern machine learning suffers from catastrophic forgetting when learning new classes incrementally. The performance dramatically degrades due to the missing data of old classes. Incremental learning methods have been proposed to retain the knowledge acquired from the old classes, by using knowledge distilling and keeping a few exemplars from the old classes. However, these methods struggle to scale up to a large number of classes. We believe this is because of the combination of two factors: (a) the data imbalance between the old and new classes, and (b) the increasing number of visually similar classes. Distinguishing between an increasing number of visually similar classes is particularly challenging, when the training data is unbalanced. We propose a simple and effective method to address this data imbalance issue. We found that the last fully connected layer has a strong bias towards the new classes, and this bias can be corrected by a linear model. With two bias parameters, our method performs remarkably well on two large datasets: ImageNet (1000 classes) and MS-Celeb-1M (10000 classes), outperforming the state-of-the-art algorithms by 11.1% and 13.2% respectively."}, "keywords": ["fine-grained continual learning"], "citation_intent": "result"}
{"citing_id": "2304.10041v1", "cited_id": "1104.1159", "section_title": "Vi. Related Work", "citation": "Researchers are especially interested in the problem of synthesizing a policy that maximizes the satisfaction probability of given high-level specifications expressed in LTL. This satisfaction problem is first introduced by #REFR .", "text_before_citation": ["Formal policy synthesis for stochastic systems has received substantial research attention."], "text_after_citation": ["Approaches to solving this satisfaction problem often fall into two categories: (a) automaton-based approach and (b) non-automaton-based approach.", "Non-automaton-based approaches #OTHEREFR formulate a constrained optimization, where LTL formulas are encoded in constraints.", "Such constrained optimization approaches leverage the empirical performance of state-of-art solvers and enjoy great successes in theoretical results #OTHEREFR as well as the application #OTHEREFR within the control community.", "Specifically, authors #OTHEREFR propose a mixed-integer linear programming (MILP) for optimal control of nonlinear systems.", "They encode LTL specifications as mixed-integer linear constraints to avoid the construction of the task automaton."], "citing_paper_content": {"title": "Topological Guided Actor-Critic Modular Learning Of Continuous Systems With Temporal Objectives", "abstract": "This work investigates the formal policy synthesis of continuous-state stochastic dynamic systems given high-level specifications in linear temporal logic. To learn an optimal policy that maximizes the satisfaction probability, we take a product between a dynamic system and the translated automaton to construct a product system on which we solve an optimal planning problem. Since this product system has a hybrid product state space that results in reward sparsity, we introduce a generalized optimal backup order, in reverse to the topological order, to guide the value backups and accelerate the learning process. We provide the optimality proof for using the generalized optimal backup order in this optimal planning problem. Further, this paper presents an actor-critic reinforcement learning algorithm when topological order applies. This algorithm leverages advanced mathematical techniques and enjoys the property of hyperparameter self-tuning. We provide proof of the optimality and convergence of our proposed reinforcement learning algorithm. We use neural networks to approximate the value function and policy function for hybrid product state space. Furthermore, we observe that assigning integer numbers to automaton states can rank the value or policy function approximated by neural networks. To break the ordinal relationship, we use an individual neural network for each automaton state's value (policy) function, termed modular learning. We conduct two experiments. First, to show the efficacy of our reinforcement learning algorithm, we compare it with baselines on a classic control task, CartPole. Second, we demonstrate the empirical performance of our formal policy synthesis framework on motion planning of a Dubins car with a temporal specification."}, "cited_paper_content": {"title": "Ltl Control In Uncertain Environments With Probabilistic Satisfaction Guarantees", "abstract": "We present a method to generate a robot control strategy that maximizes the probability to accomplish a task. The task is given as a Linear Temporal Logic (LTL) formula over a set of properties that can be satisfied at the regions of a partitioned environment. We assume that the probabilities with which the properties are satisfied at the regions are known, and the robot can determine the truth value of a proposition only at the current region. Motivated by several results on partitioned-based abstractions, we assume that the motion is performed on a graph. To account for noisy sensors and actuators, we assume that a control action enables several transitions with known probabilities. We show that this problem can be reduced to the problem of generating a control policy for a Markov Decision Process (MDP) such that the probability of satisfying an LTL formula over its states is maximized. We provide a complete solution for the latter problem that builds on existing results from probabilistic model checking. We include an illustrative case study."}, "keywords": ["given high-level specifications"], "citation_intent": "background"}
{"citing_id": "2303.02781v1", "cited_id": "1810.03993", "section_title": "Evaluation Of Deployed Models", "citation": "Alternative practical evaluation measures are desired, especially with the increasing prevalence of ML models #REFR .", "text_before_citation": ["The performance of an ML model can vary under domain shifts.", "Different users may need to deploy the model on very different data distributions, with possibly widely different accuracy.", "Yet, traditional evaluation methods report a single aggregate number on benchmarks or through leaderboards."], "text_after_citation": ["We propose to evaluate black-box classification model, not as one or few aggregate numbers, but as a surface defined on a space of input instance attributes that capture the variability of user expectations.", "Indoor/outdoor, day/night, urban/rural may be attributes of input images for visual object recognition tasks.", "Speaker age, gender, ethnicity/accent may be attributes of input audio for speech recognition tasks.", "We call a combination of attributes in their Cartesian space an arm .", "We present a Gaussian Process (GP)-based probabilistic estimator of the accuracy surface."], "citing_paper_content": {"title": "Robustness, Evaluation And Adaptation Of Machine Learning Models In The Wild", "abstract": "This is to certify that Vihari Piratla (Roll No. 173050039) was admitted to the candidacy of Ph.D. degree on July 2017, after successfully completing all the courses required for the Ph.D. programme. The details of the course work done are given below."}, "cited_paper_content": {"title": "Model Cards For Model Reporting", "abstract": "Trained machine learning models are increasingly used to perform high-impact tasks in areas such as law enforcement, medicine, education, and employment. In order to clarify the intended use cases of machine learning models and minimize their usage in contexts for which they are not well suited, we recommend that released models be accompanied by documentation detailing their performance characteristics. In this paper, we propose a framework that we call model cards, to encourage such transparent model reporting. Model cards are short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups (e.g., race, geographic location, sex, Fitzpatrick skin type [15]) and intersectional groups (e.g., age and race, or sex and Fitzpatrick skin type) that are relevant to the intended application domains. Model cards also disclose the context in which models are intended to be used, details of the performance evaluation procedures, and other relevant information. While we focus primarily on human-centered machine learning models in the application fields of computer vision and natural language processing, this framework can be used to document any trained machine learning model. To solidify the concept, we provide cards for two supervised models: One trained to detect smiling faces in images, and one trained to detect toxic comments in text. We propose model cards as a step towards the responsible democratization of machine learning and related artificial intelligence technology, increasing transparency into how well artificial intelligence technology works. We hope this work encourages those releasing trained machine learning models to accompany model releases with similar detailed evaluation numbers and other relevant documentation."}, "keywords": ["ML models"], "citation_intent": "background"}
{"citing_id": "2304.07803v1", "cited_id": "1703.06211", "section_title": "( ) ,", "citation": "Inspired by their findings, we propose a distance-based attention score for each i-th horizontal (Das h i \u2208 R 1\u00d7W \u00d7W ) and vertical (Das v i \u2208 R 1\u00d7H\u00d7H ) local window, defined by Eqs. (10) and #REFR respectively. Here, N represents L 1 normalization.", "text_before_citation": ["However, softmax is computationally inefficient #OTHEREFR and lacks relevance to the natural language processing (NLP) or vision tasks.", "For these reasons, some studies have attempted to replace softmax with alternative functions #OTHEREFR .", "These studies empirically determined that the essential roles performed by softmax are as follows.", "\u2022 Ensuring attention score to get non-negative value", "\u2022 Re-weighting the attention score"], "text_after_citation": ["Then, the local attention of each i-th local window is obtained using Eq.(12)", "EQUATION", "Das", "EQUATION", "EQUATION"], "citing_paper_content": {"title": "Egformer: Equirectangular Geometry-Biased Transformer For 360 Depth Estimation", "abstract": "Estimating the depths of equirectangular (i.e. 360 \u2022) images (EIs) is challenging given the distorted 180 \u2022 \u00d7 360 \u2022 field-of-view, which is hard to be addressed via convolutional neural network (CNN). Although a transformer with global attention achieves significant improvements over CNN for EI depth estimation task, it is computationally inefficient, which raises the need for transformer with local attention. However, to apply local attention successfully for EIs, a specific strategy, which addresses distorted equirectangular geometry and limited receptive field simultaneously, is required. Prior works have only cared either of them, resulting in unsatisfactory depths occasionally. In this paper, we propose an equirectangular geometry-biased transformer termed EGformer, which enables local attention extraction in a global manner considering the equirectangular geometry. To achieve this, we actively utilize the equirectangular geometry as the bias for the local attention instead of struggling to reduce the distortion of EIs. As compared to the most recent transformer based EI depth estimation studies, the proposed approach yields the best depth outcomes overall with the lowest computational cost and the fewest parameters, demonstrating the effectiveness of the proposed methods."}, "cited_paper_content": {"title": "Deformable Convolutional Networks", "abstract": "Convolutional neural networks (CNNs) are inherently limited to model geometric transformations due to the fixed geometric structures in their building modules. In this work, we introduce two new modules to enhance the transformation modeling capability of CNNs, namely, deformable convolution and deformable RoI pooling. Both are based on the idea of augmenting the spatial sampling locations in the modules with additional offsets and learning the offsets from the target tasks, without additional supervision. The new modules can readily replace their plain counterparts in existing CNNs and can be easily trained end-to-end by standard back-propagation, giving rise to deformable convolutional networks. Extensive experiments validate the performance of our approach. For the first time, we show that learning dense spatial transformation in deep CNNs is effective for sophisticated vision tasks such as object detection and semantic segmentation. The code is released at https://github.com/msracver/Deformable-ConvNets."}, "keywords": ["distance-based attention score"], "citation_intent": "method"}
{"citing_id": "2303.01978v1", "cited_id": "cs/9901013", "section_title": "Ocsdf For Implicit Shape Parametrization", "citation": "Efficient data structures (e.g., using K-dtrees #REFR or Octrees (Meagher, 1980)) can mitigate this effect but do not scale well to high dimensions.", "text_before_citation": ["Our approach to learning the SDF contrasts with the computer graphics literature, where SDF is used to obtain the distance of a point to a surface (here defined as \u2202X ).", "In-deed, SDFs are usually learned in a supervised fashion, requiring the ground truth of l2 distance.", "This is classically achieved using Nearest-Neighbor algorithms, which can be cumbersome, especially when the number of points is high."], "text_after_citation": ["Instead, OCSDF learns the SDF solely based on points contained in the support.", "To illustrate this, we use models from Princeton's Model-Net10 dataset #OTHEREFR .", "We sample n = 2048 within each mesh to obtain a 3d point cloud.", "We fit the SDF on the point cloud with the same hyperparameters as the tabular experiment.", "We use Lewiner marching algorithm #OTHEREFR as the level set of the iso-surface we plot."], "citing_paper_content": {"title": "Robust One-Class Classification With Signed Distance Function Using 1-Lipschitz Neural Networks", "abstract": "We propose a new method, dubbed One Class Signed Distance Function (OCSDF), to perform One Class Classification (OCC) by provably learning the Signed Distance Function (SDF) to the boundary of the support of any distribution. The distance to the support can be interpreted as a normality score, and its approximation using 1-Lipschitz neural networks provides robustness bounds against l2 adversarial attacks, an underexplored weakness of deep learning-based OCC algorithms. As a result, OCSDF comes with a new metric, certified AUROC, that can be computed at the same cost as any classical AUROC. We show that OCSDF is competitive against concurrent methods on tabular and image data while being way more robust to adversarial attacks, illustrating its theoretical properties. Finally, as exploratory research perspectives, we theoretically and empirically show how OCSDF connects OCC with image generation and implicit neural surface parametrization. Our code is available at https://anonymous.4open. science/r/CSDFL"}, "cited_paper_content": {"title": "Analysis Of Approximate Nearest Neighbor Searching With Clustered Point Sets", "abstract": "We present an empirical analysis of data structures for approximate nearest neighbor searching. We compare the well-known optimized kd-tree splitting method against two alternative splitting methods. The first, called the sliding-midpoint method, which attempts to balance the goals of producing subdivision cells of bounded aspect ratio, while not producing any empty cells. The second, called the minimum-ambiguity method is a query-based approach. In addition to the data points, it is also given a training set of query points for preprocessing. It employs a simple greedy algorithm to select the splitting plane that minimizes the average amount of ambiguity in the choice of the nearest neighbor for the training points. We provide an empirical analysis comparing these two methods against the optimized kd-tree construction for a number of synthetically generated data and query sets. We demonstrate that for clustered data and query sets, these algorithms can provide significant improvements over the standard kd-tree construction for approximate nearest neighbor searching."}, "keywords": ["K-dtrees"], "citation_intent": "background"}
{"citing_id": "2303.02014v1", "cited_id": "1807.07878", "section_title": "Information-Theoretic Approaches", "citation": "On the privacy front, prior work has shown that I(U ; Y ) can be reduced while allowing the attacker to guess S correctly from Y with higher probability (see Example 1 in #REFR ).", "text_before_citation": ["To find a good privacy-preserving data processing strategy P Y |X , the privacy funnel solves the optimization", "min P Y |X :I(X;Y )\u2265R I(U ; Y ) ,", "where R is a desired threshold on the utility of Y .", "To apply it in data sharing problems, we could regard X as the original data, Y as the released data, and U as the secret data holder wants to protect (e.g., the fraction of a specific server type).", "However, mutual information is not a good metric for either privacy or utility."], "text_after_citation": ["On the utility front, higher mutual information I(X; Y ) does not mean that the released data Y is a useful representation of X.", "For example, Y could be an arbitrary one-to-one transformation of X.", "In that case, I(X; Y ) is maximized, but the data structure could be completely destroyed.", "In addition, privacy funnel #OTHEREFR only considers X and Y in discrete supports, which is too restrictive for our setting."], "citing_paper_content": {"title": "Summary Statistic Privacy In Data Sharing", "abstract": "Data sharing between different parties has become increasingly common across industry and academia. An important class of privacy concerns that arises in data sharing scenarios regards the underlying distribution of data. For example, the total traffic volume of data from a networking company can reveal the scale of its business, which may be considered a trade secret. Unfortunately, existing privacy frameworks (e.g., differential privacy, anonymization) do not adequately address such concerns. In this paper, we propose summary statistic privacy, a framework for analyzing and protecting these summary statistic privacy concerns. We propose a class of quantization mechanisms that can be tailored to various data distributions and statistical secrets, and analyze their privacy-distortion tradeoffs under our framework. We prove corresponding lower bounds on the privacy-utility tradeoff, which match the tradeoffs of the quantization mechanism under certain regimes, up to small constant factors. Finally, we demonstrate that the proposed quantization mechanisms achieve better privacy-distortion tradeoffs than alternative privacy mechanisms on real-world datasets."}, "cited_paper_content": {"title": "An Operational Approach To Information Leakage", "abstract": "Given two random variables $X$ and $Y$ , an operational approach is undertaken to quantify the \u201cleakage\u201d of information from $X$ to $Y$ . The resulting measure $\\mathcal {L}\\left ({X \\!\\! \\to \\!\\! Y}\\right)$ is called maximal leakage , and is defined as the multiplicative increase, upon observing $Y$ , of the probability of correctly guessing a randomized function of $X$ , maximized over all such randomized functions. A closed-form expression for $\\mathcal {L}\\left ({X \\!\\! \\to \\!\\! Y}\\right)$ is given for discrete $X$ and $Y$ , and it is subsequently generalized to handle a large class of random variables. The resulting properties are shown to be consistent with an axiomatic view of a leakage measure, and the definition is shown to be robust to variations in the setup. Moreover, a variant of the Shannon cipher system is studied, in which performance of an encryption scheme is measured using maximal leakage. A single-letter characterization of the optimal limit of (normalized) maximal leakage is derived and asymptotically-optimal encryption schemes are demonstrated. Furthermore, the sample complexity of estimating maximal leakage from data is characterized up to subpolynomial factors. Finally, the guessing framework used to define maximal leakage is used to give operational interpretations of commonly used leakage measures, such as Shannon capacity, maximal correlation, and local differential privacy."}, "keywords": ["privacy front"], "citation_intent": "background"}
{"citing_id": "2304.14298v1", "cited_id": "1811.11127", "section_title": "Ablation Studies", "citation": "The reason may be that the steps in the image processing pipeline can make the noise of sRGB more complex than RAW #REFR , which leads to accuracy degradation.", "text_before_citation": ["The image processing pipeline includes digital gain, white balance, demosaicing, color correction, and gamma correction.", "It is worth noting that quantizing the captured 14-bit RAW images to RAW images of various color encodings can be different from directly capturing them.", "However, since a specific commercial camera model only support RAW capture with constant bit depth (typically 14 bits in high-end DSLR), here, we simulate their results with quantization for fast verification.", "As shown in Table 4 , we can see the performance of both sRGB-dark and RAW-dark gradually increase from 8bit to 14-bit, and the results of sRGB-dark are very similar to RAW-dark.", "Besides, we also notice that the results of RAWdark are slightly better than the sRGB-dark."], "text_after_citation": ["These quantitative results show that the high-bit property of RAW images plays a crucial part in low-light instance segmentation.", "Low-light Synthetic RAW Pipeline.", "The unprocessing operation inverts sRGB images to synthetical RAW images, and noise injection simulates the corruption caused by limited photon count and imperfection of photodetectors.", "As shown in Table 2 , they bring in 2.2 AP and 3.5 AP performance improvements, respectively.", "Moreover, when we combine these two steps together, the accuracy increases from 31.6 AP to 38.0 AP, which is very close to the result of training with real RAW-dark images 39.0 AP."], "citing_paper_content": {"title": "Instance Segmentation In The Dark", "abstract": "Existing instance segmentation techniques are primarily tailored for high-visibility inputs, but their performance significantly deteriorates in extremely low-light environments. In this work, we take a deep look at instance segmentation in the dark and introduce several techniques that substantially boost the low-light inference accuracy. The proposed method is motivated by the observation that noise in low-light images introduces high-frequency disturbances to the feature maps of neural networks, thereby significantly degrading performance. To suppress this \"feature noise\", we propose a novel learning method that relies on an adaptive weighted downsampling layer, a smooth-oriented convolutional block, and disturbance suppression learning. These components effectively reduce feature noise during downsampling and convolution operations, enabling the model to learn disturbance-invariant features. Furthermore, we discover that high-bit-depth RAW images can better preserve richer scene information in low-light conditions compared to typical camera sRGB outputs, thus supporting the use of RAW-input algorithms. Our analysis indicates that high bit"}, "cited_paper_content": {"title": "Unprocessing Images For Learned Raw Denoising", "abstract": "Machine learning techniques work best when the data used for training resembles the data used for evaluation. This holds true for learned single-image denoising algorithms, which are applied to real raw camera sensor readings but, due to practical constraints, are often trained on synthetic image data. Though it is understood that generalizing from synthetic to real images requires careful consideration of the noise properties of camera sensors, the other aspects of an image processing pipeline (such as gain, color correction, and tone mapping) are often overlooked, despite their significant effect on how raw measurements are transformed into finished images. To address this, we present a technique to \u201cunprocess\u201d images by inverting each step of an image processing pipeline, thereby allowing us to synthesize realistic raw sensor measurements from commonly available Internet photos. We additionally model the relevant components of an image processing pipeline when evaluating our loss function, which allows training to be aware of all relevant photometric processing that will occur after denoising. By unprocessing and processing training data and model outputs in this way, we are able to train a simple convolutional neural network that has 14%-38% lower error rates and is 9\u00d7-18\u00d7 faster than the previous state of the art on the Darmstadt Noise Dataset, and generalizes to sensors outside of that dataset as well."}, "keywords": ["sRGB", "image processing pipeline"], "citation_intent": "background"}
{"citing_id": "2304.12571v1", "cited_id": "1508.00271", "section_title": "Related Work", "citation": "In this paper, we adopt the attention strategy to model the motion as a Gaussian distribution similar to #REFR .", "text_before_citation": ["Scheduled sampling #OTHEREFR related techniques are also adopted #OTHEREFR , #OTHEREFR for the same purpose, while applying such techniques will significantly slow down the training. Henter et al.", "#OTHEREFR adopt normalizing flow in LSTM to get powerful and implicit conditional motion distributions, which greatly improves the diversity of synthesized motions.", "Introducing extra temporally-related variables called \"phase\" provides another option to cope with the over-smoothness problem.", "The phase variables can be classified into three categories: the contact-based global phase #OTHEREFR for cyclic motions like running or walking, the local phase #OTHEREFR for acyclic motions like basketball motions, and the phase measured in the frequency domain that is automatically fitted in DeepPhase #OTHEREFR by neural networks.", "However, even with the DeepPhase method, different dimensions of phase manifolds still have to be fitted for different motions."], "text_after_citation": ["The human skeleton is inherently tree-structured, and different body parts may influence human motion differently.", "Training in separate body parts and combining them back is similar to matching curves with piece-wise functions, making the fitting easier.", "Many methods #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR have applied this idea to different motion synthesis tasks, such as motion prediction and style transfer. Wang et al.", "#OTHEREFR decompose the human body into seven parts, which are encoded into a latent space through four fully-connected layers.", "Then they decode the motions of the seven body parts from the latent space and concatenate them to form motions for the entire human body."], "citing_paper_content": {"title": "A Two-Part Transformer Network For Controllable Motion Synthesis", "abstract": "Although part-based motion synthesis networks have been investigated to reduce the complexity of modeling heterogeneous human motions, their computational cost remains prohibitive in interactive applications. To this end, we propose a novel two-part transformer network that aims to achieve high-quality, controllable motion synthesis results in real-time. Our network separates the skeleton into the upper and lower body parts, reducing the expensive cross-part fusion operations, and models the motions of each part separately through two streams of auto-regressive modules formed by multi-head attention layers. However, such a design might not sufficiently capture the correlations between the parts. We thus intentionally let the two parts share the features of the root joint and design a consistency loss to penalize the difference in the estimated root features and motions by these two auto-regressive modules, significantly improving the quality of synthesized motions. After training on our motion dataset, our network can synthesize a wide range of heterogeneous motions, like cartwheels and twists. Experimental and user study results demonstrate that our network is superior to state-of-the-art human motion synthesis networks in the quality of generated motions."}, "cited_paper_content": {"title": "Recurrent Network Models For Human Dynamics", "abstract": "We propose the Encoder-Recurrent-Decoder (ERD) model for recognition and prediction of human body pose in videos and motion capture. The ERD model is a recurrent neural network that incorporates nonlinear encoder and decoder networks before and after recurrent layers. We test instantiations of ERD architectures in the tasks of motion capture (mocap) generation, body pose labeling and body pose forecasting in videos. Our model handles mocap training data across multiple subjects and activity domains, and synthesizes novel motions while avoiding drifting for long periods of time. For human pose labeling, ERD outperforms a per frame body part detector by resolving left-right body part confusions. For video pose forecasting, ERD predicts body joint displacements across a temporal horizon of 400ms and outperforms a first order motion model based on optical flow. ERDs extend previous Long Short Term Memory (LSTM) models in the literature to jointly learn representations and their dynamics. Our experiments show such representation learning is crucial for both labeling and prediction in space-time. We find this is a distinguishing feature between the spatio-temporal visual domain in comparison to 1D text, speech or handwriting, where straightforward hard coded representations have shown excellent results when directly combined with recurrent units [31]."}, "keywords": ["motion"], "citation_intent": "method"}
{"citing_id": "2304.09172v1", "cited_id": "1508.07909", "section_title": "Training Details", "citation": "We use the same byte-pair encoding tokenizer #REFR as CLIP, and truncate input text at maximum 77 tokens.", "text_before_citation": ["Refer Appendix B for details about our CLIP baseline.", "Our implementation is based on PyTorch #OTHEREFR and timm #OTHEREFR libraries.", "Models.", "We use the Vision Transformer #OTHEREFR as image encoder, considering three models of varying capacity -ViT-S #OTHEREFR , ViT-B, and ViT-L. All use a patch size of 16.", "The text encoder is same as CLIP -a 12-layer, 512 dimensions wide Transformer #OTHEREFR language model."], "text_after_citation": ["Data augmentation.", "We randomly crop 50-100% area of images and resize them to 224 \u00d7 224, following #OTHEREFR .", "For text augmentation, we randomly prefix the subreddit names to captions as '{subreddit} : {caption}'.", "Initialization.", "We initialize image/text encoders in the same style as CLIP, except for one change: we use a sine-cosine position embedding in ViT, like #OTHEREFR He et al., 2022) , and keep it frozen while training."], "citing_paper_content": {"title": "Hyperbolic Image-Text Representations", "abstract": "Visual and linguistic concepts naturally organize themselves in a hierarchy, where a textual concept \"dog\" entails all images that contain dogs. Despite being intuitive, current large-scale vision and language models such as CLIP (Radford et al., 2021) do not explicitly capture such hierarchy. We propose MERU, a contrastive model that yields hyperbolic representations of images and text. Hyperbolic spaces have suitable geometric properties to embed tree-like data, so MERU can better capture the underlying hierarchy in image-text datasets. Our results show that MERU learns a highly interpretable and structured representation space while being competitive with CLIP's performance on standard multi-modal tasks like image classification and image-text retrieval."}, "cited_paper_content": {"title": "Neural Machine Translation Of Rare Words With Subword Units", "abstract": "Neural machine translation (NMT) models typically operate with a fixed vocabulary, but translation is an open-vocabulary problem. Previous work addresses the translation of out-of-vocabulary words by backing off to a dictionary. In this paper, we introduce a simpler and more effective approach, making the NMT model capable of open-vocabulary translation by encoding rare and unknown words as sequences of subword units. This is based on the intuition that various word classes are translatable via smaller units than words, for instance names (via character copying or transliteration), compounds (via compositional translation), and cognates and loanwords (via phonological and morphological transformations). We discuss the suitability of different word segmentation techniques, including simple character ngram models and a segmentation based on the byte pair encoding compression algorithm, and empirically show that subword models improve over a back-off dictionary baseline for the WMT 15 translation tasks English!German and English!Russian by up to 1.1 and 1.3 BLEU, respectively."}, "keywords": ["byte-pair encoding tokenizer"], "citation_intent": "method"}
{"citing_id": "2303.01471v1", "cited_id": "1603.04245", "section_title": "A.3 Time Dilation", "citation": "This means the continuous-time QHD can converge at any speed along the same evolution path. This result generalizes #REFR Theorem 2.2] .", "text_before_citation": ["In this section, we show that QHD is closed under time dilation, i.e., time-dilated QHD evolution is also described by the QHD equation, but with different time-dependent parameters."], "text_after_citation": ["The quantum evolution \u03a8(x, t) for t \u2208 [0, T ] forms a curve in the Hilbert space", "L 2 (R d ).", "We introduce a smooth increasing function \u03c4 : [0, T ] \u2192 R to represent the reparametrization of time. The reparametrized wave function is", "\u03a8(x, t) := \u03a8(x, \u03c4 (t)), (A.26)", "which is the same curve in the Hilbert space L 2 (R d ) but with a different speed."], "citing_paper_content": {"title": "Quantum Hamiltonian Descent *", "abstract": "Gradient descent is a fundamental algorithm in both theory and practice for continuous optimization. Identifying its quantum counterpart would be appealing to both theoretical and practical quantum applications. A conventional approach to quantum speedups in optimization relies on the quantum acceleration of intermediate steps of classical algorithms, while keeping the overall algorithmic trajectory and solution quality unchanged. We propose Quantum Hamiltonian Descent (QHD), which is derived from the path integral of dynamical systems referring to the continuous-time limit of classical gradient descent algorithms, as a truly quantum counterpart of classical gradient methods where the contribution from classically-prohibited trajectories can significantly boost QHD's performance for non-convex optimization. Moreover, QHD is described as a Hamiltonian evolution efficiently simulatable on both digital and analog quantum computers. By embedding the dynamics of QHD into the evolution of the socalled Quantum Ising Machine (including D-Wave and others), we empirically observe that the D-Wave-implemented QHD outperforms a selection of state-of-the-art gradient-based classical solvers and the standard quantum adiabatic algorithm, based on the time-to-solution metric, on non-convex constrained quadratic programming instances up to 75 dimensions. Finally, we propose a \"three-phase picture\" to explain the behavior of QHD, especially its difference from the quantum adiabatic algorithm."}, "cited_paper_content": {"title": "A Variational Perspective On Accelerated Methods In Optimization", "abstract": "Accelerated gradient methods play a central role in optimization, achieving optimal rates in many settings. Although many generalizations and extensions of Nesterov's original acceleration method have been proposed, it is not yet clear what is the natural scope of the acceleration concept. In this paper, we study accelerated methods from a continuous-time perspective. We show that there is a Lagrangian functional that we call the Bregman Lagrangian, which generates a large class of accelerated methods in continuous time, including (but not limited to) accelerated gradient descent, its non-Euclidean extension, and accelerated higher-order gradient methods. We show that the continuous-time limit of all of these methods corresponds to traveling the same curve in spacetime at different speeds. From this perspective, Nesterov's technique and many of its generalizations can be viewed as a systematic way to go from the continuous-time curves generated by the Bregman Lagrangian to a family of discrete-time accelerated algorithms."}, "keywords": ["continuous-time QHD"], "citation_intent": "background"}
{"citing_id": "2303.04942v1", "cited_id": "1808.01400", "section_title": "I. Introduction", "citation": "We then train a popular code intelligence model, Code2Seq #REFR , using the role-augmented dataset.", "text_before_citation": ["#OTHEREFR , who classified variables in programs into role categories based on how they are used (e.g., fixed value, temporary variable, stepper, etc.).", "They applied the role concepts in programming learning tasks and found that roles facilitated the abilities of students to mentally process program information and apply programming steps in writing programs.", "Inspired by this work, we are interested in how data enrichment with variable role information can impact neural models of code.", "In particular, we seek to answer the following question: can explicitly injecting the information about variable roles to the input data help neural models of code achieve better performance and robustness with lesser effort spent on training?", "To pursue this investigation, we (1) build a static analyzer to identify two kinds of variable roles, and (2) augment the dataset with explicit roles information."], "text_after_citation": ["Next, we compare our trained model against a pretrained version of Code2Seq, released by the creators of the model, who had pretrained it with the original (unaugmented) dataset.", "In addition, we evaluate the models' robustness by testing them on noise-induced data (data in which variables in a program are randomly transformed).", "For both the role augmentation and noise induction steps, we use the semantic-preserving transformation technique #OTHEREFR . Contributions. This paper provides the following contributions:", "\u2022 We present a novel study to investigate the impact of augmenting variable roles information in a large code dataset on the performance and robustness of a state-of-the-art neural program model.", "\u2022 We present and implement a technique to automatically detect certain roles of variables in programs, and consequently embed the roles information into variables in the input programs."], "citing_paper_content": {"title": "A Study Of Variable-Role-Based Feature Enrichment In Neural Models Of Code", "abstract": "Although deep neural models substantially reduce the overhead of feature engineering, the features readily available in the inputs might significantly impact training cost and the performance of the models. In this paper, we explore the impact of an unsuperivsed feature enrichment approach based on variable roles on the performance of neural models of code. The notion of variable roles (as introduced in the works of Sajaniemi et al. [1, 2]) has been found to help students' abilities in programming. In this paper, we investigate if this notion would improve the performance of neural models of code. To the best of our knowledge, this is the first work to investigate how Sajaniemi et al.'s concept of variable roles can affect neural models of code. In particular, we enrich a source code dataset by adding the role of individual variables in the dataset programs, and thereby conduct a study on the impact of variable role enrichment in training the Code2Seq model. In addition, we shed light on some challenges and opportunities in feature enrichment for neural code intelligence models. Index Terms-feature enrichment, variable roles, neural models of code"}, "cited_paper_content": {"title": "Code2Seq: Generating Sequences From Structured Representations Of Code", "abstract": "The ability to generate natural language sequences from source code snippets has a variety of applications such as code summarization, documentation, and retrieval. Sequence-to-sequence (seq2seq) models, adopted from neural machine translation (NMT), have achieved state-of-the-art performance on these tasks by treating source code as a sequence of tokens. We present ${\\rm {\\scriptsize CODE2SEQ}}$: an alternative approach that leverages the syntactic structure of programming languages to better encode source code. Our model represents a code snippet as the set of compositional paths in its abstract syntax tree (AST) and uses attention to select the relevant paths while decoding. We demonstrate the effectiveness of our approach for two tasks, two programming languages, and four datasets of up to $16$M examples. Our model significantly outperforms previous models that were specifically designed for programming languages, as well as state-of-the-art NMT models."}, "keywords": ["popular code intelligence"], "citation_intent": "method"}
{"citing_id": "2303.09527v1", "cited_id": "1905.12101", "section_title": "User-Side Fairness", "citation": "This bias can be amplified when differential privacy is incorporated in the model as demonstrated in #REFR .", "text_before_citation": ["Given a recommendation evaluation metric M, the use group fairness with respect to groups and is defined as", "E [M (R )| \u2208 ] = E [M (R )| \u2208 ]", ". Empirically, the user group fairness is measured by", "F (R ; , ) = 1 | | \u2211\ufe01 \u2208 M (R ) \u2212 1 | | \u2211\ufe01 \u2208 M (R ) .", "Due to different user activity levels, recommender systems would usually underperform against users who have less historical interactions."], "text_after_citation": ["Following the classical 80/20 rule, we select the top 20% users based on their engagement activities as the frequent/active group and the rest as the infrequent/inactive group ."], "citing_paper_content": {"title": "Fairness-Aware Differentially Private Collaborative Filtering", "abstract": "Recently, there has been an increasing adoption of differential privacy guided algorithms for privacy-preserving machine learning tasks. However, the use of such algorithms comes with trade-offs in terms of algorithmic fairness, which has been widely acknowledged. Specifically, we have empirically observed that the classical collaborative filtering method, trained by differentially private stochastic gradient descent (DP-SGD), results in a disparate impact on user groups with respect to different user engagement levels. This, in turn, causes the original unfair model to become even more biased against inactive users. To address the above issues, we propose DP-Fair, a two-stage framework for collaborative filtering based algorithms. Specifically, it combines differential privacy mechanisms with fairness constraints to protect user privacy while ensuring fair recommendations. The experimental results, based on Amazon datasets, and user history logs collected from Etsy, one of the largest e-commerce platforms, demonstrate that our proposed method exhibits superior performance in terms of both overall accuracy and user group fairness on both shallow and deep recommendation models compared to vanilla DP-SGD."}, "cited_paper_content": {"title": "Differential Privacy Has Disparate Impact On Model Accuracy", "abstract": "Differential privacy (DP) is a popular mechanism for training machine learning models with bounded leakage about the presence of specific points in the training data. The cost of differential privacy is a reduction in the model's accuracy. We demonstrate that this cost is not borne equally: accuracy of DP models drops much more for the underrepresented classes and subgroups. ::: For example, a DP gender classification model exhibits much lower accuracy for black faces than for white faces. Critically, this gap is bigger in the DP model than in the non-DP model, i.e., if the original model is unfair, the unfairness becomes worse once DP is applied. We demonstrate this effect for a variety of tasks and models, including sentiment analysis of text and image classification. We then explain why DP training mechanisms such as gradient clipping and noise addition have disproportionate effect on the underrepresented and more complex subgroups, resulting in a disparate reduction of model accuracy."}, "keywords": ["differential privacy"], "citation_intent": "background"}
{"citing_id": "2304.11781v1", "cited_id": "0807.3917", "section_title": "I. Introduction", "citation": "The formalism introduced by Ar\u0131kan #REFR came with a large panel of theoretical and practical challenges.", "text_before_citation": ["The discovery of polar coding enabled 5G and other coding tasks."], "text_after_citation": ["A fundamental one is about the reliability of the synthetic channels W \u03b1 , where W is the underlying physical channel and \u03b1 is a binary string, and how to order the set {W \u03b1 : \u03b1 \u2208 {0, 1} m }.", "Further, we shall write \u03b1 \u03b3 when W \u03b1 is favored over W \u03b3 for all W .", "From general channel models (W being a BMS channel) to particular ones (W being a BEC), several rules were proposed.", "Rule Set A (RS-A) was initially used by Mori and Tanaka #OTHEREFR to construct polar codes over BMS channels and is generated by 1 0. Later, Schurch #OTHEREFR and Bardet et al.", "#OTHEREFR added 10 01 over BMS channels, which we call Rule Set B (RS-B)."], "citing_paper_content": {"title": "Fast Methods For Ranking Synthetic Becs", "abstract": "We gather existing methods that are used to compare and rank the BECs synthesized by a polar code constructor, compare them, and propose new methods that compare synthetic BECs faster."}, "cited_paper_content": {"title": "Channel Polarization: A Method For Constructing Capacity-Achieving Codes For Symmetric Binary-Input Memoryless Channels", "abstract": "A method is proposed, called channel polarization, to construct code sequences that achieve the symmetric capacity $I(W)$ of any given binary-input discrete memoryless channel (B-DMC) $W$. The symmetric capacity is the highest rate achievable subject to using the input letters of the channel with equal probability. Channel polarization refers to the fact that it is possible to synthesize, out of $N$ independent copies of a given B-DMC $W$, a second set of $N$ binary-input channels $\\{W_N^{(i)}:1\\le i\\le N\\}$ such that, as $N$ becomes large, the fraction of indices $i$ for which $I(W_N^{(i)})$ is near 1 approaches $I(W)$ and the fraction for which $I(W_N^{(i)})$ is near 0 approaches $1-I(W)$. The polarized channels $\\{W_N^{(i)}\\}$ are well-conditioned for channel coding: one need only send data at rate 1 through those with capacity near 1 and at rate 0 through the remaining. Codes constructed on the basis of this idea are called polar codes. The paper proves that, given any B-DMC $W$ with $I(W)>0$ and any target rate $R<I(W)$, there exists a sequence of polar codes $\\{{\\mathscr C}_n;n\\ge 1\\}$ such that ${\\mathscr C}_n$ has block-length $N=2^n$, rate $\\ge R$, and probability of block error under successive cancellation decoding bounded as $P_{e}(N,R) \\le \\bigoh(N^{-\\frac14})$ independently of the code rate. This performance is achievable by encoders and decoders with complexity $O(N\\log N)$ for each."}, "keywords": ["large panel", "practical challenges"], "citation_intent": "background"}
{"citing_id": "2303.15916v2", "cited_id": "1802.06739", "section_title": "Experiment 5: T-Sne Visualization Of Generated Data", "citation": "In the case of the CharacterTrajectories dataset, it is visible that for the DPWGAN #REFR the class distribution shows a clear separation between the classes for both the private and public data (Figure 1a ).", "text_before_citation": ["In Figure 1 the results of four datasets are shown.", "The first column of each subfigure shows the private train data and the public train data.", "Ideally, the data overlaps, as this states that the private and generated (public) dataset are indistinguishable.", "The CharacterTrajectories dataset shows this behavior for DPWGAN #OTHEREFR whereas, for the other three datasets, this is not the case.", "The second and third columns of the subfigure show the class distribution within the private and public train datasets."], "text_after_citation": ["Although the data shows a small offset, this did not hinder the performance by a large margin.", "Both models show a similar position of the 20 classes within the space.", "Furthermore, the results provide evidence that the dataset generated using GSWGAN #OTHEREFR shows a much better overlap compared to the private data.", "In addition, the class separation is much better compared to the DPWGAN #OTHEREFR .", "The results provide evidence that the generated data for the CharacterTrajectories and FaceDetection dataset using the GSWGAN #OTHEREFR share the same distribution as the original data (Figure 1b and Figure 1f )."], "citing_paper_content": {"title": "From Private To Public: Benchmarking Gans In The Context Of Private Time Series Classification", "abstract": "Deep learning has proven to be successful in various domains and for different tasks. However, when it comes to private data several restrictions are making it difficult to use deep learning approaches in these application fields. Recent approaches try to generate data privately instead of applying a privacy-preserving mechanism directly, on top of the classifier. The solution is to create public data from private data in a manner that preserves the privacy of the data. In this work, two very prominent GAN-based architectures were evaluated in the context of private time series classification. In contrast to previous work, mostly limited to the image domain, the scope of this benchmark was the time series domain. The experiments show that especially GSWGAN performs well across a variety of public datasets outperforming the competitor DPWGAN. An analysis of the generated datasets further validates the superiority of GSWGAN in the context of time series generation."}, "cited_paper_content": {"title": "Differentially Private Generative Adversarial Network", "abstract": "Generative Adversarial Network (GAN) and its variants have recently attracted intensive research interests due to their elegant theoretical foundation and excellent empirical performance as generative models. These tools provide a promising direction in the studies where data availability is limited. One common issue in GANs is that the density of the learned generative distribution could concentrate on the training data points, meaning that they can easily remember training samples due to the high model complexity of deep networks. This becomes a major concern when GANs are applied to private or sensitive data such as patient medical records, and the concentration of distribution may divulge critical patient information. To address this issue, in this paper we propose a differentially private GAN (DPGAN) model, in which we achieve differential privacy in GANs by adding carefully designed noise to gradients during the learning procedure. We provide rigorous proof for the privacy guarantee, as well as comprehensive empirical evidence to support our analysis, where we demonstrate that our method can generate high quality data points at a reasonable privacy level."}, "keywords": ["public data"], "citation_intent": "background"}
{"citing_id": "2303.17713v1", "cited_id": "1702.08423", "section_title": "E Experiment Details E.1 Dataset Details", "citation": "UTKFace UTKFace dataset #REFR consists of faces images in the wild with annotations of age, gender, and ethnicity. Features are extracted using CLIP embeddings from images.", "text_before_citation": ["The dataset has annotations about the target communities -race, religion, gender, LGBTQ. We used the race group as the group variable.", "The total number of instances is 17281, where the training set has 15360 rows and the test set has 1921 rows.", "CelebA The CelebA dataset #OTHEREFR is the image datasets that has images of faces with various attributes. Features are extracted using CLIP #OTHEREFR embeddings from images.", "We tried gender classification task from this dataset-various other tasks can be constructed. We used \"Young\" as the group variable.", "The total number of instances is 202599, where the training set has 162770 rows and the test set has 39829 rows."], "text_after_citation": ["This dataset has attributes of age, gender, race and we chose gender as label (Male:0, Female: 1), solving gender classification.", "We used the race as the group variable, encoding \"Asian\" group as 1, other race groups as 0.", "The total number of instances is 23705, where the training set has 18964 rows and the test set has 4741 rows."], "citing_paper_content": {"title": "Mitigating Source Bias For Fairer Weak Supervision", "abstract": "Weak supervision overcomes the label bottleneck, enabling efficient development of training sets. Millions of models trained on such datasets have been deployed in the real world and interact with users on a daily basis. However, the techniques that make weak supervision attractive-such as integrating any source of signal to estimate unknown labels-also ensure that the pseudolabels it produces are highly biased. Surprisingly, given everyday use and the potential for increased bias, weak supervision has not been studied from the point of view of fairness. This work begins such a study. Our departure point is the observation that even when a fair model can be built from a dataset with access to ground-truth labels, the corresponding dataset labeled via weak supervision can be arbitrarily unfair. Fortunately, not all is lost: we propose and empirically validate a model for source unfairness in weak supervision, then introduce a simple counterfactual fairness-based technique that can mitigate these biases. Theoretically, we show that it is possible for our approach to simultaneously improve both accuracy and fairness metrics-in contrast to standard fairness approaches that suffer from tradeoffs. Empirically, we show that our technique improves accuracy on weak supervision baselines by as much as 32% while reducing demographic parity gap by 82.5%."}, "cited_paper_content": {"title": "Age Progression/Regression By Conditional Adversarial Autoencoder", "abstract": "If I provide you a face image of mine (without telling you the actual age when I took the picture) and a large amount of face images that I crawled (containing labeled faces of different ages but not necessarily paired), can you show me what I would look like when I am 80 or what I was like when I was 5? The answer is probably a No. Most existing face aging works attempt to learn the transformation between age groups and thus would require the paired samples as well as the labeled query image. In this paper, we look at the problem from a generative modeling perspective such that no paired samples is required. In addition, given an unlabeled image, the generative model can directly produce the image with desired age attribute. We propose a conditional adversarial autoencoder (CAAE) that learns a face manifold, traversing on which smooth age progression and regression can be realized simultaneously. In CAAE, the face is first mapped to a latent vector through a convolutional encoder, and then the vector is projected to the face manifold conditional on age through a deconvolutional generator. The latent vector preserves personalized face features (i.e., personality) and the age condition controls progression vs. regression. Two adversarial networks are imposed on the encoder and generator, respectively, forcing to generate more photo-realistic faces. Experimental results demonstrate the appealing performance and flexibility of the proposed framework by comparing with the state-of-the-art and ground truth."}, "keywords": ["ethnicity", "faces images"], "citation_intent": "method"}
{"citing_id": "2304.00202v1", "cited_id": "1706.06083", "section_title": "Introduction", "citation": "However, most of them adopt a multi-step adversarial attack, i.e., Projected Gradient Descent (PGD) #REFR , to generate adversarial examples for training.", "text_before_citation": ["D EEP neural networks (DNNs) #OTHEREFR , #OTHEREFR , [59] have achieved cutting-edge performance in a lot of problems and tasks.", "However, they have been shown to be vulnerable to adversarial perturbations that are often imperceptible to human observers #OTHEREFR , #OTHEREFR .", "This vulnerability poses a significant security risk for sensitive applications of DNNs.", "To address this issue, a great deal of defense methods #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR have been developed to improve the adversarial robustness of DNNs against adversarial attacks.", "Standard adversarial training #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR has been demonstrated as one of the most effective methods to enhance the robustness of DNNs against adversarial examples."], "text_after_citation": ["This approach incurs a significant training cost, which limits the practical application of standard adversarial training.", "To reduce training costs, fast adversarial training #OTHEREFR , which can be formulated as a mini-max optimization problem, has been proposed to improve adversarial robustness.", "During the inner maximization step, a single-step adversarial attack, such as the Fast Gradient Sign Method (FGSM), is used to generate adversarial examples for training.", "However, after just a few training epochs, the model suddenly loses its robustness. This phenomenon is described as catastrophic overfitting #OTHEREFR .", "To address this issue, several fast adversarial training variants #OTHEREFR , #OTHEREFR , #OTHEREFR have been proposed from different perspectives to enhance the quality of adversarial examples and prevent overfitting. These approaches arXiv:2304.00202v1 [cs. LG] 1 Apr 2023 Fig. 2 ."], "citing_paper_content": {"title": "Improving Fast Adversarial Training With Prior-Guided Knowledge", "abstract": "Fast adversarial training (FAT) is an efficient method to improve robustness. However, the original FAT suffers from catastrophic overfitting, which dramatically and suddenly reduces robustness after a few training epochs. Although various FAT variants have been proposed to prevent overfitting, they require high training costs. In this paper, we investigate the relationship between adversarial example quality and catastrophic overfitting by comparing the training processes of standard adversarial training and FAT. We find that catastrophic overfitting occurs when the attack success rate of adversarial examples becomes worse. Based on this observation, we propose a positive prior-guided adversarial initialization to prevent overfitting by improving adversarial example quality without extra training costs. This initialization is generated by using high-quality adversarial perturbations from the historical training process. We provide theoretical analysis for the proposed initialization and propose a prior-guided regularization method that boosts the smoothness of the loss function. Additionally, we design a prior-guided ensemble FAT method that averages the different model weights of historical models using different decay rates. Our proposed method, called FGSM-PGK, assembles the prior-guided knowledge, i.e., the prior-guided initialization and model weights, acquired during the historical training process. Evaluations of four datasets demonstrate the superiority of the proposed method."}, "cited_paper_content": {"title": "Towards Deep Learning Models Resistant To Adversarial Attacks", "abstract": "Recent work has demonstrated that neural networks are vulnerable to adversarial examples, i.e., inputs that are almost indistinguishable from natural data and yet classified incorrectly by the network. In fact, some of the latest findings suggest that the existence of adversarial attacks may be an inherent weakness of deep learning models. To address this problem, we study the adversarial robustness of neural networks through the lens of robust optimization. This approach provides us with a broad and unifying view on much of the prior work on this topic. Its principled nature also enables us to identify methods for both training and attacking neural networks that are reliable and, in a certain sense, universal. In particular, they specify a concrete security guarantee that would protect against any adversary. These methods let us train networks with significantly improved resistance to a wide range of adversarial attacks. They also suggest the notion of security against a first-order adversary as a natural and broad security guarantee. We believe that robustness against such well-defined classes of adversaries is an important stepping stone towards fully resistant deep learning models."}, "keywords": ["adversarial examples"], "citation_intent": "method"}
{"citing_id": "2304.06908v1", "cited_id": "1312.6199", "section_title": "A. Adversarial Attack", "citation": "Although DNNs have demonstrated to be successful on many tasks, they are vulnerable to adversarial examples #REFR .", "text_before_citation": [], "text_after_citation": ["Let x \u2208 R d denote an image with dimension d, y denote its ground truth label.", "Given a pretrained DNN \u03b8, the adversarial attack is implemented by solving the following optimization problem:", "EQUATION", "where the L(x, y, \u03b8) denotes the loss function that guides the DNN \u03b8 to predict the class of image x as ground-truth y, is the maximum allowed magnitude of perturbation, and x * is the generated adversarial example. Some gradient-based optimizers were proposed to solve (1). Szegedy et al.", "were the first to use a box-constrained L-BFGS optimizer #OTHEREFR ."], "citing_paper_content": {"title": "Generating Adversarial Examples With Better Transferability Via Masking Unimportant Parameters Of Surrogate Model", "abstract": "Deep neural networks (DNNs) have been shown to be vulnerable to adversarial examples. Moreover, the transferability of the adversarial examples has received broad attention in recent years, which means that adversarial examples crafted by a surrogate model can also attack unknown models. This phenomenon gave birth to the transfer-based adversarial attacks, which aim to improve the transferability of the generated adversarial examples. In this paper, we propose to improve the transferability of adversarial examples in the transfer-based attack via masking unimportant parameters (MUP). The key idea in MUP is to refine the pretrained surrogate models to boost the transfer-based attack. Based on this idea, a Taylor expansionbased metric is used to evaluate the parameter importance score and the unimportant parameters are masked during the generation of adversarial examples. This process is simple, yet can be naturally combined with various existing gradient-based optimizers for generating adversarial examples, thus further improving the transferability of the generated adversarial examples. Extensive experiments are conducted to validate the effectiveness of the proposed MUP-based methods."}, "cited_paper_content": {"title": "Intriguing Properties Of Neural Networks", "abstract": "Deep neural networks are highly expressive models that have recently achieved state of the art performance on speech and visual recognition tasks. While their expressiveness is the reason they succeed, it also causes them to learn uninterpretable solutions that could have counter-intuitive properties. In this paper we report two such properties. ::: First, we find that there is no distinction between individual high level units and random linear combinations of high level units, according to various methods of unit analysis. It suggests that it is the space, rather than the individual units, that contains of the semantic information in the high layers of neural networks. ::: Second, we find that deep neural networks learn input-output mappings that are fairly discontinuous to a significant extend. We can cause the network to misclassify an image by applying a certain imperceptible perturbation, which is found by maximizing the network's prediction error. In addition, the specific nature of these perturbations is not a random artifact of learning: the same perturbation can cause a different network, that was trained on a different subset of the dataset, to misclassify the same input."}, "keywords": ["adversarial examples"], "citation_intent": "background"}
{"citing_id": "2303.10533v1", "cited_id": "1612.01474", "section_title": "Discussion", "citation": "Additionally, the ensemble design also provides a new tool to quantify the segmentation uncertainty via examining the variation of sub-model softmax probability maps #REFR .", "text_before_citation": ["The comparison of our deep ensemble model results against mp-MRI+4 PCs model results highlights the benefit of ensemble learning design.", "The four sub-models utilized different input designs that incorporated different 4 RFM PCs.", "Since these 4 RFM PCs were uncorrelated, the sub-model results were likely to be complementary, and thus the synthesized result could be further improved.", "As shown in Figures 4-6 , owing to the adopted Ostu's method, the ensemble result of the ensemble segmentation result captured all key morphological details shown in the sub-model softmax probability maps.", "The simple averaging operation would be unable to retain the complementary image information."], "text_after_citation": ["We adopted the entropy metric below to represent this data-related segmentation uncertainty:", "EQUATION", "Where \u0302( , ) represents the intensity at position ( , ) of the softmax probability map .", "Figure 8 below shows an example uncertainty map (normalized to 0-100) of WT segmentation.", "As seen, high levels of uncertainty are observed at the edges of the WT segment, whereas the segmentation core exhibits minimal levels of uncertainty."], "citing_paper_content": {"title": "A Radiomics-Incorporated Deep Ensemble Learning Model For Multi-Parametric Mri-Based Glioma Segmentation", "abstract": "To develop a deep ensemble learning model with a radiomics spatial encoding execution for improved glioma segmentation accuracy using multi-parametric MRI (mp-MRI). Methods: This model was developed using 369 glioma patients with a 4-modality mp-MRI protocol: T1, contrast-enhanced T1 (T1-Ce), T2, and FLAIR. In each modality volume, a 3D sliding kernel was implemented across the brain to capture image heterogeneity: fifty-six radiomic features were extracted within the kernel, resulting in a 4 th order tensor. Each radiomic feature can then be encoded as a 3D image volume, namely a radiomic feature map (RFM). For each patient, all RFMs extracted from all 4 modalities were processed by the Principal Component Analysis (PCA) for dimension reduction, and the first 4 principal components (PCs) were selected. Next, four deep neural networks as sub-models following the U-Net's architecture were trained for the segmenting of a region-of-interest (ROI): each sub-model utilizes the mp-MRI and 1 of the 4 PCs as a 5channel input for 2D execution. Last, the 4 softmax probability results given by the U-net ensemble were superimposed and binarized by Otsu's method as the segmentation result. Three ensemble models were trained to segment enhancing tumor (ET), tumor core (TC), and whole tumor (WT), respectively. Segmentation results given by the proposed ensemble were compared to the mp-MRI-only U-net results. Results: All 3 radiomics-incorporated deep ensemble learning models were successfully implemented: comparing to mp-MRI-only U-net results, the dice coefficients of ET (0.777\u21920.817), TC (0.742 \u2192 0.757), and WT (0.823 \u2192 0.854) demonstrated improvements. Accuracy, sensitivity, and specificity results demonstrated the same patterns. 3 Conclusion: The adopted radiomics spatial encoding execution enriches the image heterogeneity information that leads to the successful demonstration of the proposed deep ensemble model, which offers a new tool for mp-MRI based medical image segmentation."}, "cited_paper_content": {"title": "Simple And Scalable Predictive Uncertainty Estimation Using Deep Ensembles", "abstract": "Deep neural networks (NNs) are powerful black box predictors that have recently achieved impressive performance on a wide spectrum of tasks. Quantifying predictive uncertainty in NNs is a challenging and yet unsolved problem. Bayesian NNs, which learn a distribution over weights, are currently the state-of-the-art for estimating predictive uncertainty; however these require significant modifications to the training procedure and are computationally expensive compared to standard (non-Bayesian) NNs. We propose an alternative to Bayesian NNs that is simple to implement, readily parallelizable, requires very little hyperparameter tuning, and yields high quality predictive uncertainty estimates. Through a series of experiments on classification and regression benchmarks, we demonstrate that our method produces well-calibrated uncertainty estimates which are as good or better than approximate Bayesian NNs. To assess robustness to dataset shift, we evaluate the predictive uncertainty on test examples from known and unknown distributions, and show that our method is able to express higher uncertainty on out-of-distribution examples. We demonstrate the scalability of our method by evaluating predictive uncertainty estimates on ImageNet."}, "keywords": ["segmentation uncertainty", "sub-model softmax probability"], "citation_intent": "method"}
{"citing_id": "2304.09024v1", "cited_id": "1611.07004", "section_title": "Metrics", "citation": "Additionally, the FCN score #REFR is computed to further evaluate details of the performance in the Cityscapes dataset.", "text_before_citation": ["We follow recent practices #OTHEREFR and report the Kernel Inception Distance (KID) between feature representations of real and fake images.", "The feature extraction #OTHEREFR is done by the Inception v3 model."], "text_after_citation": ["This measures the accuracy of the FCN-8s semantic classifier #OTHEREFR (pre-trained on real images) after segmenting the generated images and comparing the results against the labels these images were synthesized from."], "citing_paper_content": {"title": "Look Atme: The Discriminator Mean Entropy Needs Attention", "abstract": "Figure 1. ATME is a GAN where, for each iteration at epoch t, the input images for the generator are corrupted with a noisy representation of the discriminator's decision map at epoch t \u2212 1. The adversarial game enforces this noise to be removed, sending the proper signals for equilibration by encouraging the discriminator to converge towards its maximum entropy state. Convergence is often slower in the cases where the noise significantly affects the semantic content of the input (a) and is faster in the opposite cases (b)."}, "cited_paper_content": {"title": "Image-To-Image Translation With Conditional Adversarial Networks", "abstract": "We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. This makes it possible to apply the same generic approach to problems that traditionally would require very different loss formulations. We demonstrate that this approach is effective at synthesizing photos from label maps, reconstructing objects from edge maps, and colorizing images, among other tasks. Indeed, since the release of the pix2pix software associated with this paper, a large number of internet users (many of them artists) have posted their own experiments with our system, further demonstrating its wide applicability and ease of adoption without the need for parameter tweaking. As a community, we no longer hand-engineer our mapping functions, and this work suggests we can achieve reasonable results without hand-engineering our loss functions either."}, "keywords": ["Cityscapes dataset"], "citation_intent": "method"}
{"citing_id": "2303.09174v1", "cited_id": "1908.04729", "section_title": "C. Detailed Descriptions Of Datasets", "citation": "Additionally, there are 2,885 and 716 complicated tables are extracted to constitute the SciTSR-COMP #REFR .", "text_before_citation": ["Consequently, it is extended to the partial version, i.e., ICDAR-2013-P.", "ICDAR Considering it is also no training set provided, similar division with ICDAR-2013 dataset is also performed to extend it to the partial version, UNLV-P.", "TableBank #OTHEREFR collects 145,000 tables for training and 1,000 ones for testing, where the logical coordinates of cell boxes are provided.", "PubTabNet #OTHEREFR directly annotates tables as HTML format, which consists of 339,000 training and 114,000 testing samples.", "SciTSR #OTHEREFR is a large-scale dataset containing 12,000 tables for training and 3,000 testing samples, which are mostly from PDF."], "text_after_citation": ["To evaluate the capacity of TSR methods under more challenging scenes, in work #OTHEREFR it is further augmented with synthesizing distortion algorithms to simulate distractors brought by capture device, which is called SciTSR-COMP-A.", "WTW #OTHEREFR is another complex dataset, which contains 10,970 training images and 3,611 testing images collected from wild complex scenes.", "Specifically, it is mainly composed of the tabular objects with explicit boundaries, with information of table id, tabular cell coordinates and row/column information annotated."], "citing_paper_content": {"title": "Grab What You Need: Rethinking Complex Table Structure Recognition With Flexible Components Deliberation", "abstract": "Recently, Table Structure Recognition (TSR) task, aiming at identifying table structure into machine readable formats, has received increasing interest in the community. While impressive success, most single table componentbased methods can not perform well on unregularized table cases distracted by not only complicated inner structure but also exterior capture distortion. In this paper, we raise it as Complex TSR problem, where the performance degeneration of existing methods is attributable to their inefficient component usage and redundant post-processing. To mitigate it, we shift our perspective from table component extraction towards the efficient multiple components leverage, which awaits further exploration in the field. Specifically, we propose a seminal method, termed GrabTab, equipped with newly proposed Component Deliberator. Thanks to its progressive deliberation mechanism, our GrabTab can flexibly accommodate to most complex tables with reasonable components selected but without complicated postprocessing involved. Quantitative experimental results on public benchmarks demonstrate that our method significantly outperforms the state-of-the-arts, especially under more challenging scenes."}, "cited_paper_content": {"title": "Complicated Table Structure Recognition", "abstract": "The task of table structure recognition aims to recognize the internal structure of a table, which is a key step to make machines understand tables. Currently, there are lots of studies on this task for different file formats such as ASCII text and HTML. It also attracts lots of attention to recognize the table structures in PDF files. However, it is hard for the existing methods to accurately recognize the structure of complicated tables in PDF files. The complicated tables contain spanning cells which occupy at least two columns or rows. To address the issue, we propose a novel graph neural network for recognizing the table structure in PDF files, named GraphTSR. Specifically, it takes table cells as input, and then recognizes the table structures by predicting relations among cells. Moreover, to evaluate the task better, we construct a large-scale table structure recognition dataset from scientific papers, named SciTSR, which contains 15,000 tables from PDF files and their corresponding structure labels. Extensive experiments demonstrate that our proposed model is highly effective for complicated tables and outperforms state-of-the-art baselines over a benchmark dataset and our new constructed dataset."}, "keywords": ["716 complicated tables"], "citation_intent": "method"}
{"citing_id": "2303.13863v1", "cited_id": "1804.02767", "section_title": "4) Loss", "citation": "We compared our model to YoloV3 #REFR and a fully unfrozen YoloV5. The results are illustrated in Fig. 10 .", "text_before_citation": ["As previously indicated, we leveraged YoloV5's architecture to adapt its robust performance and precision.", "However, we compared our framework to other cutting-edge architecture, and MagicEye outperforms every comparison.", "MagicEye, in particular, employs YoloV5 architecture with a frozen backbone.", "By freezing weights, the gradient computations of the network are minimized, thus lowering the model's dimensionality.", "This increases efficiency by significantly reducing storage, transfer costs, and computation required for both the forward and backward passes."], "text_after_citation": ["In addition, a thorough evaluation was performed on the currency detection and facial recognition capabilities of MagicEye.", "As part of the evaluation process, we utilized various metrics such as precision, recall, and f1 score, as demonstrated in Table III .", "These metrics serve to assess the robustness of the model and provide a clear understanding of its performance.", "In conclusion, the MagicEye system represents a significant advancement in the field of visual aids for the visually impaired.", "Its real-time obstacle detection, turn-by-turn navigation, and advanced facial recognition and currency detection capabilities make it a comprehensive solution for the challenges faced by the visually impaired community."], "citing_paper_content": {"title": "Magiceye: An Intelligent Wearable Towards Independent Living Of Visually Impaired", "abstract": "Individuals with visual impairments often face a multitude of challenging obstacles in their daily lives. Vision impairment can severely impair a person's ability to work, navigate, and retain independence. This can result in educational limits, a higher risk of accidents, and a plethora of other issues. To address these challenges, we present MagicEye, a state-of-the-art intelligent wearable device designed to assist visually impaired individuals. MagicEye employs a custom-trained CNNbased object detection model, capable of recognizing a wide range of indoor and outdoor objects frequently encountered in daily life. With a total of 35 classes, the neural network employed by MagicEye has been specifically designed to achieve high levels of efficiency and precision in object detection. The device is also equipped with facial recognition and currency identification modules, providing invaluable assistance to the visually impaired. In addition, MagicEye features a GPS sensor for navigation, allowing users to move about with ease, as well as a proximity sensor for detecting nearby objects without physical contact. In summary, MagicEye is an innovative and highly advanced wearable device that has been designed to address the many challenges faced"}, "cited_paper_content": {"title": "Yolov3: An Incremental Improvement", "abstract": "We present some updates to YOLO! We made a bunch of little design changes to make it better. We also trained this new network that's pretty swell. It's a little bigger than last time but more accurate. It's still fast though, don't worry. At 320x320 YOLOv3 runs in 22 ms at 28.2 mAP, as accurate as SSD but three times faster. When we look at the old .5 IOU mAP detection metric YOLOv3 is quite good. It achieves 57.9 mAP@50 in 51 ms on a Titan X, compared to 57.5 mAP@50 in 198 ms by RetinaNet, similar performance but 3.8x faster. As always, all the code is online at this https URL"}, "keywords": ["YoloV3"], "citation_intent": "result"}
{"citing_id": "2303.01429v1", "cited_id": "1411.1792", "section_title": "Introduction", "citation": "Another popular and simple transfer learning strategy is keeping frozen most of the network #REFR , and training to the target task only some of the model parameters.", "text_before_citation": ["Indeed, fine-tuning the already learned meaningful representation to the target task allows one to work with dramatically smaller dataset sizes while keeping the same level of accuracy.", "Unfortunately, deep neural networks tend to be sensitive even to tiny distribution shifts #OTHEREFR .", "Therefore, performing transfer learning starting from the right initialization and with the right protocol is crucial for success.", "One of the major risks is to incur in overfitting #OTHEREFR due to the limited size of the target dataset.", "Common strategies to avoid it include damping the learning rate in the fine-tuning stage #OTHEREFR , which forces the network to change its weights only by a small amount."], "text_after_citation": ["Transferring the layers directly from another network allows, in principle, filtering out irrelevant information present in the input of both the source and the target data sets, effectively reducing the dimensionality of the problem and increasing data efficiency.", "But how should one choose the layers that should be kept frozen? Ideally, one should be able to identify, in the source network, a layer in which the representation is general enough to be meaningful also for the target task, but not too general, as otherwise training it to a specialized task might be hindered by the scarcity of data in the target set.", "Given the relevance of transfer learning in realworld applications, it is important to define simple protocols that can be applied without too much finetuning or domain knowledge.", "In this work, we propose an approach capable of identifying the layer in which #OTHEREFR to CIFAR-10 (a,d) with 128 images per class and full dataset respectively, from ImageNet #OTHEREFR to Breast Cancer #OTHEREFR with 10 4 images per class (b), from CIFAR-10 to GM with 32 images per class(c).", "We can see samples of each dataset in the lower part of the panels, while the curves above -the defrosting profiles -represent the test performance of the network."], "citing_paper_content": {"title": "Optimal Transfer Protocol By Incremental Layer Defrosting", "abstract": "Transfer learning is a powerful tool enabling model training with limited amounts of data. This technique is particularly useful in real-world problems where data availability is often a serious limitation. The simplest transfer learning protocol is based on \"freezing\" the feature-extractor layers of a network pretrained on a data-rich source task, and then adapting only the last layers to a data-poor target task. This workflow is based on the assumption that the feature maps of the pre-trained model are qualitatively similar to the ones that would have been learned with enough data on the target task. In this work, we show that this protocol is often sub-optimal, and the largest performance gain may be achieved when smaller portions of the pre-trained network are kept frozen. In particular, we make use of a controlled framework to identify the optimal transfer depth, which turns out to depend non-trivially on the amount of available training data and on the degree of source-target task correlation. We then characterize transfer optimality by analyzing the internal representations of two networks trained from scratch on the source and the target task through multiple established similarity measures."}, "cited_paper_content": {"title": "How Transferable Are Features In Deep Neural Networks?", "abstract": "Many deep neural networks trained on natural images exhibit a curious phenomenon in common: on the first layer they learn features similar to Gabor filters and color blobs. Such first-layer features appear not to be specific to a particular dataset or task, but general in that they are applicable to many datasets and tasks. Features must eventually transition from general to specific by the last layer of the network, but this transition has not been studied extensively. In this paper we experimentally quantify the generality versus specificity of neurons in each layer of a deep convolutional neural network and report a few surprising results. Transferability is negatively affected by two distinct issues: (1) the specialization of higher layer neurons to their original task at the expense of performance on the target task, which was expected, and (2) optimization difficulties related to splitting networks between co-adapted neurons, which was not expected. In an example network trained on ImageNet, we demonstrate that either of these two issues may dominate, depending on whether features are transferred from the bottom, middle, or top of the network. We also document that the transferability of features decreases as the distance between the base task and target task increases, but that transferring features even from distant tasks can be better than using random features. A final surprising result is that initializing a network with transferred features from almost any number of layers can produce a boost to generalization that lingers even after fine-tuning to the target dataset."}, "keywords": ["simple transfer learning"], "citation_intent": "background"}
{"citing_id": "2303.01331v1", "cited_id": "1907.10043", "section_title": "B. Canonical Mapping", "citation": "Canonical Surface Mapping #REFR approach demonstrated the ability to learn dense pixelto-3D correspondences with limited supervision by ensuring that the projection of a predicted 3D point back onto the image maps close to the original pixel (cycle of geometric consistency).", "text_before_citation": ["The problem can be viewed as an example of a more general area of estimation of dense semantic correspondences.", "The latter area includes more established tasks, such as computing pixel-to-pixel correspondences (e.g., SIFT).", "The advantage of using a dense mesh as one of the mapping domains is that it contains a complete representation of the object (rather than a partial view), includes 3D information, and provides a convenient modality to select and define manipulation areas and strategy.", "To estimate canonical mapping, DensePose #OTHEREFR partitions a 3D shape into surface parts (parameterized as 2D charts) and learns to regress to those parts' uv coordinates (pixel-to-surface correspondences).", "Training such a model requires millions of annotated correspondences between images and 3D models."], "text_after_citation": ["The work was extended in #OTHEREFR to the prediction of articulations of predefined object parts (parametrized as 6D poses of vertex groups).", "Continuous Surface Embeddings (CSE) #OTHEREFR method used as part of our approach, addresses computing dense correspondences by learning to predict, for each foreground pixel in an image, an embedding vector of the corresponding vertex in the object's mesh.", "CSE achieves state-of-the-art performance while using fewer annotations and conceptually simpler architecture.", "Additionally, the method shows the generalizability potential and the ability to relate and share computation for several object categories."], "citing_paper_content": {"title": "Canonical Mapping As A General-Purpose Object Descriptor For Robotic Manipulation", "abstract": "Perception is an essential part of robotic manipulation in a semi-structured environment. Traditional approaches produce a narrow task-specific prediction (e.g., object's 6D pose), that cannot be adapted to other tasks and is ill-suited for deformable objects. In this paper, we propose using canonical mapping as a near-universal and flexible object descriptor. We demonstrate that common object representations can be derived from a single pre-trained canonical mapping model, which in turn can be generated with minimal manual effort using an automated data generation and training pipeline. We perform a multi-stage experiment using two robot arms that demonstrate the robustness of the perception approach and the ways it can inform the manipulation strategy, thus serving as a powerful foundation for general-purpose robotic manipulation."}, "cited_paper_content": {"title": "Canonical Surface Mapping Via Geometric Cycle Consistency", "abstract": "We explore the task of Canonical Surface Mapping (CSM). Specifically, given an image, we learn to map pixels on the object to their corresponding locations on an abstract 3D model of the category. But how do we learn such a mapping? A supervised approach would require extensive manual labeling which is not scalable beyond a few hand-picked categories. Our key insight is that the CSM task (pixel to 3D), when combined with 3D projection (3D to pixel), completes a cycle. Hence, we can exploit a geometric cycle consistency loss, thereby allowing us to forgo the dense manual supervision. Our approach allows us to train a CSM model for a diverse set of classes, without sparse or dense keypoint annotation, by leveraging only foreground mask labels for training. We show that our predictions also allow us to infer dense correspondence between two images, and compare the performance of our approach against several methods that predict correspondence by leveraging varying amount of supervision."}, "keywords": ["dense pixelto-3D correspondences"], "citation_intent": "background"}
{"citing_id": "2304.00648v1", "cited_id": "1406.2661", "section_title": "E. Deep Learning Architectures", "citation": "During GAN training, the G network intends to learn the distribution P g over the training data X by simultaneously maximizing the correct decision probability D(x) and minimizing the term (1 \u2212 D(G(z))) corresponding to the G network #REFR .", "text_before_citation": ["GAN training can be viewed as a minimax two-player game in which the G attempts to learn the training data distribution and estimate a mapping function capable of generating new samples that increase the D's probability of making the wrong decision, while D tries to maximize its probability of making the right decision (i.e., differentiating a training sample from a generated one) #OTHEREFR , #OTHEREFR .", "The training results in a unique solution when the D's output is uniformly distributed with a probability of one-half everywhere.", "When MLP networks are used for both the G and D networks, backpropagation can be used to train the entire system where the NN representing the G recovers the data distribution without access to the training data and D maximizes the probability of making the right binary decision #OTHEREFR , #OTHEREFR .", "If the prior probability of G input z is P z (z), then the generator mapping function can be given by G(z; \u03b8 g ) where \u03b8 g is the MLP parameters of G.", "The D function is given by D(x; \u03b8 d ) for input x and a single output D(x), which represents the probability that x came from the training data and not the G #OTHEREFR ."], "text_after_citation": ["The GAN minimax optimization problem can be described by the following objective function,", "min G max D V (D, G) = E x\u223cPd(x) {log[D(x)]} + E z\u223cPz(z) {log[1 \u2212 D(G(z))]}, (15)", "where E is the expected value.", "The optimum point is reached when the G perfectly recovers the training data distribution (i.e.", "P g = P data ) #OTHEREFR , #OTHEREFR ."], "citing_paper_content": {"title": "Improving Rf-Dna Fingerprinting Performance In An Indoor Multipath Environment Using Semi-Supervised Learning", "abstract": "The number of Internet of Things (IoT) deployments is expected to reach 75.4 billion by 2025. Roughly 70% of all IoT devices employ weak or no encryption; thus, putting them and their connected infrastructure at risk of attack by devices that are wrongly authenticated or not authenticated at all. A physical layer-based security approach-known as Specific Emitter Identification (SEI)-has been proposed and is being pursued as a viable IoT security mechanism. SEI is advantageous because it is a passive technique that exploits inherent and distinct features that are unintentionally imparted upon the signal during its formation and transmission within and by the IoT device's Radio Frequency (RF) front-end. SEI's passive exploitation of unintentional signal features removes any need to modify the IoT device, which makes it ideal for existing and future IoT deployments. Despite the amount of SEI research conducted there remains challenges that must be addressed to make SEI a viable IoT security approach. One of these challenges is the extraction of SEI features from signals collected under multipath fading conditions. Multipath corrupts the inherent SEI exploited features that are used to discriminate one IoT device from another; thus, degrading authentication performance and increasing the chance of attack. This work presents two semi-supervised Deep Learning (DL) equalization approaches and compares their performance with the current state of the art. The two approaches are the Conditional Generative Adversarial Network (CGAN) and Joint Convolutional Auto-Encoder and Convolutional Neural Network (JCAECNN). Both approaches learn the channel distribution to enable multipath correction while simultaneously preserving the SEI exploited signal features. CGAN and JCAECNN performance is assessed using a Rayleigh fading channel under degrading SNR, up to thirty-two IoT devices, as well as two publicly available signal sets. The JCAECNN improves SEI performance by 10% beyond that of the current state of the art."}, "cited_paper_content": {"title": "Generative Adversarial Nets", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to \u00bd everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["GAN training"], "citation_intent": "background"}
{"citing_id": "2303.01123v1", "cited_id": "1907.10250", "section_title": "Ii. Related Work", "citation": "It was shown in #REFR that the introduced quadric loss in combination with the Chamfer loss improves 3D reconstruction accuracy by preserving sharp edges of scene objects.", "text_before_citation": ["The authors utilize the model for lidars and cameras calibration scenario.", "However, the application of the model in localization and mapping scenarios was not studied in their paper.", "As we pose the task of global point cloud map consistency estimation as optimization problem, a differentiable loss function is required as minimization objective.", "The Chamfer loss introduced in #OTHEREFR is one of the most famous differentiable function for point sets.", "It is usually utilized to evaluate a distance between two point clouds, thus is typically suitable for supervised tasks."], "text_after_citation": ["The loss is designed to reduce quadric error between the reconstructed points and the input surface.", "However, as it was mentioned by the authors of #OTHEREFR , the quadric loss can not preserve the input point distribution, since the quadric loss is zero anywhere on the plane.", "On the contrary, in order to reduce the lidar measurements bias caused by high incidence angle, we target to correct mainly the points which belong to planar surfaces.", "Therefore a designed global map consistency loss should be able to encourage points from different scans that represent such objects as walls, floors, and roads to belong to the same plane.", "Moreover, as shown in the next Section III, we do not require ground-truth data to evaluate the point cloud consistency, and the introduced losses are computed directly on the global map independently on the number of lidar scans and view-points forming the input point cloud."], "citing_paper_content": {"title": "Self-Supervised Depth Correction Of Lidar Measurements From Map Consistency Loss", "abstract": "Depth perception is considered an invaluable source of information in the context of 3D mapping and various robotics applications. However, point cloud maps acquired using consumer-level light detection and ranging sensors (lidars) still suffer from bias related to local surface properties such as measuring beam-to-surface incidence angle, distance, texture, reflectance, or illumination conditions. This fact has recently motivated researchers to exploit traditional filters, as well as the deep learning paradigm, in order to suppress the aforementioned depth sensors error while preserving geometric and map consistency details. Despite the effort, depth correction of lidar measurements is still an open challenge mainly due to the lack of clean 3D data that could be used as ground truth. In this paper, we introduce two novel point cloud map consistency losses, which facilitate self-supervised learning on real data of lidar depth correction models. Specifically, the models exploit multiple point cloud measurements of the same scene from different viewpoints in order to learn to reduce the bias based on the constructed map consistency signal. Complementary to the removal of the bias from the measurements, we demonstrate that the depth correction models help to reduce localization drift. Additionally, we release a data set that contains point cloud data captured in an indoor corridor environment with precise localization and ground truth mapping information 1 ."}, "cited_paper_content": {"title": "Learning Embedding Of 3D Models With Quadric Loss", "abstract": "Sharp features such as edges and corners play an important role in the perception of 3D models. In order to capture them better, we propose quadric loss, a point-surface loss function, which minimizes the quadric error between the reconstructed points and the input surface. Computation of Quadric loss is easy, efficient since the quadric matrices can be computed apriori, and is fully differentiable, making quadric loss suitable for training point and mesh based architectures. Through extensive experiments we show the merits and demerits of quadric loss. When combined with Chamfer loss, quadric loss achieves better reconstruction results as compared to any one of them or other point-surface loss functions."}, "keywords": ["3D reconstruction accuracy", "introduced quadric loss"], "citation_intent": "background"}
{"citing_id": "2304.06021v1", "cited_id": "1808.01050", "section_title": "Fully-Supervised Crowd Counting", "citation": "In UCF-QNRF dataset #REFR , it costs over 2000 human hours to annotate only 1535 images, in which the average cost per image is over one hour.", "text_before_citation": ["In addition, #OTHEREFR proposed Mean Count Proxies to select the best count proxy for each interval to represent its count value during inference.", "#OTHEREFR introduces transformer into crowd localization, proposing a crowd localization transformer to directly predict the localization.", "#OTHEREFR proposes a local matching point-based framework to alleviate instability.", "These fully supervised methods achieve incredible performance on various datasets.", "However, the complete annotation for a crowd image is time-consuming and laborious."], "text_after_citation": [], "citing_paper_content": {"title": "Crowd Counting With Sparse Annotation", "abstract": "This paper presents a new annotation method called Sparse Annotation (SA) for crowd counting, which reduces human labeling efforts by sparsely labeling individuals in an image. We argue that sparse labeling can reduce the redundancy of full annotation and capture more diverse information from distant individuals that is not fully captured by Partial Annotation methods. Besides, we propose a pointbased Progressive Point Matching network (PPM) to better explore the crowd from the whole image with sparse annotation, which includes a Proposal Matching Network (PMN) and a Performance Restoration Network (PRN). The PMN generates pseudo-point samples using a basic point classifier, while the PRN refines the point classifier with the pseudo points to maximize performance. Our experimental results show that PPM outperforms previous semisupervised crowd counting methods with the same amount of annotation by a large margin and achieves competitive performance with state-of-the-art fully-supervised methods."}, "cited_paper_content": {"title": "Composition Loss For Counting, Density Map Estimation And Localization In Dense Crowds", "abstract": "With multiple crowd gatherings of millions of people every year in events ranging from pilgrimages to protests, concerts to marathons, and festivals to funerals; visual crowd analysis is emerging as a new frontier in computer vision. In particular, counting in highly dense crowds is a challenging problem with far-reaching applicability in crowd safety and management, as well as gauging political significance of protests and demonstrations. In this paper, we propose a novel approach that simultaneously solves the problems of counting, density map estimation and localization of people in a given dense crowd image. Our formulation is based on an important observation that the three problems are inherently related to each other making the loss function for optimizing a deep CNN decomposable. Since localization requires high-quality images and annotations, we introduce UCF-QNRF dataset that overcomes the shortcomings of previous datasets, and contains 1.25 million humans manually marked with dot annotations. Finally, we present evaluation measures and comparison with recent deep CNN networks, including those developed specifically for crowd counting. Our approach significantly outperforms state-of-the-art on the new dataset, which is the most challenging dataset with the largest number of crowd annotations in the most diverse set of scenes."}, "keywords": ["1535 images", "UCF-QNRF dataset"], "citation_intent": "background"}
{"citing_id": "2303.07598v1", "cited_id": "1805.12152", "section_title": "Generalization Of Adpe In", "citation": "The results also confirm the existing observation in literature #REFR that the instance-wise perturbations cannot improve the standard accuracy for the downstream tasks.", "text_before_citation": ["It adds pixel-wise perturbations on the raw inputs that maximize the loss to make the worst-case attacks, and uses the resultant perturbed inputs to train the backbone network.", "We extend the FGSM #OTHEREFR that applies instance-wise perturbations to image pixels.", "It follows the classic FGSM except the MAE reconstruction loss with the multi-crop tokenization in MAE+ are adopted to compute the adversarial perturbations.", "Table 7 reports the fine-tuning results of the FGSM with various for \u221e constraint over 400 epochs of pretraining ViT-B. FGSM does not perform better than MAE+ (83.51%).", "This suggests that a straight extension of FGSM cannot improve the accuracy of a MIM-pretrained model in downstream tasks."], "text_after_citation": [], "citing_paper_content": {"title": "Adpe: Adversarial Positional Embeddings For Pretraining Vision Transformers Via Mae+", "abstract": "Unsupervised learning of vision transformers seeks to pretrain an encoder via pretext tasks without labels. Among them is the Masked Image Modeling (MIM) aligned with pretraining of language transformers by predicting masked patches as a pretext task. A criterion in unsupervised pretraining is the pretext task needs to be sufficiently hard to prevent the transformer encoder from learning trivial low-level features not generalizable well to downstream tasks. For this purpose, we propose an Adversarial Positional Embedding (AdPE) approach-It distorts the local visual structures by perturbing the position encodings so that the learned transformer cannot simply use the locally correlated patches to predict the missing ones. We hypothesize that it forces the transformer encoder to learn more discriminative features in a global context with stronger generalizability to downstream tasks. We will consider both absolute and relative positional encodings, where adversarial positions can be imposed both in the embedding mode and the coordinate mode. We will also present a new MAE+ baseline that brings the performance of the MIM pretraining to a new level with the AdPE. The experiments demonstrate that our approach can improve the fine-tuning accuracy of MAE by 0.8% and 0.4% over 1600 epochs of pretraining ViT-B and ViT-L on Imagenet1K. For the transfer learning task, it outperforms the MAE with the ViT-B backbone by 2.6% in mIoU on ADE20K, and by 3.2% in AP bbox and 1.6% in AP mask on COCO, respectively. These results are obtained with the AdPE being a pure MIM approach that does not use any extra models or external datasets for pretraining."}, "cited_paper_content": {"title": "Robustness May Be At Odds With Accuracy", "abstract": "We show that there exists an inherent tension between the goal of adversarial robustness and that of standard generalization. Specifically, training robust models may not only be more resource-consuming, but also lead to a reduction of standard accuracy. We demonstrate that this trade-off between the standard accuracy of a model and its robustness to adversarial perturbations provably exists even in a fairly simple and natural setting. These findings also corroborate a similar phenomenon observed in practice. Further, we argue that this phenomenon is a consequence of robust classifiers learning fundamentally different feature representations than standard classifiers. These differences, in particular, seem to result in unexpected benefits: the representations learned by robust models tend to align better with salient data characteristics and human perception."}, "keywords": ["instance-wise perturbations"], "citation_intent": "result"}
{"citing_id": "2304.03159v1", "cited_id": "1911.02116", "section_title": "B. Main Results", "citation": "In the results on cross-lingual classification in #REFR , performance on Arabic and Chinese is obviously worse than in German and Spanish.", "text_before_citation": ["Baseline model XLM-R-Base has the best performance when the questions are set to German and Spanish.", "Results on cross-lingual classification reported in #OTHEREFR indicate that the performance on Spanish and German are also the highest except English.", "We can also see from Fig.5 that the token coverage of these two languages is only a little different from the others.", "Take all these into consideration, and we think the limited improvement is caused by the deep understanding of German and Spanish in the baseline PLM.", "Significantly, in the Arabic and Chinese cases, the baseline performance is the worst, indicating less knowledge storage and shallow understanding of the two languages."], "text_after_citation": ["In some cases, the cough may not be present or may only occur occasionally and may not be productive.", "Some people with COPD attribute the symptoms to a \"smoker's cough\".", "Sputum may be swallowed or spat out, depending often on social and cultural factors.", "Vigorous coughing may lead to rib fractures or a brief loss of consciousness.", "Those with COPD often have a history of \"common colds\" that last a long time."], "citing_paper_content": {"title": "Bridging The Language Gap: Knowledge Injected Multilingual Question Answering", "abstract": "Question Answering (QA) is the task of automatically answering questions posed by humans in natural languages. There are different settings to answer a question, such as abstractive, extractive, boolean, and multiple-choice QA. As a popular topic in natural language processing tasks, extractive question answering task (extractive QA) has gained extensive attention in the past few years. With the continuous evolvement of the world, generalized cross-lingual transfer (G-XLT), where question and answer context are in different languages, poses some unique challenges over cross-lingual transfer (XLT), where question and answer context are in the same language. With the boost of corresponding development of related benchmarks, many works have been done to improve the performance of various language QA tasks. However, only a few works are dedicated to the G-XLT task. In this work, we propose a generalized crosslingual transfer framework to enhance the model's ability to understand different languages. Specifically, we first assemble triples from different languages to form multilingual knowledge. Since the lack of knowledge between different languages greatly limits models' reasoning ability, we further design a knowledge injection strategy via leveraging link prediction techniques to enrich the model storage of multilingual knowledge. In this way, we can profoundly exploit rich semantic knowledge. Experiment results on real-world datasets MLQA demonstrate that the proposed method can improve the performance by a large margin, outperforming the baseline method by 13.18%/12.00% F1/EM on average. Our code and models will be open-sourced."}, "cited_paper_content": {"title": "Unsupervised Cross-Lingual Representation Learning At Scale", "abstract": "This paper shows that pretraining multilingual language models at scale leads to significant performance gains for a wide range of cross-lingual transfer tasks. We train a Transformer-based masked language model on one hundred languages, using more than two terabytes of filtered CommonCrawl data. Our model, dubbed XLM-R, significantly outperforms multilingual BERT (mBERT) on a variety of cross-lingual benchmarks, including +13.8% average accuracy on XNLI, +12.3% average F1 score on MLQA, and +2.1% average F1 score on NER. XLM-R performs particularly well on low-resource languages, improving 11.8% in XNLI accuracy for Swahili and 9.2% for Urdu over the previous XLM model. We also present a detailed empirical evaluation of the key factors that are required to achieve these gains, including the trade-offs between (1) positive transfer and capacity dilution and (2) the performance of high and low resource languages at scale. Finally, we show, for the first time, the possibility of multilingual modeling without sacrificing per-language performance; XLM-Ris very competitive with strong monolingual models on the GLUE and XNLI benchmarks. We will make XLM-R code, data, and models publicly available."}, "keywords": ["cross-lingual classification"], "citation_intent": "result"}
{"citing_id": "2303.14304v1", "cited_id": "1910.14667", "section_title": "Related Work", "citation": "These methods have mainly focused on classification models, and ensemble attacks on dense prediction tasks such as object detection and semantic segmentation are relatively less studied, especially for targeted attacks #REFR . Attacks against object detectors and segmentation.", "text_before_citation": ["Transfer-based attacks rely on the assumption that surrogate models share similarities with the victim model, such that an adversarial example generated for the surrogate model can also fool the victim model.", "Query-based methods generate attacks by searching the adversarial examples space based on the feedback obtained from the victim model through queries.", "They can often achieve higher success rate but may require a large number of queries. Ensemble-based attacks.", "Ensemble-based attacks leverage the idea of transfer attack and assume that if an adversarial example can fool multiple models simultaneously, the chances of fooling an unseen model are higher #OTHEREFR .", "Recently, some methods have combined ensemble-based transfer attacks with limited feedback from the victim models to improve the overall success rate #OTHEREFR ."], "text_after_citation": ["Dense (pixel-level) prediction tasks such as object detection and semantic segmentation have higher task complexities #OTHEREFR compared to classification tasks.", "Existing attacks on object detectors mainly focus on whitebox setting, although there are a few exceptions #OTHEREFR .", "A recent study #OTHEREFR generates blackbox attacks on object detectors by using a surrogate ensemble and context-aware attack-based queries.", "Another approach #OTHEREFR trains a generative model to generate transferable attacks.", "While some patch-based attacks #OTHEREFR are effective, the patches are easily noticeable."], "citing_paper_content": {"title": "Ensemble-Based Blackbox Attacks On Dense Prediction", "abstract": "We propose an approach for adversarial attacks on dense prediction models (such as object detectors and segmentation). It is well known that the attacks generated by a single surrogate model do not transfer to arbitrary (blackbox) victim models. Furthermore, targeted attacks are often more challenging than the untargeted attacks. In this paper, we show that a carefully designed ensemble can create effective attacks for a number of victim models. In particular, we show that normalization of the weights for individual models plays a critical role in the success of the attacks. We then demonstrate that by adjusting the weights of the ensemble according to the victim model can further improve the performance of the attacks. We performed a number of experiments for object detectors and segmentation to highlight the significance of the our proposed methods. Our proposed ensemble-based method outperforms existing blackbox attack methods for object detection and segmentation. Finally we show that our proposed method can also generate a single perturbation that can fool multiple blackbox detection and segmentation models simultaneously. Code is available at https://github.com/CSIPlab/EBAD."}, "cited_paper_content": {"title": "Making An Invisibility Cloak: Real World Adversarial Attacks On Object Detectors", "abstract": "We present a systematic study of adversarial attacks on state-of-the-art object detection frameworks. Using standard detection datasets, we train patterns that suppress the objectness scores produced by a range of commonly used detectors, and ensembles of detectors. Through extensive experiments, we benchmark the effectiveness of adversarially trained patches under both white-box and black-box settings, and quantify transferability of attacks between datasets, object classes, and detector models. Finally, we present a detailed study of physical world attacks using printed posters and wearable clothes, and rigorously quantify the performance of such attacks with different metrics."}, "keywords": ["ensemble attacks", "object detectors"], "citation_intent": "background"}
{"citing_id": "2304.14837v1", "cited_id": "1511.07247", "section_title": "B.4. Qualitative Results On Aachen", "citation": "By comparing the results of EIMP in #REFR and 3, we see that the number of preserved keypoints are based on the number of potential inliers in the image pair: more potential inliers result in more retrained keypoints. That is because our sampling strategy is fully adaptive.", "text_before_citation": ["Besides, the number of keypoints in each image are shown at the top of each pair.", "For simple case (1), all methods give a large number of inliers, but IMP and EIMP report smaller pose errors even when using fewer inliers.", "Note that instead of keeping all keypoints, EIMP effectively reduces the number of keypoints from 2,0000 to 865 and 255 in the iteration process, significantly decreasing the time complexity for self and cross attention computation 7.", "For tough cases ((2), (3)), due to large viewpoint and illumination changes, SuperGlue* and SGMNet fail to report comparable number of inliers to our models, resulting in higher rotation and translation errors.", "In contrast, our models still progressively increase the number of inliers from different regions in the iteration process. These well distributed inliers lead to smaller pose errors."], "text_after_citation": [], "citing_paper_content": {"title": "Imp: Iterative Matching And Pose Estimation With Adaptive Pooling", "abstract": "Previous methods solve feature matching and pose estimation using a two-stage process by first finding matches and then estimating the pose. As they ignore the geometric relationships between the two tasks, they focus on either improving the quality of matches or filtering potential outliers, leading to limited efficiency or accuracy. In contrast, we propose an iterative matching and pose estimation framework (IMP) leveraging the geometric connections between the two tasks: a few good matches are enough for a roughly accurate pose estimation; a roughly accurate pose can be used to guide the matching by providing geometric constraints. To this end, we implement a geometry-aware recurrent attention-based module which jointly outputs sparse matches and camera poses. Specifically, for each iteration, we first implicitly embed geometric information into the module via a pose-consistency loss, allowing it to predict geometry-aware matches progressively. Second, we introduce an efficient IMP, called EIMP, to dynamically discard keypoints without potential matches, avoiding redundant updating and significantly reducing the quadratic time complexity of attention computation in transformers. Experiments on YFCC100m, Scannet, and Aachen Day-Night datasets demonstrate that the proposed method outperforms previous approaches in terms of accuracy and efficiency."}, "cited_paper_content": {"title": "Netvlad: Cnn Architecture For Weakly Supervised Place Recognition", "abstract": "We tackle the problem of large scale visual place recognition, where the task is to quickly and accurately recognize the location of a given query photograph. We present the following four principal contributions. First, we develop a convolutional neural network (CNN) architecture that is trainable in an end-to-end manner directly for the place recognition task. The main component of this architecture, NetVLAD, is a new generalized VLAD layer, inspired by the \u201cVector of Locally Aggregated Descriptors\u201d image representation commonly used in image retrieval. The layer is readily pluggable into any CNN architecture and amenable to training via backpropagation. Second, we create a new weakly supervised ranking loss, which enables end-to-end learning of the architecture's parameters from images depicting the same places over time downloaded from Google Street View Time Machine. Third, we develop an efficient training procedure which can be applied on very large-scale weakly labelled tasks. Finally, we show that the proposed architecture and training procedure significantly outperform non-learnt image representations and off-the-shelf CNN descriptors on challenging place recognition and image retrieval benchmarks."}, "keywords": ["retrained keypoints"], "citation_intent": "result"}
{"citing_id": "2303.14173v2", "cited_id": "1809.02104", "section_title": "Comparison With Existing Theoretical Predictions", "citation": "Those of Item (iv) are not obviously consistent with our experimental data, although we refrain from saying that they are inconsistent since the analysis of #REFR involves a series of inequalities, and it is unclear how the predictions would change if one used slightly different approximations.", "text_before_citation": ["(ii) #OTHEREFR predict that adversarial success is a function of ( dim V dim X ) 1 2 (for p = 2).", "(iii) From the analysis in #OTHEREFR one can predict that adversarial success is a function of ( dim V dim X ) 1 q (for 1 p + 1 q = 1).", "However, arriving at such a functional form from the derivations in #OTHEREFR requires multiple non-trivial steps not carried out in those works.", "(iv) #OTHEREFR predicts that adversarial success is a function of 2,p) .", "The predictions of Items (i) to (iii) are all consistent with our experimental results, suggesting that the situation where adversarial examples are constrained to a subspace of dimension d is effectively equivalent to the unconstrained situation where data is found in an ambient space of dimension d."], "text_after_citation": ["#OTHEREFR The dependence of adversarial success on ( dim V dim X ) 1 q can be derived from the simplest possible toy model, namely binary linear classification. Let X = R n and suppose", "dim V 1 2 \u2212 1 min(", "EQUATION", "Let V \u2286 R n be a subspace with dim V = d. We will assume there exists an isometry U :", "R d \u2212 \u2192 V"], "citing_paper_content": {"title": "How Many Dimensions Are Required To Find An Adversarial Example?", "abstract": "Past work exploring adversarial vulnerability have focused on situations where an adversary can perturb all dimensions of model input. On the other hand, a range of recent works consider the case where either (i) an adversary can perturb a limited number of input parameters or (ii) a subset of modalities in a multimodal problem. In both of these cases, adversarial examples are effectively constrained to a subspace V in the ambient input space X. Motivated by this, in this work we investigate how adversarial vulnerability depends on dim(V). In particular, we show that the adversarial success of standard PGD attacks with p norm constraints behaves like a monotonically increasing function of (dim(V) dim X) 1 q where is the perturbation budget and 1 p + 1 q = 1, provided p > 1 (the case p = 1 presents additional subtleties which we analyze in some detail). This functional form can be easily derived from a simple toy linear model, and as such our results land further credence to arguments that adversarial examples are endemic to locally linear models on high dimensional spaces."}, "cited_paper_content": {"title": "Are Adversarial Examples Inevitable?", "abstract": "A wide range of defenses have been proposed to harden neural networks against adversarial attacks. However, a pattern has emerged in which the majority of adversarial defenses are quickly broken by new attacks. Given the lack of success at generating robust defenses, we are led to ask a fundamental question: Are adversarial attacks inevitable? This paper analyzes adversarial examples from a theoretical perspective, and identifies fundamental bounds on the susceptibility of a classifier to adversarial attacks. We show that, for certain classes of problems, adversarial examples are inescapable. Using experiments, we explore the implications of theoretical guarantees for real-world problems and discuss how factors such as dimensionality and image complexity limit a classifier's robustness against adversarial examples."}, "keywords": ["experimental data"], "citation_intent": "result"}
{"citing_id": "2303.11143v1", "cited_id": "1910.07517", "section_title": "Attacks Against Models For Source Code Analysis", "citation": "The general white-box attack of #REFR iteratively substitutes a target variable name in all of its occurrences with an alternative name until a misclassification occurs.", "text_before_citation": ["This section covers some prominent attacks against models that work on source code.", "The techniques from these and other works have limited applicability to binary similarity, as their perturbations may not survive compilation (e.g., variable renaming) or result in marginal differences in compiled code (e.g., turning a while-loop into a for-loop)."], "text_after_citation": ["The attack against plagiarism detection from #OTHEREFR uses genetic programming to augment a program with code lines picked from a pool and validated for program equivalence by checking that an optimizer compiler removes them.", "The attack against clone detection from #OTHEREFR combines several semantics-preserving transformations of source code using different optimization heuristic strategies."], "citing_paper_content": {"title": "Adversarial Attacks Against Binary Similarity Systems", "abstract": "In recent years, binary analysis gained traction as a fundamental approach to inspect software and guarantee its security. Due to the exponential increase of devices running software, much research is now moving towards new autonomous solutions based on deep learning models, as they have been showing state-of-the-art performances in solving binary analysis problems. One of the hot topics in this context is binary similarity, which consists in determining if two functions in assembly code are compiled from the same source code. However, it is unclear how deep learning models for binary similarity behave in an adversarial context. In this paper, we study the resilience of binary similarity models against adversarial examples, showing that they are susceptible to both targeted and untargeted attacks (w.r.t. similarity goals) performed by black-box and white-box attackers. In more detail, we extensively test three current state-of-the-art solutions for binary similarity against two black-box greedy attacks, including a new technique that we call Spatial Greedy, and one white-box attack in which we repurpose a gradient-guided strategy used in attacks to image classifiers."}, "cited_paper_content": {"title": "Adversarial Examples For Models Of Code", "abstract": "Neural models of code have shown impressive performance for tasks such as predicting method names and identifying certain kinds of bugs. In this paper, we show that these models are vulnerable to adversarial examples, and introduce a novel approach for attacking trained models of code with adversarial examples. The main idea is to force a given trained model to make an incorrect prediction as specified by the adversary by introducing small perturbations that do not change the program's semantics. To find such perturbations, we present a new technique for Discrete Adversarial Manipulation of Programs (DAMP). DAMP works by deriving the desired prediction with respect to the model's inputs while holding the model weights constant and following the gradients to slightly modify the code. ::: To defend a model against such attacks, we propose placing a defensive model (Anti-DAMP) in front of it. Anti-DAMP detects unlikely mutations and masks them before feeding the input to the downstream model. ::: We show that our DAMP attack is effective across three neural architectures: code2vec, GGNN, and GNN-FiLM, in both Java and C#. We show that DAMP has up to 89% success rate in changing a prediction to the adversary's choice (\"targeted attack\"), and a success rate of up to 94% in changing a given prediction to any incorrect prediction (\"non-targeted attack\"). By using Anti-DAMP, the success rate of the attack drops drastically for both targeted and non-targeted attacks, with a minor penalty of 2% relative degradation in accuracy while not performing under attack."}, "keywords": ["general white-box attack"], "citation_intent": "background"}
{"citing_id": "2303.07645v1", "cited_id": "2003.02583", "section_title": "Case 3.2:", "citation": "Our main contribution here is to show that a nice ordering for the disks exists to make the co-2-subdivision approach to work, which at the same time improves the inapproximability factor of #REFR .", "text_before_citation": ["We now create a rectangle R b to represent b.", "We place the top-left corner of R b near the intersection point of the boundaries of the disks for t i , t i+1 such that R b does not intersect these disks but intersects all other type-t disks.", "We place the bottom-right corner of R b near the circular segment determined by the disk for s j on I t such that R b does not intersect the disk for s j but intersects all other type-s disks. We refer to #OTHEREFR for a formal reduction. The reason that Bonnet et al.", "#OTHEREFR thought such a reduction with a co-2-subdivision approach is unlikely is that without the pair-ordered labeling, the rectangle for R b may require to avoid three arbitrary disks.", "However, such realization may not be possible for all the rectangles if the disks are not carefully ordered."], "text_after_citation": [], "citing_paper_content": {"title": "Finding A Maximum Clique In A Disk Graph", "abstract": "A disk graph is an intersection graph of disks in the Euclidean plane, where the disks correspond to the vertices of the graph and a pair of vertices are adjacent if and only if their corresponding disks intersect. The problem of determining the time complexity of computing a maximum clique in a disk graph is a long-standing open question that has been very well studied in the literature. The problem is known to be open even when the radii of all the disks are in the interval [1, (1 + \u03b5)], where \u03b5 > 0. If all the disks are unit disks then there exists an O(n 3 log n)-time algorithm to compute a maximum clique, which is the best-known running time for over a decade. Although the problem of computing a maximum clique in a disk graph remains open, it is known to be APX-hard for the intersection graphs of many other convex objects such as intersection graphs of ellipses, triangles, and a combination of unit disks and axis-parallel rectangles. Here we obtain the following results.-We give an algorithm to compute a maximum clique in a unit disk graph in O(n 2.5 log n)-time, which improves the previously best known running time of O(n 3 log n) [Eppstein '09].-We extend a widely used 'co-2-subdivision approach' to prove that computing a maximum clique in a combination of unit disks and axis-parallel rectangles is NP-hard to approximate within 4448/4449 \u2248 0.9997. The use of a 'co-2-subdivision approach' was previously thought to be unlikely in this setting [Bonnet et al. '20]. Our result improves the previously known inapproximability factor of 7633010347/7633010348 \u2248 0.9999.-We show that the parameter minimum lens width of the disk arrangement may be used to make progress in the case when disk radii are in [1, (1 + \u03b5)]. For example, if the minimum lens width is at least 0.265 and \u03b5 \u2264 0.0001, which still allows for non-Helly triples in the arrangement, then one can find a maximum clique in polynomial time."}, "cited_paper_content": {"title": "Maximum Clique In Disk-Like Intersection Graphs", "abstract": "We study the complexity of Maximum Clique in intersection graphs of convex objects in the plane. On the algorithmic side, we extend the polynomial-time algorithm for unit disks [Clark '90, Raghavan and Spinrad '03] to translates of any fixed convex set. We also generalize the efficient polynomial-time approximation scheme (EPTAS) and subexponential algorithm for disks [Bonnet et al. '18, Bonamy et al. '18] to homothets of a fixed centrally symmetric convex set. The main open question on that topic is the complexity of Maximum Clique in disk graphs. It is not known whether this problem is NP-hard. We observe that, so far, all the hardness proofs for Maximum Clique in intersection graph classes $\\mathcal I$ follow the same road. They show that, for every graph $G$ of a large-enough class $\\mathcal C$, the complement of an even subdivision of $G$ belongs to the intersection class $\\mathcal I$. Then they conclude invoking the hardness of Maximum Independent Set on the class $\\mathcal C$, and the fact that the even subdivision preserves that hardness. However there is a strong evidence that this approach cannot work for disk graphs [Bonnet et al. '18]. We suggest a new approach, based on a problem that we dub Max Interval Permutation Avoidance, which we prove unlikely to have a subexponential-time approximation scheme. We transfer that hardness to Maximum Clique in intersection graphs of objects which can be either half-planes (or unit disks) or axis-parallel rectangles. That problem is not amenable to the previous approach. We hope that a scaled down (merely NP-hard) variant of Max Interval Permutation Avoidance could help making progress on the disk case, for instance by showing the NP-hardness for (convex) pseudo-disks."}, "keywords": ["disks"], "citation_intent": "background"}
{"citing_id": "2303.00400v1", "cited_id": "1912.04696", "section_title": "Conclusion & Future Work", "citation": "One exception to this is the popularity lift metric in the case of music recommendations; however, this result is in line with our previous work #REFR , in which repeat consumption settings have been studied.", "text_before_citation": ["In this paper, we have studied the interconnection between accuracy, miscalibration, and popularity bias for different user groups in three different domains.", "Here, we measured popularity bias in terms of popularity lift, a metric that compares the popularity of items in recommendation lists to the popularity of items in user profiles.", "Additionally, we investigated miscalibration, a metric that compares the genre spectrums in user profiles with the ones in recommendation lists.", "We find that, in general, the inconsistency of recommendations in terms of miscalibration and popularity lift is aligned with lower accuracy performance."], "text_after_citation": ["Additionally, we find that different genres contribute differently to miscalibration and popularity lift.", "That finding is particularly pronounced in the case of anime recommendations for LowPop users and for genres that are unpopular among other user groups.", "Another contribution of our work is that we publicly share our datasets and source code investigated in this study with the research community.", "Limitations & future work.", "One limitation of our work is that we have focused solely on datasets from the multimedia/entertainment domains, namely music, movies, and animes."], "citing_paper_content": {"title": "A Study On Accuracy, Miscalibration, And Popularity Bias In Recommendations", "abstract": "Recent research has suggested different metrics to measure the inconsistency of recommendation performance, including the accuracy difference between user groups, miscalibration, and popularity lift. However, a study that relates miscalibration and popularity lift to recommendation accuracy across different user groups is still missing. Additionally, it is unclear if particular genres contribute to the emergence of inconsistency in recommendation performance across user groups. In this paper, we present an analysis of these three aspects of five well-known recommendation algorithms for user groups that differ in their preference for popular content. Additionally, we study how different genres affect the inconsistency of recommendation performance, and how this is aligned with the popularity of the genres. Using data from Last.fm, MovieLens, and MyAnimeList, we present two key findings. First, we find that users with little interest in popular content receive the worst recommendation accuracy and that this is aligned with miscalibration and popularity lift. Second, our experiments show that particular genres contribute to a different extent to the inconsistency of recommendation performance, especially in terms of miscalibration in the case of the MyAnimeList dataset."}, "cited_paper_content": {"title": "The Unfairness Of Popularity Bias In Music Recommendation: A Reproducibility Study", "abstract": "Research has shown that recommender systems are typically biased towards popular items, which leads to less popular items being underrepresented in recommendations. The recent work of Abdollahpouri et al. in the context of movie recommendations has shown that this popularity bias leads to unfair treatment of both long-tail items as well as users with little interest in popular items. In this paper, we reproduce the analyses of Abdollahpouri et al. in the context of music recommendation. Specifically, we investigate three user groups from the this http URL music platform that are categorized based on how much their listening preferences deviate from the most popular music among all this http URL users in the dataset: (i) low-mainstream users, (ii) medium-mainstream users, and (iii) high-mainstream users. In line with Abdollahpouri et al., we find that state-of-the-art recommendation algorithms favor popular items also in the music domain. However, their proposed Group Average Popularity metric yields different results for this http URL than for the movie domain, presumably due to the larger number of available items (i.e., music artists) in the this http URL dataset we use. Finally, we compare the accuracy results of the recommendation algorithms for the three user groups and find that the low-mainstreaminess group significantly receives the worst recommendations."}, "keywords": ["music recommendations"], "citation_intent": "result"}
{"citing_id": "2304.08789v2", "cited_id": "1908.03418", "section_title": "B. Integrated Sensing And Communication (Isac)", "citation": "To prevent RX saturation in mono-static shared-antenna BS deployments and reduce the inherent masking effect of the direct leakage sidelobes, efficient analog and digital SIC solutions, tailored to the orthogonal frequency division multiplexing (OFDM) radar use, were introduced in #REFR .", "text_before_citation": ["In-band FD is being considered as a key enabling technology for ISAC applications due to its STAR capability, with promising theoretical results demonstrating advantages of joint design #OTHEREFR .", "State-of-the-art: In #OTHEREFR , #OTHEREFR , single-antenna FD systems were considered for ISAC operation, where both communication and radar waveforms were optimized for sensing performance.", "Frequency-domain radar processing building on the' LTE or the 5G New Radio (NR) time-frequency resource grid, and complemented with interpolation to account for missing samples due to the null subcarriers within the transmit waveform passband, was presented in #OTHEREFR .", "It was shown that the large channel bandwidths supported by 5G NR provide good sensing performance already at below 6 GHz frequencies.", "With 100 MHz channel bandwidth, range estimation accuracy in the order of 1 m and target detection probability exceeding 90% were shown to be feasible at signalto-noise ratio (SNR) values lower than \u221230 dB."], "text_after_citation": ["The various provided measurements verified successful sensing of static and moving targets, such as drones and cars, while evidencing measured TX/RX isolation at the FD node of approximately 100 dB.", "Apart from the typical optimization of the UL and DL sum rate, FD radios have been considered to continuously update the channel state information at the transmitter (CSIT), which is required to compute the DL precoding matrix in multiuser (massive) MIMO systems, as schematically shown in Fig. 9 .", "By design principle, FD operation leverages channel reciprocity for open-loop UL training to estimate the DL channels. However, the UL transmission of training might Fig. 9 .", "FD-enabled simultaneous multi-user DL data transmission and channel estimation in the UL, leveraging TDD. create interference at the DL receiving mobile nodes.", "Most predominantly, the DL transmission contaminates the reception of the training symbols in the UL."], "citing_paper_content": {"title": "Full-Duplex Wireless For 6G: Progress Brings New Opportunities And Challenges", "abstract": "The use of in-band full-duplex (FD) enables nodes to simultaneously transmit and receive on the same frequency band, which challenges the traditional assumption in wireless network design. The full-duplex capability enhances spectral efficiency and decreases latency, which are two key drivers pushing the performance expectations of next-generation mobile networks. In less than ten years, in-band FD has advanced from being demonstrated in research labs to being implemented in standards, presenting new opportunities to utilize its foundational concepts. Some of the most significant opportunities include using FD to enable wireless networks to sense the physical environment, integrate sensing and communication applications, develop integrated access and backhaul solutions, and work with smart signal propagation environments powered by reconfigurable intelligent surfaces. However, these new opportunities also come with new challenges for large-scale commercial deployment of FD technology, such as managing self-interference, combating cross-link interference in multi-cell networks, and coexistence of dynamic time division duplex, subband FD and FD networks."}, "cited_paper_content": {"title": "Full-Duplex Ofdm Radar With Lte And 5G Nr Waveforms: Challenges, Solutions, And Measurements", "abstract": "This article studies the processing principles, implementation challenges, and performance of orthogonal frequency-division multiplexing (OFDM)-based radars, with particular focus on the fourth-generation Long-Term Evolution (LTE) and fifth-generation (5G) New Radio (NR) mobile networks\u2019 base stations and their utilization for radar/sensing purposes. First, we address the problem stemming from the unused subcarriers within the LTE and NR transmit signal passbands and their impact on frequency-domain radar processing. In particular, we formulate and adopt a computationally efficient interpolation approach to mitigate the effects of such empty subcarriers in the radar processing. We evaluate the target detection and the corresponding range and velocity estimation performance through computer simulations and show that high-quality target detection as well as high-precision range and velocity estimation can be achieved. In particular, 5G NR waveforms, through their impressive channel bandwidths and configurable subcarrier spacing, are shown to provide very good radar/sensing performance. Then, a fundamental implementation challenge of transmitter\u2013receiver (TX\u2013RX) isolation in OFDM radars is addressed, with specific emphasis on shared-antenna cases, where the TX\u2013RX isolation challenges are the largest. It is confirmed that from the OFDM radar processing perspective, limited TX\u2013RX isolation is primarily a concern in the detection of static targets, while moving targets are inherently more robust to transmitter self-interference (SI). Properly tailored analog/RF and digital SI cancellation solutions for OFDM radars are also described and implemented and shown through RF measurements to be key technical ingredients for practical deployments, particularly from static and slowly moving targets\u2019 point of view."}, "keywords": ["mono-static shared-antenna BS", "OFDM) radar"], "citation_intent": "method"}
{"citing_id": "2303.17870v1", "cited_id": "1406.2661", "section_title": "Introduction", "citation": "In contrast to generative adversarial network (GAN) based methods #REFR , diffusion models offer greater stability during training without relying on the complex adversarial training process, and enable precise control over generation quality and diversity during the diffusion process.", "text_before_citation": ["The field of text-to-image synthesis has greatly benefited from recent advancements in diffusion models."], "text_after_citation": ["By leveraging the semantics of textual input, large-scale text-guided diffusion models such as DALLE-2 #OTHEREFR , Imagen #OTHEREFR , and Stable Diffusion #OTHEREFR have revolutionized the way we generate high-quality images.", "These models are capable of following user instructions, generating customized high-fidelity images with specified content, style, and attributes.", "Recently, several studies #OTHEREFR have focused on gaining more fine-grained control over the generation process by leveraging auxiliary information such as edge maps, segmentation maps, color maps, etc.", "Users can specify desired attributes to guide and restrict the generation process by including such additional input, resulting in more precise control over the generated images.", "Despite some limitations in image diversity due to auxiliary input dependency, these methods have produced impressive results in generating high-quality images that meet specific requirements."], "citing_paper_content": {"title": "Glyphdraw: Learning To Draw Chinese Characters In Image Synthesis Models Coherently", "abstract": "Recent breakthroughs in the field of language-guided image generation have yielded impressive achievements, enabling the creation of high-quality and diverse images based on user instructions. Although the synthesis performance is fascinating, one significant limitation of current image generation models is their insufficient ability to generate coherent text within images, particularly for complex glyph structures like Chinese characters. To address this problem, we introduce GlyphDraw, a general learning framework aiming at endowing image generation models with the capacity to generate images embedded with coherent text. To the best of our knowledge, this is the first work in the field of image synthesis to address the generation of Chinese characters. We first sophisticatedly design the image-text dataset's construction strategy, then build our model specifically on a diffusionbased image generator and carefully modify the network structure to allow the model to learn drawing Chinese characters with the help of glyph and position information. Furthermore, we maintain the model's open-domain image synthesis capability by preventing catastrophic forgetting by using a variety of training techniques. Extensive qualitative and quantitative experiments demonstrate that our method not only produces accurate Chinese characters as in prompts, but also naturally blends the generated text into the background. Please refer to our project page."}, "cited_paper_content": {"title": "Generative Adversarial Networks", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["generative adversarial network"], "citation_intent": "background"}
{"citing_id": "2303.16482v1", "cited_id": "1912.04645", "section_title": "Point-Based Rendering", "citation": "NPCR #REFR represents point features by 3D Convolution Neural Network (CNN), and converts the 3D representation to multi-plane images.", "text_before_citation": ["Traditional Point Rendering [38, #OTHEREFR is based on computer graphics, which generates images from point clouds by simulating the physical process of imaging, considering geometry #OTHEREFR , material #OTHEREFR , BRDF #OTHEREFR , and lighting #OTHEREFR . The rendering pipeline is general for arbitrary scenes.", "But it cannot fill missing points and thus generate vacant pixels.", "In the deep learning era, neural-based point renderers #OTHEREFR made great progress in generating images from point clouds.", "They first extract multi-scale projected features from point clouds and then use a decoder to generate images.", "NPBG #OTHEREFR augments each point by a neural descriptor to encode local geometry and appearance."], "text_after_citation": ["Different from these methods, our Point2Pix combines a more discriminative point encoder with the renderer pipeline of NeRF, thus achieving better performance."], "citing_paper_content": {"title": "Point2Pix: Photo-Realistic Point Cloud Rendering Via Neural Radiance Fields", "abstract": "Synthesizing photo-realistic images from a point cloud is challenging because of the sparsity of point cloud representation. Recent Neural Radiance Fields and extensions are proposed to synthesize realistic images from 2D input. In this paper, we present Point2Pix as a novel point renderer to link the 3D sparse point clouds with 2D dense image pixels. Taking advantage of the point cloud 3D prior and NeRF rendering pipeline, our method can synthesize high-quality images from colored point clouds, generally for novel indoor scenes. To improve the efficiency of ray sampling, we propose point-guided sampling, which focuses on valid samples. Also, we present Point Encoding to build Multiscale Radiance Fields that provide discriminative 3D point features. Finally, we propose Fusion Encoding to efficiently synthesize high-quality images. Extensive experiments on the ScanNet and ArkitScenes datasets demonstrate the effectiveness and generalization."}, "cited_paper_content": {"title": "Neural Point Cloud Rendering Via Multi-Plane Projection", "abstract": "We present a new deep point cloud rendering pipeline through multi-plane projections. The input to the network is the raw point cloud of a scene and the output are image or image sequences from a novel view or along a novel camera trajectory. Unlike previous approaches that directly project features from 3D points onto 2D image domain, we propose to project these features into a layered volume of camera frustum. In this way, the visibility of 3D points can be automatically learnt by the network, such that ghosting effects due to false visibility check as well as occlusions caused by noise interferences are both avoided successfully. Next, the 3D feature volume is fed into a 3D CNN to produce multiple layers of images w.r.t. the space division in the depth directions. The layered images are then blended based on learned weights to produce the final rendering results. Experiments show that our network produces more stable renderings compared to previous methods, especially near the object boundaries. Moreover, our pipeline is robust to noisy and relatively sparse point cloud for a variety of challenging scenes."}, "keywords": ["point features", "3D Convolution Neural"], "citation_intent": "method"}
{"citing_id": "2304.12437v1", "cited_id": "1406.2661", "section_title": "Variational Autoencoder (Vae)", "citation": "A thorough description of DNNs and their training can be found in #REFR , in the work herein the DNNs utilised only made use of fully connected layers.", "text_before_citation": ["where J denotes the dimension of the latent space and", "Z i,j = \u00b5 \u03b8 (X i ) + \u03b7 l \u03c3 \u03b8 (X i ).", "The number of samples to take from the latent space to evaluate the expectation can even be 1 as in the original formulation of Kingma et al. #OTHEREFR .", "As mentioned above, in the VAE model the encoder and decoder functions are approximated using deep neural networks (DNNs).", "DNNs are a very widely used class of models that exploit multiple neural network layers applied one after another in order to approximate very complex functions more efficiently than shallow networks #OTHEREFR ."], "text_after_citation": ["In fully connected layers, the transform performed by each layer consists of the matrix multiplication of the input vector with a trainable weights matrix and the addition of a trainable bias vector.", "A nonlinear activation function, often a tanh or sigmoid is then applied element-wise to the output of this operation.", "This results in a very flexible and powerful model for learning general nonlinear relations (32)."], "citing_paper_content": {"title": "Vprom: A Novel Variational Autoencoder-Boosted Reduced Order Model For The Treatment Of Parametric Dependencies In Nonlinear Systems A Preprint", "abstract": "Reduced Order Models (ROMs) are of considerable importance in many areas of engineering in which computational time presents difficulties. Established approaches employ projection-based reduction such as Proper Orthogonal Decomposition, however, such methods can become inefficient or fail in the case of parameteric or strongly nonlinear models. Such limitations are usually tackled via a library of local reduction bases each of which being valid for a given parameter vector. The success of such methods, however, is strongly reliant upon the method used to relate the parameter vectors to the local bases, this is typically achieved using clustering or interpolation methods. We propose the replacement of these methods with a Variational Autoencoder (VAE) to be used as a generative model which can infer the local basis corresponding to a given parameter vector in a probabilistic manner. The resulting VAE-boosted parametric ROM VpROM still retains the physical insights of a projection-based method but also allows for better treatment of problems where model dependencies or excitation traits cause the dynamic behavior to span multiple response regimes. Moreover, the probabilistic treatment of the VAE representation allows for uncertainty quantification on the reduction bases which may then be propagated to the ROM response. The performance of the proposed approach is validated on an open-source simulation benchmark featuring hysteresis and multi-parametric dependencies, and on a large-scale wind turbine tower characterised by nonlinear material behavior and model uncertainty."}, "cited_paper_content": {"title": "Generative Adversarial Nets", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to \u00bd everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["fully connected layers", "DNNs"], "citation_intent": "method"}
{"citing_id": "2303.10383v1", "cited_id": "1912.01703", "section_title": "Implementation Details", "citation": "Our model is implemented based on Pytorch #REFR and trained on a PC with an RTX 3090 GPU.", "text_before_citation": [], "text_after_citation": ["The input sources are all resized to 384 \u00d7 384 and all the three stages use the mini-batch of size 4 for training.", "First, we use some RGB-D saliency datasets #OTHEREFR to train the static predictor network.", "Once the first stage of training is finished, we adopt the DAVIS 16 training set to train the multi-source fusion network.", "In the process, the parameters of the static predictor are frozen, and we just train the moving object predictor.", "We use a total of 8, 398 labeled samples including video data (2, 000 + frames) and RGB-D SOD data (6, 300 + images) in the first and second phases. The depth map is obtained by the depth camera."], "citing_paper_content": {"title": "Adaptive Multi-Source Predictor For Zero-Shot Video Object Segmentation", "abstract": "Both static and moving objects usually exist in real-life videos. Most video object segmentation methods only focus on exacting and exploiting motion cues to perceive moving objects. Once faced with static objects frames, moving object predictors may predict failed results caused by uncertain motion information, such as low-quality optical flow maps. Besides, many sources such as RGB, depth, optical flow and static saliency can provide useful information about the objects. However, existing approaches only utilize the RGB or RGB and optical flow. In this paper, we propose a novel adaptive multi-source predictor for zero-shot video object segmentation. In the static object predictor, the RGB source is converted to depth and static saliency sources, simultaneously. In the moving object predictor, we propose the multi-source fusion structure. First, the spatial importance of each source is highlighted with the help of the interoceptive spatial attention module (ISAM). Second, the motion-enhanced module (MEM) is designed to generate pure foreground motion attention for improving both static and moving features used in the decoder. Furthermore, we design a feature purification module (FPM) to filter the inter-source incompatible features. By the ISAM, MEM and FPM, the multi-source features are effectively fused. In addition, we put forward an adaptive predictor fusion network (APF) to evaluate the quality of optical flow and fuse the predictions from the static object predictor and the moving object predictor in order to prevent over-reliance on the failed results caused by low-quality optical flow maps. Experiments show that the proposed model outperforms the state-of-the-art methods on three challenging ZVOS benchmarks. And, the static object predictor can precisely predicts a high-quality depth map and static saliency map at the same time."}, "cited_paper_content": {"title": "Pytorch: An Imperative Style, High-Performance Deep Learning Library", "abstract": "Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. ::: In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. ::: We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks."}, "keywords": ["Pytorch"], "citation_intent": "method"}
{"citing_id": "2303.17671v1", "cited_id": "1910.12478", "section_title": "Conclusion And Future Work", "citation": "Extending the results of #REFR to the continuous setting, or proving the uniform convergence in partition size of the infinite-width limit are both promising directions.", "text_before_citation": ["We showed that in the infinite-width-then-depth limit, these converge weakly to GPs indexed on path space endowed with neural signature kernels satisfying certain (possibly non-linear) PDEs varying according to the choice of activation function \u03d5.", "In the special case where \u03d5 is the identity, we showed that the equation reduces to a linear PDE and the limiting kernel agrees with the signature kernel.", "In this setting, we also showed that the width-depth limits commute.", "Finally, we showed that in the infinite-depth regime, finite-width controlled ResNets converge in distribution to Neural CDEs with random vector fields which are either time-independent and Gaussian, if the system is homogeneous, or behave like a matrix-valued Brownian motion, if the system is inhomogeneous.", "An interesting question that remains to be answered is that of commutativity of the limits in the general setting."], "text_after_citation": ["We believe that a rigorous investigation of the functional analytic properties of the reproducing kernel Hilbert spaces (RKHSs) associated to the new family of neural signature kernels is also a compelling future research direction.", "In particular, it would allow to build an understanding of the expressivity and generalization properties of these kernels.", "Last but not least, establishing expressions and analyzing the associated Neural Tangent Kernels (NTK) #OTHEREFR would provide quantitative insights on the training mechanism of Neural CDEs by gradient descent."], "citing_paper_content": {"title": "Neural Signature Kernels As Infinite-Width-Depth-Limits Of Controlled Resnets", "abstract": "Motivated by the paradigm of reservoir computing, we consider randomly initialized controlled ResNets defined as Euler-discretizations of neural controlled differential equations (Neural CDEs). We show that in the infinite-width-then-depth limit and under proper scaling, these architectures converge weakly to Gaussian processes indexed on some spaces of continuous paths and with kernels satisfying certain partial differential equations (PDEs) varying according to the choice of activation function \u03d5. In the special case where \u03d5 is the identity, we show that the equation reduces to a linear PDE and the limiting kernel agrees with the signature kernel of Salvi et al. (2021a). In this setting, we also show that the width-depth limits commute. We name this new family of limiting kernels neural signature kernels. Finally, we show that in the infinite-depth regime, finite-width controlled ResNets converge in distribution to Neural CDEs with random vector fields which, depending on whether the weights are shared across layers, are either time-independent and Gaussian or behave like a matrix-valued Brownian motion."}, "cited_paper_content": {"title": "Tensor Programs I: Wide Feedforward Or Recurrent Neural Networks Of Any Architecture Are Gaussian Processes", "abstract": "Wide neural networks with random weights and biases are Gaussian processes, as observed by Neal (1995) for shallow networks, and more recently by Lee et al.~(2018) and Matthews et al.~(2018) for deep fully-connected networks, as well as by Novak et al.~(2019) and Garriga-Alonso et al.~(2019) for deep convolutional networks. We show that this Neural Network-Gaussian Process correspondence surprisingly extends to all modern feedforward or recurrent neural networks composed of multilayer perceptron, RNNs (e.g. LSTMs, GRUs), (nD or graph) convolution, pooling, skip connection, attention, batch normalization, and/or layer normalization. More generally, we introduce a language for expressing neural network computations, and our result encompasses all such expressible neural networks. This work serves as a tutorial on the \\emph{tensor programs} technique formulated in Yang (2019) and elucidates the Gaussian Process results obtained there. We provide open-source implementations of the Gaussian Process kernels of simple RNN, GRU, transformer, and batchnorm+ReLU network at github.com/thegregyang/GP4A. Please see our arxiv version for the complete and up-to-date version of this paper."}, "keywords": ["infinite-width limit", "directions"], "citation_intent": "background"}
{"citing_id": "2303.14096v1", "cited_id": "1608.03983", "section_title": ". Training Details", "citation": "We set a weight decay of 10 \u22124 , and use the cosine learning rate scheduling #REFR from the initial learning rate of 0.1.", "text_before_citation": ["Unless otherwise noted, we train each model for 200K updates for CIFAR-10 models, and 1M updates for ImageNet models.", "For training AENIB models, we use \u03b1 = 10.0, \u03b2 = 0.0001 unless otherwise noted.", "We use different training configurations depending on the encoder architecture, i.e., whether is it ResNet or ViT: (a) For ResNet-based models, we train the encoder part (f ) via stochastic gradient descent (SGD) with batch size of 64 using Nesterov momentum of weight 0.9 without dampening."], "text_after_citation": ["For the remainder parts of our AENIB architecture, e.g., the decoder g and discriminator MLPs, on the other hand, we follow the training practices of GAN instead: specifically, we use Adam #OTHEREFR with (\u03b1, \u03b2 1 , \u03b2 2 ) = (0.0002, 0.5, 0.999), following the hyperparameter practices explored by #OTHEREFR .", "(b) For ViT-based models, on the other hand, we train both (transformer-based) encoder and decoder models via AdamW #OTHEREFR with a weight decay of 10 \u22124 , using batch size 128 and (\u03b1, \u03b2 1 , \u03b2 2 ) = (0.0002, 0.9, 0.999) with the cosine learning rate scheduling #OTHEREFR .", "We use 2K and 100K steps of a linear warm-up phase in learning rate for CIFAR and ImageNet models, respectively.", "Overall, we observe that a stable training of ViT (even for CIFAR-10) requires much stronger regularization compared to ResNets, otherwise they often significantly suffer from overfitting.", "In this respect, we apply the regularization practices those are now widely used for ViTs on ImageNet, namely mixup #OTHEREFR , CutMix #OTHEREFR , and RandAugment #OTHEREFR , following those established in #OTHEREFR ."], "citing_paper_content": {"title": "Enhancing Multiple Reliability Measures Via Nuisance-Extended Information Bottleneck", "abstract": "In practical scenarios where training data is limited, many predictive signals in the data can be rather from some biases in data acquisition (i.e., less generalizable), so that one cannot prevent a model from co-adapting on such (socalled) \"shortcut\" signals: this makes the model fragile in various distribution shifts. To bypass such failure modes, we consider an adversarial threat model under a mutual information constraint to cover a wider class of perturbations in training. This motivates us to extend the standard information bottleneck to additionally model the nuisance information. We propose an autoencoder-based training to implement the objective, as well as practical encoder designs to facilitate the proposed hybrid discriminative-generative training concerning both convolutional-and Transformerbased architectures. Our experimental results show that the proposed scheme improves robustness of learned representations (remarkably without using any domain-specific knowledge), with respect to multiple challenging reliability measures. For example, our model could advance the stateof-the-art on a recent challenging OBJECTS benchmark in novelty detection by 78.4% \u2192 87.2% in AUROC, while simultaneously enjoying improved corruption, background and (certified) adversarial robustness. Code is available at https://github.com/jh-jeong/nuisance_ib."}, "cited_paper_content": {"title": "Sgdr: Stochastic Gradient Descent With Warm Restarts", "abstract": "Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets, where we demonstrate new state-of-the-art results at 3.14% and 16.21%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at https://github.com/loshchil/SGDR"}, "keywords": ["cosine learning rate", "initial learning rate"], "citation_intent": "method"}
{"citing_id": "2303.14009v1", "cited_id": "1206.6389", "section_title": "Backdoor Attacks On Gnns", "citation": "Traditional data poisoning attacks corrupt training samples to downgrade the overall performance of ML models #REFR .", "text_before_citation": ["Backdoor attacks are special types of data poisoning attacks on machine learning (ML) systems."], "text_after_citation": ["However, backdoor attacks maintain original performance until the model is provided with an input sample containing a \"backdoor trigger.\" Such a backdoor trigger causes a pre-determined output, y t , beneficial to an adversary #OTHEREFR .", "An adversary can deploy backdoor attacks by manipulating the training data and the corresponding labels.", "For the case of GNNs, where the input samples are graphs, existing backdoor attacks inject triggers in the form of subgraphs g t . Fig.", "3 illustrates the flow of a subgraphbased backdoor attack against GNNs #OTHEREFR .", "First, a backdoor trigger and a target label y t (e.g., class 1, i.e., y t = 1) are determined ( Fig. 3 1 ) ."], "citing_paper_content": {"title": "Poisonedgnn: Backdoor Attack On Graph Neural Networks-Based Hardware Security Systems", "abstract": "Graph neural networks (GNNs) have shown great success in detecting intellectual property (IP) piracy and hardware Trojans (HTs). However, the machine learning community has demonstrated that GNNs are susceptible to data poisoning attacks, which result in GNNs performing abnormally on graphs with pre-defined backdoor triggers (realized using crafted subgraphs). Thus, it is imperative to ensure that the adoption of GNNs should not introduce security vulnerabilities in critical security frameworks. Existing backdoor attacks on GNNs generate random subgraphs with specific sizes/densities to act as backdoor triggers. However, for Boolean circuits, backdoor triggers cannot be randomized since the added structures should not affect the functionality of a design. We explore this threat and develop PoisonedGNN as the first backdoor attack on GNNs in the context of hardware design. We design and inject backdoor triggers into the register-transfer-or the gate-level representation of a given design without affecting the functionality to evade some GNN-based detection procedures. To demonstrate the effectiveness of PoisonedGNN, we consider two case studies: (i) Hiding HTs and (ii) IP piracy. Our experiments on TrustHub datasets demonstrate that PoisonedGNN can hide HTs and IP piracy from advanced GNN-based detection platforms with an attack success rate of up to 100%."}, "cited_paper_content": {"title": "Poisoning Attacks Against Support Vector Machines", "abstract": "We investigate a family of poisoning attacks against Support Vector Machines (SVM). Such attacks inject specially crafted training data that increases the SVM's test error. Central to the motivation for these attacks is the fact that most learning algorithms assume that their training data comes from a natural or well-behaved distribution. However, this assumption does not generally hold in security-sensitive settings. As we demonstrate, an intelligent adversary can, to some extent, predict the change of the SVM's decision function due to malicious input and use this ability to construct malicious data. ::: ::: The proposed attack uses a gradient ascent strategy in which the gradient is computed based on properties of the SVM's optimal solution. This method can be kernelized and enables the attack to be constructed in the input space even for non-linear kernels. We experimentally demonstrate that our gradient ascent procedure reliably identifies good local maxima of the non-convex validation error surface, which significantly increases the classifier's test error."}, "keywords": ["Traditional data poisoning"], "citation_intent": "background"}
{"citing_id": "2303.16094v1", "cited_id": "1903.11027", "section_title": "Dataset", "citation": "We evaluate the 3D detection performance on the widely used benchmark, nuScenes #REFR , a public dataset for autonomous driving.", "text_before_citation": [], "text_after_citation": ["It is collected from challenging urban scenes, consisting of 1000 annotated sequences.", "Among them, 700 sequences are used as the training phase, 150 as validation, and the remaining 150 as the test phase for online evaluation.", "The bounding boxes are labeled with not only categories but also some attributes like the velocity, scale, orientation and translation.", "Besides the point cloud data from LiDAR, it also provides the 360 \u2022 image modality from cameras and signal from the Radar sensor.", "In this paper, we only utilize the point cloud from LiDAR."], "citing_paper_content": {"title": "Link: Linear Kernel For Lidar-Based 3D Perception", "abstract": "Extending the success of 2D Large Kernel to 3D perception is challenging due to: 1. the cubically-increasing overhead in processing 3D data; 2. the optimization difficulties from data scarcity and sparsity. Previous work has taken the first step to scale up the kernel size from 3 \u00d7 3 \u00d7 3 to 7 \u00d7 7 \u00d7 7 by introducing block-shared weights. However, to reduce the feature variations within a block, it only employs modest block size and fails to achieve larger kernels like the 21 \u00d7 21 \u00d7 21. To address this issue, we propose a new method, called LinK, to achieve a wider-range perception receptive field in a convolution-like manner with two core designs. The first is to replace the static kernel matrix with a linear kernel generator, which adaptively provides weights only for non-empty voxels. The second is to reuse the precomputed aggregation results in the overlapped blocks to reduce computation complexity. The proposed method successfully enables each voxel to perceive context within a range of 21 \u00d7 21 \u00d7 21. Extensive experiments on two basic perception tasks, 3D object detection and 3D semantic segmentation, demonstrate the effectiveness of our method. Notably, we rank 1st on the public leaderboard of the 3D detection benchmark of nuScenes (LiDAR track), by simply incorporating a LinK-based backbone into the basic detector, CenterPoint. We also boost the strong segmentation baseline's mIoU with 2.7% in the SemanticKITTI test set."}, "cited_paper_content": {"title": "Nuscenes: A Multimodal Dataset For Autonomous Driving", "abstract": "Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology. Image-based benchmark datasets have driven development in computer vision tasks such as object detection, tracking and segmentation of agents in the environment. Most autonomous vehicles, however, carry a combination of cameras and range sensors such as lidar and radar. As machine learning based methods for detection and tracking become more prevalent, there is a need to train and evaluate such methods on datasets containing range sensor data along with images. In this work we present nuTonomy scenes (nuScenes), the first published dataset to carry the full autonomous vehicle sensor suite: 6 cameras, 5 radars and 1 lidar, all with full 360 degree field of view. nuScenes comprises 1000 scenes, each 20s long and fully annotated with 3D bounding boxes for 23 classes and 8 attributes. It has 7x as many annotations and 100x as many images as the pioneering KITTI dataset. We define novel 3D detection and tracking metrics. We also provide careful dataset analysis as well as baselines for lidar and image based detection and tracking. Data, development kit and more information are available online at this http URL."}, "keywords": ["3D detection performance", "autonomous driving"], "citation_intent": "method"}
{"citing_id": "2304.11072v1", "cited_id": "2001.02334", "section_title": "Performance Evaluation", "citation": "Our work was also compared with the results from \u00b5VulDeePecker #REFR , which evaluated 40 classes.", "text_before_citation": ["We prevented training a biased model by using the Focal Loss function during the model's training.", "We tested VulF dataset with our proposed RoBERTa-PFGCN and two other publicly available models, Devign and VELVET.", "Moreover, we also implemented \u00b5VulDeePecker #OTHEREFR ourselves based on their proposed architecture since their model is not publicly available.", "Our experimental results show that our model achieves the highest accuracy compared to Devign, VELVET, and \u00b5VulDeePecker.", "Table 4 shows that our model improved accuracy, precision, recall, and F1 score by 11.79%, 11.06%, 9.98%, and 11.39%, respectively, which is at least 11% higher than other models."], "text_after_citation": ["Table  4 shows that our model outperforms \u00b5VulDeePecker on their MVD dataset by almost 3.80% on F1 score.", "We provide an F1 score comparison for all 40 classes by our model against \u00b5VulDeePecker with the MVD dataset in Figure 4 .", "This bar chart shows a comparative analysis of the F1 score with our model vs.", "\u00b5VulDeePecker, and we 100  100  100  100  100  100  100  100  100  100  100  100  100   95   100   98  98  97  98  99  99.5  100   81   96   90   98   90   98   85   94  95  95  95  95   90   83   94  95  95  93   50   98   85   90   100   80   95  93  95   84   95  94  94  95  94   40 can clearly see our model consistently generates higher F1 scores for all the 40 CWE classes.", "Table 4 demonstrates our model's performance with two other datasets."], "citing_paper_content": {"title": "An Unbiased Transformer Source Code Learning With Semantic Vulnerability Graph", "abstract": "Over the years, open-source software systems have become prey to threat actors. Even highly-adopted software has been crippled by unforeseeable attacks, leaving millions of devices exposed. Even as open-source communities act quickly to patch the breach, code vulnerability screening should be an integral part of agile software development from the beginning. Unfortunately, current vulnerability screening techniques are ineffective at identifying novel vulnerabilities or providing developers with code vulnerability and classification. Furthermore, the datasets used for vulnerability learning often exhibit distribution shifts from the real-world testing distribution due to novel attack strategies deployed by adversaries and as a result, the machine learning model's performance may be hindered or biased. To address these issues, we propose a joint interpolated multitasked unbiased vulnerability classifier comprising a transformer \"RoBERTa\" and graph convolution neural network (GCN). We present a training process utilizing a semantic vulnerability graph (SVG) representation from source code, created by integrating edges from a sequential flow, control flow, and data flow, as well as a novel flow dubbed Poacher Flow (PF). Poacher flow edges reduce the gap between dynamic and static program analysis and handle complex long-range dependencies. Moreover, our approach reduces biases of classifiers regarding unbalanced datasets by integrating Focal Loss objective function along with SVG. Remarkably, experimental results show that our classifier outperforms state-of-the-art results on vulnerability detection with fewer false negatives and false positives. After testing our model across multiple datasets, it shows an improvement of at least 2.41% and 18.75% in the bestcase scenario. Evaluations using N-day program samples demonstrate that our proposed approach achieves a 93% accuracy and was able to detect 4, zero-day vulnerabilities from popular GitHub repositories. Our code and data are"}, "cited_paper_content": {"title": "$\\Mu$Vuldeepecker: A Deep Learning-Based System For Multiclass Vulnerability Detection", "abstract": "Fine-grained software vulnerability detection is an important and challenging problem. Ideally, a detection system (or detector) not only should be able to detect whether or not a program contains vulnerabilities, but also should be able to pinpoint the type of a vulnerability in question. Existing vulnerability detection methods based on deep learning can detect the presence of vulnerabilities (i.e., addressing the binary classification or detection problem), but cannot pinpoint types of vulnerabilities (i.e., incapable of addressing multiclass classification). In this paper, we propose the first deep learning-based system for multiclass vulnerability detection, dubbed $\\mu$VulDeePecker. The key insight underlying $\\mu$VulDeePecker is the concept of code attention, which can capture information that can help pinpoint types of vulnerabilities, even when the samples are small. For this purpose, we create a dataset from scratch and use it to evaluate the effectiveness of $\\mu$VulDeePecker. Experimental results show that $\\mu$VulDeePecker is effective for multiclass vulnerability detection and that accommodating control-dependence (other than data-dependence) can lead to higher detection capabilities."}, "keywords": ["40 classes"], "citation_intent": "result"}
{"citing_id": "2304.06815v1", "cited_id": "1909.09436", "section_title": "Dataset", "citation": "We use the CodeSearchNet #REFR dataset for our experiments: it is a carefully de-duplicated, multi-project dataset, which allows cross-project testing.", "text_before_citation": [], "text_after_citation": ["De-duplication is key: Code duplication in machine learning models can deceptively inflate performance metrics by as much as 100% when measured on duplicated code datasets, compared to de-duplicated datasets #OTHEREFR .", "It is part of the CodeXGLUE #OTHEREFR benchmark, which comprises 14 datasets for 10 software engineering tasks. Many models have been evaluated on this dataset.", "The CodeXGLUE dataset contains thousands of samples from six different programming languages (i.e., Java, Python, JavaScript, Ruby, Go, PHP).", "However, we did not use the entire test dataset; instead, we selected 250 samples uniformly at random from each language.", "Since the original dataset is cross-project and we sampled it uniformly, our subsample remains cross-project."], "citing_paper_content": {"title": "Improving Few-Shot Prompts With Relevant Static Analysis Products", "abstract": "Large Language Models (LLM) are a new class of computation engines, \"programmed\" via prompt engineering. Researchers are still learning how to best \"program\" these LLMs to help developers. We start with the intuition that developers tend to consciously and unconsciously have a collection of semantics facts in mind when working on coding tasks. Mostly these are shallow, simple facts arising from a quick read. For a function, examples of facts might include parameter and local variable names, return expressions, simple pre-and post-conditions, and basic control and data flow, etc. One might assume that the powerful multi-layer architecture of transformer-style LLMs makes them inherently capable of doing this simple level of \"code analysis\" and extracting such information, implicitly, while processing code: but are they, really? If they aren't, could explicitly adding this information help? Our goal here is to investigate this question, using the code summarization task and evaluate whether automatically augmenting an LLM's prompt with semantic facts explicitly, actually helps. Prior work shows that LLM performance on code summarization benefits from few-shot samples drawn either from the same-project or from examples found via information retrieval methods (such as BM25). While summarization performance has steadily increased since the early days, there is still room for improvement: LLM performance on code summarization still lags its performance on natural-language tasks like translation and text summarization. We find that adding semantic facts actually does help! This approach improves performance in several different settings suggested by prior work, including for two different Large Language Models."}, "cited_paper_content": {"title": "Codesearchnet Challenge: Evaluating The State Of Semantic Code Search", "abstract": "Semantic code search is the task of retrieving relevant code given a natural language query. While related to other information retrieval tasks, it requires bridging the gap between the language used in code (often abbreviated and highly technical) and natural language more suitable to describe vague concepts and ideas. ::: To enable evaluation of progress on code search, we are releasing the CodeSearchNet Corpus and are presenting the CodeSearchNet Challenge, which consists of 99 natural language queries with about 4k expert relevance annotations of likely results from CodeSearchNet Corpus. The corpus contains about 6 million functions from open-source code spanning six programming languages (Go, Java, JavaScript, PHP, Python, and Ruby). The CodeSearchNet Corpus also contains automatically generated query-like natural language for 2 million functions, obtained from mechanically scraping and preprocessing associated function documentation. In this article, we describe the methodology used to obtain the corpus and expert labels, as well as a number of simple baseline solutions for the task. ::: We hope that CodeSearchNet Challenge encourages researchers and practitioners to study this interesting task further and will host a competition and leaderboard to track the progress on the challenge. We are also keen on extending CodeSearchNet Challenge to more queries and programming languages in the future."}, "keywords": ["cross-project testing", "dataset"], "citation_intent": "method"}
{"citing_id": "2303.06697v1", "cited_id": "1810.04805", "section_title": "Related Works", "citation": "BERT #REFR can be seen as a development of DAE, which masks input tokens and trains the model to predict the missing content.", "text_before_citation": ["mmTransformer #OTHEREFR is designed to hierarchically aggregate the past trajectories, the road information, and the social interaction.", "For predicting multiagents future trajectories, AgentFormer #OTHEREFR and AutoBots #OTHEREFR have given solutions to model the time dimension and social dimension simultaneously.", "The enhancement of the encoder's ability to model information in both dimensions is an interesting and central focus of this work.", "Self-supervised Learning has shown significant success in natural language processing and computer vision fields recently, especially the autoencoding method.", "Denoising autoencoders (DAE) #OTHEREFR is a learning representation method that reconstructs original signals from corrupted inputs."], "text_after_citation": ["With the Masked Language Modeling (MLM) task proposed in BERT, many MLM variants #OTHEREFR are proposed to improve the performance of transformer pretraining.", "Similarly, in computer vision, the autoencoding method often focuses on different pretext tasks for pre- training #OTHEREFR .", "One of the most popular methods is MAE #OTHEREFR , which randomly masks input patches and trains the model to recover masked patches in pixel space.", "Continuous progress #OTHEREFR based on MAE has verified its effectiveness.", "Following the concept of MAE, our approach concentrate on utilizing MAE as a tool to pre-train model encoders with powerful feature extraction capability."], "citing_paper_content": {"title": "Traj-Mae: Masked Autoencoders For Trajectory Prediction", "abstract": "Trajectory prediction has been a crucial task in building a reliable autonomous driving system by anticipating possible dangers. One key issue is to generate consistent trajectory predictions without colliding. To overcome the challenge, we propose an efficient masked autoencoder for trajectory prediction (Traj-MAE) that better represents the complicated behaviors of agents in the driving environment. Specifically, our Traj-MAE employs diverse masking strategies to pre-train the trajectory encoder and map encoder, allowing for the capture of social and temporal information among agents while leveraging the effect of environment from multiple granularities. To address the catastrophic forgetting problem that arises when pre-training the network with multiple masking strategies, we introduce a continual pre-training framework, which can help Traj-MAE learn valuable and diverse information from various strategies efficiently. Our experimental results in both multi-agent and single-agent settings demonstrate that Traj-MAE achieves competitive results with state-of-the-art methods and significantly outperforms our baseline model. The code will be made publicly available upon publication."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["model", "BERT"], "citation_intent": "background"}
{"citing_id": "2304.12624v1", "cited_id": "1709.06158", "section_title": "Results On Occlusion Dataset", "citation": "To verify the robustness of our model to occlusion, we evaluated the models on 132 images with occlusion from the test split of the Matteport3D dataset #REFR .", "text_before_citation": [], "text_after_citation": ["In this experiment, occlusion was defined as a situation where one or more corners of the ground truth are not visible, as illustrated in Fig. 6 .", "The Pano S2D3D dataset #OTHEREFR was not included in this experiment because of its cuboid nature, which does not cause occlusion.", "Quantitative results on the occlusion dataset are presented in Tab. 3.", "We reported the best performance of our model, a 2.37% improvement in both 2D and 3D IoU compared with LGT-Net #OTHEREFR . In comparison to Tab.", "2, the difference in 2D and 3D IoU between our model and other models was greater on the occlusion dataset."], "citing_paper_content": {"title": "Shape-Net: Room Layout Estimation From Panoramic Images Robust To Occlusion Using Knowledge Distillation With 3D Shapes As Additional Inputs", "abstract": "Estimating the layout of a room from a single-shot panoramic image is important in virtual/augmented reality and furniture layout simulation. This involves identifying three-dimensional (3D) geometry, such as the location of corners and boundaries, and performing 3D reconstruction. However, occlusion is a common issue that can negatively impact room layout estimation, and this has not been thoroughly studied to date. It is possible to obtain 3D shape information of rooms as drawings of buildings and coordinates of corners from image datasets, thus we propose providing both 2D panoramic and 3D information to a model to effectively deal with occlusion. However, simply feeding 3D information to a model is not sufficient to utilize the shape information for an occluded area. Therefore, we improve the model by introducing 3D Intersection over Union (IoU) loss to effectively use 3D information. In some cases, drawings are not available or the construction deviates from a drawing. Considering such practical cases, we propose a method for distilling knowledge from a model trained with both images and 3D information to a model that takes only images as input. The proposed model, which is called Shape-Net, achieves state-of-the-art (SOTA) performance on benchmark datasets. We also confirmed its effectiveness in dealing with occlusion through significantly improved accuracy on images with occlusion compared with existing models."}, "cited_paper_content": {"title": "Matterport3D: Learning From Rgb-D Data In Indoor Environments", "abstract": "Access to large, diverse RGB-D datasets is critical for training RGB-D scene understanding algorithms. However, existing datasets still cover only a limited number of views or a restricted scale of spaces. In this paper, we introduce Matterport3D, a large-scale RGB-D dataset containing 10,800 panoramic views from 194,400 RGB-D images of 90 building-scale scenes. Annotations are provided with surface reconstructions, camera poses, and 2D and 3D semantic segmentations. The precise global alignment and comprehensive, diverse panoramic set of views over entire buildings enable a variety of supervised and self-supervised computer vision tasks, including keypoint matching, view overlap prediction, normal prediction from color, semantic segmentation, and region classification."}, "keywords": ["occlusion", "Matteport3D dataset"], "citation_intent": "method"}
{"citing_id": "2304.07147v1", "cited_id": "1807.02011", "section_title": "Testing", "citation": "We compare our results to that of a Dense, Spatial AE and Dense VAE model (Baur et al., 2020) and an SSIM AE #REFR .", "text_before_citation": ["After obtaining the optimal models from each hyperparameter study, we test the proposed models on the hold-out testing data as a final ablation study to showcase the added contribution of each proposed method."], "text_after_citation": ["These models are trained on PET only, however we also implement them with 2 channel inputs with PET and CT to showcase the effect of CT conditioning on the baseline models in comparison to the effects of conditioning for our proposed method.", "We additionally implement F-AnoGan for a further baseline comparison #OTHEREFR .", "Testing is carried out on 36 samples of which the best achievable DICE score is calculated in addition to the area under the precision-recall curve (AUPRC) as a suitable measure for segmentation performance under class imbalance.", "These results can be seen tabulated in Table  2 and visualised in Fig.", "7 qualitatively showcasing the best two state-of-the-art models against our added contributions."], "citing_paper_content": {"title": "Cross Attention Transformers For Multi-Modal Unsupervised Whole-Body Pet Anomaly Detection", "abstract": "Cancer is a highly heterogeneous condition that can occur almost anywhere in the human body. 18 F-fluorodeoxyglucose (18 F-FDG PET) is an imaging modality commonly used to detect cancer due to its high sensitivity and clear visualisation of the pattern of metabolic activity. Nonetheless, as cancer is highly heterogeneous, it is challenging to train general-purpose discriminative cancer detection models, with data availability and disease complexity often cited as a limiting factor. Unsupervised learning methods, more specifically anomaly detection models, have been suggested as a putative solution. These models learn a healthy representation of tissue and detect cancer by predicting deviations from the healthy norm, which requires models capable of accurately learning long-range interactions between organs, their imaging patterns, and other abstract features with high levels of expressivity. Such characteristics are suitably satisfied by transformers, which have been shown to generate state-of-the-art results in unsupervised anomaly detection by training on normal data. This work expands upon such approaches by introducing multi-modal conditioning of the transformer via cross-attention i.e. supplying anatomical reference information from paired CT images to aid the PET anomaly detection task. Furthermore, we show the importance and impact of codebook sizing within a Vector Quantized Variational Autoencoder, on the ability of the transformer network to fulfill the task of anomaly detection. Using 294 whole-body PET/CT samples containing various cancer types, we show that our anomaly detection method is robust and capable of achieving accurate cancer localization results even in cases where normal training data is unavailable. In addition, we show the efficacy of this approach on out-of-sample data showcasing the generalizability of this approach even with limited training data. Lastly, we propose to combine model uncertainty with a new kernel density estimation approach, and show that it provides clinically and statistically significant improvements in accuracy and robustness, when compared to the classic residual-based anomaly maps. Overall, a superior performance is demonstrated against leading state-of-the-art alternatives, drawing attention to the potential of these approaches."}, "cited_paper_content": {"title": "Improving Unsupervised Defect Segmentation By Applying Structural Similarity To Autoencoders", "abstract": "Convolutional autoencoders have emerged as popular methods for unsupervised defect segmentation on image data. Most commonly, this task is performed by thresholding a per-pixel reconstruction error based on an p-distance. This procedure, however, leads to large residuals whenever the reconstruction includes slight localization inaccuracies around edges. It also fails to reveal defective regions that have been visually altered when intensity values stay roughly consistent. We show that these problems prevent these approaches from being applied to complex real-world scenarios and that they cannot be easily avoided by employing more elaborate architectures such as variational or feature matching autoencoders. We propose to use a perceptual loss function based on structural similarity that examines inter-dependencies between local image regions, taking into account luminance, contrast, and structural information, instead of simply comparing single pixel values. It achieves significant performance gains on a challenging real-world dataset of nanofibrous materials and a novel dataset of two woven fabrics over state-of-the-art approaches for unsupervised defect segmentation that use per-pixel reconstruction error metrics."}, "keywords": ["Dense", "Dense VAE model"], "citation_intent": "result"}
{"citing_id": "2304.10297v1", "cited_id": "1911.11298", "section_title": "B. Meta Learning-Based Few-Shot Kg Reasoning", "citation": "FSRL #REFR proposed a novel method that captures both the different relation types and the different impacts of local neighbors.", "text_before_citation": ["Meta-learning is an effective learning paradigm across curating meta-tasks in order to transfer them to a new task #OTHEREFR .", "A large portion of FS-KGR methods employ the meta-learning paradigm to effectively tackle the issue of data scarcity in the target few-shot task #OTHEREFR [16] #OTHEREFR .", "These models can be broadly categorized into two types, i.e., metric-based models, and optimization-based models.", "The former attempts to acquire a universal measure and the corresponding matching functions from a group of training examples and GMatching #OTHEREFR focus on learning a matching metric that relies on entity embedding and subgraph structure."], "text_after_citation": ["However, such approaches typically rely on many manual-created meta-training tasks towards specific relations during the training process, leading to poor robustness if the meta-training set is not welldesigned.", "Moreover, creating an excessive number of metatraining sets from the background knowledge graph would eliminate a large proportion of edges, result in sparsity in the graph that hard to learn since the number of relations is limited in real KG, which limits the expressive ability of the model."], "citing_paper_content": {"title": "Sarf: Aliasing Relation Assisted Self-Supervised Learning For Few-Shot Relation Reasoning", "abstract": "Few-shot relation reasoning on knowledge graphs (FS-KGR) aims to infer long-tail data-poor relations, which has drawn increasing attention these years due to its practicalities. The pre-training of previous methods needs to manually construct the meta-relation set, leading to numerous labor costs. Self-supervised learning (SSL) is treated as a solution to tackle the issue, but still at an early stage for FS-KGR task. Moreover, most of the existing methods ignore leveraging the beneficial information from aliasing relations (AR), i.e., data-rich relations with similar contextual semantics to the target data-poor relation. Therefore, we proposed a novel Self-Supervised Learning model by leveraging Aliasing Relations to assist FS-KGR, termed SARF. Concretely, four main components are designed in our model, i.e., SSL reasoning module, AR-assisted mechanism, fusion module, and scoring function. We first generate the representation of the co-occurrence patterns in a generative manner. Meanwhile, the representations of aliasing relations are learned to enhance reasoning in the AR-assist mechanism. Besides, multiple strategies i.e., simple summation and learnable fusion, are offered for representation fusion. Finally, the generated representation is used for scoring. Extensive experiments on three few-shot benchmarks demonstrate that SARF achieves state-of-the-art performance compared with other methods in most cases."}, "cited_paper_content": {"title": "Few-Shot Knowledge Graph Completion", "abstract": "Knowledge graphs (KGs) serve as useful resources for various natural language processing applications. Previous KG completion approaches require a large number of training instances (i.e., head-tail entity pairs) for every relation. The real case is that for most of the relations, very few entity pairs are available. Existing work of one-shot learning limits method generalizability for few-shot scenarios and does not fully use the supervisory information; however, few-shot KG completion has not been well studied yet. In this work, we propose a novel few-shot relation learning model (FSRL) that aims at discovering facts of new relations with few-shot references. FSRL can effectively capture knowledge from heterogeneous graph structure, aggregate representations of few-shot references, and match similar entity pairs of reference set for every relation. Extensive experiments on two public datasets demonstrate that FSRL outperforms the state-of-the-art."}, "keywords": ["different relation types"], "citation_intent": "method"}
{"citing_id": "2304.04161v1", "cited_id": "2003.09871", "section_title": "Discussion", "citation": "A deep learning-based COVID-Net model designed by Wang and Wong #REFR acquired an accuracy of 92.4% with 53 COVID-19 samples out of the 16,756 total images. Ozturk et al.", "text_before_citation": ["As no benchmark dataset is yet available to analyze COVID-19 cases in chest x-ray images, authors collected images from different sources to build the datasets and conduct experiments. Table 5 . presents comparison of the proposed framework with state-of-the-arts."], "text_after_citation": ["#OTHEREFR tested the proposed DarkCovidNet by using 125 COVID-19, 500 pneumonia, and 500 normal images and achieved 98.08% and 87.02% for binary and ternary class scenario. Ioannis et al. #OTHEREFR detection of COVID-19 in two datasets.", "The first dataset comprises 224 COVID-19, 504 normal, and 700 bacterial pneumonia images and achieved an overall accuracy of 98.75% and 93.48% for 2-class and 3-class respectively with the VGG-19 model.", "The second dataset contains 224 COVID-19, 504 normal and, 714 pneumonia includes both viral and bacterial images.", "And for 2-class and 3-class scenarios achieved an accuracy of 96.78% and 94.72% using pretrained MobileNet-V2.", "In the proposed work, however, we used different datasets for 2-class and 3-class classification problem and achieved an accuracy of 94.8% and 92.8% by fine-tuning the pre-trained VGG-16 model."], "citing_paper_content": {"title": "Detection Of Covid-19 In Chest X-Ray Images Using Transfer Learning", "abstract": "COVID-19a highly contagious disease infected millions of people worldwide. With limited testing components, screening tools such as chest radiography can assist the clinicians in the diagnosis and assessing the progress of disease. The performance of deep learning-based systems for diagnosis of COVID-19 disease in radiograph images has been encouraging. This paper investigates the concept of transfer learning using two of the most well-known VGGNet architectures, namely VGG-16 and VGG-19. The classifier block and hyperparameters are fine-tuned to adopt the models for automatic detection of Covid-19 in chest x-ray images. We generated two different datasets to evaluate the performance of the proposed system for the identification of positive Covid-19 instances in a multiclass and binary classification problems. The experimental outcome demonstrates the usefulness of transfer learning for small-sized datasets particularly in the field of medical imaging, not only to prevent over-fitting and convergence problems but also to attain optimal classification performance as well."}, "cited_paper_content": {"title": "Covid-Net: A Tailored Deep Convolutional Neural Network Design For Detection Of Covid-19 Cases From Chest Radiography Images", "abstract": "The COVID-19 pandemic continues to have a devastating effect on the health and well-being of the global population. A critical step in the fight against COVID-19 is effective screening of infected patients, with one of the key screening approaches being radiological imaging using chest radiography. Motivated by this, a number of artificial intelligence (AI) systems based on deep learning have been proposed and results have been shown to be quite promising in terms of accuracy in detecting patients infected with COVID-19 using chest radiography images. However, to the best of the authors' knowledge, these developed AI systems have been closed source and unavailable to the research community for deeper understanding and extension, and unavailable for public access and use. Therefore, in this study we introduce COVID-Net, a deep convolutional neural network design tailored for the detection of COVID-19 cases from chest radiography images that is open source and available to the general public. We also describe the chest radiography dataset leveraged to train COVID-Net, which we will refer to as COVIDx and is comprised of 16,756 chest radiography images across 13,645 patient cases from two open access data repositories. Furthermore, we investigate how COVID-Net makes predictions using an explainability method in an attempt to gain deeper insights into critical factors associated with COVID cases, which can aid clinicians in improved screening. By no means a production-ready solution, the hope is that the open access COVID-Net, along with the description on constructing the open source COVIDx dataset, will be leveraged and build upon by both researchers and citizen data scientists alike to accelerate the development of highly accurate yet practical deep learning solutions for detecting COVID-19 cases and accelerate treatment of those who need it the most."}, "keywords": ["deep learning-based COVID-Net"], "citation_intent": "background"}
{"citing_id": "2303.06488v1", "cited_id": "1503.00805", "section_title": "Conclusion", "citation": "This differs from the objective studied in previous work, which is the \"standard\" Binary Search objective of minimizing the worst-case total number of queries, though #REFR also established hardness results for the case when nodes have known query costs.", "text_before_citation": ["Through the user's interactions with q, the system learns about the user's preferences, which can be modeled as the direction to t from q.", "However, serving the user q may impose a cost which increases in the distance between q and t.", "For example, repeatedly serving content that the user dislikes may cause them to stop using the service altogether with some probability.", "Hence, such a scenario fits well into the generalization of binary search that we consider.", "With these potential motivations in mind, in combination with the applications given in #OTHEREFR , it is natural that the cost of a proposal should increase with the distance in G from the proposal to the correct node."], "text_after_citation": ["This naturally motivates extending a study of our non-uniform cost model to the case of arbitrary undirected graphs.", "We ask: can any non-trivial approximation guarantees (or hardness) be achieved for interesting classes of cost functions? Even for the cost function that is identically 1 (i.e., the goal is to minimize the number of queries), minimizing the number of queries is not possible in polynomial time under the ETH.", "However, no approximation hardness or nontrivial positive approximation results (beyond the trivial O(log n)) are known; obtaining such results for more general cost functions would be of great interest.", "Another generalization of standard binary search that has been extensively studied is the case when the answers received by the algorithm could be erroneous.", "Indeed, optimal algorithms for this problem have been studied under different error models, on the line, on trees, and on general graphs. #OTHEREFR ."], "citing_paper_content": {"title": "Binary Search With Distance-Dependent Costs", "abstract": "We introduce a search problem generalizing the typical setting of Binary Search on the line. Similar to the setting for Binary Search, a target is chosen adversarially on the line, and in response to a query, the algorithm learns whether the query was correct, too high, or too low. Different from the Binary Search setting (which is the special case of a constant function), the cost of a query is a monotone non-decreasing function of the distance between the query and the correct answer; different functions can be used for queries that are too high vs. those that are too low. The algorithm's goal is to identify an adversarially chosen target with minimum total cost. Note that the algorithm does not even know the cost it incurred until the end, when the target is revealed. This abstraction captures many natural settings in which a principal experiments by setting a quantity (such as an item price, bandwidth, tax rate, medicine dosage, etc.) where the cost or regret increases the further the chosen setting is from the optimal one. First, we show that for arbitrary symmetric cost functions (i.e., the case when overshooting vs. undershooting by the same amount leads to the same cost), the standard Binary Search algorithm is a 4-approximation. We then show that when the cost functions are bounded-degree polynomials of the distance (including the asymmetric case), the problem can be solved optimally using Dynamic Programming; this relies on a careful encoding of the combined cost of past queries (which, recall, will only be revealed in the future). We then generalize the setting to finding a node on a tree; here, the response to a query is the direction on the tree in which the target is located, and the cost is increasing in the distance on the tree from the query to the target. Using the k-cut search tree framework of Berendsohn and Kozma and the ideas we developed for the case of the line, we give a PTAS when the cost function is a bounded-degree polynomial."}, "cited_paper_content": {"title": "Deterministic And Probabilistic Binary Search In Graphs", "abstract": "We consider the following natural generalization of Binary Search: in a given undirected, positively weighted graph, one vertex is a target. The algorithm's task is to identify the target by adaptively querying vertices. In response to querying a node $q$, the algorithm learns either that $q$ is the target, or is given an edge out of $q$ that lies on a shortest path from $q$ to the target. We study this problem in a general noisy model in which each query independently receives a correct answer with probability $p>\\frac{1}{2}$ (a known constant), and an (adversarial) incorrect one with probability $1-p$. Our main positive result is that when $p = 1$ (i.e., all answers are correct), $\\log_2 n$ queries are always sufficient. For general $p$, we give an (almost information-theoretically optimal) algorithm that uses, in expectation, no more than $(1 - \\delta) \\frac{\\log n}{1 - H(p)} + o(\\log n) + O(\\log^2 (1/\\delta))$ queries, and identifies the target correctly with probability at leas $1-\\delta$. Here, $H(p) = -(p \\log p + (1-p) \\log(1-p))$ denotes the entropy. The first bound is achieved by the algorithm that iteratively queries a 1-median of the nodes not ruled out yet; the second bound by careful repeated invocations of a multiplicative weights algorithm. Even for $p = 1$, we show several hardness results for the problem of determining whether a target can be found using $K$ queries. Our upper bound of $\\log_2 n$ implies a quasipolynomial-time algorithm for undirected connected graphs; we show that this is best-possible under the Strong Exponential Time Hypothesis (SETH). Furthermore, for directed graphs, or for undirected graphs with non-uniform node querying costs, the problem is PSPACE-complete. For a semi-adaptive version, in which one may query $r$ nodes each in $k$ rounds, we show membership in $\\Sigma_{2k-1}$ in the polynomial hierarchy, and hardness for $\\Sigma_{2k-5}$."}, "keywords": ["\"standard\" Binary Search"], "citation_intent": "result"}
{"citing_id": "2303.04137v1", "cited_id": "1911.05722", "section_title": "B. Visual Encoder", "citation": "This is important when the normalization layer is used in conjunction with Exponential Moving Average #REFR (commonly used in DDPMs).", "text_before_citation": ["The visual encoder maps the raw image sequence into a latent embedding O t and is trained end-to-end with the diffusion policy.", "Different camera views use separate encoders, and images in each timestep are encoded independently and then concatenated to form O t .", "We used a standard ResNet-18 (without pretraining) as the encoder with the following modifications: 1) Replace the global average pooling with a spatial softmax pooling to maintain spatial information #OTHEREFR . 2) Replace BatchNorm with GroupNorm #OTHEREFR for stable training."], "text_after_citation": [], "citing_paper_content": {"title": "Diffusion Policy: Visuomotor Policy Learning Via Action Diffusion", "abstract": "Fig. 1: Policy Representations. a) Explicit policy with different types of action representations. b) Implicit policy learns an energy function conditioned on both action and observation and optimizes for actions that minimize the energy landscape c) Diffusion policy refines noise into actions via a learned gradient field. This formulation provides stable training, allows the learned policy to accurately model multimodal action distributions, and accommodates high-dimensional action sequences."}, "cited_paper_content": {"title": "Momentum Contrast For Unsupervised Visual Representation Learning", "abstract": "We present Momentum Contrast (MoCo) for unsupervised visual representation learning. From a perspective on contrastive learning as dictionary look-up, we build a dynamic dictionary with a queue and a moving-averaged encoder. This enables building a large and consistent dictionary on-the-fly that facilitates contrastive unsupervised learning. MoCo provides competitive results under the common linear protocol on ImageNet classification. More importantly, the representations learned by MoCo transfer well to downstream tasks. MoCo can outperform its supervised pre-training counterpart in 7 detection/segmentation tasks on PASCAL VOC, COCO, and other datasets, sometimes surpassing it by large margins. This suggests that the gap between unsupervised and supervised representation learning has been largely closed in many vision tasks."}, "keywords": ["normalization layer"], "citation_intent": "method"}
{"citing_id": "2303.16068v1", "cited_id": "1802.09728", "section_title": "Preference Shifts In Recommendation", "citation": "In particular, Aspect-MF #REFR analyses the dynamicity of temporal preference aspects using a component-based approach, and identifies the aspects that are easy to drift.", "text_before_citation": ["User preference may shift over time for many reasons, including the changes of user features (e.g., income and pregnancy) and environment factors (e.g., seasonal variation).", "For example, in the scenario of food recommendation, a user might become liking expensive but healthy food if the user's income increases.", "Since such preference shifts are frequent in the real-world scenarios, the recommender models should update the user representations adaptively over time.", "Ignoring the shifting nature of preference to learn user representation will lead to inappropriate recommendations.", "In addition to disentangled recommendation and sequential recommendation, there exists some work with the potential of addressing this issue."], "text_after_citation": ["ST-LDA #OTHEREFR learns region-dependent personal interests and crowd preference to adapt to preference shifts.", "Lastly, MTUPD #OTHEREFR utilizes a forgetting curve function to calculate the correlations of user preference in different time periods.", "\u2022 Long-and short-term interest modeling.", "Another possibly related direction considers the modeling of both long-and short-term interests, where short-term interests are inferred from recent interactions to capture the preference shifts, and long-term interests represent the stable preference over time.", "Many studies have shed light on the modeling of long-and short-term user interests, for instance, PLSPL #OTHEREFR utilizes attention mechanism to characterize the long-term preference while integrating the location and category information to capture the short-term preference."], "citing_paper_content": {"title": "Causal Disentangled Recommendation Against User Preference Shifts", "abstract": "Recommender systems easily face the issue of user preference shifts. User representations will become outof-date and lead to inappropriate recommendations if user preference has shifted over time. To solve the issue, existing work focuses on learning robust representations or predicting the shifting pattern. There lacks a comprehensive view to discover the underlying reasons for user preference shifts. To understand the preference shift, we abstract a causal graph to describe the generation procedure of user interaction sequences. Assuming user preference is stable within a short period, we abstract the interaction sequence as a set of chronological environments. From the causal graph, we find that the changes of some unobserved factors (e.g., becoming pregnant) cause preference shifts between environments. Besides, the fine-grained user preference over categories sparsely affects the interactions with different items. Inspired by the causal graph, our key considerations to handle preference shifts lie in modeling the interaction generation procedure by: 1) capturing the preference shifts across environments for accurate preference prediction, and 2) disentangling the sparse influence from user preference to interactions for accurate effect estimation of preference. To this end, we propose a Causal Disentangled Recommendation (CDR) framework, which captures preference shifts via a temporal variational autoencoder and learns the sparse influence from multiple environments. Specifically, an encoder is adopted to infer the unobserved factors from user interactions while a decoder is to model the interaction generation process. Besides, we introduce two learnable matrices to disentangle the sparse influence from user preference to interactions. Lastly, we devise a multi-objective loss to optimize CDR. Extensive experiments on three datasets show the superiority of CDR in enhancing the generalization ability under user preference shifts."}, "cited_paper_content": {"title": "Modelling And Analysis Of Temporal Preference Drifts Using A Component-Based Factorised Latent Approach", "abstract": "textabstractIn recommender systems, human preferences are identified by a number of individual components with complicated interactions and properties. Recently, the dynamicity of preferences has been the focus of several studies. The changes in user preferences can originate from substantial reasons, like personality shift, or transient and circumstantial ones, like seasonal changes in item popularities. Disregarding these temporal drifts in modelling user preferences can result in unhelpful recommendations. Moreover, different temporal patterns can be associated with various preference domains, and preference components and their combinations. These components comprise preferences over features, preferences over feature values, conditional dependencies between features, socially-influenced preferences, and bias. For example, in the movies domain, the user can change his rating behaviour (bias shift), her preference for genre over language (feature preference shift), or start favouring drama over comedy (feature value preference shift). In this paper, we first propose a novel latent factor model to capture the domain-dependent component-specific temporal patterns in preferences. The component-based approach followed in modelling the aspects of preferences and their temporal effects enables us to arbitrarily switch components on and off. We evaluate the proposed method on three popular recommendation datasets and show that it significantly outperforms the most accurate state-of-the-art static models. The experiments also demonstrate the greater robustness and stability of the proposed dynamic model in comparison with the most successful models to date. We also analyse the temporal behaviour of different preference components and their combinations and show that the dynamic behaviour of preference components is highly dependent on the preference dataset and domain. Therefore, the results also highlight the importance of modelling temporal effects but also underline the advantages of a component-based architecture that is better suited to capture domain-specific balances in the contributions of the aspects."}, "keywords": ["temporal preference aspects"], "citation_intent": "method"}
{"citing_id": "2303.03037v1", "cited_id": "1904.07850", "section_title": "C. Quantitative Results", "citation": "In addition, we train and also report the results of CenterNet #REFR for comparing the detection performance and execution time, since the code is available.", "text_before_citation": ["We compare the results of EvCenterNet with four baseline methods.", "We include two sampling-based baselines, which utilize MC dropout and an Ensemble of 5 networks for uncertainty estimation.", "Note that we use #OTHEREFR for both the sampling-based baselines. For MC dropout, we perform five forward passes.", "For a sampling-free baseline, we implemented the CertainNet * baseline ourselves as no code was available.", "When available, we also report the results presented by the authors of CertainNet #OTHEREFR in their paper."], "text_after_citation": ["Tab.", "I compares the detection performance of the models on the validation set of the KITTI dataset.", "EvCenterNet performs slightly better than the sampling-free #OTHEREFR baseline on almost all the categories.", "The largest difference can be seen in the cyclist class, where EvCenterNet obtains 4.48pp (easy), 2.62pp (medium), and 2.96pp (hard) in comparison to CertainNet.", "We attribute the improvement in the small class to the proposed uncertainty-aware loss with weighting approach."], "citing_paper_content": {"title": "Evcenternet: Uncertainty Estimation For Object Detection Using Evidential Learning", "abstract": "Uncertainty estimation is crucial in safety-critical settings such as automated driving as it provides valuable information for several downstream tasks including high-level decision-making and path planning. In this work, we propose EvCenterNet, a novel uncertainty-aware 2D object detection framework utilizing evidential learning to directly estimate both classification and regression uncertainties. To employ evidential learning for object detection, we devise a combination of evidential and focal loss functions for the sparse heatmap inputs. We introduce class-balanced weighting for regression and heatmap prediction to tackle the class imbalance encountered by evidential learning. Moreover, we propose a learning scheme to actively utilize the predicted heatmap uncertainties to improve the detection performance by focusing on the most uncertain points. We train our model on the KITTI dataset and evaluate it on challenging out-of-distribution datasets including BDD100K and nuImages. Our experiments demonstrate that our approach improves the precision and minimizes the execution time loss in relation to the base model."}, "cited_paper_content": {"title": "Objects As Points", "abstract": "Detection identifies objects as axis-aligned boxes in an image. Most successful object detectors enumerate a nearly exhaustive list of potential object locations and classify each. This is wasteful, inefficient, and requires additional post-processing. In this paper, we take a different approach. We model an object as a single point --- the center point of its bounding box. Our detector uses keypoint estimation to find center points and regresses to all other object properties, such as size, 3D location, orientation, and even pose. Our center point based approach, CenterNet, is end-to-end differentiable, simpler, faster, and more accurate than corresponding bounding box based detectors. CenterNet achieves the best speed-accuracy trade-off on the MS COCO dataset, with 28.1% AP at 142 FPS, 37.4% AP at 52 FPS, and 45.1% AP with multi-scale testing at 1.4 FPS. We use the same approach to estimate 3D bounding box in the KITTI benchmark and human pose on the COCO keypoint dataset. Our method performs competitively with sophisticated multi-stage methods and runs in real-time."}, "keywords": ["detection performance", "CenterNet"], "citation_intent": "method"}
{"citing_id": "2304.03635v1", "cited_id": "1908.09999", "section_title": "Results", "citation": "Compared with A2J #REFR , certain performance improvement can be achieved on two datasets, which proves that A2J-Transformer has a strong generalization ability.", "text_before_citation": ["NYU Table 5 . Performance comparison on HANDS 2017 dataset #OTHEREFR . Our method can get state-of-the-art performance on this dataset.", "state-of-the-art depth based single hand estimation methods on NYU and HANDS 2017 dataset are given in Table 4 and Table 5 .", "Since A2J-Transformer is not specifically designed for single hand estimation on depth image, we just changed the input channel to verify the generalization ability of our model through this experiment. We can summarize that:", "\u2022 Although A2J-Transformer is based on the RGB image of interacting hands, it still achieves state-of-the-art performance on HANDS 2017 dataset and gets comparable result on NYU dataset.", "This relys on the strong ability of A2J-Transformer to grasp the articulated hand information and the fitting ability of 3D anchor points."], "text_after_citation": [], "citing_paper_content": {"title": "A2J-Transformer: Anchor-To-Joint Transformer Network For 3D Interacting Hand Pose Estimation From A Single Rgb Image", "abstract": "3D interacting hand pose estimation from a single RGB image is a challenging task, due to serious self-occlusion and inter-occlusion towards hands, confusing similar appearance patterns between 2 hands, ill-posed joint position mapping from 2D to 3D, etc.. To address these, we propose to extend A2J-the state-of-the-art depth-based 3D single hand pose estimation method-to RGB domain under interacting hand condition. Our key idea is to equip A2J with strong local-global aware ability to well capture interacting hands' local fine details and global articulated clues among joints jointly. To this end, A2J is evolved under Transformer's non-local encoding-decoding framework to build A2J-Transformer. It holds 3 main advantages over A2J. First, self-attention across local anchor points is built to make them global spatial context aware to better capture joints' articulation clues for resisting occlusion. Secondly, each anchor point is regarded as learnable query with adaptive feature learning for facilitating pattern fitting capacity, instead of having the same local representation with the others. Last but not least, anchor point locates in 3D space instead of 2D as in A2J, to leverage 3D pose prediction. Experiments on challenging InterHand 2.6M demonstrate that, A2J-Transformer can achieve state-ofthe-art model-free performance (3.38mm MPJPE advancement in 2-hand case) and can also be applied to depth domain with strong generalization. The code is avaliable at https://github.com/ChanglongJiangGit/ A2J-Transformer."}, "cited_paper_content": {"title": "A2J: Anchor-To-Joint Regression Network For 3D Articulated Pose Estimation From A Single Depth Image", "abstract": "For 3D hand and body pose estimation task in depth image, a novel anchor-based approach termed Anchor-to-Joint regression network (A2J) with the end-to-end learning ability is proposed. Within A2J, anchor points able to capture global-local spatial context information are densely set on depth image as local regressors for the joints. They contribute to predict the positions of the joints in ensemble way to enhance generalization ability. The proposed 3D articulated pose estimation paradigm is different from the state-of-the-art encoder-decoder based FCN, 3D CNN and point-set based manners. To discover informative anchor points towards certain joint, anchor proposal procedure is also proposed for A2J. Meanwhile 2D CNN (i.e., ResNet-50) is used as backbone network to drive A2J, without using time-consuming 3D convolutional or deconvolutional layers. The experiments on 3 hand datasets and 2 body datasets verify A2J's superiority. Meanwhile, A2J is of high running speed around 100 FPS on single NVIDIA 1080Ti GPU."}, "keywords": ["strong generalization ability"], "citation_intent": "result"}
{"citing_id": "2303.02207v1", "cited_id": "1505.07427", "section_title": "V. Results And Discussions", "citation": "While each approach is based on PoseNet #REFR , they can be applied to other pose estimation models.", "text_before_citation": ["This section presents a chronological review of the results from the four frameworks proposed for conformalized visual odometry (CVO) and compared in Table I .", "The aim of these methods is to develop reliable prediction intervals that can capture true uncertainty in data-driven pose estimation with guaranteed confidence."], "text_after_citation": ["In highlighting the differences between the methods, we use the data from scene 2 of RGB-D Scenes Dataset v.2 #OTHEREFR as visuals of its pose trajectory are easy to comprehend.", "All results illustrate uncertainty bands that exhibit at least 90% guaranteed marginal coverage through CI.", "In Sec.", "III-A, we introduced univariate conformalized quantile regression (CQR) to generate uncertainty bands. However, as shown in Fig.", "5 , the resulting prediction interval is overly cautious and box-shaped despite its ability to adjust to the data's heteroscedasticity."], "citing_paper_content": {"title": "Lightweight, Uncertainty-Aware Conformalized Visual Odometry", "abstract": "Data-driven visual odometry (VO) is a critical subroutine for autonomous edge robotics, and recent progress in the field has produced highly accurate point predictions in complex environments. However, emerging autonomous edge robotics devices like insect-scale drones and surgical robots lack a computationally efficient framework to estimate VO's predictive uncertainties. Meanwhile, as edge robotics continue to proliferate into mission-critical application spaces, awareness of model's the predictive uncertainties has become crucial for risk-aware decision-making. This paper addresses this challenge by presenting a novel, lightweight, and statistically robust framework that leverages conformal inference (CI) to extract VO's uncertainty bands. Our approach represents the uncertainties using flexible, adaptable, and adjustable prediction intervals that, on average, guarantee the inclusion of the ground truth across all degrees of freedom (DOF) of pose estimation. We discuss the architectures of generative deep neural networks for estimating multivariate uncertainty bands along with point (mean) prediction. We also present techniques to improve the uncertainty estimation accuracy, such as leveraging Monte Carlo dropout (MC-dropout) for data augmentation. Finally, we propose a novel training loss function that combines interval scoring and calibration loss with traditional training metrics-mean-squared error and KLdivergence-to improve uncertainty-aware learning. Our simulation results demonstrate that the presented framework consistently captures true uncertainty in pose estimations across different datasets, estimation models, and applied noise types, indicating its wide applicability."}, "cited_paper_content": {"title": "Posenet: A Convolutional Network For Real-Time 6-Dof Camera Relocalization", "abstract": "We present a robust and real-time monocular six degree of freedom relocalization system. Our system trains a convolutional neural network to regress the 6-DOF camera pose from a single RGB image in an end-to-end manner with no need of additional engineering or graph optimisation. The algorithm can operate indoors and outdoors in real time, taking 5ms per frame to compute. It obtains approximately 2m and 3 degrees accuracy for large scale outdoor scenes and 0.5m and 5 degrees accuracy indoors. This is achieved using an efficient 23 layer deep convnet, demonstrating that convnets can be used to solve complicated out of image plane regression problems. This was made possible by leveraging transfer learning from large scale classification data. We show that the PoseNet localizes from high level features and is robust to difficult lighting, motion blur and different camera intrinsics where point based SIFT registration fails. Furthermore we show how the pose feature that is produced generalizes to other scenes allowing us to regress pose with only a few dozen training examples."}, "keywords": ["pose estimation models"], "citation_intent": "method"}
{"citing_id": "2305.02627v1", "cited_id": "2004.01658", "section_title": "Experimental Results", "citation": "As Table 7 shows, B-Seg performs favorably over other methods for most sub-categories, except that it performs slightly worse than PointGroup #REFR on three categories in the Campus.", "text_before_citation": ["2022] , though it is clusteringfree, it needs to encode every instance proposal into a kernel feature for the dynamic convolution. While DyCo3D #OTHEREFR", "2021 ] also adopts a dynamic-convolution-based model, it still requires a pointwise grouping first.", "On the contrary, B-Seg adopts a simple but effective strategy by aggregating points with a relation matrix and merging points only through a small set of candidates. The visualization result can be seen in Fig. 6 .", "Fine-grained Building Instance Segmentation Performance.", "We further evaluate the segmentation performance of all methods on the building sub-categories."], "text_after_citation": ["These categories are overall harder to distinguish and their features are more similar in Campus, so requiring a stronger feature extractor.", "Note also that we do not have results for some categories in Wuhu and Longhua (see \"-\" in the table), as the number of associated buildings in these cases is either too small or not existent.", "From the experiments, we can see the strength of B-Seg in analyzing buildings in 3D urban scenes and its potential for supporting building-aware 3D machine learning tasks.", "Joint-training on all scenes.", "We train the methods collectively on samples of all scenes."], "citing_paper_content": {"title": "Urbanbis: A Large-Scale Benchmark For Fine-Grained Urban Building Instance Segmentation", "abstract": "Figure 1: UrbanBIS provides 2.5 billion 3D point samples over six scenes, covering a vast area of 10.78 km 2. Particularly, this large 3D dataset is annotated not only in the urban semantic level (top right) but also in the building instance level (bottom right) with fine-grained building categories (bottom left)."}, "cited_paper_content": {"title": "Pointgroup: Dual-Set Point Grouping For 3D Instance Segmentation", "abstract": "Instance segmentation is an important task for scene understanding. Compared to the fully-developed 2D, 3D instance segmentation for point clouds have much room to improve. In this paper, we present PointGroup, a new end-to-end bottom-up architecture, specifically focused on better grouping the points by exploring the void space between objects. We design a two-branch network to extract point features and predict semantic labels and offsets, for shifting each point towards its respective instance centroid. A clustering component is followed to utilize both the original and offset-shifted point coordinate sets, taking advantage of their complementary strength. Further, we formulate the ScoreNet to evaluate the candidate instances, followed by the Non-Maximum Suppression (NMS) to remove duplicates. We conduct extensive experiments on two challenging datasets, ScanNet v2 and S3DIS, on which our method achieves the highest performance, 63.6% and 64.0%, compared to 54.9% and 54.4% achieved by former best solutions in terms of mAP with IoU threshold 0.5."}, "keywords": ["PointGroup"], "citation_intent": "result"}
{"citing_id": "2304.07362v1", "cited_id": "1310.3235", "section_title": "Related Work", "citation": "However classical and quantum error correction are fundamentally different #REFR , and these results do not directly translate to the quantum case.", "text_before_citation": ["#OTHEREFR ) obtains the largest system size and threshold among neural decoders by combining a convolutional neural network backbone with a union find decoder.", "In our work we show that our model, which does not rely on a handcrafted decoder, achieves higher accuracy.", "From the perspective of equivariant architectures #OTHEREFR , our work studies a generalized form of equivariance, where the output representation depends on the values of the inputs to the neural network.", "To the best of our knowledge, this type of symmetry properties for a neural network have not been considered before.", "Finally, neural decoders for classical error correction were discussed as a form of generalized belief propagation in #OTHEREFR ."], "text_after_citation": ["See #OTHEREFR for an attempt which however does not achieve good accuracy for the toric code."], "citing_paper_content": {"title": "The End: An Equivariant Neural Decoder For Quantum Error Correction", "abstract": "Quantum error correction is a critical component for scaling up quantum computing. Given a quantum code, an optimal decoder maps the measured code violations to the most likely error that occurred, but its cost scales exponentially with the system size. Neural network decoders are an appealing solution since they can learn from data an efficient approximation to such a mapping and can automatically adapt to the noise distribution. In this work, we introduce a data efficient neural decoder that exploits the symmetries of the problem. To this end, we characterize the symmetries of the optimal decoder for the toric code and propose a novel equivariant architecture that achieves state of the art reconstruction accuracy compared to previous neural decoders."}, "cited_paper_content": {"title": "Hardness Of Decoding Quantum Stabilizer Codes", "abstract": "In this paper, we address the computational hardness of optimally decoding a quantum stabilizer code. Much like classical linear codes, errors are detected by measuring certain check operators which yield an error syndrome, and the decoding problem consists of determining the most likely recovery given the syndrome. The corresponding classical problem is known to be NP-complete, and are appropriate a similar decoding problem for quantum codes is also known to be NP-complete. However, this decoding strategy is not optimal in the quantum setting as it does not consider error degeneracy, which causes distinct errors to have the same effect on the code. Here, we show that optimal decoding of stabilizer codes (previously known to be NP-hard) is in fact computationally much harder than optimal decoding of classical linear codes, it is #P-complete."}, "keywords": ["quantum error correction"], "citation_intent": "result"}
{"citing_id": "2303.06296v1", "cited_id": "1802.05957", "section_title": "Introduction", "citation": "We then provide a simple fix, \u03c3Reparam, which reparameterizes all weight matrices by sequentially applying Spectral Normalization #REFR and a learned multiplicative scalar.", "text_before_citation": ["We see that interventions inducing the largest drop in attention entropy result in the sharpness exceeding the stability threshold, whereas the later interventions do not cause the threshold to be crossed, explaining how they can recover.", "For details on the empirical setup and additional results see Appendix B.", "The empirical correlation of entropy collapse and training instability leads to the following questions: 1) How do we prevent entropy collapse? 2) Can we improve training stability by doing so? We answer these by showing that entropy collapse can be effectively prevented by controlling the spectral norms of the query and key projections.", "In particular, we prove a tight lower bound on the attention entropy, which decreases exponentially fast with the growth of the spectral norm of the attention matrix logits.", "This bound suggests that entropy collapse can occur swiftly when letting the spectral norm of the weights increase uncontrollably."], "text_after_citation": ["Intuitively, \u03c3Reparam decouples the update of the spectral norms of weights from their dimensionality, which allows them to update smoothly and in a controlled way.", "Also note that \u03c3Reparam does not change the model space, which allows one to learn an equally expressive model.", "We evaluate five tasks: image classification, self-supervised learning (SSL), machine translation, automatic speech recognition (Appendix E), and language modeling (Appendix G). We highlight the empirical results as follows:", "1.", "We show that entropy collapse is commonly observed in the baseline models of various benchmarks."], "citing_paper_content": {"title": "Stabilizing Transformer Training By Preventing Attention Entropy Collapse", "abstract": "Training stability is of great importance to Transformers. In this work, we investigate the training dynamics of Transformers by examining the evolution of the attention layers. In particular, we track the attention entropy for each attention head during the course of training, which is a proxy for model sharpness. We identify a common pattern across different architectures and tasks, where low attention entropy is accompanied by high training instability, which can take the form of oscillating loss or divergence. We denote the pathologically low attention entropy, corresponding to highly concentrated attention scores, as entropy collapse. As a remedy, we propose \u03c3Reparam, a simple and efficient solution where we reparametrize all linear layers with spectral normalization and an additional learned scalar. We demonstrate that the proposed reparameterization successfully prevents entropy collapse in the attention layers, promoting more stable training. Additionally, we prove a tight lower bound of the attention entropy, which decreases exponentially fast with the spectral norm of the attention logits, providing additional motivation for our approach. We conduct experiments with \u03c3Reparam on image classification, image self-supervised learning, machine translation, automatic speech recognition, and language modeling tasks, across Transformer architectures. We show that \u03c3Reparam provides stability and robustness with respect to the choice of hyperparameters, going so far as enabling training (a) a Vision Transformer to competitive performance without warmup, weight decay, layer normalization or adaptive optimizers; (b) deep architectures in machine translation and (c) speech recognition to competitive performance without warmup and adaptive optimizers."}, "cited_paper_content": {"title": "Spectral Normalization For Generative Adversarial Networks", "abstract": "One of the challenges in the study of generative adversarial networks is the instability of its training. In this paper, we propose a novel weight normalization technique called spectral normalization to stabilize the training of the discriminator. Our new normalization technique is computationally light and easy to incorporate into existing implementations. We tested the efficacy of spectral normalization on CIFAR10, STL-10, and ILSVRC2012 dataset, and we experimentally confirmed that spectrally normalized GANs (SN-GANs) is capable of generating images of better or equal quality relative to the previous training stabilization techniques."}, "keywords": ["Spectral Normalization"], "citation_intent": "method"}
{"citing_id": "2304.14541v1", "cited_id": "1511.06335", "section_title": "Experiment Results", "citation": "As shown in Table 2 , the proposed DSC algorithm outperforms all the baseline methods except that the DEC algorithm #REFR has a lower RM SE mean score.", "text_before_citation": ["To compute these evaluation results we trained each model several times and picked the best result from them.", "The results of DEC and DTC methods are generated by running their published codes.", "The hyperparameters of these models like batch size, learning rate, maximum iteration, etc.", "are finetuned to get better results on the applied dataset.", "The k-means and hierarchical clustering method are taken from sklearn python library and applied to preprocessed data without any dimension reduction."], "text_after_citation": ["Note that the DEC method does not consider spatial features and uses a greedy pretraining process.", "In contrast, our proposed method achieved RM SE mean close to DEC without any preprocessing step and significant improvement for the other four evaluation metrics.", "In Figure 2 , we visualize the clusters generated by the DEC and the proposed DSC method.", "As the latent feature space is not suitable for visualization we use t-SNE #OTHEREFR to reduce the dimension of the feature space.", "The cluster plot of the DEC method shows some clear overlap between the three clusters represented by red, orange, and cyan, even though DEC method achieved a better RM SE mean score."], "citing_paper_content": {"title": "Deep Spatiotemporal Clustering: A Temporal Clustering Approach For Multi-Dimensional Climate Data", "abstract": "Clustering high-dimensional spatiotemporal data using an unsupervised approach is a challenging problem for many data-driven applications. Existing state-of-the-art methods for unsupervised clustering use different similarity and distance functions but focus on either spatial or temporal features of the data. Concentrating on joint deep representation learning of spatial and temporal features, we propose Deep Spatiotemporal Clustering (DSC), a novel algorithm for the temporal clustering of high-dimensional spatiotemporal data using an unsupervised deep learning method. Inspired by the U-net architecture, DSC utilizes an autoencoder integrating CNN-RNN layers to learn latent representations of the spatiotemporal data. DSC also includes a unique layer for cluster assignment on latent representations that uses the Student's t-distribution. By optimizing the clustering loss and data reconstruction loss simultaneously, the algorithm gradually improves clustering assignments and the nonlinear mapping between low-dimensional latent feature space and high-dimensional original data space. A multivariate spatiotemporal climate dataset is used to evaluate the efficacy of the proposed method. Our extensive experiments show our approach outperforms both conventional and deep learning-based unsupervised clustering algorithms. Additionally, we compared the proposed model with its various variants (CNN encoder, CNN autoencoder, CNN-RNN encoder, CNN-RNN autoencoder, etc.) to get insight into using both the CNN and RNN layers in the autoencoder, and our proposed technique outperforms these variants in terms of clustering results."}, "cited_paper_content": {"title": "Unsupervised Deep Embedding For Clustering Analysis", "abstract": "Clustering is central to many data-driven application domains and has been studied extensively in terms of distance functions and grouping algorithms. Relatively little work has focused on learning representations for clustering. In this paper, we propose Deep Embedded Clustering (DEC), a method that simultaneously learns feature representations and cluster assignments using deep neural networks. DEC learns a mapping from the data space to a lower-dimensional feature space in which it iteratively optimizes a clustering objective. Our experimental evaluations on image and text corpora show significant improvement over state-of-the-art methods."}, "keywords": ["RM SE"], "citation_intent": "result"}
{"citing_id": "2303.16985v1", "cited_id": "1902.00751", "section_title": "Language Adapters, Task Adapters And Adapterhub", "citation": "Adapters also have the benefit of requiring as little as 1MB of storage space per task #REFR .", "text_before_citation": ["Fine-tuning all or a majority of a pre-trained model needs a lot of computational resources and also it depends on the availability of the dataset for the specific task which is the problem in the African context. #OTHEREFR", "(2020) introduced AdapterHub, which is a central repository for pretrained adapter modules.", "Adapters refers to a set of newly introduced weights, typically within the layers of a transformer model.", "Adapters provide an alternative to fully fine-tuning the model for each downstream task while maintaining performance."], "text_after_citation": ["Rather than pre-training a large language model, a language adapter can be pre-trained for each language.", "Adapter modules are parameter-efficient, sustainable, and achieve near SOTA results on low-resource and cross-lingual tasks #OTHEREFR .", "In this work, we leverage Adapter-based tuning for African languages because it has been shown to mitigate forgetting issues better than fine-tuning, as it yields representations with less deviation from those generated by the original pre-training #OTHEREFR .", "Moreover, Adapter fine-tuning does not take a lot of time because of its lightweight nature: we believe this will be one of the main advantages for people who are highly limited with computational resources."], "citing_paper_content": {"title": "Adapting To The Low-Resource Double-Bind: Investigating Low-Compute Methods On Low-Resource African Languages", "abstract": "Many natural language processing (NLP) tasks make use of massively pre-trained language models, which are computationally expensive. However, access to high computational resources added to the issue of data scarcity of African languages constitutes a real barrier to research experiments on these languages. In this work, we explore the applicability of low-compute approaches such as language adapters in the context of this low-resource double-bind. We intend to answer the following question: do language adapters allow those who are doubly bound by data and compute to practically build useful models? Through fine-tuning experiments on African languages, we evaluate their effectiveness as cost-effective approaches to low-resource African NLP. Using solely free compute resources, our results show that language adapters achieve comparable performances to massive pretrained language models which are heavy on computational resources. This opens the door to further experimentation and exploration on full-extent of language adapters capacities."}, "cited_paper_content": {"title": "Parameter-Efficient Transfer Learning For Nlp", "abstract": "Fine-tuning large pre-trained models is an effective transfer mechanism in NLP. However, in the presence of many downstream tasks, fine-tuning is parameter inefficient: an entire new model is required for every task. As an alternative, we propose transfer with adapter modules. Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones. The parameters of the original network remain fixed, yielding a high degree of parameter sharing. To demonstrate adapter's effectiveness, we transfer the recently proposed BERT Transformer model to 26 diverse text classification tasks, including the GLUE benchmark. Adapters attain near state-of-the-art performance, whilst adding only a few parameters per task. On GLUE, we attain within 0.4% of the performance of full fine-tuning, adding only 3.6% parameters per task. By contrast, fine-tuning trains 100% of the parameters per task."}, "keywords": ["Adapters"], "citation_intent": "background"}
{"citing_id": "2304.00900v1", "cited_id": "1612.04183", "section_title": "Visual Stability", "citation": "Additionally, this GIF supplements our findings by illustrating how increasing n out contributes to improved stability, a logical yet interesting result #REFR .", "text_before_citation": ["When examining the MAE, M4 frequently performs the worst, followed by either EveryNth or LTTB.", "Regarding the MEAP, EveryNth demonstrates to be the most unstable, followed by M4 and then LTTB.", "Notably, when focusing solely on the peaks (MEAP), the margin between MinMax and the other algorithms became even more pronounced.", "For the generated noise in the last column, Min-Max once again emerges as the most stable algorithm, while LTTB, in contrast to previous templates, is the least stable.", "Another note- worthy observation is the minimal variation in the results across the three different cases."], "text_after_citation": ["A plausible explanation for MinMax's superior stability lies in its high locality, referring to the extent that data point selection depends on individual bin values without influence from neighboring bins.", "High locality results in greater stability and reduced alignment issues susceptibility.", "MinMax has the highest locality among all considered algorithms, as it selects the minimum and maximum values for each bin with no interplay of neighboring bins.", "In contrast, LTTB exhibits #OTHEREFR Further details regarding visual stability, showing that at least 30% of the data points are included for the MEAP metric, can be found in the GitHub repository.", "lower locality, initializing its sequential data point selection process with the first data point and then using the previously selected data point from the preceding bin (together with the average point of the next bin) as input for the local algorithm."], "citing_paper_content": {"title": "Data Point Selection For Line Chart Visualization: Methodological Assessment And Evidence-Based Guidelines", "abstract": "Figure 1: Visual analysis of four time series data point selection algorithms. Each row corresponds to a unique template, with the rightmost column presenting a reference image, constructed from 200k data points. The four adjacent columns demonstrate the varying visual representations, each generated by a data point selection algorithm, all employing an equal number of output data points (n out = 300). Visualization properties used include: Plotly as toolkit, line-width set to 2, and anti-aliasing enabled. A GIF further demonstrates the above visualization with varying values for n out ."}, "cited_paper_content": {"title": "Interactive (Statistical) Visualisation And Exploration Of A Billion Objects With Vaex", "abstract": "With new catalogues arriving such as the Gaia DR1, containing more than a billion objects, new methods of handling and visualizing these data volumes are needed. In visualization, one problem is that the number of datapoints can become so large, that a scatter plot becomes cluttered. Another problem is that with over a billion objects, only a few cpu cycles are available per object if one wants to process them within a second, making traditional methods by rendering glyphs not viable. Instead, we show that by calculating statistics on a regular (N-dimensional) grid, visualizations of a billion objects can be done within a second on a modern desktop computer. This is achieved using memory mapping of hdf5 files together with a simple binning algorithm, which are part of a Python library called vaex. This enables efficient exploration or large datasets interactively, making science exploration of large catalogues feasible. Vaex is a Python library, which also integrates well in the Jupyter/Numpy/Astropy/matplotlib stack. Build on top of this is the vaex application, which allows for interactive exploration and visualization. The motivation for developing vaex is the catalogue of the Gaia satellite, however, vaex can also be used on SPH or N-body simulations, any other (future) catalogues such as SDSS, Pan-STARRS, LSST, WISE, 2MASS, etc. or other tabular data. The homepage for vaex is http://vaex.astro.rug.nl."}, "keywords": ["improved stability", "findings"], "citation_intent": "result"}
{"citing_id": "2303.10677v1", "cited_id": "1604.07316", "section_title": "B. Steering Wheel Angle Prediction", "citation": "The CNN models have been demonstrated to efficiently perform a lane following task on an unmarked road #REFR .", "text_before_citation": ["As of 2020 in the United States, about 30% of the roads are still unpaved #OTHEREFR .", "Recent technologies like the end-to-end learning approach enable AVs to drive safely even under challenging circumstances, such as unpaved and unmarked roads.", "The prediction of steering wheel angle is one of the critical aspects of an autonomous vehicle (Society of Automotive Engineers (SAE) Level 4 and Level 5 #OTHEREFR ) that controls the lateral position of the vehicle.", "The steering wheel angle is predicted from the RGB images collected from the front-facing camera as input.", "The dataset consisting of camera images and the steering wheel angle is trained on a CNN model #OTHEREFR ."], "text_after_citation": ["Related literature has demonstrated that FL can collaboratively learn the neural network model at a significantly lower communication cost while also preserving privacy #OTHEREFR - #OTHEREFR ."], "citing_paper_content": {"title": "A Survey Of Federated Learning For Connected And Automated Vehicles", "abstract": "Connected and Automated Vehicles (CAVs) are one of the emerging technologies in the automotive domain that has the potential to alleviate the issues of accidents, traffic congestion, and pollutant emissions, leading to a safe, efficient, and sustainable transportation system. Machine learningbased methods are widely used in CAVs for crucial tasks like perception, motion planning, and motion control, where machine learning models in CAVs are solely trained using the local vehicle data, and the performance is not certain when exposed to new environments or unseen conditions. Federated learning (FL) is an effective solution for CAVs that enables a collaborative model development with multiple vehicles in a distributed learning framework. FL enables CAVs to learn from a wide range of driving environments and improve their overall performance while ensuring the privacy and security of local vehicle data. In this paper, we review the progress accomplished by researchers in applying FL to CAVs. A broader view of the various data modalities and algorithms that have been implemented on CAVs is provided. Specific applications of FL are reviewed in detail, and an analysis of the challenges and future scope of research are presented."}, "cited_paper_content": {"title": "End To End Learning For Self-Driving Cars", "abstract": "We trained a convolutional neural network (CNN) to map raw pixels from a single front-facing camera directly to steering commands. This end-to-end approach proved surprisingly powerful. With minimum training data from humans the system learns to drive in traffic on local roads with or without lane markings and on highways. It also operates in areas with unclear visual guidance such as in parking lots and on unpaved roads. ::: The system automatically learns internal representations of the necessary processing steps such as detecting useful road features with only the human steering angle as the training signal. We never explicitly trained it to detect, for example, the outline of roads. ::: Compared to explicit decomposition of the problem, such as lane marking detection, path planning, and control, our end-to-end system optimizes all processing steps simultaneously. We argue that this will eventually lead to better performance and smaller systems. Better performance will result because the internal components self-optimize to maximize overall system performance, instead of optimizing human-selected intermediate criteria, e.g., lane detection. Such criteria understandably are selected for ease of human interpretation which doesn't automatically guarantee maximum system performance. Smaller networks are possible because the system learns to solve the problem with the minimal number of processing steps. ::: We used an NVIDIA DevBox and Torch 7 for training and an NVIDIA DRIVE(TM) PX self-driving car computer also running Torch 7 for determining where to drive. The system operates at 30 frames per second (FPS)."}, "keywords": ["CNN models"], "citation_intent": "method"}
{"citing_id": "2304.14907v1", "cited_id": "1606.04838", "section_title": "Consider The Random Variable", "citation": "Results similar to Theorem 3.2 have been proved for the stochastic gradient method in the unconstrained setting #REFR .", "text_before_citation": [", which is the first desired conclusion.", "The second desired conclusion follow from the fact that if r exists as stated, then", "\u2022 2 H \u22121 k \u2265 r \u22121 \u2022 2", "2 for all k \u2208 N.", "The last desired conclusion follows using the same argument as in the proof of Theorem 3.1."], "text_after_citation": ["In fact, for the stochastic gradient method in an unconstrained setting, one can prove under conditions that are similar to ours that the gradient of the objective function convergences to zero almost surely; this is stronger than a lim inf result of the type in Theorem 3.2.", "However, such results rely on the gradient of the objective function being Lipschitz continuous.", "In our setting, we have the Lipschitzcontinuity-type property in Lemma 3.2, but the gradient of the (shifted) barrier function is not globally Lipschitz over B, meaning that such a result in the unconstrained setting does not carry over to our present setting."], "citing_paper_content": {"title": "A Stochastic-Gradient-Based Interior-Point Algorithm For Solving Smooth Bound-Constrained Optimization Problems", "abstract": "A stochastic-gradient-based interior-point algorithm for minimizing a continuously differentiable objective function (that may be nonconvex) subject to bound constraints is presented, analyzed, and demonstrated through experimental results. The algorithm is unique from other interior-point methods for solving smooth (nonconvex) optimization problems since the search directions are computed using stochastic gradient estimates. It is also unique in its use of inner neighborhoods of the feasible region-defined by a positive and vanishing neighborhood-parameter sequence-in which the iterates are forced to remain. It is shown that with a careful balance between the barrier, step-size, and neighborhood sequences, the proposed algorithm satisfies convergence guarantees in both deterministic and stochastic settings. The results of numerical experiments show that in both settings the algorithm can outperform a projected-(stochastic)-gradient method."}, "cited_paper_content": {"title": "Optimization Methods For Large-Scale Machine Learning", "abstract": "This paper provides a review and commentary on the past, present, and future of numerical optimization algorithms in the context of machine learning applications. Through case studies on text classification and the training of deep neural networks, we discuss how optimization problems arise in machine learning and what makes them challenging. A major theme of our study is that large-scale machine learning represents a distinctive setting in which the stochastic gradient (SG) method has traditionally played a central role while conventional gradient-based nonlinear optimization techniques typically falter. Based on this viewpoint, we present a comprehensive theory of a straightforward, yet versatile SG algorithm, discuss its practical behavior, and highlight opportunities for designing algorithms with improved performance. This leads to a discussion about the next generation of optimization methods for large-scale machine learning, including an investigation of two main streams of research on techniques that diminish noise in the stochastic directions and methods that make use of second-order derivative approximations."}, "keywords": ["stochastic gradient method"], "citation_intent": "result"}
{"citing_id": "2304.06672v1", "cited_id": "1910.08519", "section_title": "Related Work", "citation": "Ringer et al., study the effects of texture bias illustrated by Convolutional Neural Networks (CNN) in FSL #REFR and address the same by modifying the training data with a combination of non-texture and texture-based images.", "text_before_citation": ["IER incorporates both invariance and equivariance inductive biases of certain geometric transformations.", "Unlike RFS, IER employs a multi-head distillation over standard supervised distillation. Liu et al.", "#OTHEREFR learn a base embedding model by unifying the two-stage training in RFS-Distill using online self-distillation.", "We incorporate shape bias by distilling features in latent space and model decisions between two networks trained synchronously.", "Overcoming Biases in FSL: The biases affecting the generalization of conventional DNNs create a larger impact on the FSL setup."], "text_after_citation": ["The image background also serves as a shortcut in the FSL scenario and impacts classification performance. Luo et al.", "#OTHEREFR mitigates the problem by sampling the foreground and original images in the training and testing phases.", "Finally, even though the feature distribution of the base classes used in pretraining FSL models is well defined, the distribution is affected by the domain shift incurred by the novel classes. Tao et al.", "#OTHEREFR propose Distribution Calibration Module (DCM) and Selected Sampling (SS) to mitigate class-agnostic and class-specific bias, respectively, in the meta-testing phase.", "However, LSFSL utilizes the implicit shape information in the training data with alignment objectives that aid in improving the generalization of few-shot models by learning global features."], "citing_paper_content": {"title": "Lsfsl: Leveraging Shape Information In Few-Shot Learning", "abstract": "Few-shot learning (FSL) techniques seek to learn the underlying patterns in data using fewer samples, analogous to how humans learn from limited experience. In this limiteddata scenario, the challenges associated with deep neural networks, such as shortcut learning and texture bias behaviors, are further exacerbated. Moreover, the significance of addressing shortcut learning is not yet fully explored in the few-shot setup. To address these issues, we propose LS-FSL, which enforces the model to learn more generalizable features utilizing the implicit prior information present in the data. Through comprehensive analyses, we demonstrate that LSFSL-trained models are less vulnerable to alteration in color schemes, statistical correlations, and adversarial perturbations leveraging the global semantics in the data. Our findings highlight the potential of incorporating relevant priors in few-shot approaches to increase robustness and generalization."}, "cited_paper_content": {"title": "Texture Bias Of Cnns Limits Few-Shot Classification Performance", "abstract": "Accurate image classification given small amounts of labelled data (few-shot classification) remains an open problem in computer vision. In this work we examine how the known texture bias of Convolutional Neural Networks (CNNs) affects few-shot classification performance. Although texture bias can help in standard image classification, in this work we show it significantly harms few-shot classification performance. After correcting this bias we demonstrate state-of-the-art performance on the competitive miniImageNet task using a method far simpler than the current best performing few-shot learning approaches."}, "keywords": ["Convolutional Neural Networks", "texture bias"], "citation_intent": "background"}
{"citing_id": "2303.01235v1", "cited_id": "1708.09603", "section_title": "E. Fast-Ssc Path Metric", "citation": "State-of-the-art high-throughput low-latency decoder architectures are based on the Fast-SSC decoding algorithm #REFR .", "text_before_citation": [], "text_after_citation": ["Since Fast-SSC decoding operates on higher stages in the PFT, the LLR values of the frozen leaf nodes are not calculated.", "Therefore, the PM calculation presented previously is not directly applicable.", "Thus, we derive PM calculation methods for the specialized nodes analogous to Fast Simplified SCL (SSCL) decoding #OTHEREFR .", "The value of PM ( ) is accumulated sequentially, where denotes the index of the leaf node in the pruned PFT and is initialized with PM (0) = 0.", "\u03c0 1 \u03c0 2 \u03c0 3 \u03c0 4 \u03c0 1 -1 \u03c0 2 -1 \u03c0 3 -1 \u03c0 4 -1 max y T x /"], "citing_paper_content": {"title": "Automorphism Ensemble Polar Code Decoders For 6G Urllc", "abstract": "The URLLC scenario in the upcoming 6G standard requires low latency and ultra reliable transmission, i.e., error correction towards ML performance. Achieving near-ML performance is very challenging especially for short block lengths. Polar codes are a promising candidate and already part of the 5G standard. The Successive Cancellation List (SCL) decoding algorithm provides very good error correction performance but at the cost of high computational decoding complexity resulting in large latency and low area and energy efficiency. Recently, Automorphism Ensemble Decoding (AED) gained a lot of attention to improve the error correction capability. In contrast to SCL, AED performs several low-complexity (e.g., SC) decoding in parallel. However, it is an open question whether AED can compete with sophisticated SCL decoders, especially from an implementation perspective in state of the art silicon technologies. In this paper we present an elaborated AED architecture that uses an advanced path metric based candidate selection to reduce the implementation complexity and compare it to state of the art SCL decoders in a 12nm FinFET technology. Our AED implementation outperform state of the art SCL decoders by up to 4.4x in latency, 8.9x in area efficiency and 4.6x in energy efficiency, while providing the same or even better error correction performance."}, "cited_paper_content": {"title": "Polarbear: A 28-Nm Fd-Soi Asic For Decoding Of Polar Codes", "abstract": "Polar codes are a recently proposed class of block codes that provably achieve the capacity of various communication channels. They received a lot of attention as they can do so with low-complexity encoding and decoding algorithms, and they have an explicit construction. Their recent inclusion in a 5G communication standard will only spur more research. However, only a couple of ASICs featuring decoders for polar codes were fabricated, and none of them implements a list-based decoding algorithm. In this paper, we present ASIC measurement results for a fabricated 28-nm CMOS chip that implements two different decoders: the first decoder is tailored toward error-correction performance and flexibility. It supports any code rate as well as three different decoding algorithms: successive cancellation (SC), SC flip, and SC list (SCL). The flexible decoder can also decode both non-systematic and systematic polar codes. The second decoder targets speed and energy efficiency. We present measurement results for the first silicon-proven SCL decoder, where its coded throughput is shown to be of 306.8 Mbps with a latency of 3.34 us and an energy per bit of 418.3 pJ/b at a clock frequency of 721 MHz for a supply of 1.3 V. The energy per bit drops down to 178.1 pJ/b with a more modest clock frequency of 308 MHz, lower throughput of 130.9 Mbps and a reduced supply voltage of 0.9 V. For the other two operating modes, the energy per bit is shown to be of approximately 95 pJ/b. The less flexible high-throughput unrolled decoder can achieve a coded throughput of 9.2 Gbps and a latency of 628 ns for a measured energy per bit of 1.15 pJ/b at 451 MHz."}, "keywords": ["Fast-SSC"], "citation_intent": "method"}
{"citing_id": "2303.02273v1", "cited_id": "1710.00925", "section_title": "Datasets:", "citation": "We use the evaluation methodology followed by prior works #REFR Yang et al., 2019) . Two protocols are used for evaluation.", "text_before_citation": [], "text_after_citation": ["Protocol 1 (LFH1): This protocol uses the BIWI #OTHEREFR Evaluation metrics: We report the Mean Absolute Error (MAE) between the targets (y i ) and predictions (\u0177 i ).", "Let N represent the number of samples, and P represent the number of labels (three in head pose estimation). The MAE is defined as:", "EQUATION", "Network architecture and training parameters: Table 12 summarizes the hyperparameters used for RLEL .", "The learning rate of the decoding matrix D is kept 10\u00d7 higher than the learning rate of the feature extractor."], "citing_paper_content": {"title": "Learning Label Encodings For Deep Regres-Sion", "abstract": "Deep regression networks are widely used to tackle the problem of predicting a continuous value for a given input. Task-specialized approaches for training regression networks have shown significant improvement over generic approaches, such as direct regression. More recently, a generic approach based on regression by binary classification using binary-encoded labels has shown significant improvement over direct regression. The space of label encodings for regression is large. Lacking heretofore have been automated approaches to find a good label encoding for a given application. This paper introduces Regularized Label Encoding Learning (RLEL) for end-to-end training of an entire network and its label encoding. RLEL provides a generic approach for tackling regression. Underlying RLEL is our observation that the search space of label encodings can be constrained and efficiently explored by using a continuous search space of real-valued label encodings combined with a regularization function designed to encourage encodings with certain properties. These properties balance the probability of classification error in individual bits against error correction capability. Label encodings found by RLEL result in lower or comparable errors to manually designed label encodings. Applying RLEL results in 10.9% and 12.4% improvement in Mean Absolute Error (MAE) over direct regression and multiclass classification, respectively. Our evaluation demonstrates that RLEL can be combined with off-the-shelf feature extractors and is suitable across different architectures, datasets, and tasks. Code is available at https://github.com/ubc-aamodt-group/RLEL_regression."}, "cited_paper_content": {"title": "Fine-Grained Head Pose Estimation Without Keypoints", "abstract": "Estimating the head pose of a person is a crucial problem that has a large amount of applications such as aiding in gaze estimation, modeling attention, fitting 3D models to video and performing face alignment. Traditionally head pose is computed by estimating some keypoints from the target face and solving the 2D to 3D correspondence problem with a mean human head model. We argue that this is a fragile method because it relies entirely on landmark detection performance, the extraneous head model and an ad-hoc fitting step. We present an elegant and robust way to determine pose by training a multi-loss convolutional neural network on 300W-LP, a large synthetically expanded dataset, to predict intrinsic Euler angles (yaw, pitch and roll) directly from image intensities through joint binned pose classification and regression. We present empirical tests on common in-the-wild pose benchmark datasets which show state-of-the-art results. Additionally we test our method on a dataset usually used for pose estimation using depth and start to close the gap with state-of-the-art depth pose methods. We open-source our training and testing code as well as release our pre-trained models."}, "keywords": ["prior works Yang"], "citation_intent": "method"}
{"citing_id": "2304.14104v1", "cited_id": "2004.04696", "section_title": "A.6. Ngram-Based Metrics", "citation": "This suggests that METEOR is biased towards measuring surface similarity rather than underlying semantics, consistent with the findings of #REFR who explicitly compare METEOR and BLEURT.", "text_before_citation": ["We additionally provide scores for the METEOR met- ric #OTHEREFR , which uses unigram alignment statistics and incorporates both precision and recall.", "It also uses stemming and synonym matching to provide some robustness relative to changes in the surface forms of semantically similar texts.", "Although METEOR does not ignore most predictions as does BLEU, we find that it underperforms BLEURT in capturing semantic similarity in our setting.", "The base-lines in Table 10 are calculated by (1) using the constant text \"[NAME] meeting with [NAME]\", and (2) randomizing the order of the predictions of the model fine-tuned on our pseudo-labels.", "Both baselines achieve a relatively high METEOR score, while BLEURT decreases significantly and approaches the BLEURT score of the plain captioning model."], "text_after_citation": ["This can also be seen in the qualitative examples in Table 11 of prediction (CLIPCap CC+pHHI) and ground truth pairs from Waldo and Wenda where METEOR and BLEURT differ strongly in magnitude.", "Note we replace [NAME] with the text \"person\" as needed to calculate BLEU and METEOR scores, just as we do for calculating BLEURT scores (see Section B.9)."], "citing_paper_content": {"title": "Learning Human-Human Interactions In Images From Weak Textual Supervision", "abstract": "Interactions between humans are diverse and contextdependent, but previous works have treated them as categorical, disregarding the heavy tail of possible interactions. We propose a new paradigm of learning human-human interactions as free text from a single still image, allowing for flexibility in modeling the unlimited space of situations and relationships between people. To overcome the absence of data labelled specifically for this task, we use knowledge distillation applied to synthetic caption data produced by a large language model without explicit supervision. We show that the pseudo-labels produced by this procedure can be used to train a captioning model to effectively understand human-human interactions in images, as measured by a variety of metrics that measure textual and semantic faithfulness and factual groundedness of our predictions. We further show that our approach outperforms SOTA image captioning and situation recognition models on this task. We will release 1 our code and pseudo-labels along with Waldo and Wenda, a manually-curated test set for still image human-human interaction understanding."}, "cited_paper_content": {"title": "Bleurt: Learning Robust Metrics For Text Generation", "abstract": "Text generation has made significant advances in the last few years. Yet, evaluation metrics have lagged behind, as the most popular choices (e.g., BLEU and ROUGE) may correlate poorly with human judgments. We propose BLEURT, a learned evaluation metric based on BERT that can model human judgments with a few thousand possibly biased training examples. A key aspect of our approach is a novel pre-training scheme that uses millions of synthetic examples to help the model generalize. BLEURT provides state-of-the-art results on the last three years of the WMT Metrics shared task and the WebNLG Competition dataset. In contrast to a vanilla BERT-based approach, it yields superior results even when the training data is scarce and out-of-distribution."}, "keywords": ["underlying semantics", "BLEURT"], "citation_intent": "result"}
{"citing_id": "2305.01333v1", "cited_id": "1206.4657", "section_title": "Oco With Long-Term Constraints", "citation": "Projection-Free Online Learning and Optimization Hazan and Kale #REFR developed the first online projectionfree algorithm, based on the Frank-Wolfe method [5] , that guarantees O(T 3/4 ) regret for non-smooth loss functions.", "text_before_citation": ["However, for the general setting, it is known that it is impossible to simultaneously bound the regret and constraint violation by a sublinear function in T #OTHEREFR . Mahdavi et al.", "#OTHEREFR then considered the special case where g t = g for some fixed function g for all t \u2208 [T ], and they provided an augmented-Lagrangian-based algorithm that achieves O( \u221a T ) regret and O(T 3/4 ) constraint violation. Jenatton et al.", "#OTHEREFR modified the algorithm to obtain O(T max{\u03b2,1\u2212\u03b2} ) regret and O(T 1\u2212\u03b2/2 ) constraint violation where \u03b2 \u2208 (0, 1) is an algorithm parameter.", "Yu et al.", "#OTHEREFR considered the case where g t 's are time-varying but are i."], "text_after_citation": ["Garber and Hazan [6] provided an improved algorithm for smooth and strongly convex loss functions.", "After these works, there has been many results on projection-free algorithms for online convex optimization with no long-term"], "citing_paper_content": {"title": "Projection-Free Online Convex Optimization With Stochastic Constraints", "abstract": "This paper develops projection-free algorithms for online convex optimization with stochastic constraints. We design an online primal-dual projection-free framework that can take any projection-free algorithms developed for online convex optimization with no long-term constraint. With this general template, we deduce sublinear regret and constraint violation bounds for various settings. Moreover, for the case where the loss and constraint functions are smooth, we develop a primal-dual conditional gradient method that achieves O(\u221a T) regret and O(T 3/4) constraint violations. Furthermore, for the setting where the loss and constraint functions are stochastic and strong duality holds for the associated offline stochastic optimization problem, we prove that the constraint violation can be reduced to have the same asymptotic growth as the regret."}, "cited_paper_content": {"title": "Projection-Free Online Learning", "abstract": "The computational bottleneck in applying online learning to massive data sets is usually the projection step. We present efficient online learning algorithms that eschew projections in favor of much more efficient linear optimization steps using the Frank-Wolfe technique. We obtain a range of regret bounds for online convex optimization, with better bounds for specific cases such as stochastic online smooth convex optimization. ::: Besides the computational advantage, other desirable features of our algorithms are that they are parameter-free in the stochastic case and produce sparse decisions. We apply our algorithms to computationally intensive applications of collaborative filtering, and show the theoretical improvements to be clearly visible on standard datasets."}, "keywords": ["Projection-Free Online Learning"], "citation_intent": "method"}
{"citing_id": "2303.10087v1", "cited_id": "2003.08934", "section_title": "Neural Feature Synthesizer", "citation": "To render the color of a given pixel, the original volume rendering formulation in NeRF #REFR is augmented to include the transient colors and densities, and the color of a given image-pixel (\u0108(r)) is computed as a composite of the static and transient components.", "text_before_citation": ["We propose a Neural Feature Synthesizer (NeFeS) model that directly renders dense feature maps of a given viewpoint to refine the predictions of an underlying APR network.", "Similar to NeRF-W #OTHEREFR , our NeFeS architecture uses a base MLP module with static and transient heads that predict the static and transient density (\u03c3 (s) and \u03c3 (\u03c4 ) ) and view-dependent color (c (s) and c (\u03c4 ) ) respectively, given an input 3D position (x) and viewing direction (d).", "We use the frequency encoding #OTHEREFR to encode all 3D positions and view directions.", "The transient head models the colors of the 3D points using an isotropic normal distribution and predicts a view-dependent variance value (\u03b2 2 ) for the transient color distribution."], "text_after_citation": ["Here, r denotes the ray (corresponding to the pixel) on which points are sampled to compute the volume rendering quadrature approximation #OTHEREFR .", "The variances of sampled points along the corresponding ray are also rendered using only the tran- i .", "In addition, the color map adopts exposure-adaptive ACT to compensate for exposure differences between images.", "The final feature map F f usion is the concatenation of rendered RGB and feature map processed by the convolutional fusing layer.", "sient densities (and not the static densities) to obtain a perpixel color variance \u03b2(r) 2 ."], "citing_paper_content": {"title": "Refinement For Absolute Pose Regression With Neural Feature Synthesis", "abstract": "Absolute Pose Regression (APR) methods use deep neural networks to directly regress camera poses from RGB images. Despite their advantages in inference speed and simplicity, these methods still fall short of the accuracy achieved by geometry-based techniques. To address this issue, we propose a new model called the Neural Feature Synthesizer (NeFeS). Our approach encodes 3D geometric features during training and renders dense novel view features at test time to refine estimated camera poses from arbitrary APR methods. Unlike previous APR works that require additional unlabeled training data, our method leverages implicit geometric constraints during test time using a robust feature field. To enhance the robustness of our NeFeS network, we introduce a feature fusion module and a progressive training strategy. Our proposed method improves the state-of-the-art single-image APR accuracy by as much as 54.9% on indoor and outdoor benchmark datasets without additional time-consuming unlabeled data training."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["static", "color"], "citation_intent": "method"}
{"citing_id": "2303.09003v1", "cited_id": "1906.05453", "section_title": "19:", "citation": "Get \u03c0 i (k) by #REFR The numerical simulation is implemented on MATLAB R2020b, and Fig. 10 is one example of the simulation scenarios.", "text_before_citation": [], "text_after_citation": ["And we also perform the hardwarein-the-loop simulations, which are described in Section 7.3.", "Assuming that each target can be tracked by a maximum of 2 UAVs simultaneously, set the parameters related to grid visiting requirements as w p1 = 1, w p2 = 1.2, w p3 = 0.8 and the related parameters of sensor detection performance as \u03b1 = 1.1, \u03b2 = 0.8.", "Considering the randomness of the experiment, including the initial positions of the target and the UAV, and the random movement of the target, we performed 50 repetitions of each set of experiments with 1500 simulation steps to get average results."], "citing_paper_content": {"title": "Integrated Design Of Cooperative Area Coverage And Target Tracking With Multi-Uav System", "abstract": "This paper systematically studies the cooperative area coverage and target tracking problem of multiple-unmanned aerial vehicles (multi-UAVs). The problem is solved by decomposing into three sub-problems: information fusion, task assignment, and multi-UAV behavior decision-making. Specifically, in the information fusion process, we use the maximum consistency protocol to update the joint estimation states of multi-targets (JESMT) and the area detection information. The area detection information is represented by the equivalent visiting time map (EVTM), which is built based on the detection probability and the actual visiting time of the area. Then, we model the task assignment problem of multi-UAV searching and tracking multi-targets as a network flow model with upper and lower flow bounds. An algorithm named task assignment minimum-cost maximum-flow (TAMM) is proposed. Cooperative behavior decision-making uses Fisher information as the mission reward to obtain the optimal tracking action of the UAV. Furthermore, a coverage behavior decision-making algorithm based on the anti-flocking method is designed for those UAVs assigned the coverage task. Finally, a distributed multi-UAV cooperative area coverage and target tracking algorithm is designed, which integrates information fusion, task assignment, and behavioral decision-making. Numerical and hardware-in-the-loop simulation results show that the proposed method can achieve persistent area coverage and cooperative target tracking."}, "cited_paper_content": {"title": "Coordinated Path Following Control Of Fixed-Wing Unmanned Aerial Vehicles", "abstract": "In this paper, we investigate the problem of coordinated path following for fixed-wing UAVs with speed constraints in 2D plane. The objective is to steer a fleet of UAVs along the path(s) while achieving the desired sequenced inter-UAV arc distance. In contrast to the previous coordinated path following studies, we are able through our proposed hybrid control law to deal with the forward speed and the angular speed constraints of fixed-wing UAVs. More specifically, the hybrid control law makes all the UAVs work at two different levels: those UAVs whose path following errors are within an invariant set (i.e., the designed coordination set) work at the coordination level; and the other UAVs work at the single-agent level. At the coordination level, we prove that even with speed constraints, the proposed control law can make sure the path following errors reduce to zero, while the desired arc distances converge to the desired value. At the single-agent level, the convergence analysis for the path following error entering the coordination set is provided. We develop a hardware-in-the-loop simulation testbed of the multi-UAV system by using actual autopilots and the X-Plane simulator. The effectiveness of the proposed approach is corroborated with both MATLAB and the testbed."}, "keywords": ["simulation scenarios", "MATLAB R2020b"], "citation_intent": "method"}
{"citing_id": "2303.10774v2", "cited_id": "2002.03754", "section_title": "Discussion", "citation": "Limitations First, similar to other optimization-based attribute discovery approaches #REFR , [43] , there is no guarantee that all prevalent factors are captured, though our controlled empirical studies clearly demonstrate the efficacy of xGA over existing approaches.", "text_before_citation": ["We introduced the first cross-GAN auditing framework, which utilizes a novel optimization technique to jointly infer common, novel and missing attributes for a client GAN w.r.t any reference GAN.", "Through a large suite of datasets and GAN models, we demonstrate that the proposed method (i) consistently leads to higher quality (disentangled & diverse) attributes, (ii) effectively reveals shared attributes even across challenging distribution shifts, and (iii) accurately identifies the novel/missing attributes in our controlled experiments (i.e., known ground truth)."], "text_after_citation": ["Second, while using an external feature space enhances the performance of attribute discovery, this becomes an additional component that must be tuned.", "While we found advBN ResNet-50 to be a reasonable choice for a variety of face datasets (and AFHQ), a more systematic solution will expand the utility of our approach to other applications. A.", "Ablation study First, we investigate the effect of the \u03bb b parameter on KLIEP loss that allows us to discover novel attributes.", "In addition to KLIEP loss presented in the main text, we analyze a model trained with simple log loss used to train binary classifiers to predict the likelihood of a given sample."], "citing_paper_content": {"title": "Cross-Gan Auditing: Unsupervised Identification Of Attribute Level Similarities And Differences Between Pretrained Generative Models", "abstract": "Figure 1. We introduce (xGA) an approach for fully unsupervised cross-GAN auditing and validation. Given two pre-trained GANs (Reference & Client), xGA evaluates the client by identifying three types of semantic attributes-(a) Common: those that exist in both models, (b) Novel: those only present in the client and (c) Missing: those that only exist in the reference. On the right, we show results across multiple studies, that among others include notable shifts in distribution between the Reference (CelebA) to Client (Toon, Disney, Met Faces). xGA also lends itself easily to comparing models with different properties on the same dataset as shown on the bottom right for StyleGAN3-T vs. StyleGAN-R. And CelebA-M is a control dataset we create that does not contain glasses, ties and smiles."}, "cited_paper_content": {"title": "Unsupervised Discovery Of Interpretable Directions In The Gan Latent Space", "abstract": "The latent spaces of typical GAN models often have semantically meaningful directions. Moving in these directions corresponds to human-interpretable image transformations, such as zooming or recoloring, enabling a more controllable generation process. However, the discovery of such directions is currently performed in a supervised manner, requiring human labels, pretrained models, or some form of self-supervision. These requirements can severely limit a range of directions existing approaches can discover. In this paper, we introduce an unsupervised method to identify interpretable directions in the latent space of a pretrained GAN model. By a simple model-agnostic procedure, we find directions corresponding to sensible semantic manipulations without any form of (self-)supervision. Furthermore, we reveal several non-trivial findings, which would be difficult to obtain by existing methods, e.g., a direction corresponding to background removal. As an immediate practical benefit of our work, we show how to exploit this finding to achieve a new state-of-the-art for the problem of saliency detection."}, "keywords": ["optimization-based attribute discovery"], "citation_intent": "result"}
{"citing_id": "2304.09575v1", "cited_id": "1806.04167", "section_title": "C. Results", "citation": "In the stir tank example, a relatively small NN approximates the dataset well (more than 99 % of predictions are valid solutions to (5) on test data), which is in line with numerical results in #REFR .", "text_before_citation": ["First, we show the training effort and approximation quality achieved.", "Second, performance in terms of computation time is compared with online optimization.", "Finally, we show that our method is safe in all three examples, while naively (always) applying the NN leads to constraint violations in closed loop.", "1) Dataset generation and NN training: We now describe the size and computation times of generated datasets, give an indication of NN sizes, and provide NN training times for all three systems.", "To gain some intuition into how well each approximator generalizes, it is tested on 10 5 points not used during training by forward simulating and checking whether the NN provides a feasible solution to #OTHEREFR . The results are summarized in Table I ."], "text_after_citation": ["For the ten-state quadcopter example, a dataset corresponding to approximately 5 gridpoints per state dimension suffices to fit a relatively large approximator, which yields a feasible solution to the MPC problem in over 70 % of initial conditions.", "Contributing factors to this could be that the quadcopter system is almost linear (only non-linearity is the tan(\u03c6) in the velocity derivative), only loosely coupled in three Cartesian directions by the MPC cost function, and has few constraints that become active.", "The chain mass system, on the other hand, is \"very\" non-linear and chaotic, which makes the approximation problem inherently harder.", "This is reflected in a large dataset and approximator, which yield a feasible sequence for less than 10 % of initial conditions during testing.", "The total time required to evaluate samples of the MPC on random initial conditions is the dominant time factor offline. However, this was executed in parallel."], "citing_paper_content": {"title": "Approximate Non-Linear Model Predictive Control With Safety-Augmented Neural Networks", "abstract": "Model predictive control (MPC) achieves stability and constraint satisfaction for general nonlinear systems, but requires computationally expensive online optimization. This paper studies approximations of such MPC controllers via neural networks (NNs) to achieve fast online evaluation. We propose safety augmentation that yields deterministic guarantees for convergence and constraint satisfaction despite approximation inaccuracies. We approximate the entire input sequence of the MPC with NNs, which allows us to verify online if it is a feasible solution to the MPC problem. We replace the NN solution by a safe candidate based on standard MPC techniques whenever it is infeasible or has worse cost. Our method requires a single evaluation of the NN and forward integration of the input sequence online, which is fast to compute on resource-constrained systems. The proposed control framework is illustrated on three non-linear MPC benchmarks of different complexity, demonstrating computational speedups orders of magnitudes higher than online optimization. In the examples, we achieve deterministic safety through the safetyaugmented NNs, where naive NN implementation fails."}, "cited_paper_content": {"title": "Learning An Approximate Model Predictive Controller With Guarantees", "abstract": "A supervised learning framework is proposed to approximate a model predictive controller (MPC) with reduced computational complexity and guarantees on stability and constraint satisfaction. The framework can be used for a wide class of nonlinear systems. Any standard supervised learning technique (e.g. neural networks) can be employed to approximate the MPC from samples. In order to obtain closed-loop guarantees for the learned MPC, a robust MPC design is combined with statistical learning bounds. The MPC design ensures robustness to inaccurate inputs within given bounds, and Hoeffding's Inequality is used to validate that the learned MPC satisfies these bounds with high confidence. The result is a closed-loop statistical guarantee on stability and constraint satisfaction for the learned MPC. The proposed learning-based MPC framework is illustrated on a nonlinear benchmark problem, for which we learn a neural network controller with guarantees."}, "keywords": ["predictions"], "citation_intent": "result"}
{"citing_id": "2303.15417v1", "cited_id": "1612.02177", "section_title": "S2. Results From Various Deblurring Methods", "citation": "In Table S2 , we additionally compare the results from another widely used deblurring method, DeepDeblur #REFR , as the final mesh estimation results might be dependent on the performance of deblurring methods.", "text_before_citation": ["In Tables 2 and 3 in our main manuscript, we compared our BlurHandNet with the combination of state-of-the-art 3D hand mesh estimation methods #OTHEREFR and offthe-shelf deblurring method #OTHEREFR ."], "text_after_citation": ["Please note that NAFNet [3] is a deblurring method that we used in the main manuscript.", "Our BlurHandNet still outperforms the case when we use Deep-Deblur #OTHEREFR as the deblurring method.", "The results again demonstrate that utilizing temporal information is useful rather than simply adopting deblurring methods. Table S2 . Comparison results on various deblurring methods #OTHEREFR .", "Instead of adopting deblurring methods, the proposed BlurHandNet (last row) directly utilizes temporal information from a single blurry image."], "citing_paper_content": {"title": "Recovering 3D Hand Mesh Sequence From A Single Blurry Image: A New Dataset And Temporal Unfolding", "abstract": "Hands, one of the most dynamic parts of our body, suffer from blur due to their active movements. However, previous 3D hand mesh recovery methods have mainly focused on sharp hand images rather than considering blur due to the absence of datasets providing blurry hand images. We first present a novel dataset BlurHand, which contains blurry hand images with 3D groundtruths. The BlurHand is constructed by synthesizing motion blur from sequential sharp hand images, imitating realistic and natural motion blurs. In addition to the new dataset, we propose BlurHandNet, a baseline network for accurate 3D hand mesh recovery from a blurry hand image. Our BlurHandNet unfolds a blurry input image to a 3D hand mesh sequence to utilize temporal information in the blurry input image, while previous works output a static single hand mesh. We demonstrate the usefulness of BlurHand for the 3D hand mesh recovery from blurry images in our experiments. The proposed BlurHandNet produces much more robust results on blurry images while generalizing well to in-the-wild images. The training codes and BlurHand dataset are available at https://github.com/JaehaKim97/BlurHand_RELEASE."}, "cited_paper_content": {"title": "Deep Multi-Scale Convolutional Neural Network For Dynamic Scene Deblurring", "abstract": "Non-uniform blind deblurring for general dynamic scenes is a challenging computer vision problem as blurs arise not only from multiple object motions but also from camera shake, scene depth variation. To remove these complicated motion blurs, conventional energy optimization based methods rely on simple assumptions such that blur kernel is partially uniform or locally linear. Moreover, recent machine learning based methods also depend on synthetic blur datasets generated under these assumptions. This makes conventional deblurring methods fail to remove blurs where blur kernel is difficult to approximate or parameterize (e.g. object motion boundaries). In this work, we propose a multi-scale convolutional neural network that restores sharp images in an end-to-end manner where blur is caused by various sources. Together, we present multi-scale loss function that mimics conventional coarse-to-fine approaches. Furthermore, we propose a new large-scale dataset that provides pairs of realistic blurry image and the corresponding ground truth sharp image that are obtained by a high-speed camera. With the proposed model trained on this dataset, we demonstrate empirically that our method achieves the state-of-the-art performance in dynamic scene deblurring not only qualitatively, but also quantitatively."}, "keywords": ["DeepDeblur", "deblurring method"], "citation_intent": "result"}
{"citing_id": "2304.09019v1", "cited_id": "2002.09270", "section_title": "C. Uplink Transmission", "citation": "These coefficients, can thus, be easily computed and stored at the CPU #REFR . We now explain the two-step decoding in detail.", "text_before_citation": ["The RF chain output is then fed to the dynamic-resolution ADCs, whose quantized and noisy output, based on the Bussgang model #OTHEREFR , is given as follows: #OTHEREFR .", "Principles of two-layer decoding: The CF system considered herein, as shown in Fig. 2 , employs two-layer decoding to mitigate IUI.", "In the first-stage, each AP combines its received signal by using local channel estimates, which only partially mitigate the IUI.", "To mitigate the residual IUI, all APs sends their locally-combined received signal to the CPU, which performs the secondlayer LSFD.", "The CPU computes the LSFD weights based on the large-scale fading coefficients, which remains constant for 100s of coherence intervals #OTHEREFR , #OTHEREFR ."], "text_after_citation": ["The mth AP first uses channel estimate\u0125 mk [\u03bb] to combine the distorted received signal as", "y ADC,m [n] = A m y RF,m [n]+n ADC,m [n] = A m K k=1 h mk [n]s RF,k [n]+\u03b7 RF,m [n]+z m [n] +n ADC,m [n].", "The matrix A m = diag{\u03b9 m,1 , \u2022 \u2022 \u2022 , \u03b9 m,N },", "B m = A m (I N \u2212 A m ) and S m [n] = diag(E{y RF,m [n]y H RF,m [n]|h mk [n]})", "EQUATION"], "citing_paper_content": {"title": "Hardware-Impaired Rician-Faded Cell-Free Massive Mimo Systems With Channel Aging", "abstract": "We study the impact of channel aging on the uplink of a cell-free (CF) massive multiple-input multiple-output (mMIMO) system by considering i) spatially-correlated Rician-faded channels; ii) hardware impairments at the access points and user equipments (UEs); and iii) two-layer large-scale fading decoding (LSFD). We first derive a closed-form spectral efficiency (SE) expression for this system, and later propose two novel optimization techniques to optimize the non-convex SE metric by exploiting the minorization-maximization (MM) method. The first one requires a numerical optimization solver, and has a high computation complexity. The second one with closed-form transmit power updates, has a trivial computation complexity. We numerically show that i) the two-layer LSFD scheme effectively mitigates the interference due to channel aging for both low-and high-velocity UEs; and ii) increasing the number of AP antennas does not mitigate the SE deterioration due to channel aging. We numerically characterize the optimal pilot length required to maximize the SE for various UE speeds. We also numerically show that the proposed closed-form MM optimization yields the same SE as that of the first technique, which requires numerical solver, and that too with a much reduced time-complexity. Index Terms Cell-free, channel aging, hardware impairments, minorization-maximization (MM), Rician fading."}, "cited_paper_content": {"title": "Joint Power Control And Lsfd For Wireless-Powered Cell-Free Massive Mimo", "abstract": "This paper considers wireless uplink information and downlink power transfer in cell-free massive multiple-input multiple-output systems. The single-antenna user equipments (UEs) utilize the energy harvested in the downlink to transmit uplink pilot and information signals to the multiple-antenna access points (APs). We consider Rician fading and maximum ratio processing based on either linear minimum mean-squared error (LMMSE) or least-squares (LS) channel estimation. We derive the average harvested energy by using a practical non-linear energy harvesting circuit model for both coherent and non-coherent transmission schemes. Furthermore, the uplink spectral efficiency (SE) is derived for all the considered methods and the max-min fairness problem is cast where the optimization variables are the AP and UE power control coefficients together with the large-scale fading decoding vectors. The objective is to maximize the minimum SE of the UEs' under APs' and UEs' transmission power constraints. A novel alternating optimization algorithm with guaranteed convergence and improvement at each step is proposed to solve the highly-coupled non-convex problem."}, "keywords": ["coefficients"], "citation_intent": "method"}
{"citing_id": "2303.16990v1", "cited_id": "1906.02549", "section_title": "Supervised Spatio-Temporal Grounding.", "citation": "This conceptually differs from previous works #REFR that typically use short videos of around 5-10 seconds.", "text_before_citation": ["This problem has been addressed by various approaches Tube-DETR #OTHEREFR , STCAT #OTHEREFR , STVGBert #OTHEREFR , STVGBert #OTHEREFR , STGVT #OTHEREFR , STGRN #OTHEREFR .", "These methods rely on proposal networks such as Faster R-CNN #OTHEREFR or MDETR #OTHEREFR to predict bounding box coordinates for learning text-toregion interaction.", "All those approaches rely on supervised training with the human-annotated sentence and bounding box supervision, provided, e.g., by datasets such as Vid-STG #OTHEREFR and HC-STVG #OTHEREFR .", "While those datasets provide a temporal aspect, temporal detection is usually limited to identifying the start and end frame of a single action in a video.", "Compared to that, an untrimmed setting usually comprises multiple actions in a longer video that can be separated by longer background sequences."], "text_after_citation": [], "citing_paper_content": {"title": "What, When, And Where? -Self-Supervised Spatio-Temporal Grounding In Untrimmed Multi-Action Videos From Narrated Instructions", "abstract": "Spatio-temporal grounding describes the task of localizing events in space and time, e.g., in video data, based on verbal descriptions only. Models for this task are usually trained with human-annotated sentences and bounding box supervision. This work addresses this task from a multimodal supervision perspective, proposing a framework for spatio-temporal action grounding trained on loose video and subtitle supervision only, without human annotation. To this end, we combine local representation learning, which focuses on leveraging fine-grained spatial information, with a global representation encoding that captures higher-level representations and incorporates both in a joint approach. To evaluate this challenging task in a real-life setting, a new benchmark dataset is proposed providing dense spatiotemporal grounding annotations in long, untrimmed, multiaction instructional videos for over 5K events. We evaluate the proposed approach and other methods on the proposed and standard downstream tasks showing that our method improves over current baselines in various settings, including spatial, temporal, and untrimmed multi-action spatiotemporal grounding."}, "cited_paper_content": {"title": "Weakly-Supervised Spatio-Temporally Grounding Natural Sentence In Video", "abstract": "In this paper, we address a novel task, namely weakly-supervised spatio-temporally grounding natural sentence in video. Specifically, given a natural sentence and a video, we localize a spatio-temporal tube in the video that semantically corresponds to the given sentence, with no reliance on any spatio-temporal annotations during training. First, a set of spatio-temporal tubes, referred to as instances, are extracted from the video. We then encode these instances and the sentence using our proposed attentive interactor which can exploit their fine-grained relationships to characterize their matching behaviors. Besides a ranking loss, a novel diversity loss is introduced to train the proposed attentive interactor to strengthen the matching behaviors of reliable instance-sentence pairs and penalize the unreliable ones. Moreover, we also contribute a dataset, called VID-sentence, based on the ImageNet video object detection dataset, to serve as a benchmark for our task. Extensive experimental results demonstrate the superiority of our model over the baseline approaches."}, "keywords": ["short videos"], "citation_intent": "result"}
{"citing_id": "2304.06653v1", "cited_id": "1703.01488", "section_title": "Bayesian Probabilistic Topic Models", "citation": "ProdLDA #REFR based on the VAE has better performance in topic coherence and computational efficiency than original LDA.", "text_before_citation": ["BPTMs are usually built on generative probabilistic models and can generate a document with a structure of latent variables sampled from prespecified distributions (Dirichlet for LDA). These latent variables are used to obtain topics. The BPTMs are learned through a Bayesian inference process.", "LDA performs the inference using Monte Carlo Markov chain sampling #OTHEREFR or Gibbs sampling.", "Neural topic models (NTMs) #OTHEREFR incorporate neural components for Bayesian inference, and most NTMs are based on the variational autoencoder (VAE) inference framework #OTHEREFR ).", "The VAE adopts an encoder-decoder structure: the encoder trains a neural inference network to directly map a BoW representation of documents into a continuous latent embedding, while the decoder reconstructs the BoW from the latent embedding."], "text_after_citation": ["More recently, powerful contextualized embeddings were also incorporated into NTMs, which further improved the topic coherence compared to conventional NTMs that use bag-of-words (BoW) as document representations #OTHEREFR .", "However, most of the NTMs are based on the variational autoencoder inference framework #OTHEREFR , which suffers from hyperparameter tuning, computational complexity #OTHEREFR , and an inability to choose the proper number of topics."], "citing_paper_content": {"title": "G2T: A Simple But Versatile Framework For Topic Modeling Based On Pretrained Language Model And Community Detection", "abstract": "It has been reported that clustering-based topic models, which cluster high-quality sentence embeddings with an appropriate word selection method, can generate better topics than generative probabilistic topic models. However, these approaches suffer from the inability to select appropriate parameters and incomplete models that overlook the quantitative relation between words with topics and topics with text. To solve these issues, we propose graph to topic (G2T), a simple but effective framework for topic modelling. The framework is composed of four modules. First, document representation is acquired using pretrained language models. Second, a semantic graph is constructed according to the similarity between document representations. Third, communities in document semantic graphs are identified, and the relationship between topics and documents is quantified accordingly. Fourth, the word-topic distribution is computed based on a variant of TFIDF. Automatic evaluation suggests that G2T achieved state-of-the-art performance on both English and Chinese documents with different lengths. Human judgements demonstrate that G2T can produce topics with better interpretability and coverage than baselines. In addition, G2T can not only determine the topic number automatically but also give the probabilistic distribution of words in topics and topics in documents. Finally, G2T is publicly available, and the distillation experiments provide instruction on how it works."}, "cited_paper_content": {"title": "Autoencoding Variational Inference For Topic Models", "abstract": "Topic models are one of the most popular methods for learning representations of ::: text, but a major challenge is that any change to the topic model requires mathematically ::: deriving a new inference algorithm. A promising approach to address ::: this problem is autoencoding variational Bayes (AEVB), but it has proven diffi- ::: cult to apply to topic models in practice. We present what is to our knowledge the ::: first effective AEVB based inference method for latent Dirichlet allocation (LDA), ::: which we call Autoencoded Variational Inference For Topic Model (AVITM). This ::: model tackles the problems caused for AEVB by the Dirichlet prior and by component ::: collapsing. We find that AVITM matches traditional methods in accuracy ::: with much better inference time. Indeed, because of the inference network, we ::: find that it is unnecessary to pay the computational cost of running variational ::: optimization on test data. Because AVITM is black box, it is readily applied ::: to new topic models. As a dramatic illustration of this, we present a new topic ::: model called ProdLDA, that replaces the mixture model in LDA with a product ::: of experts. By changing only one line of code from LDA, we find that ProdLDA ::: yields much more interpretable topics, even if LDA is trained via collapsed Gibbs ::: sampling."}, "keywords": ["topic coherence"], "citation_intent": "method"}
{"citing_id": "2304.12228v1", "cited_id": "1710.09412", "section_title": "Generate Harder Negative Samples", "citation": "HeCo MU MixUp #REFR is proposed to efficiently improve results in supervised learning by adding arbitrary two samples to create a new one.", "text_before_citation": ["First, we train D to identify embeddings from two views as positives and that generated from G as negatives.", "Then, we train G to generate samples to fool D.", "The two steps are alternated for some interactions to make D and G trained.", "(2) We utilize a well-trained G to generate samples, which can be viewed as the new negative samples with high quality.", "Then, we continue to train HeCo with the newly generated and original negative samples for some epochs."], "text_after_citation": ["MoCHi #OTHEREFR introduces this strategy into contrastive learning , who mixes the hard negatives to make harder negatives.", "Inspired by them, we bring this strategy into HIN field for the first time.", "We can get cosine similarities between node i and nodes from N i during calculating eq.(11), and sort them in the descending order.", "Then, we select first top k negative samples as the hardest negatives, and randomly add them to create new k negatives, which are involved in training.", "It is worth mentioning that there are no learnable parameters in this version, which is very efficient."], "citing_paper_content": {"title": "Hierarchical Contrastive Learning Enhanced Heterogeneous Graph Neural Network", "abstract": "Heterogeneous graph neural networks (HGNNs) as an emerging technique have shown superior capacity of dealing with heterogeneous information network (HIN). However, most HGNNs follow a semi-supervised learning manner, which notably limits their wide use in reality since labels are usually scarce in real applications. Recently, contrastive learning, a self-supervised method, becomes one of the most exciting learning paradigms and shows great potential when there are no labels. In this paper, we study the problem of self-supervised HGNNs and propose a novel co-contrastive learning mechanism for HGNNs, named HeCo. Different from traditional contrastive learning which only focuses on contrasting positive and negative samples, HeCo employs cross-view contrastive mechanism. Specifically, two views of a HIN (network schema and meta-path views) are proposed to learn node embeddings, so as to capture both of local and high-order structures simultaneously. Then the cross-view contrastive learning, as well as a view mask mechanism, is proposed, which is able to extract the positive and negative embeddings from two views. This enables the two views to collaboratively supervise each other and finally learn high-level node embeddings. Moreover, to further boost the performance of HeCo, two additional methods are designed to generate harder negative samples with high quality. The essence of HeCo is to make positive samples from different views close to each other by cross-view contrast, and learn the factors invariant to two proposed views. However, besides the invariant factors, view-specific factors complementally provide the diverse structure information between different nodes, which also should be contained into the final embeddings. Therefore, we need to further explore each view independently and propose a modified model, called HeCo++. Specifically, HeCo++ conducts hierarchical contrastive learning, including cross-view and intra-view contrasts, which aims to enhance the mining of respective structures. Extensive experiments conducted on a variety of real-world networks show the superior performance of the proposed methods over the state-of-the-arts."}, "cited_paper_content": {"title": "Mixup: Beyond Empirical Risk Minimization", "abstract": "Large deep neural networks are powerful, but exhibit undesirable behaviors such as memorization and sensitivity to adversarial examples. In this work, we propose mixup, a simple learning principle to alleviate these issues. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularizes the neural network to favor simple linear behavior in-between training examples. Our experiments on the ImageNet-2012, CIFAR-10, CIFAR-100, Google commands and UCI datasets show that mixup improves the generalization of state-of-the-art neural network architectures. We also find that mixup reduces the memorization of corrupt labels, increases the robustness to adversarial examples, and stabilizes the training of generative adversarial networks."}, "keywords": ["HeCo MU MixUp", "arbitrary two samples"], "citation_intent": "method"}
{"citing_id": "2303.07909v1", "cited_id": "1905.01164", "section_title": "Model Training With A Single Image", "citation": "SinGAN #REFR can generate similar images and perform well on multiple tasks (e.g., image editing) after training on a single image.", "text_before_citation": ["The success of diffusion models in the text-to-image synthesis area relies on a large amount of training samples.", "An interesting topic is how to train a generative model with a single image, e.g., SinGAN #OTHEREFR ."], "text_after_citation": ["There are also research on diffusion models with a single image #OTHEREFR , #OTHEREFR , #OTHEREFR .", "Single Denoising Diffusion Models (SinDDM) #OTHEREFR proposes a hierarchical diffusion model inspired by the multi-scale SinGAN #OTHEREFR .", "The convolutional denoiser is trained on various scales of images which are corrupted by multiple levels of noise.", "Compared to SDEdit #OTHEREFR , SinDDM #OTHEREFR can generating images with different dimensions.", "By contrast, Single-image Diffusion Model (Sin-Diffusion) #OTHEREFR is trained on a single image at a single scale, avoiding the accumulation of errors."], "citing_paper_content": {"title": "Text-To-Image Diffusion Model In Generative Ai: A Survey", "abstract": "This survey reviews text-to-image diffusion models in the context that diffusion models have emerged to be popular for a wide range of generative tasks. As a self-contained work, this survey starts with a brief introduction of how a basic diffusion model works for image synthesis, followed by how condition or guidance improves learning. Based on that, we present a review of state-of-the-art methods on text-conditioned image synthesis, i.e. text-to-image. We further summarize applications beyond text-to-image generation: text-guided creative generation and text-guided image editing. Beyond the progress made so far, we discuss existing challenges and promising future directions."}, "cited_paper_content": {"title": "Singan: Learning A Generative Model From A Single Natural Image", "abstract": "We introduce SinGAN, an unconditional generative model that can be learned from a single natural image. Our model is trained to capture the internal distribution of patches within the image, and is then able to generate high quality, diverse samples that carry the same visual content as the image. SinGAN contains a pyramid of fully convolutional GANs, each responsible for learning the patch distribution at a different scale of the image. This allows generating new samples of arbitrary size and aspect ratio, that have significant variability, yet maintain both the global structure and the fine textures of the training image. In contrast to previous single image GAN schemes, our approach is not limited to texture images, and is not conditional (i.e. it generates samples from noise). User studies confirm that the generated samples are commonly confused to be real images. We illustrate the utility of SinGAN in a wide range of image manipulation tasks."}, "keywords": ["similar images", "(e.g., image editing"], "citation_intent": "background"}
{"citing_id": "2303.08240v1", "cited_id": "1907.10844", "section_title": "Surface", "citation": "Evaluation Metrics: We use three widely adopted metrics in previous work to evaluate our performance: Chamfer Distance (CD), Point-to-Surface Distance (P2F) w.r.t ground truth meshes, and Uniformity Score #REFR . For these metrics, a smaller value means better performance.", "text_before_citation": ["The outputs of the feature extractor block are the point-wise features F 0 and the global shape code G.", "Note that the design of feature extractor is not the major contribution of this paper and we can exploit any other suitable networks.", "We then feed these outputs along with the original position of the points P 0 into stacks of upsampling blocks to generate denser point clouds.", "To upsample the point cloud 4 times, we arbitrarily set two upsampler blocks with upscale ratios of 1 and 4, respectively.", "Note that other combinations of upscale ratios and the number of upsampler blocks are also feasible."], "text_after_citation": ["Training Detail: To train this network, we use 2 Tesla V100 GPUs.", "We set the batch size to 16 and the total epoch number to 150.", "Besides, we use Adam as the optimization function with a learning rate of 0.0005 at the beginning, and we decrease the learning rate by a factor of 0.5 every 50 epochs.", "Experiment Results: Table 1 shows the quantitative upsampling results on the PU1K dataset.", "We find that our algorithm achieves the best performance over all its counterparts with large improvements."], "citing_paper_content": {"title": "Parametric Surface Constrained Upsampler Network For Point Cloud", "abstract": "Designing a point cloud upsampler, which aims to generate a clean and dense point cloud given a sparse point representation, is a fundamental and challenging problem in computer vision. A line of attempts achieves this goal by establishing a point-to-point mapping function via deep neural networks. However, these approaches are prone to produce outlier points due to the lack of explicit surface-level constraints. To solve this problem, we introduce a novel surface regularizer into the upsampler network by forcing the neural network to learn the underlying parametric surface represented by bicubic functions and rotation functions, where the new generated points are then constrained on the underlying surface. These designs are integrated into two different networks for two tasks that take advantages of upsampling layers-point cloud upsampling and point cloud completion for evaluation. The state-of-the-art experimental results on both tasks demonstrate the effectiveness of the proposed method. The implementation code will be available at https://github.com/corecai163/PSCU."}, "cited_paper_content": {"title": "Pu-Gan: A Point Cloud Upsampling Adversarial Network", "abstract": "Point clouds acquired from range scans are often sparse, noisy, and non-uniform. This paper presents a new point cloud upsampling network called PU-GAN, which is formulated based on a generative adversarial network (GAN), to learn a rich variety of point distributions from the latent space and upsample points over patches on object surfaces. To realize a working GAN network, we construct an up-down-up expansion unit in the generator for upsampling point features with error feedback and self-correction, and formulate a self-attention unit to enhance the feature integration. Further, we design a compound loss with adversarial, uniform and reconstruction terms, to encourage the discriminator to learn more latent patterns and enhance the output point distribution uniformity. Qualitative and quantitative evaluations demonstrate the quality of our results over the state-of-the-arts in terms of distribution uniformity, proximity-to-surface, and 3D reconstruction quality."}, "keywords": ["ground truth meshes"], "citation_intent": "method"}
{"citing_id": "2304.09059v2", "cited_id": "1611.10080", "section_title": "B. Backbone And Baseline", "citation": "In this paper, we choose the commonly used WideResNet-38 #REFR as the backbone network for feature extraction, which is used in ablation studies and performance comparisons with state-of-the-art methods.", "text_before_citation": ["Backbone."], "text_after_citation": ["In particular, to ensure a high-resolution output scale for feature maps and not have too much parameter growth, we replace the traditional convolutions of the last two blocks with dilated/atrous convolutions as in #OTHEREFR , such that the spatial scale of output feature maps is 1/8 of the input image.", "All our implementations are deployed on the PyTorch #OTHEREFR deep learning platform. Baseline.", "Our baseline is the single-stage semantic segmentation #OTHEREFR model, which is based on the self-supervised training manner with only image-level class labels.", "In the training stage, the backbone network first generates the initial CAMs.", "Then, a mask refinement module named PAMR is used to refine CAMs as pseudo ground-truth to complete the selfsupervised training process."], "citing_paper_content": {"title": "Coupling Global Context And Local Contents For Weakly-Supervised Semantic Segmentation", "abstract": "Thanks to the advantages of the friendly annotations and the satisfactory performance, Weakly-Supervised Semantic Segmentation (WSSS) approaches have been extensively studied. Recently, the single-stage WSSS was awakened to alleviate problems of the expensive computational costs and the complicated training procedures in multi-stage WSSS. However, results of such an immature model suffer from problems of background incompleteness and object incompleteness. We empirically find that they are caused by the insufficiency of the global object context and the lack of the local regional contents, respectively. Under these observations, we propose a single-stage WSSS model with only the image-level class label supervisions, termed as Weakly-Supervised Feature Coupling Network (WS-FCN), which can capture the multi-scale context formed from the adjacent feature grids, and encode the fine-grained spatial information from the low-level features into the high-level ones. Specifically, a flexible context aggregation module is proposed to capture the global object context in different granular spaces. Besides, a semantically consistent feature fusion module is proposed in a bottom-up parameter-learnable fashion to aggregate the finegrained local contents. Based on these two modules, WS-FCN lies in a self-supervised end-to-end training fashion. Extensive experimental results on the challenging PASCAL VOC 2012 and MS COCO 2014 demonstrate the effectiveness and efficiency of WS-FCN, which can achieve state-of-the-art results by 65.02% and 64.22% mIoU on PASCAL VOC 2012 val set and test set, 34.12% mIoU on MS COCO 2014 val set, respectively. The code and weight have been released at: WS-FCN."}, "cited_paper_content": {"title": "Wider Or Deeper: Revisiting The Resnet Model For Visual Recognition", "abstract": "The community has been going deeper and deeper in designing one cutting edge network after another, yet some works are there suggesting that we may have gone too far in this dimension. Some researchers unravelled a residual network into an exponentially wider one, and assorted the success of residual networks to fusing a large amount of relatively shallow models. Since some of their early claims are still not settled, we in this paper dig more on this topic, i.e., the unravelled view of residual networks. Based on that, we try to find a good compromise between the depth and width. Afterwards, we walk through a typical pipeline of developing a deep-learning-based algorithm. We start from a group of relatively shallow networks, which perform as well or even better than the current (much deeper) state-of-the-art models on the ImageNet classification dataset. Then, we initialize fully convolutional networks (FCNs) using our pre-trained models, and tune them for semantic image segmentation. Results show that the proposed networks, as pre-trained features, can boost existing methods a lot. Even without exhausting the sophistical techniques to improve the classic FCN model, we achieve comparable results with the best performers on four widely-used datasets, i.e., Cityscapes, PASCAL VOC, ADE20k and PASCAL-Context. The code and pre-trained models are released for public access 1 ."}, "keywords": ["backbone network", "WideResNet-38"], "citation_intent": "method"}
{"citing_id": "2305.02885v1", "cited_id": "1606.06160", "section_title": "Svhn", "citation": "We used two instances of the network, an 18-layer and a 34-layer Bi-Real net #REFR .", "text_before_citation": ["The VGG-Small model adopted uses the straight-through-estimator (STE) to approximate the gradient on non-differentiable layers #OTHEREFR .", "VGG-11 #OTHEREFR Network structure is the following: 64\u2212C3+M P 2+128\u2212C3+ M P 2 + 2 \u00d7 (256 \u2212 C3) + M P 2 + 2 \u00d7 (512 \u2212 C3) + M P 2 + 2 \u00d7 (512 \u2212 C3) + M P 2 + Sof tmax.", "Even VGG-11 uses the STE estimator for binarization operation during back-propagation.", "BiRealNet #OTHEREFR It is a modified version of classical ResNet that proposes to preserve the real activations before the sign function to increase the representational capability of the 1-bit CNN, through a simple shortcut.", "Bi-RealNet adopts a tight approximation to the derivative of the non-differentiable sign function with respect to activation and a magnitude-aware gradient to update weight parameters."], "text_after_citation": ["ReactNet #OTHEREFR To further compress compact networks, this model constructs a baseline based on MobileNetV1 #OTHEREFR and add shortcut to bypass every 1-bit convolutional layer that has the same number of input and output channels.", "The 3 \u00d7 3 depth-wise and the 1 \u00d7 1 point-wise convolutional blocks of MobileNet are replaced by the 3 \u00d7 3 and 1 \u00d7 1 vanilla convolutions in parallel with shortcuts in React Net #OTHEREFR .", "As for Bi-Real Net, we tested two different versions of React Net: a 18-layer and a 34-layer. of our binarization method w.r.t.", "the state-of-the-arts input layer binarization approaches, keeping unaltered the structure of the network except for the input data binarization layer (first half of Tables 2, 3 and 4) .", "In this first scenario all the 8-bits planes are exploited, layer F1 ( Fig."], "citing_paper_content": {"title": "Input Layer Binarization With Bit-Plane Encoding", "abstract": "Binary Neural Networks (BNNs) use 1-bit weights and activations to efficiently execute deep convolutional neural networks on edge devices. Nevertheless, the binarization of the first layer is conventionally excluded, as it leads to a large accuracy loss. The few works addressing the first layer binarization, typically increase the number of input channels to enhance data representation; such data expansion raises the amount of operations needed and it is feasible only on systems with enough computational resources. In this work, we present a new method to binarize the first layer using directly the 8-bit representation of input data; we exploit the standard bit-planes encoding to extract features bit-wise (using depth-wise convolutions); after a re-weighting stage, features are fused again. The resulting model is fully binarized and our first layer binarization approach is model independent. The concept is evaluated on three classification datasets (CIFAR10, SVHN and CIFAR100) for different model architectures (VGG and ResNet) and, the proposed technique outperforms state of the art methods both in accuracy and BMACs reduction."}, "cited_paper_content": {"title": "Dorefa-Net: Training Low Bitwidth Convolutional Neural Networks With Low Bitwidth Gradients", "abstract": "We propose DoReFa-Net, a method to train convolutional neural networks that have low bitwidth weights and activations using low bitwidth parameter gradients. In particular, during backward pass, parameter gradients are stochastically quantized to low bitwidth numbers before being propagated to convolutional layers. As convolutions during forward/backward passes can now operate on low bitwidth weights and activations/gradients respectively, DoReFa-Net can use bit convolution kernels to accelerate both training and inference. Moreover, as bit convolutions can be efficiently implemented on CPU, FPGA, ASIC and GPU, DoReFa-Net opens the way to accelerate training of low bitwidth neural network on these hardware. Our experiments on SVHN and ImageNet datasets prove that DoReFa-Net can achieve comparable prediction accuracy as 32-bit counterparts. For example, a DoReFa-Net derived from AlexNet that has 1-bit weights, 2-bit activations, can be trained from scratch using 6-bit gradients to get 46.1\\% top-1 accuracy on ImageNet validation set. The DoReFa-Net AlexNet model is released publicly."}, "keywords": ["34-layer Bi-Real net"], "citation_intent": "method"}
{"citing_id": "2305.02350v1", "cited_id": "1903.05987", "section_title": "Results", "citation": "We argue that this can be attributed to the much higher similarity between pre-training and downstream task compared to RoBERTa which was also observed by #REFR .", "text_before_citation": ["The results of feature extraction and fine-tuning are also compared to the current state-of-the-art results of the chosen datasets.", "On both single-label and multi-label datasets, fine-tuning on average performs better than feature extraction.", "In most cases, the smallest language model, BERT-Tiny, trained with feature extraction is on par with the word embeddings.", "However, when increasing model size or using fine-tuning, the word embeddings fall behind regarding recall.", "DBpedia is the only dataset on which BERT achieves better results with fine-tuning and feature extraction."], "text_after_citation": ["While feature extraction can achieve good results on single-label data compared to fine-tuning it unfortunately falls short when training on multi-label data.", "Table 1 : Test Accuracy (%) averaged over 3 runs on the single-label datasets.", "We compare feature extraction (FE) with our baselines and state-of-the-art (SOTA) models."], "citing_paper_content": {"title": "Using Language Models On Low-End Hardware", "abstract": "This paper evaluates the viability of using fixed language models for training text classification networks on low-end hardware. We combine language models with a CNN architecture and put together a comprehensive benchmark with 8 datasets covering single-label and multi-label classification of topic, sentiment, and genre. Our observations are distilled into a list of trade-offs, concluding that there are scenarios, where not fine-tuning a language model yields competitive effectiveness at faster training, requiring only a quarter of the memory compared to fine-tuning."}, "cited_paper_content": {"title": "To Tune Or Not To Tune? Adapting Pretrained Representations To Diverse Tasks", "abstract": "While most previous work has focused on different pretraining objectives and architectures for transfer learning, we ask how to best adapt the pretrained model to a given target task. We focus on the two most common forms of adaptation, feature extraction (where the pretrained weights are frozen), and directly fine-tuning the pretrained model. Our empirical results across diverse NLP tasks with two state-of-the-art models show that the relative performance of fine-tuning vs. feature extraction depends on the similarity of the pretraining and target tasks. We explore possible explanations for this finding and provide a set of adaptation guidelines for the NLP practitioner."}, "keywords": ["pre-training"], "citation_intent": "result"}
{"citing_id": "2303.04142v1", "cited_id": "1705.11087", "section_title": "I. Introduction", "citation": "However, what is more difficult to envision are the problems that are harder to fix because straightforward corrections may not exist, and rules for finding problems are more challenging to specify than those in smell detectors or linters #REFR like language idioms and code smells.", "text_before_citation": ["It involves complex challenges like following the best practices, avoiding code smells, using design patterns, and many more decisions before writing code.", "The biggest challenge with using tools like Copilot is their training data.", "These tools are trained on existing software source code, and training costs are expensive.", "Several classes of errors like securuty vlunerabilities #OTHEREFR , license compliance and copyright violation #OTHEREFR , and also simple coding mistakes, such as not allowing for an empty array in a sort routine #OTHEREFR have been discovered, which follow from the presence of these same errors in public (training) data.", "Although these are clearly challenges, Copilot seems already to be on its way to fixing them, like a filter introduced by GitHub to suppress code suggestions containing code that matches public code on GitHub."], "text_after_citation": ["Developers often discuss software architecture and actual source code implementations in online forums, chat rooms, mailing lists, or in person.", "Programming tasks can be solved in more than one way.", "The best way to proceed can be determined based on case-specific conditions, limits, and conventions.", "Strong standards and a shared vocabulary make communication easier while fostering a shared understanding of the issues and solutions related to software development.", "However, this takes time and experience to learn and use idiomatic approaches #OTHEREFR ."], "citing_paper_content": {"title": "From Copilot To Pilot: Towards Ai Supported Software Development", "abstract": "AI-supported programming has arrived, as shown by the introduction and successes of large language models for code, such as Copilot/Codex (Github/OpenAI) and AlphaCode (DeepMind). Above human average performance on programming challenges is now possible. However, software engineering is much more than solving programming contests. Moving beyond code completion to AI-supported software engineering will require an AI system that can, among other things, understand how to avoid code smells, to follow language idioms, and eventually (maybe!) propose rational software designs. In this study, we explore the current limitations of AI-supported code completion tools like Copilot and offer a simple taxonomy for understanding the classification of AI-supported code completion tools in this space. We first perform an exploratory study on Copilot's code suggestions for language idioms and code smells. Copilot does not follow language idioms and avoid code smells in most of our test scenarios. We then conduct additional investigation to determine the current boundaries of AI-supported code completion tools like Copilot by introducing a taxonomy of software abstraction hierarchies where 'basic programming functionality' such as code compilation and syntax checking is at the least abstract level, software architecture analysis and design are at the most abstract level. We conclude by providing a discussion on challenges for future development of AI-supported code completion tools to reach the design level of abstraction in our taxonomy."}, "cited_paper_content": {"title": "What To Fix? Distinguishing Between Design And Non-Design Rules In Automated Tools", "abstract": "Design problems, frequently the result of optimizing for delivery speed, are a critical part of long-term software costs. Automatically detecting such design issues is a high priority for software practitioners. Software quality tools promise automatic detection of common software quality rule violations. However, since these tools bundle a number of rules, including rules for code quality, it is hard for users to understand which rules identify design issues in particular. Research has focused on comparing these tools on open source projects, but these comparisons have not looked at whether the rules were relevant to design. We conducted an empirical study using a structured categorization approach, and manually classified 466 software quality rules from three industry tools\u2014CAST, SonarQube, and NDepend. We found that most of these rules were easily labeled as either non-design (55%) or design (19%). The remainder (26%) resulted in disagreements among the labelers. Our results are a first step in formalizing a definition of a design rule, to support automatic detection."}, "keywords": ["code smells"], "citation_intent": "background"}
{"citing_id": "2303.13887v1", "cited_id": "1312.6199", "section_title": "Introduction", "citation": "In the white box attack, the adversary has complete access to the target model, including the model architecture and parameters, allowing the attacker to use the model's gradient to produce the most effective adversarial examples #REFR .", "text_before_citation": ["The adversarial example is an input to machine learning models that an attacker has deliberately designed to cause the model to make a mistake.", "In particular, it is a corrupted version of a valid input, where the corruption is done by adding a perturbation of a small magnitude to it.", "The adversarial example aims to appear \"normal\" to humans but causes misclassification by the targeted machine learning model.", "An adversarial attack is then a method to generate adversarial examples.", "We distinguish two kinds of adversarial attacks: a white box attack and a black box attack."], "text_after_citation": ["On the contrary, in the black box attack scenario, the attackers do not have full access to the model and can only observe the output of its prediction #OTHEREFR . Fig. 1 . The proposed scheme.", "We consider three neural networks (NN): classification, simulation, and adversarial.", "The classification NN is a black box trained on an unknown 'target' dataset and presented by an unknown architecture.", "The simulation NN is not trainable and has one layer with 'simulation' kernels and serves for computing loss difference between the original and the adversarial output.", "The adversarial NN is a simple architecture, highly distinct from the black box one and trained on a different 'adversarial' dataset."], "citing_paper_content": {"title": "Effective Black Box Adversarial Attack With Handcrafted Kernels", "abstract": "We propose a new, simple framework for crafting adversarial examples for black box attacks. The idea is to simulate the substitution model with a non-trainable model compounded of just one layer of handcrafted convolutional kernels and then train the generator neural network to maximize the distance of the outputs for the original and generated adversarial image. We show that fooling the prediction of the first layer causes the whole network to be fooled and decreases its accuracy on adversarial inputs. Moreover, we do not train the neural network to obtain the first convolutional layer kernels, but we create them using the technique of F-transform. Therefore, our method is very time and resource effective."}, "cited_paper_content": {"title": "Intriguing Properties Of Neural Networks", "abstract": "Deep neural networks are highly expressive models that have recently achieved state of the art performance on speech and visual recognition tasks. While their expressiveness is the reason they succeed, it also causes them to learn uninterpretable solutions that could have counter-intuitive properties. In this paper we report two such properties. ::: First, we find that there is no distinction between individual high level units and random linear combinations of high level units, according to various methods of unit analysis. It suggests that it is the space, rather than the individual units, that contains of the semantic information in the high layers of neural networks. ::: Second, we find that deep neural networks learn input-output mappings that are fairly discontinuous to a significant extend. We can cause the network to misclassify an image by applying a certain imperceptible perturbation, which is found by maximizing the network's prediction error. In addition, the specific nature of these perturbations is not a random artifact of learning: the same perturbation can cause a different network, that was trained on a different subset of the dataset, to misclassify the same input."}, "keywords": ["effective adversarial examples"], "citation_intent": "method"}
{"citing_id": "2305.00521v1", "cited_id": "1806.05622", "section_title": "Dataset", "citation": "We train our model on Voxceleb2 #REFR , which consists of in-the-wild talking face videos collected from YouTube.", "text_before_citation": [], "text_after_citation": ["It contains more than 145,000 videos of about 6,100 identities in the train set and more than 4,900 videos of more than 110 identities in the test set.", "We convert all videos into frames with 25 fps and then crop and resize those frames into 256 \u00d7 256 resolution, following the method of #OTHEREFR .", "For more semantically rich face priors, we only use videos where the detected face bounding boxes' height (and width) is longer than 256.", "After pre-processing, the remaining 11051 videos and 340 videos are used as the training set and the test set, respectively.", "All audios are re-sampled to 16kHz and converted into a mel-spectrogram to be used as our audio representation similar to the method in #OTHEREFR ."], "citing_paper_content": {"title": "Stylelipsync: Style-Based Personalized Lip-Sync Video Generation", "abstract": "In this paper, we present StyleLipSync, a style-based personalized lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from arbitrary audio. To generate a video of arbitrary identities, we leverage expressive lip prior from the semantically rich latent space of a pre-trained StyleGAN, where we can also design a video consistency with a linear transformation. In contrast to the previous lip-sync methods, we introduce pose-aware masking that dynamically locates the mask to improve the naturalness over frames by utilizing a 3D parametric mesh predictor frame by frame. Moreover, we propose a few-shot lip-sync adaptation method for an arbitrary person by introducing a sync regularizer that preserves lips-sync generalization while enhancing the person-specific visual information. Extensive experiments demonstrate that our model can generate accurate lip-sync videos even with the zeroshot setting and enhance characteristics of an unseen face using a few seconds of target video through the proposed adaptation method. Please refer to our project page."}, "cited_paper_content": {"title": "Voxceleb2: Deep Speaker Recognition", "abstract": "The objective of this paper is speaker recognition under noisy and unconstrained conditions. ::: We make two key contributions. First, we introduce a very large-scale audio-visual speaker recognition dataset collected from open-source media. Using a fully automated pipeline, we curate VoxCeleb2 which contains over a million utterances from over 6,000 speakers. This is several times larger than any publicly available speaker recognition dataset. ::: Second, we develop and compare Convolutional Neural Network (CNN) models and training strategies that can effectively recognise identities from voice under various conditions. The models trained on the VoxCeleb2 dataset surpass the performance of previous works on a benchmark dataset by a significant margin."}, "keywords": ["in-the-wild talking face", "Voxceleb2"], "citation_intent": "method"}
{"citing_id": "2303.06266v1", "cited_id": "1010.2441", "section_title": "B. Bp Based User Identification", "citation": "In contrast to the previous studies in #REFR , our noise model as given in (5) incorporating channel fading effects and non-coherent detection at the BS side is more intricate.", "text_before_citation": ["In #OTHEREFR , the authors considered BP based GT decoding under a noise model that combines the addition and dilution models.", "In the context of mMTC, BP has been previously studied as a means for unsourced random access the users share the same codebook and the DRAFT March 14, 2023 decoder is only required to provide an unordered list of user messages #OTHEREFR .", "Recognizing the suitability of BP in sparse recovery problems including GT, we aim to employ a BP based strategy for the problem of active device identification in the non-coherent", "( , k)\u2212MnAC.", "In BP based user identification, we estimate the set of active devices by approximately performing bitwise-MAP detection through message passing on a bipartite graph with devices on one side and the binary energy measurements on the other side #OTHEREFR ."], "text_after_citation": ["Note that in #OTHEREFR , Polyanskiy proposed an alternating BP scheme as a candidate practical solution for random access in a quasi-static fading multiple access channel where each user's transmissions are attenuated by random fading gains, unknown to the receiver.", "However, #OTHEREFR considers iterative coding scheme based on LDPC codes and slow fading scenarios whereas our work focus on energy-efficient OOK signaling in a fast fading non-coherent ( , k)\u2212MnAC.", "To identify the activity status vector \u03b2 using the binary vector of energy detector outputs Z = (Z 1 , . . .", ", Z n ), BP relies on an approximate Maximum Aposteriori Probability (MAP)", "based decoding where instead of maximizing the posterior as a function of all the \u03b2 i 's, we use the marginals of the posterior distribution. i.e., \u03b2 i = arg max"], "citing_paper_content": {"title": "Non-Coherent Active Device Identification For Massive Random Access", "abstract": "Massive Machine-Type Communications (mMTC) is a key service category in the current generation of wireless networks featuring an extremely high density of energy and resource-limited devices with sparse and sporadic activity patterns. In order to enable random access in such mMTC networks, base station needs to identify the active devices while operating within stringent access delay constraints. In this paper, an energy efficient active device identification protocol is proposed in which active devices transmit On-Off Keying (OOK) modulated preambles jointly and base station employs non-coherent energy detection avoiding channel estimation overheads. The minimum number of channel-uses required by the active user identification protocol is characterized in the asymptotic regime of total number of devices when the number of active devices k scales as k = \u0398(1) along with an achievability scheme relying on the equivalence of activity detection to a group testing problem. Several practical schemes based on Belief Propagation (BP) and Combinatorial Orthogonal Matching Pursuit (COMP) are also proposed. Simulation results show that BP strategies outperform COMP significantly and can operate close to the theoretical achievability bounds. In a partial-recovery setting where few misdetections are allowed, BP continues to perform well."}, "cited_paper_content": {"title": "Note On Noisy Group Testing: Asymptotic Bounds And Belief Propagation Reconstruction", "abstract": "An information theoretic perspective on group testing problems has recently been proposed by Atia and Saligrama, in order to characterise the optimal number of tests. Their results hold in the noiseless case, where only false positives occur, and where only false negatives occur. We extend their results to a model containing both false positives and false negatives, developing simple information theoretic bounds on the number of tests required. Based on these bounds, we obtain an improved order of convergence in the case of false negatives only. Since these results are based on (computationally infeasible) joint typicality decoding, we propose a belief propagation algorithm for the detection of defective items and compare its actual performance to the theoretical bounds."}, "keywords": ["non-coherent detection"], "citation_intent": "result"}
{"citing_id": "2305.01386v1", "cited_id": "1604.01685", "section_title": "Related Works", "citation": "Most of these models have been benchmarked on large datasets such as COCO, Cityscapes (Lin et al., 2014; #REFR .", "text_before_citation": ["Among them, Semantic Segmentation is an important task where the model has to learn the classification of each pixel to a semantic class. In 2014, Jonathan Long et al.", "proposed a Fully Convolutional Network (FCN), which was the first CNN with only convolutional and transposed convolutional layers and without any fully connected (or dense) layers #OTHEREFR .", "With the advent of FCNs, researchers have proposed and developed several other states of the art models for the semantic segmentation task.", "Among them, the popular ones include -UNet, LinkNet, PSPNet, DeepLabV3+ #OTHEREFR .", "These models have one thing in common -i.e., they are all variations of encoder-decoder architectures."], "text_after_citation": ["Cityscapes, a labeled dataset used as a benchmark for semantic segmentation tasks, contains 5000 high quality labeled images and 20000 weakly labeled images. Marios Krestenitis et al.", "benchmarked some of these models on the Oil Spill Detection Dataset, a dataset developed by compiling the images extracted from the satellite Synthetic Aperture Radar (SAR) data #OTHEREFR .", "This dataset contains a meager 1002 training image, which is relatively much lower than the Cityscapes dataset. In their research, Marios Krestenitis et al.", "trained the models by dividing the original images of size 1250 \u00d7 650 into multiple patches.", "They used the image patches as input for the models. They used different image patch sizes for different models."], "citing_paper_content": {"title": "Oil Spill Segmentation Using Deep Encoder-Decoder Models", "abstract": "Crude oil is an integral component of the modern world economy. With the growing demand for crude oil due to its widespread applications, accidental oil spills are unavoidable. Even though oil spills are in and themselves difficult to clean up, the first and foremost challenge is to detect spills. In this research, the authors test the feasibility of deep encoder-decoder models that can be trained effectively to detect oil spills. The work compares the results from several segmentation models on high dimensional satellite Synthetic Aperture Radar (SAR) image data. Multiple combinations of models are used in running the experiments. The best-performing model is the one with the ResNet-50 encoder and DeepLabV3+ decoder 1. It achieves a mean Intersection over Union (IoU) of 64.868% and a class IoU of 61.549% for the \"oil spill\" class when compared with the current benchmark model, which achieved a mean IoU of 65.05% and a class IoU of 53.38% for the \"oil spill\" class."}, "cited_paper_content": {"title": "The Cityscapes Dataset For Semantic Urban Scene Understanding", "abstract": "Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations, 20 000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark."}, "keywords": ["large datasets", "Cityscapes"], "citation_intent": "background"}
{"citing_id": "2303.05581v1", "cited_id": "1706.04599", "section_title": "Related Work", "citation": "This method is limited by the often inaccurate (uncalibrated) predictive confidence learned by the neural network #REFR .", "text_before_citation": ["Boundary Calibration The classical local outlier factor (LOF) #OTHEREFR method is a custom metric that calculates the local density deviation of a data point from its neighbors.", "However, there is not a principled rule on how to choose the outlier detection threshold when using LOF. #OTHEREFR ; #OTHEREFR", "(2022) added open category data into the validation set to estimate or grid-search the proper threshold.", "So motivated, #OTHEREFR fit the output logits of the classifier to a Weibull distribution, but still use a validation set that contains the open category to select the confidence threshold. Further, #OTHEREFR", "(2017) employed one-versus-rest binary classifiers and then calculates the threshold over the confidence score space by fitting it to a Gaussian distribution."], "text_after_citation": ["Adaptive decision boundary #OTHEREFR , illustrated in Figure 1 (c), was recently proposed to learn bounded spherical regions for known categories to contain known class samples.", "Though this post-processing approach achieves state-of-theart performance, it still suffers from the issue that the tight-bounded spheres may not exist or cannot be well-defined in representation space.", "Due to the fact that high-dimensional data representations usually lie on a low-dimensional manifold #OTHEREFR , a sphere defined in a Euclidean space can be restrictive as a decision boundary.", "Moreover, the issue can be more severe if certain categories follow multimodal or skewed distributions.", "Representation Learning DeepUnk (Lin and Xu, 2019a) trains the feature extractor with Large Margin Cosine Loss #OTHEREFR ."], "citing_paper_content": {"title": "Open World Classification With Adaptive Negative Samples", "abstract": "Open world classification is a task in natural language processing with key practical relevance and impact. Since the open or unknown category data only manifests in the inference phase, finding a model with a suitable decision boundary accommodating for the identification of known classes and discrimination of the open category is challenging. The performance of existing models is limited by the lack of effective open category data during the training stage or the lack of a good mechanism to learn appropriate decision boundaries. We propose an approach based on adaptive negative samples (ANS) designed to generate effective synthetic open category samples in the training stage and without requiring any prior knowledge or external datasets. Empirically, we find a significant advantage in using auxiliary one-versus-rest binary classifiers, which effectively utilize the generated negative samples and avoid the complex thresholdseeking stage in previous works. Extensive experiments on three benchmark datasets show that ANS achieves significant improvements over state-of-the-art methods."}, "cited_paper_content": {"title": "On Calibration Of Modern Neural Networks", "abstract": "Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions."}, "keywords": ["predictive confidence", "neural network"], "citation_intent": "method"}
{"citing_id": "2304.01565v1", "cited_id": "1802.04364", "section_title": "Variational Auto Encoder", "citation": "By contrast, JT-VAE #REFR , a motif-sequence-based generator, generates molecular graphs by a procedure that generates a treestructured scaffold over chemical substructures first and then combines them into a molecule.", "text_before_citation": ["When generating a graph, there is often a constraint demand that only specific combinations of the nodes and edges are valid in context #OTHEREFR , #OTHEREFR , #OTHEREFR .", "RGVAE #OTHEREFR generates a semantically valid graph by adding a penalty term that regularizes the output distribution of the decoder to the objective function of VAEs.", "Graphite #OTHEREFR repeats the process that constructs an intermediate weighted graph and iteratively refines the reconstructed feature matrix via message passing.", "The final distribution of graph is obtained via the inner product of feature matrix. The other category is sequential generation method.", "CGVAE #OTHEREFR , node-by-node generator, repeats the process that chooses which node is connected to the current node and selects and labels edges between nodes to generate a graph."], "text_after_citation": ["But Autoencoderbassed graph generation methods have limitations in that they are not scalable to be applied on large graphs."], "citing_paper_content": {"title": "A Survey On Graph Diffusion Models: Generative Ai In Science For Molecule, Protein And Material", "abstract": "Diffusion models have become a new SOTA generative modeling method in various fields, for which there are multiple survey works that provide an overall survey. With the number of articles on diffusion models increasing exponentially in the past few years, there is an increasing need for survey of diffusion models on specific fields. In this work, we are committed to conducting a survey on the graph diffusion models. Even though our focus is to cover the progress of diffusion models in graphs, we first briefly summarize how other generative modeling methods are used for graphs. After that, we introduce the mechanism of diffusion models in various forms, which facilitates the discussion on the graph diffusion models. The applications of graph diffusion models mainly fall into the catogory of AI-generated content (AIGC) in science, for which we mainly focus on how graph diffusion models are utilized for generating molecules and proteins but also cover other cases, including materials design. Moreover, we discuss the issue of evaluating diffusion models in the graph domain and the existing challenges."}, "cited_paper_content": {"title": "Junction Tree Variational Autoencoder For Molecular Graph Generation", "abstract": "We seek to automate the design of molecules based on specific chemical properties. In computational terms, this task involves continuous embedding and generation of molecular graphs. Our primary contribution is the direct realization of molecular graphs, a task previously approached by generating linear SMILES strings instead of graphs. Our junction tree variational autoencoder generates molecular graphs in two phases, by first generating a tree-structured scaffold over chemical substructures, and then combining them into a molecule with a graph message passing network. This approach allows us to incrementally expand molecules while maintaining chemical validity at every step. We evaluate our model on multiple tasks ranging from molecular generation to optimization. Across these tasks, our model outperforms previous state-of-the-art baselines by a significant margin."}, "keywords": ["molecular graphs"], "citation_intent": "method"}
{"citing_id": "2304.09704v1", "cited_id": "1904.09970", "section_title": "Additional Qualitative Results", "citation": "Since SuperQuadrics #REFR uses a restricted family of 3D shapes to reconstruct an input scene, it has worst qualitative performances for instance segmentation when compared to our model, which learns scene-specific prototypes and can provide semantic information. Table 7 : Results on the Earth Parser Dataset.", "text_before_citation": ["Earth Parser Dataset results.", "We show in Figure 9 the ground truth semantic segmentation, our predicted semantic segmentation, our reconstruction and our learned prototypes.", "They showcase the quality, interpretability, and diversity of use cases of our model on this dataset of aerial LiDAR scans.", "Instance segmentation.", "We show in Figure 10 a comparison of the instance segmentation produced by Su-perQuadrics #OTHEREFR and our Learnable Earth Parser."], "text_after_citation": ["We report the quality of the reconstruction (Cham.) and semantic segmentation (mIoU) for the models presented in the main paper and other variations.", "Bold numbers indicate the best results of the models shown in the main paper, while green bold numbers indicate the best results across all variations.", "\u2193 indicates that the variation results in a significant drop in performance, while \u2191 indicates a performance boost.", "We also show the number of prototypes selected by our post-processing selection algorithm. Ours Figure 10 : Instance Segmentation.", "We represent the instances predicted with our algorithm and by SuperQuadric #OTHEREFR ."], "citing_paper_content": {"title": "Learnable Earth Parser: Discovering 3D Prototypes In Aerial Scans", "abstract": "We propose an unsupervised method for parsing large 3D scans of real-world scenes into interpretable parts. Our goal is to provide a practical tool for analyzing 3D scenes with unique characteristics in the context of aerial surveying and mapping, without relying on applicationspecific user annotations. Our approach is based on a probabilistic reconstruction model that decomposes an input 3D point cloud into a small set of learned prototypical shapes. Our model provides an interpretable reconstruction of complex scenes and leads to relevant instance and semantic segmentations. To demonstrate the usefulness of our results, we introduce a novel dataset of seven diverse aerial LiDAR scans. We show that our method outperforms state-of-the-art unsupervised methods in terms of decomposition accuracy while remaining visually interpretable. Our method offers significant advantage over existing approaches, as it does not require any manual annotations, making it a practical and efficient tool for 3D scene analysis. Our code and dataset are available at https://imagine.enpc.fr/\u02dcloiseaur/ learnable-earth-parser. Windturbines Forest Crop fields Urban Power plant Greenhouses Marina"}, "cited_paper_content": {"title": "Superquadrics Revisited: Learning 3D Shape Parsing Beyond Cuboids", "abstract": "ing complex 3D shapes with parsimonious part-based representations has been a long standing goal in computer vision. This paper presents a learning-based solution to this problem which goes beyond the traditional 3D cuboid representation by exploiting superquadrics as atomic elements. We demonstrate that superquadrics lead to more expressive 3D scene parses while being easier to learn than 3D cuboid representations. Moreover, we provide an analytical solution to the Chamfer loss which avoids the need for computational expensive reinforcement learning or iterative prediction. Our model learns to parse 3D objects into consistent superquadric representations without supervision. Results on various ShapeNet categories as well as the SURREAL human body dataset demonstrate the flexibility of our model in capturing fine details and complex poses that could not have been modelled using cuboids."}, "keywords": ["Earth Parser Dataset", "3D shapes"], "citation_intent": "result"}
{"citing_id": "2303.12528v1", "cited_id": "1606.05250", "section_title": "Question Answering", "citation": "For each task, we calculate the Exact Match and F1 score as defined in #REFR .", "text_before_citation": ["First, for the few-shot examples we only provide the line within the paragraph containing the answer as the context.", "Second, for the test example, we index the chunks of the context using the embeddings from the text-embedding-ada-002 model.", "Given the question, the closest chunk in the full context is retrieved and used in the prompt for the test example.", "We use a maximum chunk size of 100 in our experiments and use the implementation for retrieval provided in the LangChain 3 library.", "By doing this,we minimize the space taken by the context tokens in our prompt."], "text_after_citation": ["For our experiments we consider the following three tasks: 1.", "TyDiQA (Clark et al., 2020) is a QA dataset covering 11 typologically diverse languages.", "The task consists of two sub-tasks -passage selection and minimum answer span (Gold-P).", "For our experiments, we consider the Gold-P task and evaluate Monolingual and Zero-Shot Cross-Lingual prompting strategies.", "Since the labels do not directly transfer one-to-one across translation for QA tasks as they do for classification and require the use of alignment algorithms, we skip translate-test prompting for this task. 2."], "citing_paper_content": {"title": "Mega: Multilingual Evaluation Of Generative Ai", "abstract": "Generative AI models have impressive performance on many Natural Language Processing tasks such as language understanding, reasoning and language generation. One of the most important questions that is being asked by the AI community today is about the capabilities and limits of these models, and it is clear that evaluating generative AI is very challenging. Most studies on generative LLMs are restricted to English and it is unclear how capable these models are at understanding and generating other languages. We present the first comprehensive benchmarking of generative LLMs-MEGA, which evaluates models on standard NLP benchmarks, covering 8 diverse tasks and 33 typologically diverse languages. We also compare the performance of generative LLMs to State of the Art (SOTA) non-autoregressive models on these tasks to determine how well generative models perform compared to the previous generation of LLMs. We present a thorough analysis of the performance of models across languages and discuss some of the reasons why generative LLMs are currently not optimal for all languages. We create a framework for evaluating generative LLMs in the multilingual setting and provide directions for future progress in the field."}, "cited_paper_content": {"title": "Squad: 100,000+ Questions For Machine Comprehension Of Text", "abstract": "We present the Stanford Question Answering Dataset (SQuAD), a new reading comprehension dataset consisting of 100,000+ questions posed by crowdworkers on a set of Wikipedia articles, where the answer to each question is a segment of text from the corresponding reading passage. We analyze the dataset to understand the types of reasoning required to answer the questions, leaning heavily on dependency and constituency trees. We build a strong logistic regression model, which achieves an F1 score of 51.0%, a significant improvement over a simple baseline (20%). However, human performance (86.8%) is much higher, indicating that the dataset presents a good challenge problem for future research. ::: The dataset is freely available at this https URL"}, "keywords": ["task"], "citation_intent": "method"}
{"citing_id": "2303.16529v1", "cited_id": "1412.6980", "section_title": "Experimental Results", "citation": "This result is not surprising, as RMSProp and ADAM are known for yielding higher convergence speed than SGD since they use more information: the moving average of gradients helps RMSProp and ADAM driving the optimization towards the optimum #REFR .", "text_before_citation": ["This probability distribution verifies the hypothesis of our theorem, hence it brings theoretically an improvement on the convergence speed compared to the uniform sampling scheme. The experiments show this improvement is not only theoretical.", "For SGD and SGD with momentum performed with p 1/2 , the loss decrease is faster than the loss decrease obtained with the uniform sampling, and slower than the one obtained with the optimal sampling scheme (in terms of iterations). Fig.", "2 lower right compares SGD, SGD with momentum, RMSProp and ADAM with the uniform sampling for completeness.", "Even though this comparison is not the core of our study, the following can be noted.", "First, RMSProp and ADAM provide a faster loss decrease in the early stages of the optimization process than SGD (with or without momentum)."], "text_after_citation": ["However, more surprising is the final result of the optimization: the final loss is lower for SGD than for RMSProp and ADAM.", "This can be justified by the bias induced by the update equations of RMSProp and ADAM.", "Whereas SGD minimizes the empirical loss, as proven by #OTHEREFR , it is not the case for RMSProp and ADAM. Fig.", "2 cannot be fully appreciated without Table 1 , where the computational time of each sampling scheme used in this work is reported.", "Whatever the optimizer, the computational time is approximately the same for a given sampling scheme."], "citing_paper_content": {"title": "Importance Sampling For Stochastic Gradient Descent In Deep Neural Networks", "abstract": "Stochastic gradient descent samples uniformly the training set to build an unbiased gradient estimate with a limited number of samples. However, at a given step of the training process, some data are more helpful than others to continue learning. Importance sampling for training deep neural networks has been widely studied to propose sampling schemes yielding better performance than the uniform sampling scheme. After recalling the theory of importance sampling for deep learning, this paper reviews the challenges inherent to this research area. In particular, we propose a metric allowing the assessment of the quality of a given sampling scheme; and we study the interplay between the sampling scheme and the optimizer used. Preprint. Under review."}, "cited_paper_content": {"title": "Adam: A Method For Stochastic Optimization", "abstract": "We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm."}, "keywords": ["higher convergence speed", "RMSProp"], "citation_intent": "result"}
{"citing_id": "2304.02291v1", "cited_id": "1706.03762", "section_title": "Internal Lr Feature Guided Multiscale Attention", "citation": "In this study, scaled dot-product attention #REFR is chosen to implement multiscale attention. In Fig.", "text_before_citation": ["The details of multiscale attention for internal LR and HR feature fusion are shown in Fig. 7 .", "Unlike the original attention model in Transformer #OTHEREFR , three types of inputs, key (K), query (Q), and value (V), are visual feature maps, and K has a different scale than Q and V.", "In other words, K is small-scale and takes linearly transformed internal LR features \u210b(\u2113) as the input. In addition, K represents deformed LR features. In contrast, Q and V are large-scale.", "The internal HR feature maps are assigned to Q and V after applying linear transformation."], "text_after_citation": ["6 , the HR hourglass follows the encoder-decoder framework, and thus, multiscale attention is implemented slightly differently depending on the encoder and decoder.", "For the encoder of the HR hourglass, multiscale attention is designed as follows: Here, represents the internal HR features transferred by the encoder via within-hourglass skip connection, as shown in Fig.", "6 , and then added to the scaled dot-product attention result.", "EQUATION", "In the proposed multiscale attention, internal LR features are jointly learned with internal HR features via between-hourglass skip connection."], "citing_paper_content": {"title": "Trap-Based Pest Counting: Multiscale And Deformable Attention Centernet Integrating Internal Lr And Hr Joint Feature Learning", "abstract": "Pest counting, which predicts the number of pests in the early stage, is very important because it enables rapid pest control, reduces damage to crops, and improves productivity. In recent years, light traps have been increasingly used to lure and photograph pests for pest counting. However, pest images have a wide range of variability in pest appearance owing to severe occlusion, wide pose variation, and even scale variation. This makes pest counting more challenging. To address these issues, this study proposes a new pest counting model referred to as multiscale and deformable attention CenterNet (Mada-CenterNet) for internal low-resolution (LR) and high-resolution (HR) joint feature learning. Compared with the conventional CenterNet, the proposed Mada-CenterNet adopts a multiscale heatmap generation approach in a two-step fashion to predict LR and HR heatmaps adaptively learned to scale variations, that is, changes in the number of pests. In addition, to overcome the pose and occlusion problems, a new between-hourglass skip connection based on deformable and multiscale attention is designed to ensure internal LR and HR joint feature learning and incorporate geometric deformation, thereby resulting in an improved pest counting accuracy. Through experiments, the proposed Mada-CenterNet is verified to generate the HR heatmap more accurately and improve pest counting accuracy owing to multiscale heatmap generation, joint internal feature learning, and deformable and multiscale attention. In addition, the proposed model is confirmed to be effective in overcoming severe occlusions and variations in pose and scale. The experimental results show that the proposed model outperforms state-of-the-art crowd counting and object detection models."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["multiscale attention"], "citation_intent": "method"}
{"citing_id": "2303.03339v1", "cited_id": "1703.00420", "section_title": "B. Action Projection", "citation": "As the projection only checks against intersection in the endconfiguration of the validation trajectory, it is not guaranteed that the resulting projected trajectory is intersection-free in #REFR .", "text_before_citation": ["EQUATION", "subject to", "G ([0, \u03b1 opt ]) O (o) exp = \u2205, \u03b1 opt < 1 \u2212 \u03b1 opt < \u03b1 opt, min , 2 A \u2295 B = {a + v | b \u2208 A, v \u2208 B}", "where we adapt \u03b1 opt, min to the robot dynamics.", "Finally, we use a trajectory planner to plan a trajectory \u03c7 P from p (r) 0 to an ISS in an -bound around c (r) V , which ensures that the first L steps of the trajectory have a constant control input."], "text_after_citation": ["Therefore, if \u03c7 P would result in an intersection, we can repetitively reduce \u03b1 for M times to find a more conservative projection point.", "If no temporarily safe trajectory \u03c7 P can be found, we execute the zero action and continue to ensure safety with our safety shield."], "citing_paper_content": {"title": "Reducing Safety Interventions In Provably Safe Reinforcement Learning", "abstract": "Deep Reinforcement Learning (RL) has shown promise in addressing complex robotic challenges. In real-world applications, RL is often accompanied by failsafe controllers as a last resort to avoid catastrophic events. While necessary for safety, these interventions can result in undesirable behaviors, such as abrupt braking or aggressive steering. This paper proposes two safety intervention reduction methods: action replacement and projection, which change the agent's action if it leads to an unsafe state. These approaches are compared to the state-of-the-art constrained RL on the OpenAI safety gym benchmark and a human-robot collaboration task. Our study demonstrates that the combination of our method with provably safe RL leads to high-performing policies with zero safety violations and a low number of failsafe interventions. Our versatile method can be applied to a wide range of realworld robotics tasks, while effectively improving safety without sacrificing task performance."}, "cited_paper_content": {"title": "Virtual-To-Real Deep Reinforcement Learning: Continuous Control Of Mobile Robots For Mapless Navigation", "abstract": "We present a learning-based mapless motion planner by taking the sparse 10-dimensional range findings and the target position with respect to the mobile robot coordinate frame as input and the continuous steering commands as output. Traditional motion planners for mobile ground robots with a laser range sensor mostly depend on the obstacle map of the navigation environment where both the highly precise laser sensor and the obstacle map building work of the environment are indispensable. We show that, through an asynchronous deep reinforcement learning method, a mapless motion planner can be trained end-to-end without any manually designed features and prior demonstrations. The trained planner can be directly applied in unseen virtual and real environments. The experiments show that the proposed mapless motion planner can navigate the nonholonomic mobile robot to the desired targets without colliding with any obstacles."}, "keywords": ["validation trajectory"], "citation_intent": "background"}
{"citing_id": "2303.14284v1", "cited_id": "1805.08571", "section_title": "Approximating Logistic Loss", "citation": "This type of error guarantees were previously provided by a body of work on coresets for logistic regression; see the works of Munteanu et al. #REFR , Mai et al.", "text_before_citation": ["In this section, we consider approximations to the logistic loss function."], "text_after_citation": ["#OTHEREFR for state-of-the-art results and bounds and references therein for motivation and prior work on this topic.", "We begin by proving a general lower bound on the space complexity of any data structure which approximates logistic loss to \u01eb-relative error for every parameter vector \u03b2 \u2208 R d .", "Our lower bound shows that the coreset construction of Mai et al.", "#OTHEREFR has optimal space complexity up to an\u00d5(d \u2022 \u00b5 y (X) 2 ) factor, where \u00b5 y (X) 2 is a data complexity measure introduced by Munteanu et al. #OTHEREFR (see also Definition 1 below).", "Furthermore, we provide an efficient linear programming formulation to compute this data complexity measure on real data sets, refuting an earlier conjecture that the factor was hard to compute #OTHEREFR ."], "citing_paper_content": {"title": "Feature Space Sketching For Logistic Regression", "abstract": "We present novel bounds for coreset construction, feature selection, and dimensionality reduction for logistic regression. All three approaches can be thought of as sketching the logistic regression inputs. On the coreset construction front, we resolve open problems from prior work and present novel bounds for the complexity of coreset construction methods. On the feature selection and dimensionality reduction front, we initiate the study of forward error bounds for logistic regression. Our bounds are tight up to constant factors and our forward error bounds can be extended to Generalized Linear Models."}, "cited_paper_content": {"title": "On Coresets For Logistic Regression", "abstract": "Coresets are one of the central methods to facilitate the analysis of large data sets. We continue a recent line of research applying the theory of coresets to logistic regression. First, we show a negative result, namely, that no strongly sublinear sized coresets exist for logistic regression. To deal with intractable worst-case instances we introduce a complexity measure $\\mu(X)$, which quantifies the hardness of compressing a data set for logistic regression. $\\mu(X)$ has an intuitive statistical interpretation that may be of independent interest. For data sets with bounded $\\mu(X)$-complexity, we show that a novel sensitivity sampling scheme produces the first provably sublinear $(1\\pm\\varepsilon)$-coreset. We illustrate the performance of our method by comparing to uniform sampling as well as to state of the art methods in the area. The experiments are conducted on real world benchmark data for logistic regression."}, "keywords": ["logistic regression"], "citation_intent": "background"}
{"citing_id": "2303.15472v1", "cited_id": "1409.0575", "section_title": "Visualization Of The Consistency Of Orientation Estimation", "citation": "To visualize Figure 9 , we create a sequence of 480 \u00d7 640 images augmented by random in-plane rotation with Gaussian noise sourced by ILSVRC2012 #REFR .", "text_before_citation": ["We provide more examples for Figure 5 of the main paper, which visualize the consistency of orientation estimation. Additionally, we show the similarity map w.r.t. a keypoint under varying rotations."], "text_after_citation": ["Figure 9 shows the qualitative comparison of the estimated orientation consistency.", "Given the dominant orientations estimated from the image pair, we calculate the relative angle between the corresponding keypoint orientations and measure the difference between the relative angle and the ground-truth rotation.", "We evaluate the relative angle to be correct i.e., the dominant orientation estimation is consistent if the difference with the ground-truth rotation is within a 30 \u2022 threshold.", "Our rotation-equivariant model trained with the orientation alignment loss inspired by #OTHEREFR consistently estimates more correct keypoint orientations than LF-Net #OTHEREFR and RF-Net #OTHEREFR .", "Figure 10 shows the similarity maps with respect to a keypoint under varying rotations of images with a resolution of 180 \u00d7 180, with uniform rotation intervals of 45 \u2022 ."], "citing_paper_content": {"title": "Learning Rotation-Equivariant Features For Visual Correspondence", "abstract": "Extracting discriminative local features that are invariant to imaging variations is an integral part of establishing correspondences between images. In this work, we introduce a self-supervised learning framework to extract discriminative rotation-invariant descriptors using groupequivariant CNNs. Thanks to employing group-equivariant CNNs, our method effectively learns to obtain rotationequivariant features and their orientations explicitly, without having to perform sophisticated data augmentations. The resultant features and their orientations are further processed by group aligning, a novel invariant mapping technique that shifts the group-equivariant features by their orientations along the group dimension. Our group aligning technique achieves rotation-invariance without any collapse of the group dimension and thus eschews loss of discriminability. The proposed method is trained end-to-end in a self-supervised manner, where we use an orientation alignment loss for the orientation estimation and a contrastive descriptor loss for robust local descriptors to geometric/photometric variations. Our method demonstrates state-of-the-art matching accuracy among existing rotationinvariant descriptors under varying rotation and also shows competitive results when transferred to the task of keypoint matching and camera pose estimation."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["ILSVRC2012"], "citation_intent": "method"}
{"citing_id": "2303.07645v1", "cited_id": "0908.3916", "section_title": "Our Contribution", "citation": "Unit disk graph: We give an algorithm to compute a maximum clique in a unit disk graph in O(n 2.5 log n)-time, which improves the previously best known running time of O(n 3 log n) #REFR .", "text_before_citation": ["In this paper we make significant progress on the maximum clique problem for unit disk graphs, disk graphs with disk radii lying in the interval [1, 1 + \u03b5] , and intersection graphs of unit disks and axis-parallel rectangles."], "text_after_citation": ["Our algorithm is based on a divide-and-conquer approach that, unlike the previous algorithms that search a clique over all the lenses, shows how to efficiently merge solutions to the subproblems to achieve a faster time complexity.", "Such techniques have previously been used to accelerate computation for other computational geometry problems, e.g., when finding a closest pair in a point set #OTHEREFR , but appeared to be highly non-trivial while adapting it for the unit disk graph setting.", "Intersection graph of unit disks and axis-parallel rectangles: We extend the co-2-subdivision approach to prove a (4448/4449 \u2248 0.9997)-inapproximability result for computing a maximum clique in an intersection graph that contains both unit disks and axis-parallel rectangles, and thus improve the previously known inapproximability factor of 7633010347/7633010348 \u2248 0.9999 #OTHEREFR .", "Note that the use of a co-2-subdivision approach was previously thought to be unlikely in this setting by Bonnet et al. #OTHEREFR .", "The key idea behind our NP-hardness reduction is to show that every Hamiltonian cubic graph admits a well-behaved edge orientation and vertex labeling, i.e., its vertices can be labeled and the edges can be oriented such that every vertex has two outgoing or two incoming edges where the labels of these corresponding neighbors are consecutive."], "citing_paper_content": {"title": "Finding A Maximum Clique In A Disk Graph", "abstract": "A disk graph is an intersection graph of disks in the Euclidean plane, where the disks correspond to the vertices of the graph and a pair of vertices are adjacent if and only if their corresponding disks intersect. The problem of determining the time complexity of computing a maximum clique in a disk graph is a long-standing open question that has been very well studied in the literature. The problem is known to be open even when the radii of all the disks are in the interval [1, (1 + \u03b5)], where \u03b5 > 0. If all the disks are unit disks then there exists an O(n 3 log n)-time algorithm to compute a maximum clique, which is the best-known running time for over a decade. Although the problem of computing a maximum clique in a disk graph remains open, it is known to be APX-hard for the intersection graphs of many other convex objects such as intersection graphs of ellipses, triangles, and a combination of unit disks and axis-parallel rectangles. Here we obtain the following results.-We give an algorithm to compute a maximum clique in a unit disk graph in O(n 2.5 log n)-time, which improves the previously best known running time of O(n 3 log n) [Eppstein '09].-We extend a widely used 'co-2-subdivision approach' to prove that computing a maximum clique in a combination of unit disks and axis-parallel rectangles is NP-hard to approximate within 4448/4449 \u2248 0.9997. The use of a 'co-2-subdivision approach' was previously thought to be unlikely in this setting [Bonnet et al. '20]. Our result improves the previously known inapproximability factor of 7633010347/7633010348 \u2248 0.9999.-We show that the parameter minimum lens width of the disk arrangement may be used to make progress in the case when disk radii are in [1, (1 + \u03b5)]. For example, if the minimum lens width is at least 0.265 and \u03b5 \u2264 0.0001, which still allows for non-Helly triples in the arrangement, then one can find a maximum clique in polynomial time."}, "cited_paper_content": {"title": "Graph-Theoretic Solutions To Computational Geometry Problems", "abstract": "Many problems in computational geometry are not stated in graph-theoretic terms, but can be solved efficiently by constructing an auxiliary graph and performing a graph-theoretic algorithm on it. Often, the efficiency of the algorithm depends on the special properties of the graph constructed in this way. We survey the art gallery problem, partition into rectangles, minimum-diameter clustering, rectilinear cartogram construction, mesh stripification, angle optimization in tilings, and metric embedding from this perspective."}, "keywords": ["maximum clique", "algorithm"], "citation_intent": "method"}
{"citing_id": "2303.01410v1", "cited_id": "1908.10084", "section_title": "Components", "citation": "In addition to name similarity, the ranking of candidates utilizes the cosine similarity between the sentence embeddings #REFR of the context and the descriptions of the candidate entity from Wikipedia and Wikidata.", "text_before_citation": ["Existing tools and models in NLP Workbench include: Named Entity Recognition The task, known as NER, is to identify mentions to entities such as people, organizations, and locations.", "We incorporated the NER model from PURE #OTHEREFR , which achieved good performance by simply fine-tuning BERT #OTHEREFR .", "Coreference Resolution To determine which entity a pronoun refers to, we adopted the heuristic algorithm by #OTHEREFR that is based on recency and type agreement.", "Entity Linking Mentions to entities in the text are disambiguated and linked to Wikidata (Vrande\u010di\u0107 and Kr\u00f6tzsch, 2014) entities.", "Candidate entities are generated by a fuzzy match on name."], "text_after_citation": [], "citing_paper_content": {"title": "Nlp Workbench: Efficient And Extensible Integration Of State-Of-The-Art Text Mining Tools", "abstract": "NLP Workbench is a web-based platform for text mining that allows non-expert users to obtain semantic understanding of large-scale corpora using state-of-the-art text mining models. The platform is built upon latest pretrained models and open source systems from academia that provide semantic analysis functionalities, including but not limited to entity linking, sentiment analysis, semantic parsing, and relation extraction. Its extensible design enables researchers and developers to smoothly replace an existing model or integrate a new one. To improve efficiency, we employ a microservice architecture that facilitates allocation of acceleration hardware and parallelization of computation. This paper presents the architecture of NLP Workbench and discusses the challenges we faced in designing it. We also discuss diverse use cases of NLP Workbench and the benefits of using it over other approaches. The platform is under active development, with its source code released under the MIT license 1. A website 2 and a short video 3 demonstrating our platform are also available."}, "cited_paper_content": {"title": "Sentence-Bert: Sentence Embeddings Using Siamese Bert-Networks", "abstract": "BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019) has set a new state-of-the-art performance on sentence-pair regression tasks like semantic textual similarity (STS). However, it requires that both sentences are fed into the network, which causes a massive computational overhead: Finding the most similar pair in a collection of 10,000 sentences requires about 50 million inference computations (~65 hours) with BERT. The construction of BERT makes it unsuitable for semantic similarity search as well as for unsupervised tasks like clustering. ::: In this publication, we present Sentence-BERT (SBERT), a modification of the pretrained BERT network that use siamese and triplet network structures to derive semantically meaningful sentence embeddings that can be compared using cosine-similarity. This reduces the effort for finding the most similar pair from 65 hours with BERT / RoBERTa to about 5 seconds with SBERT, while maintaining the accuracy from BERT. ::: We evaluate SBERT and SRoBERTa on common STS tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embeddings methods."}, "keywords": ["sentence embeddings"], "citation_intent": "method"}
{"citing_id": "2303.09789v2", "cited_id": "1903.07789", "section_title": "Introduction", "citation": "To address this limitation, we propose to divide the city into irregular areas based on its road network #REFR , and create two traffic flow prediction datasets incorporating POI data.", "text_before_citation": ["The POI-MetaBlock adopts the self-attention structure #OTHEREFR and integrates time information to achieve dynamic attention, enabling our module to capture traffic patterns dynamically.", "We treat the POI distribution as metadata for parameter generation #OTHEREFR , which enables the model to accommodate different traffic patterns across various regions.", "To model functionality correlations that are overlooked by the original geographic topology, we establish a second topological structure (represented by the red line in Figure 1 ) using cosine similarity of POI distribution and apply a graph convolution network #OTHEREFR to refine attention results in each region based on the functionality topological structure.", "It is noteworthy that the proposed module can be easily integrated into conventional models through a residual connection, allowing existing models to additionally account for the impact of functional attributes on the original basis.", "In commonly used traffic prediction datasets, the urban area is often obtained through the grid method, which can lead to a loss of the original functional areas of the city."], "text_after_citation": ["In summary, our contributions are mainly summarized into the following aspects:", "\u2022 We propose a novel module named POI-MetaBlock, which leverages POI&time data to dynamically generate prediction parameters and fit traffic patterns of different areas at different times.", "\u2022 The proposed POI-MetaBlock is a lightweight module that can be easily integrated into other traffic flow prediction models.", "It allows the original model to take advantage of functional correlations during traffic flow prediction.", "\u2022 We constructed two new datasets by partitioning the city into irregular areas using POI information."], "citing_paper_content": {"title": "Urban Regional Function Guided Traffic Flow Prediction", "abstract": "The prediction of traffic flow is a challenging yet crucial problem in spatial-temporal analysis, which has recently gained increasing interest. In addition to spatial-temporal correlations, the functionality of urban areas also plays a crucial role in traffic flow prediction. However, the exploration of regional functional attributes mainly focuses on adding additional topological structures, ignoring the influence of functional attributes on regional traffic patterns. Different from the existing works, we propose a novel module named POI-MetaBlock, which utilizes the functionality of each region (represented by Point of Interest distribution) as metadata to further mine different traffic characteristics in areas with different functions. Specifically, the proposed POI-MetaBlock employs a self-attention architecture and incorporates POI and time information to generate dynamic attention parameters for each region, which enables the model to fit different traffic patterns of various areas at different times. Furthermore, our lightweight POI-MetaBlock can be easily integrated into conventional traffic flow prediction models. Extensive experiments demonstrate that our module significantly improves the performance of traffic flow prediction and outperforms state-of-the-art methods that use metadata."}, "cited_paper_content": {"title": "Predicting Citywide Crowd Flows In Irregular Regions Using Multi-View Graph Convolutional Networks", "abstract": "Being able to predict the crowd flows in each and every part of a city, especially in irregular regions, is strategically important for traffic control, risk assessment, and public safety. However, it is very challenging because of interactions and spatial correlations between different regions. In addition, it is affected by many factors: i) multiple temporal correlations among different time intervals: closeness, period, trend; ii) complex external influential factors: weather, events; iii) meta features: time of the day, day of the week, and so on. In this paper, we formulate crowd flow forecasting in irregular regions as a spatio-temporal graph (STG) prediction problem in which each node represents a region with time-varying flows. By extending graph convolution to handle the spatial information, we propose using spatial graph convolution to build a multi-view graph convolutional network (MVGCN) for the crowd flow forecasting problem, where different views can capture different factors as mentioned above. We evaluate MVGCN using four real-world datasets (taxicabs and bikes) and extensive experimental results show that our approach outperforms the adaptations of state-of-the-art methods. And we have developed a crowd flow forecasting system for irregular regions that can now be used internally."}, "keywords": ["prediction datasets", "two traffic flow"], "citation_intent": "method"}
{"citing_id": "2303.16974v1", "cited_id": "1910.02655", "section_title": "Conclusion", "citation": "Despite being similar to previous works in structure #REFR and utilizing little in terms of novel improvements, our system was able to achieve SOTA performance on FEVER and the highest label accuracy on SciFact.", "text_before_citation": ["We presented BEVERS, a strong baseline approach for the FEVER and SciFact datasets."], "text_after_citation": ["We primarily attribute these improvements to diligent hyperparameter tuning and error analysis.", "While several previous works have shown novel contributions to portions of the pipeline can yield improvements, in this work we show a well-tuned baseline is very strong."], "citing_paper_content": {"title": "Bevers: A General, Simple, And Performant Framework For Automatic Fact Verification", "abstract": "Automatic fact verification has become an increasingly popular topic in recent years and among datasets the Fact Extraction and VERification (FEVER) dataset is one of the most popular. In this work we present BEVERS, a tuned baseline system for the FEVER dataset. Our pipeline uses standard approaches for document retrieval, sentence selection, and final claim classification, however, we spend considerable effort ensuring optimal performance for each component. The results are that BEV-ERS achieves the highest FEVER score and label accuracy among all systems, published or unpublished. We also apply this pipeline to another fact verification dataset, Scifact, and achieve the highest label accuracy among all systems on that dataset as well. We also make our full code available 1 ."}, "cited_paper_content": {"title": "Bert For Evidence Retrieval And Claim Verification", "abstract": "Motivated by the promising performance of pre-trained language models, we investigate BERT in an evidence retrieval and claim verification pipeline for the FEVER fact extraction and verification challenge. To this end, we propose to use two BERT models, one for retrieving potential evidence sentences supporting or rejecting claims, and another for verifying claims based on the predicted evidence sets. To train the BERT retrieval system, we use pointwise and pairwise loss functions, and examine the effect of hard negative mining. A second BERT model is trained to classify the samples as supported, refuted, and not enough information. Our system achieves a new state of the art recall of 87.1 for retrieving top five sentences out of the FEVER documents consisting of 50K Wikipedia pages, and scores second in the official leaderboard with the FEVER score of 69.7."}, "keywords": ["highest label accuracy"], "citation_intent": "result"}
{"citing_id": "2303.14736v1", "cited_id": "1706.03762", "section_title": "Chinese Handwriting Generation", "citation": "Following the origginal Transformer #REFR , each Transformer layer contains the multi-head attention with c = 512 dimensional states and 8 attention heads.", "text_before_citation": ["We use Dynamic Time Warping (DTW) #OTHEREFR , an elastic matching technique for aligning the given two sequences, to calculate the distance between the generated and real characters.", "Moreover, we use Content Score #OTHEREFR to measure the structure correctness of generated characters, and adopt Style Score #OTHEREFR to quantify the style similarity between the generated and real handwritings.", "We also conduct user preference studies to quantify the subjective quality of the output characters. More details are provided in Appendix A.1.1. Implementation details.", "We set the number of the style reference to K = 15, and resize the reference style and content images to 64\u00d764.", "Moreover, each Transformer encoder contains 2 self-attention layers, while the Transformer decoder has 4 layers for obtaining style representations (2 for writer-wise and 2 for character-wise)."], "text_after_citation": ["Contrary to HWT #OTHEREFR , where F is concatenated before the Transformer encoder, we process each feature sequence f \u2208 F individually.", "Moreover, we apply sinusoidal positional encoding #OTHEREFR to input tokens before feeding them to the Transformer encoder and decoder.", "For training, we first pre-train the content encoder with 138k iterations (batch size of 256) for character classification on training samples and then train the whole model with 148k iterations (batch size of 128), on a single RTX3090 GPU.", "The optimizer is Adam #OTHEREFR , with a learning rate of 0.0002 and gradient clipping of 5.0.", "The sampling ratio \u03b1 is determined through a search over 0.25, 0.5, 0.75, 1.00, and 0.25 is chosen. Following #OTHEREFR , we set \u03bb = 2. Further implementation details are provided in Appendix A.1.2. Compared methods."], "citing_paper_content": {"title": "Disentangling Writer And Character Styles For Handwriting Generation", "abstract": "Training machines to synthesize diverse handwritings is an intriguing task. Recently, RNN-based methods have been proposed to generate stylized online Chinese characters. However, these methods mainly focus on capturing a person's overall writing style, neglecting subtle style inconsistencies between characters written by the same person. For example, while a person's handwriting typically exhibits general uniformity (e.g., glyph slant and aspect ratios), there are still small style variations in finer details (e.g., stroke length and curvature) of characters. In light of this, we propose to disentangle the style representations at both writer and character levels from individual handwritings to synthesize realistic stylized online handwritten characters. Specifically, we present the styledisentangled Transformer (SDT), which employs two complementary contrastive objectives to extract the style commonalities of reference samples and capture the detailed style patterns of each sample, respectively. Extensive experiments on various language scripts demonstrate the effectiveness of SDT. Notably, our empirical findings reveal that the two learned style representations provide information at different frequency magnitudes, underscoring the importance of separate style extraction. Our source code is public at: https://github.com/dailenson/SDT."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["multi-head attention"], "citation_intent": "method"}
{"citing_id": "2303.13203v1", "cited_id": "1710.10589", "section_title": "Details Of The Proposed Model", "citation": "We also evaluated the binary classification of KL-0 vs KL-2 on #REFR modifying the output units of their Siamese-based model from five to two.", "text_before_citation": ["As presented in Section 2.1, instead of using max pooling layers for down-sampling, we used stride and more convolutional layers to preserve more spatial information, which increased the nunber of parameters to a certain extant compared to the work of Tiulpin et al. #OTHEREFR ."], "text_after_citation": ["Using the same dataset, the accuracy reached 86.2% with 0.15M parameters.", "As shown in Table 4 , our proposed Siamese-GAP model achieved an accuracy of 88.38% with 2.71M parameters.", "It is noteworthy that our model only needs a very small number of parameters compared to common evaluated CNN models (i.e., DenseNet, ResNet, and VGG).", "Hence, compared to #OTHEREFR , the greater number of parameters in our proposed model appears to be acceptable, given the significant improvement in performance."], "citing_paper_content": {"title": "A Confident Labelling Strategy Based On Deep Learning For Improving Early Detection Of Knee Osteoarthritis", "abstract": "Knee OsteoArthritis (KOA) is a prevalent musculoskeletal disorder that causes decreased mobility in seniors. The diagnosis provided by physicians is subjective, however, as it relies on personal experience and the semi-quantitative Kellgren-Lawrence (KL) scoring system. KOA has been successfully diagnosed by Computer-Aided Diagnostic (CAD) systems that use deep learning techniques like Convolutional Neural Networks (CNN). In this paper, we propose a novel Siamese-based network, and we introduce a new hybrid loss strategy for the early detection of KOA. The model extends the classical Siamese network by integrating a collection of Global Average Pooling (GAP) layers for feature extraction at each level. Then, to improve the classification performance, a novel training strategy that partitions each training batch into low-, medium-and high-confidence subsets, and a specific hybrid loss function are used for each new label attributed to each sample. The final loss function is then derived by combining the latter loss functions with optimized weights. Our test results demonstrate that our proposed approach significantly improves the detection performance."}, "cited_paper_content": {"title": "Automatic Knee Osteoarthritis Diagnosis From Plain Radiographs: A Deep Learning-Based Approach", "abstract": "Knee osteoarthritis (OA) is the most common musculoskeletal disorder. OA diagnosis is currently conducted by assessing symptoms and evaluating plain radiographs, but this process suffers from subjectivity. In this study, we present a new transparent computer-aided diagnosis method based on the Deep Siamese Convolutional Neural Network to automatically score knee OA severity according to the Kellgren-Lawrence grading scale. We trained our method using the data solely from the Multicenter Osteoarthritis Study and validated it on randomly selected 3,000 subjects (5,960 knees) from Osteoarthritis Initiative dataset. Our method yielded a quadratic Kappa coefficient of 0.83 and average multiclass accuracy of 66.71% compared to the annotations given by a committee of clinical experts. Here, we also report a radiological OA diagnosis area under the ROC curve of 0.93. Besides this, we present attention maps highlighting the radiological features affecting the network decision. Such information makes the decision process transparent for the practitioner, which builds better trust toward automatic methods. We believe that our model is useful for clinical decision making and for OA research; therefore, we openly release our training codes and the data set created in this study."}, "keywords": ["binary classification"], "citation_intent": "method"}
{"citing_id": "2304.03266v1", "cited_id": "1912.04838", "section_title": "Datasets And Evaluation Setting", "citation": "The first scene is from the Waymo Open Dataset (WOD) #REFR , and has a 20-second clip acquired by five pinhole cameras and one 64-beam LiDAR sensor at 10 Hz. We use all five camera views for our experiments.", "text_before_citation": ["In our evaluation, we follow the setting proposed in #OTHEREFR .", "Specifically, we use three scenes for quantitative evaluation and use 13/12/11 sessions respectively to optimize the parameters of our neural scene representation.", "We then use environment maps from five other recording sessions to relight each scene and measure average PSNR and MSE between the rendered and ground-truth images.", "To remove dynamic objects, sky and vegetation pixels we again follow #OTHEREFR and use the segmentation masks predicted by an off-the-shelf semantic segmentation network #OTHEREFR .", "Driving dataset includes two scenes captured by autonomous vehicles (AV) in an urban environment."], "text_after_citation": ["The second set of scenes is also from a high-quality AV dataset (dubbed RoadData) acquired in-house.", "It is captured using eight high resolution (3848x2168 pixel) cameras with calibrated distortion and one 128-beam LiDAR.", "We only use images from the front-facing 120 FoV camera.", "For both scenes, we additionally rely on LiDAR point clouds for depth supervision.", "The Driving dataset is challenging as it records large street environments with complex geometry, lighting, and occlusion, and typically with a fast camera motion."], "citing_paper_content": {"title": "Neural Fields Meet Explicit Geometric Representations For Inverse Rendering Of Urban Scenes", "abstract": "Figure 1. We present FEGR, an approach for reconstructing scene geometry and recovering intrinsic properties of the scene from posed camera images. Our approach works both for single and multi-illumination captured data. FEGR enables various downstream applications such as VR and AR where users may want to control the lighting of the environment and insert desired 3D objects into the scene."}, "cited_paper_content": {"title": "Scalability In Perception For Autonomous Driving: Waymo Open Dataset", "abstract": "The research community has increasing interest in autonomous driving research, despite the resource intensity of obtaining representative real world data. Existing self-driving datasets are limited in the scale and variation of the environments they capture, even though generalization within and between operating regions is crucial to the overall viability of the technology. In an effort to help align the research community's contributions with real-world self-driving problems, we introduce a new large scale, high quality, diverse dataset. Our new dataset consists of 1150 scenes that each span 20 seconds, consisting of well synchronized and calibrated high quality LiDAR and camera data captured across a range of urban and suburban geographies. It is 15x more diverse than the largest camera+LiDAR dataset available based on our proposed diversity metric. We exhaustively annotated this data with 2D (camera image) and 3D (LiDAR) bounding boxes, with consistent identifiers across frames. Finally, we provide strong baselines for 2D as well as 3D detection and tracking tasks. We further study the effects of dataset size and generalization across geographies on 3D detection methods. Find data, code and more up-to-date information at http://www.waymo.com/open."}, "keywords": ["one 64-beam LiDAR", "Waymo Open Dataset"], "citation_intent": "method"}
{"citing_id": "2303.01936v1", "cited_id": "1911.04429", "section_title": "Computer Simulations", "citation": "We set the perturbation bound in #REFR to = 4 for MNIST and = 1.5 for CIFAR10.", "text_before_citation": ["EQUATION", "which is consistent with the perturbation computed from the fast gradient method (FGM) #OTHEREFR .", "In our experiments, we use both the MNIST #OTHEREFR and CIFAR10 #OTHEREFR datasets, and randomly generate a graph with 20 nodes, shown in Fig. 1 .", "We limit our simulations to binary classification in this example.", "For this reason, we consider samples with digits 0 and 1 from MNIST, and images for airplanes and automobiles from CI-FAR10."], "text_after_citation": ["In the test phase, we compute the average classification error across the network to measure the performance of the multi-agent system against perturbations of different strengths.", "We first illustrate the convergence of our algorithm, as anticipated by Theorem 1. From Fig.", "2 , we observe a steady decrease in the classification error towards a limiting value.", "The robust behavior of the proposed algorithm is illustrated in Fig. 3 for both MNIST and CIFAR10.", "We explain the curves for MNIST and similar remarks hold for CIFAR10."], "citing_paper_content": {"title": "Multi-Agent Adversarial Training Using Diffusion Learning", "abstract": "This work focuses on adversarial learning over graphs. We propose a general adversarial training framework for multi-agent systems using diffusion learning. We analyze the convergence properties of the proposed scheme for convex optimization problems, and illustrate its enhanced robustness to adversarial attacks."}, "cited_paper_content": {"title": "Graphdefense: Towards Robust Graph Convolutional Networks", "abstract": "In this paper, we study the robustness of graph convolutional networks (GCNs). Despite the good performance of GCNs on graph semi-supervised learning tasks, previous works have shown that the original GCNs are very unstable to adversarial perturbations. In particular, we can observe a severe performance degradation by slightly changing the graph adjacency matrix or the features of a few nodes, making it unsuitable for security-critical applications. Inspired by the previous works on adversarial defense for deep neural networks, and especially adversarial training algorithm, we propose a method called GraphDefense to defend against the adversarial perturbations. In addition, for our defense method, we could still maintain semi-supervised learning settings, without a large label rate. We also show that adversarial training in features is equivalent to adversarial training for edges with a small perturbation. Our experiments show that the proposed defense methods successfully increase the robustness of Graph Convolutional Networks. Furthermore, we show that with careful design, our proposed algorithm can scale to large graphs, such as Reddit dataset."}, "keywords": ["MNIST"], "citation_intent": "method"}
{"citing_id": "2304.06708v1", "cited_id": "1905.00561", "section_title": "Related Works", "citation": "Similarly to #REFR , we find that training with verb phrase supervision helps for zero-shot performance on tasks with shorter sentences. Temporal understanding in videos.", "text_before_citation": ["#OTHEREFR learn a separate multi-modal embedding space for each PoS tag and then combine these embeddings for fine-grained action retrieval.", "#OTHEREFR construct a hierarchical semantic graph and use graph reasoning for local-global alignments.", "Most closely related to our work, #OTHEREFR use a PoS based token contrastive loss.", "Our work differs in that: (i) we apply a verb phrase contrastive loss, as opposed to separate verb and noun losses; (ii) we extract verb phrases using a LLM and show this performs better than PoS tagging with NLTK #OTHEREFR (Tab.", "5); (iii) we evaluate our methods on verb-focused downstream tasks."], "text_after_citation": ["A long term goal in computer vision is temporal understanding in videos #OTHEREFR .", "However, current training and test datasets have a strong visual bias towards objects and backgrounds as well as static actions #OTHEREFR , with some works #OTHEREFR demonstrating strong results with a single frame.", "Despite these challenges, many recent works in video-only self-supervised learning propose pretext tasks for improving temporal modelling #OTHEREFR .", "Unlike these works that use only video, #OTHEREFR focus on fine-grained temporal videotext alignment via localization of text sub-tokens.", "#OTHEREFR also leverage before/after relations in captions to create artifical training samples for video-text."], "citing_paper_content": {"title": "Verbs In Action: Improving Verb Understanding In Video-Language Models", "abstract": "two brown horses eating grass LLM Original caption Video Add prompt two brown horses running on the grass two brown horses fighting on the grass two brown horses lying on the grass eating grass Verb phrase Hard verb negative captions two brown horses sleeping on the grass two brown horses playing on the grass eating grass cleaning camera standing squatting Verb phrase loss two brown horses eating grass two brown horses lying on the grass a woman standing in a post office person squatting at the gym a man cleaning his camera Hard negative loss Generated hard negative caption Negative captions in the batch Negative verb phrases in the batch Batch Figure 1. Verb-Focused Contrastive (VFC) learning: (Left): Given a video and its corresponding caption, we leverage a Large Language Model (LLM) to output (1) hard negative captions, where only the verb has been changed while keeping the remaining context, and (2) verb phrases which succinctly describe the action in the video. (Right): To encourage better verb reasoning, we subsequently enforce (1) a calibrated hard negative loss, using our generated hard negative captions and other captions in the batch, and (2) a fine-grained, verb phrase loss. We show that VFC improves verb understanding of video-language models compared to the standard contrastive loss."}, "cited_paper_content": {"title": "Large-Scale Weakly-Supervised Pre-Training For Video Action Recognition", "abstract": "Current fully-supervised video datasets consist of only a few hundred thousand videos and fewer than a thousand domain-specific labels. This hinders the progress towards advanced video architectures. This paper presents an in-depth study of using large volumes of web videos for pre-training video models for the task of action recognition. Our primary empirical finding is that pre-training at a very large scale (over 65 million videos), despite on noisy social-media videos and hashtags, substantially improves the state-of-the-art on three challenging public action recognition datasets. Further, we examine three questions in the construction of weakly-supervised video action datasets. First, given that actions involve interactions with objects, how should one construct a verb-object pre-training label space to benefit transfer learning the most? Second, frame-based models perform quite well on action recognition; is pre-training for good image features sufficient or is pre-training for spatio-temporal features valuable for optimal transfer learning? Finally, actions are generally less well-localized in long videos vs. short videos; since action labels are provided at a video level, how should one choose video clips for best performance, given some fixed budget of number or minutes of videos?"}, "keywords": ["verb phrase supervision", "videos"], "citation_intent": "result"}
{"citing_id": "2303.17987v1", "cited_id": "1908.07873", "section_title": "Federated Learning", "citation": "Federated learning #REFR takes the distributed data and trains the model locally for data security.", "text_before_citation": ["This type of federated learning is suitable for scenarios where the samples are different, but the features between the samples are similar.", "\u2022 The applicable scenario of vertical federated learning is the opposite of horizontal federated learning, i.e., the union of features of cross-users under different businesses.", "During the learning process, each participant is unaware of the existence of other parties and only uploads their model parameters, but the prediction requires joint modeling and collaboration to complete the task.", "\u2022 Federated transfer learning can be considered useful when there is no significant overlap between either the samples or their features.", "The key problem is to identify similarities between the source and target domains."], "text_after_citation": ["With a trusted channel established by secure algorithms such as homomorphic encryption, the results of the model can be aggregated to a central node for training to obtain the final training model.", "Federated learning ensures both parity and independence among the participants.", "Moreover, the entire process of learning and training achieves data isolation to prevent data leakage, which well protects the privacy of users #OTHEREFR .", "In such a distributed machine learning framework, the various participants can be coordinated to build more accurate global models. Until now, there are many studies about federated learning."], "citing_paper_content": {"title": "Federated Learning For Metaverse: A Survey", "abstract": "The metaverse, which is at the stage of innovation and exploration, faces the dilemma of data collection and the problem of private data leakage in the process of development. This can seriously hinder the widespread deployment of the metaverse. Fortunately, federated learning (FL) is a solution to the above problems. FL is a distributed machine learning paradigm with privacy-preserving features designed for a large number of edge devices. Federated learning for metaverse (FL4M) will be a powerful tool. Because FL allows edge devices to participate in training tasks locally using their own data, computational power, and model-building capabilities. Applying FL to the metaverse not only protects the data privacy of participants but also reduces the need for high computing power and high memory on servers. Until now, there have been many studies about FL and the metaverse, respectively. In this paper, we review some of the early advances of FL4M, which will be a research direction with unlimited development potential. We first introduce the concepts of metaverse and FL, respectively. Besides, we discuss the convergence of key metaverse technologies and FL in detail, such as big data, communication technology, the Internet of Things, edge computing, blockchain, and extended reality. Finally, we discuss some key challenges and promising directions of FL4M in detail. In summary, we hope that our up-to-date brief survey can help people better understand FL4M and build a fair, open, and secure metaverse. CCS CONCEPTS \u2022 Computing methodologies \u2192 Federated learning."}, "cited_paper_content": {"title": "Federated Learning: Challenges, Methods, And Future Directions", "abstract": "Federated learning involves training statistical models over remote devices or siloed data centers, such as mobile phones or hospitals, while keeping data localized. Training in heterogeneous and potentially massive networks introduces novel challenges that require a fundamental departure from standard approaches for large-scale machine learning, distributed optimization, and privacy-preserving data analysis. In this article, we discuss the unique characteristics and challenges of federated learning, provide a broad overview of current approaches, and outline several directions of future work that are relevant to a wide range of research communities."}, "keywords": ["Federated learning"], "citation_intent": "background"}
{"citing_id": "2305.01177v1", "cited_id": "1905.04273", "section_title": "A.4 Iterative Exponential Mechanism", "citation": "It is folklore that the peeling exponential mechanism is equivalent to taking the top-after adding Gumbel noise, as formally shown in Lemma 4.2 of #REFR .", "text_before_citation": [], "text_after_citation": ["We show a similar property here for AboveThreshold with Gumbel noise that it is equivalent to iteratively running the exponential mechanism."], "citing_paper_content": {"title": "Unbounded Di Erentially Private Quantile And Maximum Estimation", "abstract": "In this work we consider the problem of di erentially private computation of quantiles for the data, especially the highest quantiles such as maximum, but with an unbounded range for the dataset. We show that this can be done e ciently through a simple invocation of AboveThreshold, a subroutine that is iteratively called in the fundamental Sparse Vector Technique, even when there is no upper bound on the data. In particular, we show that this procedure can give more accurate and robust estimates on the highest quantiles with applications towards clipping that is essential for di erentially private sum and mean estimation. In addition, we show how two invocations can handle the fully unbounded data setting. Within our study, we show that an improved analysis of AboveThreshold can improve the privacy guarantees for the widely used Sparse Vector Technique that is of independent interest. We give a more general characterization of privacy loss for AboveThreshold which we immediately apply to our method for improved privacy guarantees. Our algorithm only requires one () pass through the data, which can be unsorted, and each subsequent query takes (1) time. We empirically compare our unbounded algorithm with the state-of-the-art algorithms in the bounded setting. For inner quantiles, we nd that our method often performs better on non-synthetic datasets. For the maximal quantiles, which we apply to di erentially private sum computation, we nd that our method performs signi cantly better."}, "cited_paper_content": {"title": "Practical Differentially Private Top-$K$ Selection With Pay-What-You-Get Composition", "abstract": "We study the problem of top-k selection over a large domain universe subject to user-level differential privacy. Typically, the exponential mechanism or report noisy max are the algorithms used to solve this problem. However, these algorithms require querying the database for the count of each domain element. We focus on the setting where the data domain is unknown, which is different than the setting of frequent itemsets where an apriori type algorithm can help prune the space of domain elements to query. We design algorithms that ensures (approximate) differential privacy and only needs access to the true top-k' elements from the data for any chosen k' \u2265 k. This is a highly desirable feature for making differential privacy practical, since the algorithms require no knowledge of the domain. We consider both the setting where a user's data can modify an arbitrary number of counts by at most 1, i.e. unrestricted sensitivity, and the setting where a user's data can modify at most some small, fixed number of counts by at most 1, i.e. restricted sensitivity. Additionally, we provide a pay-what-you-get privacy composition bound for our algorithms. That is, our algorithms might return fewer than k elements when the top-k elements are queried, but the overall privacy budget only decreases by the size of the outcome set."}, "keywords": ["top"], "citation_intent": "background"}
{"citing_id": "2304.11077v1", "cited_id": "1508.07909", "section_title": "Pre-Training Hero", "citation": "We began by constructing a BPE tokenizer #REFR ) with a vocabulary size of 50,265 tokens using the HeDC4 dataset.", "text_before_citation": [], "text_after_citation": ["The model was trained in two stages: in the first stage the model was trained on four NVIDIA GeForce GTX 1080 Ti 12GB GPUs for 35 days. The checkpoint was then saved for LongHeRo initialization.", "The second stage of training was identical to the first.", "Each stage started with the learning rate of 1e-4 with linear decay and the AdamW optimizer (Loshchilov and Hutter, 2017) ."], "citing_paper_content": {"title": "Hero: Roberta And Longformer Hebrew Language Models", "abstract": "In this paper, we fill in an existing gap in resources available to the Hebrew NLP community by providing it with the largest so far pre-train dataset HeDC4, a state-ofthe-art pre-trained language model HeRo for standard length inputs and an efficient transformer LongHeRo for long input sequences. The HeRo model was evaluated on the sentiment analysis, the named entity recognition, and the question answering tasks while the LongHeRo model was evaluated on the document classification task with a dataset composed of long documents. Both HeRo and LongHeRo presented state-of-the-art performance. The dataset and model checkpoints used in this work are publicly available 1 ."}, "cited_paper_content": {"title": "Neural Machine Translation Of Rare Words With Subword Units", "abstract": "Neural machine translation (NMT) models typically operate with a fixed vocabulary, but translation is an open-vocabulary problem. Previous work addresses the translation of out-of-vocabulary words by backing off to a dictionary. In this paper, we introduce a simpler and more effective approach, making the NMT model capable of open-vocabulary translation by encoding rare and unknown words as sequences of subword units. This is based on the intuition that various word classes are translatable via smaller units than words, for instance names (via character copying or transliteration), compounds (via compositional translation), and cognates and loanwords (via phonological and morphological transformations). We discuss the suitability of different word segmentation techniques, including simple character ngram models and a segmentation based on the byte pair encoding compression algorithm, and empirically show that subword models improve over a back-off dictionary baseline for the WMT 15 translation tasks English!German and English!Russian by up to 1.1 and 1.3 BLEU, respectively."}, "keywords": ["HeDC4 dataset", "vocabulary size"], "citation_intent": "method"}
{"citing_id": "2304.01554v2", "cited_id": "1710.09412", "section_title": "Domain Mixup Module", "citation": "Inspired by the mixup #REFR approach for 2D data, we propose to mix the feature embeddings obtained by F , but from multiple domains in the latent space.", "text_before_citation": [], "text_after_citation": ["Unlike the methods for 2D data where the input images are blended by an alpha factor #OTHEREFR , we propose mixing the feature embeddings, since the feature embeddings from the deeper lay-ers of the network contains information about the global shape of the point cloud and local point-point interaction, as demonstrated in #OTHEREFR applied to STDA set-up.", "Specifically, we linearly interpolate the source (F s ) and target feature (F Ti ) embeddings to obtain F m i and the corresponding mixed soft domain labels L m i as:", "EQUATION", "EQUATION", "where L s and L Ti denote the domain labels of source and target domain which are set to 1 and 0, respectively."], "citing_paper_content": {"title": "Mensa: Mix-Up Ensemble Average For Unsupervised Multi Target Domain Adaptation On 3D Point Clouds", "abstract": "Unsupervised domain adaptation (UDA) addresses the problem of distribution shift between the unlabeled target domain and labelled source domain. While the single target domain adaptation (STDA) is well studied in both 2D and 3D vision literature, multi-target domain adaptation (MTDA) is barely explored for 3D data despite its wide realworld applications such as autonomous driving systems for various geographical and climatic conditions. We establish an MTDA baseline for 3D point cloud data by proposing to mix the feature representations from all domains together to achieve better domain adaptation performance by an ensemble average, which we call Mixup Ensemble Average or MEnsA. With the mixed representation, we use a domain classifier to improve at distinguishing the feature representations of source domain from those of target domains in a shared latent space. In extensive empirical validations on the challenging PointDA-10 dataset, we showcase a clear benefit of our simple method over previous unsupervised STDA and MTDA methods by large margins (up to 17.10% and 4.76% on averaged over all domain shifts). We make the code publicly available here."}, "cited_paper_content": {"title": "Mixup: Beyond Empirical Risk Minimization", "abstract": "Large deep neural networks are powerful, but exhibit undesirable behaviors such as memorization and sensitivity to adversarial examples. In this work, we propose mixup, a simple learning principle to alleviate these issues. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularizes the neural network to favor simple linear behavior in-between training examples. Our experiments on the ImageNet-2012, CIFAR-10, CIFAR-100, Google commands and UCI datasets show that mixup improves the generalization of state-of-the-art neural network architectures. We also find that mixup reduces the memorization of corrupt labels, increases the robustness to adversarial examples, and stabilizes the training of generative adversarial networks."}, "keywords": ["feature embeddings"], "citation_intent": "method"}
{"citing_id": "2303.08485v1", "cited_id": "1603.06212", "section_title": "Automl And Fairness", "citation": "The AutoML system TPOT uses multi-objective optimization to balance pipeline size and performance #REFR .", "text_before_citation": ["In this section we discuss multiobjective and constrained optimization, then describe how it is used to handle fairness as an additional objective, and lastly, highlight other findings that relate to fairness and AutoML.", "Multi-objective and constrained optimization HPO methods, AutoML systems, and NAS methods have been extended to the multi-objective optimization case to take into account additional objectives.", "A related solution is given by constrained optimization, in which one or more objectives are optimized to meet a certain constraint value #OTHEREFR .", "HPO has been used to optimize ML models to take multiple cost functions #OTHEREFR or resource consumption #OTHEREFR into account.", "The goal of multi-objective optimization is to approximate the Pareto set representing the best possible combination of outcomes -solutions that can not be improved w.r.t. one objective without sacrificing performance w.r.t. another."], "text_after_citation": ["AutoXGBoostMC proposes tuning several additional metrics, such as fairness, robustness, or interpretability metrics #OTHEREFR .", "The NAS technique LEMONADE #OTHEREFR searches for neural architectures that balance performance, latency and size. We refer to Morales-Hern\u00e1ndez et al. 2022and #OTHEREFR", "(2022) for reviews of multi-objective HPO and to Benmeziane et al.", "(2021) for a review on hardware-aware NAS, a subfield of multi-objective NAS.", "Fairness as a constraint or additional objective Many fairness-aware ML algorithms consider a constrained optimization approach."], "citing_paper_content": {"title": "Can Fairness Be Automated? Guidelines And Opportunities For Fairness-Aware Automl", "abstract": "The problem is scoped and translated into an ML task including real-world success criteria, requirements, and constraints. 2. Data understanding and preparation. 3 It is determined which data is required, followed by an iterative sequence of data collection, exploration, cleaning, and aggregation."}, "cited_paper_content": {"title": "Evaluation Of A Tree-Based Pipeline Optimization Tool For Automating Data Science", "abstract": "As the field of data science continues to grow, there will be an ever-increasing demand for tools that make machine learning accessible to non-experts. In this paper, we introduce the concept of tree-based pipeline optimization for automating one of the most tedious parts of machine learning--pipeline design. We implement an open source Tree-based Pipeline Optimization Tool (TPOT) in Python and demonstrate its effectiveness on a series of simulated and real-world benchmark data sets. In particular, we show that TPOT can design machine learning pipelines that provide a significant improvement over a basic machine learning analysis while requiring little to no input nor prior knowledge from the user. We also address the tendency for TPOT to design overly complex pipelines by integrating Pareto optimization, which produces compact pipelines without sacrificing classification accuracy. As such, this work represents an important step toward fully automating machine learning pipeline design."}, "keywords": ["multi-objective optimization"], "citation_intent": "method"}
{"citing_id": "2304.04168v1", "cited_id": "1812.10528", "section_title": "O D Formula", "citation": "Inspired by the success of current defensive approaches #REFR , we conclude the properties of operations on graph structure for robustness and design representative defensive operations in our search space accordingly.", "text_before_citation": ["Identity", "M (l) = A (l\u22121) LRA A (l\u22121) = U (l\u22121) S (l\u22121) (V (l\u22121) ) T , M (l) = U (l\u22121) r S (l\u22121) r (V (l\u22121) r ) T NFS m (l) ij = 0, if a (l\u22121) ij > 0 and J ij < \u03c4 a (l\u22121) ij , otherwise NIE m (l) ij = \u03b2m (l\u22121) ij + (1 \u2212 \u03b2)\u03b1 (l) ij VPO M (l) = V v=1 \u03b8 v (A (l\u22121) ) v", "j where m ij = 0 indicates a complete edge pruning and m ij = 1 means no modification to the original edge.", "The defensive operation aims to assign fewer weights to potential perturbed edges."], "text_after_citation": ["In this way, we can choose the most appropriate defensive strategies when confronting perturbed graphs.", "To our best knowledge, this is the first time the search space to be designed with a specific purpose to enhance the robustness of GNNs.", "Specifically, we include five graph structure mask operations in the search space.", "Identity keeps the same graph structure as the previous layer.", "Low Rank Approximation (LRA) reconstructs the adjacency matrix in the l-th layer from the top-r components of singular value decomposition from adjacency matrix in the previous layer."], "citing_paper_content": {"title": "Adversarially Robust Neural Architecture Search For Graph Neural Networks", "abstract": "Graph Neural Networks (GNNs) obtain tremendous success in modeling relational data. Still, they are prone to adversarial attacks, which are massive threats to applying GNNs to risk-sensitive domains. Existing defensive methods neither guarantee performance facing new data/tasks or adversarial attacks nor provide insights to understand GNN robustness from an architectural perspective. Neural Architecture Search (NAS) has the potential to solve this problem by automating GNN architecture designs. Nevertheless, current graph NAS approaches lack robust design and are vulnerable to adversarial attacks. To tackle these challenges, we propose a novel Robust Neural Architecture search framework for GNNs (G-RNA). Specifically, we design a robust search space for the message-passing mechanism by adding graph structure mask operations into the search space, which comprises various defensive operation candidates and allows us to search for defensive GNNs. Furthermore, we define a robustness metric to guide the search procedure, which helps to filter robust architectures. In this way, G-RNA helps understand GNN robustness from an architectural perspective and effectively searches for optimal adversarial robust GNNs. Extensive experimental results on benchmark datasets show that G-RNA significantly outperforms manually designed robust GNNs and vanilla graph NAS baselines by 12.1% to 23.4% under adversarial attacks."}, "cited_paper_content": {"title": "Adversarial Attack And Defense On Graph Data: A Survey", "abstract": "Deep neural networks (DNNs) have been widely applied to various applications including image classification, text generation, audio recognition, and graph data analysis. However, recent studies have shown that DNNs are vulnerable to adversarial attacks. Though there are several works studying adversarial attack and defense strategies on domains such as images and natural language processing, it is still difficult to directly transfer the learned knowledge to graph structure data due to its representation challenges. Given the importance of graph analysis, an increasing number of works start to analyze the robustness of machine learning models on graph data. Nevertheless, current studies considering adversarial behaviors on graph data usually focus on specific types of attacks with certain assumptions. In addition, each work proposes its own mathematical formulation which makes the comparison among different methods difficult. Therefore, in this paper, we aim to survey existing adversarial learning strategies on graph data and first provide a unified formulation for adversarial learning on graph data which covers most adversarial learning studies on graph. Moreover, we also compare different attacks and defenses on graph data and discuss their corresponding contributions and limitations. In this work, we systemically organize the considered works based on the features of each topic. This survey not only serves as a reference for the research community, but also brings a clear image researchers outside this research domain. Besides, we also create an online resource and keep updating the relevant papers during the last two years. More details of the comparisons of various studies based on this survey are open-sourced at https://github.com/YingtongDou/graph-adversarial-learning-literature."}, "keywords": ["graph structure"], "citation_intent": "background"}
{"citing_id": "2303.08250v1", "cited_id": "1904.00310", "section_title": "Introduction", "citation": "To that end, we adapt the learn-to-grow method #REFR , but do not apply the NAS with respect to the four operations uniformly across layers in a ViT.", "text_before_citation": ["It has been observed in neuroscience that learning and memory are entangled together in a highly sophisticated way #OTHEREFR .", "For lifelong learning, the longterm memory (LM) maintained by the hippocampal system plays an important role.", "A question naturally arises: What would be the counterpart, Artificial Hippocampi (Arti-Hippo), in ViTs to facilitate resilient lifelong learning?", "The L2P methods exploit external LM memory to store the learned task prompts/tokens while using frozen Transformer backbones, which may lack the plasticity needed in resilient lifelong learning, increases the cost (due to the quadratic complexity of Transformer models with respect to the number of tokens), and unnecessarily enforces all tasks encountered in lifelong learning (regardless of their underlying difficulty levels) to use the same frozen network.", "In this paper, we seek more integrative memory mechanisms that introduce learnable parts into Transformers (in contrast to be entirely frozen) to induce reconfigurability, selectivity and plasticity in lifelong learning."], "text_after_citation": ["Instead, we aim to find the ArtiHippo inside the ViT.", "As illustrated in Figure 1 , the final projection layer in the multi-head self-attention (MHSA) block of a ViT is identified and selected as the ArtiHippo (Sec. 3.1).", "The learn-to-grow NAS is only applied in maintaining ArtiHippo layers, while other components are frozen to maintain the stability of core functions, as illustrated in Figure 2 .", "Rather than adopting the DARTS #OTHEREFR NAS used in #OTHEREFR , the learn-to-grow NAS in this paper is built on the single-path one-shot (SPOS) NAS #OTHEREFR , in which we propose a hierarchical exploration-exploitation sampling (Figure 3 ) strategy for lifelong learning (Sec. 3.2).", "In experiments, this paper considers lifelong learning with task indices available in both training and inference, which is often referred to as task-incremental setup."], "citing_paper_content": {"title": "Learning To Grow Artificial Hippocampi In Vision Transformers For Resilient Lifelong Learning", "abstract": "Lifelong learning without catastrophic forgetting (i.e., resiliency) possessed by human intelligence is entangled with sophisticated memory mechanisms in the brain, especially the long-term memory (LM) maintained by Hippocampi. To a certain extent, Transformers have emerged as the counterpart \"Brain\" of Artificial Intelligence (AI), and yet leave the LM component under-explored for lifelong learning settings. This paper presents a method of learning to grow Artificial Hippocampi (ArtiHippo) in Vision Transformers (ViTs) for resilient lifelong learning. With a comprehensive ablation study, the final linear projection layer in the multi-head self-attention (MHSA) block is selected in realizing and growing ArtiHippo. ArtiHippo is represented by a mixture of experts (MoEs). Each expert component is an on-site variant of the linear projection layer, which is maintained via neural architecture search (NAS) with the search space defined by four basic growing operationsskip, reuse, adapt, and new in lifelong learning. The LM of a task consists of two parts: the dedicated expert components (as model parameters) at different layers of a ViT learned via NAS, and the mean class-tokens (as stored latent vectors for measuring task similarity) associated with the expert components. For a new task, a hierarchical task-similarity-oriented exploration-exploitation sampling based NAS is proposed to learn the expert components. The task similarity is measured based on the normalized cosine similarity between the mean class-token of the new task and those of old tasks. The proposed method is complementary to prompt-based lifelong learningwith ViTs. In experiments, the proposed method is tested on the challenging Visual Domain Decathlon (VDD) benchmark and the recently proposed 5-Dataset benchmark. It obtains consistently better performance than the prior art with sensible ArtiHippo learned continually."}, "cited_paper_content": {"title": "Learn To Grow: A Continual Structure Learning Framework For Overcoming Catastrophic Forgetting", "abstract": "Addressing catastrophic forgetting is one of the key challenges in continual learning where machine learning systems are trained with sequential or streaming tasks. Despite recent remarkable progress in state-of-the-art deep learning, deep neural networks (DNNs) are still plagued with the catastrophic forgetting problem. This paper presents a conceptually simple yet general and effective framework for handling catastrophic forgetting in continual learning with DNNs. The proposed method consists of two components: a neural structure optimization component and a parameter learning and/or fine-tuning component. By separating the explicit neural structure learning and the parameter estimation, not only is the proposed method capable of evolving neural structures in an intuitively meaningful way, but also shows strong capabilities of alleviating catastrophic forgetting in experiments. Furthermore, the proposed method outperforms all other baselines on the permuted MNIST dataset, the split CIFAR100 dataset and the Visual Domain Decathlon dataset in continual learning setting."}, "keywords": ["learn-to-grow method"], "citation_intent": "method"}
{"citing_id": "2304.07298v1", "cited_id": "1810.02244", "section_title": "Dual-Channel Aggregating Mechanism", "citation": "Therefore, we conduct message aggregation via two channels, namely edge channel and hyperedge channel, and devise a novel dual-channel aggregating mechanism. Note that some existing GNN methods #REFR ; Y.", "text_before_citation": ["An overview of the Dual-channel Aggregating Mechanism is shown in Fig. 2 (the middle part).", "In this module, it iteratively aggregates the information from related roads to encode the structures.", "The intuition is that the edge level neighborhoods in the simple graph G R capture pairwise relationships among roads and local receptive fields, and the hyperedge level neighborhoods in the hypergraph G H capture the complex high-order relationships. Both are indispensable and complementary for road modeling.", "For example, the hyperedge context (e.g., highway loop) and local environment (e.g., residential area) can jointly determine the function of a ramp road."], "text_after_citation": ["Wang and Derr 2021] also model high-order node relationships to help expand the receptive fields and/or improve the expressiveness of GNN model. For example, -GNNs #OTHEREFR", "2019] aggregates the information from high-order neighbors defined by the -order cliques, and TDGNN [Y.", "#OTHEREFR aggregates the information from multi-hop neighbors defined by the tree decomposed subgraphs.", "These methods define different structures on the same simple graph.", "In contrast, our aggregation mechanism takes a dual graph perspective and aggregates information from both simple graph and hypergraph."], "citing_paper_content": {"title": "Road Network Representation Learning: A Dual Graph Based Approach", "abstract": "Road network is a critical infrastructure powering many applications including transportation, mobility and logistics in real life. To leverage the input of a road network across these different applications, it is necessary to learn the representations of the roads in the form of vectors, which is named road network representation learning (RNRL). While several models have been proposed for RNRL, they capture the pairwise relationships/connections among roads only (i.e., as a simple graph), and fail to capture among roads the high-order relationships (e.g., those roads that jointly form a local region usually have similar features such as speed limit) and long-range relationships (e.g., some roads that are far apart may have similar semantics such as being roads in residential areas). Motivated by this, we propose to construct a hypergraph, where each hyperedge corresponds to a set of multiple roads forming a region. The constructed hypergraph would naturally capture the high-order relationships among roads with hyperedges. We then allow information propagation via both the edges in the simple graph and the hyperedges in the hypergraph in a graph neural network context. In addition, we introduce different pretext tasks based on both the simple graph (i.e., graph reconstruction) and the hypergraph (including hypergraph reconstruction and hyperedge classification) for optimizing the representations of roads. The graph reconstruction and hypergraph reconstruction tasks are conventional ones and can capture structural information. The hyperedge classification task can capture long-range relationships between pairs of roads that belong to hyperedges with the same label. We call the resulting model HyperRoad. We further extend HyperRoad to problem settings when additional inputs of road attributes and/or trajectories that are generated on the roads are available. We conduct extensive experiments on two real datasets, for five downstream tasks, and under four problem settings, which demonstrate that our model achieves impressive improvements compared with existing baselines across datasets, tasks, problem settings and performance metrics."}, "cited_paper_content": {"title": "Weisfeiler And Leman Go Neural: Higher-Order Graph Neural Networks", "abstract": "In recent years, graph neural networks (GNNs) have emerged as a powerful neural architecture to learn vector representations of nodes and graphs in a supervised, end-to-end fashion. Up to now, GNNs have only been evaluated empirically\u2014showing promising results. The following work investigates GNNs from a theoretical point of view and relates them to the 1-dimensional Weisfeiler-Leman graph isomorphism heuristic (1-WL). We show that GNNs have the same expressiveness as the 1-WL in terms of distinguishing non-isomorphic (sub-)graphs. Hence, both algorithms also have the same shortcomings. Based on this, we propose a generalization of GNNs, so-called k-dimensional GNNs (k-GNNs), which can take higher-order graph structures at multiple scales into account. These higher-order structures play an essential role in the characterization of social networks and molecule graphs. Our experimental evaluation confirms our theoretical findings as well as confirms that higher-order information is useful in the task of graph classification and regression."}, "keywords": ["hyperedge channel", "message aggregation"], "citation_intent": "method"}
{"citing_id": "2303.15616v1", "cited_id": "1812.01717", "section_title": "Video Generation", "citation": "FVD #REFR is used to measure the generated video from captions and our fine-grained descriptions.", "text_before_citation": ["Quantitative results.", "To compare the performance between the caption-generated and the description-generated videos, We report the quantitative results over Frechet Video Distance (FVD) #OTHEREFR metric in Table 5 .", "We employ Cogvideo #OTHEREFR to generate the video based on captions and our fine-grained descriptions.", "The FVD score reflects that our fine-grained description can generate videos that are more close to the referenced videos. Qualitative results. In Fig. 10 and Fig. 11 , we provide Table 5 . Video generation performance on FAVDBench."], "text_after_citation": ["As the text-video model is used to directly infer without fine-tuning, the score is higher than normal ranges. Besides, the smaller FVD is the better."], "citing_paper_content": {"title": "Fine-Grained Audible Video Description", "abstract": "We explore a new task for audiovisual language modeling called fine-grained audible video description (FAVD). It aims to provide detailed textual descriptions for the given audible videos, including the appearance and spatial locations of each object, the actions of moving objects, and the sounds in videos. Existing visual-language modeling tasks often concentrate on visual cues in videos while undervaluing the language and audio modalities. On the other hand, FAVD requires not only audiovisual language modeling skills but also paragraph-level language generation abilities. We construct the first fine-grained audible video description benchmark (FAVDBench) to facilitate this research. For each video clip, we first provide a one-sentence summary of the video, i.e., the caption, followed by 4-6 sentences describing the visual details and 1-2 audio-related descriptions at the end. The descriptions are provided in both English and Chinese. We create two new metrics for this task: an EntityScore to gauge the completeness of entities in the visual descriptions, and an Au-dioScore to assess the audio descriptions. As a preliminary approach to this task, we propose an audio-visuallanguage transformer that extends existing video captioning model with an additional audio branch. We combine the masked language modeling and auto-regressive language modeling losses to optimize our model so that it can produce paragraph-level descriptions. We illustrate the efficiency of our model in audiovisual language modeling by evaluating it against the proposed benchmark using both conventional captioning metrics and our proposed metrics. We further put our benchmark to the test in video generation models, demonstrating that employing finegrained video descriptions can create more intricate videos than using captions."}, "cited_paper_content": {"title": "Towards Accurate Generative Models Of Video: A New Metric&Challenges", "abstract": "Recent advances in deep generative models have lead to remarkable progress in synthesizing high quality images. Following their successful application in image processing and representation learning, an important next step is to consider videos. Learning generative models of video is a much harder task, requiring a model to capture the temporal dynamics of a scene, in addition to the visual presentation of objects. Although recent attempts at formulating generative models of video have had some success, current progress is hampered by (1) the lack of qualitative metrics that consider visual quality, temporal coherence, and diversity of samples, and (2) the wide gap between purely synthetic video datasets and challenging real-world datasets in terms of complexity. To this extent we propose Frechet Video Distance (FVD), a new metric for generative models of video based on FID, and StarCraft 2 Videos (SCV), a collection of progressively harder datasets that challenge the capabilities of the current iteration of generative models for video. We conduct a large-scale human study, which confirms that FVD correlates well with qualitative human judgment of generated videos, and provide initial benchmark results on SCV."}, "keywords": ["fine-grained descriptions", "generated video"], "citation_intent": "method"}
{"citing_id": "2305.02350v1", "cited_id": "1810.04805", "section_title": "Arxiv:2305.02350V1 [Cs.Cl] 3 May 2023 2 Related Work", "citation": "Typically, a pre-trained language model is used, attaching an extra layer for projection to task-specific target variables, and fine-tuning both #REFR .", "text_before_citation": ["Currently, text classification is mostly tackled using neural networks.", "Virtually every state-of-the-art result for text classification has been achieved using neural language models #OTHEREFR ."], "text_after_citation": ["This either requires graphics cards with large memory, or low batch sizes.", "A substantial amount of research focuses on methods to downsize these pre-trained language models #OTHEREFR .", "But, while these approaches reduce the number of parameters drastically, the memory during fine-tuning can still exceed 10 GiB due to gradient calculations.", "Prior to language models, the most popular neural approach was to leverage pre-trained word embeddings, like GloVe #OTHEREFR and Word2Vec #OTHEREFR , feed them to task-specific neural architectures #OTHEREFR , and then train on the task data at hand.", "Although pre-trained word embeddings also range at around 100M parameters, they are computationally efficient, since no gradient calculations are required."], "citing_paper_content": {"title": "Using Language Models On Low-End Hardware", "abstract": "This paper evaluates the viability of using fixed language models for training text classification networks on low-end hardware. We combine language models with a CNN architecture and put together a comprehensive benchmark with 8 datasets covering single-label and multi-label classification of topic, sentiment, and genre. Our observations are distilled into a list of trade-offs, concluding that there are scenarios, where not fine-tuning a language model yields competitive effectiveness at faster training, requiring only a quarter of the memory compared to fine-tuning."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["pre-trained language model"], "citation_intent": "method"}
{"citing_id": "2304.06354v1", "cited_id": "1710.11063", "section_title": "Odam: Object Detector Activation Maps", "citation": "In Grad-CAM (Selvaraju et al., 2017) and its generalization Grad-CAM++ #REFR , the final score for a particular class Y c is predicted from the whole image and the algorithm Figure 2: Our proposed framework.", "text_before_citation": ["EQUATION", "1 , x", "2 , y", "2 ).", "Our goal is to generate heat maps to indicate the important regions that have a positive influence on the output of each prediction."], "text_after_citation": ["(a) ODAM generates instance-specific heat maps for explaining the predictions of an object detector; (b) ODAM-Train uses auxiliary losses to encourage heat maps for predictions on the same object to be consistent, and for different objects to be distinct; (c) ODAM-NMS uses the box IoU and the normalized correlation between heat maps to determine the duplicate detections.", "The bounding box shows the detection proposal corresponding to the heat map.", "ignores distinguishing object instances within.", "Their explanation starts from the assumption that the score to be interpreted can be written as a linear combination of its global pooled last convolutional layer feature maps", "{A k } k , Y c = k w c k ij A ijk = ij k w c k A ijk , where A ijk indexes location (i, j) of A k ."], "citing_paper_content": {"title": "Odam: Gradient-Based Instance-Specific Vi-Sual Explanations For Object Detection", "abstract": "We propose the gradient-weighted Object Detector Activation Maps (ODAM), a visualized explanation technique for interpreting the predictions of object detectors. Utilizing the gradients of detector targets flowing into the intermediate feature maps, ODAM produces heat maps that show the influence of regions on the detector's decision for each predicted attribute. Compared to previous works classification activation maps (CAM), ODAM generates instance-specific explanations rather than class-specific ones. We show that ODAM is applicable to both one-stage detectors and two-stage detectors with different types of detector backbones and heads, and produces higher-quality visual explanations than the state-of-the-art both effectively and efficiently. We next propose a training scheme, Odam-Train, to improve the explanation ability on object discrimination of the detector through encouraging consistency between explanations for detections on the same object, and distinct explanations for detections on different objects. Based on the heat maps produced by ODAM with Odam-Train, we propose Odam-NMS, which considers the information of the model's explanation for each prediction to distinguish the duplicate detected objects. We present a detailed analysis of the visualized explanations of detectors and carry out extensive experiments to validate the effectiveness of the proposed ODAM."}, "cited_paper_content": {"title": "Grad-Cam++: Improved Visual Explanations For Deep Convolutional Networks", "abstract": "Over the last decade, Convolutional Neural Network (CNN) models have been highly successful in solving complex vision based problems. However, deep models are perceived as \"black box\" methods considering the lack of understanding of their internal functioning. There has been a significant recent interest to develop explainable deep learning models, and this paper is an effort in this direction. Building on a recently proposed method called Grad-CAM, we propose Grad-CAM++ to provide better visual explanations of CNN model predictions (when compared to Grad-CAM), in terms of better localization of objects as well as explaining occurrences of multiple objects of a class in a single image. We provide a mathematical explanation for the proposed method, Grad-CAM++, which uses a weighted combination of the positive partial derivatives of the last convolutional layer feature maps with respect to a specific class score as weights to generate a visual explanation for the class label under consideration. Our extensive experiments and evaluations, both subjective and objective, on standard datasets showed that Grad-CAM++ indeed provides better visual explanations for a given CNN architecture when compared to Grad-CAM."}, "keywords": ["whole image"], "citation_intent": "method"}
{"citing_id": "2304.03646v1", "cited_id": "1912.06883", "section_title": "Experimental Results", "citation": "Our results complement the findings by Binns #REFR , who suggest that group and individual fairness may not always be conflicting objectives.", "text_before_citation": ["For GAIA LS and GAIA ALS, the disparity between fairness is less pronounced since both versions perform comparatively better than random chance. Fig.", "3 also illustrates the value of weighted training which considers both utility and fairness objectives (GAIA).", "Even though BNN with distribution shift by itself shows competitive performance compared to the baselines, GAIA consistently outperforms the BNN in terms of utility while matching it in terms of fairness.", "This consistent improvement is more pronounced in the Adult dataset, where there are more samples for GAIA to leverage the disparity between ambiguous and non-ambiguous subsets of data. Our experimental results highlight the viability of GAIA (RQ1).", "Since prior works have drawn a distinction between group and individual fairness #OTHEREFR , we study how our approach compares against both (RQ2)."], "text_after_citation": ["Our experiments suggest that ALS introduces an improvement over LS, with a more pronounced improvement in the German dataset compared to the Adult dataset.", "In addition, we observe that GAIA outperforms BNN consistently in terms of utility, while the BNN has a minuscule advantage in terms of fairness.", "The BNN has a coherent representation due to the regularization effect of the variational inference on the encoding space, where the encoder must output a probabilistic distribution over the latent variables that approximates the true posterior.", "This encourages similar samples to have similar encodings, leading to a more organized and smoother latent space representation.", "Therefore, it is not surprising that BNNs demonstrate high performance on individual fairness metrics, as these metrics evaluate the consistency in the treatment of similar covariates."], "citing_paper_content": {"title": "Fairness Through Aleatoric Uncertainty", "abstract": "We propose a unique solution to tackle the often-competing goals of fairness and utility in machine learning classification tasks. While fairness ensures that the model's predictions are unbiased and do not discriminate against any particular group, utility focuses on maximizing the accuracy of the model's predictions. Our aim is to investigate the relationship between uncertainty and fairness. Our approach leverages this concept by employing Bayesian learning to estimate the uncertainty in sample predictions where the estimation is independent of confounding effects related to the protected attribute. Through empirical evidence, we show that samples with low classification uncertainty are modeled more accurately and fairly than those with high uncertainty, which may have biased representations and higher prediction errors. To address the challenge of balancing fairness and utility, we propose a novel fairness-utility objective that is defined based on uncertainty quantification. The weights in this objective are determined by the level of uncertainty, allowing us to optimize both fairness and utility simultaneously. Experiments on real-world datasets demonstrate the effectiveness of our approach. Our results show that our method outperforms state-of-the-art methods in terms of the fairness-utility tradeoff and this applies to both group and individual fairness metrics. This work presents a fresh perspective on the trade-off between accuracy and fairness in machine learning and highlights the potential of using uncertainty as a means to achieve optimal fairness and utility. Preprint. Under review."}, "cited_paper_content": {"title": "On The Apparent Conflict Between Individual And Group Fairness", "abstract": "A distinction has been drawn in fair machine learning research between `group' and `individual' fairness measures. Many technical research papers assume that both are important, but conflicting, and propose ways to minimise the trade-offs between these measures. This paper argues that this apparent conflict is based on a misconception. It draws on theoretical discussions from within the fair machine learning research, and from political and legal philosophy, to argue that individual and group fairness are not fundamentally in conflict. First, it outlines accounts of egalitarian fairness which encompass plausible motivations for both group and individual fairness, thereby suggesting that there need be no conflict in principle. Second, it considers the concept of individual justice, from legal philosophy and jurisprudence which seems similar but actually contradicts the notion of individual fairness as proposed in the fair machine learning literature. The conclusion is that the apparent conflict between individual and group fairness is more of an artifact of the blunt application of fairness measures, rather than a matter of conflicting principles. In practice, this conflict may be resolved by a nuanced consideration of the sources of `unfairness' in a particular deployment context, and the carefully justified application of measures to mitigate it."}, "keywords": ["individual fairness"], "citation_intent": "result"}
{"citing_id": "2304.03113v1", "cited_id": "0908.3817", "section_title": "C.1 Overview Of Graphs", "citation": "Table 2 shows the hyperparameter settings used for CSL relying on the bnlearn package #REFR for R (R Core Team, 2022).", "text_before_citation": ["In Table 1 we provide an overview of all twelve graphs used in Section 5, the randomly sampled target, the adjacency degree of the target and the share of d-separations w.r.t. the target.", "This gives further insight into the relation of graph sparsity, degree of target and share of d-separations.", "The latter can be regarded as the potential relative runtime decrease for SAGE approximation."], "text_after_citation": ["In Algorithm 2 we explicate how we inferred the number of true positive, false positive, true negative and false negative d-separations within an estimated graph and especially how we dealt with the exponential number of potential conditioning sets for the larger graphs."], "citing_paper_content": {"title": "Efficient Sage Estimation Via Causal Structure Learning", "abstract": "The Shapley Additive Global Importance (SAGE) value is a theoretically appealing interpretability method that fairly attributes global importance to a model's features. However, its exact calculation requires the computation of the feature's surplus performance contributions over an exponential number of feature sets. This is computationally expensive, particularly because estimating the surplus contributions requires sampling from conditional distributions. Thus, SAGE approximation algorithms only take a fraction of the feature sets into account. We propose d-SAGE, a method that accelerates SAGE approximation. d-SAGE is motivated by the observation that conditional independencies (CIs) between a feature and the model target imply zero surplus contributions, such that their computation can be skipped. To identify CIs, we leverage causal structure learning (CSL) to infer a graph that encodes (conditional) independencies in the data as d-separations. This is computationally more efficient because the expense of the one-time graph inference and the d-separation queries is negligible compared to the expense of surplus contribution evaluations. Empirically we demonstrate that d-SAGE enables the efficient and accurate estimation of SAGE values."}, "cited_paper_content": {"title": "Learning Bayesian Networks With The Bnlearn R Package", "abstract": "bnlearn is an R package (R Development Core Team 2010) which includes several algorithms for learning the structure of Bayesian networks with either discrete or continuous variables. Both constraint-based and score-based algorithms are implemented, and can use the functionality provided by the snow package (Tierney et al. 2008) to improve their performance via parallel computing. Several network scores and conditional independence algorithms are available for both the learning algorithms and independent use. Advanced plotting options are provided by the Rgraphviz package (Gentry et al. 2010)."}, "keywords": ["bnlearn package"], "citation_intent": "method"}
{"citing_id": "2303.14139v1", "cited_id": "1510.06479", "section_title": "Comparison With State-Of-The-Art Models", "citation": "We speculate that this could be due to the fact that NSD contains more complex natural scenes compared to the previous Generic Object Decoding ( GOD ) dataset #REFR .", "text_before_citation": ["The second and third lines represent the results obtained using our proposed MindDiffuser with and without stage 2, respectively.", "The fourth and fifth lines represent the results obtained by reproducing the experiment described in Takagi's #OTHEREFR paper and by using Ozcelik's #OTHEREFR provided random version code, respectively.", "compared to recent work, our approach produces reconstructed results on NSD that are visually more similar to the ground truth both in terms of semantic and structure.", "This suggests that our approach outperforms existing methods in controllable image reconstruction on NSD.", "We only have access to the random version code released by Ozcelik #OTHEREFR , and based on their paper, the quantitative results between the random and dense versions are similar. However, this model did not perform well on NSD."], "text_after_citation": ["Extracted instance features using only self-supervised pre-trained feature extractor SWAV #OTHEREFR on the visual modality may not contain complete semantic and structural information, leading to unsatisfactory results even if the decoding accuracy of these features is high (with an average accuracy of around 0.6 on the test set). It has been observed that although Takagi et al.", "#OTHEREFR and MindDiffuser without stage 2 are able to achieve semantically similar reconstruction results to the original images, such as airplanes, plants on a table, people surfing at a beach, bathrooms, and people skiing, it is clear that these images cannot be aligned with the original images in terms of details such as shape and orientation.", "This indicates that the CLIP-based modulation used in stage 2 of this paper can effectively reconstruct results that are both semantically and structurally aligned with the original images.", "In order to further quantitatively compare the reconstruction performance of our method with some state-of-the-art methods, we utilize three evaluation metrics to compare from both semantic and structural aspects.", "We use cosine similarity in the last layer of CLIP visual branch (512 dimensions) to measure the semantic similarity between the reconstructed results and the original images."], "citing_paper_content": {"title": "Minddiffuser: Controlled Image Reconstruction From Human Brain Activity With Semantic And Structural Diffusion", "abstract": "Reconstructing visual stimuli from measured functional magnetic resonance imaging (fMRI) has been a meaningful and challenging task. Previous studies have successfully achieved reconstructions with structures similar to the original images, such as the outlines and size of some natural images. However, these reconstructions lack explicit semantic information and are difficult to discern. In recent years, many studies have utilized multi-modal pre-trained models with stronger generative capabilities to reconstruct images that are semantically similar to the original ones. However, these images have uncontrollable structural information such as position and orientation. To address both of the aforementioned issues simultaneously, we propose a two-stage image reconstruction model called MindDiffuser, utilizing Stable Diffusion. In Stage 1, the VQ-VAE latent representations and the CLIP text embeddings decoded from fMRI are put into the image-to-image process of Stable Diffusion, which yields a preliminary image that contains semantic and structural information. In Stage 2, we utilize the low-level CLIP visual features decoded from fMRI as supervisory information, and continually adjust the two features in Stage 1 through backpropagation to align the structural information. The results of both qualitative and quantitative analyses demonstrate that our proposed model has surpassed the current state-of-the-art models in terms of reconstruction results on Natural Scenes Dataset (NSD). Furthermore, the results of ablation experiments indicate that each component of our model is effective for image reconstruction."}, "cited_paper_content": {"title": "Generic Decoding Of Seen And Imagined Objects Using Hierarchical Visual Features", "abstract": "Object recognition is a key function in both human and machine vision. While recent studies have achieved fMRI decoding of seen and imagined contents, the prediction is limited to training examples. We present a decoding approach for arbitrary objects, using the machine vision principle that an object category is represented by a set of features rendered invariant through hierarchical processing. We show that visual features including those from a convolutional neural network can be predicted from fMRI patterns and that greater accuracy is achieved for low/high-level features with lower/higher-level visual areas, respectively. Predicted features are used to identify seen/imagined object categories (extending beyond decoder training) from a set of computed features for numerous object images. Furthermore, the decoding of imagined objects reveals progressive recruitment of higher to lower visual representations. Our results demonstrate a homology between human and machine vision and its utility for brain-based information retrieval."}, "keywords": ["complex natural scenes"], "citation_intent": "result"}
{"citing_id": "2303.04445v2", "cited_id": "1912.07418", "section_title": "I. Introduction", "citation": "In this sense, our results represent a substantial generalization of the work in #REFR while maintaining the core features of L 0/1 -SVM.", "text_before_citation": ["Recently in the literature, there is a resurging interest in the original SVM problem with the (0, 1) loss, abbreviated as \"L 0/1 -SVM\", following theoretical and algorithmic developments for optimization problems with the \"\u2113 0 -norm\", see e.g., #OTHEREFR and the references therein.", "In particular, #OTHEREFR proposed KKT-like optimality conditions for the L 0/1 -SVM optimization problem and a efficient ADMM solver to obtain an approximate solution.", "In this work, we draw inspiration from the aforementioned papers and present a kernelized version of the theory in which the ambient functional space has a richer structure than the usual Euclidean space.", "More precisely, we shall formulate the L 0/1 -SVM problem in the context of Multiple Kernel Learning #OTHEREFR and describe a first-order optimality theory as well as a numerical procedure for the optimization problem via the ADMM.", "Obviously, the MKL framework can offer much more flexibility than the single-kernel formulation by letting the optimization algorithm determine the best combination of different kernel functions."], "text_after_citation": ["The remainder of this paper is organized as follows.", "Section II reviews the classic L 0/1 -SVM in the singlekernel case, while Section III discusses the MKL framework.", "Section IV establishes the optimality theory for the MKL-L 0/1 -SVM problem, and in Section V we propose an ADMM algorithm to solve the optimization problem.", "Finally, numerical experiments and concluding remarks are provided in Sections VI and VII, respectively.", "Notation R + denotes the set of nonnegative reals, and R n + := R + \u00d7 \u2022 \u2022 \u2022 \u00d7 R + the n-fold Cartesian product. N m := {1, 2, . . ."], "citing_paper_content": {"title": "", "abstract": "We formulate the Multiple Kernel Learning (abbreviated as MKL) problem for the support vector machine with the infamous (0, 1)-loss function. Some first-order optimality conditions are given and then exploited to develop a fast ADMM solver for the nonconvex and nonsmooth optimization problem. A simple numerical experiment on synthetic planar data shows that our MKL-L 0/1-SVM framework could be promising."}, "cited_paper_content": {"title": "Support Vector Machine Classifier Via $L_{0/1}$ Soft-Margin Loss", "abstract": "Support vector machine (SVM) has attracted great attentions for the last two decades due to its extensive applications, and thus numerous optimization models have been proposed. To distinguish all of them, in this paper, we introduce a new model equipped with an $L_{0/1}$ soft-margin loss (dubbed as $L_{0/1}$-SVM) which well captures the nature of the binary classification. Many of the existing convex/non-convex soft-margin losses can be viewed as a surrogate of the $L_{0/1}$ soft-margin loss. Despite the discrete nature of $L_{0/1}$, we manage to establish the existence of global minimizer of the new model as well as revealing the relationship among its minimizers and KKT/P-stationary points. These theoretical properties allow us to take advantage of the alternating direction method of multipliers. In addition, the $L_{0/1}$-support vector operator is introduced as a filter to prevent outliers from being support vectors during the training process. Hence, the method is expected to be relatively robust. Finally, numerical experiments demonstrate that our proposed method generates better performance in terms of much shorter computational time with much fewer number of support vectors when against with some other leading methods in areas of SVM. When the data size gets bigger, its advantage becomes more evident."}, "keywords": ["-SVM"], "citation_intent": "result"}
{"citing_id": "2303.10449v1", "cited_id": "1812.04606", "section_title": "Ood Detection With Extra Data", "citation": "OE #REFR uses purified unlabeled OOD data, which is encouraged to produce a uniform softmax distribution during training.", "text_before_citation": ["The mainstream OOD detection works primarily focus on detecting samples with semantic shifts, i.e.OOD samples drawn from classes that do not overlap with known classes.", "To enhance the ability to understand ID/OOD discrepancy, some methods synthesize OOD training data such as #OTHEREFR , and another series of OOD detection methods introduce extra OOD data into training."], "text_after_citation": ["MCD [32] maximizes the entropy discrepancy between ID and extra OOD samples outputted by two parallel classifiers.", "The survey #OTHEREFR indicates that OOD detection with external data generally performs better.", "However, massive unlabeled samples could be very easily obtained considering their accessibility, and ID data would inevitably exist among them.", "Hence, 'purified unlabeled OOD data' means the mixed ID data need to be manually removed, which brings a high cost and is impractical.", "Against the above issues, UDG #OTHEREFR proposes a realistic benchmark that retains ID samples in the unlabeled set, namely SCOOD (Semantically Coherent Out-of-Distribution Detection) benchmarks."], "citing_paper_content": {"title": "Uncertainty-Aware Optimal Transport For Semantically Coherent Out-Of-Distribution Detection", "abstract": "Semantically coherent out-of-distribution (SCOOD) detection aims to discern outliers from the intended data distribution with access to unlabeled extra set. The coexistence of in-distribution and out-of-distribution samples will exacerbate the model overfitting when no distinction is made. To address this problem, we propose a novel uncertainty-aware optimal transport scheme. Our scheme consists of an energy-based transport (ET) mechanism that estimates the fluctuating cost of uncertainty to promote the assignment of semantic-agnostic representation, and an inter-cluster extension strategy that enhances the discrimination of semantic property among different clusters by widening the corresponding margin distance. Furthermore, a T-energy score is presented to mitigate the magnitude gap between the parallel transport and classifier branches. Extensive experiments on two standard SCOOD benchmarks demonstrate the above-par OOD detection performance, outperforming the state-of-the-art methods by a margin of 27.69% and 34.4% on FPR@95, respectively. Code is available at https://github.com/LuFan31/ET-OOD. * Co-first Author. \u2020Corresponding Author. \u2021Work done during an internship at Ant Group."}, "cited_paper_content": {"title": "Deep Anomaly Detection With Outlier Exposure", "abstract": "It is important to detect anomalous inputs when deploying machine learning systems. The use of larger and more complex inputs in deep learning magnifies the difficulty of distinguishing between anomalous and in-distribution examples. At the same time, diverse image and text data are available in enormous quantities. We propose leveraging these data to improve deep anomaly detection by training anomaly detectors against an auxiliary dataset of outliers, an approach we call Outlier Exposure (OE). This enables anomaly detectors to generalize and detect unseen anomalies. In extensive experiments on natural language processing and small- and large-scale vision tasks, we find that Outlier Exposure significantly improves detection performance. We also observe that cutting-edge generative models trained on CIFAR-10 may assign higher likelihoods to SVHN images than to CIFAR-10 images; we use OE to mitigate this issue. We also analyze the flexibility and robustness of Outlier Exposure, and identify characteristics of the auxiliary dataset that improve performance."}, "keywords": ["unlabeled OOD data"], "citation_intent": "method"}
{"citing_id": "2303.15681v1", "cited_id": "1905.05178", "section_title": "Methods", "citation": "This approach resembles the graph U-Net architecture used in #REFR for image processing tasks such as image classification and segmentation.", "text_before_citation": ["This edge augmentation significantly reduces the number of hops required for the nodes to gather messages from distant nodes, thus resulting in a smaller number of GN blocks for faster message propagation.", "(b) Multi Graph: This approach draws inspiration from multigrid methods for PDEs, whose foundation is a hierarchy of discretizations (or meshes) 40 .", "By performing a series of graph down-sampling operations followed by up-sampling operations, we apply the multigrid concept to graphs and pass information across the computational domain in a hierarchical manner.", "In practice, the down-sampling operations correspond to mesh coarsening and result in smaller graphs connecting points that are far away in the domain.", "On the other hand, during up-sampling the information is redistributed across the entire graph."], "text_after_citation": ["For both approaches, we assume that the computational domain is a triangular mesh which undergoes a transformation into a graph as detailed in Section 4.1 and illustrated in Figure 1 ."], "citing_paper_content": {"title": "Gnn-Based Physics Solver For Time-Independent Pdes", "abstract": "Physics-based deep learning frameworks have shown to be effective in accurately modeling the dynamics of complex physical systems with generalization capability across problem inputs. However, time-independent problems pose the challenge of requiring long-range exchange of information across the computational domain for obtaining accurate predictions. In the context of graph neural networks (GNNs), this calls for deeper networks, which, in turn, may compromise or slow down the training process. In this work, we present two GNN architectures to overcome this challenge-the Edge Augmented GNN and the Multi-GNN. We show that both these networks perform significantly better (by a factor of 1.5 to 2) than baseline methods when applied to time-independent solid mechanics problems. Furthermore, the proposed architectures generalize well to unseen domains, boundary conditions, and materials. Here, the treatment of variable domains is facilitated by a novel coordinate transformation that enables rotation and translation invariance. By broadening the range of problems that neural operators based on graph neural networks can tackle, this paper provides the groundwork for their application to complex scientific and industrial settings."}, "cited_paper_content": {"title": "Graph U-Nets", "abstract": "We consider the problem of representation learning for graph data. Convolutional neural networks can naturally operate on images, but have significant challenges in dealing with graph data. Given images are special cases of graphs with nodes lie on 2D lattices, graph embedding tasks have a natural correspondence with image pixel-wise prediction tasks such as segmentation. While encoder-decoder architectures like U-Nets have been successfully applied on many image pixel-wise prediction tasks, similar methods are lacking for graph data. This is due to the fact that pooling and up-sampling operations are not natural on graph data. To address these challenges, we propose novel graph pooling (gPool) and unpooling (gUnpool) operations in this work. The gPool layer adaptively selects some nodes to form a smaller graph based on their scalar projection values on a trainable projection vector. We further propose the gUnpool layer as the inverse operation of the gPool layer. The gUnpool layer restores the graph into its original structure using the position information of nodes selected in the corresponding gPool layer. Based on our proposed gPool and gUnpool layers, we develop an encoder-decoder model on graph, known as the graph U-Nets. Our experimental results on node classification and graph classification tasks demonstrate that our methods achieve consistently better performance than previous models."}, "keywords": ["graph U-Net architecture"], "citation_intent": "method"}
{"citing_id": "2305.02567v1", "cited_id": "1505.04597", "section_title": "Introduction", "citation": "First, diffusion models often use convolutional neural networks such as U-Net #REFR to learn the reverse process to construct desired data samples from the noise.", "text_before_citation": ["VAEs can capture more diversity and are typically easier to scale and train than GANs, but still fall short in terms of visual sample quality and sampling efficiency #OTHEREFR .", "Recently, diffusion models such as denoising diffusion probabilistic model (DDPM) #OTHEREFR have emerged as a powerful class of generative models, capable of producing highquality images comparable to those of GANs.", "Importantly, they additionally offer desirable properties such as strong sample diversity, faithful distribution coverage, a stationary training objective, and easy scalability.", "This implies that diffusion models are well suited for learning models of complex and diverse data, which also motivates us to explore the potential of diffusion-based generative models for graphic layout generation.", "Though diffusion models have shown splendid performance in high-fidelity image generation #OTHEREFR , it is still a sparsely explored area and provides unique challenges to develop diffusion-based generative models for layout generation."], "text_after_citation": ["However, a layout is a non-sequential data structure consisting of varying length samples with discrete (classes) and continuous (coordinates) elements simultaneously, instead of pixels laid on a regular lattice.", "Obviously, convolutional neural networks are not suitable for layout denoising, which prevents diffusion models from being directly applied to layout generation.", "Second, the placement and sizing of a given element depend not only on its attributes (such as category label) but also on its relationship to other elements.", "How to incorporate the attributes knowledge and model the elements' relationship in diffusion models is still an open problem.", "Since diffusion models are general frameworks, they leave room for adapting the underlying neural architectures to exploit the properties of the data."], "citing_paper_content": {"title": "Layoutdm: Transformer-Based Diffusion Model For Layout Generation", "abstract": "Automatic layout generation that can synthesize highquality layouts is an important tool for graphic design in many applications. Though existing methods based on generative models such as Generative Adversarial Networks (GANs) and Variational Auto-Encoders (VAEs) have progressed, they still leave much room for improving the quality and diversity of the results. Inspired by the recent success of diffusion models in generating high-quality images, this paper explores their potential for conditional layout generation and proposes Transformer-based Layout Diffusion Model (LayoutDM) by instantiating the conditional denoising diffusion probabilistic model (DDPM) with a purely transformer-based architecture. Instead of using convolutional neural networks, a transformer-based conditional Layout Denoiser is proposed to learn the reverse diffusion process to generate samples from noised layout data. Benefitting from both transformer and DDPM, our Lay-outDM is of desired properties such as high-quality generation, strong sample diversity, faithful distribution coverage, and stationary training in comparison to GANs and VAEs. Quantitative and qualitative experimental results show that our method outperforms state-of-the-art generative models in terms of quality and diversity. * Corresponding author. generate design layouts given a set of design components with user-specified attributes. Though meaningful attempts are made [1, 10, 11, 15, 18, 21, 23-25, 33, 44-46], it is still challenging to generate realistic and complex layouts, because many factors need to be taken into consideration, such as design elements, their attributes, and their relationships to other elements."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["convolutional neural networks"], "citation_intent": "method"}
{"citing_id": "2304.11928v1", "cited_id": "2003.04614", "section_title": "1) Style Transfer", "citation": "The style transfer is similar to that of LDR #REFR , but it replaces the transposed convolutions with bilinear upsampling and convolutions.", "text_before_citation": ["For example, CDGA #OTHEREFR , SIM #OTHEREFR , MCSSF #OTHEREFR , and BDL+ESL #OTHEREFR use this method or the already transferred images.", "In contrast to previous works, the authors of LDR #OTHEREFR train a style translation model that transfers the target domain images in order to make them look like source domain images.", "They employ the general translation model of BDL #OTHEREFR but add a cycle-reconstruction loss to enforce semantic consistency between the image and the image reconstructed from the labels.", "The active pseudo-labeling (APL) method #OTHEREFR first adapts the target domain images to the source domain using a style transfer.", "Afterward, the style-transferred images are used to create pseudo labels that are later used for selfsupervised training in the target domain (cf. III-E)."], "text_after_citation": ["Ramirez et al.", "#OTHEREFR employ a CycleGAN for style transfer from the source to the target domain in their image-level domain adaptation (ILDA) method.", "They enforce the similarity of segmentation masks based on style-transferred images and unaltered synthetic images using a discriminator in the generation process to avoid artifacts and guide the synthesis.", "The DISE-CT method #OTHEREFR is based on DISE #OTHEREFR but adds a cycle consistency to the generator training.", "It also adapts the zero loss #OTHEREFR to a zero-style loss."], "citing_paper_content": {"title": "Survey On Unsupervised Domain Adaptation For Semantic Segmentation For Visual Perception In Automated Driving", "abstract": "The research leading to these results is funded by the German Federal Ministry for Economic Affairs and Energy within the project \"KI Delta Learning\" (F\u00f6rderkennzeichen 19A19013C and 19A19013K). The authors would like to thank the consortium for the successful cooperation."}, "cited_paper_content": {"title": "Label-Driven Reconstruction For Domain Adaptation In Semantic Segmentation", "abstract": "Unsupervised domain adaptation enables to alleviate the need for pixel-wise annotation in the semantic segmentation. One of the most common strategies is to translate images from the source domain to the target domain and then align their marginal distributions in the feature space using adversarial learning. However, source-to-target translation enlarges the bias in translated images, owing to the dominant data size of the source domain. Furthermore, consistency of the joint distribution in source and target domains cannot be guaranteed through global feature alignment. Here, we present an innovative framework, designed to mitigate the image translation bias and align cross-domain features with the same category. This is achieved by 1) performing the target-to-source translation and 2) reconstructing both source and target images from their predicted labels. Extensive experiments on adapting from synthetic to real urban scene understanding demonstrate that our framework competes favorably against existing state-of-the-art methods."}, "keywords": ["style transfer", "transposed convolutions"], "citation_intent": "method"}
{"citing_id": "2304.14845v1", "cited_id": "1512.03385", "section_title": "D. Network", "citation": "To increase the representation ability of our model, we introduce 3 Res-Blocks #REFR . Details of the network are shown in Fig. 10 .", "text_before_citation": ["Alike to SuperPoint #OTHEREFR , we adopt 8 times downsampling to reduce the resolution of high-dimension features, making the model efficient at test time."], "text_after_citation": [], "citing_paper_content": {"title": "Sfd2: Semantic-Guided Feature Detection And Description", "abstract": "Visual localization is a fundamental task for various applications including autonomous driving and robotics. Prior methods focus on extracting large amounts of often redundant locally reliable features, resulting in limited efficiency and accuracy, especially in large-scale environments under challenging conditions. Instead, we propose to extract globally reliable features by implicitly embedding high-level semantics into both the detection and description processes. Specifically, our semantic-aware detector is able to detect keypoints from reliable regions (e.g. building, traffic lane) and suppress unreliable areas (e.g. sky, car) implicitly instead of relying on explicit semantic labels. This boosts the accuracy of keypoint matching by reducing the number of features sensitive to appearance changes and avoiding the need of additional segmentation networks at test time. Moreover, our descriptors are augmented with semantics and have stronger discriminative ability, providing more inliers at test time. Particularly, experiments on long-term large-scale visual localization Aachen Day-Night and RobotCar-Seasons datasets demonstrate that our model outperforms previous local features and gives competitive accuracy to advanced matchers but is about 2 and 3 times faster when using 2k and 4k keypoints, respectively. Code is available at https://github.com/feixue94/sfd2."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["representation ability"], "citation_intent": "method"}
{"citing_id": "2303.12368v1", "cited_id": "2003.08934", "section_title": "E. View Synthesis", "citation": "This application consists of two steps: (1) background rendering with NeRF #REFR , and (2) object and mask rendering with our renderer.", "text_before_citation": ["While image-based rendering(IBR) can perform view interpolation excellently, the view-dependent effect of highly specular objects, such as chrome spheres, is difficult to reproduce using IBR.", "Physically-based rendering(PBR) can handle this view-dependent effect realistically, but PBR requires scene material, geometry, and spatially-varying lighting that is difficult to obtain in the real-world.", "Because MAIR can perform accurate inverse rendering in real-world scenes, and can be easily applied to existing view synthesis methods with multi-view images, we can take advantage of IBR and PBR.", "The view synthesis result of the scene with chrome sphere inserted is in the accompanied video."], "text_after_citation": ["We render the shadow of an object in all images and we train NeRF with these images.", "Background including shadow in novel view is rendered with NeRF, and chrome sphere in novel view is rendered with our lighting and renderer.", "Among the variants of NeRF, we use DirectVoxGO #OTHEREFR for fast training."], "citing_paper_content": {"title": "Mair: Multi-View Attention Inverse Rendering With 3D Spatially-Varying Lighting Estimation", "abstract": "We propose a scene-level inverse rendering framework that uses multi-view images to decompose the scene into geometry, a SVBRDF, and 3D spatially-varying lighting. Because multi-view images provide a variety of information about the scene, multi-view images in object-level inverse rendering have been taken for granted. However, owing to the absence of multi-view HDR synthetic dataset, scene-level inverse rendering has mainly been studied using single-view image. We were able to successfully perform scene-level inverse rendering using multi-view images by expanding OpenRooms dataset and designing efficient pipelines to handle multi-view images, and splitting spatially-varying lighting. Our experiments show that the proposed method not only achieves better performance than single-view-based methods, but also achieves robust performance on unseen real-world scene. Also, our sophisticated 3D spatially-varying lighting volume allows for photorealistic object insertion in any 3D location."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["renderer"], "citation_intent": "method"}
{"citing_id": "2304.03535v1", "cited_id": "1801.01290", "section_title": "Implementation Details", "citation": "We use off-policy Soft Actor Critic #REFR algorithm for optimizing RL objective in our experiments.", "text_before_citation": ["The lower primitive action a which is directly executed on the environment, is a 4 dimensional vector with every dimension a i \u2208 [0, 1].", "The first 3 dimensions provide offsets to be scaled and added to gripper position for moving it to the intended position.", "The last dimension provides gripper control(0 implies a fully closed gripper, 0.5 implies a half closed gripper and 1 implies a fully open gripper).", "We select 100 randomly generated mazes each for training, testing and validation.", "For selecting train, test and validation mazes, we first randomly generate 300 distinct mazes, and then randomly divide them into 100 train, test and validation mazes each. Each experiment is run on 4 parallel workers."], "text_after_citation": ["We keep the regularization weight hyperparameter as \u03a8 = 0.0078. We use Adam #OTHEREFR optimizer.", "The hyperparameter u which is the number of training iterations after which the replay buffer is flushed and re-populated is set as 10. The experiments are run for 2.93e6 timesteps.", "The method for generating expert demonstrations is provided in Appendix A.3."], "citing_paper_content": {"title": "Crisp: Curriculum Inducing Primitive Informed Subgoal Prediction For Hierarchical Reinforcement Learning", "abstract": "Hierarchical reinforcement learning is a promising approach that uses temporal abstraction to solve complex long horizon problems. However, simultaneously learning a hierarchy of policies is unstable as it is challenging to train higherlevel policy when the lower-level primitive is nonstationary. In this paper, we propose a novel hierarchical algorithm by generating a curriculum of achievable subgoals for evolving lower-level primitives using reinforcement learning and imitation learning. The lower level primitive periodically performs data relabeling on a handful of expert demonstrations using our primitive informed parsing approach. We provide expressions to bound the sub-optimality of our method and develop a practical algorithm for hierarchical reinforcement learning. Since our approach uses a handful of expert demonstrations, it is suitable for most robotic control tasks. Experimental evaluation on complex maze navigation and robotic manipulation environments show that inducing hierarchical curriculum learning significantly improves sample efficiency, and results in efficient goal conditioned policies for solving temporally extended tasks."}, "cited_paper_content": {"title": "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning With A Stochastic Actor", "abstract": "Model-free deep reinforcement learning (RL) algorithms have been demonstrated on a range of challenging decision making and control tasks. However, these methods typically suffer from two major challenges: very high sample complexity and brittle convergence properties, which necessitate meticulous hyperparameter tuning. Both of these challenges severely limit the applicability of such methods to complex, real-world domains. In this paper, we propose soft actor-critic, an off-policy actor-critic deep RL algorithm based on the maximum entropy reinforcement learning framework. In this framework, the actor aims to maximize expected reward while also maximizing entropy - that is, succeed at the task while acting as randomly as possible. Prior deep RL methods based on this framework have been formulated as Q-learning methods. By combining off-policy updates with a stable stochastic actor-critic formulation, our method achieves state-of-the-art performance on a range of continuous control benchmark tasks, outperforming prior on-policy and off-policy methods. Furthermore, we demonstrate that, in contrast to other off-policy algorithms, our approach is very stable, achieving very similar performance across different random seeds."}, "keywords": ["RL objective", "policy"], "citation_intent": "method"}
{"citing_id": "2304.00648v1", "cited_id": "1608.06993", "section_title": "B. Multipath Equalization Using Jcaecnn", "citation": "Motivated by the results presented in #REFR , a Densely Connected Convolutional Network (DenseNet) is used to implement the \"shared\" encoder shown in Fig. 4 .", "text_before_citation": ["This equalization approach uses a joint CAE and CNN (a.k.a., JCAECNN) architecture similar to that used in #OTHEREFR ; however, the approach described here differs from that in #OTHEREFR in that the CAE architecture is modified to consist of multiple decoder heads instead of one.", "The decoder heads decompose a received preamble-that undergoes Rayleigh fading-into its L weighted and delayed copies of the transmitted signal, x[n]. As illustrated in Fig.", "4 , the JCAECNN equalization process is built on a Single Input Multiple Output (SIMO) system that includes: signal collection and detection (Sect. III-A), multipath fading and AWGN scaling (Sect.", "III-C), and signal preparation-in accordance with our prior work #OTHEREFR -prior to equalization and classification.", "Equalization is performed using a CAE consisting of a single encoder and L decoder heads (i.e., one for each path of the Rayleigh fading channel) using the preambles' IQ+NL representations at the selected SNR."], "text_after_citation": ["DenseNet is comprised of multiple Dense blocks where each block is created by connecting each convolutional layer to all preceding convolutional layers of the same size.", "These dense connections grant each convolutional layer access to all previously generated feature maps to enable feature reuse.", "In this work, two densely connected blocks are used to construct the encoder.", "The DenseNet encoder generates compressed representation h(R[m, n]), which is the input shared with each of the following NN architectures: the L decoder heads and the classifier The resulting JCAECNN architecture is derived from the square error function of equation #OTHEREFR .", "The optimization goal is to minimize the error between the received signal r(m) and its delayed and scaled copies C k = \u03b1 k x(m \u2212 \u03c4 k ) generated using the TDL of equation #OTHEREFR ."], "citing_paper_content": {"title": "Improving Rf-Dna Fingerprinting Performance In An Indoor Multipath Environment Using Semi-Supervised Learning", "abstract": "The number of Internet of Things (IoT) deployments is expected to reach 75.4 billion by 2025. Roughly 70% of all IoT devices employ weak or no encryption; thus, putting them and their connected infrastructure at risk of attack by devices that are wrongly authenticated or not authenticated at all. A physical layer-based security approach-known as Specific Emitter Identification (SEI)-has been proposed and is being pursued as a viable IoT security mechanism. SEI is advantageous because it is a passive technique that exploits inherent and distinct features that are unintentionally imparted upon the signal during its formation and transmission within and by the IoT device's Radio Frequency (RF) front-end. SEI's passive exploitation of unintentional signal features removes any need to modify the IoT device, which makes it ideal for existing and future IoT deployments. Despite the amount of SEI research conducted there remains challenges that must be addressed to make SEI a viable IoT security approach. One of these challenges is the extraction of SEI features from signals collected under multipath fading conditions. Multipath corrupts the inherent SEI exploited features that are used to discriminate one IoT device from another; thus, degrading authentication performance and increasing the chance of attack. This work presents two semi-supervised Deep Learning (DL) equalization approaches and compares their performance with the current state of the art. The two approaches are the Conditional Generative Adversarial Network (CGAN) and Joint Convolutional Auto-Encoder and Convolutional Neural Network (JCAECNN). Both approaches learn the channel distribution to enable multipath correction while simultaneously preserving the SEI exploited signal features. CGAN and JCAECNN performance is assessed using a Rayleigh fading channel under degrading SNR, up to thirty-two IoT devices, as well as two publicly available signal sets. The JCAECNN improves SEI performance by 10% beyond that of the current state of the art."}, "cited_paper_content": {"title": "Densely Connected Convolutional Networks", "abstract": "Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections - one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less computation to achieve high performance. Code and pre-trained models are available at this https URL ."}, "keywords": ["DenseNet", "Densely Connected Convolutional"], "citation_intent": "method"}
{"citing_id": "2303.09859v2", "cited_id": "2002.12327", "section_title": "C Layer Interpretation", "citation": "On the other hand, NER and CR are represented more strongly in the higher layers, which confirms the related findings in the literature #REFR .", "text_before_citation": ["On the other hand, our NormFormer-like architecture (Section 4) defines each layer k as:", "a k (x) = x + LN(a k (x)) k =\u00e2 k ( k\u22121 ) + ff k (\u00e2 k ( k\u22121 ))", "Then it is trivial to calculate the contribution of each layer as s k = k \u2212 k\u22121 .", "We use this s k entities to compute the learnt convex combination of all layers\u015d = 12 k=1 \u03b3 k s k .", "Interpreting \u03b3 k as the amount of 'knowledge' of a particular task in layer k, we see that POS information is contained primarily in the lowest layers, followed by SRL and DP."], "text_after_citation": [], "citing_paper_content": {"title": "Trained On 100 Million Words And Still In Shape: Bert Meets British National Corpus", "abstract": "While modern masked language models (LMs) are trained on ever larger corpora, we here explore the effects of down-scaling training to a modestly-sized but representative, wellbalanced, and publicly available English text source-the British National Corpus. We show that pre-training on this carefully curated corpus can reach better performance than the original BERT model. We argue that this type of corpora has great potential as a language modeling benchmark. To showcase this potential, we present fair, reproducible and data-efficient comparative studies of LMs, in which we evaluate several training objectives and model architectures and replicate previous empirical results in a systematic way. We propose an optimized LM architecture called LTG-BERT. bert#pre-training-data)."}, "cited_paper_content": {"title": "A Primer In Bertology: What We Know About How Bert Works", "abstract": "Transformer-based models are now widely used in NLP, but we still do not understand a lot about their inner workings. This paper describes what is known to date about the famous BERT model (Devlin et al. 2019), synthesizing over 40 analysis studies. We also provide an overview of the proposed modifications to the model and its training regime. We then outline the directions for further research."}, "keywords": ["higher layers", "related findings"], "citation_intent": "result"}
{"citing_id": "2305.01643v1", "cited_id": "1912.04838", "section_title": "Introduction", "citation": "Real data from the Waymo #REFR dataset is used to evaluate NFL against real scans at held-out viewpoints, including real-world intensities, ray drops and secondary returns.", "text_before_citation": ["This includes the possibilities that there are multiple returns 1 from the emitted ray, or no return at all.", "Our formulation closely adheres to the principles of the LiDAR measurement process and incorporates them into the neural field framework.", "Specifically, we (i) devise volume rendering for LiDAR sensors; (ii) incorporate beam divergence and (iii) propose truncated volume rendering to account for secondary returns and improve range prediction.", "We evaluate our method on both synthetic and real Li-DAR data.", "To this end, we (iv) develop a LiDAR simulator for synthesizing scenes from 3D assets that serve as a test bed for viewpoints far from the original scan locations, and to study the effect of different scan patterns."], "text_after_citation": ["Additionally, we (v) propose a novel closed-loop evaluation protocol that leverages real data to evaluate view synthesis in challenging views.", "As an end-to-end test for downstream tasks, we further evaluate the performance of state-of-the-art segmentation and registration networks when trained on real scans and tested on novel views generated by NFL."], "citing_paper_content": {"title": "Neural Lidar Fields For Novel View Synthesis", "abstract": "We present Neural Fields for LiDAR (NFL), a method to optimise a neural field scene representation from LiDAR measurements, with the goal of synthesizing realistic LiDAR scans from novel viewpoints. NFL combines the rendering power of neural fields with a detailed, physically motivated model of the LiDAR sensing process, thus enabling it to accurately reproduce key sensor behaviors like beam divergence, secondary returns, and ray dropping. We evaluate NFL on synthetic and real LiDAR scans and show that it outperforms explicit reconstruct-then-simulate methods as well as other NeRF-style methods on LiDAR novel view synthesis task. Moreover, we show that the improved realism of the synthesized views narrows the domain gap to real scans and translates to better registration and semantic segmentation performance."}, "cited_paper_content": {"title": "Scalability In Perception For Autonomous Driving: Waymo Open Dataset", "abstract": "The research community has increasing interest in autonomous driving research, despite the resource intensity of obtaining representative real world data. Existing self-driving datasets are limited in the scale and variation of the environments they capture, even though generalization within and between operating regions is crucial to the overall viability of the technology. In an effort to help align the research community's contributions with real-world self-driving problems, we introduce a new large scale, high quality, diverse dataset. Our new dataset consists of 1150 scenes that each span 20 seconds, consisting of well synchronized and calibrated high quality LiDAR and camera data captured across a range of urban and suburban geographies. It is 15x more diverse than the largest camera+LiDAR dataset available based on our proposed diversity metric. We exhaustively annotated this data with 2D (camera image) and 3D (LiDAR) bounding boxes, with consistent identifiers across frames. Finally, we provide strong baselines for 2D as well as 3D detection and tracking tasks. We further study the effects of dataset size and generalization across geographies on 3D detection methods. Find data, code and more up-to-date information at http://www.waymo.com/open."}, "keywords": ["real scans", "Waymo dataset"], "citation_intent": "method"}
{"citing_id": "2303.12360v1", "cited_id": "1409.1556", "section_title": "Model", "citation": "We designed a deep learning model based on VGG-Net #REFR to meet our goals.", "text_before_citation": ["Although many methods such as molecular dynamics, have been evaluated #OTHEREFR and used in microstructure analysis convolution neutral network #OTHEREFR (CNN) is a better choice in automatic image recognition since the development of image datasets (such as ImageNet #OTHEREFR ) and computational power.", "Some kinds of model architectures with CNN module have been proposed."], "text_after_citation": ["At the same time, we applied some competing models to our dataset, including other deep learning algorithms such as Dense-Net #OTHEREFR and Res-Net #OTHEREFR and a traditional edge detection algorithm like Sobel. Details of these models are introduced as follows.", "2.2.1.", "VGG-Net CNN extracts different morphological features of the image through the convolution, and then pools the image, and finally gets the predicted result. VGG-Net was proposed by Oxford's Visual Geometry Group.", "VGG-Net uses two or three 2\u00d72 or 3\u00d73 convolution kernels to replace the larger convolution nuclei in the previous convolutional neural network.", "The stacked small convolution kernel can improve the depth of the network while maintaining the same receptive field, and more complex models can be learned through multi-layer nonlinear layers."], "citing_paper_content": {"title": "Automatically Predict Material Properties With Microscopic Image: Example Polymer Compatibility", "abstract": "Many material properties are manifested in the morphological appearance and characterized with microscopic image, such as scanning electron microscopy (SEM). Polymer compatibility is a key physical quantity of polymer material and commonly and intuitively judged by SEM images. However, human observation and judgement for the images is time-consuming, labor-intensive and hard to be quantified. Computer image recognition with machine learning method can make up the defects of artificial judging, giving accurate and quantitative judgement. We achieve automatic compatibility recognition utilizing convolution neural network and transfer learning method, and the model obtains up to 94% accuracy. We also put forward a quantitative criterion for polymer compatibility with this model. The proposed method can be widely applied to the quantitative characterization of the microstructure and properties of various materials."}, "cited_paper_content": {"title": "Very Deep Convolutional Networks For Large-Scale Image Recognition", "abstract": "In this work we investigate the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. Our main contribution is a thorough evaluation of networks of increasing depth using an architecture with very small (3x3) convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers. These findings were the basis of our ImageNet Challenge 2014 submission, where our team secured the first and the second places in the localisation and classification tracks respectively. We also show that our representations generalise well to other datasets, where they achieve state-of-the-art results. We have made our two best-performing ConvNet models publicly available to facilitate further research on the use of deep visual representations in computer vision."}, "keywords": ["deep learning model"], "citation_intent": "method"}
{"citing_id": "2304.07140v1", "cited_id": "1704.03847", "section_title": "Potential Of Existing Point Cloud Benchmarks For Fa\u00e7ade Segmentation", "citation": "Still, even simulated total points numbers were lower than the ones of TLS datasets: the semantic3D.net TLS dataset consisted of 4 BN points #REFR .", "text_before_citation": ["Interestingly, our set consisted not only of real-world point clouds but also included synthetic-world point clouds.", "This was the case with both SynthCity #OTHEREFR and Paris-CARLA-3D #OTHEREFR .", "The former presented a completely simulated MLS point cloud based on a vector model and covering a combination of European mainland cities and New York, USA #OTHEREFR .", "The latter combined acquired point clouds with simulated ones using the CARLA environment #OTHEREFR .", "One distinct advantage of synthetic point clouds was that they can easily outnumber the real ones: 700 M to 60 M in the case of Paris-CARLA-3D #OTHEREFR ."], "text_after_citation": ["On the other hand, the KITTI-360 MLS dataset featured 1 BN points, with 73.7 km of roads being measured in Karlsruhe, Germany #OTHEREFR .", "It was thus clear that quantity of points was directly linked to the pace of acquisition (e.g., MLS is intuitively faster than TLS) and the total covered area.", "The latter is particularly difficult to acquire and compare, since the various datasets have different ways of quantifing this measure, namely: as a number of scenes, the length of road driven, approximate area extent, or else it is unpublished.", "Moreover, three obtained datasets #OTHEREFR did not reveal their total number of points, thereby this statistic is omitted in these cases.", "Regarding point clouds georeferencing, ou analysis showed that the set was equally divided between those published in a local CRS and in those in a global CRS, both had a score of eight datasets."], "citing_paper_content": {"title": "Tum-Fa\u00e7 Ade: Reviewing And Enriching Point Cloud Benchmarks For Fa\u00e7 Ade Segmentation", "abstract": "Point clouds are widely regarded as one of the best dataset types for urban mapping purposes. Hence, point cloud datasets are commonly investigated as benchmark types for various urban interpretation methods. Yet, few researchers have addressed the use of point cloud benchmarks for fa\u00e7ade segmentation. Robust fa\u00e7ade segmentation is becoming a key factor in various applications ranging from simulating autonomous driving functions to preserving cultural heritage. In this work, we present a method of enriching existing point cloud datasets with fa\u00e7ade-related classes that have been designed to facilitate fa\u00e7ade segmentation testing. We propose how to efficiently extend existing datasets and comprehensively assess their potential for fa\u00e7ade segmentation. We use the method to create the TUM-FA\u00c7 ADE dataset, which extends the capabilities of TUM-MLS-2016. Not only can TUM-FA\u00c7 ADE facilitate the development of point-cloud-based fa\u00e7ade segmentation tasks, but our procedure can also be applied to enrich further datasets."}, "cited_paper_content": {"title": "Semantic3D.Net: A New Large-Scale Point Cloud Classification Benchmark", "abstract": "This paper presents a new 3D point cloud classification benchmark data set with over four billion manually labelled points, meant as input for data-hungry (deep) learning methods. We also discuss first submissions to the benchmark that use deep convolutional neural networks (CNNs) as a work horse, which already show remarkable performance improvements over state-of-the-art. CNNs have become the de-facto standard for many tasks in computer vision and machine learning like semantic segmentation or object detection in images, but have no yet led to a true breakthrough for 3D point cloud labelling tasks due to lack of training data. With the massive data set presented in this paper, we aim at closing this data gap to help unleash the full potential of deep learning methods for 3D labelling tasks. Our semantic3D.net data set consists of dense point clouds acquired with static terrestrial laser scanners. It contains 8 semantic classes and covers a wide range of urban outdoor scenes: churches, streets, railroad tracks, squares, villages, soccer fields and castles. We describe our labelling interface and show that our data set provides more dense and complete point clouds with much higher overall number of labelled points compared to those already available to the research community. We further provide baseline method descriptions and comparison between methods submitted to our online system. We hope semantic3D.net will pave the way for deep learning methods in 3D point cloud labelling to learn richer, more general 3D representations, and first submissions after only a few months indicate that this might indeed be the case."}, "keywords": ["semantic3D.net TLS dataset"], "citation_intent": "result"}
{"citing_id": "2303.10428v2", "cited_id": "1908.05739", "section_title": "Related Works", "citation": "Specifically, the ART #REFR generates the most likely hypothesis (text) to explain what has happened between the two observations (texts).", "text_before_citation": ["The main idea behind the VL foundation models is to learn transferable visual representation with noisy text supervision through a two-tower structure.", "We follow the current VAR pipeline and adopt the foundation model CLIP to generate initial visual & textual features.", "Abductive Reasoning Tasks.", "Humans reckon plausible inferences or hypotheses from incomplete observations every day #OTHEREFR .", "To teach the machine with this ability, researchers proposed several new Tasks, like ART #OTHEREFR for NLP, Sherlock #OTHEREFR for vision, and VideoVAR #OTHEREFR , VideoABC #OTHEREFR for video."], "text_after_citation": ["For Sherlock, Video-VAR, and VideoABC, the observations are represented by regional or whole images, while inference could be either text or middle frames.", "There are several similar tasks, such as Visual Commonsense Reasoning (VCR) #OTHEREFR and Vi-sual7W #OTHEREFR .", "Abductive reasoning differs from them in having non-definitely correct, plausible inferences as human beings.", "Prompt Tuning comes from adapting the LLM to downstream NLP tasks by adding a few \"virtual tokens\" as the prefix or suffix to original sentences.", "According to whether using trainable continuous or handcrafted discrete \"virtual tokens\", we could classify prompts into soft #OTHEREFR or hard #OTHEREFR 26] prompts."], "citing_paper_content": {"title": "Fine-Grained Regional Prompt Tuning For Visual Abductive Reasoning", "abstract": "Visual Abductive Reasoning (VAR) is an emerging vision-language (VL) topic where the model needs to retrieve/generate a likely textual hypothesis from a visual input (image or part of an image) using backward reasoning based on prior knowledge or commonsense. Unlike in conventional VL retrieval or captioning tasks, where entities of texts appear in the image, in abductive inferences, the relevant facts about inferences are not directly visible in the input images. Besides, the inferences are causally relevant to regional visual hints and vary with the latter. Existing works highlight visual parts from a global background with specific prompt tuning techniques (e.g., colorful prompt tuning) on top of foundation models, like CLIP. However, these methods uniformly patchify \"regional hints\" and \"global context\" at the same granularity level and may lose finegrained visual details significant for abductive reasoning. To tackle this, we propose a simple yet effective Regional Prompt Tuning, which encodes \"regional visual hints\" and \"global contexts\" separately at fine and coarse-grained levels. Specifically, our model explicitly upsamples, then patchify local hints to get fine-grained regional prompts. These prompts are concatenated with coarse-grained contextual tokens from whole images. We also equip our model with a new Dual-Contrastive Loss to regress the visual feature simultaneously toward features of factual description (a.k.a. clue text) and plausible hypothesis (abductive inference text) during training. Extensive experiments on the Sherlock dataset demonstrate that our fully fine-tuned RGP/RGP s with Dual-Contrastive Loss significantly outperforms previous SOTAs, achieving the 1 st rank on abductive reasoning leaderboards among all submissions, under all metrics (e.g., P@1 i\u2192t : RGP s 38.78 vs CPT-CLIP 33.44, higher=better). We would open-source our codes for further research. (a) Performance improvements obtained by RGP, RGPs and Dual-Contrastive Loss using two ViT backbone architectures. X axis is Image \u21c6 T ext Mean Rank (i.e., observation \u21c6 hypothesis; lower=better). (b) Overview of Regional Prompt Tuning and Dual-Contrastive Loss Figure 1: Effects and Methodology of Regional Prompt Tuning on the Visual Abductive Reasoning Task. (a). Performance boosts on the Sherlock testing set; (b) The RGP patchify regional observation in more fine-grained level than its colorful counterpart (CPT), and Dual-Contrasive Loss serves to shrink gaps between \"Vision-Inference/-Clue\" modalities simultaneously."}, "cited_paper_content": {"title": "Abductive Commonsense Reasoning", "abstract": "Abductive reasoning is inference to the most plausible explanation. For example, if Jenny finds her house in a mess when she returns from work, and remembers that she left a window open, she can hypothesize that a thief broke into her house and caused the mess, as the most plausible explanation. While abduction has long been considered to be at the core of how people interpret and read between the lines in natural language (Hobbs et al., 1988), there has been relatively little research in support of abductive natural language inference and generation. We present the first study that investigates the viability of language-based abductive reasoning. We introduce a challenge dataset, ART, that consists of over 20k commonsense narrative contexts and 200k explanations. Based on this dataset, we conceptualize two new tasks \u2013 (i) Abductive NLI: a multiple-choice question answering task for choosing the more likely explanation, and (ii) Abductive NLG: a conditional generation task for explaining given observations in natural language. On Abductive NLI, the best model achieves 68.9% accuracy, well below human performance of 91.4%. On Abductive NLG, the current best language generators struggle even more, as they lack reasoning capabilities that are trivial for humans. Our analysis leads to new insights into the types of reasoning that deep pre-trained language models fail to perform\u2014despite their strong performance on the related but more narrowly defined task of entailment NLI\u2014pointing to interesting avenues for future research."}, "keywords": ["text", "likely hypothesis"], "citation_intent": "background"}
{"citing_id": "2303.17392v1", "cited_id": "1602.09028", "section_title": "I. Introduction", "citation": "The work in #REFR aims to optimize the beamformers associated with the common message and private messages to maximize the sum rate of CUs in multi-user multiple input single output (MISO) systems under imperfect channel state information.", "text_before_citation": ["However, the CUs in the guard zone have bad QoS experience since they are out of the coverage of the communication system.", "To address this issue, the work in #OTHEREFR proposes to deploy intelligent reflecting surface (IRS) in the guard zone to enhance the network coverage and QoS of the users.", "There also exist several works related to RSMA, and the readers are referred to a comprehensive survey on RSMA in #OTHEREFR . However, the existing works do not investigate CRC.", "In particular, the authors in #OTHEREFR and #OTHEREFR evaluate the performance obtained by RSMA.", "These works demonstrate that RSMA outperforms both NOMA and SDMA in terms of spectrum efficiency, energy efficiency, QoS, and computational complexity in different network loads and channel conditions."], "text_after_citation": ["The authors in #OTHEREFR analyze the data rate obtained by RSMA for two-receiver MISO broadcast channel with finite rate feedback.", "Considering a multi-user single input single output (SISO), the authors in #OTHEREFR aim to allocate the common rate, transmit power of the common message, and transmit power of the private messages by accounting for successful SIC power requirements.", "Different from #OTHEREFR , #OTHEREFR , and #OTHEREFR , the works in #OTHEREFR and #OTHEREFR aim to achieve max-min fairness among the users, i.e., maximizing the user capacity with the minimum data rate, in the RSMA-based network.", "Recently, RSMA can also be combined with enabling technologies in 6G.", "In particular, RSMA is shown to efficiently manage the inter-beam interference for satellite communications #OTHEREFR ."], "citing_paper_content": {"title": "Joint Rate Allocation And Power Control For Rsma-Based Communication And Radar Coexistence Systems", "abstract": "We consider a rate-splitting multiple access (RSMA)-based communication and radar coexistence (CRC) system. The proposed system allows an RSMA-based communication system to share spectrum with multiple radars. Furthermore, RSMA enables flexible and powerful interference management by splitting messages into common parts and private parts to partially decode interference and partially treat interference as noise. The RSMA-based CRC system thus significantly improves spectral efficiency, energy efficiency and quality of service (QoS) of communication users (CUs). However, the RSMA-based CRC system raises new challenges. Due to the spectrum sharing, the communication network and the radars cause interference to each other, which reduces the signal-to-interference-plus-noise ratio (SINR) of the radars as well as the data rate of the CUs in the communication network. Therefore, a major problem is to maximize the sum rate of the CUs while guaranteeing their QoS requirements of data transmissions and the SINR requirements of multiple radars. To achieve these objectives, we formulate a problem that optimizes i) the common rate allocation to the CUs, transmit power of common message and transmit power of private messages of the CUs, and ii) transmit power of the radars. The problem is non-convex with multiple decision parameters, which is challenging to be solved. We propose two algorithms. The first sequential quadratic programming (SQP) can quickly return a local optimal solution, and has been known to be the state-of-the-art in nonlinear programming methods. The second is an additive approximation scheme (AAS) which solves the problem globally in a reasonable amount of time, based on the technique of applying exhaustive enumeration to a modified instance. The simulation results show the improvement of the AAS compared with the SQP in terms of sum rate. Furthermore, with the AAS, the sum rate of the CUs only slightly decreases when the radars' SINR is significantly increased. This implies that the AAS supports the RSMA-based communication system which allows to well coexist with the radars."}, "cited_paper_content": {"title": "Sum-Rate Maximization For Linearly Precoded Downlink Multiuser Miso Systems With Partial Csit: A Rate-Splitting Approach", "abstract": "This paper considers the Sum-Rate (SR) maximization problem in downlink MU-MISO systems under imperfect Channel State Information at the Transmitter (CSIT). Contrary to existing works, we consider a rather unorthodox transmission scheme. In particular, the message intended to one of the users is split into two parts: a common part which can be recovered by all users, and a private part recovered by the corresponding user. On the other hand, the rest of users receive their information through private messages. This Rate-Splitting (RS) approach was shown to boost the achievable Degrees of Freedom (DoF) when CSIT errors decay with increased SNR. In this work, the RS strategy is married with linear precoder design and optimization techniques to achieve a maximized Ergodic SR (ESR) performance over the entire range of SNRs. Precoders are designed based on partial CSIT knowledge by solving a stochastic rate optimization problem using means of Sample Average Approximation (SAA) coupled with the Weighted Minimum Mean Square Error (WMMSE) approach. Numerical results show that in addition to the ESR gains, the benefits of RS also include relaxed CSIT quality requirements and enhanced achievable rate regions compared to conventional transmission with NoRS."}, "keywords": ["sum rate", "imperfect channel state"], "citation_intent": "background"}
{"citing_id": "2304.06813v1", "cited_id": "1705.08500", "section_title": "Related Works", "citation": "Equipping a model with the option to reject has also been studied in selective classification #REFR .", "text_before_citation": ["We argue that whether an example with covariate shift should be detected or not depends on whether the deployed classifier would misclassify it or not.", "By taking a modelspecific perspective, our MS-OOD DETECTION resolves the dilemma between OOD detection #OTHEREFR and OOD generalization #OTHEREFR : a robust model should generalizes to examples with covariate shift; a weak model should reject them. OOD generalization.", "Covariate shift is commonly studied in model generalization and robustness #OTHEREFR .", "Instead of rejecting examples with covariate shift, the community aims to improve the robustness of a neural network model so that the model could classify them correctly.", "However, given the notable accuracy gap between classifying ID examples and covariate-shift examples #OTHEREFR , we argue that it is desirable to also consider covariate shift in OOD detection. Selective classification."], "text_after_citation": ["Different from OOD detection, selective classification focuses on rejecting uncertain \"ID\" examples.", "Recently, #OTHEREFR proposed to integrate selective classification with OOD detection (SCOD), aiming to detect both misclassified ID and semantic shift data (see Figure 2 (d)).", "In this context, our work can be seen as a generalized version, further taking covariate shifts into account (see Figure 2 (e)).", "Compared to #OTHEREFR , we provide a more comprehensive study, further emphasizing the role of models in evaluation.", "OOD detection methods have roughly two categories: posthoc and training-based #OTHEREFR ."], "citing_paper_content": {"title": "Unified Out-Of-Distribution Detection: A Model-Specific Perspective", "abstract": "Out-of-distribution (OOD) detection aims to identify test examples that do not belong to the training distribution and are thus unlikely to be predicted reliably. Despite a plethora of existing works, most of them focused only on the scenario where OOD examples come from semantic shift (e.g., unseen categories), ignoring other possible causes (e.g., covariate shift). In this paper, we present a novel, unifying framework to study OOD detection in a broader scope. Instead of detecting OOD examples from a particular cause, we propose to detect examples that a deployed machine learning model (e.g., an image classifier) is unable to predict correctly. That is, whether a test example should be detected and rejected or not is \"model-specific\". We show that this framework unifies the detection of OOD examples caused by semantic shift and covariate shift, and closely addresses the concern of applying a machine learning model to uncontrolled environments. We provide an extensive analysis that involves a variety of models (e.g., different architectures and training strategies), sources of OOD examples, and OOD detection approaches, and reveal several insights into improving and understanding OOD detection in uncontrolled environments."}, "cited_paper_content": {"title": "Selective Classification For Deep Neural Networks", "abstract": "Selective classification techniques (also known as reject option) have not yet been considered in the context of deep neural networks (DNNs). These techniques can potentially significantly improve DNNs prediction performance by trading-off coverage. In this paper we propose a method to construct a selective classifier given a trained neural network. Our method allows a user to set a desired risk level. At test time, the classifier rejects instances as needed, to grant the desired risk (with high probability). Empirical results over CIFAR and ImageNet convincingly demonstrate the viability of our method, which opens up possibilities to operate DNNs in mission-critical applications. For example, using our method an unprecedented 2% error in top-5 ImageNet classification can be guaranteed with probability 99.9%, and almost 60% test coverage."}, "keywords": ["selective classification"], "citation_intent": "background"}
{"citing_id": "2304.02160v1", "cited_id": "1911.13254", "section_title": "Separation Results", "citation": "However, we found that Demucs V2 performs better than HuBERT-SEP N , possibly due to the more complex encoder and decoder design of Demucs V2, which was also reported in #REFR .", "text_before_citation": ["All experiments on the original Res-U-Net and Demucs V2 were reproduced by us to ensure a fair comparison.", "Time-domain models were trained at 16 kHz to utilize the pretrained HuBERT model, so the baseline Demucs V2 results differ from #OTHEREFR .", "Additionally, we provide the reported performance of several state-of-the-art models, although some of them (denoted in grey) used additional training data or were developed for the MDX challenge #OTHEREFR , which directly used the MusDB18 test set as a validation set for model selection.", "From the comparison between the original Demucs V2, Res-U-Net, HuBERT-SEP N , and Pa -HuBERT N models (i.e., without pretraining), we observe that Pa -HuBERT N generally outperforms Res-U-Net.", "The potential reason is that the transformer bottleneck processes the encoder features more efficiently than the CNN bottleneck model in Res-U-Net."], "text_after_citation": ["When fine-tuning the pretrained HuBERT and Pa -HuBERT models on the separation task, we observe an improvement in SDR for all four sources compared to models without pretraining.", "This demonstrates the effectiveness of HuBERT in contributing to audio separation tasks beyond speech, and also shows that our proposed Pa -HuBERT model, pretrained with primitive auditory features from music data, can improve the performance of TF-domain separation models.", "When comparing between Pa -HuBERT-SEP SSL and Pa -HuBERT-SEP 2SSL , we observe that 2SSL yields better performance on vocals and drums separation, but is degraded on bass and other separation.", "We believe that the design of iterative pretraining can be further improved, including exploring different layers of latent features for pretraining and training varying numbers of K-Means models for different separation sources. This will be a focus of our future work.", "Specifically, HuBERT shows a larger improvement than Pa -HuBERT, which can potentially be attributed to the size of the pretraining dataset (60 000 hours for HuBERT vs. 890 hours for Pa -HuBERT)."], "citing_paper_content": {"title": "Pa -Hubert: Self-Supervised Music Source Separation Via Primitive Auditory Clustering And Hidden-Unit Bert", "abstract": "In spite of the progress in music source separation research, the small amount of publicly-available clean source data remains a constant limiting factor for performance. Thus, recent advances in selfsupervised learning present a largely-unexplored opportunity for improving separation models by leveraging unlabelled music data. In this paper, we propose a self-supervised learning framework for music source separation inspired by the HuBERT speech representation model. We first investigate the potential impact of the original HuBERT model by inserting an adapted version of it into the wellknown Demucs V2 time-domain separation model architecture. We then propose a time-frequency-domain self-supervised model, Pa-HuBERT (pronounced Pac-HuBERT, for primitive auditory clustering HuBERT), that we later use in combination with a Res-U-Net decoder for source separation. Pa-HuBERT uses primitive auditory features of music as unsupervised clustering labels to initialize the self-supervised pretraining process using the Free Music Archive (FMA) dataset. The resulting framework achieves better source-todistortion ratio (SDR) performance on the MusDB18 test set than the original Demucs V2 and Res-U-Net models. We further demonstrate that it can boost performance with small amounts of supervised data. Ultimately, our proposed framework is an effective solution to the challenge of limited clean source data for music source separation."}, "cited_paper_content": {"title": "Music Source Separation In The Waveform Domain", "abstract": "Source separation for music is the task of isolating contributions, or stems, from different instruments recorded individually and arranged together to form a song. Such components include voice, bass, drums and any other accompaniments. Contrarily to many audio synthesis tasks where the best performances are achieved by models that directly generate the waveform, the state-of-the-art in source separation for music is to compute masks on the magnitude spectrum. In this paper, we first show that an adaptation of Conv-Tasnet (Luo & Mesgarani, 2019), a waveform-to-waveform model for source separation for speech, significantly beats the state-of-the-art on the MusDB dataset, the standard benchmark of multi-instrument source separation. Second, we observe that Conv-Tasnet follows a masking approach on the input signal, which has the potential drawback of removing parts of the relevant source without the capacity to reconstruct it. We propose Demucs, a new waveform-to-waveform model, which has an architecture closer to models for audio generation with more capacity on the decoder. Experiments on the MusDB dataset show that Demucs beats previously reported results in terms of signal to distortion ratio (SDR), but lower than Conv-Tasnet. Human evaluations show that Demucs has significantly higher quality (as assessed by mean opinion score) than Conv-Tasnet, but slightly more contamination from other sources, which explains the difference in SDR. Additional experiments with a larger dataset suggest that the gap in SDR between Demucs and Conv-Tasnet shrinks, showing that our approach is promising."}, "keywords": ["complex encoder"], "citation_intent": "result"}
{"citing_id": "2303.12818v1", "cited_id": "1803.08494", "section_title": "Group Normalization", "citation": "GroupNorm was created to allow for smaller batch sizes, as compared to standard BatchNorm #REFR .", "text_before_citation": [], "text_after_citation": ["For high-resolution images, smaller batches of size one or two are preferred, whereas BatchNorm requires larger batch sizes to perform well.", "GroupNorm, does not normalize in batches, but instead normalizes along the feature dimension by considering groups of features.", "GroupNorm has been shown to work well for batches of size two, and it may enhance object segmentation and detection."], "citing_paper_content": {"title": "An Empirical Analysis Of The Shift And Scale Parameters In Batchnorm", "abstract": "Batch Normalization (BatchNorm) is a technique that improves the training of deep neural networks, especially Convolutional Neural Networks (CNN). It has been empirically demonstrated that BatchNorm increases performance, stability, and accuracy, although the reasons for such improvements are unclear. BatchNorm includes a normalization step as well as trainable shift and scale parameters. In this paper, we empirically examine the relative contribution to the success of BatchNorm of the normalization step, as compared to the re-parameterization via shifting and scaling. To conduct our experiments, we implement two new optimizers in PyTorch, namely, a version of BatchNorm that we refer to as AffineLayer, which includes the re-parameterization step without normalization, and a version with just the normalization step, that we call BatchNorm-minus. We compare the performance of our AffineLayer and BatchNorm-minus implementations to standard BatchNorm, and we also compare these to the case where no batch normalization is used. We experiment with four ResNet architectures (ResNet18, ResNet34, ResNet50, and ResNet101) over a standard image dataset and multiple batch sizes. Among other findings, we provide empirical evidence that the success of BatchNorm may derive primarily from improved weight initialization."}, "cited_paper_content": {"title": "Group Normalization", "abstract": "Batch Normalization (BN) is a milestone technique in the development of deep learning, enabling various networks to train. However, normalizing along the batch dimension introduces problems --- BN's error increases rapidly when the batch size becomes smaller, caused by inaccurate batch statistics estimation. This limits BN's usage for training larger models and transferring features to computer vision tasks including detection, segmentation, and video, which require small batches constrained by memory consumption. In this paper, we present Group Normalization (GN) as a simple alternative to BN. GN divides the channels into groups and computes within each group the mean and variance for normalization. GN's computation is independent of batch sizes, and its accuracy is stable in a wide range of batch sizes. On ResNet-50 trained in ImageNet, GN has 10.6% lower error than its BN counterpart when using a batch size of 2; when using typical batch sizes, GN is comparably good with BN and outperforms other normalization variants. Moreover, GN can be naturally transferred from pre-training to fine-tuning. GN can outperform its BN-based counterparts for object detection and segmentation in COCO, and for video classification in Kinetics, showing that GN can effectively replace the powerful BN in a variety of tasks. GN can be easily implemented by a few lines of code in modern libraries."}, "keywords": ["standard BatchNorm"], "citation_intent": "method"}
{"citing_id": "2304.09869v1", "cited_id": "1805.11074", "section_title": "B. Compared Algorithms", "citation": "In addition to 0.00001, 1 and 100 tested in previous work #REFR , we also test \u03bb = 0.0001, 0.001, 0.01, 0.1, 1 and 10 to demonstrate the impact brought by the choice of penalty coefficient \u03bb.", "text_before_citation": ["ECRL is compared to several state-of-the-art algorithms including ERL #OTHEREFR , RCPO #OTHEREFR , and IPO #OTHEREFR .", "In addition, we implement two variants of ERL with RCPO #OTHEREFR and a reward shape method #OTHEREFR using different coefficient values for constrained optimisation, as compared algorithms.", "RCPO implemented with SAC is denoted as \"RCPO\" and IPO implemented with PPO is denoted as \"IPO\" in the experiment results, respectively.", "Adopting RCPO to ERL is straightforward and not detailed here. The resulted algorithm is referred to as \"RCPO-ERL\".", "When adopting the reward shape method to ERL, it is hard to determine the best value for penalty coefficient \u03bb."], "text_after_citation": ["Instances of ERL with reward shape using different values of \u03bb are referred to as \"ERL * \", where * indicates the value of penalty coefficient \u03bb.", "\"ERL\" refers to the original ERL without reward shaping or RCPO."], "citing_paper_content": {"title": "Evolving Constrained Reinforcement Learning Policy", "abstract": "Evolutionary algorithms have been used to evolve a population of actors to generate diverse experiences for training reinforcement learning agents, which helps to tackle the temporal credit assignment problem and improves the exploration efficiency. However, when adapting this approach to address constrained problems, balancing the trade-off between the reward and constraint violation is hard. In this paper, we propose a novel evolutionary constrained reinforcement learning (ECRL) algorithm, which adaptively balances the reward and constraint violation with stochastic ranking, and at the same time, restricts the policy's behaviour by maintaining a set of Lagrange relaxation coefficients with a constraint buffer. Extensive experiments on robotic control benchmarks show that our ECRL achieves outstanding performance compared to state-of-the-art algorithms. Ablation analysis shows the benefits of introducing stochastic ranking and constraint buffer. Index Terms-Evolutionary constrained reinforcement learning, evolutionary reinforcement learning, constrained reinforcement learning, stochastic ranking, robotic control This work is accepted by the 2023 International Joint Conference on Neural Networks (IJCNN)."}, "cited_paper_content": {"title": "Reward Constrained Policy Optimization", "abstract": "Solving tasks in Reinforcement Learning is no easy feat. As the goal of the agent is to maximize the accumulated reward, it often learns to exploit loopholes and misspecifications in the reward signal resulting in unwanted behavior. While constraints may solve this issue, there is no closed form solution for general constraints. In this work we present a novel multi-timescale approach for constrained policy optimization, called `Reward Constrained Policy Optimization' (RCPO), which uses an alternative penalty signal to guide the policy towards a constraint satisfying one. We prove the convergence of our approach and provide empirical evidence of its ability to train constraint satisfying policies."}, "keywords": ["penalty coefficient"], "citation_intent": "method"}
{"citing_id": "2303.00244v1", "cited_id": "1512.04150", "section_title": "Introduction", "citation": "The pioneering method in the family is CAM #REFR , which produces saliency maps by linearly combining the convolutional feature maps Figure 1 . An example of SUNY explanations. SUNY highlights sufficient and necessary input regions w.r.t. the model's prediction towards the target class.", "text_before_citation": ["In computer vision, an important research direction is to give Convolutional Neural Network (CNN) more transparency to extend its deployment on a broader basis.", "This paper addresses the eXplainable Artificial Intelligence (XAI) #OTHEREFR problem corresponding to CNN for natural image classification, i.e., reasoning why a classifier makes particular decisions.", "Specifically, we study a leading family of techniques called Class-Activation-Maps (CAMs), which present heatmaps highlighting image portions associated with a model's class prediction."], "text_after_citation": ["The 2D saliency map is the first-of-itskind visual explanation design to the best of our knowledge.", "with weights generated by a global average pooling layer.", "CAM's structural restriction on the global average pooling layer prohibits its use for general CNNs, and a series of gradient-weighted CAMs #OTHEREFR 31] have been proposed to overcome the restriction.", "However, such CAMs suffer from gradients' saturation and vanishing issues, which cause the resulting explanations to be noisy #OTHEREFR .", "To bypass the shortcomings of gradients, Score-CAM #OTHEREFR and Group-CAM #OTHEREFR weight the feature maps by contribution scores, referring to the corresponding input features' importance to the model output."], "citing_paper_content": {"title": "Suny: A Visual Interpretation Framework For Convolutional Neural Networks From A Necessary And Sufficient Perspective", "abstract": "Researchers have proposed various methods for visually interpreting the Convolutional Neural Network (CNN) via saliency maps, which include Class-Activation-Map (CAM) based approaches as a leading family. However, in terms of the internal design logic, existing CAM-based approaches often overlook the causal perspective that answers the core \"why\" question to help humans understand the explanation. Additionally, current CNN explanations lack the consideration of both necessity and sufficiency, two complementary sides of a desirable explanation. This paper presents a causality-driven framework, SUNY, designed to rationalize the explanations toward better human understanding. Using the CNN model's input features or internal filters as hypothetical causes, SUNY generates explanations by bi-directional quantifications on both the necessary and sufficient perspectives. Extensive evaluations justify that SUNY not only produces more informative and convincing explanations from the angles of necessity and sufficiency, but also achieves performances competitive to other approaches across different CNN architectures over largescale datasets, including ILSVRC2012 and CUB-200-2011."}, "cited_paper_content": {"title": "Learning Deep Features For Discriminative Localization", "abstract": "In this work, we revisit the global average pooling layer proposed in [13], and shed light on how it explicitly enables the convolutional neural network (CNN) to have remarkable localization ability despite being trained on imagelevel labels. While this technique was previously proposed as a means for regularizing training, we find that it actually builds a generic localizable deep representation that exposes the implicit attention of CNNs on an image. Despite the apparent simplicity of global average pooling, we are able to achieve 37.1% top-5 error for object localization on ILSVRC 2014 without training on any bounding box annotation. We demonstrate in a variety of experiments that our network is able to localize the discriminative image regions despite just being trained for solving classification task1."}, "keywords": ["convolutional feature maps"], "citation_intent": "method"}
{"citing_id": "2303.06870v1", "cited_id": "1903.05134", "section_title": "Introduction", "citation": "Although using inplace distillation #REFR for sub-networks prevents the model from collapsing, there is still a big gap between the results of S-Net+Distill and training each model individually for SimSiam.", "text_before_citation": ["To address this problem in the context of supervised learning, the family of slimmable networks (S-Net) and universally slimmable networks (US-Net) #OTHEREFR were proposed, which can switch freely among different widths by training only once.", "Driven by the success of slimmable networks, a question arises: Can we train a self-supervised model that can run at arbitrary width? A na\u00efve solution is to replace the supervised loss with self-supervised loss based on the US-Net framework.", "However, we find that this solution doesn't work directly after empirical studies.", "Table 1 shows that the phenomenon in self-supervised scenarios is very different.", "The model directly collapses after applying the popular SSL method SimSiam #OTHEREFR to slimmable networks #OTHEREFR ."], "text_after_citation": ["So why is the situation so different in SSL and how to further improve the performance (i.e., close the gap)?", "In this paper, we present a unified perspective to explain the differences and propose corresponding measures to bridge the gap.", "From a unified gradient perspective, we find that the key is that the guidance to sub-networks should be consistent between iterations, and we analyze which components of SSL incur the temporal inconsistency problem and why US-Net works in supervised learning.", "Based on these theoretical analyses, we propose three guidelines for the loss design of US-Net training to ensure temporal consistency.", "As long as one of them is satisfied, US-Net can work well, no matter in supervised or self-supervised scenarios."], "citing_paper_content": {"title": "Three Guidelines You Should Know For Universally Slimmable Self-Supervised Learning", "abstract": "We propose universally slimmable self-supervised learning (dubbed as US3L) to achieve better accuracy-efficiency trade-offs for deploying self-supervised models across different devices. We observe that direct adaptation of selfsupervised learning (SSL) to universally slimmable networks misbehaves as the training process frequently collapses. We then discover that temporal consistent guidance is the key to the success of SSL for universally slimmable networks, and we propose three guidelines for the loss design to ensure this temporal consistency from a unified gradient perspective. Moreover, we propose dynamic sampling and group regularization strategies to simultaneously improve training efficiency and accuracy. Our US3L method has been empirically validated on both convolutional neural networks and vision transformers. With only once training and one copy of weights, our method outperforms various state-ofthe-art methods (individually trained or not) on benchmarks including recognition, object detection and instance segmentation. Our code is available at https://github.com/ megvii-research/US3L-CVPR2023."}, "cited_paper_content": {"title": "Universally Slimmable Networks And Improved Training Techniques", "abstract": "Slimmable networks are a family of neural networks that can instantly adjust the runtime width. The width can be chosen from a predefined widths set to adaptively optimize accuracy-efficiency trade-offs at runtime. In this work, we propose a systematic approach to train universally slimmable networks (US-Nets), extending slimmable networks to execute at arbitrary width, and generalizing to networks both with and without batch normalization layers. We further propose two improved training techniques for US-Nets, named the sandwich rule and inplace distillation, to enhance training process and boost testing accuracy. We show improved performance of universally slimmable MobileNet v1 and MobileNet v2 on ImageNet classification task, compared with individually trained ones and 4-switch slimmable network baselines. We also evaluate the proposed US-Nets and improved training techniques on tasks of image super-resolution and deep reinforcement learning. Extensive ablation experiments on these representative tasks demonstrate the effectiveness of our proposed methods. Our discovery opens up the possibility to directly evaluate FLOPs-Accuracy spectrum of network architectures. Code and models are available at: this https URL"}, "keywords": ["Net+Distill"], "citation_intent": "result"}
{"citing_id": "2303.06972v1", "cited_id": "1710.04340", "section_title": "Experiments", "citation": "The Lorenz system is a well-known 3-dimensional chaotic system with a strange attractor, driven by the equations: #REFR with parameters \u03c3 = 10, r = 28, b = 8/3.", "text_before_citation": ["We perform experiments on three dynamical systems: the simple pendulum, a 3-dimensional dynamical system modeling the lowdimensional attractor of a nonlinear high-dimensional fluid flow, and the chaotic Lorenz-63 system.", "The first two benchmarks are taken from #OTHEREFR and we refer the reader to this paper for further detail."], "text_after_citation": ["dx dt = \u03c3(y \u2212 x); dy dt = x(r \u2212 z) \u2212 y; dz dt = xy \u2212 bz", "1 Should we constrain K to be perfectly orthogonal, a real logarithm would always exist since O(d) is a matrix Lie group, whose Lie algebra is the set of skew-symmetric matrices #OTHEREFR As for the Lorenz-63 system, despite the simplicity of our architecture, it is still able to learn to behave in a chaotic way, or at least to simulate a chaotic behavior, even when the initialization is outside of the attractor.", "We obtain Lyapunov exponents 0.61, 0.07, \u221214.34 for simulated trajectories of our model, which is close to the estimated real exponents 0.9056, 0, \u221214.5721 of the Lorenz system #OTHEREFR .", "We also manage to model the attractor when initializing our predictions outside of it.", "Note that, while the method from #OTHEREFR is also able to model the Lorenz-63 dynamics faithfully, classical models like recurrent neural networks are known to fail at this task. We plot our simulated trajectories in figure 1."], "citing_paper_content": {"title": "Leveraging Neural Koopman Operators To Learn Continuous Representations Of Dynamical Systems From Scarce Data", "abstract": "Over the last few years, several works have proposed deep learning architectures to learn dynamical systems from observation data with no or little knowledge of the underlying physics. A line of work relies on learning representations where the dynamics of the underlying phenomenon can be described by a linear operator, based on the Koopman operator theory. However, despite being able to provide reliable long-term predictions for some dynamical systems in ideal situations, the methods proposed so far have limitations, such as requiring to discretize intrinsically continuous dynamical systems, leading to data loss, especially when handling incomplete or sparsely sampled data. Here, we propose a new deep Koopman framework that represents dynamics in an intrinsically continuous way, leading to better performance on limited training data, as exemplified on several datasets arising from dynamical systems."}, "cited_paper_content": {"title": "Learning Koopman Invariant Subspaces For Dynamic Mode Decomposition", "abstract": "Spectral decomposition of the Koopman operator is attracting attention as a tool for the analysis of nonlinear dynamical systems. Dynamic mode decomposition is a popular numerical algorithm for Koopman spectral analysis; however, we often need to prepare nonlinear observables manually according to the underlying dynamics, which is not always possible since we may not have any a priori knowledge about them. In this paper, we propose a fully data-driven method for Koopman spectral analysis based on the principle of learning Koopman invariant subspaces from observed data. To this end, we propose minimization of the residual sum of squares of linear least-squares regression to estimate a set of functions that transforms data into a form in which the linear regression fits well. We introduce an implementation with neural networks and evaluate performance empirically using nonlinear dynamical systems and applications."}, "keywords": ["well-known 3-dimensional chaotic"], "citation_intent": "background"}
{"citing_id": "2303.09171v1", "cited_id": "1512.02479", "section_title": "Introduction", "citation": "LRP and its variants, which are based on deep Taylor decomposition #REFR , backpropagate the relevance score from output layer to input layer through specific rules to directly obtain the contribution of each pixel in the input image.", "text_before_citation": ["Therefore, as an extension of CAM, Grad-CAM #OTHEREFR , Score-CAM #OTHEREFR and other CAM-based methods have been proposed.", "However, before Layer-CAM #OTHEREFR , all CAM-based methods could only be applied to the final convolutional layer of the model, and if the target layer became shallow, the faithfulness (measure how accurately the explantion reveals the model decision) of the explanations would be significantly reduced, thus only coarse-grained visual explanations could be generated, which provides coarse location of the target object, limiting the performance on tasks that need pixel-accurate object locations #OTHEREFR .", "Layer-CAM, on the other hand, although outperforming other CAM-based methods on shallow convolutional layers, suffers from shattered gradient problem #OTHEREFR as the network gets deep and the target layer becomes shallow, which can lead to problems such as making visual explanation noisy and reduced faithfulness (not as dramatically as other CAM-based methods, but still not negligible).", "As shown in Figure 1 , as the target layer becomes shallower, the accuracy of the explanations generated by Grad-CAM and Score-CAM will decrease significantly and even become unusable (unable to extract contributing pixels or filter out non-contributing pixels).", "The visual explanation of Layer-CAM has more and more noise, and the accuracy has decreased, while the FG-Grad-CAM and FG-Score-CAM still show good performance."], "text_after_citation": ["They can generate the explanations at the same resolution as the input image. However, the explanations have low faithfulness.", "In order to generate fine-grained visual explanation with high faithfulness, this paper proposes a novel method: FG-CAM.", "Different from other CAM-based methods which are directly applied to shallow layers to obtain high-resolution explanations, FG-CAM uses the relationship between two adjacent layers of feature maps with resolution difference to gradually increase the resolution of explanations produced on the last convolutional layer, while finding the contributing pixels and filtering out the pixels that do not contribute at each step.", "FG-CAM not only solves the problem that CAM-based methods can only provide rough explanations without changing their characteristics, but also generates fine-grained explanations that have higher faithfulness than LRP and its variants.", "This paper also presents FG-CAM with denoising, which is able to generate less noisy explanations with almost no change in explanation faithfulness."], "citing_paper_content": {"title": "Fine-Grained And High-Faithfulness Explanations For Convolutional Neural Networks", "abstract": "Recently, explaining CNNs has become a research hotspot. CAM (Class Activation Map)-based methods and LRP (Layer-wise Relevance Propagation) method are two common explanation methods. However, due to the small spatial resolution of the last convolutional layer, the CAMbased methods can often only generate coarse-grained visual explanations that provide a coarse location of the target object. LRP and its variants, on the other hand, can generate fine-grained explanations. But the faithfulness of the explanations is too low. In this paper, we propose FG-CAM (fine-grained CAM), which extends the CAM-based methods to generate fine-grained visual explanations with high faithfulness. FG-CAM uses the relationship between two adjacent layers of feature maps with resolution difference to gradually increase the explanation resolution, while finding the contributing pixels and filtering out the pixels that do not contribute at each step. Our method not only solves the shortcoming of CAM-based methods without changing their characteristics, but also generates fine-grained explanations that have higher faithfulness than LRP and its variants. We also present FG-CAM with denoising, which is a variant of FG-CAM and is able to generate less noisy explanations with almost no change in explanation faithfulness. Experimental results show that the performance of FG-CAM is almost unaffected by the explanation resolution. FG-CAM outperforms existing CAM-based methods significantly in the both shallow and intermediate convolutional layers, and outperforms LRP and its variations significantly in the input layer."}, "cited_paper_content": {"title": "Explaining Nonlinear Classification Decisions With Deep Taylor Decomposition", "abstract": "Nonlinear methods such as Deep Neural Networks (DNNs) are the gold standard for various challenging machine learning problems such as image recognition. Although these methods perform impressively well, they have a significant disadvantage, the lack of transparency, limiting the interpretability of the solution and thus the scope of application in practice. Especially DNNs act as black boxes due to their multilayer nonlinear structure. In this paper we introduce a novel methodology for interpreting generic multilayer neural networks by decomposing the network classification decision into contributions of its input elements. Although our focus is on image classification, the method is applicable to a broad set of input data, learning tasks and network architectures. Our method called deep Taylor decomposition efficiently utilizes the structure of the network by backpropagating the explanations from the output to the input layer. We evaluate the proposed method empirically on the MNIST and ILSVRC data sets. HighlightsA novel method to explain nonlinear classification decisions in terms of input variables is introduced.The method is based on Taylor expansions and decomposes the output of a deep neural network in terms of input variables.The resulting deep Taylor decomposition can be applied directly to existing neural networks without retraining.The method is tested on two large-scale neural networks for image classification: BVLC CaffeNet and GoogleNet."}, "keywords": ["deep Taylor decomposition"], "citation_intent": "method"}
{"citing_id": "2304.13949v1", "cited_id": "1901.08971", "section_title": "Settings", "citation": "We use a modified version of Xception #REFR as the backbone network, with model parameters initialized by pre-training on ImageNet.", "text_before_citation": ["Note that there are three versions of FF++ in terms of compression level, i.e., raw, lightly compressed (HQ), and heavily compressed (LQ).", "Since realistic forgeries often have a limited quality, the HQ and LQ versions are used in experiments.", "By default, the HQ version is adopted, in line with other works #OTHEREFR .", "If there is any deviation from this default, it will be explicitly stated.", "Implementation."], "text_after_citation": ["Face extraction and alignment are performed using DLIB #OTHEREFR .", "Following previous works #OTHEREFR , the aligned faces are resized to 256 \u00d7 256 for both the training and testing.", "We use the Adam #OTHEREFR for optimization with the learning rate of 0.0002, and the batch size is fixed as 32. In the overall loss function in Eq.", "(9), we set \u03bb 1 to \u03bb 3 as 0.1, 0.3, 0.05 empirically. The margin \u03b1 in Eq. (5) is set to 3.", "We also apply some widely used data augmentations, i.e., image compression, horizontal flip, and random brightness contrast."], "citing_paper_content": {"title": "Ucf: Uncovering Common Features For Generalizable Deepfake Detection", "abstract": "Deepfake detection remains a challenging task due to the difficulty of generalizing to new types of forgeries. This problem primarily stems from the overfitting of existing detection methods to forgery-irrelevant features and methodspecific patterns. The latter is often ignored by previous works. This paper presents a novel approach to address the two types of overfitting issues by uncovering common forgery features. Specifically, we first propose a disentanglement framework that decomposes image information into three distinct components: forgery-irrelevant, methodspecific forgery, and common forgery features. To ensure the decoupling of method-specific and common forgery features, a multi-task learning strategy is employed, including a multi-class classification that predicts the category of the forgery method and a binary classification that distinguishes the real from the fake. Additionally, a conditional decoder is designed to utilize forgery features as a condition along with forgery-irrelevant features to generate reconstructed images. Furthermore, a contrastive regularization technique is proposed to encourage the disentanglement of the common and specific forgery features. Ultimately, we only utilize the common forgery features for the purpose of generalizable deepfake detection. Extensive evaluations demonstrate that our framework can perform superior generalization than current state-of-the-art methods."}, "cited_paper_content": {"title": "Faceforensics++: Learning To Detect Manipulated Facial Images", "abstract": "The rapid progress in synthetic image generation and manipulation has now come to a point where it raises significant concerns for the implications towards society. At best, this leads to a loss of trust in digital content, but could potentially cause further harm by spreading false information or fake news. This paper examines the realism of state-of-the-art image manipulations, and how difficult it is to detect them, either automatically or by humans. To standardize the evaluation of detection methods, we propose an automated benchmark for facial manipulation detection. In particular, the benchmark is based on DeepFakes, Face2Face, FaceSwap and NeuralTextures as prominent representatives for facial manipulations at random compression level and size. The benchmark is publicly available and contains a hidden test set as well as a database of over 1.8 million manipulated images. This dataset is over an order of magnitude larger than comparable, publicly available, forgery datasets. Based on this data, we performed a thorough analysis of data-driven forgery detectors. We show that the use of additional domainspecific knowledge improves forgery detection to unprecedented accuracy, even in the presence of strong compression, and clearly outperforms human observers."}, "keywords": ["ImageNet"], "citation_intent": "method"}
{"citing_id": "2303.00996v1", "cited_id": "1904.04232", "section_title": "Problem Statement: Unsupervised Few-Shot Learning", "citation": "Note that D meta train and D meta test can come from the same domain (i.e., the standard in-domain setting) or different domains (i.e., cross-domain) as suggested by #REFR .", "text_before_citation": ["At the meta-test phase, we transfer the model to new few-shot tasks {T i } \u223c D meta test where each task T i aims to classify query samples {x q } among N labels using support (i.e., training)", "samples S = {(x s , y s )} N K s=1", ".", "We here assume the task T i consists of K support samples for each label y \u2208 {1, . . .", ", N }, which is referred to as N -way K-shot classification."], "text_after_citation": [], "citing_paper_content": {"title": "Unsupervised Meta-Learning Via Few-Shot Pseudo-Supervised Contrastive Learning", "abstract": "Unsupervised meta-learning aims to learn generalizable knowledge across a distribution of tasks constructed from unlabeled data. Here, the main challenge is how to construct diverse tasks for meta-learning without label information; recent works have proposed to create, e.g., pseudo-labeling via pretrained representations or creating synthetic samples via generative models. However, such a task construction strategy is fundamentally limited due to heavy reliance on the immutable pseudo-labels during meta-learning and the quality of the representations or the generated samples. To overcome the limitations, we propose a simple yet effective unsupervised meta-learning framework, coined Pseudo-supervised Contrast (PsCo), for few-shot classification. We are inspired by the recent self-supervised learning literature; PsCo utilizes a momentum network and a queue of previous batches to improve pseudo-labeling and construct diverse tasks in a progressive manner. Our extensive experiments demonstrate that PsCo outperforms existing unsupervised meta-learning methods under various in-domain and cross-domain few-shot classification benchmarks. We also validate that PsCo is easily scalable to a large-scale benchmark, while recent prior-art meta-schemes are not."}, "cited_paper_content": {"title": "A Closer Look At Few-Shot Classification", "abstract": "Few-shot classification aims to learn a classifier to recognize unseen classes during training with limited labeled examples. While significant progress has been made, the growing complexity of network designs, meta-learning algorithms, and differences in implementation details make a fair comparison difficult. In this paper, we present 1) a consistent comparative analysis of several representative few-shot classification algorithms, with results showing that deeper backbones significantly reduce the performance differences among methods on datasets with limited domain differences, 2) a modified baseline method that surprisingly achieves competitive performance when compared with the state-of-the-art on both the \\miniI and the CUB datasets, and 3) a new experimental setting for evaluating the cross-domain generalization ability for few-shot classification algorithms. Our results reveal that reducing intra-class variation is an important factor when the feature backbone is shallow, but not as critical when using deeper backbones. In a realistic cross-domain evaluation setting, we show that a baseline method with a standard fine-tuning practice compares favorably against other state-of-the-art few-shot learning algorithms."}, "keywords": ["meta train"], "citation_intent": "background"}
{"citing_id": "2303.15147v1", "cited_id": "1811.09497", "section_title": "Semi-Supervised Learning Under Extremely Low Data Regimes", "citation": "As can be seen in Table 5 , the proposed method significantly outperforms MURAUER #REFR under the three out of the four scenarios.", "text_before_citation": ["We compare the proposed method with MURAUER #OTHEREFR , which to the best of our knowledge is the only depth-based 3D hand pose estimation method examined under scenarios of severe labeled data scarcity."], "text_after_citation": ["The only scenario where the proposed method is inferior is when there are only 10 labeled examples.", "This is because such a small amount of labeled training data lacks adequate information about the nature of the task.", "MURAUER #OTHEREFR compensates for this lack of information by pre-training on synthetic data. However, the proposed method does not use any pre-training.", "More importantly, in contrast to MURAUER #OTHEREFR , the proposed method achieves this performance without the application-limiting requirement for multi-view real-data.", "Remarkably, in the case where there are only 100 labeled samples, the proposed method achieves 12.11 mm, which clearly demonstrates the practicality of the proposed method in cases of severely limited access to labeled data."], "citing_paper_content": {"title": "Pushing The Envelope For Depth-Based Semi-Supervised 3D Hand Pose Estimation With Consistency Training", "abstract": "Despite the significant progress that depth-based 3D hand pose estimation methods have made in recent years, they still require a large amount of labeled training data to achieve high accuracy. However, collecting such data is both costly and time-consuming. To tackle this issue, we propose a semi-supervised method to significantly reduce the dependence on labeled training data. The proposed method consists of two identical networks trained jointly: a teacher network and a student network. The teacher network is trained using both the available labeled and unlabeled samples. It leverages the unlabeled samples via a loss formulation that encourages estimation equivariance under a set of affine transformations. The student network is trained using the unlabeled samples with their pseudo-labels provided by the teacher network. For inference at test time, only the student network is used. Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art semi-supervised methods by large margins."}, "cited_paper_content": {"title": "Murauer: Mapping Unlabeled Real Data For Label Austerity", "abstract": "Data labeling for learning 3D hand pose estimation models is a huge effort. Readily available, accurately labeled synthetic data has the potential to reduce the effort. However, to successfully exploit synthetic data, current state-of-the-art methods still require a large amount of labeled real data. In this work, we remove this requirement by learning to map from the features of real data to the features of synthetic data mainly using a large amount of synthetic and unlabeled real data. We exploit unlabeled data using two auxiliary objectives, which enforce that (i) the mapped representation is pose specific and (ii) at the same time, the distributions of real and synthetic data are aligned. While pose specifity is enforced by a self-supervisory signal requiring that the representation is predictive for the appearance from different views, distributions are aligned by an adversarial term. In this way, we can significantly improve the results of the baseline system, which does not use unlabeled data and outperform many recent approaches already with about 1% of the labeled real data. This presents a step towards faster deployment of learning based hand pose estimation, making it accessible for a larger range of applications."}, "keywords": ["Table"], "citation_intent": "result"}
{"citing_id": "2304.01443v1", "cited_id": "1907.06724", "section_title": "Introduction", "citation": "The current implementation enables two users to engage in a virtual video call while displaying their avatar faces. arXiv:2304.01443v1 [cs.HC] 4 Apr 2023 #REFR .", "text_before_citation": ["Furthermore, our solution leads to other applications, such as a virtual Omegle chat and a setup for virtual YouTubers.", "The former provides a safe space for one-on-one video chats with random users while displaying facial movements and expressions on avatar mesh.", "The latter eliminates the need for costly tracking headsets by using software to track and animate characters with a webcam.", "To offer a partially immersive Metaverse experience, we provide a solution to replace the 3D conference call feature from Meta, which requires a costly VR headset #OTHEREFR .", "In brief, we use an AI Face Mesh model to track the user's facial expressions, position, and orientation, to render 3D mesh data into a 2D screen through a rendering pipeline done on WebGL, allowing for a simulated 3D conference experience using only a webcam, microphone, and web browser."], "text_after_citation": ["In summary, we demonstrate that companies like Meta can use this product to offer an entry-level Metaverse experience through a web portal rather than requiring consumers to invest in expensive hardware.", "By providing these accessible applications, our solution is expected to bridge the gap for consumers to experience the immersive Metaverse."], "citing_paper_content": {"title": "Virtual Avatar Stream: A Cost-Down Approach To The Metaverse Experience", "abstract": "The Metaverse through VR headsets is a rapidly growing concept, but the high cost of entry currently limits access for many users. This project aims to provide an accessible entry point to the immersive Metaverse experience by leveraging web technologies. The platform developed allows users to engage with rendered avatars using only a web browser, microphone, and webcam. By employing the WebGL and MediaPipe face tracking AI model from Google, the application generates real-time 3D face meshes for users. It uses a client-to-client streaming cluster to establish a connection, and clients negotiate SRTP protocol through WebRTC for direct data streaming. Additionally, the project addresses backend challenges through an architecture that is serverless, distributive, auto-scaling, highly resilient, and secure. The platform offers a scalable, hardware-free solution for users to experience a near-immersive Metaverse, with the potential for future integration with game server clusters. This project provides an important step toward a more inclusive Metaverse accessible to a wider audience."}, "cited_paper_content": {"title": "Real-Time Facial Surface Geometry From Monocular Video On Mobile Gpus", "abstract": "We present an end-to-end neural network-based model for inferring an approximate 3D mesh representation of a human face from single camera input for AR applications. The relatively dense mesh model of 468 vertices is well-suited for face-based AR effects. The proposed model demonstrates super-realtime inference speed on mobile GPUs (100-1000+ FPS, depending on the device and model variant) and a high prediction quality that is comparable to the variance in manual annotations of the same image."}, "keywords": ["avatar"], "citation_intent": "background"}
{"citing_id": "2303.01224v1", "cited_id": "1505.03132", "section_title": "Definition 2. For A Matrix", "citation": "Additionally, it was shown by Gribanov & Veselov #REFR that it is possible to prove a variant of the Flatness Theorem for simplices such that the function \u03c9 will depend only on the sub-determinants spectrum of A instead of n.", "text_before_citation": ["Due to #OTHEREFR Corollary 9] (see also #OTHEREFR ), the integer linear optimization problem can be solved with O n + \u2206 2 \u2022 log(\u2206) \u2022 log(\u2206 gcd ) operations.", "Due to Gribanov & Malyshev #OTHEREFR and a modification from Gribanov, Malyshev & Zolotykh #OTHEREFR , the integer points counting problem can be solved with O(n 4 \u2022\u2206 3 ) operations.", "Due to Gribanov, Malyshev, Pardalos & Veselov #OTHEREFR (see also #OTHEREFR ), the lattice-width can be computed with poly(n, \u2206, \u2206 ext )-operations, where \u2206 ext = \u2206(A b) and (A b) is the system's extended matrix.", "Additionally, for empty simplices, the complexity dependence on \u2206 ext can be avoided, which gives the bound poly(n, \u2206).", "The analogous result for lattice-simplices defined by convex hulls of their vertices is presented in Gribanov, Malyshev & Veselov #OTHEREFR ."], "text_after_citation": ["More precisely, the following statement is true: for any empty simplex S, the inequality width(S) < \u2206 min (A) holds, where \u2206 min (A) is the minimum between all rankorder nonzero sub-determinants of A taken by an absolute value.", "Additionally, if the inequality is not satisfied, then some integer point inside S can be found by a polynomial-time algorithm.", "The last inequality for empty simplices was improved to width(S) < \u230a\u2206 min /2\u230b in the work #OTHEREFR of Henk, Kuhlmann & Weismantel.", "For a general \u2206-modular empty polytope P, the inequality width", "(S) \u2264 (\u2206 lcm (A) \u2212 1)\u2022 \u2206(A)"], "citing_paper_content": {"title": "Enumeration And Unimodular Equivalence Of Empty Delta-Modular Simplices \u22c6", "abstract": "Consider a class of simplices defined by systems Ax \u2264 b of linear inequalities with \u2206-modular matrices. A matrix is called \u2206modular, if all its rank-order sub-determinants are bounded by \u2206 in an absolute value. In our work we call a simplex \u2206-modular, if it can be defined by a system Ax \u2264 b with a \u2206-modular matrix A. And we call a simplex empty, if it contains no points with integer coordinates. In literature, a simplex is called lattice-simplex, if all its vertices have integer coordinates. And a lattice-simplex called empty, if it contains no points with integer coordinates excluding its vertices. Recently, assuming that \u2206 is fixed, it was shown in [37] that the number of \u2206-modular empty simplices modulo the unimodular equivalence relation is bounded by a polynomial on dimension. We show that the analogous fact holds for the class of \u2206-modular empty lattice-simplices. As the main result, assuming again that the value of the parameter \u2206 is fixed, we show that all unimodular equivalence classes of simplices of the both types can be enumerated by a polynomial-time algorithm. As the secondary result, we show the existence of a polynomial-time algorithm for the problem to check the unimodular equivalence relation for a given pair of \u2206-modular, not necessarily empty, simplices."}, "cited_paper_content": {"title": "On Integer Programing With Bounded Determinants", "abstract": "Let A be an \\((m \\times n)\\) integral matrix, and let \\(P=\\{ x :A x \\le b\\}\\) be an n-dimensional polytope. The width of P is defined as \\( w(P)=min\\{ x\\in \\mathbb {Z}^n{\\setminus }\\{0\\} :max_{x \\in P} x^\\top u - min_{x \\in P} x^\\top v \\}\\). Let \\(\\varDelta (A)\\) and \\(\\delta (A)\\) denote the greatest and the smallest absolute values of a determinant among all \\(r(A) \\times r(A)\\) sub-matrices of A, where r(A) is the rank of the matrix A. We prove that if every \\(r(A) \\times r(A)\\) sub-matrix of A has a determinant equal to \\(\\pm \\varDelta (A)\\) or 0 and \\(w(P)\\ge (\\varDelta (A)-1)(n+1)\\), then P contains n affine independent integer points. Additionally, we present similar results for the case of k-modular matrices. The matrix A is called totally k-modular if every square sub-matrix of A has a determinant in the set \\(\\{0,\\, \\pm k^r :r \\in \\mathbb {N} \\}\\). When P is a simplex and \\(w(P)\\ge \\delta (A)-1\\), we describe a polynomial time algorithm for finding an integer point in P."}, "keywords": ["simplices"], "citation_intent": "background"}
{"citing_id": "2304.11141v1", "cited_id": "1804.03728", "section_title": "Ii. The Proposed H2Tf", "citation": "Sometimes the transform \u03c6 \u03b8 (\u2022) may not be learnable (e.g., the fixed DFT #REFR ), and in those situations the optimization variable only includes Z.", "text_before_citation": ["Suppose that the noisy HSI Y \u2208 R h\u00d7w\u00d7b admits Y = X + N , where X denotes the clean HSI and N denotes noise.", "To infer the underlying clean HSI X from the observed Y, t-SVD method generally formulates the following model:", "EQUATION", "Here, L(Y, X ) denotes the fidelity term and \u03c8(Z (k) ) represents the low-rank characterization of Z (k) (which denotes the k-th frontal (spatial) slice of Z \u2208 R h\u00d7w\u00d7b #OTHEREFR ).", "\u03c6 \u03b8 (\u2022) : R h\u00d7w\u00d7b \u2192 R h\u00d7w\u00d7b denotes a transform with learnable parameters \u03b8, which transforms the low-rank representation Z into the original domain."], "text_after_citation": ["The philosophy of the t-SVD model (1) is to minimize the rank in the transformed domain, which can model the implicit low-rankness of HSI.", "There are naturally two key building blocks for exactly modeling the implicit low-rankness, i.e., the selection of the transform \u03c6 \u03b8 (\u2022) and the exact low-rank characterization \u03c8(\u2022) of the transformed frontal slice Z (k) .", "Most t-SVD-based methods focus on the design of different transforms \u03c6 \u03b8 (\u2022) (see examples in #OTHEREFR , #OTHEREFR , #OTHEREFR ), but all of them pay less attention to the exact characterization of the transformed frontal slice.", "They mostly adopt shallow representations to characterize Z (k) , e.g., MF #OTHEREFR , #OTHEREFR , QR factorization #OTHEREFR , and nuclear norm #OTHEREFR , #OTHEREFR .", "However, these shallow representations may not be expressive enough to capture fine details of the clean HSI."], "citing_paper_content": {"title": "H2Tf For Hyperspectral Image Denoising: Where Hierarchical Nonlinear Transform Meets Hierarchical Matrix Factorization", "abstract": "Recently, tensor singular value decomposition (t-SVD) has emerged as a promising tool for hyperspectral image (HSI) processing. In the t-SVD, there are two key building blocks: (i) the low-rank enhanced transform and (ii) the accompanying low-rank characterization of transformed frontal slices. Previous t-SVD methods mainly focus on the developments of (i), while neglecting the other important aspect, i.e., the exact characterization of transformed frontal slices. In this letter, we exploit the potentiality in both building blocks by leveraging the Hierarchical nonlinear transform and the Hierarchical matrix factorization to establish a new Tensor Factorization (termed as H2TF). Compared to shallow counter partners, e.g., low-rank matrix factorization or its convex surrogates, H2TF can better capture complex structures of transformed frontal slices due to its hierarchical modeling abilities. We then suggest the H2TF-based HSI denoising model and develop an alternating direction method of multipliers-based algorithm to address the resultant model. Extensive experiments validate the superiority of our method over state-of-the-art HSI denoising methods."}, "cited_paper_content": {"title": "Tensor Robust Principal Component Analysis With A New Tensor Nuclear Norm", "abstract": "In this paper, we consider the Tensor Robust Principal Component Analysis (TRPCA) problem, which aims to exactly recover the low-rank and sparse components from their sum. Our model is based on the recently proposed tensor-tensor product (or t-product) [15]. Induced by the t-product, we first rigorously deduce the tensor spectral norm, tensor nuclear norm, and tensor average rank, and show that the tensor nuclear norm is the convex envelope of the tensor average rank within the unit ball of the tensor spectral norm. These definitions, their relationships and properties are consistent with matrix cases. Equipped with the new tensor nuclear norm, we then solve the TRPCA problem by solving a convex program and provide the theoretical guarantee for the exact recovery. Our TRPCA model and recovery guarantee include matrix RPCA as a special case. Numerical experiments verify our results, and the applications to image recovery and background modeling problems demonstrate the effectiveness of our method."}, "keywords": ["transform", "\u2022"], "citation_intent": "background"}
{"citing_id": "2303.08825v1", "cited_id": "1911.03084", "section_title": "Ii. System Model", "citation": "It is responsible for adaptively adjusting the phase shift of each reflecting element in terms of the acquired CSI through periodic channel estimation #REFR .", "text_before_citation": ["with the Rician factor \u0393, the LOS component H LOS , the multipath component H N LOS consisting of independent entries that follow CN (0, 1), and the BS-IRS path loss \u03c3 2 h .", "Consider a macro-cell scenario with far-field assumption, where there is no reflector surrounding the BS and IRS, the signal transmission more likes free-space propagation, which can be computed by", "EQUATION", "where L 0 stands for the path loss at the reference distance of 1 m, and \u03b1 means the path loss exponent.", "A smart controller of the IRS is connected to the BS with a wired or wireless link."], "text_after_citation": ["The signal reflection of a typical element n for user k is mathematically modeled by a reflection coefficient kn = a kn e j\u03c6 kn , where \u03c6 kn \u2208 [0, 2\u03c0) denotes an induced phase shift, and a kn \u2208 [0, 1] stands for amplitude attenuation.", "As mentioned by #OTHEREFR , a kn = 1, \u2200n, k is the optimal attenuation that maximizes the strength of the received signal and simplifies the implementation complexity.", "Hence, the reflection optimization only focuses on the phase shifts \u03c6 kn , \u2200n, k.", "By ignoring hardware impairments such as quantified phase shifts #OTHEREFR and phase noise #OTHEREFR , the k th UE observes the received signal", "EQUATION"], "citing_paper_content": {"title": "Orthogonal And Non-Orthogonal Multiple Access For Intelligent Reflection Surface In 6G Systems", "abstract": "Intelligent reflecting surface (IRS) is envisioned to become a key technology for the upcoming six-generation (6G) wireless system due to its potential of reaping high performance in a power-efficient and cost-efficient way. With its disruptive capability and hardware constraint, the integration of IRS imposes some fundamental particularities on the coordination of multiuser signal transmission. Consequently, the conventional orthogonal and non-orthogonal multiple-access schemes are hard to directly apply because of the joint optimization of active beamforming at the base station and passive reflection at the IRS. Relying on an alternating optimization method, we develop novel schemes for efficient multiple access in IRS-aided multiuser multi-antenna systems in this paper. Achievable performance in terms of the sum spectral efficiency is theoretically analyzed. A comprehensive comparison of different schemes and configurations is conducted through Monte-Carlo simulations to clarify which scheme is favorable for this emerging 6G paradigm."}, "cited_paper_content": {"title": "Channel Estimation For Intelligent Reflecting Surface Assisted Multiuser Communications", "abstract": "In intelligent reflecting surface (IRS) assisted communication systems, the acquisition of channel state information (CSI) is a crucial impediment for achieving the beamforming gain of IRS because of the considerable overhead required for channel estimation. Specifically, under the current beamforming design for IRS-assisted communications, $KMN+KM$ channel coefficients should be estimated, where $K$, $N$ and $M$ denote the numbers of users, IRS reflecting elements, and antennas at the base station (BS), respectively. To accurately estimate such a large number of channel coefficients within a short time interval, we propose a novel three-phase pilot-based channel estimation framework in this paper for IRS-assisted uplink multiuser communications. Under this framework, we analytically prove that a time duration consisting of $K+N+\\max(K-1,\\lceil (K-1)N/M \\rceil)$ pilot symbols is sufficient for the BS to perfectly recover all the $KMN+KM$ channel coefficients for the case without receiver noise at the BS. In contrast to the channel estimation for conventional uplink communications without IRS where the minimum channel estimation time is independent of the number of receive antennas at the BS, our result reveals the crucial role of massive MIMO (multiple-input multiple-output) in reducing the channel estimation time for IRS-assisted communications. Further, for the case with receiver noise, the user pilot sequences, IRS reflecting coefficients, and BS linear minimum mean-squared error (LMMSE) channel estimators are characterized in closed-form, and the corresponding estimation mean-squared error (MSE) is quantified."}, "keywords": ["periodic channel estimation"], "citation_intent": "background"}
{"citing_id": "2304.12951v1", "cited_id": "1812.02822", "section_title": "Semantic Editing", "citation": "Figure 3 illustrates a few of such basis functions for the generative decoder of the IM-Net model #REFR .", "text_before_citation": ["In representation learning the aim is to explain observations with a small number of latent variables #OTHEREFR , which we will denote by l.", "Generative models f (x; l, \u0398), such as (variational) auto-encoders or generative adversarial networks, attempt to map novel latent codes to novel outputs within the same distribution as the observations.", "When interpolating between two latent codes, the appearance of the generated output changes continuously #OTHEREFR .", "For a generative neural implicit shape model, the basis for the boundary movement is locally described by b = \u2212\u2207 l f / \u2207 x f (Equation 3)."], "text_after_citation": ["The decoder is trained as part of an auto-encoder reconstructing the entire ShapeNet dataset #OTHEREFR from l \u2208 R 256 latent variables.", "We use a pretrained model available at https://github.com/czq142857/ IM-NET-pytorch, which is implemented as an MLP with 8 layers of 1024 neurons each and leaky-ReLU activations.", "These basis functions can be observed to encode semantics, such as rotation and reflection symmetries, and segment semantically meaningful parts, such as windscreen on the car, rim of the lamp and different surfaces on chairs.", "For similar parts, such as the two chairs, the bases show similar behaviour.", "However, it is known that not all directions in the latent space are (equally) viable #OTHEREFR , hence, a similar argument can be made for the bases."], "citing_paper_content": {"title": "Neural Implicit Shape Editing Using Boundary Sensitivity", "abstract": "Neural fields are receiving increased attention as a geometric representation due to their ability to compactly store detailed and smooth shapes and easily undergo topological changes. Compared to classic geometry representations, however, neural representations do not allow the user to exert intuitive control over the shape. Motivated by this, we leverage boundary sensitivity to express how perturbations in parameters move the shape boundary. This allows to interpret the effect of each learnable parameter and study achievable deformations. With this, we perform geometric editing: finding a parameter update that best approximates a globally prescribed deformation. Prescribing the deformation only locally allows the rest of the shape to change according to some prior, such as semantics or deformation rigidity. Our method is agnostic to the model its training and updates the NN in-place. Furthermore, we show how boundary sensitivity helps to optimize and constrain objectives (such as surface area and volume), which are difficult to compute without first converting to another representation, such as a mesh."}, "cited_paper_content": {"title": "Learning Implicit Fields For Generative Shape Modeling", "abstract": "We advocate the use of implicit fields for learning generative models of shapes and introduce an implicit field decoder, called IM-NET, for shape generation, aimed at improving the visual quality of the generated shapes. An implicit field assigns a value to each point in 3D space, so that a shape can be extracted as an iso-surface. IM-NET is trained to perform this assignment by means of a binary classifier. Specifically, it takes a point coordinate, along with a feature vector encoding a shape, and outputs a value which indicates whether the point is outside the shape or not. By replacing conventional decoders by our implicit decoder for representation learning (via IM-AE) and shape generation (via IM-GAN), we demonstrate superior results for tasks such as generative shape modeling, interpolation, and single-view 3D reconstruction, particularly in terms of visual quality. Code and supplementary material are available at https://github.com/czq142857/implicit-decoder."}, "keywords": ["generative decoder"], "citation_intent": "background"}
{"citing_id": "2303.08583v1", "cited_id": "1702.06370", "section_title": "Factorized Ring Computation", "citation": "Figure 5 gives the contents of the views in the view tree from Figure 2 , where R, S, and T are relations over a ring D with payloads p i \u2208 D for i \u2208 #REFR .", "text_before_citation": ["If X is a free variable, however, then we retain it in the view schema without applying the lifting function to its values.", "The schema of the view consists of dep(X) and the free variables in the subtree of \u03c9 rooted at X.", "\u03c4 (variable order \u03c9, free variables F) : view tree switch \u03c9: Figure 3 : Creating a view tree \u03c4 (\u03c9, F) for a variable order \u03c9 and a set of free variables F. Example 8.", "Figure 2 shows the view tree constructed by the function \u03c4 from Figure 3 over the variable order \u03c9 and the empty set of free variables.", "Figure 4 depicts the view tree constructed over the same variable order but for the set F = {A, C} of free variables."], "text_after_citation": ["Assume that D is the Z ring, each tuple in these relations is mapped to 1, i.e., p i = 1 for i \u2208 #OTHEREFR , and the lifting functions map all values to 1.", "Then, the view tree computes the COUNT query from Example 3 and the root view V @A", "R return R[sch(R)] X \u03c9 1 . .", ".\u03c9 k let T i = \u03c4 (\u03c9 i , F), \u2200i \u2208 [k] let V @\u03c9 i rels i [keys i ] = root of T i , \u2200i \u2208 [k] let keys = dep(X) \u222a (F \u2229 vars(\u03c9)) let rels = i\u2208[k] rels i if X / \u2208 F V @X rels [keys] = X i\u2208[k] V @\u03c9 i rels i [keys i ] else V @X rels [keys] = i\u2208[k] V @\u03c9 i rels i [keys i ] return \uf8f1 \uf8f4 \uf8f2 \uf8f4 \uf8f3 V @X rels [keys] T 1 . . . T k", "V @A RST [A, C ] V @B R [A] V @C ST [A, C ] V @D T [C ] V @E S [A, C ] R[A, B ] S[A, C , E ] T[C , D] V @A RST [A, C ] = V @B R [A] \u2297 V @C ST [A, C ] V @C ST [A, C ] = V @D T [C ] \u2297 V @E S [A, C ] V @E S [A, C ] = E S[A, C , E ] V @B R [A] = B R[A, B ] V @D T [C ] = D T[C , D]"], "citing_paper_content": {"title": "F-Ivm: Analytics Over Relational Databases Under Updates", "abstract": "This article describes F-IVM, a unified approach for maintaining analytics over changing relational data. We exemplify its versatility in four disciplines: processing queries with group-by aggregates and joins; learning linear regression models using the covariance matrix of the input features; building Chow-Liu trees using pairwise mutual information of the input features; and matrix chain multiplication. F-IVM has three main ingredients: higher-order incremental view maintenance; factorized computation; and ring abstraction. F-IVM reduces the maintenance of a task to that of a hierarchy of simple views. Such views are functions mapping keys, which are tuples of input values, to payloads, which are elements from a ring. F-IVM also supports efficient factorized computation over keys, payloads, and updates. Finally, F-IVM treats uniformly seemingly disparate tasks. In the key space, all tasks require joins and variable marginalization. In the payload space, tasks differ in the definition of the sum and product ring operations. We implemented F-IVM on top of DBToaster and show that it can outperform classical first-order and fully recursive higher-order incremental view maintenance by orders of magnitude while using less memory."}, "cited_paper_content": {"title": "Answering Conjunctive Queries Under Updates", "abstract": "We consider the task of enumerating and counting answers to k-ary conjunctive queries against relational databases that may be updated by inserting or deleting tuples. We exhibit a new notion of q-hierarchical conjunctive queries and show that these can be maintained efficiently in the following sense. During a linear time pre-processing phase, we can build a data structure that enables constant delay enumeration of the query results; and when the database is updated, we can update the data structure and restart the enumeration phase within constant time. For the special case of self-join free conjunctive queries we obtain a dichotomy: if a query is not q-hierarchical, then query enumeration with sublinear *) delay and sublinear update time (and arbitrary preprocessing time) is impossible. For answering Boolean conjunctive queries and for the more general problem of counting the number of solutions of k-ary queries we obtain complete dichotomies: if the query's homomorphic core is q-hierarchical, then size of the the query result can be computed in linear time and maintained with constant update time. Otherwise, the size of the query result cannot be maintained with sublinear update time. All our lower bounds rely on the OMv-conjecture, a conjecture on the hardness of online matrix-vector multiplication that has recently emerged in the field of fine-grained complexity to characterise the hardness of dynamic problems. The lower bound for the counting problem additionally relies on the orthogonal vectors conjecture, which in turn is implied by the strong exponential time hypothesis.*) By sublinear we mean O(n(1-e) for some e > 0, where n is the size of the active domain of the current database."}, "keywords": ["view tree", "relations"], "citation_intent": "background"}
{"citing_id": "2304.13242v1", "cited_id": "2002.11477", "section_title": "V. Experiments", "citation": "We conclude that the probabilistic objective (12) substantially reduces bias compared with the non-probabilistic DSLA affordance objective #REFR .", "text_before_citation": ["Ablation studies We evaluate the advantage of our proposed predictive world modeling approach [15] for learning navigational patterns from sampled plausible completed worldsx instead of partially observed worlds x.", "We conduct an experiment using unaugmented samples to quantify the performance contribution of our geometric data augmentation method #OTHEREFR on real-world data.", "We conduct experiments on dataset splits including a different number of regions to estimate how performance increases with additional data or experience.", "Table I presents evaluation results for the predicted probability fields.", "Our proposed DSLP model optimized with the information balance regularizer \u03b1 IB (9) predicts the least biased probability field among all models trained and evaluated on accumulated past observation inputs."], "text_after_citation": ["Training and evaluating on accumulated past and future observation inputs in an offline map creation manner (i.e.", "full obs) reduces bias, demonstrating that more comprehensively observed environments result in better performance.", "We performed experiments with different constant \u03b1 values to #OTHEREFR .", "The best constant weight \u03b1 value 0.1, found over five hyperparameter experiments, result in worse performance than using \u03b1 IB .", "We demonstrate the merit of dynamic, per-sample computed \u03b1 IB values (9) by running an experiment with the constant mean \u03b1 IB value 0.122 computed over all training samples, which results in worse performance. See Fig. 7 for probability field visualizations."], "citing_paper_content": {"title": "Learning To Predict Navigational Patterns From Partial Observations", "abstract": "Human beings cooperatively navigate ruleconstrained environments by adhering to mutually known navigational patterns, which may be represented as directional pathways or road lanes. Inferring these navigational patterns from incompletely observed environments is required for intelligent mobile robots operating in unmapped locations. However, algorithmically defining these navigational patterns is nontrivial. This paper presents the first self-supervised learning (SSL) method for learning to infer navigational patterns in realworld environments from partial observations only. We explain how geometric data augmentation, predictive world modeling, and an information-theoretic regularizer enables our model to predict an unbiased local directional soft lane probability (DSLP) field in the limit of infinite data. We demonstrate how to infer global navigational patterns by fitting a maximum likelihood graph to the DSLP field. Experiments show that our SSL model outperforms two SOTA supervised lane graph prediction models on the nuScenes dataset. We propose our SSL method as a scalable and interpretable continual learning paradigm for navigation by perception. Code released upon publication."}, "cited_paper_content": {"title": "Learning A Directional Soft Lane Affordance Model For Road Scenes Using Self-Supervision", "abstract": "Humans navigate complex environments in an organized yet flexible manner, adapting to the context and implicit social rules. Understanding these naturally learned patterns of behavior is essential for applications such as autonomous vehicles. However, algorithmically defining these implicit rules of human behavior remains difficult. This work proposes a novel self-supervised method for training a probabilistic network model to estimate the regions humans are most likely to drive in as well as a multimodal representation of the inferred direction of travel at each point. The model is trained on individual human trajectories conditioned on a representation of the driving environment. The model is shown to successfully generalize to new road scenes, demonstrating potential for real-world application as a prior for socially acceptable driving behavior in challenging or ambiguous scenarios which are poorly handled by explicit traffic rules."}, "keywords": ["non-probabilistic DSLA affordance"], "citation_intent": "background"}
{"citing_id": "2305.00521v1", "cited_id": "1801.03924", "section_title": "Conclusion", "citation": "In all cases of generator tuning, CSIM scores are higher than the zero-shot case and comparable with each other since we minimize the perceptual distance between the reference and generated frame #REFR .", "text_before_citation": ["In contrast to existing lip-sync generative models, we introduce poseaware masking for lip region by utilizing a 3D parametric mesh predictor, which inherits the pose information in the mask itself.", "Designing a smooth lip motion by using the moving-average based latent smoothing in the continuous and linear latent space, StyleLipSync can generate temporally consistent lip motion.", "Furthermore, we propose a fewshot lip-sync adaptation method for a single person who lies in the out-of-distribution, which uses a few seconds of the target person's video.", "Experimental results show that our StyleLipSync can generate realistic lip-sync video from arbitrary audio even with the zero-shot setting and the proposed adaptation method enhance the person-specific information without losing the lip-sync generality. Unseen Face Adaptation.", "We plot the CSIM [9] scores in Figure 9 under the same set-up in Figure 7 (b) to further support the necessity of the sync regularizer R sync ."], "text_after_citation": ["However, the sync regularizer R sync enforces maintaining the lip-sync generality, as illustrated in Figure 7 Figure 9 .", "Cosine similarity between face embedding (CSIM) #OTHEREFR tendency of the adaptation method."], "citing_paper_content": {"title": "Stylelipsync: Style-Based Personalized Lip-Sync Video Generation", "abstract": "In this paper, we present StyleLipSync, a style-based personalized lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from arbitrary audio. To generate a video of arbitrary identities, we leverage expressive lip prior from the semantically rich latent space of a pre-trained StyleGAN, where we can also design a video consistency with a linear transformation. In contrast to the previous lip-sync methods, we introduce pose-aware masking that dynamically locates the mask to improve the naturalness over frames by utilizing a 3D parametric mesh predictor frame by frame. Moreover, we propose a few-shot lip-sync adaptation method for an arbitrary person by introducing a sync regularizer that preserves lips-sync generalization while enhancing the person-specific visual information. Extensive experiments demonstrate that our model can generate accurate lip-sync videos even with the zeroshot setting and enhance characteristics of an unseen face using a few seconds of target video through the proposed adaptation method. Please refer to our project page."}, "cited_paper_content": {"title": "The Unreasonable Effectiveness Of Deep Features As A Perceptual Metric", "abstract": "While it is nearly effortless for humans to quickly assess the perceptual similarity between two images, the underlying processes are thought to be quite complex. Despite this, the most widely used perceptual metrics today, such as PSNR and SSIM, are simple, shallow functions, and fail to account for many nuances of human perception. Recently, the deep learning community has found that features of the VGG network trained on the ImageNet classification task has been remarkably useful as a training loss for image synthesis. But how perceptual are these so-called\"perceptual losses\"? What elements are critical for their success? To answer these questions, we introduce a new Full Reference Image Quality Assessment (FR-IQA) dataset of perceptual human judgments, orders of magnitude larger than previous datasets. We systematically evaluate deep features across different architectures and tasks and compare them with classic metrics. We find that deep features outperform all previous metrics by huge margins. More surprisingly, this result is not restricted to ImageNet-trained VGG features, but holds across different deep architectures and levels of supervision (supervised, self-supervised, or even unsupervised). Our results suggest that perceptual similarity is an emergent property shared across deep visual representations."}, "keywords": ["generated frame", "perceptual distance"], "citation_intent": "result"}
{"citing_id": "2304.04336v1", "cited_id": "1812.02713", "section_title": "Shape Instance Segmentation", "citation": "Note that instance segmentation for the airplane category is not reported because PartNet #REFR does not have ground-truth labels. The results are reported in Tab. 3.", "text_before_citation": ["Since our bounding boxes are based on part-level segmentation, we also measure its performance by measuring the mean average precision (mAP) of instance segmentation with the labels obtained from PartNet #OTHEREFR .", "In addition to HA [34] and CA [37], we also compare with BSP-Net #OTHEREFR which was used to get our initial segment.", "We follow the evaluation setting of PartNet #OTHEREFR to report the instance segmentation results.", "The segmentation was obtained by identifying the nearest point for each primitive and assigning them to obtain the segmentation mask.", "Subsequently, we computed the IoU between each segmentation mask and the ground-truth mask and considered a segmentation mask as true positive only when IoU was greater than 0.5."], "text_after_citation": ["Compared to other baselines we achieve the best result showing that SMART captures the individual parts in the shape better than other works having more alignment with the human perception of the shape decomposition."], "citing_paper_content": {"title": "Split, Merge, And Refine: Fitting Tight Bounding Boxes Via Learned Over-Segmentation And Iterative Search", "abstract": "Figure 1: Our framework dubbed SMART finds a set of tight bounding boxes of a 3D shape. Given a 3D shape represented as a volumetric mesh, it first takes over-segmentation from the output of BSP-Net [8] and post-processing and then performs hierarchical merging with tightness-aware criteria (second row). Then, it refines the bounding box parameters with a sequence of actions in a discrete space (third row). The results are further improved with long-sighted exploration by our MCTS accelerated with our technique (last row)."}, "cited_paper_content": {"title": "Partnet: A Large-Scale Benchmark For Fine-Grained And Hierarchical Part-Level 3D Object Understanding", "abstract": "We present PartNet: a consistent, large-scale dataset of 3D objects annotated with fine-grained, instance-level, and hierarchical 3D part information. Our dataset consists of 573,585 part instances over 26,671 3D models covering 24 object categories. This dataset enables and serves as a catalyst for many tasks such as shape analysis, dynamic 3D scene modeling and simulation, affordance analysis, and others. Using our dataset, we establish three benchmarking tasks for evaluating 3D part recognition: fine-grained semantic segmentation, hierarchical semantic segmentation, and instance segmentation. We benchmark four state-of-the-art 3D deep learning algorithms for fine-grained semantic segmentation and three baseline methods for hierarchical semantic segmentation. We also propose a baseline method for part instance segmentation and demonstrate its superior performance over existing methods."}, "keywords": ["instance segmentation"], "citation_intent": "result"}
{"citing_id": "2304.03659v1", "cited_id": "1405.0312", "section_title": "Probe-B: Contextual Understanding", "citation": "To generate a dataset that probes for model understanding on objects and their relationship to contextual cues found in an image's background, we collected samples from MSCOCO #REFR consisting of 80 objects.", "text_before_citation": [], "text_after_citation": ["These samples are used to probe model reliance on background cues and reliance on co-occurrence between objects.", "An overview of Probe-B is shown in Figure 2 .", "For each group, there is an unmodified image x 0 , an image with a random patch on the backgroundx 0 , a modified image where the background is removedx 1 and 80 or less prompts.", "We have two splits in this data, the first removing the background but keeping all objects Probe-B M R and the other removing both background and all other objects Probe-B R .", "Probe-B R aims to probe models on whether they use conceptual maps on object co-occurrence to improve recognition."], "citing_paper_content": {"title": "Probing Conceptual Understanding Of Large Visual-Language Models", "abstract": "We present a novel framework for probing and improving relational, compositional and contextual understanding of large visual-language models (V+L). While large V+L models have achieved success in various downstream tasks, it is not clear if they have a conceptual grasp of the content. We propose a novel benchmarking dataset for probing three aspects of content understanding. Our probes are grounded in cognitive science and help determine if a V+L model can, for example, determine if snow garnished with a man is implausible, or if it can identify beach furniture by knowing it is located on a beach. We have experimented with 5 well known models, such as CLIP and ViLT, and found that they mostly fail to demonstrate a conceptual understanding. That said, we find interesting insights such as cross-attention helps learning conceptual understandingWe use these insights to propose a new finetuning technique that rewards the three conceptual understanding measures we proposed. We hope that the presented benchmarks will help the community assess and improve the conceptual understanding capabilities of large V+L models. Visual-language (V+L) models are recently and greatly successful deep learning models that learn representations of image and text in a shared space. These representations are useful for downstream tasks like image classification, visual-question answering, image retrieval and more [42, 37, 27, 26, 39, 1]. However, for use in real-world applications, it is also vital that models \"understand\" rather than memorize to perform on more general tasks [24]. While large-language models have been shown to have a moderate amount of \"theory of mind\", as measured by conceptual consistency [29], V+L models have not"}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["contextual cues"], "citation_intent": "method"}
{"citing_id": "2305.01387v1", "cited_id": "2002.11613", "section_title": "Global Model Generation With Lth", "citation": "Different from DPLTM #REFR , our method does not have the privacy budget as we do not use the private data while generating the model structure.", "text_before_citation": ["Unlike the selection method in DPLTM #OTHEREFR , for the winning ticket selection, there are M alternative tickets, and each ticket is given a score V (w j , D pub ) that equals to the correct number of samples for inference.", "To adjust the trade-off between the accuracy and the degree of network pruning, we use the softmax function #OTHEREFR over the preference value V (w j , D pub ) also known as the score to select WT, which ensures that all WTs are explored:", "EQUATION", "where j denotes one of possibly many winning tickets and P j is its corresponding probability.", "It is worth noting that we use LTH to generate WTs on the public dataset, and each client performs the training process with DP on their private data."], "text_after_citation": [], "citing_paper_content": {"title": "Efficient Federated Learning With Enhanced Privacy Via Lottery Ticket Pruning In Edge Computing", "abstract": "Federated learning (FL) is a collaborative learning paradigm for decentralized private data from mobile terminals (MTs). However, it suffers from issues in terms of communication, resource of MTs, and privacy. Existing privacy-preserving FL methods usually adopt the instance-level differential privacy (DP), which provides a rigorous privacy guarantee but with several bottlenecks: severe performance degradation, transmission overhead, and resource constraints of edge devices such as MTs. To overcome these drawbacks, we propose Fed-LTP, an efficient and privacy-enhanced FL framework with Lottery Ticket Hypothesis (LTH) and zero-concentrated DP (zCDP). It generates a pruned global model on the server side and conducts sparse-to-sparse training from scratch with zCDP on the client side. On the server side, two pruning schemes are proposed: (i) the weight-based pruning (LTH) determines the pruned global model structure; (ii) the iterative pruning further shrinks the size of the pruned model's parameters. Meanwhile, the performance of Fed-LTP is also boosted via model validation based on the Laplace mechanism. On the client side, we use sparse-to-sparse training to solve the resource-constraints issue and provide tighter privacy analysis to reduce the privacy budget. We evaluate the effectiveness of Fed-LTP on several real-world datasets in both independent and identically distributed (IID) and non-IID settings. The results clearly confirm the superiority of Fed-LTP over state-of-the-art (SOTA) methods in communication, computation, and memory efficiencies while realizing a better utility-privacy trade-off."}, "cited_paper_content": {"title": "The Differentially Private Lottery Ticket Mechanism", "abstract": "We propose the differentially private lottery ticket mechanism (DPLTM). An end-to-end differentially private training paradigm based on the lottery ticket hypothesis. Using \"high-quality winners\", selected via our custom score function, DPLTM significantly improves the privacy-utility trade-off over the state-of-the-art. We show that DPLTM converges faster, allowing for early stopping with reduced privacy budget consumption. We further show that the tickets from DPLTM are transferable across datasets, domains, and architectures. Our extensive evaluation on several public datasets provides evidence to our claims."}, "keywords": ["private data"], "citation_intent": "method"}
{"citing_id": "2304.06908v1", "cited_id": "1607.02533", "section_title": "A. Adversarial Attack", "citation": "It was later extended to the iterative version of FGSM (I-FGSM) #REFR , which generates adversarial examples that fool the model \u03b8 with a higher success rate.", "text_before_citation": ["Given a pretrained DNN \u03b8, the adversarial attack is implemented by solving the following optimization problem:", "EQUATION", "where the L(x, y, \u03b8) denotes the loss function that guides the DNN \u03b8 to predict the class of image x as ground-truth y, is the maximum allowed magnitude of perturbation, and x * is the generated adversarial example. Some gradient-based optimizers were proposed to solve (1). Szegedy et al.", "were the first to use a box-constrained L-BFGS optimizer #OTHEREFR .", "Then, a one-step optimizer, i.e., fast gradient sign method (FGSM) was proposed to efficiently generate adversarial examples #OTHEREFR ."], "text_after_citation": ["The generated adversarial example is shown to have the ability to attack an unknown model, and this ability is called the transferability of the adversarial example.", "In this scenario, the model \u03b8 used to generate the adversarial example is called the surrogate model, and the unknown model to be attacked is called the victim model.", "This type of adversarial attack is called the transfer-based attack. Dong et al.", "systematically investigated the transferability of adversarial examples and found an interesting phenomenon that is although the iterative method (I-FGSM) is a stronger optimizer than the one-step method (FGSM), the transferability of the adversarial examples generated by I-FGSM is worse than FGSM #OTHEREFR .", "Specifically, their experimental results on the white-box attacks show that an Inception-v3 has only a 72.3% probability of being fooled by FGSM, but a 100% probability of being fooled by I-FGSM."], "citing_paper_content": {"title": "Generating Adversarial Examples With Better Transferability Via Masking Unimportant Parameters Of Surrogate Model", "abstract": "Deep neural networks (DNNs) have been shown to be vulnerable to adversarial examples. Moreover, the transferability of the adversarial examples has received broad attention in recent years, which means that adversarial examples crafted by a surrogate model can also attack unknown models. This phenomenon gave birth to the transfer-based adversarial attacks, which aim to improve the transferability of the generated adversarial examples. In this paper, we propose to improve the transferability of adversarial examples in the transfer-based attack via masking unimportant parameters (MUP). The key idea in MUP is to refine the pretrained surrogate models to boost the transfer-based attack. Based on this idea, a Taylor expansionbased metric is used to evaluate the parameter importance score and the unimportant parameters are masked during the generation of adversarial examples. This process is simple, yet can be naturally combined with various existing gradient-based optimizers for generating adversarial examples, thus further improving the transferability of the generated adversarial examples. Extensive experiments are conducted to validate the effectiveness of the proposed MUP-based methods."}, "cited_paper_content": {"title": "Adversarial Examples In The Physical World", "abstract": "Most existing machine learning classifiers are highly vulnerable to adversarial examples. An adversarial example is a sample of input data which has been modified very slightly in a way that is intended to cause a machine learning classifier to misclassify it. In many cases, these modifications can be so subtle that a human observer does not even notice the modification at all, yet the classifier still makes a mistake. Adversarial examples pose security concerns because they could be used to perform an attack on machine learning systems, even if the adversary has no access to the underlying model. Up to now, all previous work have assumed a threat model in which the adversary can feed data directly into the machine learning classifier. This is not always the case for systems operating in the physical world, for example those which are using signals from cameras and other sensors as an input. This paper shows that even in such physical world scenarios, machine learning systems are vulnerable to adversarial examples. We demonstrate this by feeding adversarial images obtained from cell-phone camera to an ImageNet Inception classifier and measuring the classification accuracy of the system. We find that a large fraction of adversarial examples are classified incorrectly even when perceived through the camera."}, "keywords": ["adversarial examples"], "citation_intent": "method"}
{"citing_id": "2304.13192v1", "cited_id": "1706.04599", "section_title": "F. Model Calibration", "citation": "It is an important step towards improving model interpretability as most of the deep neural networks are typically over-confident in their predictions #REFR .", "text_before_citation": ["Confidence calibration is the problem of matching the output confidence level with the actual likelihood of the model."], "text_after_citation": ["Of note, A model is said to be perfectly calibrated when the confidence level of a prediction represents the true probability of the prediction being correct #OTHEREFR .", "Mathematically speaking, if input X is considered with class labels Y , the predicted class is\u0176 andP is its associated confidence, then for perfect calibration, the probability Pis:", "P(\u0176 = Y |P = p) = p, \u2200 p \u2208 [0, 1]", "where the probability is over the joint distribution.", "1) Temperature Scaling: Temperature scaling is the simplest extension of Platt scaling #OTHEREFR and uses a single parameter T > 0 for all cases. Guo et al."], "citing_paper_content": {"title": "Towards Reliable Colorectal Cancer Polyps Classification Via Vision Based Tactile Sensing And Confidence-Calibrated Neural Networks", "abstract": "In this study, toward addressing the over-confident outputs of existing artificial intelligence-based colorectal cancer (CRC) polyp classification techniques, we propose a confidencecalibrated residual neural network. Utilizing a novel visionbased tactile sensing (VS-TS) system and unique CRC polyp phantoms, we demonstrate that traditional metrics such as accuracy and precision are not sufficient to encapsulate model performance for handling a sensitive CRC polyp diagnosis. To this end, we develop a residual neural network classifier and address its over-confident outputs for CRC polyps classification via the post-processing method of temperature scaling. To evaluate the proposed method, we introduce noise and blur to the obtained textural images of the VSTS and test the model's reliability for non-ideal inputs through reliability diagrams and other statistical metrics."}, "cited_paper_content": {"title": "On Calibration Of Modern Neural Networks", "abstract": "Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions."}, "keywords": ["deep neural networks"], "citation_intent": "background"}
{"citing_id": "2304.02009v1", "cited_id": "1909.03752", "section_title": "Pose Estimation By Template Matching", "citation": "When the distribution is mostly unimodal, we can obtain a measure of uncertainty as the covariance of P around \u03be * #REFR .", "text_before_citation": ["EQUATION", "where \u03be(p) transforms a 2D point p from BEV to map coordinate frame.", "The confidence C masks the correlation to ignore some parts of the BEV space, such as occluded areas.", "This formulation benefits from an efficient implementation by rotating T K times and performing a single convolution as a batched multiplication in the Fourier domain #OTHEREFR .", "Pose inference: We estimate a single pose by maximum likelihood: \u03be * = argmax \u03be P (\u03be|I, map, \u03be prior )."], "text_after_citation": ["We train a single model that generalizes well across many datasets.", "OrienterNet handles different cameras, street-level viewpoints, and unseen cities and is thus suitable for both AR and robotics.", "Overlayed on the input maps, the single-image predictions (black arrow ) are close to the ground truth (red arrow) and more accurate than the noisy GPS (blue dot \u2022).", "The model effectively leverages building corners and boundaries, crosswalks, sidewalks, road intersections, trees, and other common urban objects.", "Likelihood maps can be multi-modal in scenes with repeated elements (last two rows) -we show the predicted orientations at local maxima as small arrows. \u2022GPS \u2022single-frame \u2022sequence \u2022GT Figure 5 . Multi-frame fusion resolves ambiguities."], "citing_paper_content": {"title": "Orienternet: Visual Localization In 2D Public Maps With Neural Matching", "abstract": "Humans can orient themselves in their 3D environments using simple 2D maps. Differently, algorithms for visual localization mostly rely on complex 3D point clouds that are expensive to build, store, and maintain over time. We bridge this gap by introducing OrienterNet, the first deep neural network that can localize an image with sub-meter accuracy using the same 2D semantic maps that humans use. Orienter-Net estimates the location and orientation of a query image by matching a neural Bird's-Eye View with open and globally available maps from OpenStreetMap, enabling anyone to localize anywhere such maps are available. OrienterNet is supervised only by camera poses but learns to perform semantic matching with a wide range of map elements in an end-to-end manner. To enable this, we introduce a large crowd-sourced dataset of images captured across 12 cities from the diverse viewpoints of cars, bikes, and pedestrians. OrienterNet generalizes to new datasets and pushes the state of the art in both robotics and AR scenarios. The code and trained model will be released publicly."}, "cited_paper_content": {"title": "Masking By Moving: Learning Distraction-Free Radar Odometry From Pose Information", "abstract": "This paper presents an end-to-end radar odometry system which delivers robust, real-time pose estimates based on a learned embedding space free of sensing artefacts and distractor objects. The system deploys a fully differentiable, correlation-based radar matching approach. This provides the same level of interpretability as established scan-matching methods and allows for a principled derivation of uncertainty estimates. The system is trained in a (self-)supervised way using only previously obtained pose information as a training signal. Using 280km of urban driving data, we demonstrate that our approach outperforms the previous state-of-the-art in radar odometry by reducing errors by up 68% whilst running an order of magnitude faster."}, "keywords": ["covariance"], "citation_intent": "background"}
{"citing_id": "2304.11072v1", "cited_id": "1909.03496", "section_title": "Rq2: Can Our Classifier Learn Vulnerability In A Biased Setting?", "citation": "Compared to the next best model #REFR , our model with Focal Loss shows an improvement of 22.91% in Precision, 23.53% in Recall, and 18.04% on F1 score.", "text_before_citation": ["Observing the last two rows in Table 5 , we find that, initially, we tested our model using Weighted Loss (WL) only, and later we tested with Focal Loss, which is a combination of weighted loss \u03b1 with the hyperparameter \u03b3.", "In both cases, we observe that our model has surpassed previous models in terms of Precision, Recall, and F1 score, indicating lower false positive and false negative rates with WL and with FL.", "However, these numbers improved slightly when compared with performance between weighted loss and Focal Loss, with the exception of the precision metric.", "From the results, it is observed that with Focal Loss, we achieved an improvement of 11.18% in Precision, 1.06% in Recall, and 0.61% on our F1 score compared to the weighted loss.", "Nevertheless, using either weighted loss or Focal Loss, the accuracy of our model drops by almost 1% compared to #OTHEREFR , indicating that these previous models are highly biased with an assumption of non-vulnerability."], "text_after_citation": ["Thus, Focal Loss improves precision, recall and F1 on imbalanced data while not causing a negating impact on a balanced dataset only by adjusting the parameters \u03b1 and \u03b3 RQ3: Is our classifier generalized enough to detect vulnerabilities in N-day and zero-day program sam- We evaluated our classifier's performance based on its ability to accurately predict vulnerability with 273 Nday real-world sample programs. These sample programs are never used during training.", "We also used 4 zero-day examples in order to evaluate our classifier on predicting zero-day vulnerabilities as well.", "The classifier predicts the vulnerability classif the vulnerability exists in the code and predicts non-vulnerable when the vulnerability does not exist.", "Out of these 273 N-day and 4, zero-day code samples, some vulnerability classes exist that are not part of our VulF dataset from table 1.", "For example, the VulF dataset has no code samples for CWE-787 for training."], "citing_paper_content": {"title": "An Unbiased Transformer Source Code Learning With Semantic Vulnerability Graph", "abstract": "Over the years, open-source software systems have become prey to threat actors. Even highly-adopted software has been crippled by unforeseeable attacks, leaving millions of devices exposed. Even as open-source communities act quickly to patch the breach, code vulnerability screening should be an integral part of agile software development from the beginning. Unfortunately, current vulnerability screening techniques are ineffective at identifying novel vulnerabilities or providing developers with code vulnerability and classification. Furthermore, the datasets used for vulnerability learning often exhibit distribution shifts from the real-world testing distribution due to novel attack strategies deployed by adversaries and as a result, the machine learning model's performance may be hindered or biased. To address these issues, we propose a joint interpolated multitasked unbiased vulnerability classifier comprising a transformer \"RoBERTa\" and graph convolution neural network (GCN). We present a training process utilizing a semantic vulnerability graph (SVG) representation from source code, created by integrating edges from a sequential flow, control flow, and data flow, as well as a novel flow dubbed Poacher Flow (PF). Poacher flow edges reduce the gap between dynamic and static program analysis and handle complex long-range dependencies. Moreover, our approach reduces biases of classifiers regarding unbalanced datasets by integrating Focal Loss objective function along with SVG. Remarkably, experimental results show that our classifier outperforms state-of-the-art results on vulnerability detection with fewer false negatives and false positives. After testing our model across multiple datasets, it shows an improvement of at least 2.41% and 18.75% in the bestcase scenario. Evaluations using N-day program samples demonstrate that our proposed approach achieves a 93% accuracy and was able to detect 4, zero-day vulnerabilities from popular GitHub repositories. Our code and data are"}, "cited_paper_content": {"title": "Devign: Effective Vulnerability Identification By Learning Comprehensive Program Semantics Via Graph Neural Networks", "abstract": "Vulnerability identification is crucial to protect the software systems from attacks for cyber security. It is especially important to localize the vulnerable functions among the source code to facilitate the fix. However, it is a challenging and tedious process, and also requires specialized security expertise. Inspired by the work on manually-defined patterns of vulnerabilities from various code representation graphs and the recent advance on graph neural networks, we propose Devign, a general graph neural network based model for graph-level classification through learning on a rich set of code semantic representations. It includes a novel Conv module to efficiently extract useful features in the learned rich node representations for graph-level classification. The model is trained over manually labeled datasets built on 4 diversified large-scale open-source C projects that incorporate high complexity and variety of real source code instead of synthesis code used in previous works. The results of the extensive evaluation on the datasets demonstrate that Devign outperforms the state of the arts significantly with an average of 10.51% higher accuracy and 8.68\\% F1 score, increases averagely 4.66% accuracy and 6.37% F1 by the Conv module."}, "keywords": ["Focal Loss", "Recall"], "citation_intent": "result"}
{"citing_id": "2303.17930v1", "cited_id": "1810.04805", "section_title": "Skill Extraction From Job Description", "citation": "Word vector would be obtained from this component by converting text into low-dimensional dense vectors based on pre-trained Bert model #REFR .", "text_before_citation": ["For skill extraction from job description, the model train on the AutoDL platform [27] and it will be loaded on my laptop."], "text_after_citation": ["As it is considered as classification task in Bert, token [CLS] would be added to the start of the sentence and token [SEP] would be added to the end of the sentence.", "Generated input id is obtained by looking up words in the pre-trained Bert model and replacing it with sub-word and specific token, which is based on WordPiece algorithm.", "The processed input sentence, generated input id, will be fed into Bert model.", "A list of skills id related to the job description will be provided and remain the first 20 th skills id, which will be changed into skills label."], "citing_paper_content": {"title": "Table Of Contents", "abstract": "Due to the increasing number of graduates, many applicants experience the situation about finding a job, and employers experience difficulty filtering job applicants, which might negatively impact their effectiveness. However, most job-hunting websites lack job recommendation and CV filtering or ranking functionality, which are not integrated into the system. Thus, an smart job hunter combined with the above functionality will be conducted in this project, which contains job recommendations, CV ranking and even a job dashboard for skills and job applicant's functionality. Job recommendation and CV ranking starts from the automatic keyword extraction and end with the Job/CV ranking algorithm. Automatic keyword extraction is implemented by Job2Skill and the CV2Skill model based on Bert. Job2Skill consists of two components, text encoder and Gru-based layers, while CV2Skill is mainly based on Bert and fine-tunes the pre-trained model by the Resume-Entity dataset. Besides, to match skills from CV and job description and rank lists of job and candidates, job/CV ranking algorithms have been provided to compute the occurrence ratio of skill words based on TFIDF score and match ratio of the total skill numbers. Besides, some advanced features have been integrated into the website to improve user experiences, such as the calendar and sweetalert2 plugin. And some basic features to go through job application processes, such as job application tracking and interview arrangement. This paper has been divided into seven parts. The first part introduces the aims, objectives and challenges of the project. The second part provided this project's three requirements: essential, recommended and optional. The third part introduces the background of the project and some related techniques that will be used in the project. The fourth part describes the project's system design, including the frontend, backend, model architecture and ranking algorithm. The fifth part shows the details of implementing the functionality and automatic keyword extraction. The sixth part describes the evaluation methods for the website, and the last part summarises the project and future work."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["pre-trained Bert model"], "citation_intent": "method"}
{"citing_id": "2303.11114v1", "cited_id": "1905.04899", "section_title": "B.5. Robustness Benchmarks", "citation": "As the original DeiT is trained on strong augmentation, such as RandAugment or 3-Augment, we also compare our method with \"weak augmented\" ViT-S, where it only employs resized random crop (RRC) and CutMix #REFR .", "text_before_citation": ["We compare ViT-S models trained on ImageNet-1k with different training strategies using robustness benchmarks.", "We employ three scenarios: (1) noise and blur scenario (2) domain shift scenario (3) adversarial attack scenario.", "For the first scenario, we add Gaussian noise and Gaussian blur to the validation images.", "We use ImageNet-R #OTHEREFR and Sketch-ImageNet #OTHEREFR for testing the robustness against domain shifts.", "Finally, we use a weak version of AutoAttak #OTHEREFR for measuring adversarial robustness."], "text_after_citation": ["Our assumption is that because the pixel-trained models are sensitive to imperceptible details, they will be less robust than our approach in noise or adversarial attack scenarios.", "However, on the other hand, because our method relies on the encoding power of the pre-trained tokenizer, if the employed tokenizer is not a robust feature extractor, our method could be more vulnerable than pixel-trained counterparts.", "Table B .4 shows the results of the first and the second scenarios. Here, we observe two important findings.", "First, when we use the same augmentations with the same strength (ViT-S Weak Aug vs.", "ViT-S SeiT), SeiT shows smaller performance drops on both noise scenarios and do- main shift scenarios."], "citing_paper_content": {"title": "Seit: Storage-Efficient Vision Training With Tokens Using 1% Of Pixel Storage", "abstract": "We need billion-scale images to achieve more generalizable and groundbreaking vision models, as well as massive dataset storage to ship the images (e.g., the LAION-4B dataset needs 240TB storage space). However, it has become challenging to deal with unlimited dataset storage with limited storage infrastructure. A number of storageefficient training methods have been proposed to tackle the problem, but they are rarely scalable or suffer from severe damage to performance. In this paper, we propose a storage-efficient training strategy for vision classifiers for large-scale datasets (e.g., ImageNet) that only uses 1024 tokens per instance without using the raw level pixels; our token storage only needs <1% of the original JPEG-compressed raw pixels. We also propose token augmentations and a Stem-adaptor module to make our approach able to use the same architecture as pixel-based approaches with only minimal modifications on the stem layer and the carefully tuned optimization settings. Our experimental results on ImageNet-1k show that our method significantly outperforms other storage-efficient training methods with a large gap. We further show the effectiveness of our method in other practical scenarios, storage-efficient pre-training, and continual learning. Code is available at https://github.com/naver-ai/seit Adjusting image resolution Adjusting JPEG quality C-score based sampling Uniform random sampling Original DeiT SeiT SeiT (pixel fine-tuned) Storage (GB, log scale) Top 1 Accuracy (%)"}, "cited_paper_content": {"title": "Cutmix: Regularization Strategy To Train Strong Classifiers With Localizable Features", "abstract": "Regional dropout strategies have been proposed to enhance the performance of convolutional neural network classifiers. They have proved to be effective for guiding the model to attend on less discriminative parts of objects (e.g. leg as opposed to head of a person), thereby letting the network generalize better and have better object localization capabilities. On the other hand, current methods for regional dropout remove informative pixels on training images by overlaying a patch of either black pixels or random noise. Such removal is not desirable because it leads to information loss and inefficiency during training. We therefore propose the CutMix augmentation strategy: patches are cut and pasted among training images where the ground truth labels are also mixed proportionally to the area of the patches. By making efficient use of training pixels and retaining the regularization effect of regional dropout, CutMix consistently outperforms the state-of-the-art augmentation strategies on CIFAR and ImageNet classification tasks, as well as on the ImageNet weakly-supervised localization task. Moreover, unlike previous augmentation methods, our CutMix-trained ImageNet classifier, when used as a pretrained model, results in consistent performance gains in Pascal detection and MS-COCO image captioning benchmarks. We also show that CutMix improves the model robustness against input corruptions and its out-of-distribution detection performances. Source code and pretrained models are available at this https URL ."}, "keywords": ["strong augmentation"], "citation_intent": "method"}
{"citing_id": "2303.18027v2", "cited_id": "2003.05002", "section_title": "Clinical Natural Language Processing Beyond", "citation": "To mitigate this English-centric problem, some datasets only sample questions from native speakers of each langauge #REFR .", "text_before_citation": ["Though there are variants that use crosslingual supervision (e.g., #OTHEREFR , many of these polyglot models can benefit from joint training of different languages without any explicit supervision.", "We suspect that similar polyglot language modeling is happenning in LLMs, such as ChatGPT and GPT-4, which we tested on our IGAKU QA benchmark in Japanese, a language typologically distant from English.", "Open-Domain and Multilingual Question Answering Much prior work proposed datasets for open-domain QA for English and beyond #OTHEREFR .", "Several works pointed out the problem of translation-based question answering evaluations #OTHEREFR : questions raised mainly by English speakers can diverge from information needs from speakers of other languages.", "For instance, these translationbased benchmarks can overly represent Englishcentric topics, such as American politics, sports, and culture."], "text_after_citation": ["Consistent with such data creation methods, our IGAKU QA consists of problems that are written by native Japanese speakers to evaluate the qualifica-tions and skills for medical practice in the country."], "citing_paper_content": {"title": "Evaluating Gpt-4 And Chatgpt On Japanese Medical Licensing Examinations", "abstract": "As large language models (LLMs) gain popularity among speakers of diverse languages, we believe that it is crucial to benchmark them to better understand model behaviors, failures, and limitations in languages beyond English. In this work, we evaluate LLM APIs (Chat-GPT, GPT-3, and GPT-4) on the Japanese national medical licensing examinations from the past five years, including the current year. Our team comprises native Japanese-speaking NLP researchers and a practicing cardiologist based in Japan. Our experiments show that GPT-4 outperforms ChatGPT and GPT-3 and passes all six years of the exams, highlighting LLMs' potential in a language that is typologically distant from English. However, our evaluation also exposes critical limitations of the current LLM APIs. First, LLMs sometimes select prohibited choices (\u7981\u5fcc\u80a2) that should be strictly avoided in medical practice in Japan, such as suggesting euthanasia. Further, our analysis shows that the API costs are generally higher and the maximum context size is smaller for Japanese because of the way non-Latin scripts are currently tokenized in the pipeline. We release our benchmark as IGAKU QA as well as all model outputs and exam metadata. We hope that our results and benchmark will spur progress on more diverse applications of LLMs."}, "cited_paper_content": {"title": "Tydi Qa: A Benchmark For Information-Seeking Question Answering In Typologically Diverse Languages", "abstract": "Confidently making progress on multilingual modeling requires challenging, trustworthy evaluations. We present TyDi QA---a question answering dataset covering 11 typologically diverse languages with 204K question-answer pairs. The languages of TyDi QA are diverse with regard to their typology---the set of linguistic features each language expresses---such that we expect models performing well on this set to generalize across a large number of the world's languages. We present a quantitative analysis of the data quality and example-level qualitative linguistic analyses of observed language phenomena that would not be found in English-only corpora. To provide a realistic information-seeking task and avoid priming effects, questions are written by people who want to know the answer, but don't know the answer yet, and the data is collected directly in each language without the use of translation."}, "keywords": ["English-centric problem"], "citation_intent": "background"}
{"citing_id": "2303.01986v1", "cited_id": "2002.05709", "section_title": "The Impact Of Mini-Batch Size For Simclr And Barlowtwins", "citation": "This belief was confirmed by one of the Figure in #REFR that display a 7% accuracy gap on ImageNet between a model trained with a small (256) versus large batch size (8192).", "text_before_citation": ["2that SimCLR uses negative examples in the denominator term of its loss.", "It has been reasoned that the role of this term is to perform negative sampling #OTHEREFR and thus that is will only be effective when the number of samples i.e. mini-batch size is large. Equivalently, BarlowTwins (recall Eq.", "(3)) estimates correlation along the sample dimensions and thus also is expected to benefit from large mini-batch size for more accurate estimation.", "In this section, we propose to refute that large mini-batch sizes are required to successfully train JE-SSL with these two methods.", "The belief that many methods rely on large mini-batch size has been mentioned in many recent studies, e.g. #OTHEREFR ."], "text_after_citation": ["However, an important point that people often miss when reading #OTHEREFR is the critical influence of the optimizer and the choice of the learning rate when training with different batch size.", "In their appendix, #OTHEREFR show that when using a better optimization technique, the gap in performances between bigger and smaller batch size are significantly reduced.", "The impact of the downstream task One caveat of the batch size analysis in #OTHEREFR is that it concerns only the performances on ImageNet.", "Since one of the main motivation behind SSL is to learn a model whose representations can generalize to different tasks, we analyse the performances with different batch sizes across several downstream tasks: ImageNet-1K [19] , CIFAR10 #OTHEREFR , CLEVR #OTHEREFR , Eurosat #OTHEREFR , Inaturalist #OTHEREFR and Places #OTHEREFR .", "In Figure 2 , we plot the performances with SimCLR trained with several batch sizes."], "citing_paper_content": {"title": "Towards Democratizing Joint-Embedding Self-Supervised Learning", "abstract": "Joint Embedding Self-Supervised Learning (JE-SSL) has seen rapid developments in recent years, due to its promise to effectively leverage large unlabeled data. The development of JE-SSL methods was driven primarily by the search for ever increasing downstream classification accuracies, using huge computational resources, and typically built upon insights and intuitions inherited from a close parent JE-SSL method. This has led unwittingly to numerous preconceived ideas that carried over across methods e.g. that SimCLR requires very large mini batches to yield competitive accuracies; that strong and computationally slow data augmentations are required. In this work, we debunk several such ill-formed a priori ideas in the hope to unleash the full potential of JE-SSL free of unnecessary limitations. In fact, when carefully evaluating performances across different downstream tasks and properly optimizing hyper-parameters of the methods, we most often-if not always-see that these widespread misconceptions do not hold. For example we show that it is possible to train SimCLR to learn useful representations, while using a single image patch as negative example, and simple Gaussian noise as the only data augmentation for the positive pair. Along these lines, in the hope to democratize JE-SSL and to allow researchers to easily make more extensive evaluations of their methods, we introduce an optimized PyTorch library for SSL https: //github.com/facebookresearch/FFCV-SSL."}, "cited_paper_content": {"title": "A Simple Framework For Contrastive Learning Of Visual Representations", "abstract": "This paper presents SimCLR: a simple framework for contrastive learning of visual representations. We simplify recently proposed contrastive self-supervised learning algorithms without requiring specialized architectures or a memory bank. In order to understand what enables the contrastive prediction tasks to learn useful representations, we systematically study the major components of our framework. We show that (1) composition of data augmentations plays a critical role in defining effective predictive tasks, (2) introducing a learnable nonlinear transformation between the representation and the contrastive loss substantially improves the quality of the learned representations, and (3) contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. By combining these findings, we are able to considerably outperform previous methods for self-supervised and semi-supervised learning on ImageNet. A linear classifier trained on self-supervised representations learned by SimCLR achieves 76.5% top-1 accuracy, which is a 7% relative improvement over previous state-of-the-art, matching the performance of a supervised ResNet-50. When fine-tuned on only 1% of the labels, we achieve 85.8% top-5 accuracy, outperforming AlexNet with 100X fewer labels."}, "keywords": ["ImageNet"], "citation_intent": "result"}
{"citing_id": "2303.06338v1", "cited_id": "1811.10652", "section_title": "Ablation Studies", "citation": "From Table 3b , we can observe that: 1) Either ComPro use mask attention or segment embedding significantly outperformed SCT #REFR .", "text_before_citation": ["\"C-CS\", \"S-CS\", \"T-CS\", and \"Cap\" denote content-related control signal, structure-related control signal, tense control signal #OTHEREFR , and generated captions, respectively.", "Different colors show the corresponding control region sets and the associations between generated words and regions. responding visual guidance after removing masks.", "Instead, structure variables designed to learn global information are barely affected.", "Segment Embedding.", "To validate the effectiveness of mask attention and segment embedding 2 , we run several ablations with content-related control signals."], "text_after_citation": ["2) ComPro with mask attention can achieve the best performance, while the model with only segment embedding or both are not as effective and require more training sources.", "Prompt Length.", "We evaluated the performance of Com-Pro on structure control and content control by setting different structure prompt lengths M \u2208 {1, 10, 40, 80} and content prompt lengths N \u2208 {1, 5, 10, 20}, respectively.", "From Table 3c , we can observe that: The performance of ComPro keeps improving when we increase the prompt length.", "Then, the metrics reach the best scores and even slightly drop with longer prompt length."], "citing_paper_content": {"title": "Learning Combinatorial Prompts For Universal Controllable Image Captioning", "abstract": "Controllable Image Captioning (CIC)-generating natural language descriptions about images under the guidance of given control signals-is one of the most promising directions towards next-generation captioning systems. Till now, various kinds of control signals for CIC have been proposed, ranging from content-related control to structurerelated control. However, due to the format and target gaps of different control signals, all existing CIC works (or architectures) only focus on one certain control signal, and overlook the human-like combinatorial ability. By \"combinatorial\", we mean that our humans can easily meet multiple needs (or constraints) simultaneously when generating descriptions. To this end, we propose a novel prompt-based framework for CIC by learning Combinatorial Prompts, dubbed as ComPro. Specifically, we directly utilize a pretrained language model GPT-2 [36] as our language model, which can help to bridge the gap between different signalspecific CIC architectures. Then, we reformulate the CIC as a prompt-guide sentence generation problem, and propose a new lightweight prompt generation network to generate the combinatorial prompts for different kinds of control signals. For different control signals, we further design a new mask attention mechanism to realize the prompt-based CIC. Due to its simplicity, our ComPro can easily be extended to more complex combined control signals by concatenating these prompts. Extensive experiments on two prevalent CIC benchmarks have verified the effectiveness and efficiency of our ComPro on both single and combined control signals."}, "cited_paper_content": {"title": "Show, Control And Tell: A Framework For Generating Controllable And Grounded Captions", "abstract": "Current captioning approaches can describe images using black-box architectures whose behavior is hardly controllable and explainable from the exterior. As an image can be described in infinite ways depending on the goal and the context at hand, a higher degree of controllability is needed to apply captioning algorithms in complex scenarios. In this paper, we introduce a novel framework for image captioning which can generate diverse descriptions by allowing both grounding and controllability. Given a control signal in the form of a sequence or set of image regions, we generate the corresponding caption through a recurrent architecture which predicts textual chunks explicitly grounded on regions, following the constraints of the given control. Experiments are conducted on Flickr30k Entities and on COCO Entities, an extended version of COCO in which we add grounding annotations collected in a semi-automatic manner. Results demonstrate that our method achieves state of the art performances on controllable image captioning, in terms of caption quality and diversity. Code and annotations are publicly available at: https://github.com/aimagelab/show-control-and-tell."}, "keywords": ["mask attention"], "citation_intent": "result"}
{"citing_id": "2304.10439v1", "cited_id": "1909.06674", "section_title": "Related Work", "citation": "Raff #REFR directly compared the reported results of 255 AI papers without using the code from the original papers, and found out that 162 papers were at least 75% consistent with reported results while 93 were not.", "text_before_citation": ["#OTHEREFR proposed a hybrid classifier to automatically extract open-access datasets and software from scientific papers. Gunderson et al.", "#OTHEREFR investigated 400 AI papers, and found that none contain documentation of published experiments, methods, and data altogether.", "Although the work referenced above have highlighted the importance of including codes and data alongside published findings, studies were limited to meta-level indicators.", "There are a handful of studies that directly compare reproduced results of AI algorithms with published results. For example, Olorisade et al. #OTHEREFR Figure 2 : Workflow of our study.", "papers using the codes and datasets from the original papers, reporting inconsistent results with published findings."], "text_after_citation": ["Lack of transparency and reproducibility is particularly critical given the standard for AI papers to evaluate performance of proposed methods against baselines.", "Such problems have been found in AI research on machine learning analyses on clinical research #OTHEREFR and deep metric learning #OTHEREFR .", "Replication studies that provide side-by-side comparisons between AI papers addressing the same topic are rarely conducted."], "citing_paper_content": {"title": "A Study On Reproducibility And Replicability Of Table Structure Recognition Methods", "abstract": "Concerns about reproducibility in artificial intelligence (AI) have emerged, as researchers have reported unsuccessful attempts to directly reproduce published findings in the field. Replicability, the ability to affirm a finding using the same procedures on new data, has not been well studied. In this paper, we examine both reproducibility and replicability of a corpus of 16 papers on table structure recognition (TSR), an AI task aimed at identifying cell locations of tables in digital documents. We attempt to reproduce published results using codes and datasets provided by the original authors. We then examine replicability using a dataset similar to the original as well as a new dataset, GenTSR, consisting of 386 annotated tables extracted from scientific papers. Out of 16 papers studied, we reproduce results consistent with the original in only four. Two of the four papers are identified as replicable using the similar dataset under certain IoU values. No paper is identified as replicable using the new dataset. We offer observations on the causes of irreproducibility and irreplicability. All code and data are available on Codeocean at https://codeocean.com/capsule/6680116/tree."}, "cited_paper_content": {"title": "A Step Toward Quantifying Independently Reproducible Machine Learning Research", "abstract": "What makes a paper independently reproducible? Debates on reproducibility center around intuition or assumptions but lack empirical results. Our field focuses on releasing code, which is important, but is not sufficient for determining reproducibility. We take the first step toward a quantifiable answer by manually attempting to implement 255 papers published from 1984 until 2017, recording features of each paper, and performing statistical analysis of the results. For each paper, we did not look at the authors code, if released, in order to prevent bias toward discrepancies between code and paper."}, "keywords": ["255 AI papers"], "citation_intent": "result"}
{"citing_id": "2304.03159v1", "cited_id": "1910.07475", "section_title": "Iv. Experiments A. Experiment Setup", "citation": "Dataset We evaluate our model's performance on MLQA #REFR , a highly-parallel multilingual extractive QA dataset.", "text_before_citation": [], "text_after_citation": ["MLQA has QA samples in 7 languages including English, Arabic, German, Spanish, Hindi, Vietnamese and Simplified Chinese.", "This dataset has 12,738 extractive QA instances in English and over 5K in target language.", "9,019 instances are 4-way parallel, 2,930 are 3-way parallel and 789 are 2way parallel.", "MLQA is constructed from Wikipedia articles which have similar meaning in parallel languages.", "Labeling work is done by crowd-sourcing and translation is done by professional translators. In total, there are over 46,000 QA instances."], "citing_paper_content": {"title": "Bridging The Language Gap: Knowledge Injected Multilingual Question Answering", "abstract": "Question Answering (QA) is the task of automatically answering questions posed by humans in natural languages. There are different settings to answer a question, such as abstractive, extractive, boolean, and multiple-choice QA. As a popular topic in natural language processing tasks, extractive question answering task (extractive QA) has gained extensive attention in the past few years. With the continuous evolvement of the world, generalized cross-lingual transfer (G-XLT), where question and answer context are in different languages, poses some unique challenges over cross-lingual transfer (XLT), where question and answer context are in the same language. With the boost of corresponding development of related benchmarks, many works have been done to improve the performance of various language QA tasks. However, only a few works are dedicated to the G-XLT task. In this work, we propose a generalized crosslingual transfer framework to enhance the model's ability to understand different languages. Specifically, we first assemble triples from different languages to form multilingual knowledge. Since the lack of knowledge between different languages greatly limits models' reasoning ability, we further design a knowledge injection strategy via leveraging link prediction techniques to enrich the model storage of multilingual knowledge. In this way, we can profoundly exploit rich semantic knowledge. Experiment results on real-world datasets MLQA demonstrate that the proposed method can improve the performance by a large margin, outperforming the baseline method by 13.18%/12.00% F1/EM on average. Our code and models will be open-sourced."}, "cited_paper_content": {"title": "Mlqa: Evaluating Cross-Lingual Extractive Question Answering", "abstract": "Question answering (QA) models have shown rapid progress enabled by the availability of large, high-quality benchmark datasets. Such annotated datasets are difficult and costly to collect, and rarely exist in languages other than English, making training QA systems in other languages challenging. An alternative to building large monolingual training datasets is to develop cross-lingual systems which can transfer to a target language without requiring training data in that language. In order to develop such systems, it is crucial to invest in high quality multilingual evaluation benchmarks to measure progress. We present MLQA, a multi-way aligned extractive QA evaluation benchmark intended to spur research in this area. MLQA contains QA instances in 7 languages, namely English, Arabic, German, Spanish, Hindi, Vietnamese and Simplified Chinese. It consists of over 12K QA instances in English and 5K in each other language, with each QA instance being parallel between 4 languages on average. MLQA is built using a novel alignment context strategy on Wikipedia articles, and serves as a cross-lingual extension to existing extractive QA datasets. We evaluate current state-of-the-art cross-lingual representations on MLQA, and also provide machine-translation-based baselines. In all cases, transfer results are shown to be significantly behind training-language performance."}, "keywords": ["highly-parallel multilingual extractive"], "citation_intent": "method"}
{"citing_id": "2303.12304v1", "cited_id": "1409.0575", "section_title": "C. Implementation Details", "citation": "And it is initialized using weights which is trained on ImageNet #REFR , and the parameters of the first two layers are frozen throughout training.", "text_before_citation": ["Our approach is implemented under PyTorch 1.8.0 framework on a Intel(R) Xeon(R) Silver 4210R CPU(2.40GHz) along with a Nvidia Geforce RTX 3090GPU. The backbone is modified ResNet-50 #OTHEREFR ."], "text_after_citation": ["Because the shallow layers of the network are more generalized and high layers are more related to specific tasks.", "GOT-10k #OTHEREFR , COCO #OTHEREFR , ImageNet VID #OTHEREFR , and ImageNet DET #OTHEREFR provided the data for the training set.", "All of the images are cropped and scaled to 127\u00d7127 and 511\u00d7511 according to the ground truth provided in the dataset. We utilize corrective loss to train SiamTHN.", "The whole network is trained with 20 epcohs using Stochastic Gradient Descent (SGD) with a momentum of 0.9. Batch size is 28.", "Learning rate for the first 5 warm-up epochs varies from 0.001 to 0.005; for the following 15 epochs, it ranges from 0.005 to 0.00005."], "citing_paper_content": {"title": "Siamthn: Siamese Target Highlight Network For Visual Tracking", "abstract": "Siamese network based trackers develop rapidly in the field of visual object tracking in recent years. The majority of siamese network based trackers now in use treat each channel in the feature maps generated by the backbone network equally, making the similarity response map sensitive to background influence and hence challenging to focus on the target region. Additionally, there are no structural links between the classification and regression branches in these trackers, and the two branches are optimized separately during training. Therefore, there is a misalignment between the classification and regression branches, which results in less accurate tracking results. In this paper, a Target Highlight Module is proposed to help the generated similarity response maps to be more focused on the target region. To reduce the misalignment and produce more precise tracking results, we propose a corrective loss to train the model. The two branches of the model are jointly tuned with the use of corrective loss to produce more reliable prediction results. Experiments on 5 challenging benchmark datasets reveal that the method outperforms current models in terms of performance, and runs at 38 fps, proving its effectiveness and efficiency."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["ImageNet"], "citation_intent": "method"}
{"citing_id": "2303.16068v1", "cited_id": "1802.05814", "section_title": "Effect Of Multi-Objective Loss.", "citation": "From Figure 8 , we find that 1 \u2208 [0.5, 0.7] usually leads to a good balance and achieves superior performance, which is consistent with the findings in #REFR .", "text_before_citation": ["To analyze the influence of KL divergence, the sparsity, and variance regularizers in the multi-objective loss, we conduct experiments to compare the performance by changing their coefficients 1 , 2 , and 3 . The results w.r.t.", "the three coefficients on Yelp and Book are presented in Figure 8 .", "The results on Electronics and Book are similar so that we omit the one on Electronics. From Figure 8 , we have the following observations:", "\u2022 Increasing 1 of KL divergence is useful to improve the performance.", "Actually, as indicated by #OTHEREFR , a large 1 for KL divergence regulates the independence of latent factors to facilitate disentangled representations; while a small 1 helps the model to fit the user interactions better #OTHEREFR ."], "text_after_citation": ["\u2022 The sparsity regularizer is essential to improve the performance because Recall and NDCG drop significantly when 2 is changed from 0.5 to 0.", "Besides, 2 cannot be too large, which will limit the representation capability of and .", "\u2022 The decreased performance from 3 = 1 \u22124 to 3 = 0 justifies the usefulness of the variance regularizer, which balances the gradient optimization across multiple environments.", "Moreover, it should be noted that 3 is expected to be tuned in a small magnitude [0, 1 \u22123 ] because the stronger gradient regularization will inevitably disturb the normal optimization of parameters."], "citing_paper_content": {"title": "Causal Disentangled Recommendation Against User Preference Shifts", "abstract": "Recommender systems easily face the issue of user preference shifts. User representations will become outof-date and lead to inappropriate recommendations if user preference has shifted over time. To solve the issue, existing work focuses on learning robust representations or predicting the shifting pattern. There lacks a comprehensive view to discover the underlying reasons for user preference shifts. To understand the preference shift, we abstract a causal graph to describe the generation procedure of user interaction sequences. Assuming user preference is stable within a short period, we abstract the interaction sequence as a set of chronological environments. From the causal graph, we find that the changes of some unobserved factors (e.g., becoming pregnant) cause preference shifts between environments. Besides, the fine-grained user preference over categories sparsely affects the interactions with different items. Inspired by the causal graph, our key considerations to handle preference shifts lie in modeling the interaction generation procedure by: 1) capturing the preference shifts across environments for accurate preference prediction, and 2) disentangling the sparse influence from user preference to interactions for accurate effect estimation of preference. To this end, we propose a Causal Disentangled Recommendation (CDR) framework, which captures preference shifts via a temporal variational autoencoder and learns the sparse influence from multiple environments. Specifically, an encoder is adopted to infer the unobserved factors from user interactions while a decoder is to model the interaction generation process. Besides, we introduce two learnable matrices to disentangle the sparse influence from user preference to interactions. Lastly, we devise a multi-objective loss to optimize CDR. Extensive experiments on three datasets show the superiority of CDR in enhancing the generalization ability under user preference shifts."}, "cited_paper_content": {"title": "Variational Autoencoders For Collaborative Filtering", "abstract": "We extend variational autoencoders (VAEs) to collaborative filtering for implicit feedback. This non-linear probabilistic model enables us to go beyond the limited modeling capacity of linear factor models which still largely dominate collaborative filtering research. We introduce a generative model with multinomial likelihood and use Bayesian inference to learn this powerful generative model. Despite widespread use in language modeling and economics, the multinomial likelihood receives less attention in the recommender systems literature. We introduce a different regularization parameter for the learning objective, which proves to be crucial for achieving competitive performance. Remarkably, there is an efficient way to tune the parameter using annealing. The resulting model and learning algorithm has information-theoretic connections to maximum entropy discrimination and the information bottleneck principle. Empirically, we show that the proposed approach significantly outperforms several state-of-the-art baselines, including two recently-proposed neural network approaches, on several real-world datasets. We also provide extended experiments comparing the multinomial likelihood with other commonly used likelihood functions in the latent factor collaborative filtering literature and show favorable results. Finally, we identify the pros and cons of employing a principled Bayesian inference approach and characterize settings where it provides the most significant improvements."}, "keywords": ["findings", "superior performance"], "citation_intent": "result"}
{"citing_id": "2304.05260v1", "cited_id": "1602.05629", "section_title": "Ablations", "citation": "Fraction of Participating Clients One of the characteristics of FL is that data is massively distributed and nodes may have limited communication with the central server as nodes may be offline or have slow connections limiting their communication #REFR .", "text_before_citation": ["From Figure 6 we observe WSM has a more significant effect as \u03b1 decreases with the biggest margin of improvement over vanilla FedAvg occurring when \u03b1 = 0.01.", "As the data becomes increasingly homogeneous, the gap between cross entropy with and without WSM shrinks until the data is i.i.d. and the two methods are equivalent.", "From Figure 8 we see that by the time \u03b1 = 0.5 most of the clients possess all of the classes, albeit in very skewed proportions.", "This is also the point at which the performance of FedAvg and FedAvg+WSM become very close.", "While WSM continues to offer a performance increase over the entire range of \u03b1 except for \u03b1 = 100, we conclude WSM is most advantageous when clients labels distributions are imbalanced."], "text_after_citation": ["The direct result of the limited capacity for connection is that only a small number of client nodes may participate in updates at any given time.", "These ablations focus on the fraction of clients participating in an update round, for each experiment the fraction of participating clients is constant for all rounds of training and we explore fractions from 1%, where only one client participates in the update, up to 60%.", "For the fraction of participating clients p, we observe the largest performance gap between FedAvg+WSM and FedAvg when the number of participating clients is low.", "This feature is significant since as explained above, low client participation is a known feature of real world federated learning settings.", "We hypothesize the performance gap as a function of client fraction between the FedAvg and FedAvg+WSM is due to the larger impact of local client forgetting when we limit communication capability."], "citing_paper_content": {"title": "Re-Weighted Softmax Cross-Entropy To Control Forgetting In Federated Learning", "abstract": "In Federated Learning, a global model is learned by aggregating model updates computed at a set of independent client nodes, to reduce communication costs multiple gradient steps are performed at each node prior to aggregation. A key challenge in this setting is data heterogeneity across clients resulting in differing local objectives which can lead clients to overly minimize their own local objective, diverging from the global solution. We demonstrate that individual client models experience a catastrophic forgetting with respect to data from other clients and propose an efficient approach that modifies the cross-entropy objective on a per-client basis by re-weighting the softmax logits prior to computing the loss. This approach shields classes outside a client's label set from abrupt representation change and we empirically demonstrate it can alleviate client forgetting and provide consistent improvements to standard federated learning algorithms. Our method is particularly beneficial under the most challenging federated learning settings where data heterogeneity is high and client participation in each round is low."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["Participating Clients", "data"], "citation_intent": "background"}
{"citing_id": "2303.12892v1", "cited_id": "1912.05372", "section_title": "B. Language Models For Clinical Narratives", "citation": "FlauBERT #REFR : It is based on the original BERT architecture and was trained on a large corpus of the French text.", "text_before_citation": ["In this context, we will explore some of the most popular pre-trained BERT-based models for French language processing available from Huggingface.", "CamemBERT #OTHEREFR : This is a pre-trained Transformer-based language model designed explicitly for processing French text.", "It is based on the Roberta architecture and was trained on a large corpus of French text that was filtered and pre-processed to improve the data quality.", "Its pre-training objective is a masked language model, where some input tokens are masked, and the model is trained to predict the missing tokens.", "Overall, CamemBERT is a highly effective tool for processing French language text and can be fine-tuned for specific downstream tasks or used for transfer learning in multilingual settings."], "text_after_citation": ["It has been shown to perform strongly on several natural language processing tasks in French, including named entity recognition and sentiment analysis.", "It also performs well on tasks related to French morphosyntaxes, such as part-of-speech tagging and dependency parsing.", "It was trained using a masked language model objective, where a portion of the input tokens are masked, and the model is trained to predict the missing tokens.", "FlauBERT is a powerful language model for processing French text that can be fine-tuned for specific downstream tasks.", "FrALBERT #OTHEREFR is a Transformer-based language model designed explicitly for text classification tasks in French."], "citing_paper_content": {"title": "A Small-Scale Switch Transformer And Nlp-Based Model For Clinical Narratives Classification", "abstract": "In recent years, Transformer-based models such as the Switch Transformer have achieved remarkable results in natural language processing tasks. However, these models are often too complex and require extensive pre-training, which limits their effectiveness for small clinical text classification tasks with limited data. In this study, we propose a simplified Switch Transformer framework and train it from scratch on a small French clinical text classification dataset at CHU Sainte-Justine hospital. Our results demonstrate that the simplified small-scale Transformer models outperform pre-trained BERT-based models, including DistillBERT, CamemBERT, FlauBERT, and FrALBERT. Additionally, using a mixture of expert mechanisms from the Switch Transformer helps capture diverse patterns; hence, the proposed approach achieves better results than a conventional Transformer with the self-attention mechanism. Finally, our proposed framework achieves an accuracy of 87%, precision at 87%, and recall at 85%, compared to the third-best pre-trained BERT-based model, FlauBERT, which achieved an accuracy of 84%, precision at 84%, and recall at 84%. However, Switch Transformers have limitations, including a generalization gap and sharp minima. We compare it with a multi-layer perceptron neural network for small French clinical narratives classification and show that the latter outperforms all other models."}, "cited_paper_content": {"title": "Flaubert: Unsupervised Language Model Pre-Training For French", "abstract": "Language models have become a key step to achieve state-of-the art results in many different Natural Language Processing (NLP) tasks. Leveraging the huge amount of unlabeled texts nowadays available, they provide an efficient way to pre-train continuous word representations that can be fine-tuned for a downstream task, along with their contextualization at the sentence level. This has been widely demonstrated for English using contextualized representations (Dai and Le, 2015; Peters et al., 2018; Howard and Ruder, 2018; Radford et al., 2018; Devlin et al., 2019; Yang et al., 2019b). In this paper, we introduce and share FlauBERT, a model learned on a very large and heterogeneous French corpus. Models of different sizes are trained using the new CNRS (French National Centre for Scientific Research) Jean Zay supercomputer. We apply our French language models to diverse NLP tasks (text classification, paraphrasing, natural language inference, parsing, word sense disambiguation) and show that most of the time they outperform other pre-training approaches. Different versions of FlauBERT as well as a unified evaluation protocol for the downstream tasks, called FLUE (French Language Understanding Evaluation), are shared to the research community for further reproducible experiments in French NLP."}, "keywords": ["large corpus"], "citation_intent": "method"}
{"citing_id": "2304.03890v1", "cited_id": "2001.06139", "section_title": "Identify Best-Qualified High-Speed Lossy Compressor", "citation": "In comparison with FXR mode, ZFP's ABS mode features a much higher compression quality on scientific datasets (i.e., higher compression ratio with the same reconstructed data quality), as verified in #REFR .", "text_before_citation": ["As shown in the previous analysis, in addition to controlling the data distortion by error bounds, the compression throughput and compression ratio are two critical metrics to consider.", "According to the prior literature #OTHEREFR , ZFP and SZx exhibit much higher compression speed than other compressors, including SZ2 #OTHEREFR , SZ3 #OTHEREFR , FPZIP #OTHEREFR , and Auto-SZ #OTHEREFR .", "Therefore, we focus on ZFP and SZx in particular and select the best-qualified one for compression-enabled collective communication.", "In ZFP, users can set a fixed compression rate in the fixed-rate (FXR) mode or a fixed absolute error bound in the fixed-accuracy (ABS) mode. The two modes have their particular pros and cons.", "The advantage of the FXR is that it brings up substantial convenience in some use cases that require knowing the compressed data size in advance."], "text_after_citation": ["We briefly analyze the key reason why ZFP's ABS mode has higher compression quality than its FXR mode in the following text.", "ZFP divides the dataset into small data blocks during the compression.", "Unlike ZFP(ABS) which allows variable compressed size for each block, ZFP(FXR) strictly forces each block to have the same compression ratio, which may inevitably degrade the quality of reconstructed data in turn.", "Moreover, the FXR mode cannot control the error bound, which may cause fairly high compression errors on some data points unexpectedly.", "In comparison with ZFP, SZx is an ultra-fast error-bounded compressor that was designed driven by the fixed-accuracy mode."], "citing_paper_content": {"title": "C-Coll: Introducing Error-Bounded Lossy Compression Into Mpi Collectives", "abstract": "With the ever-increasing computing power of supercomputers and the growing scale of scientific applications, the efficiency of MPI collective communications turns out to be a critical bottleneck in large-scale distributed and parallel processing. Large message size in MPI collectives is a particularly big concern because it may significantly delay the overall parallel performance. To address this issue, prior research simply applies the off-the-shelf fix-rate lossy compressors in the MPI collectives, leading to suboptimal performance, limited generalizability, and unbounded errors. In this paper, we propose a novel solution, called C-Coll, which leverages errorbounded lossy compression to significantly reduce the message size, resulting in a substantial reduction in communication cost. The key contributions are threefold. (1) We develop two general, optimized lossy-compression-based frameworks for both types of MPI collectives (collective data movement as well as collective computation), based on their particular characteristics. Our framework not only reduces communication cost but also preserves data accuracy. (2) We customize an optimized version based on SZx, an ultra-fast error-bounded lossy compressor, which can meet the specific needs of collective communication. (3) We integrate C-Coll into multiple collectives, such as MPI_Allreduce, MPI_Scatter, and MPI_Bcast, and perform a comprehensive evaluation based on real-world scientific datasets. Experiments show that our solution outperforms the original MPI collectives as well as multiple baselines and related efforts by 3.5-9.7\u00d7."}, "cited_paper_content": {"title": "Fraz: A Generic High-Fidelity Fixed-Ratio Lossy Compression Framework For Scientific Floating-Point Data", "abstract": "With ever-increasing volumes of scientific floating-point data being produced by high-performance computing applications, significantly reducing scientific floating-point data size is critical, and error-controlled lossy compressors have been developed for years. None of the existing scientific floating-point lossy data compressors, however, support effective fixed-ratio lossy compression. Yet fixed-ratio lossy compression for scientific floating-point data not only compresses to the requested ratio but also respects a user-specified error bound with higher fidelity. In this paper, we present FRaZ: a generic fixed-ratio lossy compression framework respecting user-specified error constraints. The contribution is twofold. (1) We develop an efficient iterative approach to accurately determine the appropriate error settings for different lossy compressors based on target compression ratios. (2) We perform a thorough performance and accuracy evaluation for our proposed fixed-ratio compression framework with multiple state-of-the-art error-controlled lossy compressors, using several real-world scientific floating-point datasets from different domains. Experiments show that FRaZ effectively identifies the optimum error setting in the entire error setting space of any given lossy compressor. While fixed-ratio lossy compression is slower than fixed-error compression, it provides an important new lossy compression technique for users of very large scientific floating-point datasets."}, "keywords": ["(i.e., higher compression"], "citation_intent": "result"}
{"citing_id": "2303.17647v1", "cited_id": "1604.02878", "section_title": "Experiments Character Detection", "citation": "For images, we obtain the faces using MTCNN #REFR and resize them to 160\u00d7160. Table 3 shows the results of character detection.", "text_before_citation": ["We use the spaCy PoS tagger (https://spacy.io/api/tagger) to obtain the nouns in text and then identify the characters using WordNet #OTHEREFR ."], "text_after_citation": ["Evaluation Metric We use precision and recall to evaluate textual and visual character detection.", "For textual character detection, a predicted character phrase is considered as a correct character detection when the head word of the noun phrase is the same as the gold-standard one.", "For visual character detection, a predicted face region is considered as a correct character detection when the bounding box of the face is entirely inside of an annotated body bounding box.", "Analysis Overall, recall scores are higher than precision scores for both modalities.", "The precision of image-based character detection is only 40.5% and the gap with recall is around 28.6%."], "citing_paper_content": {"title": "Detecting And Grounding Important Characters In Visual Stories", "abstract": "Characters are essential to the plot of any story. Establishing the characters before writing a story can improve the clarity of the plot and the overall flow of the narrative. However, previous work on visual storytelling tends to focus on detecting objects in images and discovering relationships between them. In this approach, characters are not distinguished from other objects when they are fed into the generation pipeline. The result is a coherent sequence of events rather than a charactercentric story. In order to address this limitation, we introduce the VIST-Character dataset, which provides rich charactercentric annotations, including visual and textual co-reference chains and importance ratings for characters. Based on this dataset, we propose two new tasks: important character detection and character grounding in visual stories. For both tasks, we develop simple, unsupervised models based on distributional similarity and pre-trained vision-and-language models. Our new dataset, together with these models, can serve as the foundation for subsequent work on analysing and generating stories from a character-centric perspective."}, "cited_paper_content": {"title": "Joint Face Detection And Alignment Using Multitask Cascaded Convolutional Networks", "abstract": "Face detection and alignment in unconstrained environment are challenging due to various poses, illuminations, and occlusions. Recent studies show that deep learning approaches can achieve impressive performance on these two tasks. In this letter, we propose a deep cascaded multitask framework that exploits the inherent correlation between detection and alignment to boost up their performance. In particular, our framework leverages a cascaded architecture with three stages of carefully designed deep convolutional networks to predict face and landmark location in a coarse-to-fine manner. In addition, we propose a new online hard sample mining strategy that further improves the performance in practice. Our method achieves superior accuracy over the state-of-the-art techniques on the challenging face detection dataset and benchmark and WIDER FACE benchmarks for face detection, and annotated facial landmarks in the wild benchmark for face alignment, while keeps real-time performance."}, "keywords": ["character detection", "faces"], "citation_intent": "method"}
{"citing_id": "2303.02717v1", "cited_id": "1511.07247", "section_title": "Method Seen Unseen", "citation": "Pairs from the CambridgeLandmarks dataset are computed by encoding images with a pre-trained NetVLAD model #REFR , and fetching the nearest 40 neighbors based on cosine similarity.", "text_before_citation": ["At test time, images are rescaled, and the center crop is used for inference, with no further augmentations.", "Our model is optimized using Adam, with \u03b2 1 = 0.9, \u03b2 2 = 0.999 and = 10 \u221210 and a batch size of 8.", "We used an initial learning rate of \u03bb = 10 \u22124 and a weight decay of 10 \u22124 and trained for 30 (600) epochs for indoor (outdoor) localization. Our loss is initialized as in #OTHEREFR .", "The training pairs for the 7Scenes dataset are generated as in #OTHEREFR .", "We further extend this dataset with pairs from #OTHEREFR ."], "text_after_citation": ["At train time, we sample from the pool of neighbors to generate different pairs.", "We provide the training pairs, code, configuration files, and pre-trained models in order to support the reproducibility of our results.", "All models were trained on a single NVIDIA RTX A5000 24GB GPU. Implementation Details. We use EfficientNet-B0 #OTHEREFR as our convolutional backbone.", "We extract the feature maps F trans and F rot at resolutions of 14 \u00d7 14 \u00d7 112 and 28 \u00d7 28 \u00d7 40, respectively, as these endpoints were previously shown to provide useful features for camera pose regression #OTHEREFR (we reconfirm this choice in our ablation study).", "We set C h = 512 for the dimension of our projected paired feature maps and their respective positional encoding."], "citing_paper_content": {"title": "Learning To Localize In Unseen Scenes With Relative Pose Regressors", "abstract": "Relative pose regressors (RPRs) localize a camera by estimating its relative translation and rotation to a poselabelled reference. Unlike scene coordinate regression and absolute pose regression methods, which learn absolute scene parameters, RPRs can (theoretically) localize in unseen environments, since they only learn the residual pose between camera pairs. In practice, however, the performance of RPRs is significantly degraded in unseen scenes. In this work, we propose to aggregate paired feature maps into latent codes, instead of operating on global image descriptors, in order to improve the generalization of RPRs. We implement aggregation with concatenation, projection, and attention operations (Transformer Encoders) and learn to regress the relative pose parameters from the resulting latent codes. We further make use of a recently proposed continuous representation of rotation matrices, which alleviates the limitations of the commonly used quaternions. Compared to stateof-the-art RPRs, our model is shown to localize significantly better in unseen environments, across both indoor and outdoor benchmarks, while maintaining competitive performance in seen scenes. We validate our findings and architecture design through multiple ablations. Our code and pretrained models is publicly available 1 ."}, "cited_paper_content": {"title": "Netvlad: Cnn Architecture For Weakly Supervised Place Recognition", "abstract": "We tackle the problem of large scale visual place recognition, where the task is to quickly and accurately recognize the location of a given query photograph. We present the following four principal contributions. First, we develop a convolutional neural network (CNN) architecture that is trainable in an end-to-end manner directly for the place recognition task. The main component of this architecture, NetVLAD, is a new generalized VLAD layer, inspired by the \u201cVector of Locally Aggregated Descriptors\u201d image representation commonly used in image retrieval. The layer is readily pluggable into any CNN architecture and amenable to training via backpropagation. Second, we create a new weakly supervised ranking loss, which enables end-to-end learning of the architecture's parameters from images depicting the same places over time downloaded from Google Street View Time Machine. Third, we develop an efficient training procedure which can be applied on very large-scale weakly labelled tasks. Finally, we show that the proposed architecture and training procedure significantly outperform non-learnt image representations and off-the-shelf CNN descriptors on challenging place recognition and image retrieval benchmarks."}, "keywords": ["nearest 40 neighbors"], "citation_intent": "method"}
{"citing_id": "2304.09235v1", "cited_id": "1911.01686", "section_title": "5.", "citation": "Numerical results confirmed the bound's validity and showed that it is even sharper than the state-of-the-art result #REFR .", "text_before_citation": ["The main focus of this paper was on linear diffusive problems.", "In that setting, and when Assumptions 3.1 to 3.3 are satisfied, we adapted ParaOpt to work for both tracking and terminal-cost objectives, while having generic convergence bounds and a preconditioner that ensures good weak scalability.", "We summarize our main contributions in the linear diffusive setting.", "\u2022 We formulated an extension of ParaOpt to the setting of tracking problems.", "\u2022 We proved a generalized convergence bound for terminal-cost ParaOpt that is generic in the propagators used (as long as they satisfy Assumptions 3.1 to 3.3)."], "text_after_citation": ["\u2022 We proved a similar bound for the tracking setting and proved that the case of an exact fine and a one-step implicit-Euler coarse propagator guarantees ParaOpt does not diverge.", "\u2022 Thanks to the generic convergence bound, we were able to study a different type of implicit-Euler coarse propagator for terminal-cost problems and found it improves ParaOpt's convergence.", "\u2022 We proposed diagonalization-based preconditioners to improve the scaling of ParaOpt.", "Analytic results from the ParaDiag literature and numerical tests confirmed that ParaOpt is now a weakly scalable algorithm.", "For problems outside the linear diffusive category, ParaOpt is still a very promising method #OTHEREFR , although poorly understood."], "citing_paper_content": {"title": "Diagonalization-Based Preconditioners And Generalized Convergence Bounds For Paraopt", "abstract": "The ParaOpt algorithm was recently introduced as a time-parallel solver for optimalcontrol problems with a terminal-cost objective, and convergence results have been presented for the linear diffusive case with implicit-Euler time integrators. We reformulate ParaOpt for tracking problems and provide generalized convergence analyses for both objectives. We focus on linear diffusive equations and prove convergence bounds that are generic in the time integrators used. For large problem dimensions, ParaOpt's performance depends crucially on having a good preconditioner to solve the arising linear systems. For the case where ParaOpt's cheap, coarse-grained propagator is linear, we introduce diagonalization-based preconditioners, inspired by recent advances in the ParaDiag family of methods. These preconditioners not only lead to a weakly-scalable ParaOpt version, but are themselves invertible in parallel, making maximal use of available concurrency. They have proven convergence properties in the linear diffusive case that are generic in the time discretization used, similarly to our ParaOpt results. Numerical results confirm that the iteration count of the iterative solvers used for ParaOpt's linear systems becomes constant in the limit of an increasing processor count. The paper is accompanied by a sequential Matlab implementation."}, "cited_paper_content": {"title": "Paraopt: A Parareal Algorithm For Optimality Systems", "abstract": "The time parallel solution of optimality systems arising in PDE constraint optimization could be achieved by simply applying any time parallel algorithm, such as Parareal, to solve the forward and backward evolution problems arising in the optimization loop. We propose here a different strategy by devising directly a new time parallel algorithm, which we call ParaOpt, for the coupled forward and backward non-linear partial differential equations. ParaOpt is inspired by the Parareal algorithm for evolution equations, and thus is automatically a two-level method. We provide a detailed convergence analysis for the case of linear parabolic PDE constraints. We illustrate the performance of ParaOpt with numerical experiments both for linear and nonlinear optimality systems."}, "keywords": ["Numerical results"], "citation_intent": "result"}
{"citing_id": "2303.12363v1", "cited_id": "1511.04508", "section_title": "Introduction", "citation": "While our research shares some similarities with previous studies #REFR [26], we believe that our loss function can serve as a significant complement to address different scenarios.", "text_before_citation": ["But only if the training process is sufficiently, it can be beneficial to the test robustness.", "In this paper, we propose a distribution-restrained softmax loss for adversarial robustness.", "Based on adversarial attack tests, we find that pre-and post-attack softmax are highly correlated.", "According to the relationship, we come up with a loss function to suppress the distribution diversity of softmax.", "By testing of this new loss function, we suggest that our method can improve the model robustness."], "text_after_citation": ["Further comparative discussion will be conducted later."], "citing_paper_content": {"title": "Distribution-Restrained Softmax Loss For The Model Robustness", "abstract": "Recently, the robustness of deep learning models has received widespread attention, and various methods for improving model robustness have been proposed, including adversarial training, model architecture modification, design of loss functions, certified defenses, and so on. However, the principle of the robustness to attacks is still not fully understood, also the related research is still not sufficient. Here, we have identified a significant factor that affects the robustness of models: the distribution characteristics of softmax values for non-real label samples. We found that the results after an attack are highly correlated with the distribution characteristics, and thus we proposed a loss function to suppress the distribution diversity of softmax. A large number of experiments have shown that our method can improve robustness without significant time consumption."}, "cited_paper_content": {"title": "Distillation As A Defense To Adversarial Perturbations Against Deep Neural Networks", "abstract": "Deep learning algorithms have been shown to perform extremely well on many classical machine learning problems. However, recent studies have shown that deep learning, like other machine learning techniques, is vulnerable to adversarial samples: inputs crafted to force a deep neural network (DNN) to provide adversary-selected outputs. Such attacks can seriously undermine the security of the system supported by the DNN, sometimes with devastating consequences. For example, autonomous vehicles can be crashed, illicit or illegal content can bypass content filters, or biometric authentication systems can be manipulated to allow improper access. In this work, we introduce a defensive mechanism called defensive distillation to reduce the effectiveness of adversarial samples on DNNs. We analytically investigate the generalizability and robustness properties granted by the use of defensive distillation when training DNNs. We also empirically study the effectiveness of our defense mechanisms on two DNNs placed in adversarial settings. The study shows that defensive distillation can reduce effectiveness of sample creation from 95% to less than 0.5% on a studied DNN. Such dramatic gains can be explained by the fact that distillation leads gradients used in adversarial sample creation to be reduced by a factor of 10^30. We also find that distillation increases the average minimum number of features that need to be modified to create adversarial samples by about 800% on one of the DNNs we tested."}, "keywords": ["loss function"], "citation_intent": "result"}
{"citing_id": "2304.08983v1", "cited_id": "1202.6144", "section_title": "A. Related Work", "citation": "One of the main challenges for resilient state estimation is that many of the proposed algorithms are combinatorial in nature #REFR .", "text_before_citation": ["To solve the resilient state estimation problem, there have been techniques developed based on fault detection and diagnosis #OTHEREFR , #OTHEREFR , as in #OTHEREFR , #OTHEREFR ; methods motivated from compressed sensing #OTHEREFR , #OTHEREFR , as in #OTHEREFR ; and observer-based methods as in #OTHEREFR - #OTHEREFR .", "Extensions have been presented for decentralized design and distributed operation #OTHEREFR - #OTHEREFR , nonlinear systems #OTHEREFR - #OTHEREFR , and complexity analysis #OTHEREFR ."], "text_after_citation": ["Hence, the algorithms cannot solve the problem in polynomial time in general, and substantial amount of computational or storage resources may be required, especially when the number of sensors is large.", "For instance, to identify and isolate/exclude q unknown compromised sensors out of p sensors (q < p), methods in #OTHEREFR , #OTHEREFR construct not less than p q observers for the computation, and methods in #OTHEREFR , #OTHEREFR consider p q cases to search out the un-compromised sensors.", "Reducing the complexity of resilient state estimation consequently has been major interest.", "Many sufficient conditions that allow complexity reduction have been found, specifically for linear systems.", "For example, a condition for converting the problem into a convex optimization problem is introduced in #OTHEREFR , and an assumption for making use of gradient decent algorithms is found in #OTHEREFR ."], "citing_paper_content": {"title": "", "abstract": "A resilient state estimation scheme for uniformly observable nonlinear systems, based on a method for local identification of sensor attacks, is presented. The estimation problem is combinatorial in nature, and so many methods require substantial computational and storage resources as the number of sensors increases. To reduce the complexity, the proposed method performs the attack identification with local subsets of the measurements, not with the set of all measurements. A condition for nonlinear attack identification is introduced as a relaxed version of existing redundant observability condition. It is shown that an attack identification can be performed even when the state cannot be recovered from the measurements. As a result, although a portion of measurements are compromised, they can be locally identified and excluded from the state estimation, and thus the true state can be recovered. Simulation results demonstrate the effectiveness of the proposed scheme."}, "cited_paper_content": {"title": "Attack Detection And Identification In Cyber-Physical Systems", "abstract": "Cyber-physical systems integrate computation, communication, and physical capabilities to interact with the physical world and humans. Besides failures of components, cyber-physical systems are prone to malignant attacks, and specific analysis tools as well as monitoring mechanisms need to be developed to enforce system security and reliability. This paper proposes a unified framework to analyze the resilience of cyber-physical systems against attacks cast by an omniscient adversary. We model cyber-physical systems as linear descriptor systems, and attacks as exogenous unknown inputs. Despite its simplicity, our model captures various real-world cyber-physical systems, and it includes and generalizes many prototypical attacks, including stealth, (dynamic) false-data injection and replay attacks. First, we characterize fundamental limitations of static, dynamic, and active monitors for attack detection and identification. Second, we provide constructive algebraic conditions to cast undetectable and unidentifiable attacks. Third, by using the system interconnection structure, we describe graph-theoretic conditions for the existence of undetectable and unidentifiable attacks. Finally, we validate our findings through some illustrative examples with different cyber-physical systems, such as a municipal water supply network and two electrical power grids."}, "keywords": ["resilient state estimation"], "citation_intent": "background"}
{"citing_id": "2304.00012v1", "cited_id": "1907.07374", "section_title": "Discussion", "citation": "To address this gap, we will incorporate explainable AI techniques #REFR into our future work to provide a human-understandable interpretation of the predictions.", "text_before_citation": ["Our study showcases the potential of a tree-distillation multi-task learning framework from computational and clinical perspectives.", "This approach combines the benefits of a tree model and multi-task learning, resulting in competitive performance on transplant EHR data.", "The dependable outputs from CoD-MTL indicate its potential as a promising clinical decision support tool for physicians involved in organ transplant procedures, as demonstrated by our case study focusing on liver transplants.", "Additionally, we need to address some limitations and identify corresponding solutions for future improvement of our CoD-MTL framework.", "A critical constraint of our current model is its lack of interpretability, as the inability to explain predictions may impede its deployment in organ transplant scenarios."], "text_after_citation": ["Another limitation is the failure to consider fairness in our current framework.", "Equity is an ethical goal that clinical decision support systems should aim to achieve #OTHEREFR .", "Therefore, our future work will place a strong emphasis on integrating fairness constraints within our prediction framework to ensure it is suitable for multiple outcomes in organ transplants.", "The proposed framework has the potential to extend beyond our focus on organ transplants and can be applied to other medical fields that use multi-task learning.", "For instance, the prediction of complications and the length of stay in the ICU #OTHEREFR may benefit from our CoD-MTL framework, subject to future refinement."], "citing_paper_content": {"title": "Multi-Task Learning For Post-Transplant Cause Of Death Analysis: A Case Study On Liver Transplant", "abstract": "Organ transplant is the essential treatment method for some end-stage diseases, such as liver failure. Analyzing the post-transplant cause of death (CoD) after organ transplant provides a powerful tool for clinical decision making, including personalized treatment and organ allocation. However, traditional methods like Model for End-stage Liver Disease (MELD) score and conventional machine learning (ML) methods are limited in CoD analysis due to two major data and model-related challenges. To address this, we propose a novel framework called CoD-MTL leveraging multi-task learning to model the semantic relationships between various CoD prediction tasks jointly. Specifically, we develop a novel tree distillation strategy for multi-task learning, which combines the strength of both the tree model and multi-task learning. Experimental results are presented to show the precise and reliable CoD predictions of our framework. A case study is conducted to demonstrate the clinical importance of our method in the liver transplant."}, "cited_paper_content": {"title": "A Survey On Explainable Artificial Intelligence (Xai): Towards Medical Xai", "abstract": "Recently, artificial intelligence, especially machine learning has demonstrated remarkable performances in many tasks, from image processing to natural language processing, especially with the advent of deep learning. Along with research progress, machine learning has encroached into many different fields and disciplines. Some of them, such as the medical field, require high level of accountability, and thus transparency, which means we need to be able to explain machine decisions, predictions and justify their reliability. This requires greater interpretability, which often means we need to understand the mechanism underlying the algorithms. Unfortunately, the black-box nature of the deep learning is still unresolved, and many machine decisions are still poorly understood. We provide a review on interpretabilities suggested by different research works and categorize them, with the intention of providing alternative perspective that is hopefully more tractable for future adoption of interpretability standard. We explore further into interpretability in the medical field, illustrating the complexity of interpretability issue."}, "keywords": ["explainable AI techniques"], "citation_intent": "method"}
{"citing_id": "2303.10683v1", "cited_id": "1907.07432", "section_title": "Qudit Teleportation Through Quantum Switch", "citation": "Fidelity starts increasing after the threshold even as noise affecting the channel increases, thereby corroborating the results obtained in #REFR .", "text_before_citation": ["When p = q = 0.5 (like a 50% erasure channel), no quantum information can be sent through any classical trajectory traversing the channels A and B #OTHEREFR .", "Hence, achievable fidelity increases with the dimension of the quantum particle used for teleportation.", "Increased dimension refers to an enlarged Hilbert space and a larger Hilbert space offers the advantage of communicating with larger information capacity and higher noise resilience than traditional qubit-based quantum systems. It is evident from Fig.", "6 , average fidelityF QS decreases as p and q increase (i.e.", "increase in error probabilities), up to a certain threshold like p = q = 1/2 for the case of d = 2."], "text_after_citation": ["For the case when p = q \u2192 0.5 (the limiting condition of zero-capacity quantum channel), teleportation of qubit over a quantum switch offers a fidelity of approaching 0.5, i.e.F QS d=2 \u2192 0.5.", "While as we increase d to 5, fidelity approaches 0.65.", "If d = 10,F QS d=10 \u2192 0.78, d = 15, ifF QS d=15 \u2192 0.89, d = 20, and ifF QS d=20 \u2192 0.94.", "Therefore, entanglement distribution of a qudit over a quantum switch provides a remarkable gain in fidelity of the teleported qudit at Bob's side with increase in both noise plaguing the communication channel and dimensionality of the quantum particle.", "For any scenario of quantum teleportation, the EPR pair while being distributed between transmitter and receiver may lose its coherence, and become a mixed state due to interaction with the environment and noise introduced by the quantum channel."], "citing_paper_content": {"title": "Entanglement Distribution And Quantum Teleportation In Higher Dimension Over The Superposition Of Causal Orders Of Quantum Channels *", "abstract": "Multiple photonic degrees of freedom can be explored to generate high-dimensional quantum states; commonly referred to as 'qudits'. Qudits offer several advantages for quantum communications, including higher information capacity, noise resilience and data throughput, and lower information loss over different propagation mediums (free space, optical fibre, underwater) as compared to conventional qubits based communication system. However, qudits have been little exploited in literature, owing to their difficulty in transmission and detection. In this paper, for the first time, we develop and formulate the theoretical framework for transmission of classical information through entanglement distribution of qudits over two quantum channels in superposition of alternative causal order. For the first time we i) engineer quantum switch operation for 2-qudit systems and ii) formulate theoretical system model for entanglement distribution of qudits via quantum switch. Results show that entanglement distribution of a qudit provides a considerable gain in fidelity even with increase in noise."}, "cited_paper_content": {"title": "Quantum Switch For The Quantum Internet: Noiseless Communications Through Noisy Channels", "abstract": "Counter-intuitively, quantum mechanics enables quantum particles to propagate simultaneously among multiple space-time trajectories. Hence, a quantum information carrier can travel through different communication channels in a quantum superposition of different orders, so that the relative causal order of the communication channels becomes indefinite. This is realized by utilizing a quantum device known as quantum switch. In this paper, we investigate, from a communications engineering perspective, the use of the quantum switch within the quantum teleportation process, one of the key functionalities of the Quantum Internet. Specifically, a theoretical analysis is conducted to quantify the performance gain that can be achieved by employing a quantum switch for the entanglement distribution process within the quantum teleportation, with respect to the case of absence of the quantum switch. The analysis reveals that, by utilizing the quantum switch, the quantum teleportation is heralded as a noiseless communication process with a probability that, remarkably and counter-intuitively, increases with the noise levels affecting the communication channels considered in the indefinite-order combination."}, "keywords": ["channel increases"], "citation_intent": "result"}
{"citing_id": "2303.11675v1", "cited_id": "1909.12828", "section_title": "Related Work", "citation": "For instance, SPIN #REFR incorporates the optimization process into the regression framework, allowing better-optimized 3D results as supervision for the network.", "text_before_citation": ["Optimization-based approaches concentrate on the optimization algorithms to fit parametric models (e.g., SMPL #OTHEREFR ) based on 2D observations such as keypoints and silhouettes.", "These methods have been criticized for having long optimization processes and high sensitivity to initialization #OTHEREFR .", "Therefore, regression-based approaches have become a more popular paradigm in recent years, as they can directly and fast regress parameters from image features.", "HMR #OTHEREFR is a milestone work that introduces reprojection loss of keypoints as weak supervision, enabling training regression models on in-the-wild datasets.", "Recent work continues to explore more advanced weak supervision loss, such as introducing fine-grained correspondences #OTHEREFR , providing pseudo 3D ground truth #OTHEREFR , and improving reprojection loss #OTHEREFR ."], "text_after_citation": ["SPEC #OTHEREFR estimates the perspective camera from a single image and utilizes it to improve reprojection loss.", "CLIFF #OTHEREFR proposes considering the box information and computing the reprojection loss from the original images, which facilitates predicting global rotations.", "However, all these methods employ global representation for regression, making them sensitive to occlusion and partially visible humans.", "Occlusion handling.", "Handling occlusion is challenging yet crucial for human shape and pose estimation."], "citing_paper_content": {"title": "Bopr: Body-Aware Part Regressor For Human Shape And Pose Estimation", "abstract": "Figure 1: Illustration of our method's main superiority over existing approaches. When given a challenging input (a), PARE (b) [16] and CLIFF (c) [20] struggle to accurately estimate the pose of the forearm due to occlusion and depth ambiguity. Our method (d), on the other hand, tackles these challenges by exploiting body-part attention dependency in perpart estimation rather than relying on independent part regressors or global regression. This approach enables our method to more effectively reason about the relative spatial relationship between different body parts and the body as a whole, leading to improved occlusion handling and reduced depth ambiguity."}, "cited_paper_content": {"title": "Learning To Reconstruct 3D Human Pose And Shape Via Model-Fitting In The Loop", "abstract": "Model-based human pose estimation is currently approached through two different paradigms. Optimization-based methods fit a parametric body model to 2D observations in an iterative manner, leading to accurate image-model alignments, but are often slow and sensitive to the initialization. In contrast, regression-based methods, that use a deep network to directly estimate the model parameters from pixels, tend to provide reasonable, but not pixel accurate, results while requiring huge amounts of supervision. In this work, instead of investigating which approach is better, our key insight is that the two paradigms can form a strong collaboration. A reasonable, directly regressed estimate from the network can initialize the iterative optimization making the fitting faster and more accurate. Similarly, a pixel accurate fit from iterative optimization can act as strong supervision for the network. This is the core of our proposed approach SPIN (SMPL oPtimization IN the loop). The deep network initializes an iterative optimization routine that fits the body model to 2D joints within the training loop, and the fitted estimate is subsequently used to supervise the network. Our approach is self-improving by nature, since better network estimates can lead the optimization to better solutions, while more accurate optimization fits provide better supervision for the network. We demonstrate the effectiveness of our approach in different settings, where 3D ground truth is scarce, or not available, and we consistently outperform the state-of-the-art model-based pose estimation approaches by significant margins. The project website with videos, results, and code can be found at this https URL."}, "keywords": ["better-optimized 3D results"], "citation_intent": "method"}
{"citing_id": "2303.07615v1", "cited_id": "1512.03385", "section_title": "Models", "citation": "ResNet18: This model introduced by He et al #REFR , consists of 18-layers with skip connections and batch normalization.", "text_before_citation": ["Using our framework, we evaluate the following off-theshelf pretrained models:"], "text_after_citation": ["We refer to ResNet18 here as, the widely used version of this model that is trained on the Imagenet-1k classification challenge #OTHEREFR .", "ResNet50: Similar to ResNet18, but this model has 50 layers #OTHEREFR .", "We refer to ResNet50 here as specifically the model trained on the Imagenet-1k classification challenge #OTHEREFR .", "BiT-M-R50x1: This model also leverages the ResNet50 model architecture but is trained on the full Imagenet dataset containing 21k classes #OTHEREFR .", "The original full Imagenet-21k dataset might not be available #OTHEREFR , checkpoints for this model are publicly available."], "citing_paper_content": {"title": "Variation Of Gender Biases In Visual Recognition Models Before And After Finetuning", "abstract": "We introduce a framework to measure how biases change before and after fine-tuning a large scale visual recognition model for a downstream task. Deep learning models trained on increasing amounts of data are known to encode societal biases. Many computer vision systems today rely on models typically pretrained on large scale datasets. While bias mitigation techniques have been developed for tuning models for downstream tasks, it is currently unclear what are the effects of biases already encoded in a pretrained model. Our framework incorporates sets of canonical images representing individual and pairs of concepts to highlight changes in biases for an array of off-the-shelf pretrained models across model sizes, dataset sizes, and training objectives. Through our analyses, we find that (1) supervised models trained on datasets such as ImageNet-21k are more likely to retain their pretraining biases regardless of the target dataset compared to self-supervised models. We also find that (2) models finetuned on larger scale datasets are more likely to introduce new biased associations. Our results also suggest that (3) biases can transfer to finetuned models and the finetuning objective and dataset can impact the extent of transferred biases."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["batch normalization", "ResNet18"], "citation_intent": "method"}
{"citing_id": "2304.13799v1", "cited_id": "1502.05767", "section_title": "Introduction", "citation": "It takes advantage of automatic differentiation #REFR in formulating a physical loss in the loss function along with data loss.", "text_before_citation": ["Modelling solutions that leverage the strengths of theory-guided as well as data-driven models have the potential to reduce data needs, increase robustness, and effectively use theoretical and practical knowledge of the system.", "To investigate model architectures, balancing the strengths of both theory-based models and data-driven models, this work explores the application of physics-informed neural networks (PINNs) to a diesel internal combustion engine model for the purposes of simultaneous parameter and state estimation.", "The physical portion is based on the mean value model of a diesel engine with a variable geometry turbocharger (VGT), and exhaust gas recirculation (EGR) proposed by Wahlstr\u00f6m and Eriksson #OTHEREFR .", "Physics Inform Neural Networks (PINNs) [2-4, 7, 8] is a new method of training neural networks, which takes into account the physics of a problem while evaluating the parameters of the neural network.", "The method is suitable for both evaluation of the solution of PDF (forward problem) and the data-driven identification of parameters of PDF (inverse problem)."], "text_after_citation": ["In the present study, we formulate a PINN model for the data-driven identification of parameters and prediction of dynamics of system variables of a diesel engine.", "Specifically, we are interested in estimating unknown parameters and states when we know a few state variables from field data.", "The dynamics of the state variables of the mean value engine #OTHEREFR are described by first-order differential equations.", "We will utilize these equations in the formulation of the physics-informed loss function.", "The unknown parameters are considered trainable and updated in the training process along with the neural network parameters."], "citing_paper_content": {"title": "Physics-Informed Neural Networks For Predicting Gas Flow Dynamics And Unknown Parameters In Diesel Engines", "abstract": "This paper presents a physics-informed neural network (PINN) approach for monitoring the health of diesel engines. The aim is to evaluate the engine dynamics, identify unknown parameters in a \"mean value\" model, and anticipate maintenance requirements. The PINN model is applied to diesel engines with a variable-geometry turbocharger and exhaust gas recirculation, using measurement data of selected state variables. The results demonstrate the ability of the PINN model to predict simultaneously both unknown parameters and dynamics accurately with both clean and noisy data, and the importance of the self-adaptive weight in the loss function for faster convergence. The input data for these simulations are derived from actual engine running conditions, while the outputs are simulated data, making this a practical case study of PINN's ability to predict real-world dynamical systems. The mean value model of the diesel engine incorporates empirical formulae to represent certain states, but these formulae may not be generalizable to other engines. To address this, the study considers the use of deep neural networks (DNNs) in addition to the PINN model. The DNNs are trained using laboratory test data and are used to model the engine-specific empirical formulae in the mean value model, allowing for a more flexible and adaptive representation of the engine's states. In other words, the mean value model uses both the PINN model and the DNNs to represent the engine's states, with the PINN providing a physics-based understanding of the engine's overall dynamics and the DNNs offering a more engine-specific and adaptive representation of the empirical formulae. By combining these two approaches, the study aims to offer a comprehensive and versatile approach to monitoring the health and performance of diesel engines."}, "cited_paper_content": {"title": "Automatic Differentiation In Machine Learning: A Survey", "abstract": "Derivatives, mostly in the form of gradients and Hessians, are ubiquitous in machine learning. Automatic differentiation (AD), also called algorithmic differentiation or simply \u201cauto-diff\u201d, is a family of techniques similar to but more general than backpropagation for efficiently and accurately evaluating derivatives of numeric functions expressed as computer programs. AD is a small but established field with applications in areas including computational fluid dynamics, atmospheric sciences, and engineering design optimization. Until ::: very recently, the fields of machine learning and AD have largely been unaware of each other and, in some cases, have independently discovered each other\u2019s results. Despite its ::: relevance, general-purpose AD has been missing from the machine learning toolbox, a situation slowly changing with its ongoing adoption under the names \u201cdynamic computational ::: graphs\u201d and \u201cdifferentiable programming\u201d. We survey the intersection of AD and machine learning, cover applications where AD has direct relevance, and address the main imple- ::: mentation techniques. By precisely defining the main differentiation techniques and their interrelationships, we aim to bring clarity to the usage of the terms \u201cautodiff\u201d, \u201cautomatic differentiation\u201d, and \u201csymbolic differentiation\u201d as these are encountered more and more in machine learning settings."}, "keywords": ["automatic differentiation"], "citation_intent": "method"}
{"citing_id": "2304.06819v1", "cited_id": "1706.03762", "section_title": "I. Introduction", "citation": "Capturing dense multimodal interactions: Early fusion of histology and pathway tokens can be done with a Transformer that uses self-attention to capture pairwise similarities between all tokens #REFR .", "text_before_citation": ["Instead, we propose tokenizing genes according to established biological pathways #OTHEREFR - #OTHEREFR .", "Pathways are gene sets with known interactions that relate to specific cellular functions, such as the TGF-\u03b2 signaling cascade, which contributes to the epithelial-mesenchymal transition in breast cancer #OTHEREFR .", "Compared to coarse sets (e.g., N P = 6 #OTHEREFR ), pathwaybased gene grouping can yield hundreds to thousands of tokens that represent unique molecular processes (N P = 331 in our work), which we hypothesize are more suitable representations for multimodal fusion and alignment with fine-grained pathology features.", "In addition, as pathways represent unique cellular functions, they constitute appropriate basic reasoning units suitable for interpretability (see Fig. 1 ).", "2."], "text_after_citation": ["However, modeling pairwise interactions between large sets of histology patch tokens (e.g., N H = 15, 000) and pathway tokens (N P = 331) poses scalability challenges for fusion and alignment.", "Due to the quadratic complexity of the Transformer attention with respect to the number of tokens, modeling all possible interactions imposes substantial computational and memory requirements.", "To tackle this issue, we introduce a new unified, memory-efficient attention mechanism that can successfully model patch-to-pathway, pathway-to-patch, and pathway-to-pathway interactions.", "Modeling these three forms of interaction is achieved by the following: (1) designing the queries, keys, and values to share parameters across token types #OTHEREFR , #OTHEREFR , and (2) simplifying the attention layer to ignore patch-to-patch interactions, which we find through experimentation to be not as effective for survival analysis.", "Adapting Transformer attention in such a manner leads to significant performance increase over prior works, as well as new applications for model interpretability in multimodal CPATH, especially by understanding interactions between pathways and morphological patterns."], "citing_paper_content": {"title": "Modeling Dense Multimodal Interactions Between Biological Pathways And Histology For Survival Prediction", "abstract": "Integrating whole-slide images (WSIs) and bulk transcriptomics for predicting patient survival can improve our understanding of patient prognosis. However, this multimodal task is particularly challenging due to the different nature of these data: WSIs represent a very high-dimensional spatial description of a tumor, while bulk transcriptomics represent a global description of gene expression levels within that tumor. In this context, our work aims to address two key challenges: (1) how can we tokenize transcriptomics in a semantically meaningful and interpretable way?, and (2) how can we capture dense multimodal interactions between these two modalities? Specifically, we propose to learn biological pathway tokens from transcriptomics that can encode specific cellular functions. Together with histology patch tokens that encode the different morphological patterns in the WSI, we argue that they form appropriate reasoning units for downstream interpretability analyses. We propose fusing both modalities using a memory-efficient multimodal Transformer that can model interactions between pathway and histology patch tokens. Our proposed model, SURVPATH, achieves stateof-the-art performance when evaluated against both unimodal and multimodal baselines on five datasets from The Cancer Genome Atlas. Our interpretability framework identifies key multimodal prognostic factors, and, as such, can provide valuable insights into the interaction between genotype and phenotype, enabling a deeper understanding of the underlying biological mechanisms at play. We make our code public at https://github.com/ajv012/SurvPath."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["dense multimodal interactions", "Transformer"], "citation_intent": "method"}
{"citing_id": "2304.01192v1", "cited_id": "1712.07629", "section_title": "Proposed Modules", "citation": "We extract these features using SuperPoint #REFR , a single-pass convolutional neural network trained using homographic adaptation, a self-supervised consistency method.", "text_before_citation": ["FBE greedily selects the nearest frontier in the map to navigate to.", "Upon reaching that frontier, the process repeats with the selection of the next nearest frontier.", "This policy enables an efficient expansion of coverage with demonstrated effectiveness in both simulation and reality #OTHEREFR . Goal Instance Re-Identification.", "We employ a keypointbased re-identification method that performs binary classification conditioned on the goal image and egocentric image.", "First, we extract the pixel-wise (x, y) location of keypoints K * \u2208 R n\u00d72 in the goal image and their associated vector descriptions V * \u2208 R n\u00d7256 ."], "text_after_citation": ["We repeat this for the egocentric image, resulting in K t \u2208 R m\u00d72 and V t \u2208 R m\u00d7256 . Concretely:", "EQUATION", "We then compute correspondences between (K * , V * ) and", "(K t , V t ), resulting in a matched subset (K * \u2208 R w\u00d72 ,K t \u2208 R w\u00d72", ") such that the i th keypoint ofK * corresponds to the i th keypoint ofK t with w \u2264 min(n, m)."], "citing_paper_content": {"title": "Navigating To Objects Specified By Images", "abstract": "Images are a convenient way to specify which particular object instance an embodied agent should navigate to. Solving this task requires semantic visual reasoning and exploration of unknown environments. We present a system that can perform this task in both simulation and the real world. Our modular method solves sub-tasks of exploration, goal instance re-identification, goal localization, and local navigation. We re-identify the goal instance in egocentric vision using feature-matching and localize the goal instance by projecting matched features to a map. Each sub-task is solved using off-the-shelf components requiring zero finetuning. On the HM3D InstanceImageNav benchmark, this system outperforms a baseline end-to-end RL policy 7x and a state-of-the-art ImageNav model 2.3x (56% vs. 25% success). We deploy this system to a mobile robot platform and demonstrate effective real-world performance, achieving an 88% success rate across a home and an office environment."}, "cited_paper_content": {"title": "Superpoint: Self-Supervised Interest Point Detection And Description", "abstract": "This paper presents a self-supervised framework for training interest point detectors and descriptors suitable for a large number of multiple-view geometry problems in computer vision. As opposed to patch-based neural networks, our fully-convolutional model operates on full-sized images and jointly computes pixel-level interest point locations and associated descriptors in one forward pass. We introduce Homographic Adaptation, a multi-scale, multi-homography approach for boosting interest point detection repeatability and performing cross-domain adaptation (e.g., synthetic-to-real). Our model, when trained on the MS-COCO generic image dataset using Homographic Adaptation, is able to repeatedly detect a much richer set of interest points than the initial pre-adapted deep model and any other traditional corner detector. The final system gives rise to state-of-the-art homography estimation results on HPatches when compared to LIFT, SIFT and ORB."}, "keywords": ["self-supervised consistency method", "homographic adaptation"], "citation_intent": "method"}
{"citing_id": "2303.11052v2", "cited_id": "2002.05709", "section_title": "D.1. Additional Ablation Study", "citation": "As shown in Tab.5, the performance of our method increases as the number of negative pairs increases, with the best performance when the number reaches 512, which is consistent with the phenomenon in #REFR .", "text_before_citation": ["Ablation on the number of negative pairs.", "The number of negative pairs is shown to have a larger effect on the performance of contrastive learning as described in #OTHEREFR , and a Table 5 .", "Ablation study on the ScanNet dataset #OTHEREFR with respect to the number of negative pairs.", "larger number of negative pairs has a significant advantage over the smaller ones.", "Here in our method, we conduct ablation studies with different numbers of negative pairs to see the effect of the number of negative pairs on the model performance."], "text_after_citation": ["We also tried another way of multi-view consistency optimization, that is, we directly optimize the similarity between positive pairs p \u2212q + 2 , where p and q + are defined in Eq.7. The result is shown in row 1 of Tab.5.", "We can see that the above optimization method performs worse than our GeoContrast, which further reflects the importance of negative pairs.", "Ablation on proportion of real data.", "Here we present the SSIM and LPIPS under varying proportions of real and synthetic data as shown in Fig.10 .", "Similar to the curve of PSNR in Fig.6 , the performance of our method continues to improve as the proportion of real data increases, but the performance saturates when the proportion of real data reaches 40%."], "citing_paper_content": {"title": "Contranerf: Generalizable Neural Radiance Fields For Synthetic-To-Real Novel View Synthesis Via Contrastive Learning", "abstract": "Although many recent works have investigated generalizable NeRF-based novel view synthesis for unseen scenes, they seldom consider the synthetic-to-real generalization, which is desired in many practical applications. In this work, we first investigate the effects of synthetic data in synthetic-to-real novel view synthesis and surprisingly observe that models trained with synthetic data tend to produce sharper but less accurate volume densities. For pixels where the volume densities are correct, fine-grained details will be obtained. Otherwise, severe artifacts will be produced. To maintain the advantages of using synthetic data while avoiding its negative effects, we propose to introduce geometry-aware contrastive learning to learn multiview consistent features with geometric constraints. Meanwhile, we adopt cross-view attention to further enhance the geometry perception of features by querying features across input views. Experiments demonstrate that under the synthetic-to-real setting, our method can render images with higher quality and better fine-grained details, outperforming existing generalizable novel view synthesis methods in terms of PSNR, SSIM, and LPIPS. When trained on real data, our method also achieves state-of-the-art results. https://haoy945.github.io/contranerf/"}, "cited_paper_content": {"title": "A Simple Framework For Contrastive Learning Of Visual Representations", "abstract": "This paper presents SimCLR: a simple framework for contrastive learning of visual representations. We simplify recently proposed contrastive self-supervised learning algorithms without requiring specialized architectures or a memory bank. In order to understand what enables the contrastive prediction tasks to learn useful representations, we systematically study the major components of our framework. We show that (1) composition of data augmentations plays a critical role in defining effective predictive tasks, (2) introducing a learnable nonlinear transformation between the representation and the contrastive loss substantially improves the quality of the learned representations, and (3) contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. By combining these findings, we are able to considerably outperform previous methods for self-supervised and semi-supervised learning on ImageNet. A linear classifier trained on self-supervised representations learned by SimCLR achieves 76.5% top-1 accuracy, which is a 7% relative improvement over previous state-of-the-art, matching the performance of a supervised ResNet-50. When fine-tuned on only 1% of the labels, we achieve 85.8% top-5 accuracy, outperforming AlexNet with 100X fewer labels."}, "keywords": ["negative pairs increases", "best performance"], "citation_intent": "result"}
{"citing_id": "2304.03452v1", "cited_id": "1903.04154", "section_title": "Comparison With Fishergcn", "citation": "The other parameters remained the same in EigLearn as in Section 4.3.1, which is consistent with those in #REFR .", "text_before_citation": ["The second assessment focuses on the comparison between EigLearn and FisherGCN #OTHEREFR that is arguably more sophisticated than LanczosNet.", "In this comparison, we adopted the same settings for data splitting and training as in #OTHEREFR .", "In particular, we split the data into a training set with 20 samples per class, a validation set with 500 samples, and a testing set with 1000 samples.", "We ran experiments with 20 random splits and for each data split there were 10 different initializations per split.", "As thoroughly studied in #OTHEREFR , a random split is a less biased evaluation setting for GCN performance, and hence is preferable over the public fixed split."], "text_after_citation": ["Table 4 .2 reports a comparison among the baseline GCN, FisherGCN, and EigLearn using the graph filter matrices based on S SNA and S DW , respectively.", "We used both classification accuracy and loss on testing set as the evaluation metrics.", "For the baselines, we report the performances of both TensorFlow-based implementations of GCN and GCN T in #OTHEREFR and our PyTorch-based implementation.", "to implement and fits more naturally into the neural network architectures.", "In addition, EigLearnGCN is significantly faster than FisherGCN in that the additional cost associated with EigLearn is only a small fraction of the cost for training a regular GCN, while FisherGCN is several times slower than GCN as reported in #OTHEREFR ."], "citing_paper_content": {"title": "Many Thanks Go To Professor Keli Xiao, Professor Xiangmin Jiao, Dr. Dimitrios Katramatos, Professor Eden Figueroa And Professor Yao Ma. I Am Grateful For The Collaboration Opportunities And Benefit Significantly From Their Guidance On My Research Exploration And Technical Writing. I Am Thankful For Dr. Yufei Ren, Professor", "abstract": "The world has never been more connected, led by the information technology revolution in the past decades that has fundamentally changed the way people interact with each other using social networks. Consequently, enormous human activity data are collected from the business world and machine learning techniques are widely adopted to aid our decision processes. Despite of the success of machine learning in various application scenarios, there are still many questions that need to be well answered, such as optimizing machine learning outcomes when desired knowledge cannot be extracted from the available data. This naturally drives us to ponder if one can leverage some side information to populate the knowledge domain of their interest, such that the problems within that knowledge domain can be better tackled. In this work, such problems are investigated and practical solutions are proposed. To leverage machine learning in any decision-making process, one must Simplicity is the final achievement. Frederic Chopin v ACKNOWLEDGMENT First and foremost, I would like to express my sincere gratitude to my advisor, Prof. Dantong Yu. Throughout my Ph.D. journey, Prof. Yu patiently listened my thoughts on various research topics and gave me numerous pieces of priceless advice. He gave me substantial flexibility to explore and broaden my research interests, and shaped me into a better researcher and person over the years. I could not be more thankful for his support, guidance and encouragement to complete this journey. I would like to thank my committee members, Baruch M. Schieber, Ioannis Koutis, Yi Chen and Junmin Shi, for their kindness and insightful feedback during my journey on the dissertation. I would like to thank Prof. Schieber and Prof. Koutis for the generous office discussions and making complicated stuff simple and fun."}, "cited_paper_content": {"title": "Fisher-Bures Adversary Graph Convolutional Networks", "abstract": "In a graph convolutional network, we assume that the graph $G$ is generated with respect to some observation noise. We make small random perturbations $\\Delta{}G$ of the graph and try to improve generalization. Based on quantum information geometry, we can have quantitative measurements on the scale of $\\Delta{}G$. We try to maximize the intrinsic scale of the permutation with a small budget while minimizing the loss based on the perturbed $G+\\Delta{G}$. Our proposed model can consistently improve graph convolutional networks on semi-supervised node classification tasks with reasonable computational overhead. We present two different types of geometry on the manifold of graphs: one is for measuring the intrinsic change of a graph; the other is for measuring how such changes can affect externally a graph neural network. These new analytical tools will be useful in developing a good understanding of graph neural networks and fostering new techniques."}, "keywords": ["EigLearn"], "citation_intent": "result"}
{"citing_id": "2303.02827v1", "cited_id": "1910.00172", "section_title": "Introduction", "citation": "A fully discrete discontinuous Galerkin scheme is proposed in #REFR , where the time step as well as the mesh size are irrespective.", "text_before_citation": ["As one of the paradigms of nonlinear dynamical system, the Swift-Hohenberg equation has been widely applied in physics #OTHEREFR , materials science #OTHEREFR , dynamics of ecological systems #OTHEREFR , and many other nonlinear fields #OTHEREFR Bifurcation analysis #OTHEREFR and pattern selection #OTHEREFR of the equilibrium solutions for the Swift-Hohenberg equation are theoretically analyzed.", "Shi and Han #OTHEREFR obtained a reversible homoclinic solution approaching to a periodic solution of the Swift-Hohenberg equation, and the existence of nontrivial periodic solutions was proved by Lai and Zhang #OTHEREFR .", "In #OTHEREFR , the effects of rapid oscillations of the forcing term on the long-time behaviour of the solutions are studied.", "We also refer the reader to the recent article #OTHEREFR on nonstationary case and the references therein.", "Recently, the efficient numerical methods for the Swift-Hohenberg equation are considered since the analytically solutions cannot be generally obtained."], "text_after_citation": ["The scheme is proved uniquely solvable and unconditionally energy stable.", "A second-order energy-stable time-integration method that suppresses numerical instabilities is presented by Sarmiento et al.", "#OTHEREFR and a detailed proof of the unconditional energy stability is provided.", "Liu #OTHEREFR considered two linear, second-order and unconditionally energy stable schemes by linear invariant energy quadratization and modified scalar auxiliary variable approaches. Dehghan et al.", "#OTHEREFR combined the proper orthogonal decomposition approach and the local discontinuous Galerkin technique, and discussed the energy stability. Sun et al."], "citing_paper_content": {"title": "Energy Stable And L 2 Norm Convergent Bdf3 Scheme For The Swift-Hohenberg Equation", "abstract": "A fully discrete implicit scheme is proposed for the Swift-Hohenberg model, combining the third-order backward differentiation formula (BDF3) for the time discretization and the second-order finite difference scheme for the space discretization. Applying the Brouwer fixed-point theorem and the positive definiteness of the convolution coefficients of BDF3, the presented numerical algorithm is proved to be uniquely solvable and unconditionally energy stable, further, the numerical solution is shown to be bounded in the maximum norm. The proposed scheme is rigorously proved to be convergent in L 2 norm by the discrete orthogonal convolution (DOC) kernel, which transfer the four-level-solution form into the three-levelgradient form for the approximation of the temporal derivative. Consequently, the error estimate for the numerical solution is established by utilization of the discrete Gronwall inequality. Numerical examples in 2D and 3D cases are provided to support the theoretical results."}, "cited_paper_content": {"title": "Unconditionally Energy Stable Dg Schemes For The Swift-Hohenberg Equation", "abstract": "The Swift\u2013Hohenberg equation as a central nonlinear model in modern physics has a gradient flow structure. Here we introduce fully discrete discontinuous Galerkin (DG) schemes for a class of fourth order gradient flow problems, including the nonlinear Swift\u2013Hohenberg equation, to produce free-energy-decaying discrete solutions, irrespective of the time step and the mesh size. We exploit and extend the mixed DG method introduced in Liu and Yin (J Sci Comput 77:467\u2013501, 2018) for the spatial discretization, and the \u201cInvariant Energy Quadratization\u201d method for the time discretization. The resulting IEQ-DG algorithms are linear, thus they can be efficiently solved without resorting to any iteration method. We actually prove that these schemes are unconditionally energy stable. We present several numerical examples that support our theoretical results and illustrate the efficiency, accuracy and energy stability of our new algorithm. The numerical results on two dimensional pattern formation problems indicate that the method is able to deliver comparable patterns of high accuracy."}, "keywords": ["Galerkin scheme"], "citation_intent": "method"}
{"citing_id": "2304.08640v1", "cited_id": "1603.02754", "section_title": "Baselines", "citation": "We compare our TRAVEL model with two generic machine learning models and eleven state-of-the-art GNN models: (1) XGBoost: A scalable gradient boosting tree-based approach #REFR .", "text_before_citation": [], "text_after_citation": ["(2) MLP: A multi-layer perceptron is a classic feedforward artificial neural network.", "(3) GCN: Graph Convolutional Networks, which generalize CNNs from low-dimensional regular grids to graph data using neighborhood-based filters #OTHEREFR .", "(4) ChebNet: Chebyshev spectral graph convolution networks, which are spectral graph convolutional architectures with fast localized spectral filtering #OTHEREFR .", "(5) ARMANet: Graph neural networks with convolutional autoregressive moving average (ARMA) filters #OTHEREFR .", "(6) GraphSAGE: A general inductive framework for inductive representation learning on graphs #OTHEREFR ."], "citing_paper_content": {"title": "Tap: A Comprehensive Data Repository For Traffic Accident Prediction In Road Networks", "abstract": "Road safety is a major global public health concern. Effective traffic crash prediction can play a critical role in reducing road traffic accidents. However, Existing machine learning approaches tend to focus on predicting traffic accidents in isolation, without considering the potential relationships between different accident locations within road networks. To incorporate graph structure information, graph-based approaches such as Graph Neural Networks (GNNs) can be naturally applied. However, applying GNNs to the accident prediction problem faces challenges due to the lack of suitable graph-structured traffic accident datasets. To bridge this gap, we have constructed a real-world graph-based Traffic Accident Prediction (TAP) data repository, along with two representative tasks: accident occurrence prediction and accident severity prediction. With nationwide coverage, real-world network topology, and rich geospatial features, this data repository can be used for a variety of traffic-related tasks. We further comprehensively evaluate eleven state-of-the-art GNN variants and two non-graph-based machine learning methods using the created datasets. Significantly facilitated by the proposed data, we develop a novel Traffic Accident Vulnerability Estimation via Linkage (TRAVEL) model, which is designed to capture angular and directional information from road networks. We demonstrate that the proposed model consistently outperforms the baselines. The data and code are available on GitHub 1 ."}, "cited_paper_content": {"title": "Xgboost: A Scalable Tree Boosting System", "abstract": "Tree boosting is a highly effective and widely used machine learning method. In this paper, we describe a scalable end-to-end tree boosting system called XGBoost, which is used widely by data scientists to achieve state-of-the-art results on many machine learning challenges. We propose a novel sparsity-aware algorithm for sparse data and weighted quantile sketch for approximate tree learning. More importantly, we provide insights on cache access patterns, data compression and sharding to build a scalable tree boosting system. By combining these insights, XGBoost scales beyond billions of examples using far fewer resources than existing systems."}, "keywords": ["learning models", "1) XGBoost"], "citation_intent": "method"}
{"citing_id": "2303.00423v1", "cited_id": "1601.06473", "section_title": "Teaching Of Robots And Machines", "citation": "These include teaching assembly tasks #REFR , where the robot learns from human demonstration: First the robot observes the human, then it imitates the human's movements.", "text_before_citation": ["Robots are employed in a wide range of applications, especially in industry."], "text_after_citation": ["A different way of learning through user interaction is proposed in #OTHEREFR .", "In this approach, natural language is used to explain to the robot, which tasks it should perform. One such typical task is grasping objects.", "Solving this challenge is part of current research such as #OTHEREFR , where an object detection approach was used to learn good grasping poses.", "Data driven approaches, as stated in #OTHEREFR , are often addressed by providing training data in form of labeled examples, by trial-and-error, or through human demonstration.", "This means that communication between human and robot is also important here for a flexible learning progress without prior offline data generation."], "citing_paper_content": {"title": "Multiperspective Teaching Of Unknown Objects Via Shared-Gaze-Based Multimodal Human-Robot Interaction", "abstract": "For successful deployment of robots in multifaceted situations, an understanding of the robot for its environment is indispensable. With advancing performance of state-of-the-art object detectors, the capability of robots to detect objects within their interaction domain is also enhancing. However, it binds the robot to a few trained classes and prevents it from adapting to unfamiliar surroundings beyond predefined scenarios. In such scenarios, humans could assist robots amidst the overwhelming number of interaction entities and impart the requisite expertise by acting as teachers. We propose a novel pipeline that effectively harnesses human gaze and augmented reality in a human-robot collaboration context to teach a robot novel objects in its surrounding environment. By intertwining gaze (to guide the robot's attention to an object of interest) with augmented reality (to convey the respective class information) we enable the robot to quickly acquire a significant amount of automatically labeled training data on its own. Training in a transfer learning fashion, we demonstrate the robot's capability to detect recently learned objects and evaluate the influence of different machine learning models and learning procedures as well as the amount of training data involved. Our multimodal approach proves to be an efficient and natural way to teach the robot novel objects based on a few instances and allows it to detect classes for which no training dataset is available. In addition, we make our dataset publicly available to the research community, which consists of RGB and depth data, intrinsic and extrinsic camera parameters, along with regions of interest."}, "cited_paper_content": {"title": "Teaching Robots To Do Object Assembly Using Multi-Modal 3D Vision", "abstract": "The motivation of this paper is to develop an intelligent robot assembly system using multi-modal vision for next-generation industrial assembly. The system includes two phases where in the first phase human beings demonstrate assembly to robots and in the second phase robots detect objects, plan grasps, and assemble objects following human demonstration using AI searching. A notorious difficulty to implement such a system is the bad precision of 3D visual detection. This paper presents multi-modal approaches to overcome the difficulty: It uses AR markers in the teaching phase to detect human operation, and uses point clouds and geometric constraints in the robot execution phase to avoid unexpected occlusion and noises. The paper presents several experiments to examine the precision and correctness of the approaches. It demonstrates the applicability of the approaches by integrating them with graph model-based motion planning, and by executing the results on industrial robots in real-world scenarios."}, "keywords": ["robot", "assembly tasks"], "citation_intent": "background"}
{"citing_id": "2303.10030v1", "cited_id": "1902.11156", "section_title": "Nuclear Norm Minimization", "citation": "Moreover, the instability result is not expected to hold if one picks a generic isometry B, thus randomizing over both B and C #REFR .", "text_before_citation": ["1.X is feasible: A(X) \u2212 y 2 \u2264 \u03c4 ; 2.X is preferred to X 0 : X * \u2264 X 0 * ;", "3.X is far from the true solution in Frobenius norm:", "X \u2212 X 0 F \u2265 \u03c4 C 4 KN L . (2.7)", "Theorem 2.4 suggests that the dimension factor in the recovery guarantees (2.6) is not merely a proof artifact.", "AlthoughX constructed in the proof of Theorem 2.4 is preferred to the true solution, it is not a minimizer of the SDP (2.5), see Remark 3.4. in #OTHEREFR ."], "text_after_citation": ["Still, Theorem 2.4 provides a lower bound for possible recovery guarantees which hold for all preferred solutions.", "However, Theorem 2.4 gives no information on the noise level \u03c4 0 from which the instability is to be expected.", "In particular, it does not exclude the possibility that for larger noise levels \u03c4 , stronger recovery guarantees can be shown.", "Indeed, #OTHEREFR made a first step in this direction by showing near-optimal error scaling whenever the adversarial noise magnitude is not too small:", "Theorem 2.5 ([KS20]). Let \u03c9 > 0, \u00b5 \u2265 1."], "citing_paper_content": {"title": "How Robust Is Randomized Blind Deconvolution Via Nuclear Norm Minimization Against Adversarial Noise?", "abstract": "In this paper, we study the problem of recovering two unknown signals from their convolution, which is commonly referred to as blind deconvolution. Reformulation of blind deconvolution as a low-rank recovery problem has led to multiple theoretical recovery guarantees in the past decade due to the success of the nuclear norm minimization heuristic. In particular, in the absence of noise, exact recovery has been established for sufficiently incoherent signals contained in lower-dimensional subspaces. However, if the convolution is corrupted by additive bounded noise, the stability of the recovery problem remains much less understood. In particular, existing reconstruction bounds involve large dimension factors and therefore fail to explain the empirical evidence for dimension-independent robustness of nuclear norm minimization. Recently, theoretical evidence has emerged for ill-posed behavior of low-rank matrix recovery for sufficiently small noise levels. In this work, we develop improved recovery guarantees for blind deconvolution with adversarial noise which exhibit square-root scaling in the noise level. Hence, our results are consistent with existing counterexamples which speak against linear scaling in the noise level as demonstrated for related low-rank matrix recovery problems."}, "cited_paper_content": {"title": "On The Convex Geometry Of Blind Deconvolution And Matrix Completion", "abstract": "Low-rank matrix recovery from structured measurements has been a topic of intense study in the last decade and many important problems like matrix completion and blind deconvolution have been formulated in this framework. An important benchmark method to solve these problems is to minimize the nuclear norm, a convex proxy for the rank. A common approach to establish recovery guarantees for this convex program relies on the construction of a so-called approximate dual certificate. However, this approach provides only limited insight in various respects. Most prominently, the noise bounds exhibit seemingly suboptimal dimension factors. In this paper we take a novel, more geometric viewpoint to analyze both the matrix completion and the blind deconvolution scenario. We find that for both these applications the dimension factors in the noise bounds are not an artifact of the proof, but the problems are intrinsically badly conditioned. We show, however, that bad conditioning only arises for very small noise levels: Under mild assumptions that include many realistic noise levels we derive near-optimal error estimates for blind deconvolution under adversarial noise."}, "keywords": ["instability result", "generic isometry B"], "citation_intent": "background"}
{"citing_id": "2303.07160v1", "cited_id": "2002.10400", "section_title": "Summary Of Our Contributions", "citation": "Our lower bounds are obtained for without-replacement SGD with constant step size, which is also the case in other existing results in the literature Shamir, 2020, 2021; #REFR .", "text_before_citation": ["Towards a complete understanding of SGD-RR and permutation-based SGD in general, we seek to close the existing gaps outlined above by developing tighter lower bounds with matching upper bounds.", "We present results under two different kinds of algorithm settings: Section 3 contains lower bounds obtained from SGD-RR, and Section 4 presents lower bounds that are applicable to arbitrary permutation-based SGD algorithms."], "text_after_citation": ["While all lower bounds proved in the aforementioned papers are only applicable to the final iterate of the algorithm, many of our results in this paper apply to arbitrary weighted average of end-of-epoch iterates, which can be used to show tightness of matching upper bounds that employ iterate averaging.", "Our main contributions are as follows.", "Here we include \u03ba = \u0398 (1/\u00b5) in the convergence rates to better describe the results.", "For a complete summary of our results, please refer to Table 1.", "\u2022 Theorem 1 derives a lower bound of rate \u2126 \u03ba 2 nK 2 on the strongly convex case for the final iterate, which matches the best-known corresponding upper bound\u00d5 \u03ba 3 nK 2 up to a factor of \u03ba."], "citing_paper_content": {"title": "Tighter Lower Bounds For Shuffling Sgd: Random Permutations And Beyond", "abstract": "We study convergence lower bounds of without-replacement stochastic gradient descent (SGD) for solving smooth (strongly-)convex finite-sum minimization problems. Unlike most existing results focusing on final iterate lower bounds in terms of the number of components n and the number of epochs K, we seek bounds for arbitrary weighted average iterates that are tight in all factors including the condition number \u03ba. For SGD with Random Reshuffling, we present lower bounds that have tighter \u03ba dependencies than existing bounds. Our results are the first to perfectly close the gap between lower and upper bounds for weighted average iterates in both strongly-convex and convex cases. We also prove weighted average iterate lower bounds for arbitrary permutation-based SGD, which apply to all variants that carefully choose the best permutation. Our bounds improve the existing bounds in factors of n and \u03ba and thereby match the upper bounds shown for a recently proposed algorithm called GraB. Recent works provide evidence of designing algorithms that converge faster than SGD-RR. Concretely, Rajput et al. (2022) introduce a permutation-based SGD algorithm called FlipFlop and prove that it can outperform SGD-RR for quadratic objectives. The authors also propose a lower bound applicable to arbitrary permutation-based SGD, by proving that no algorithm can converge faster than \u2126"}, "cited_paper_content": {"title": "Closing The Convergence Gap Of Sgd Without Replacement", "abstract": "Stochastic gradient descent without replacement sampling is widely used in practice for model training. However, the vast majority of SGD analyses assumes data sampled with replacement, and when the function minimized is strongly convex, an $\\mathcal{O}\\left(\\frac{1}{T}\\right)$ rate can be established when SGD is run for $T$ iterations. A recent line of breakthrough work on SGD without replacement (SGDo) established an $\\mathcal{O}\\left(\\frac{n}{T^2}\\right)$ convergence rate when the function minimized is strongly convex and is a sum of $n$ smooth functions, and an $\\mathcal{O}\\left(\\frac{1}{T^2}+\\frac{n^3}{T^3}\\right)$ rate for sums of quadratics. On the other hand, the tightest known lower bound postulates an $\\Omega\\left(\\frac{1}{T^2}+\\frac{n^2}{T^3}\\right)$ rate, leaving open the possibility of better SGDo convergence rates in the general case. In this paper, we close this gap and show that SGD without replacement achieves a rate of $\\mathcal{O}\\left(\\frac{1}{T^2}+\\frac{n^2}{T^3}\\right)$ when the sum of the functions is a quadratic, and offer a new lower bound of $\\Omega\\left(\\frac{n}{T^2}\\right)$ for strongly convex functions that are sums of smooth functions."}, "keywords": ["lower bounds"], "citation_intent": "result"}
{"citing_id": "2303.14888v1", "cited_id": "1812.00324", "section_title": "Iv. Experiments", "citation": "The CrowdPose dataset #REFR is a challenging benchmark with the goal to evaluate the robustness of methods in crowded scenes.", "text_before_citation": ["A. Implementation Details 1) Dataset: COCO.", "The COCO dataset #OTHEREFR is a popular multi-person pose estimation benchmark containing more than 200, 000 images and 250, 000 person instances labeling with 17 keypoints.", "The COCO dataset is divided into three parts: the train set includes 57K images, the val set contains 5K images, and the test-dev set consists of 20K images.", "However, it contains rarely crowded scenarios, which we use to validate the generalization of our methods in unobstructed scenarios.", "CrowdPose."], "text_after_citation": ["The CrowdPose dataset contains 20K images and 80K persons labeled with 14 keypoints.", "It contains 10K, 2K and 8K images for train, val and test set.", "Following #OTHEREFR , we train our model on the train and validation splits combined, and report the final performance on the test set.", "2) Evaluation metrics: We use the standard evaluation metric Object Keypoint Similarity (OKS) to evaluate our models, which is defined as:", "EQUATION"], "citing_paper_content": {"title": "Global Relation Modeling And Refinement For Bottom-Up Human Pose Estimation", "abstract": "In this paper, we concern on the bottom-up paradigm in multi-person pose estimation (MPPE). Most previous bottom-up methods try to consider the relation of instances to identify different body parts during the post processing, while ignoring to model the relation among instances or environment in the feature learning process. In addition, most existing works adopt the operations of upsampling and downsampling. During the sampling process, there will be a problem of misalignment with the source features, resulting in deviations in the keypoint features learned by the model. To overcome the above limitations, we propose a convolutional neural network for bottom-up human pose estimation. It invovles two basic modules: (i) Global Relation Modeling (GRM) module globally learns relation (e.g., environment context, instance interactive information) among region of image by fusing multiple stages features in the feature learning process. It combines with the spatial-channel attention mechanism, which focuses on achieving adaptability in spatial and channel dimensions. (ii) Multi-branch Feature Align (MFA) module aggregates features from multiple branches to align fused feature and obtain refined local keypoint representation. Our model has the ability to focus on different granularity from local to global regions, which significantly boosts the performance of the multi-person pose estimation. Our results on the COCO and CrowdPose datasets demonstrate that it is an efficient framework for multi-person pose estimation."}, "cited_paper_content": {"title": "Crowdpose: Efficient Crowded Scenes Pose Estimation And A New Benchmark", "abstract": "Multi-person pose estimation is fundamental to many computer vision tasks and has made significant progress in recent years. However, few previous methods explored the problem of pose estimation in crowded scenes while it remains challenging and inevitable in many scenarios. Moreover, current benchmarks cannot provide an appropriate evaluation for such cases. In this paper, we propose a novel and efficient method to tackle the problem of pose estimation in the crowd and a new dataset to better evaluate algorithms. Our model consists of two key components: joint-candidate single person pose estimation (SPPE) and global maximum joints association. With multi-peak prediction for each joint and global association using the graph model, our method is robust to inevitable interference in crowded scenes and very efficient in inference. The proposed method surpasses the state-of-the-art methods on CrowdPose dataset by 5.2 mAP and results on MSCOCO dataset demonstrate the generalization ability of our method."}, "keywords": ["CrowdPose dataset"], "citation_intent": "background"}
{"citing_id": "2303.10802v1", "cited_id": "2003.02752", "section_title": "Introduction", "citation": "Nonetheless, we empirically note that separating clean, but hard to classify samples from noisy label samples still remains challenging for these sample selection processes #REFR , particularly for problems of large noise rates.", "text_before_citation": ["proaches incorporate robust loss functions #OTHEREFR , designed specifically to operate effectively with either clean or noisy labels, as well as probabilistic modeling approaches that model the data generation process #OTHEREFR .", "Furthermore, training regularization #OTHEREFR imposes a penalty term on the loss function during training, thereby reducing overfitting and ameliorating generalization.", "Various techniques integrate sample selection strategies as a key algorithmic step #OTHEREFR , allowing for the identification and segregation of noisy and clean labels.", "One popular criterion for this sample selection process is the loss value between the prediction of the trained classifier and its label, whereby it is typically assumed that the noisy data exhibits a large loss #OTHEREFR or greater magnitude of the gradient during training #OTHEREFR .", "Additionally, feature-based sample selection techniques based on the similarity to principal components of feature representations #OTHEREFR or K nearest neighbor (KNN) classification in the feature space #OTHEREFR have also been considered for sample selection criteria."], "text_after_citation": ["The use of peer classifiers for noisy label learning problems have been investigated for enforcing consistency in the training of the classifiers #OTHEREFR , but not for selecting clean and noisy label samples.", "We argue in this paper that the prediction agreement between peer classifiers is more effective to select clean and noisy label samples than previous approaches because intuitively, such agreement is unlikely to happen, except when the classifiers agree on the clean label.", "In this paper, we propose a new sample selection cri-terion based on the predictive probability agreement between peer classifiers, which has been shown in computational linguistics to be a sample reliability measure #OTHEREFR .", "In light of implementation, we train three classifiers simultaneously by using the agreement between two classifiers to select samples to train the remaining classifier, as shown in Fig. 1 .", "This sample selection relies on a thresholding algorithm #OTHEREFR that distinguishes samples based on how high the agreement between the peer classification predictions is."], "citing_paper_content": {"title": "Pass: Peer-Agreement Based Sample Selection For Training With Noisy Labels", "abstract": "Noisy labels present a significant challenge in deep learning because models are prone to overfitting. This problem has driven the development of sophisticated techniques to address the issue, with one critical component being the selection of clean and noisy label samples. Selecting noisy label samples is commonly based on the small-loss hypothesis or on feature-based sampling, but we present empirical evidence that shows that both strategies struggle to differentiate between noisy label and hard samples, resulting in relatively large proportions of samples falsely selected as clean. To address this limitation, we propose a novel peer-agreement based sample selection (PASS). An automated thresholding technique is then applied to the agreement score to select clean and noisy label samples. PASS is designed to be easily integrated into existing noisy label robust frameworks, and it involves training a set of classifiers in a round-robin fashion, with peer models used for sample selection. In the experiments, we integrate our PASS with several state-of-the-art (SOTA) models, including InstanceGM, DivideMix, SSR, FaMUS, AugDesc, and C2D, and evaluate their effectiveness on several noisy label benchmark datasets, such as CIFAR-100, CIFAR-N, Animal-10N, Red Mini-Imagenet, Clothing1M, Mini-Webvision, and Imagenet. Our results demonstrate that our new sample selection approach improves the existing SOTA results of algorithms."}, "cited_paper_content": {"title": "Combating Noisy Labels By Agreement: A Joint Training Method With Co-Regularization", "abstract": "Deep Learning with noisy labels is a practically challenging problem in weakly-supervised learning. The state-of-the-art approaches\"Decoupling\"and\"Co-teaching+\"claim that the\"disagreement\"strategy is crucial for alleviating the problem of learning with noisy labels. In this paper, we start from a different perspective and propose a robust learning paradigm called JoCoR, which aims to reduce the diversity of two networks during training. Specifically, we first use two networks to make predictions on the same mini-batch data and calculate a joint loss with Co-Regularization for each training example. Then we select small-loss examples to update the parameters of both two networks simultaneously. Trained by the joint loss, these two networks would be more and more similar due to the effect of Co-Regularization. Extensive experimental results on corrupted data from benchmark datasets including MNIST, CIFAR-10, CIFAR-100 and Clothing1M demonstrate that JoCoR is superior to many state-of-the-art approaches for learning with noisy labels."}, "keywords": ["noisy label samples"], "citation_intent": "background"}
{"citing_id": "2303.10703v2", "cited_id": "1512.02325", "section_title": "Related Work 2.1. Object Detection", "citation": "Another representative one-stage method is the SSD algorithm #REFR , which focuses on improving the accuracy and recall rate.", "text_before_citation": ["From left to right, the images are originally from MGD #OTHEREFR , USRT #OTHEREFR , and UCF #OTHEREFR datasets, respectively.", "one-stage detection algorithm.", "It considered detection a regression task and used an improved GoogLeNet as the backbone.", "It is later extended to YOLOv2 #OTHEREFR by drawing inspiration from Faster R-CNN and the anchor mechanism.", "A further improved version, YOLOv3 #OTHEREFR , uses better backbones to enhance performance."], "text_after_citation": ["Later, the focal loss is introduced #OTHEREFR to solve the mismatch between positive and negative samples.", "Transformer-based Methods.", "Transformers #OTHEREFR have been the de-facto choice of architecture in NLP.", "Since the advent of Vision Transformer #OTHEREFR , they have achieved remarkable results in computer vision tasks such as object detection.", "The first representative work in this category is DETR #OTHEREFR , which eliminates the intermediate steps like nonmaximal suppression and anchor generation in CNN-based detectors."], "citing_paper_content": {"title": "Cctv-Gun: Benchmarking Handgun Detection In Cctv Images", "abstract": "Gun violence is a critical security problem, and it is imperative for the computer vision community to develop effective gun detection algorithms for real-world scenarios, particularly in Closed Circuit Television (CCTV) surveillance data. Despite significant progress in visual object detection, detecting guns in real-world CCTV images remains a challenging and under-explored task. Firearms, especially handguns, are typically very small in size, nonsalient in appearance, and often severely occluded or indistinguishable from other small objects. Additionally, the lack of principled benchmarks and difficulty collecting relevant datasets further hinder algorithmic development. In this paper, we present a meticulously crafted and annotated benchmark, called CCTV-Gun, which addresses the challenges of detecting handguns in real-world CCTV images. Our contribution is threefold. Firstly, we carefully select and analyze real-world CCTV images from three datasets, manually annotate handguns and their holders, and assign each image with relevant challenge factors such as blur and occlusion. Secondly, we propose a new crossdataset evaluation protocol in addition to the standard intra-dataset protocol, which is vital for gun detection in practical settings. Finally, we comprehensively evaluate both classical and state-of-the-art object detection algorithms, providing an in-depth analysis of their generalizing abilities. The benchmark will facilitate further research and development on this topic and ultimately enhance security. Code, annotations, and trained models are available at https://github.com/srikarym/CCTV-Gun."}, "cited_paper_content": {"title": "Ssd: Single Shot Multibox Detector", "abstract": "We present a method for detecting objects in images using a single deep neural network. Our approach, named SSD, discretizes the output space of bounding boxes into a set of default boxes over different aspect ratios and scales per feature map location. At prediction time, the network generates scores for the presence of each object category in each default box and produces adjustments to the box to better match the object shape. Additionally, the network combines predictions from multiple feature maps with different resolutions to naturally handle objects of various sizes. SSD is simple relative to methods that require object proposals because it completely eliminates proposal generation and subsequent pixel or feature resampling stages and encapsulates all computation in a single network. This makes SSD easy to train and straightforward to integrate into systems that require a detection component. Experimental results on the PASCAL VOC, COCO, and ILSVRC datasets confirm that SSD has competitive accuracy to methods that utilize an additional object proposal step and is much faster, while providing a unified framework for both training and inference. For \\(300 \\times 300\\) input, SSD achieves 74.3 % mAP on VOC2007 test at 59 FPS on a Nvidia Titan X and for \\(512 \\times 512\\) input, SSD achieves 76.9 % mAP, outperforming a comparable state of the art Faster R-CNN model. Compared to other single stage methods, SSD has much better accuracy even with a smaller input image size. Code is available at https://github.com/weiliu89/caffe/tree/ssd."}, "keywords": ["accuracy"], "citation_intent": "method"}
{"citing_id": "2304.05898v1", "cited_id": "1706.04599", "section_title": "Iv. Discussion", "citation": "Such overconfidence in highly nonlinear networks is consistent with phenomena reported in large image and natural language datasets #REFR .", "text_before_citation": ["By contrast, LDA, which is widely used in EMG classification, demonstrated low accuracy for all datasets.", "The strong constraint of sharing covariance among classes in the LDA is likely to adversely affect the classification of these datasets.", "However, accuracy and confidence metrics are not necessarily positively (or negatively) correlated.", "Discriminative classifiers, including Deep MLP, which shows high accuracy, tend to have worse ECE and MCE overall.", "This can be observed from the reliability diagrams, indicating that neural network-based classifiers such as MLP and Deep MLP are overconfident."], "text_after_citation": ["By contrast, generative classifiers, in particular LDA and SMMC, showed better results in terms of ECE and MCE.", "Further analysis is required to fully understand the underlying reasons.", "However, generative classifiers involve stronger assumptions regarding the data generative process; thus, if this assumption is consistent with the actual data distribution, then the confidence may approach the true predictive uncertainty.", "Remarkably, SMMC demonstrated superior performance in terms of both accuracy and confidence.", "In particular, ECE was less than 10% for all datasets, indicating relatively few miscalibrations."], "citing_paper_content": {"title": "Evaluating Classifier Confidence For Surface Emg Pattern Recognition", "abstract": "Surface electromyogram (EMG) can be employed as an interface signal for various devices and software via pattern recognition. In EMG-based pattern recognition, the classifier should not only be accurate, but also output an appropriate confidence (i.e., probability of correctness) for its prediction. If the confidence accurately reflects the likelihood of true correctness, then it will be useful in various application tasks, such as motion rejection and online adaptation. The aim of this paper is to identify the types of classifiers that provide higher accuracy and better confidence in EMG pattern recognition. We evaluate the performance of various discriminative and generative classifiers on four EMG datasets, both visually and quantitatively. The analysis results show that while a discriminative classifier based on a deep neural network exhibits high accuracy, it outputs a confidence that differs from true probabilities. By contrast, a scale mixture model-based classifier, which is a generative classifier that can account for uncertainty in EMG variance, exhibits superior performance in terms of both accuracy and confidence."}, "cited_paper_content": {"title": "On Calibration Of Modern Neural Networks", "abstract": "Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions."}, "keywords": ["highly nonlinear networks"], "citation_intent": "result"}
{"citing_id": "2303.15647v1", "cited_id": "1806.03822", "section_title": "Freeze And Reconfigure (Far)", "citation": "The authors showed that FAR achieved similar performance to fine-tuning on five GLUE tasks and SQuAD 2.0 #REFR while updating 6% of the parameters.", "text_before_citation": ["h1 = x @ W_t h2 = x @ W_f return concat([h1, h2], dim=-1)", "While this approach creates additional compute overhead during training, it provides great flexibility in terms of parameter selection on modern hardware using standard frameworks like Py-Torch.", "After training, the parameters can be reconfigured back removing any inference overhead.", "The original paper focused on edge scenarios and used DistilBERT (66M) for their experiments.", "FAR was only applied to feed-forward layers, as these make up the majority of the DistilBERT parameters."], "text_after_citation": [], "citing_paper_content": {"title": "Scaling Down To Scale Up: A Guide To Parameter-Efficient Fine-Tuning", "abstract": "This paper presents a systematic overview and comparison of parameter-efficient finetuning methods covering over 40 papers published between February 2019 and February 2023. These methods aim to resolve the infeasibility and impracticality of fine-tuning large language models by only training a small set of parameters. We provide a taxonomy that covers a broad range of methods and present a detailed method comparison with a specific focus on real-life efficiency and fine-tuning multibillion-scale language models. 1 github.com/google-research/bert 2 A Recipe for Training Neural Networks, A. Karpathy"}, "cited_paper_content": {"title": "Know What You Don'T Know: Unanswerable Questions For Squad", "abstract": "Extractive reading comprehension systems can often locate the correct answer to a question in a context document, but they also tend to make unreliable guesses on questions for which the correct answer is not stated in the context. Existing datasets either focus exclusively on answerable questions, or use automatically generated unanswerable questions that are easy to identify. To address these weaknesses, we present SQuAD 2.0, the latest version of the Stanford Question Answering Dataset (SQuAD). SQuAD 2.0 combines existing SQuAD data with over 50,000 unanswerable questions written adversarially by crowdworkers to look similar to answerable ones. To do well on SQuAD 2.0, systems must not only answer questions when possible, but also determine when no answer is supported by the paragraph and abstain from answering. SQuAD 2.0 is a challenging natural language understanding task for existing models: a strong neural system that gets 86% F1 on SQuAD 1.1 achieves only 66% F1 on SQuAD 2.0."}, "keywords": ["fine-tuning", "authors"], "citation_intent": "result"}
{"citing_id": "2303.12454v1", "cited_id": "1904.09237", "section_title": "Experimental Results", "citation": "Non-convergence of these optimizers is documented in literature #REFR and we can confirm with our experiments that it also extends to our optimization setting.", "text_before_citation": ["In a first step, we investigated mean squared error loss by setting \u03bb = 1 in our loss function defined in equation 2for a single polynomial, which revealed a learning rate of 0.1 as a reasonable setting.", "We then ran tests with available TensorFlow optimizers listed in #OTHEREFR and compared their outcomes.", "We found that SGD with momentum, Adam, Adamax as well as AMSgrad show the lowest losses, with a declining tendency even after 5000 epochs.", "However, the training curves of Adamax and Adam exhibit recurring phases of instability every \u223c 500 epochs."], "text_after_citation": ["Using the AMSGrad variant of Adam eliminates this behavior with comparable remaining loss levels.", "With these results in mind, we chose SGD with Nesterov momentum as non-adaptive and AMSGrad as adaptive optimizer for all further experiments, in order to work with optimizers from both paradigms.", "The AMSGrad optimizer performs better on the \u03bb = 1 optimization target, however, SGD is competitive.", "The loss curves of these optimizer candidates, as well as instabilities in the Adam loss curve are shown in figure 2.", "An overview of all evaluated optimizers is given in our GitHub repository at #OTHEREFR ."], "citing_paper_content": {"title": "C K -Continuous Spline Approximation With Tensorflow Gradient Descent Optimizers", "abstract": "In this work we present an \"out-of-the-box\" application of Machine Learning (ML) optimizers for an industrial optimization problem. We introduce a piecewise polynomial model (spline) for fitting of C k-continuos functions, which can be deployed in a cam approximation setting. We then use the gradient descent optimization context provided by the machine learning framework TensorFlow to optimize the model parameters with respect to approximation quality and C k-continuity and evaluate available optimizers. Our experiments show that the problem solution is feasible using TensorFlow gradient tapes and that AMSGrad and SGD show the best results among available TensorFlow optimizers. Furthermore, we introduce a novel regularization approach to improve SGD convergence. Although experiments show that remaining discontinuities after optimization are small, we can eliminate these errors using a presented algorithm which has impact only on affected derivatives in the local spline segment."}, "cited_paper_content": {"title": "On The Convergence Of Adam And Beyond", "abstract": "Several recently proposed stochastic optimization methods that have been successfully used in training deep networks such as RMSProp, Adam, Adadelta, Nadam are based on using gradient updates scaled by square roots of exponential moving averages of squared past gradients. In many applications, e.g. learning with large output spaces, it has been empirically observed that these algorithms fail to converge to an optimal solution (or a critical point in nonconvex settings). We show that one cause for such failures is the exponential moving average used in the algorithms. We provide an explicit example of a simple convex optimization setting where Adam does not converge to the optimal solution, and describe the precise problems with the previous analysis of Adam algorithm. Our analysis suggests that the convergence issues can be fixed by endowing such algorithms with `long-term memory' of past gradients, and propose new variants of the Adam algorithm which not only fix the convergence issues but often also lead to improved empirical performance."}, "keywords": ["optimizers"], "citation_intent": "result"}
{"citing_id": "2303.04851v1", "cited_id": "1804.09132", "section_title": "Cwi-2018 At Bea", "citation": "As discussed in the CWI-2018 report #REFR , the performance obtained by all teams on the News domain was generally substantially higher than the performance obtained in the two other domains.", "text_before_citation": ["CWI-2018 received submissions by 12 teams in multiple task and track combinations.", "At the end of the competition, 10 teams wrote system description papers presented at the BEA workshop.", "In Table 10 (Appendices), we present the approaches by teams who submitted systems to the CWI-2018 English binary classification task and who also wrote system description papers.", "An observed trend was that more teams tried deep neural networks in CWI-2018 compared to CWI-2016, a trend also observed in other areas of AI and NLP research (Section 7).", "In CWI-2018's binary classification task, being sub-task 1, the organizers reported the performance from all teams in each of the three domains, namely News, WikiNews, and Wikipedia."], "text_after_citation": [], "citing_paper_content": {"title": "Lexical Complexity Prediction: An Overview", "abstract": "The occurrence of unknown words in texts significantly hinders reading comprehension. To improve accessibility for specific target populations, computational modelling has been applied to identify complex words in texts and substitute them for simpler alternatives. In this paper, we present an overview of computational approaches to lexical complexity prediction focusing on the work carried out on English data. We survey relevant approaches to this problem which include traditional machine learning classifiers (e.g. SVMs, logistic regression) and deep neural networks as well as a variety of features, such as those inspired by literature in psycholinguistics as well as word frequency, word length, and many others. Furthermore, we introduce readers to past competitions and available datasets created on this topic. Finally, we include brief sections on applications of lexical complexity prediction, such as readability and text simplification, together with related studies on languages other than English."}, "cited_paper_content": {"title": "A Report On The Complex Word Identification Shared Task 2018", "abstract": "We report the findings of the second Complex Word Identification (CWI) shared task organized as part of the BEA workshop co-located with NAACL-HLT'2018. The second CWI shared task featured multilingual and multi-genre datasets divided into four tracks: English monolingual, German monolingual, Spanish monolingual, and a multilingual track with a French test set, and two tasks: binary classification and probabilistic classification. A total of 12 teams submitted their results in different task/track combinations and 11 of them wrote system description papers that are referred to in this report and appear in the BEA workshop proceedings."}, "keywords": ["CWI-2018 report"], "citation_intent": "result"}
{"citing_id": "2304.11906v1", "cited_id": "1903.11027", "section_title": "I. Introduction", "citation": "It can be seen that DETR3D-Binocular converges to a poor 3D detection performance, which can compared neither to advanced CNN-based stereo 3D object detectors nor to the surround-view detection performance #REFR of DETR3D.", "text_before_citation": ["However, for stereo 3D object detection in driving scenes, to the best of our knowledge, there have not been any public Transformer-based detectors in the literature.", "An intuitive way to introduce vision Transformers into stereo 3D object detection is to adopt the above surround-view Transformer-based 3D detectors on the KITTI #OTHEREFR dataset.", "To arXiv:2304.11906v1 [cs.CV] 24 Apr 2023 achieve that, we train a binocular revision of the surround-view 3D object detector DETR3D #OTHEREFR , which is named DETR3D-Binocular.", "The validation AP curve of DETR3D-Binocular during training is shown in Fig.", "1a as blue circles, and its evaluation indicators after training 320 epochs are listed in Fig. 1b ."], "text_after_citation": ["We argue that the convergence problem is caused by a lack of image correspondence supervision in surround-view 3D object detection Transformers such as DETR3D, as well as in existing monocular-based and LiDAR-based 3D object detection Transformers.", "It is a fact that monocular images and LiDAR point clouds do not involve cross-view correspondence #OTHEREFR , #OTHEREFR ; and the overlapping area across images in a surroundview system is small #OTHEREFR .", "Concerning the binocular vision system, however, the overlap areas between left and right images are large.", "The image correspondence information in the large overlap areas, as mentioned above, is key to decoding the 3D scene.", "Therefore, extracting and exploiting the image correspondence information in a Transformer-based model is a task-specific challenge of stereo 3D object detection."], "citing_paper_content": {"title": "Transformer-Based Stereo-Aware 3D Object Detection From Binocular Images", "abstract": "Vision Transformers have shown promising progress in various object detection tasks, including monocular 2D/3D detection and surround-view 3D detection. However, when used in essential and classic stereo 3D object detection, directly adopting those surround-view Transformers leads to slow convergence and significant precision drops. We argue that one of the causes of this defect is that the surround-view Transformers do not consider the stereo-specific image correspondence information. In a surround-view system, the overlapping areas are small, and thus correspondence is not a primary issue. In this paper, we explore the model design of vision Transformers in stereo 3D object detection, focusing particularly on extracting and encoding the task-specific image correspondence information. To achieve this goal, we present TS3D, a Transformer-based Stereo-aware 3D object detector. In the TS3D, a Disparity-Aware Positional Encoding (DAPE) model is proposed to embed the image correspondence information into stereo features. The correspondence is encoded as normalized disparity and is used in conjunction with sinusoidal 2D positional encoding to provide the location information of the 3D scene. To extract enriched multiscale stereo features, we propose a Stereo Reserving Feature Pyramid Network (SRFPN). The SRFPN is designed to reserve the correspondence information while fusing intra-scale and aggregating cross-scale stereo features. Our proposed TS3D achieves a 41.29% Moderate Car detection average precision on the KITTI test set and takes 88 ms to detect objects from each binocular image pair. It is competitive with advanced counterparts in terms of both precision and inference speed."}, "cited_paper_content": {"title": "Nuscenes: A Multimodal Dataset For Autonomous Driving", "abstract": "Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology. Image-based benchmark datasets have driven development in computer vision tasks such as object detection, tracking and segmentation of agents in the environment. Most autonomous vehicles, however, carry a combination of cameras and range sensors such as lidar and radar. As machine learning based methods for detection and tracking become more prevalent, there is a need to train and evaluate such methods on datasets containing range sensor data along with images. In this work we present nuTonomy scenes (nuScenes), the first published dataset to carry the full autonomous vehicle sensor suite: 6 cameras, 5 radars and 1 lidar, all with full 360 degree field of view. nuScenes comprises 1000 scenes, each 20s long and fully annotated with 3D bounding boxes for 23 classes and 8 attributes. It has 7x as many annotations and 100x as many images as the pioneering KITTI dataset. We define novel 3D detection and tracking metrics. We also provide careful dataset analysis as well as baselines for lidar and image based detection and tracking. Data, development kit and more information are available online at this http URL."}, "keywords": ["3D object detectors"], "citation_intent": "result"}
{"citing_id": "2304.06370v1", "cited_id": "1911.05722", "section_title": "Introduction", "citation": "To this end, we introduce SuMoCo, a novel GPUfriendly framework based on its unsupervised counterpart, MoCo #REFR .", "text_before_citation": ["This task involves classifying a driver's actions into \"normal driving\" and several non-driving-related activities (NDRAs) such as talking and drinking.", "Existing approaches #OTHEREFR typically employ a naive fusion method that combines the multiview multimodal data at the decision level.", "However, this approach fails to exploit the complementarity of the semantic features from different views and modalities and does not consider their relative importance, leading to underperformance.", "Hence, we propose DMSs based on feature-level fusion with self-attention, as depicted in Fig. 1 , to address these limitations.", "Following the prior work #OTHEREFR , we also train our models with supervised contrastive learning (CL)."], "text_after_citation": ["Given that drivers can perform indefinite non-drivingrelated activities, we evaluate our proposed methods on the Driver Anomaly Detection Dataset (DAD) #OTHEREFR , which is designed explicitly for open-set NDRA detection.", "The dataset consists of only two categories, namely \"normal\" and \"anomalous\" (the class for all NDRAs).", "However, identifying the specific type of NDRA is critical in practice as they pose varying degrees of risk from inattention.", "Therefore, we manually annotate DAD with fine-grained labels, such as \"drinking\", \"talking\" and \"texting\", to enable the multi-classification of NDRAs.", "Our contributions in this paper are three-fold as follows: 1."], "citing_paper_content": {"title": "Robust Multiview Multimodal Driver Monitoring System Using Masked Multi-Head Self-Attention", "abstract": "Driver Monitoring Systems (DMSs) are crucial for safe handover actions in Level-2+ self-driving vehicles. Stateof-the-art DMSs leverage multiple sensors mounted at different locations to monitor the driver and the vehicle's interior scene and employ decision-level fusion to integrate these heterogenous data. However, this fusion method may not fully utilize the complementarity of different data sources and may overlook their relative importance. To address these limitations, we propose a novel multiview multimodal driver monitoring system based on feature-level fusion through multi-head self-attention (MHSA). We demonstrate its effectiveness by comparing it against four alternative fusion strategies (Sum, Conv, SE, and AFF). We also present a novel GPU-friendly supervised contrastive learning framework SuMoCo to learn better representations. Furthermore, We fine-grained the test split of the DAD dataset to enable the multi-class recognition of drivers' activities. Experiments on this enhanced database demonstrate that 1) the proposed MHSA-based fusion method (AUC-ROC: 97.0%) outperforms all baselines and previous approaches, and 2) training MHSA with patch masking can improve its robustness against modality/view collapses. The code and annotations are publicly available 1 ."}, "cited_paper_content": {"title": "Momentum Contrast For Unsupervised Visual Representation Learning", "abstract": "We present Momentum Contrast (MoCo) for unsupervised visual representation learning. From a perspective on contrastive learning as dictionary look-up, we build a dynamic dictionary with a queue and a moving-averaged encoder. This enables building a large and consistent dictionary on-the-fly that facilitates contrastive unsupervised learning. MoCo provides competitive results under the common linear protocol on ImageNet classification. More importantly, the representations learned by MoCo transfer well to downstream tasks. MoCo can outperform its supervised pre-training counterpart in 7 detection/segmentation tasks on PASCAL VOC, COCO, and other datasets, sometimes surpassing it by large margins. This suggests that the gap between unsupervised and supervised representation learning has been largely closed in many vision tasks."}, "keywords": ["unsupervised counterpart"], "citation_intent": "method"}
{"citing_id": "2305.01278v1", "cited_id": "1607.08822", "section_title": "Introduction", "citation": "Without updating the whole architecture on the on-demand data, ideally, the large backbone LLM can be kept fully frozen, with only the comparatively light VPG module being (fully or partially) updated. #REFR Motivation. However, building a VL-LLM is inevitably computational costing.", "text_before_citation": ["To construct a VL-LLM, a visual prompt generator (VPG) module (cf. Fig.", "1 ) that produces soft prompts for the input images/videos is added #OTHEREFR at the bottom of the backbone LLM for bridging the gap between vision and language modalities.", "Currently, such architecture has been frequently adopted by many popular VL-LLMs.", "For example, BLIP-2 pre-train a CLIP-ViT #OTHEREFR combined with a Q-Former as VPG; PaLM-E explores several different architectures, e.g. ViT #OTHEREFR or OSRT #OTHEREFR as VPG.", "To obtain the final VL-LLM, the VPG needs to be tuned."], "text_after_citation": ["Because even though the backbone LLM is frozen, tuning the light VPG and projector parts still requires additional huge computation resources to cover the LLM load that accounts for the main overhead.", "For example, tuning the VPG and projector for BLIP2-FlanT5 XXL still needs over 600 A100-GPU hours on over 100 million image-text pairs, where the LLM takes up 70% of the GPU memory consumption.", "Thus, whenever developing a new VL-LLM with any newly-emerging LLM, it can be prohibitively expensive to train the VPG and projector.", "Hopefully, transferring a pre-trained VPG (which is the main body of trainable parts) from an existing VL-LLM to a novel LLM instead of training from scratch #OTHEREFR , offers a promising solution for relieving the cost issue.", "Intuitively, all the VL-LLMs literally can share the same VPG infrastructure and utility, #OTHEREFR which makes the VPG transferring theoretically feasible."], "citing_paper_content": {"title": "Transfer Visual Prompt Generator Across Llms", "abstract": "While developing a new vision-language LLM (VL-LLM) by pre-training on tremendous image-text pairs from scratch can be exceedingly resource-consuming, connecting an existing LLM with a comparatively lightweight visual prompt generator (VPG) becomes a feasible paradigm. However, further tuning the VPG part of the VL-LLM still suffers from indispensable computational costs, i.e., requiring thousands of GPU hours and millions of training data. One alternative solution is to transfer an existing VPG from any existing VL-LLMs for the target VL-LLM."}, "cited_paper_content": {"title": "Spice: Semantic Propositional Image Caption Evaluation", "abstract": "There is considerable interest in the task of automatically generating image captions. However, evaluation is challenging. Existing automatic evaluation metrics are primarily sensitive to n-gram overlap, which is neither necessary nor sufficient for the task of simulating human judgment. We hypothesize that semantic propositional content is an important component of human caption evaluation, and propose a new automated caption evaluation metric defined over scene graphs coined SPICE. Extensive evaluations across a range of models and datasets indicate that SPICE captures human judgments over model-generated captions better than other automatic metrics (e.g., system-level correlation of 0.88 with human judgments on the MS COCO dataset, versus 0.43 for CIDEr and 0.53 for METEOR). Furthermore, SPICE can answer questions such as which caption-generator best understands colors? and can caption-generators count?"}, "keywords": ["large backbone LLM", "VL-LLM"], "citation_intent": "background"}
{"citing_id": "2304.13896v1", "cited_id": "1910.01047", "section_title": "Runtime", "citation": "The pathwidth pw(G) of graph G is the minimum width over all TDs of G whose trees are paths. Proposition II.4 (LB for Pathwidth #REFR ).", "text_before_citation": ["The treewidth tw(G) of G is the minimum width(T ) over all TDs T of G. For QSAT, the following tractability result is known.", "Proposition II.3 (Treewidth UB #OTHEREFR ).", "Given any CQBF Q of quantifier depth \u2113 with k=tw(G Q ).", "QSAT on Q can be decided in time tow(\u2113, O(k)) \u2022 poly(|var(Q)|).", "However, it is not expected that one can significantly improve this, since already for the weaker pathwidth there are limits."], "text_after_citation": ["Given any CQBF Q of quantifier depth \u2113 with k = pw(G Q ).", "v 1 v 2 \u2022 \u2022 \u2022 v n\u22121 v n", "Then, under ETH, QSAT on Q cannot be decided in time tow(\u2113, o(k)) \u2022 poly(|var(Q)|).", "Treedepth: Given a graph G = (V, E).", "Then, a treedepth decomposition T = (V, F ) of G is a forest of rooted trees, where for every edge {u, v} \u2208 E we require that u is an ancestor or descendant of v in T ."], "citing_paper_content": {"title": "Structure-Aware Lower Bounds And Broadening The Horizon Of Tractability For Qbf", "abstract": "The QSAT problem, which asks to evaluate a quantified Boolean formula (QBF), is of fundamental interest in approximation, counting, decision, and probabilistic complexity and is also considered the prototypical PSPACEcomplete problem. As such, it has previously been studied under various structural restrictions (parameters), most notably parameterizations of the primal graph representation of instances. Indeed, it is known that QSAT remains PSPACE-complete even when restricted to instances with constant treewidth of the primal graph, but the problem admits a double-exponential fixed-parameter algorithm parameterized by the vertex cover number (primal graph). However, prior works have left a gap in our understanding of the complexity of QSAT when viewed from the perspective of other natural representations of instances, most notably via incidence graphs. In this paper, we develop structure-aware reductions which allow us to obtain essentially tight lower bounds for highly restricted instances of QSAT, including instances whose incidence graphs have bounded treedepth or feedback vertex number. We complement these lower bounds with novel algorithms for QSAT which establish a nearly-complete picture of the problem's complexity under standard graph-theoretic parameterizations. We also show implications for other natural graph representations, and obtain novel upper as well as lower bounds for QSAT under more fine-grained parameterizations of the primal graph."}, "cited_paper_content": {"title": "Te-Eth: Lower Bounds For Qbfs Of Bounded Treewidth", "abstract": "The problem of deciding the validity (QSAT) of quantified Boolean formulas (QBF) is a vivid research area in both theory and practice. In the field of parameterized algorithmics, the well-studied graph measure treewidth turned out to be a successful parameter. A well-known result by Chen in parameterized complexity is that QSAT when parameterized by the treewidth of the primal graph of the input formula together with the quantifier depth of the formula is fixed-parameter tractable. More precisely, the runtime of such an algorithm is polynomial in the formula size and exponential in the treewidth, where the exponential function in the treewidth is a tower, whose height is the quantifier depth. A natural question is whether one can significantly improve these results and decrease the tower while assuming the Exponential Time Hypothesis (ETH). In the last years, there has been a growing interest in the quest of establishing lower bounds under ETH, showing mostly problem-specific lower bounds up to the third level of the polynomial hierarchy. Still, an important question is to settle this as general as possible and to cover the whole polynomial hierarchy. In this work, we show lower bounds based on the ETH for arbitrary QBFs parameterized by treewidth (and quantifier depth). More formally, we establish lower bounds for QSAT and treewidth, namely, that under ETH there cannot be an algorithm that solves QSAT of quantifier depth i in runtime significantly better than i-fold exponential in the treewidth and polynomial in the input size. In doing so, we provide a versatile reduction technique to compress treewidth that encodes the essence of dynamic programming on arbitrary tree decompositions. Further, we describe a general methodology for a more fine-grained analysis of problems parameterized by treewidth that are at higher levels of the polynomial hierarchy."}, "keywords": ["whose trees"], "citation_intent": "background"}
{"citing_id": "2303.00609v1", "cited_id": "1901.07031", "section_title": "A. Modalities And Datasets", "citation": "CXR to evaluate our selected models on the challenging CXR modality, we used CheXpert #REFR , a large publicly available dataset containing 224 316 chest radiographs.", "text_before_citation": ["We also expanded our evaluation with two additional modalities, namely Chest X-ray (CXR) and Retinal Fundus Photography (RF), in an attempt to establish a more comprehensive benchmark for UPD.", "Brain-MRI We used the Cambridge Centre for Ageing and Neuroscience dataset (CamCAN) #OTHEREFR as our source of lesionfree scans for all Brain-MRI datasets.", "It contains T1 and T2-weighted brain scans of 652 healthy adult subjects.", "The Anatomical Tracings of Lesions After Stroke (ATLAS) dataset #OTHEREFR provided us with 220 T1-weighted scans of stroke patients, which we use for evaluation.", "We also used the 2020 version of the Multimodal Brain Tumor Segmentation (BraTS) Challenge dataset for evaluation #OTHEREFR - #OTHEREFR ."], "text_after_citation": ["Samples are labeled as positive, negative, or uncertain for 14 observations, 12 of which are pathology related.", "The remaining two regard the presence of support devices or the absence of all pathologies, labeled as \"no finding\".", "We use the last observation for our \"healthy\" subset but we do so with caution as the label only implies the absence of the 12 pre-defined pathologies of interest, a common limitation in CXR datasets.", "We limited the task by using images annotated with exactly one out of three of the most common observations, namely \"pleural effusion\", \"opacity\" or \"enlarged cardiomediastinal contour\".", "We avoided possible confounders and biases by discarding scans with support devices, using only radiographs of anteroposterior view, and by enforcing an equal number of samples of male and female patients on all consequent subsets."], "citing_paper_content": {"title": "Unsupervised Pathology Detection: A Deep Dive Into The State Of The Art", "abstract": "Deep unsupervised approaches are gathering increased attention for applications such as pathology detection and segmentation in medical images since they promise to alleviate the need for large labeled datasets and are more generalizable than their supervised counterparts in detecting any kind of rare pathology. As the Unsupervised Anomaly Detection (UAD) literature continuously grows and new paradigms emerge, it is vital to continuously evaluate and benchmark new methods in a common framework, in order to reassess the state-ofthe-art (SOTA) and identify promising research directions. To this end, we evaluate a diverse selection of cutting-edge UAD methods on multiple medical datasets, comparing them against the established SOTA in UAD for brain MRI. Our experiments demonstrate that newly developed feature-modeling methods from the industrial and medical literature achieve increased performance compared to previous work and set the new SOTA in a variety of modalities and datasets. Additionally, we show that such methods are capable of benefiting from recently developed self-supervised pre-training algorithms, further increasing their performance. Finally, we perform a series of experiments in order to gain further insights into some unique characteristics of selected models and datasets. Our code can be found under https://github.com/iolag/UPD study/."}, "cited_paper_content": {"title": "Chexpert: A Large Chest Radiograph Dataset With Uncertainty Labels And Expert Comparison", "abstract": "Large, labeled datasets have driven deep learning methods to achieve expert-level performance on a variety of medical imaging tasks. We present CheXpert, a large dataset that contains 224,316 chest radiographs of 65,240 patients. We design a labeler to automatically detect the presence of 14 observations in radiology reports, capturing uncertainties inherent in radiograph interpretation. We investigate different approaches to using the uncertainty labels for training convolutional neural networks that output the probability of these observations given the available frontal and lateral radiographs. On a validation set of 200 chest radiographic studies which were manually annotated by 3 board-certified radiologists, we find that different uncertainty approaches are useful for different pathologies. We then evaluate our best model on a test set composed of 500 chest radiographic studies annotated by a consensus of 5 board-certified radiologists, and compare the performance of our model to that of 3 additional radiologists in the detection of 5 selected pathologies. On Cardiomegaly, Edema, and Pleural Effusion, the model ROC and PR curves lie above all 3 radiologist operating points. We release the dataset to the public as a standard benchmark to evaluate performance of chest radiograph interpretation models. ::: The dataset is freely available at this https URL ."}, "keywords": ["dataset", "radiographs"], "citation_intent": "method"}
{"citing_id": "2303.00844v1", "cited_id": "1711.07622", "section_title": "4:", "citation": "The standard OMP algorithm is a special case of Algorithm 1 when G is the least-squares loss function, i.e., G(z) = F (z) = y \u2212 Az #REFR 2 and for \u03bb = 0.", "text_before_citation": ["Let x (0) = 0 and S (0) = \u2205 5:", "for k = 1, . . .", ", K do 6: Find j (k) \u2208 arg max j\u2208[N ] \u2206(x (k\u22121) , S (k\u22121) , j), with \u2206 defined as in (2) 7: Define S (k) = S (k\u22121) \u222a {j (k) } 8:", "Compute x (k) by solving (3) return x (K) 11: end procedure Remark 1 (Standard OMP)."], "text_after_citation": ["To demonstrate that Algorithm 1 is practically implementable, we ought to show that the loss reduction factor \u2206(x, S, j) is (ideally, easily) computable.", "The main technical contribution of the paper is to show that this is indeed the case for the weighted LASSO, SR-LASSO, and LAD-LASSO loss functions (referred to as \" * -LASSO\" below).", "This is summarized in the following result, which unifies Theorems 2, 3 and 4."], "citing_paper_content": {"title": "The Greedy Side Of The Lasso: New Algorithms For Weighted Sparse Recovery Via Loss Function-Based Orthogonal Matching Pursuit", "abstract": "We propose a class of greedy algorithms for weighted sparse recovery by considering new loss function-based generalizations of Orthogonal Matching Pursuit (OMP). Given a (regularized) loss function, the proposed algorithms alternate the iterative construction of the signal support via greedy index selection and a signal update based on solving a local data-fitting problem restricted to the current support. We show that greedy selection rules associated with popular weighted sparsity-promoting loss functions admit explicitly computable and simple formulas. Specifically, we consider \u2113 0-and \u2113 1-based versions of the weighted LASSO (Least Absolute Shrinkage and Selection Operator), the Square-Root LASSO (SR-LASSO) and the Least Absolute Deviations LASSO (LAD-LASSO). Through numerical experiments on Gaussian compressive sensing and high-dimensional function approximation, we demonstrate the effectiveness of the proposed algorithms and empirically show that they inherit desirable characteristics from the corresponding loss functions, such as SR-LASSO's noise-blind optimal parameter tuning and LAD-LASSO's fault tolerance. In doing so, our study sheds new light on the connection between greedy sparse recovery and convex relaxation."}, "cited_paper_content": {"title": "Correcting For Unknown Errors In Sparse High-Dimensional Function Approximation", "abstract": "We consider sparsity-based techniques for the approximation of high-dimensional functions from random pointwise evaluations. To date, almost all the works published in this field contain some a priori assumptions about the error corrupting the samples that are hard to verify in practice. In this paper, we instead focus on the scenario where the error is unknown. We study the performance of four sparsity-promoting optimization problems: weighted quadratically-constrained basis pursuit, weighted LASSO, weighted square-root LASSO, and weighted LAD-LASSO. From the theoretical perspective, we prove uniform recovery guarantees for these decoders, deriving recipes for the optimal choice of the respective tuning parameters. On the numerical side, we compare them in the pure function approximation case and in applications to uncertainty quantification of ODEs and PDEs with random inputs. Our main conclusion is that the lesser-known square-root LASSO is better suited for high-dimensional approximation than the other procedures in the case of bounded noise, since it avoids (both theoretically and numerically) the need for parameter tuning."}, "keywords": ["least-squares loss function"], "citation_intent": "method"}
{"citing_id": "2303.16195v1", "cited_id": "1909.08418", "section_title": "Introduction", "citation": "A related phenomenon has been observed on neuromorphic chips, which optimally perform in simple tasks when in the subcritical regime, whereas harder tasks require progressively more critical dynamics #REFR .", "text_before_citation": ["A: Environment with 50 organisms (red circles with trails) foraging for food (green dots). B: A network with 12 neurons. Four sensory, four hidden and four motor neurons. All allowed edges are displayed.", "The exact topology and edge weights are subject to change by the evolutionary algorithm.", "decreasing the system size, making the task less complex, or introducing noise to the system pushes the optimal regime further into the subcritical range #OTHEREFR . Similarly #OTHEREFR", "(2007) observed, that whereas information propagation is maximized in critical RBNs, the optimal regime shifts slightly into the subcritical regime under the presence of noise.", "In recordings from the nervous systems of different animals, slightly subcritical behaviour was observed #OTHEREFR ."], "text_after_citation": ["The escape dynamics of schooling fish remain in subcritical state even under pharmacological manipulation that increases alertness, though more alert fish self-organize closer to critical state #OTHEREFR .", "Finally, for some applications the combination of systems operating at different distances from criticality has been shown to lead to optimal results #OTHEREFR .", "The disordered supercritical state has been universally observed to perform poorly (S. A. #OTHEREFR .", "There are different critical transitions in the examples mentioned above, thus different definitions of what it means to be subcritical.", "For example, for the order/disorder transition, subcritical refers to the ordered state."], "citing_paper_content": {"title": "When To Be Critical? Performance And Evolvability In Different Regimes Of Neural Ising Agents", "abstract": "It has long been hypothesized that operating close to the critical state is beneficial for natural, artificial and their evolutionary systems. We put this hypothesis to test in a system of evolving foraging agents controlled by neural networks that can adapt agents' dynamical regime throughout evolution. Surprisingly, we find that all populations that discover solutions, evolve to be subcritical. By a resilience analysis, we find that there are still benefits of starting the evolution in the critical regime. Namely, initially critical agents maintain their fitness level under environmental changes (for example, in the lifespan) and degrade gracefully when their genome is perturbed. At the same time, initially subcritical agents, even when evolved to the same fitness, are often inadequate to withstand the changes in the lifespan and degrade catastrophically with genetic perturbations. Furthermore, we find the optimal distance to criticality depends on the task complexity. To test it we introduce a hard and simple task: for the hard task, agents evolve closer to criticality whereas more subcritical solutions are found for the simple task. We verify that our results are independent of the selected evolutionary mechanisms by testing them on two principally different approaches: a genetic algorithm and an evolutionary strategy. In summary, our study suggests that although optimal behaviour in the simple task is obtained in a subcritical regime, initializing near criticality is important to be efficient at finding optimal solutions for new tasks of unknown complexity."}, "cited_paper_content": {"title": "Control Of Criticality And Computation In Spiking Neuromorphic Networks With Plasticity", "abstract": "The critical state is assumed to be optimal for any computation in recurrent neural networks, because criticality maximizes a number of abstract computational properties. We challenge this assumption by evaluating the performance of a spiking recurrent neural network on a set of tasks of varying complexity at - and away from critical network dynamics. To that end, we developed a spiking network with synaptic plasticity on a neuromorphic chip. We show that the distance to criticality can be easily adapted by changing the input strength, and then demonstrate a clear relation between criticality, task-performance and information-theoretic fingerprint. Whereas the information-theoretic measures all show that network capacity is maximal at criticality, this is not the case for performance on specific tasks: Only the complex, memory-intensive tasks profit from criticality, whereas the simple tasks suffer from it. Thereby, we challenge the general assumption that criticality would be beneficial for any task, and provide instead an understanding of how the collective network state should be tuned to task requirement to achieve optimal performance."}, "keywords": ["critical dynamics", "neuromorphic chips"], "citation_intent": "background"}
{"citing_id": "2304.06495v1", "cited_id": "1708.08012", "section_title": "Dataset", "citation": "In this study, we train and compare our models on the publicly available High-Gamma Dataset #REFR .", "text_before_citation": [], "text_after_citation": ["It contains 128-channel EEG recordings of fourteen healthy subjects executing three different movement classes and a resting condition.", "The movements are executed by the left hand, the right hand, and both feet.", "During each trial, the subject is asked to execute one of the movement or a rest for a duration of four seconds.", "The exact number of trials varied between subjects but was balanced between the four classes.", "For each subject, approximately 1000 trials were recorded in a single session structured into 13 runs."], "citing_paper_content": {"title": "An Embedding For Eeg Signals Learned Using A Triplet Loss", "abstract": "Neurophysiological time series recordings like the electroencephalogram (EEG) or local field potentials are obtained from multiple sensors. They can be decoded by machine learning models in order to estimate the ongoing brain state of a patient or healthy user. In a brain-computer interface (BCI), this decoded brain state information can be used with minimal time delay to either control an application, e.g., for communication or for rehabilitation after stroke, or to passively monitor the ongoing brain state of the subject, e.g., in a demanding work environment. A specific challenge in such decoding tasks is posed by the small dataset sizes in BCI compared to other domains of machine learning like computer vision or natural language processing. A possibility to tackle classification or regression problems in BCI despite small training data sets is through transfer learning, which utilizes data from other sessions, subjects or even datasets to train a model. In this exploratory study, we propose novel domain-specific embeddings for neurophysiological data. Our approach is based on metric learning and builds upon the recently proposed ladder loss. Using embeddings allowed us to benefit, both from the good generalisation abilities and robustness of deep learning and from the fast training of classical machine learning models for subject-specific calibration. In offline analyses using EEG data of 14 subjects, we tested the embeddings' feasibility and compared their efficiency with state-of-the-art deep learning models and conventional machine learning pipelines. For within-subject classification of motor tasks, our data suggests that embeddings significantly improve the decoding accuracy compared to an established baseline model. In between-subjects transfer scenarios, pre-trained embeddings allow for an extremely fast fine-tuning of a classifier for a novel subject, even if only a few trials per class are available, which currently can not be achieved with regular deep learning models. And in the scenarios where one subject is completely set aside for testing without specific calibration, the embeddings were not competitive. In summary, we propose the use of metric learning to obtain pre-trained embeddings of EEG-BCI data as a means to incorporate domain knowledge and to reach competitive performance on novel subjects with minimal calibration requirements."}, "cited_paper_content": {"title": "Deep Learning With Convolutional Neural Networks For Decoding And Visualization Of Eeg Pathology", "abstract": "We apply convolutional neural networks (ConvNets) to the task of distinguishing pathological from normal EEG recordings in the Temple University Hospital EEG Abnormal Corpus. We use two basic, shallow and deep ConvNet architectures recently shown to decode task-related information from EEG at least as well as established algorithms designed for this purpose. In decoding EEG pathology, both ConvNets reached substantially better accuracies (about 6% better, ~85% vs. ~79%) than the only published result for this dataset, and were still better when using only 1 minute of each recording for training and only six seconds of each recording for testing. We used automated methods to optimize architectural hyperparameters and found intriguingly different ConvNet architectures, e.g., with max pooling as the only nonlinearity. Visualizations of the ConvNet decoding behavior showed that they used spectral power changes in the delta (0-4 Hz) and theta (4-8 Hz) frequency range, possibly alongside other features, consistent with expectations derived from spectral analysis of the EEG data and from the textual medical reports. Analysis of the textual medical reports also highlighted the potential for accuracy increases by integrating contextual information, such as the age of subjects. In summary, the ConvNets and visualization techniques used in this study constitute a next step towards clinically useful automated EEG diagnosis and establish a new baseline for future work on this topic."}, "keywords": ["Dataset"], "citation_intent": "method"}
{"citing_id": "2304.00749v1", "cited_id": "1505.04597", "section_title": "Improving U-Net Architecture", "citation": "The original U-Net #REFR , which is proposed with symmetrical encoder-decoder architecture, is widely used in image segmentation.", "text_before_citation": [], "text_after_citation": ["In particular, skip connections are used to combine the high-level semantic features from the decoder and low-level details from the encoder.", "Based on this architecture, a number of improved network architectures have been further proposed.", "Res-UNet #OTHEREFR and Dense-UNet #OTHEREFR replace each sub-module of the U-Net with a residual connection or a dense connection, respectively.", "U-Net++ #OTHEREFR redesigns skip connection pathways to explore the optimal depth by fusing multiple U-Net sub-networks with varying depths, where each sub-network is connected through a series of nested, dense skip pathways.", "U-Net3+ #OTHEREFR proposes full-scale skip connections to fully exploit multi-scale feature representations."], "citing_paper_content": {"title": "Small But Mighty: Enhancing 3D Point Clouds Semantic Segmentation With U-Next Framework", "abstract": "We study the problem of semantic segmentation of largescale 3D point clouds. In recent years, significant research efforts have been directed toward local feature aggregation, improved loss functions and sampling strategies. While the fundamental framework of point cloud semantic segmentation has been largely overlooked, with most existing approaches rely on the U-Net architecture by default. In this paper, we propose U-Next, a small but mighty framework designed for point cloud semantic segmentation. The key to this framework is to learn multi-scale hierarchical representations from semantically similar feature maps. Specifically, we build our U-Next by stacking multiple U-Net L 1 codecs in a nested and densely arranged manner to minimize the semantic gap, while simultaneously fusing the feature maps across scales to effectively recover the fine-grained details. We also devised a multi-level deep supervision mechanism to further smooth gradient propagation and facilitate network optimization. Extensive experiments conducted on three large-scale benchmarks including S3DIS, Toronto3D, and SensatUrban demonstrate the superiority and the effectiveness of the proposed U-Next architecture. Our U-Next architecture shows consistent and visible performance improvements across different tasks and baseline models, indicating its great potential to serve as a general framework for future research."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["original U-Net", "image segmentation"], "citation_intent": "method"}
{"citing_id": "2304.07437v1", "cited_id": "1912.08777", "section_title": "Introduction", "citation": "The original Seq2Seq-based methods train the encoder and decoder with an attention mechanism to predict the most probable word that emerges next in the summary text sequence #REFR .", "text_before_citation": ["Among these, question summarization has gained increasing attention from the research community.", "Medical question summarization (MQS) is the task of summarizing longer consumer health questions (CHQs) into shorter ones, i.e., frequently asked questions (FAQs), which contain the essential intent of original CHQs.", "Compared with lengthy CHQs, FAQs focus on the key information of health questions and filter out unnecessary and redundant information, which is helpful for MQAs to retrieve correct answers.", "An example of MQS is shown in Table 1 .", "In recent, there have been various deep learning-based approaches for medical question summarization, which can be categorized into three groups: Seq2Seq-based methods, RL (reinforcement learning)-enhanced methods, and CL (contrastive learning)-enhanced methods #OTHEREFR ."], "text_after_citation": ["Some variants of Seq2Seq, including BART #OTHEREFR , pointer-generator network , and multi-task learning framework #OTHEREFR , have also been applied to the MQS task.", "However, these models rely on maximum likelihood estimation (MLE), which prioritizes the accuracy of masked token prediction over the preservation of semantic similarity or dissimilarity of questions #OTHEREFR .", "To address this limitation, RL-enhanced methods attempt to utilize reinforcement learning with question-aware semantic rewards to enhance the ability of Seq2Seq-based framework on the MQS task.", "These methods create reasonable rewards, such as question-type identification and question-focus recognition, to ensure the quality of generated question summary #OTHEREFR .", "However, RL-enhanced methods still suffer from noise gradient estimation problems, which results in unstable training #OTHEREFR , and require substantial annotated training data for the devised rewards."], "citing_paper_content": {"title": "Medical Question Summarization With Entity-Driven Contrastive Learning", "abstract": "By summarizing longer consumer health questions into shorter and essential ones, medical question answering (MQA) systems can more accurately understand consumer intentions and retrieve suitable answers. However, medical question summarization is very challenging due to obvious distinctions in health trouble descriptions from patients and doctors. Although existing works have attempted to utilize Seq2Seq, reinforcement learning, or contrastive learning to solve the problem, two challenges remain: how to correctly capture question focus to model its semantic intention, and how to obtain reliable datasets to fairly evaluate performance. To address these challenges, this paper proposes a novel medical question summarization framework using entity-driven contrastive learning (ECL). ECL employs medical entities in frequently asked questions (FAQs) as focuses and devises an effective mechanism to generate hard negative samples. This approach forces models to pay attention to the crucial focus information and generate more ideal question summarization. Additionally, we find that some MQA datasets suffer from serious data leakage problems, such as the iCliniq dataset's 33% duplicate rate. To evaluate the related methods fairly, this paper carefully checks leaked samples to reorganize more reasonable datasets. Extensive experiments demonstrate that our ECL method outperforms state-of-the-art methods by accurately capturing question focus and generating medical question summaries. The code and datasets are available at https://github.com/yrbobo/MQS-ECL."}, "cited_paper_content": {"title": "Pegasus: Pre-Training With Extracted Gap-Sentences For Abstractive Summarization", "abstract": "Recent work pre-training Transformers with self-supervised objectives on large text corpora has shown great success when fine-tuned on downstream NLP tasks including text summarization. However, pre-training objectives tailored for abstractive text summarization have not been explored. Furthermore there is a lack of systematic evaluation across diverse domains. In this work, we propose pre-training large Transformer-based encoder-decoder models on massive text corpora with a new self-supervised objective. In PEGASUS, important sentences are removed/masked from an input document and are generated together as one output sequence from the remaining sentences, similar to an extractive summary. We evaluated our best PEGASUS model on 12 downstream summarization tasks spanning news, science, stories, instructions, emails, patents, and legislative bills. Experiments demonstrate it achieves state-of-the-art performance on all 12 downstream datasets measured by ROUGE scores. Our model also shows surprising performance on low-resource summarization, surpassing previous state-of-the-art results on 6 datasets with only 1000 examples."}, "keywords": ["summary text sequence"], "citation_intent": "method"}
{"citing_id": "2303.10848v1", "cited_id": "1911.09516", "section_title": "Text Recognition Module", "citation": "To address this challenge, a customized backbone with FPN feature fusion #REFR is designed to extract features at three levels.", "text_before_citation": ["We build our text recognition module following the fashion of SAR #OTHEREFR , which provides the required attention location while having a light-weighted framework.", "It is worth noting that the recognition baseline can be replaced by other attention-based recognizers to pursue better performance. In the wild, texts are varying in scale."], "text_after_citation": ["Given the input image X \u2208 R H\u00d7W \u00d7C , the extracted visual features at different levels are denoted as", "F 1 \u2208 R H 2 \u00d7 W 2 \u00d7C , F 2 \u2208 R H 4 \u00d7 W 4 \u00d7C , and F 3 \u2208 R H 8 \u00d7 W 4", "\u00d7C from bottom to top, respectively.", "Subsequently, by training adaptive weights at nearly free cost, the fused features manage to keep the scale-invariance between different levels, represented as:", "F l f use = \u03b1 l \u2022 F 1\u2192l + \u03b2 l \u2022 F 2\u2192l + \u03b3 l \u2022 F 3\u2192l (1)"], "citing_paper_content": {"title": "Weakly-Supervised Text Instance Segmentation", "abstract": "Text segmentation is a challenging vision task with many downstream applications. Current text segmentation methods require pixel-level annotations, which are expensive in the cost of human labor and limited in application scenarios. In this paper, we take the first attempt to perform weaklysupervised text instance segmentation by bridging text recognition and text segmentation. The insight is that text recognition methods provide precise attention position of each text instance, and the attention location can feed to both a text adaptive refinement head (TAR) and a text segmentation head. Specifically, the proposed TAR generates pseudo labels by performing two-stage iterative refinement operations on the attention location to fit the accurate boundaries of the corresponding text instance. Meanwhile, the text segmentation head takes the rough attention location to predict segmentation masks which are supervised by the aforementioned pseudo labels. In addition, we design a mask-augmented contrastive learning by treating our segmentation result as an augmented version of the input text image, thus improving the visual representation and further enhancing the performance of both recognition and segmentation. The experimental results demonstrate that the proposed method significantly outperforms weakly-supervised instance segmentation methods on ICDAR13-FST (18.95% improvement) and TextSeg (17.80% improvement) benchmarks."}, "cited_paper_content": {"title": "Learning Spatial Fusion For Single-Shot Object Detection", "abstract": "Pyramidal feature representation is the common practice to address the challenge of scale variation in object detection. However, the inconsistency across different feature scales is a primary limitation for the single-shot detectors based on feature pyramid. In this work, we propose a novel and data driven strategy for pyramidal feature fusion, referred to as adaptively spatial feature fusion (ASFF). It learns the way to spatially filter conflictive information to suppress the inconsistency, thus improving the scale-invariance of features, and introduces nearly free inference overhead. With the ASFF strategy and a solid baseline of YOLOv3, we achieve the best speed-accuracy trade-off on the MS COCO dataset, reporting 38.1% AP at 60 FPS, 42.4% AP at 45 FPS and 43.9% AP at 29 FPS. The code is available at this https URL"}, "keywords": ["challenge", "FPN feature fusion"], "citation_intent": "method"}
{"citing_id": "2303.10859v1", "cited_id": "1703.07710", "section_title": "F Auxiliary Lemmas", "citation": "The following lemma #REFR will be useful to measure the difference between two value functions under two MDPs and reward functions.", "text_before_citation": [", \u03c0 n\u22121 ), we have the following corollary which will be intensively used in the analysis. Corollary 2.", "Given \u03b4 \u2208 (0, 1), the following inequality holds for any n, h \u2265 2 with probability at least 1 \u2212 \u03b4/2:", "E s h\u22121 \u223c(P ,\u03a0n) a h\u22121 ,a h \u223cU (A) s h \u223cP (\u2022|s h\u22121 ,a h\u22121 ) f n h (s h , a h ) 2 \u2264 \u03b6 n .", "In addition, for h = 1,", "E a1\u223cU (A) f n 1 (s 1 , a 1 ) 2 \u2264 \u03b6 n ."], "text_after_citation": ["We define P h V h+1 (s h , a h ) = E s\u223cP h (\u2022|s h ,a h ) [V (s)] for shorthand. Lemma 9. (Simulation lemma).", "Suppose P 1 and P 2 are two MDPs and r 1 , r 2 are the corresponding reward functions. Given a policy \u03c0, we have,", "V \u03c0 h,P1,r1 (s h ) \u2212 V \u03c0 h,P2,r2 (s h ) = H h =h E s h \u223c(P 2 ,\u03c0) a h \u223c\u03c0 r 1 (s h , a h ) \u2212 r 2 (s h , a h ) + (P 1,h \u2212 P 2,h )V \u03c0 h +1,P1,r (s h , a h )|s h = H h =h E s h \u223c(P 1 ,\u03c0) a h \u223c\u03c0 r 1 (s h , a h ) \u2212 r 2 (s h , a h ) + (P 1,h \u2212 P 2,h )V \u03c0 h +1,P2,r (s h , a h )|s h .", "The following lemma is a standard inequality in regret analysis for linear models in reinforcement learning (see Lemma G.2 in #OTHEREFR and Lemma 10 in #OTHEREFR ). Lemma 10. (Elliptical potential lemma).", "Consider a sequence of d\u00d7d positive semidefinite matrices X 1 , . . ."], "citing_paper_content": {"title": "Improved Sample Complexity For Reward-Free Reinforcement Learning Under Low-Rank Mdps", "abstract": "In reward-free reinforcement learning (RL), an agent explores the environment first without any reward information, in order to achieve certain learning goals afterwards for any given reward. In this paper we focus on reward-free RL under low-rank MDP models, in which both the representation and linear weight vectors are unknown. Although various algorithms have been proposed for reward-free low-rank MDPs, the corresponding sample complexity is still far from being satisfactory. In this work, we first provide the first known sample complexity lower bound that holds for any algorithm under low-rank MDPs. This lower bound implies it is strictly harder to find a near-optimal policy under low-rank MDPs than under linear MDPs. We then propose a novel model-based algorithm, coined RAFFLE, and show it can both find an-optimal policy and achieve an-accurate system identification via reward-free exploration, with a sample complexity significantly improving the previous results. Such a sample complexity matches our lower bound in the dependence on , as well as on K in the large d regime, where d and K respectively denote the representation dimension and action space cardinality. Finally, we provide a planning algorithm (without further interaction with true environment) for RAFFLE to learn a near-accurate representation, which is the first known representation learning guarantee under the same setting."}, "cited_paper_content": {"title": "Unifying Pac And Regret: Uniform Pac Bounds For Episodic Reinforcement Learning", "abstract": "Statistical performance bounds for reinforcement learning (RL) algorithms can be critical for high-stakes applications like healthcare. This paper introduces a new framework for theoretically measuring the performance of such algorithms called Uniform-PAC, which is a strengthening of the classical Probably Approximately Correct (PAC) framework. In contrast to the PAC framework, the uniform version may be used to derive high probability regret guarantees and so forms a bridge between the two setups that has been missing in the literature. We demonstrate the benefits of the new framework for finite-state episodic MDPs with a new algorithm that is Uniform-PAC and simultaneously achieves optimal regret and PAC guarantees except for a factor of the horizon."}, "keywords": ["reward functions"], "citation_intent": "background"}
{"citing_id": "2303.14365v1", "cited_id": "1908.08683", "section_title": "Example 1", "citation": "The \u03c3 sub-problems (4.3) and (4.16) are solved by NLCG method #REFR .", "text_before_citation": ["In this example, the background conductivity \u03c3 0 = 1 in \u2126 C and the object \u2126 2 = [\u22120.5, 0.5] \u00d7 [\u22120.5, 0.5] \u00d7 [\u22121.0, \u22120.5] and the exact abnormal conductivity is given by \u03c3 = 5 in \u2126 2 and \u03c3 = 0 in \u2126 C \\\u03a9 2 .", "So the exact conductivity is 6 in \u2126 2 and 1 in \u2126 C \\\u2126 2 .", "The mesh M h we used in this example has 798138 edges."], "text_after_citation": ["In the Algorithm 1, 3 NLCG iterations is used for Algorithm 1 and 5 NLCG iterations is used for Algorithm 2.", "We choose the truncation parameter m = 0.02 \u00d7 max{iter, 50} and M = 15 in the iter-th iteration.", "Figure 2 shows isosurfaces of the recovered \u03c3 with isovalue 2.3 in 20th, 40th, 60th and 80th iterations by Algorithm 1.", "The results tell us that the reconstructions approaches to the exact conductivity with iterations.", "As we cans see, the isosurface becomes stable after 40-th iteration."], "citing_paper_content": {"title": "Iterative Methods For An Inverse Eddy Current Problem With Total Variation Regularization", "abstract": "Conductivity reconstruction in an inverse eddy current problem is considered in the present paper. With the electric field measurement on part of domain boundary, we formulate the reconstruction problem to a constrained optimization problem with total variation regularization. Existence and stability are proved for the solution to the optimization problem. The finite element method is employed to discretize the optimization problem. The gradient Lipschitz properties of the objective functional are established for the the discrete optimization problems. We propose the alternating direction method of multipliers to solve the discrete problem. Based on the the gradient Lipschitz property, we prove the convergence by extending the admissible set to the whole finite element space. Finally, we show some numerical experiments to illustrate the efficiency of the proposed methods."}, "cited_paper_content": {"title": "Mathematical And Numerical Study Of A Three-Dimensional Inverse Eddy Current Problem", "abstract": "We study an inverse problem associated with an eddy current model. We first address the ill-posedness of the inverse problem by proving the compactness of the forward map with respect to the conductivity and the non-uniqueness of the recovery process. Then by virtue of non-radiating source conceptions, we establish a regularity result for the tangential trace of the true solution on the boundary, which is necessary to justify our subsequent mathematical formulation. After that, we formulate the inverse problem as a constrained optimization problem with an appropriate regularization and prove the existence and stability of the regularized minimizers. To facilitate the numerical solution of the nonlinear non-convex constrained optimization, we introduce a feasible Lagrangian and its discrete variant. Then the gradient of the objective functional is derived using the adjoint technique. By means of the gradient, a nonlinear conjugate gradient method is formulated for solving the optimization system, and a Sobolev gradient is incorporated to accelerate the iterative process. Numerical examples are provided to demonstrate the feasibility of the proposed algorithm."}, "keywords": ["NLCG method"], "citation_intent": "method"}
{"citing_id": "2305.00581v1", "cited_id": "1706.03762", "section_title": "Background On Transformers", "citation": "The Transformer layer #REFR consists of two modules: a multi-head attention and a feed-forward network (FFN).", "text_before_citation": [], "text_after_citation": ["Specifically, each head is represented by four main matrices: the query matrix", "W q i \u2208 R d m \u00d7d q /h , the key matrix W k i \u2208 R d m \u00d7 d k h , the value matrix W v i \u2208 R d m \u00d7 d v", "h , and the output matrix", "W o i \u2208 R d v h \u00d7d o", ", and takes the hidden states H \u2208 R l\u00d7d m of the previous layer as input, where d denotes the dimension of the model, h represents the number of head, and i denotes the index of layer number. The output of attention is given by:"], "citing_paper_content": {"title": "Multimodal Graph Transformer For Multimodal Question Answering", "abstract": "Despite the success of Transformer models in vision and language tasks, they often learn knowledge from enormous data implicitly and cannot utilize structured input data directly. On the other hand, structured learning approaches such as graph neural networks (GNNs) that integrate prior information can barely compete with Transformer models. In this work, we aim to benefit from both worlds and propose a novel Multimodal Graph Transformer for question answering tasks that requires performing reasoning across multiple modalities. We introduce a graph-involved plug-and-play quasi-attention mechanism to incorporate multimodal graph information, acquired from text and visual data, to the vanilla self-attention as effective prior. In particular, we construct the text graph, dense region graph, and semantic graph to generate adjacency matrices, and then compose them with input vision and language features to perform downstream reasoning. Such a way of regularizing self-attention with graph information significantly improves the inferring ability and helps align features from different modalities. We validate the effectiveness of Multimodal Graph Transformer over its Transformer baselines on GQA, VQAv2, and MultiModalQA datasets."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["multi-head attention"], "citation_intent": "method"}
{"citing_id": "2303.12795v1", "cited_id": "1602.06023", "section_title": "Ner-Based Pointer-Generator Network", "citation": "We do not use pretrained word embeddings as #REFR do; in our case token embeddings are learned from scratch during training.", "text_before_citation": ["This model consists of an NER-based tokenizer layer and a pointer-generator network.", "The NERbased tokenizer layer converts the words in the input document to a sequence of tokens, thus preserving an entity name as a single token.", "In particular, it uses the named entity recognizer in spaCy 1 However, we do not use entity types."], "text_after_citation": ["Here, the main role of NER is that instead of directly feeding the normal tokens of the input document into the encoder, we are passing the NER-based tokens.", "4 Experimental setup"], "citing_paper_content": {"title": "Named Entity Recognition Based Automatic Generation Of Research Highlights", "abstract": "A scientific paper is traditionally prefaced by an abstract that summarizes the paper. Recently, research highlights that focus on the main findings of the paper have emerged as a complementary summary in addition to an abstract. However, highlights are not yet as common as abstracts, and are absent in many papers. In this paper, we aim to automatically generate research highlights using different sections of a research paper as input. We investigate whether the use of named entity recognition on the input improves the quality of the generated highlights. In particular, we have used two deep learning-based models: the first is a pointer-generator network, and the second augments the first model with coverage mechanism. We then augment each of the above models with named entity recognition features. The proposed method can be used to produce highlights for papers with missing highlights. Our experiments show that adding named entity information improves the performance of the deep learning-based summarizers in terms of ROUGE, METEOR and BERTScore measures."}, "cited_paper_content": {"title": "Abstractive Text Summarization Using Sequence-To-Sequence Rnns And Beyond", "abstract": "In this work, we model abstractive text summarization using Attentional Encoder-Decoder Recurrent Neural Networks, and show that they achieve state-of-the-art performance on two different corpora. We propose several novel models that address critical problems in summarization that are not adequately modeled by the basic architecture, such as modeling key-words, capturing the hierarchy of sentence-to-word structure, and emitting words that are rare or unseen at training time. Our work shows that many of our proposed models contribute to further improvement in performance. We also propose a new dataset consisting of multi-sentence summaries, and establish performance benchmarks for further research."}, "keywords": ["token embeddings", "pretrained word embeddings"], "citation_intent": "method"}
{"citing_id": "2304.06783v1", "cited_id": "1904.01634", "section_title": "I. Introduction", "citation": "Using the system level parameterization #REFR , they provide equivalent reformulations of the resulting minimax regret optimal control problems as semidefinite programs (SDPs).", "text_before_citation": ["Using operator-theoretic techniques from robust control, controllers (in state-space representation) are synthesized to minimize the worst-case regret over all possible disturbances with bounded energy.", "Building on this framework, #OTHEREFR and #OTHEREFR synthesize regret optimal controllers subject to state and This work was supported in part by The Nature Conservancy, in part by the Cornell Atkinson Center for Sustainability, and in part by the Natural Sciences and Engineering Research Council of Canada.", "The authors are with the School of Electrical and Computer Engineering, Cornell Univ., Ithaca, NY, 14853, USA.", "Emails: Shuhao Yan (sy499@cornell.edu), Feras Al Taha (foa6@cornell.edu), Eilyan Bitar (eyb5@cornell.edu).", "input constraints."], "text_after_citation": ["While the resultant controllers are robust with respect to the worst-case disturbance realization, they may be conservative if the underlying disturbance is actually stochastic in nature.", "Taking a probabilistic view, we propose an alternative approach to regret optimal control, where controllers are designed to minimize the worst-case expected regret over all probability distributions in a given ambiguity set specified in terms of the type-2 Wasserstein distance.", "In this setting, the expected regret is defined as the difference between an expected cost incurred by a causal controller and the expected cost incurred by the optimal noncausal controller with perfect knowledge of the disturbance trajectory at the outset.", "The proposed approach can also be thought of as an alternative to distributionally robust control, see, i.e., #OTHEREFR , #OTHEREFR and references therein.", "While controllers minimizing the worst-case expected cost are robust with respect to the worst-case distribution, they may incur large expected regret under other distributions in the ambiguity set."], "citing_paper_content": {"title": "A Distributionally Robust Approach To Regret Optimal Control Using The Wasserstein Distance", "abstract": "This paper proposes a distributionally robust approach to regret optimal control of discrete-time linear dynamical systems with quadratic costs subject to stochastic additive disturbance on the state process. The underlying probability distribution of the disturbance process is unknown, but assumed to lie in a given ball of distributions defined in terms of the type-2 Wasserstein distance. In this framework, strictly causal linear disturbance feedback controllers are designed to minimize the worst-case expected regret. The regret incurred by a controller is defined as the difference between the cost it incurs in response to a realization of the disturbance process and the cost incurred by the optimal noncausal controller which has perfect knowledge of the disturbance process realization at the outset. Building on a well-established duality theory for optimal transport problems, we show how to equivalently reformulate this minimax regret optimal control problem as a tractable semidefinite program. The equivalent dual reformulation also allows us to characterize a worst-case distribution achieving the worst-case expected regret in relation to the distribution at the center of the Wasserstein ball."}, "cited_paper_content": {"title": "System Level Synthesis", "abstract": "This article surveys the System Level Synthesis framework, which presents a novel perspective on constrained robust and optimal controller synthesis for linear systems. We show how SLS shifts the controller synthesis task from the design of a controller to the design of the entire closed loop system, and highlight the benefits of this approach in terms of scalability and transparency. We emphasize two particular applications of SLS, namely large-scale distributed optimal control and robust control. In the case of distributed control, we show how SLS allows for localized controllers to be computed, extending robust and optimal control methods to large-scale systems under practical and realistic assumptions. In the case of robust control, we show how SLS allows for novel design methodologies that, for the first time, quantify the degradation in performance of a robust controller due to model uncertainty -- such transparency is key in allowing robust control methods to interact, in a principled way, with modern techniques from machine learning and statistical inference. Throughout, we emphasize practical and efficient computational solutions, and demonstrate our methods on easy to understand case studies."}, "keywords": ["optimal control problems"], "citation_intent": "method"}
{"citing_id": "2303.14977v1", "cited_id": "1709.01507", "section_title": "Attention Mechanisms", "citation": "SENet #REFR differs from RAM and STN in the way that SENet innovatively introduces attention mechanisms to the detector to refine the channels of features.", "text_before_citation": ["Attention mechanisms #OTHEREFR enable the selection of information that is required to obtain a target, with outputs dynamically weighted according to input characteristics.", "Demonstrates efficient, attention mechanism is added to the object detectors.", "RAM #OTHEREFR is a pioneer in introducing attention mechanisms into deep neural networks.", "RAM showed that the attention mechanisms are effective in working with neural networks, then related works have been presented sequentially. STN #OTHEREFR introduces a sub-network to select important regions."], "text_after_citation": ["Related work has been launched with CBAM #OTHEREFR , ECANet #OTHEREFR , SRM #OTHEREFR and GSop #OTHEREFR , all of which achieved successful records.", "Recently, transformer #OTHEREFR has become more and more popular.", "For example, VIT #OTHEREFR is pioneering the concept of utilizing the transformer encoder for feature extraction and prediction.", "3 Method"], "citing_paper_content": {"title": "A Novel Multi To Single Module For Small Object Detection", "abstract": "Small object detection presents a significant challenge in computer vision and object detection. The performance of small object detectors is often compromised by a lack of pixels and less significant features. This issue stems from information misalignment caused by variations in feature scale and information loss during feature processing. In response to this challenge, this paper proposes a novel the Multi to Single Module (M2S), which enhances a specific layer through improving feature extraction and refining features. Specifically, M2S includes the proposed Cross-scale Aggregation Module (CAM) and explored Dual Relationship Module (DRM) to improve information extraction capabilities and feature refinement effects. Moreover, this paper enhances the accuracy of small object detection by utilizing M2S to generate an additional detection head. The effectiveness of the proposed method is evaluated on two datasets, VisDrone2021-DET and SeaDronesSeeV2. The experimental results demonstrate its improved performance compared with existing methods. Compared to the baseline model (YOLOv5s), M2S improves the accuracy by about 1.1% on the VisDrone2021-DET testing dataset and 15.68% on the SeaDronesSeeV2 validation set. 1 Introduction Small object detection ([1],[2],[3],[4],[5]) has long been a challenge in object detection, which aims to accurately detect small objects (those under 32 pixels by 32 pixels) with very few visual features in the image. A detection model consisting of a backbone network, an encoder, and a decoder is proposed in You Only Look One-level Feature (YOLOF) [6]. It is proposed in YOLOF that by choosing the right scale features for output to a particular level with multiple-input and single-output, comparable performance to multiple-input and multiple-output can be obtained. In the small object detection task, this paper proposes a multiple-input to the single-output module that outputs with a low-level feature with high resolution. It provides the detector with more accurate and richer information, thereby improving its performance. Feature Pyramid network (FPN) [7] success is due to the idea of divide and conquer. Divide and conquer ensures that image features are available in different layers, whereas if only a unidirectional flow of information is delivered, then FPN ends up with a limitation at each layer. The result is that each layer of character can only be thinking within its own perspective. Thus, previous work has been devoted to increasing the interaction of different levels of feature and allowing more global thinking at each level. FPN appears as the backbone network of excellent detectors SSD [8], YOLO [9], and RCNN [10]. EffientDet [11]uses a cross-edge connecting BiFPN and repeats it to obtain better accuracy and efficiency. Although the previous works have shown gratifying performance, object detection still has a problem: the poor detection of small-scale instance objects. Based on this problem, this paper proposes a module to enhance communication at different network levels and improve feature extraction efficiency. The Cross-scale Fusion Module"}, "cited_paper_content": {"title": "Squeeze-And-Excitation Networks", "abstract": "Convolutional neural networks are built upon the convolution operation, which extracts informative features by fusing spatial and channel-wise information together within local receptive fields. In order to boost the representational power of a network, much existing work has shown the benefits of enhancing spatial encoding. In this work, we focus on channels and propose a novel architectural unit, which we term the\"Squeeze-and-Excitation\"(SE) block, that adaptively recalibrates channel-wise feature responses by explicitly modelling interdependencies between channels. We demonstrate that by stacking these blocks together, we can construct SENet architectures that generalise extremely well across challenging datasets. Crucially, we find that SE blocks produce significant performance improvements for existing state-of-the-art deep architectures at slight computational cost. SENets formed the foundation of our ILSVRC 2017 classification submission which won first place and significantly reduced the top-5 error to 2.251%, achieving a 25% relative improvement over the winning entry of 2016."}, "keywords": ["detector", "channels"], "citation_intent": "background"}
{"citing_id": "2303.01592v2", "cited_id": "1809.05231", "section_title": "Learning-Based Spherical Registration", "citation": "Semi-supervised methods employed additional information, like segmentation maps, to guide registration without requiring them during inference #REFR .", "text_before_citation": ["Recent learning-based registration methods can be categorized into supervised #OTHEREFR and unsupervised registration #OTHEREFR .", "Supervised registration predicts deformation fields minimizing the difference with a \"ground truth\" deformation.", "\"Ground truth\" maps are often generated using iterative methods or via simulation, which fundamentally limits the performance of supervised approaches #OTHEREFR .", "Unsupervised registration methods employ a classical loss evaluating image similarity and deformation regularity, thus forming an end-to-end training pipeline #OTHEREFR ."], "text_after_citation": ["Recent methods extend this strategy to the spherical domain by parameterizing the brain surfaces in a 2D grid that accounts for distortions #OTHEREFR or directly on the sphere using spherical kernels #OTHEREFR .", "Learning-based methods improved registration run time substantially during inference, while achieving superior or similar accuracy relative to iterative methods.", "However, these methods do not account for variations between geometry and function within a subject."], "citing_paper_content": {"title": "Joint Cortical Registration Of Geometry And Function Using Semi-Supervised Learning", "abstract": "Brain surface-based image registration, an important component of brain image analysis, establishes spatial correspondence between cortical surfaces. Existing iterative and learning-based approaches focus on accurate registration of folding patterns of the cerebral cortex, and assume that geometry predicts function and thus functional areas will also be well aligned. However, structure/functional variability of anatomically corresponding areas across subjects has been widely reported. In this work, we introduce a learning-based cortical registration framework, JOSA, which jointly aligns folding patterns and functional maps while simultaneously learning an optimal atlas. We demonstrate that JOSA can substantially improve registration performance in both anatomical and functional domains over existing methods. By employing a semi-supervised training strategy, the proposed framework obviates the need for functional data during inference, enabling its use in broad neuroscientific domains where functional data may not be observed."}, "cited_paper_content": {"title": "Voxelmorph: A Learning Framework For Deformable Medical Image Registration", "abstract": "We present VoxelMorph, a fast learning-based framework for deformable, pairwise medical image registration. Traditional registration methods optimize an objective function for each pair of images, which can be time-consuming for large datasets or rich deformation models. In contrast to this approach and building on recent learning-based methods, we formulate registration as a function that maps an input image pair to a deformation field that aligns these images. We parameterize the function via a convolutional neural network and optimize the parameters of the neural network on a set of images. Given a new pair of scans, VoxelMorph rapidly computes a deformation field by directly evaluating the function. In this paper, we explore two different training strategies. In the first (unsupervised) setting, we train the model to maximize standard image matching objective functions that are based on the image intensities. In the second setting, we leverage auxiliary segmentations available in the training data. We demonstrate that the unsupervised model\u2019s accuracy is comparable to the state-of-the-art methods while operating orders of magnitude faster. We also show that VoxelMorph trained with auxiliary data improves registration accuracy at test time and evaluate the effect of training set size on registration. Our method promises to speed up medical image analysis and processing pipelines while facilitating novel directions in learning-based registration and its applications. Our code is freely available at https://github.com/voxelmorph/voxelmorph ."}, "keywords": ["segmentation maps"], "citation_intent": "method"}
{"citing_id": "2304.04591v1", "cited_id": "1409.0575", "section_title": "B. Pre-Trained Vision Models", "citation": "In this study, we adopt the original MAE pre-trained on ImageNet #REFR for fair comparisons with other models.", "text_before_citation": ["MAE #OTHEREFR is pre-trained by the masked image modeling (MIM) task #OTHEREFR inspired by the success of masked language modeling in NLP #OTHEREFR . MAE adopts an asymmetric encoder-decoder architecture.", "The encoder is only applied to the visible subset of patches (without mask tokens).", "The lightweight decoder reconstructs the image from the encoded visible patches and mask tokens.", "The mask ratio is as high as 75% to reduce the heavy spatial redundancy of images.", "MVP #OTHEREFR ) uses an MAE model pre-trained on images in the wild, e.g., from YouTube #OTHEREFR or Egocentric videos #OTHEREFR to provide effective representations for motor control."], "text_after_citation": ["iBOT #OTHEREFR can be viewed as the combination of self-distillation (DINO) and masked image modeling.", "Self-distillation is used to train a teacher network/online tokenizer that captures high-level visual semantics.", "Masked image modeling is used to train a student network that can recover each masked patch token to its corresponding teacher network/online tokenizer output.", "CLIP #OTHEREFR ) is a simple and scalable method to learn visual representations with language supervision.", "CLIP forces the image representations to be aligned with paired captions through contrastive learning."], "citing_paper_content": {"title": "For Pre-Trained Vision Models In Motor Control, Not All Policy Learning Methods Are Created Equal", "abstract": "In recent years, increasing attention has been directed to leveraging pre-trained vision models for motor control. While existing works mainly emphasize the importance of this pre-training phase, the arguably equally important role played by downstream policy learning during controlspecific fine-tuning is often neglected. It thus remains unclear if pre-trained vision models are consistent in their effectiveness under different control policies. To bridge this gap in understanding, we conduct a comprehensive study on 14 pretrained vision models using 3 distinct classes of policy learning methods, including reinforcement learning (RL), imitation learning through behavior cloning (BC), and imitation learning with a visual reward function (VRF). Our study yields a series of intriguing results, including the discovery that the effectiveness of pre-training is highly dependent on the choice of the downstream policy learning algorithm. We show that conventionally accepted evaluation based on RL methods is highly variable and therefore unreliable, and further advocate for using more robust methods like VRF and BC. To facilitate more universal evaluations of pre-trained models and their policy learning methods in the future, we also release a benchmark of 21 tasks across 3 different environments alongside our work."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["ImageNet"], "citation_intent": "method"}
{"citing_id": "2303.02207v1", "cited_id": "1505.07427", "section_title": "A. Vo-Uncertainty By Univariate Cqr", "citation": "To test our approach, we use PoseNet #REFR with a ResNet34-based feature extractor for conformalization.", "text_before_citation": ["We employ conformalized quantile regression (CQR) #OTHEREFR to extract univariate uncertainty bands for each output coordinate in VO."], "text_after_citation": ["However, the presented techniques are generalizable to other feature-based and direct predictors for VO.", "The PoseNet model outputs translational elements x, y, and z, as well as rotational elements p, q, and r of angle w in radians.", "To reduce the number of response variables from seven to six, we convert the orientation quaternion (w, p, q, r) to Euler angles (roll \u03c6, pitch \u03b8, yaw \u03c8).", "After pre-training, we apply CQR to obtain prediction intervals along each one-dimensional variable, which reveals the predictive uncertainty. Fig. 2 depicts the process for this method.", "The main goal of CQR is to produce prediction intervals where the probability of true realizations falling within these intervals is guaranteed to be equivalent to or greater than some chosen coverage rate p, i.e., P{Y \u2208 C(X) | X = x} \u2265 p."], "citing_paper_content": {"title": "Lightweight, Uncertainty-Aware Conformalized Visual Odometry", "abstract": "Data-driven visual odometry (VO) is a critical subroutine for autonomous edge robotics, and recent progress in the field has produced highly accurate point predictions in complex environments. However, emerging autonomous edge robotics devices like insect-scale drones and surgical robots lack a computationally efficient framework to estimate VO's predictive uncertainties. Meanwhile, as edge robotics continue to proliferate into mission-critical application spaces, awareness of model's the predictive uncertainties has become crucial for risk-aware decision-making. This paper addresses this challenge by presenting a novel, lightweight, and statistically robust framework that leverages conformal inference (CI) to extract VO's uncertainty bands. Our approach represents the uncertainties using flexible, adaptable, and adjustable prediction intervals that, on average, guarantee the inclusion of the ground truth across all degrees of freedom (DOF) of pose estimation. We discuss the architectures of generative deep neural networks for estimating multivariate uncertainty bands along with point (mean) prediction. We also present techniques to improve the uncertainty estimation accuracy, such as leveraging Monte Carlo dropout (MC-dropout) for data augmentation. Finally, we propose a novel training loss function that combines interval scoring and calibration loss with traditional training metrics-mean-squared error and KLdivergence-to improve uncertainty-aware learning. Our simulation results demonstrate that the presented framework consistently captures true uncertainty in pose estimations across different datasets, estimation models, and applied noise types, indicating its wide applicability."}, "cited_paper_content": {"title": "Posenet: A Convolutional Network For Real-Time 6-Dof Camera Relocalization", "abstract": "We present a robust and real-time monocular six degree of freedom relocalization system. Our system trains a convolutional neural network to regress the 6-DOF camera pose from a single RGB image in an end-to-end manner with no need of additional engineering or graph optimisation. The algorithm can operate indoors and outdoors in real time, taking 5ms per frame to compute. It obtains approximately 2m and 3 degrees accuracy for large scale outdoor scenes and 0.5m and 5 degrees accuracy indoors. This is achieved using an efficient 23 layer deep convnet, demonstrating that convnets can be used to solve complicated out of image plane regression problems. This was made possible by leveraging transfer learning from large scale classification data. We show that the PoseNet localizes from high level features and is robust to difficult lighting, motion blur and different camera intrinsics where point based SIFT registration fails. Furthermore we show how the pose feature that is produced generalizes to other scenes allowing us to regress pose with only a few dozen training examples."}, "keywords": ["PoseNet"], "citation_intent": "method"}
{"citing_id": "2303.02970v1", "cited_id": "1805.08206", "section_title": "Concluding Remarks", "citation": "To have a unitless performance measure that can be applied across models, Geifman et al., #REFR introduce a normalized AURC, named Excess-AURC (E-AURC).", "text_before_citation": ["Finally, we propose to enlarge the confidence gap by finding flat minima, which yields strong performance on extensive experiments. AURC & E-AURC.", "AURC measures the area under the curve drawn by plotting the risk according to coverage.", "The coverage indicates the ratio of samples whose confidence estimates are higher than some confidence threshold, and the risk, also known as the selective risk #OTHEREFR , is an error rate computed by using those samples.", "A low value of AURC implies that correct and incorrect predictions can be well-separable by confidence estimates associated with samples.", "Inherently, AURC is affected by the predictive performance of a model."], "text_after_citation": ["Specifically, E-AURC can be computed by subtracting the optimal AURC, the lowest possible value for a given model, from the empirical AURC."], "citing_paper_content": {"title": "Rethinking Confidence Calibration For Failure Prediction", "abstract": "Reliable confidence estimation for the predictions is important in many safety-critical applications. However, modern deep neural networks are often overconfident for their incorrect predictions. Recently, many calibration methods have been proposed to alleviate the overconfidence problem. With calibrated confidence, a primary and practical purpose is to detect misclassification errors by filtering out low-confidence predictions (known as failure prediction). In this paper, we find a general, widely-existed but actually-neglected phenomenon that most confidence calibration methods are useless or harmful for failure prediction. We investigate this problem and reveal that popular confidence calibration methods often lead to worse confidence separation between correct and incorrect samples, making it more difficult to decide whether to trust a prediction or not. Finally, inspired by the natural connection between flat minima and confidence separation, we propose a simple hypothesis: flat minima is beneficial for failure prediction. We verify this hypothesis via extensive experiments and further boost the performance by combining two different flat minima techniques. Our code is available at https://github.com/Impression2805/FMFP."}, "cited_paper_content": {"title": "Bias-Reduced Uncertainty Estimation For Deep Neural Classifiers", "abstract": "We consider the problem of uncertainty estimation in the context of (non-Bayesian) deep neural classification. In this context, all known methods are based on extracting uncertainty signals from a trained network optimized to solve the classification problem at hand. We demonstrate that such techniques tend to introduce biased estimates for instances whose predictions are supposed to be highly confident. We argue that this deficiency is an artifact of the dynamics of training with SGD-like optimizers, and it has some properties similar to overfitting. Based on this observation, we develop an uncertainty estimation algorithm that selectively estimates the uncertainty of highly confident points, using earlier snapshots of the trained model, before their estimates are jittered (and way before they are ready for actual classification). We present extensive experiments indicating that the proposed algorithm provides uncertainty estimates that are consistently better than all known methods."}, "keywords": ["normalized AURC", "models"], "citation_intent": "method"}
{"citing_id": "2303.13401v1", "cited_id": "2003.01690", "section_title": "Numerical Optimization For Max-Loss Form", "citation": "Square Attack does not have as effective attack performance as APGDs (see Table 2 in later sections) but is included as #REFR states that the diversity in the evaluation algorithms is important to achieve reliable numerical RE.", "text_before_citation": ["For p = 2, sequential projection-first onto the box and then the norm ball, at least finds a feasible solution (see our clarification of these projections in Appendix A.1). Therefore, PGD is viable for these cases.", "For other general p metrics and nonp metrics, analytical projectors mostly cannot be derived and existing PGD methods do not apply.", "Previous work has shown that the solution quality of PGD is sensitive to hyperparameter tuning, e.g., step-size schedule and iteration budget #OTHEREFR .", "PGD variants, such as AGPD methods where 'A' stands for Auto(matic) #OTHEREFR , try to make the tuning process automatic by combining a heuristic adaptive step-size schedule and momentum acceleration under a fixed iteration budget and are built into the popular AutoAttack package #OTHEREFR .", "AutoAttack also includes non-PGD methods, e.g., Square Attack #OTHEREFR , which is based on gradientfree random search."], "text_after_citation": [], "citing_paper_content": {"title": "Optimization And Optimizers For Adversarial Robustness", "abstract": "Empirical robustness evaluation (RE) of deep learning models against adversarial perturbations entails solving nontrivial constrained optimization problems. Existing numerical algorithms that are commonly used to solve them in practice predominantly rely on projected gradient, and mostly handle perturbations modeled by the 1, 2 and \u221e distances. In this paper, we introduce a novel algorithmic framework that blends a general-purpose constrained-optimization solver PyGRANSO with Constraint-Folding (PWCF), which can add more reliability and generality to the state-of-the-art RE packages, e.g., AutoAttack. Regarding reliability, PWCF provides solutions with stationarity measures and feasibility tests to assess the solution quality. For generality, PWCF can handle perturbation models that are typically inaccessible to the existing projected gradient methods; the main requirement is the distance metric to be almost everywhere differentiable. Taking advantage of PWCF and other existing numerical algorithms, we further explore the distinct patterns in the solutions found for solving these optimization problems using various combinations of losses, perturbation models, and optimization algorithms. We then discuss the implications of these patterns on the current robustness evaluation and adversarial training."}, "cited_paper_content": {"title": "Reliable Evaluation Of Adversarial Robustness With An Ensemble Of Diverse Parameter-Free Attacks", "abstract": "The field of defense strategies against adversarial attacks has significantly grown over the last years, but progress is hampered as the evaluation of adversarial defenses is often insufficient and thus gives a wrong impression of robustness. Many promising defenses could be broken later on, making it difficult to identify the state-of-the-art. Frequent pitfalls in the evaluation are improper tuning of hyperparameters of the attacks, gradient obfuscation or masking. In this paper we first propose two extensions of the PGD-attack overcoming failures due to suboptimal step size and problems of the objective function. We then combine our novel attacks with two complementary existing ones to form a parameter-free, computationally affordable and user-independent ensemble of attacks to test adversarial robustness. We apply our ensemble to over 40 models from papers published at recent top machine learning and computer vision venues. In all except one of the cases we achieve lower robust test accuracy than reported in these papers, often by more than $10\\%$, identifying several broken defenses."}, "keywords": ["Square Attack"], "citation_intent": "background"}
{"citing_id": "2304.05260v1", "cited_id": "1909.06335", "section_title": "Settings", "citation": "For lower values of \u03b1, as client distributions become more skewed, there can be significant changes in accuracy between training runs #REFR ).", "text_before_citation": ["We use SGD as our optimizer, with weight decay of 1 \u00d7 10 \u22124 following Yao et al. 2021; #OTHEREFR .", "In these experiments, we observe that FedAvg often (though not always) performs better with group normalization as indicated by #OTHEREFR", "(2020) while FedAvg+WSM is able to perform well with both group and batch normalization, very frequently achieving the best results with batchnormalization.", "We therefore treat the normalization method as a hyperparameter and provide the best results for both batch and group normalization for both methods at each learning rate.", "Validation Throughout the training process the global model is periodically evaluated on the aggregation of the client validation sets to gauge overall training progress, the test set on the other hand is only used at the end of the training process."], "text_after_citation": ["Since we focus our analysis on the highly heterogeneous case in which the Dirichlet distributions at each client are parameterized by \u03b1 = 0.1, we observe higher variance in our results.", "This is especially true for smaller datasets such as CIFAR-10 and CIFAR-100.", "To mitigate these effects on the validation statistics, we follow the lead of and report our final accuracy as the average of the test accuracies taken over the last 100 rounds of training.", "0.92 0.9 0.7 0.92 0.66 0.9 0.9 0.9 0.9 0.8 0.9 0.88 0.76 0.88 0.68 0.9 0.82 0.9 0.9 0.86 0.9 0.86 0.38 0.94 0.64 0.92 0.88 0.82 0.94 0.64 0.9 0.9 0.64 0.9 0.68 0.9 0.92 0.88 0.94 0.74 0.9 0.88 0.58 0.96 0.", "The right column show the accuracy of client i's model after training when evaluated on client k's data."], "citing_paper_content": {"title": "Re-Weighted Softmax Cross-Entropy To Control Forgetting In Federated Learning", "abstract": "In Federated Learning, a global model is learned by aggregating model updates computed at a set of independent client nodes, to reduce communication costs multiple gradient steps are performed at each node prior to aggregation. A key challenge in this setting is data heterogeneity across clients resulting in differing local objectives which can lead clients to overly minimize their own local objective, diverging from the global solution. We demonstrate that individual client models experience a catastrophic forgetting with respect to data from other clients and propose an efficient approach that modifies the cross-entropy objective on a per-client basis by re-weighting the softmax logits prior to computing the loss. This approach shields classes outside a client's label set from abrupt representation change and we empirically demonstrate it can alleviate client forgetting and provide consistent improvements to standard federated learning algorithms. Our method is particularly beneficial under the most challenging federated learning settings where data heterogeneity is high and client participation in each round is low."}, "cited_paper_content": {"title": "Measuring The Effects Of Non-Identical Data Distribution For Federated Visual Classification", "abstract": "Federated Learning enables visual models to be trained in a privacy-preserving way using real-world data from mobile devices. Given their distributed nature, the statistics of the data across these devices is likely to differ significantly. In this work, we look at the effect such non-identical data distributions has on visual classification via Federated Learning. We propose a way to synthesize datasets with a continuous range of identicalness and provide performance measures for the Federated Averaging algorithm. We show that performance degrades as distributions differ more, and propose a mitigation strategy via server momentum. Experiments on CIFAR-10 demonstrate improved classification performance over a range of non-identicalness, with classification accuracy improved from 30.1% to 76.9% in the most skewed settings."}, "keywords": ["training runs", "accuracy"], "citation_intent": "background"}
{"citing_id": "2304.03433v1", "cited_id": "1506.03236", "section_title": "Vi. Numerical Results", "citation": "The K min values obtained from the numerical evaluations are compared with the theoretical result in #REFR , which substantiates the theoretical derivation for K min . In addition, the results in Fig.", "text_before_citation": ["It should also be noted that the minimum DEP becomes larger with the more number of cooperative users, i.e., K = 25, which results in a higher interference power and ends up making Willie more confused.", "Now, we numerically evaluate the minimum detection error probabilities, \u03b6 min for different numbers of cooperative users, K between 1 and 100 setting P a to either 0.67 or 0.83.", "The numerical evaluations are obtained by searching for the minimums of the empirical DEP for K values. The evaluations of minimum DEPs are depicted in Fig.", "3 where the minimum number of cooperative users, K min to meet the covert constraint, \u03b6 \u2265 1 \u2212 , for P a = 0.67 and 0.83 are given by 28 and 43, respectively, when = 0.05.", "Note that K min is the value on abscissa corresponding to \u03b6 min = 1 \u2212 = 0.95 on the ordinate."], "text_after_citation": ["3 tell that the minimum DEP, \u03b6 min is in proportion to the number of cooperative users, K.", "The minimum number of cooperative users, K min also grows with the increasing transmit power, P a which is predicted by the closed-form expression of K min in #OTHEREFR .", "In Fig.", "4 , the throughput, \u03b7 = RP c versus the covert rate, R is depicted for = 0.030 and 0.025 values when M = 500, and P a = 0.83. The numerical results in Fig.", "4 are simply given by the product of the empirical connection probability and the covert rate."], "citing_paper_content": {"title": "Multi-User Cooperation For Covert Communication Under Quasi-Static Fading", "abstract": "This work studies a covert communication scheme for an uplink multiuser scenario in which some users are opportunistically selected to help a covert user. In particular, the selected users emit interfering signals via an orthogonal resource dedicated to the covert user together with signals for their own communications using orthogonal resources allocated to the selected users, which helps the covert user hide the presence of the covert communication. For the covert communication scheme, we carry out extensive analysis and find system parameters in closed forms. The analytic derivation for the system parameters allow one to find the optimal combination of system parameters by performing a simple one-dimensional search. In addition, the analytic results elucidate relations among the system parameters. In particular, it will be proved that the optimal strategy for the non-covert users is an on-off scheme with equal transmit power. The theoretical results derived in this work are confirmed by comparing them with numerical results obtained with exhaustive searches. Finally, we demonstrate that the results of work can be utilized in versatile ways by demonstrating a design of covert communication with energy efficiency into account."}, "cited_paper_content": {"title": "Fundamental Limits Of Communication With Low Probability Of Detection", "abstract": "This paper considers the problem of communication over a discrete memoryless channel (DMC) or an additive white Gaussian noise (AWGN) channel subject to the constraint that the probability that an adversary who observes the channel outputs can detect the communication is low. Specifically, the relative entropy between the output distributions when a codeword is transmitted and when no input is provided to the channel must be sufficiently small. For a DMC whose output distribution induced by the\"off\"input symbol is not a mixture of the output distributions induced by other input symbols, it is shown that the maximum amount of information that can be transmitted under this criterion scales like the square root of the blocklength. The same is true for the AWGN channel. Exact expressions for the scaling constant are also derived."}, "keywords": ["K min values", "theoretical result"], "citation_intent": "result"}
{"citing_id": "2305.01461v1", "cited_id": "1802.09477", "section_title": "C. Td3Aq Algorithm", "citation": "To address this issue, we combine DDAQ and TD3 #REFR to form TD3AQ. Three key modifications to improve the performance are introduced.", "text_before_citation": ["In previous work (P-DQN), the objective function was defined as the sum of all Q-values, which posed a problem when back propagating gradients: updating the weights in the Qnetwork during back propagation affected all weights in the network, including those not related to the action with the highest Q-value, potentially causing the algorithm to update in an incorrect direction.", "DDAQ addresses this issue by using the maximum Q-value as the objective function, ensuring that only the weights corresponding to the action a d with the highest Q-value are updated.", "The back propagation process of DDAQ is illustrated in Fig. 2 .", "2) Twin Delayed Deep Deterministic Actor-Q (TD3AQ): DDPG has shown potential in achieving good performance, but its robustness is limited due to its sensitivity to hyperparameters.", "This is mainly because DDPG tends to overestimate Q-values, which results in poor policy updates."], "text_after_citation": ["a) Target policy smoothing: a noise sampled from a normal distribution is added to the action generated by the target actor network.", "This technique acts as a regularizer that prevents the actor network from converging to a suboptimal solution.", "Specifically, the action was smoothed by adding a clipped noise, defined as follows:", "EQUATION", "where a c min and a c max are the lower and upper bounds of the continuous action, respectively. The noise is defined as:"], "citing_paper_content": {"title": "Mixed-Integer Optimal Control Via Reinforcement Learning: A Case Study On Hybrid Vehicle Energy Management", "abstract": "Many optimal control problems require the simultaneous output of continuous and discrete control variables. Such problems are usually formulated as mixed-integer optimal control (MIOC) problems, which are challenging to solve due to the complexity of the solution space. Numerical methods such as branchand-bound are computationally expensive and unsuitable for realtime control. This paper proposes a novel continuous-discrete reinforcement learning (CDRL) algorithm, twin delayed deep deterministic actor-Q (TD3AQ), for MIOC problems. TD3AQ combines the advantages of both actor-critic and Q-learning methods, and can handle the continuous and discrete action spaces simultaneously. The proposed algorithm is evaluated on a hybrid electric vehicle (HEV) energy management problem, where real-time control of the continuous variable engine torque and discrete variable gear ratio is essential to maximize fuel economy while satisfying driving constraints. Simulation results on different drive cycles show that TD3AQ can achieve nearoptimal solutions compared to dynamic programming (DP) and outperforms the state-of-the-art discrete RL algorithm Rainbow, which is adopted for MIOC by discretizing continuous actions into a finite set of discrete values."}, "cited_paper_content": {"title": "Addressing Function Approximation Error In Actor-Critic Methods", "abstract": "In value-based reinforcement learning methods such as deep Q-learning, function approximation errors are known to lead to overestimated value estimates and suboptimal policies. We show that this problem persists in an actor-critic setting and propose novel mechanisms to minimize its effects on both the actor and critic. Our algorithm takes the minimum value between a pair of critics to restrict overestimation and delays policy updates to reduce per-update error. We evaluate our method on the suite of OpenAI gym tasks, outperforming the state of the art in every environment tested."}, "keywords": ["DDAQ", "performance"], "citation_intent": "method"}
{"citing_id": "2303.05221v1", "cited_id": "1506.06201", "section_title": "Data Availability", "citation": "For a detailed tutorial on linear mixed models in the Bayesian setting, see chapter 5 of Nicenboim et al. (2023), or #REFR .", "text_before_citation": ["2023results, we report only regression path duration and first-pass regressions out from the pre-critical adverb and the critical verb; for full details of all experimental results, please see the original paper.", "The effects of animacy and subjecthood (coded as sum contrasts) were analyzed using Bayesian mixed-effects models.", "Subject and item were specified as random effects in the models, with a full variance-covariance matrix for subject and item random effects.", "The models were implemented with brms #OTHEREFR , an interface to Stan #OTHEREFR .", "Priors were mildly informative Gaussian distributions for the linear model coefficients (intercept and slopes) and mildly informative regularizing Lewandowski-Kurowicka-Joe (LKJ) priors #OTHEREFR for random effects correlation matrices; setting the LKJ prior's parameter \u03bd to 2 downweights extreme correlations like \u00b11."], "text_after_citation": ["The results in #OTHEREFR", "(2023) showed reading time patterns consistent with effects of subjecthood (syntactic interference) and effects of animacy (semantic interference).", "Figure 3 shows that on the pre-critical adverb, the effect of subjecthood shows longer regression path duration (RPD) and more first-pass regressions out for conditions that have a +subject distractor (95% credible intervals (CrIs): RPD [17, 63] ms, FPR [3, 11]%).", "Similarly, the effect of animacy shows longer regression-path duration and an increase in first-pass regressions out for conditions with animate distractors compared to conditions with inanimate distractors (95% CrIs: RPD [8, 57] ms, FPR [2, 8]%).", "The subjecthood \u00d7 animacy interaction in regression-path duration is centered on zero; for first-pass regressions, the interaction has a negative sign ([\u22127, 0]%)."], "citing_paper_content": {"title": "Seam: An Integrated Activation-Coupled Model Of Sentence Processing And Eye Movements In Reading", "abstract": "Models of eye-movement control during reading, developed largely within psychology, usually focus on visual, attentional, and motor processes but neglect postlexical language processing; by contrast, models of sentence comprehension processes, developed largely within psycholinguistics, generally focus only on postlexical language processes. We present a model that combines these two research threads, by integrating eye-movement control and sentence processing. Developing such an integrated model is extremely challenging and computationally demanding, but such an integration is an important step toward complete mathematical models of natural language comprehension in reading. We combine the SWIFT model of eye-movement control (Engbert et al.,"}, "cited_paper_content": {"title": "Bayesian Linear Mixed Models Using Stan: A Tutorial For Psychologists, Linguists, And Cognitive Scientists", "abstract": "With the arrival of the R packages nlme and lme4, linear mixed models (LMMs) have come to be widely used in experimentally-driven areas like psychology, linguistics, and cognitive science. This tutorial provides a practical introduction to fitting LMMs in a Bayesian framework using the probabilistic programming language Stan. We choose Stan (rather than WinBUGS or JAGS) because it provides an elegant and scalable framework for fitting models in most of the standard applications of LMMs. We ease the reader into fitting increasingly complex LMMs, first using a two-condition repeated measures self-paced reading study, followed by a more complex $2\\times 2$ repeated measures factorial design that can be generalized to much more complex designs."}, "keywords": ["linear mixed models"], "citation_intent": "background"}
{"citing_id": "2303.00304v3", "cited_id": "1808.10654", "section_title": "B.4. Training", "citation": "We have collected 200 random navigation trajectories on 86 (training 72 + validation 14) scenes from the Gibson #REFR dataset.", "text_before_citation": [], "text_after_citation": ["We first trained the pair of encoder and decoder, and then trained the F loc and F stop with the frozen parameters of the encoder.", "The neural networks used in RNR-Map are all trained with the same dataset.", "The data sample for each neural network is also created in a similar manner.", "We sample a trajectory from the dataset and select a subset of frames from the trajectory.", "The pose information of the subset is normalized with the first frame of the sampled subset, considering the first frame as the origin."], "citing_paper_content": {"title": "Renderable Neural Radiance Map For Visual Navigation", "abstract": "We propose a novel type of map for visual navigation, a renderable neural radiance map (RNR-Map), which is designed to contain the overall visual information of a 3D environment. The RNR-Map has a grid form and consists of latent codes at each pixel. These latent codes are embedded from image observations, and can be converted to the neural radiance field which enables image rendering given a camera pose. The recorded latent codes implicitly contain visual information about the environment, which makes the RNR-Map visually descriptive. This visual information in RNR-Map can be a useful guideline for visual localization and navigation. We develop localization and navigation frameworks that can effectively utilize the RNR-Map. We evaluate the proposed frameworks on camera tracking, visual localization, and image-goal navigation. Experimental results show that the RNR-Map-based localization framework can find the target location based on a single query image with fast speed and competitive accuracy compared to other baselines. Also, this localization framework is robust to environmental changes, and even finds the most visually similar places when a query image from a different environment is given. The proposed navigation framework outperforms the existing image-goal navigation methods in difficult scenarios, under odometry and actuation noises. The navigation framework shows 65.7% success rate in curved scenarios of the NRNS [23] dataset, which is an improvement of 18.6% over the current state-of-the-art. Project page: https: //rllab-snu.github.io/projects/RNR-Map/"}, "cited_paper_content": {"title": "Gibson Env: Real-World Perception For Embodied Agents", "abstract": "Developing visual perception models for active agents and sensorimotor control are cumbersome to be done in the physical world, as existing algorithms are too slow to efficiently learn in real-time and robots are fragile and costly. This has given rise to learning-in-simulation which consequently casts a question on whether the results transfer to real-world. In this paper, we are concerned with the problem of developing real-world perception for active agents, propose Gibson Virtual Environment for this purpose, and showcase sample perceptual tasks learned therein. Gibson is based on virtualizing real spaces, rather than using artificially designed ones, and currently includes over 1400 floor spaces from 572 full buildings. The main characteristics of Gibson are: I. being from the real-world and reflecting its semantic complexity, II. having an internal synthesis mechanism,\"Goggles\", enabling deploying the trained models in real-world without needing further domain adaptation, III. embodiment of agents and making them subject to constraints of physics and space."}, "keywords": ["200 random navigation", "Gibson dataset"], "citation_intent": "method"}
{"citing_id": "2304.01096v1", "cited_id": "1406.2661", "section_title": "Introduction", "citation": "First, we propose an evolutionary algorithm-focused adaptation of the adversarial generation framework recently made popular by #REFR .", "text_before_citation": ["Gradient-based optimization of Deep Neural Networks (LeCun, Bengio & Hinton, 2015; Schmidhuber, 2015), by virtue of its ability to leverage modern computational resources, has established itself as this decade's single most influential line of work in Artificial Intelligence.", "However, given the exponentially falling cost of computation, it seems rather likely that other techniques with similarly strong scaling properties will eventually become of interest.", "An increasingly strong case is being made for evolutionary computation to be one of these.", "Indeed, when provided sufficient computation, simple evolutionary algorithms paired with deep networks have begun to rival popular modern methods #OTHEREFR .", "In this work, we attempt to further this effort through three contributions."], "text_after_citation": ["Next, we revisit the dynamic complexification of neural networks, component key to the field's prior success #OTHEREFR .", "Finally, we present a highly scalable inter-process communication protocol, designed for evolutionary algorithms to better utilize modern computing systems.", "We find that these techniques, once provided adequate resources, are able on short time-frames to produce artificial behaviour almost indistinguishable from human subjects in a video game environment."], "citing_paper_content": {"title": "Evolving Artificial Neural Networks To Imitate Human Behaviour In Shinobi Iii : Return Of The Ninja Master Pr\u00e9sent\u00e9 Par", "abstract": "Our society is increasingly fond of computational tools. This phenomenon has greatly increased over the past decade following, among other factors, the emergence of a new Artificial Intelligence paradigm. Specifically, the coupling of two algorithmic techniques, Deep Neural Networks and Stochastic Gradient Descent, thrusted by an exponentially increasing computing capacity, has and is continuing to become a major asset in many modern technologies. However, as progress takes its course, some still wonder whether other methods could similarly or even more greatly benefit from these various hardware advances. In order to further this study, we delve in this thesis into Evolutionary Algorithms and their application to Dynamic Neural Networks, two techniques which despite enjoying many advantageous properties have yet to find their niche in contemporary Artificial Intelligence. We find that by elaborating new methods while exploiting strong computational resources, it becomes possible to develop strongly performing agents on a variety of benchmarks but also some other agents behaving very similarly to human subjects on the video game Shinobi III : Return of The Ninja Master, typical complex tasks previously out of reach for non-gradient-based optimization."}, "cited_paper_content": {"title": "Generative Adversarial Nets", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to \u00bd everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["evolutionary algorithm-focused adaptation", "adversarial generation framework"], "citation_intent": "method"}
{"citing_id": "2303.10349v1", "cited_id": "2004.02980", "section_title": "Pyramid Covariance Predictor", "citation": "We follow #REFR to predict the Cholesky decomposition of the covariance matrix in case of the illegal covariance matrix from direct regression. Details are further demonstrated in the Appendix.", "text_before_citation": ["However, as aforementioned, isotropic Gaussian functions may be too simple to accurately represent the real distributions of landmarks.", "To better model the probability distributions, we use an anisotropic Gaussian function to estimate the uncertainty:", "EQUATION", "Different from Eq.", "2, we use a full two-dimensional covariance matrix \u03a3 i rather than a single value \u03c3 i to represent the heatmap, thus introducing anisotropy into the distribution."], "text_after_citation": ["According to Cholesky Factorization Theorem #OTHEREFR , the target covariance matrix can be decomposed as:", "\u03a3 i = C i C i T (4)", "where C i has the form of", "a i 0 b i c i .", "In our work, we calculate a i , b i , and c i by utilizing the multi-scale features extracted by the encoder of U-Net architecture."], "citing_paper_content": {"title": "Uncertainty-Aware U-Net For Medical Landmark Detection", "abstract": "Heatmap-based methods play an important role in anatomical landmark detection. However, most current heatmap-based methods assume that the distributions of all landmarks are the same and the distribution of each landmark is isotropic, which may not be in line with reality. For example, the landmark on the jaw is more likely to be located along the edge and less likely to be located inside or outside the jaw. Manually annotating tends to follow similar rules, resulting in an anisotropic distribution for annotated landmarks, which represents the uncertainty in the annotation. To estimate the uncertainty, we propose a module named Pyramid Covariance Predictor to predict the covariance matrices of the target Gaussian distributions, which determine the distributions of landmarks and represent the uncertainty of landmark annotation. Specifically, the Pyramid Covariance Predictor utilizes the pyramid features extracted by the encoder of the backbone U-Net and predicts the Cholesky decomposition of the covariance matrix of the landmark location distribution. Experimental results show that the proposed Pyramid Covariance Predictor can accurately predict the distributions and improve the performance of anatomical landmark detection."}, "cited_paper_content": {"title": "Luvli Face Alignment: Estimating Landmarks' Location, Uncertainty, And Visibility Likelihood", "abstract": "Modern face alignment methods have become quite accurate at predicting the locations of facial landmarks, but they do not typically estimate the uncertainty of their predicted locations nor predict whether landmarks are visible. In this paper, we present a novel framework for jointly predicting landmark locations, associated uncertainties of these predicted locations, and landmark visibilities. We model these as mixed random variables and estimate them using a deep network trained with our proposed Location, Uncertainty, and Visibility Likelihood (LUVLi) loss. In addition, we release an entirely new labeling of a large face alignment dataset with over 19,000 face images in a full range of head poses. Each face is manually labeled with the ground-truth locations of 68 landmarks, with the additional information of whether each landmark is unoccluded, self-occluded (due to extreme head poses), or externally occluded. Not only does our joint estimation yield accurate estimates of the uncertainty of predicted landmark locations, but it also yields state-of-the-art estimates for the landmark locations themselves on multiple standard face alignment datasets. Our method's estimates of the uncertainty of predicted landmark locations could be used to automatically identify input images on which face alignment fails, which can be critical for downstream tasks."}, "keywords": ["covariance matrix", "direct regression"], "citation_intent": "method"}
{"citing_id": "2304.05634v1", "cited_id": "1908.05913", "section_title": "F. Adapting Sota Methods For Our Task", "citation": "CAER (Context Aware Emotion Recognition) #REFR is a deep Convolutional Network which consists of two stream encoding networks to separately extract the facial and context features which are fused using an adaptive fusion network.", "text_before_citation": ["It uses a joint embedding training approach which uses emotional keywords associated with a given image and aligns its learned text embedding (pretrained on massive text data) with image embedding extracted from a standard feature backbone (ResNet50).", "To adapt EmotionNet for our task, we used word2vec #OTHEREFR for extracting text embeddings and ResNet50 for frames.", "Since we use a video as input, the frame features are max-pooled to generate a single representation.", "We use the proposed embedding loss and provide the emotion labels as the keywords for joint embedding training.", "This learned ResNet50 is finetuned for multilabel emotion recognition where the individual frame features are max-pooled before passing to the logits layer."], "text_after_citation": ["Detections from our extended face tracks are used as inputs for the face encoding stream and the full video frame with masked faces was used as input to context encoding stream.", "Since CAER is designed to extract emotions from images we adapt it to videos by applying max-pooling over the fused features from both the streams to generate a single representation for a video.", "This adapted model is trained to predict multiple scene-level emotions.", "M2FNet #OTHEREFR is a transformer based model originally developed for Emotion Recognition in Conversations (ERC) and features a fusion-attention mechanism to modulate the attention given to each utterance considering the audio and visual features.", "As this model is designed for utterance emotion recognition we apply a max-pooling operation over the final outputs of fusion attention module to generate a feature representation for all the utterances in a video."], "citing_paper_content": {"title": "How You Feelin'? Learning Emotions And Mental States In Movie Scenes", "abstract": "Movie story analysis requires understanding characters' emotions and mental states. Towards this goal, we formulate emotion understanding as predicting a diverse and multi-label set of emotions at the level of a movie scene and for each character. We propose EmoTx, a multimodal Transformer-based architecture that ingests videos, multiple characters, and dialog utterances to make joint predictions. By leveraging annotations from the MovieGraphs dataset [74], we aim to predict classic emotions (e.g. happy, angry) and other mental states (e.g. honest, helpful). We conduct experiments on the most frequently occurring 10 and 25 labels, and a mapping that clusters 181 labels to 26. Ablation studies and comparison against adapted stateof-the-art emotion recognition approaches shows the effectiveness of EmoTx. Analyzing EmoTx's self-attention scores reveals that expressive emotions often look at character tokens while other mental states rely on video and dialog cues."}, "cited_paper_content": {"title": "Context-Aware Emotion Recognition Networks", "abstract": "Traditional techniques for emotion recognition have focused on the facial expression analysis only, thus providing limited ability to encode context that comprehensively represents the emotional responses. We present deep networks for context-aware emotion recognition, called CAER-Net, that exploit not only human facial expression but also context information in a joint and boosting manner. The key idea is to hide human faces in a visual scene and seek other contexts based on an attention mechanism. Our networks consist of two sub-networks, including two-stream encoding networks to seperately extract the features of face and context regions, and adaptive fusion networks to fuse such features in an adaptive fashion. We also introduce a novel benchmark for context-aware emotion recognition, called CAER, that is more appropriate than existing benchmarks both qualitatively and quantitatively. On several benchmarks, CAER-Net proves the effect of context for emotion recognition. Our dataset is available at this http URL."}, "keywords": ["Context Aware Emotion"], "citation_intent": "method"}
{"citing_id": "2303.01510v1", "cited_id": "1702.05638", "section_title": "Fact Verification", "citation": "Another popular type of unstructured evidence dataset often considered is metadata which offers information complementary to textual sources #REFR .", "text_before_citation": ["In general, fact verification is the task of assessing the authenticity of a claim supported by a validated corpus of documents.", "Fact-checking introduces objective references and external knowledge to fake news detection, which is more reliable than relying simply on news text and visual features to make judgments.", "Fact verification can also increase users' awareness of precaution, which helps them try to find fact-checking information when exposed to fake news.", "There are many types of fact-verified datasets, and many of them are uni-modal datasets. For example, Thorne et al.", "#OTHEREFR considered Wikipedia as the source of textual evidence and annotated the sentences that support or refute each claim."], "text_after_citation": ["Although most uni-modal datasets focus on unstructured evidence, structured knowledge has also been used.", "Some datasets consist of semi-structured data tables with the ability to convey information concisely and flexibly. Wang et al.", "#OTHEREFR extracted tables from scientific articles and required evidence selection in the form of cells selected from tables.", "Recently, fact verification has also begun to consider building multi-modal datasets to improve the accuracy of fake news detection.", "A dataset named Mocheg #OTHEREFR consisted of textual and visual evidence and used three labels: support, refute, and NEI."], "citing_paper_content": {"title": "Ino At Factify 2: Structure Coherence Based Multi-Modal Fact Verification", "abstract": "This paper describes our approach to the multi-modal fact verification (FACTIFY) challenge at AAAI2023. In recent years, with the widespread use of social media, fake news can spread rapidly and negatively impact social security. Automatic claim verification becomes more and more crucial to combat fake news. In fact verification involving multiple modal data, there should be a structural coherence between claim and document. Therefore, we proposed a structure coherence-based multi-modal fact verification scheme to classify fake news. Our structure coherence includes the following four aspects: sentence length, vocabulary similarity, semantic similarity, and image similarity. Specifically, CLIP and Sentence BERT are combined to extract text features, and ResNet50 is used to extract image features. In addition, we also extract the length of the text as well as the lexical similarity. Then the features were concatenated and passed through the random forest classifier. Finally, our weighted average F1 score has reached 0.8079, achieving 2nd place in FACTIFY2."}, "cited_paper_content": {"title": "A Stylometric Inquiry Into Hyperpartisan And Fake News", "abstract": "This paper reports on a writing style analysis of hyperpartisan (i.e., extremely one-sided) news in connection to fake news. It presents a large corpus of 1,627 articles that were manually fact-checked by professional journalists from BuzzFeed. The articles originated from 9 well-known political publishers, 3 each from the mainstream, the hyperpartisan left-wing, and the hyperpartisan right-wing. In sum, the corpus contains 299 fake news, 97% of which originated from hyperpartisan publishers. ::: We propose and demonstrate a new way of assessing style similarity between text categories via Unmasking---a meta-learning approach originally devised for authorship verification---, revealing that the style of left-wing and right-wing news have a lot more in common than any of the two have with the mainstream. Furthermore, we show that hyperpartisan news can be discriminated well by its style from the mainstream (F1=0.78), as can be satire from both (F1=0.81). Unsurprisingly, style-based fake news detection does not live up to scratch (F1=0.46). Nevertheless, the former results are important to implement pre-screening for fake news detectors."}, "keywords": ["textual sources"], "citation_intent": "background"}
{"citing_id": "2304.13153v1", "cited_id": "1801.07698", "section_title": "Comparison With State-Of-The-Art 3D Gans", "citation": "We also compute the identity similarity #REFR between the frontal view and side views of varying yaw angles under fixed lighting to evaluate the view consistency.", "text_before_citation": ["Except for EpiGRAF and EG3D, the real images are filtered with the matting network #OTHEREFR to remove the background. As shown in Tab.", "1, LumiGAN can achieve state-of-the-art FID and KID only marginally worse than EG3D and EG3D-noBG yet much better than the relightable alternative VoluxGAN.", "The difference between our lighting distribution and the unknown actual FFHQ lighting distribution is a probable cause of the slight regression from EG3D. Geometry Quality.", "The geometry of the generated faces is evaluated by comparing the output normal maps with references predicted from a pretrained facial normal predictor #OTHEREFR .", "We compute the normal maps' cosine similarity as well as the L1 distance of their Laplacians to measure the fidelity of the geometric details #OTHEREFR ."], "text_after_citation": ["As shown in Tab.", "1, LumiGAN yields more accurate normals than non-relightable approaches thanks to the stronger geometric cues induced by the geometry-dependent physically based lighting module (also see Sec. 4.4).", "An exception is VoluxGAN in the L1 Laplacian metric, which is likely biased for smooth inputs. Indeed, the qualitative comparison in Fig. 5 Figure 4 : Qualitative Comparison.", "We compare our generated 3D faces lit in various conditions with 2 baselines: Shade-GAN #OTHEREFR and VoluxGAN #OTHEREFR .", "ShadeGAN uses a Lambertian lighting model and is limited to simple directional lights with changing hues, while VoluxGAN and LumiGAN use the environment maps shown in the last row."], "citing_paper_content": {"title": "Lumigan: Unconditional Generation Of Relightable 3D Human Faces", "abstract": "Unsupervised learning of 3D human faces from unstructured 2D image data is an active research area. While recent works have achieved an impressive level of photorealism, they commonly lack control of lighting, which prevents the generated assets from being deployed in novel environments. To this end, we introduce LumiGAN, an unconditional Generative Adversarial Network (GAN) for 3D human faces with a physically based lighting module that enables relighting under novel illumination at inference time. Unlike prior work, LumiGAN can create realistic shadow effects using an efficient visibility formulation that is learned in a selfsupervised manner. LumiGAN generates plausible physical properties for relightable faces, including surface normals, diffuse albedo, and specular tint without any ground truth data. In addition to relightability, we demonstrate significantly improved geometry generation compared to state-ofthe-art non-relightable 3D GANs and notably better photorealism than existing relightable GANs."}, "cited_paper_content": {"title": "Arcface: Additive Angular Margin Loss For Deep Face Recognition", "abstract": "One of the main challenges in feature learning using Deep Convolutional Neural Networks (DCNNs) for large-scale face recognition is the design of appropriate loss functions that can enhance the discriminative power. Centre loss penalises the distance between deep features and their corresponding class centres in the Euclidean space to achieve intra-class compactness. SphereFace assumes that the linear transformation matrix in the last fully connected layer can be used as a representation of the class centres in the angular space and therefore penalises the angles between deep features and their corresponding weights in a multiplicative way. Recently, a popular line of research is to incorporate margins in well-established loss functions in order to maximise face class separability. In this paper, we propose an Additive Angular Margin Loss (ArcFace) to obtain highly discriminative features for face recognition. The proposed ArcFace has a clear geometric interpretation due to its exact correspondence to geodesic distance on a hypersphere. We present arguably the most extensive experimental evaluation against all recent state-of-the-art face recognition methods on ten face recognition benchmarks which includes a new large-scale image database with trillions of pairs and a large-scale video dataset. We show that ArcFace consistently outperforms the state of the art and can be easily implemented with negligible computational overhead. To facilitate future research, the code has been made available."}, "keywords": ["fixed lighting", "frontal view"], "citation_intent": "method"}
{"citing_id": "2304.07002v1", "cited_id": "1412.6980", "section_title": "Complexity Prediction", "citation": "The proposed neural network uses ADAM #REFR as the solver for weight optimization, along with a constant learning rate of 0.001.", "text_before_citation": ["The output is a one-hot-encoder vector that determines the probability of a word to be complex or simple, i.e, the vector [1 0] means the word is simple, while the vector [0 1] means the word is complex. The hidden layer employs ReLU as the activation function.", "The ReLU activation function used is defined as the positive part of its argument (Equation 1).", "We choose this activation function as it solves the vanishing gradient problem #OTHEREFR .", "The final layer contains 2 units for classifying the words into complex or simple. These units employ the Sigmoid activation function (Equation 2).", "The Sigmoid function is used to predict probabilities that exist in the range [0, 1], mapping the output of the model to the on-hot-encoder class representation."], "text_after_citation": ["The ADAM optimizer is a good choice for sparse gradients as it is invariant to their diagonal rescale #OTHEREFR .", "EQUATION"], "citing_paper_content": {"title": "Simplex: A Lexical Text Simplification Architecture", "abstract": "Text simplification (TS) is the process of generating easy-to-understand sentences from a given sentence or piece of text. The aim of TS is to reduce both the lexical (which refers to vocabulary complexity and meaning) and syntactic (which refers to the sentence structure) complexity of a given text or sentence without the loss of meaning or nuance. In this paper, we present SIMPLEX, a novel simplification architecture for generating simplified English sentences. To generate a simplified sentence, the proposed architecture uses either word embeddings (i.e., Word2Vec) and perplexity, or sentence transformers (i.e., BERT, RoBERTa, and GPT2) and cosine similarity. The solution is incorporated into a user-friendly and simple-to-use software. We evaluate our system using two metrics, i.e., SARI, and Perplexity Decrease. Experimentally, we observe that the transformer models outperform the other models in terms of the SARI score. However, in terms of Perplexity, the Word-Embeddings-based models achieve the biggest decrease. Thus, the main contributions of this paper are: (1) We propose a new Word Embedding and Transformer based algorithm for text simplification; (2) We design SIMPLEX-a modular novel text simplification system-that can provide a baseline for further research; and (3) We perform an in-depth analysis of our solution and compare our results with two state-of-the-art models, i.e., LightLS [19] and NTS-w2v [44]. We also make the code publicly available online."}, "cited_paper_content": {"title": "Adam: A Method For Stochastic Optimization", "abstract": "We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm."}, "keywords": ["proposed neural network", "constant learning rate"], "citation_intent": "method"}
{"citing_id": "2304.13147v1", "cited_id": "2004.01177", "section_title": "Ii. Related Work", "citation": "Approaches using a single network for object detection and tracking bounding boxes, such as CenterTrack #REFR , either work offline or lack the ability to reconnect long-range tracks.", "text_before_citation": ["Therefore, it is critical to include such examples in the training of the tracking model.", "In contrast, multitask methods train object detection and tracking networks jointly.", "This is achieved by extracting REID features from a shared backbone #OTHEREFR and extending the bounding box prediction head with a single object tracking capability to regress the object location from the previous frame to the current frame #OTHEREFR , #OTHEREFR .", "However, the best-performing methods on MOT benchmarks, and especially self-supervised learning (SSL) methods, treat object detection and REID feature estimation as two separate tasks, presumably since they can tailor the network architecture to each task and handle scale variations in the REID task by rescaling the image crops used for extracting tracking features #OTHEREFR .", "Nevertheless, the self-supervised tracking setting requires a pre-trained object detector."], "text_after_citation": ["The latter scenarios require more advanced tracking logic accompanied by memory.", "With the rise of query-based object detectors, such as DETR #OTHEREFR , #OTHEREFR , the tracking by query propagation paradigm #OTHEREFR , #OTHEREFR - #OTHEREFR has emerged.", "These methods learn to track queries to recall the same instance across different frames in online or offline settings.", "However, there are no identity-free approaches to train such models in the literature, which we could compare against our proposed method.", "Self-supervised MOT: Self-Supervised Multi-Object Tracking (SSMOT) methods learn the association of object proposals without annotated track identities."], "citing_paper_content": {"title": "Self-Supervised Multi-Object Tracking From Consistency Across Timescales", "abstract": "Self-supervised multi-object trackers have the potential to leverage the vast amounts of raw data recorded worldwide. However, they still fall short in re-identification accuracy compared to their supervised counterparts. We hypothesize that this deficiency results from restricting self-supervised objectives to single frames or frame pairs. Such designs lack sufficient visual appearance variations during training to learn consistent re-identification features. Therefore, we propose a training objective that learns re-identification features over a sequence of frames by enforcing consistent association scores across short and long timescales. Extensive evaluations on the BDD100K and MOT17 benchmarks demonstrate that our learned ReID features significantly reduce ID switches compared to other self-supervised methods, setting the new state of the art for selfsupervised multi-object tracking and even performing on par with supervised methods on the BDD100k benchmark."}, "cited_paper_content": {"title": "Tracking Objects As Points", "abstract": "Tracking has traditionally been the art of following interest points through space and time. This changed with the rise of powerful deep networks. Nowadays, tracking is dominated by pipelines that perform object detection followed by temporal association, also known as tracking-by-detection. In this paper, we present a simultaneous detection and tracking algorithm that is simpler, faster, and more accurate than the state of the art. Our tracker, CenterTrack, applies a detection model to a pair of images and detections from the prior frame. Given this minimal input, CenterTrack localizes objects and predicts their associations with the previous frame. That's it. CenterTrack is simple, online (no peeking into the future), and real-time. It achieves 67.3% MOTA on the MOT17 challenge at 22 FPS and 89.4% MOTA on the KITTI tracking benchmark at 15 FPS, setting a new state of the art on both datasets. CenterTrack is easily extended to monocular 3D tracking by regressing additional 3D attributes. Using monocular video input, it achieves 28.3% AMOTA@0.2 on the newly released nuScenes 3D tracking benchmark, substantially outperforming the monocular baseline on this benchmark while running at 28 FPS."}, "keywords": ["object detection"], "citation_intent": "background"}
{"citing_id": "2303.15190v1", "cited_id": "1810.04805", "section_title": "Background", "citation": "Bidirectional Encoder Representations from Transformers (BERT) models #REFR is a stack of n encoders from the Transformer architecture.", "text_before_citation": ["\u2022 Q (query) is a matrix that represents the input in which the attention mechanism focuses on.", "\u2022 K (key) is a matrix that represents the different elements in the input that the attention mechanism can attend to.", "\u2022 V (value) is a matrix that represents the output of the attention mechanism.", "\u2022 d k is the dimension of matrices Q and V and allows to stabilize the model during the training phase Hence, each head has its own set of parameters, allowing the model to learn different types of attention patterns.", "The attentions resulting from each of the heads are then concatenated and projected on a dense layer."], "text_after_citation": ["Each BERT layer contains h attention heads with its own set of weights, which have been learned during training.", "These weights determine how the model will attend to different parts of the input when making a prediction.", "In this way, words are related to each other even in the case of long-term dependency.", "BERT has been widely adopted and has achieved state-of-the-art performance on a variety of benchmarks such as GLUE #OTHEREFR for natural language understanding (NLU).", "One of the key features of BERT is its bidirectional nature."], "citing_paper_content": {"title": "", "abstract": "Attention mechanisms have played a crucial role in the development of complex architectures such as Transformers in natural language processing. However, Transformers remain hard to interpret and are considered as black-boxes. This paper aims to assess how attention coefficients from Transformers can help in providing interpretability. A new attention-based interpretability method called CLaSsification-Attention (CLS-A) is proposed. CLS-A computes an interpretability score for each word based on the attention coefficient distribution related to the part specific to the classification task within the Transformer architecture. A human-grounded experiment is conducted to evaluate and compare CLS-A to other interpretability methods. The experimental protocol relies on the capacity of an interpretability method to provide explanation in line with human reasoning. Experiment design includes measuring reaction times and correct response rates by human subjects. CLS-A performs comparably to usual interpretability methods regarding average participant reaction time and accuracy. The lower computational cost of CLS-A compared to other interpretability methods and its availability by design within the classifier make it particularly interesting. Data analysis also highlights the link between the probability score of a classifier prediction and adequate explanations. Finally, our work confirms the relevancy of the use of CLS-A and shows to which extent self-attention contains rich information to explain Transformer classifiers."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["Transformers (BERT) models", "Bidirectional Encoder Representations"], "citation_intent": "background"}
{"citing_id": "2304.09987v1", "cited_id": "2003.08934", "section_title": "Sparse And Dense Point Cloud Comparison.", "citation": "From the results, we can see, that for the Blender dataset #REFR case, the dense model is much better.", "text_before_citation": ["However, dense point clouds may not be required to achieve good reconstructions.", "In this section, we compare models trained on dense reconstructions and sparse reconstructions.", "Table 5 compares PSNR, SSIM metrics on two synethetic scenes from the Blender dataset #OTHEREFR and two real-world scenes from the mip-NeRF 360 dataset #OTHEREFR .", "We also show the number of vertices of the tetrahedra which is the size of the input point cloud, including the random points added at the beginning.", "Note, that in this experiment we only trained for 100, 000 iterations to save computation time."], "text_after_citation": ["This is to be expected since the number of points obtained from the sparse reconstruction on the Blender dataset is very small and our MLP does not have enough capacity to represent fine details.", "However, even in regions with zero point coverage, we are still able to provide at least some low-frequency data (c.f . Fig. 7 ).", "On the real-world scenes, the sparse point cloud model is almost on par with the dense one.", "The sparse reconstructions on the real-world dataset provide many more points compared to the synthetic dataset case and the coverage is sufficient.", "From these results, we conclude that on real-world data the sparse model can be sufficient to achieve a good performance."], "citing_paper_content": {"title": "Tetra-Nerf: Representing Neural Radiance Fields Using Tetrahedra", "abstract": "2) scene representation: dense tetrahedra 3) barycetric interpolation of features 4) shallow MLP predicts density and colour Figure 1. The input to Tetra-NeRF is a point cloud which is triangulated to get a set of tetrahedra used to represent the radiance field. Rays are sampled, and the field is queried. The barycentric interpolation is used to interpolate tetrahedra vertices, and the resulting features are passed to a shallow MLP to get the density and colours for volumetric rendering."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["dense model"], "citation_intent": "result"}
{"citing_id": "2304.07120v1", "cited_id": "2001.08909", "section_title": "5:", "citation": "Similarly, it is observed the proposed designs achieve better AoI performance than the equal time/bandwidth allocation schemes as in #REFR .", "text_before_citation": ["Moreover, it shows that the TDMA transmission scheme achieves lower AoI than the FDMA scheme.", "This is as expected because in the former case the SNR of the currently scheduled device can be maximized via dynamically adjusted IRS passive beamforming, which thus achieves lower PER for each device.", "It also should be noted that as the number of IRS elements increases, the performance gap between our proposed design and the benchmarks becomes smaller and then approaches zero.", "This is because as M increases, the passive beamforming gain and thus the SNR of each packet is sufficiently large for approximately achieving error-free transmission, which renders the optimization of time/bandwidth allocation unnecessary.", "Fig.3(b) shows the maximum average AoI versus the total bandwidth, B."], "text_after_citation": ["More interestingly, it shows that for the case when the total bandwidth is small, the proposed (c) The maximum average AoI versus the length of information bits.", "FDMA transmission scheme even outperforms the proposed TDMA one, i.e., the latter is not always more favourable as shown in Fig. 3(a) .", "This is because when frequency resource is extremely limited, optimizing the bandwidth allocation is more efficient than dynamically adjusting the passive beamforming for improving the performance of the worst user.", "However, as the bandwidth increases, such performance gain becomes marginal, while the time-varying passive beamforming associated with each scheduled user is more useful, and thus the proposed TDMA transmission scheme outperforms the FDMA one.", "Even more, as B increases up to sufficiently large (e.g., 1000 KHz as in the considered setup), both benchmark schemes achieve almost the same AoI performance as our proposed design."], "citing_paper_content": {"title": "Resource Allocation And Passive Beamforming For Irs-Assisted Urllc Systems", "abstract": "In this correspondence, we investigate an intelligent reflective surface (IRS) assisted downlink ultra-reliable and lowlatency communication (URLLC) system, where an access point (AP) sends short packets to multiple devices with the help of an IRS. Specifically, a performance comparison between the frequency division multiple access (FDMA) and time division multiple access (TDMA) is conducted for the considered system, from the perspective of average age of information (AoI). Aiming to minimize the maximum average AoI among all devices by jointly optimizing the resource allocation and passive beamforming. However, the formulated problem is difficult to solve due to the non-convex objective function and coupled variables. Thus, we propose an alternating optimization based algorithm by dividing the original problem into two sub-problems which can be efficiently solved. Simulation results show that TDMA can achieve lower AoI by exploiting the time-selective passive beamforming of IRS for maximizing the signal to noise ratio (SNR) of each device consecutively. Moreover, it also shows that as the length of information bits becomes sufficiently large as compared to the available bandwidth, the proposed FDMA transmission scheme becomes more favorable instead, due to the more effective utilization of bandwidth."}, "cited_paper_content": {"title": "Intelligent Reflecting Surface-Assisted Multiple Access With User Pairing: Noma Or Oma?", "abstract": "The integration of intelligent reflecting surface (IRS) to multiple access networks is a cost-effective solution for boosting spectrum/energy efficiency and enlarging network coverage/ connections. However, due to the new capability of IRS in reconfiguring the wireless propagation channels, it is fundamentally unknown which multiple access scheme is superior in the IRSassisted wireless network. In this letter, we pursue a theoretical performance comparison between non-orthogonal multiple access (NOMA) and orthogonal multiple access (OMA) in the IRSassisted downlink communication, for which the transmit power minimization problems are formulated under the discrete unitmodulus reflection constraint on each IRS element. We analyze the minimum transmit powers required by different multiple access schemes and compare them numerically, which turn out to not fully comply with the stereotyped superiority of NOMA over OMA in conventional systems without IRS. Moreover, to avoid the exponential complexity of the brute-force search for the optimal discrete IRS phase shifts, we propose a low-complexity solution to achieve near-optimal performance."}, "keywords": ["equal time/bandwidth allocation"], "citation_intent": "result"}
{"citing_id": "2303.00897v1", "cited_id": "1910.00189", "section_title": "Related Work", "citation": "Following this study #REFR , we review four Non-IID scenarios that widely exist in real-world applications. They are summarized below.", "text_before_citation": ["Non-IID Data Issue.", "FedAvg (McMahan et al., 2017) could solve the decentralized data optimization problem when these decentralized data are IID.", "However, in real-world federated learning, the data distributions are usually Non-IID.", "In this case, FedAvg would take more rounds to iterate the FL model under the Non-IID data setting.", "The mathematical analysis of this issue in FL is given by work #OTHEREFR , which reveals that the local updates from Non-IID clients are divergent. It consequently damages the convergence performance of FedAvg."], "text_after_citation": ["\u2022 Feature distribution skew : The marginal distribution of label P(x) varies across clients.", "In a handwriting recognition task, the same words written by different users may differ in stroke width, slant, etc.", "\u2022 Label distribution skew : The marginal distribution of label P(y) may vary across clients.", "For instance, clients' data samples are tied to geo-regions -kangaroos are only in Australia or zoos; for an online shopping website, customers only buy certain items with specific preferences.", "\u2022 Feature concept skew : The condition distributions P(x|y) varies across clients, and P(y) is the same."], "citing_paper_content": {"title": "Stochastic Clustered Federated Learning", "abstract": "Federated learning is a distributed learning framework that takes full advantage of private data samples kept on edge devices. In real-world federated learning systems, these data samples are often decentralized and Non-Independently Identically Distributed (Non-IID), causing divergence and performance degradation in the federated learning process. As a new solution, clustered federated learning groups federated clients with similar data distributions to impair the Non-IID effects and train a better model for every cluster. This paper proposes StoCFL, a novel clustered federated learning approach for generic Non-IID issues. In detail, StoCFL implements a flexible CFL framework that supports an arbitrary proportion of client participation and newly joined clients for a varying FL system, while maintaining a great improvement in model performance. The intensive experiments are conducted by using four basic Non-IID settings and a real-world dataset. The results show that StoCFL could obtain promising cluster results even when the number of clusters is unknown. Based on the client clustering results, models trained with StoCFL outperform baseline approaches in a variety of contexts."}, "cited_paper_content": {"title": "The Non-Iid Data Quagmire Of Decentralized Machine Learning", "abstract": "Many large-scale machine learning (ML) applications need to train ML models over decentralized datasets that are generated at different devices and locations. These decentralized datasets pose a fundamental challenge to ML because they are typically generated in very different contexts, which leads to significant differences in data distribution across devices/locations (i.e., they are not independent and identically distributed (IID)). In this work, we take a step toward better understanding this challenge, by presenting the first detailed experimental study of the impact of such non-IID data on the decentralized training of deep neural networks (DNNs). Our study shows that: (i) the problem of non-IID data partitions is fundamental and pervasive, as it exists in all ML applications, DNN models, training datasets, and decentralized learning algorithms in our study; (ii) this problem is particularly difficult for DNN models with batch normalization layers; and (iii) the degree of deviation from IID (the skewness) is a key determinant of the difficulty level of the problem. With these findings in mind, we present SkewScout, a system-level approach that adapts the communication frequency of decentralized learning algorithms to the (skew-induced) accuracy loss between data partitions. We also show that group normalization can recover much of the skew-induced accuracy loss of batch normalization."}, "keywords": ["four Non-IID scenarios", "real-world applications"], "citation_intent": "background"}
{"citing_id": "2304.04591v1", "cited_id": "1409.0575", "section_title": "Imitation Learning Through Behavior Cloning", "citation": "Interestingly, VICRegL is pre-trained purely on ImageNet #REFR , without robotic manipulation tasks in mind, while R3M is pre-trained on diverse human videos.", "text_before_citation": ["We now thoroughly evaluate the large suite of pre-trained vision models using BC.", "We train BC policies with varying numbers of expert demonstrations to study how dataset size impacts performance in the low data regime.", "The results in Figure 7 demonstrate that no individual pre-trained vision model can dominate all three environments. However, considering 21 tasks across all environments ( Fig. 7 rightmost) , R3M is the best-performing model. This echoes the finding of #OTHEREFR", "(2022) that R3M enables data-efficient behavior cloning for robotic manipulation.", "Furthermore, there is a model that follows R3M closely: VICRegL."], "text_after_citation": ["This lends credence to a dominant paradigm that advancements in vision can potentially be transferred directly to visuomotor control without extensive control-specific adaptation.", "With respect to architectures, iBOT is clearly the bestperforming model among ViTs, highlighting the benefit of combining masked image modeling with self-distillation.", "Another general observation is that ResNet-based models tend to outperform ViT-based models.", "This is in spite of the fact that model size and computation complexity tend to favor ViTs.", "We hypothesize this is because transformer-based architectures contain less visual prior knowledge."], "citing_paper_content": {"title": "For Pre-Trained Vision Models In Motor Control, Not All Policy Learning Methods Are Created Equal", "abstract": "In recent years, increasing attention has been directed to leveraging pre-trained vision models for motor control. While existing works mainly emphasize the importance of this pre-training phase, the arguably equally important role played by downstream policy learning during controlspecific fine-tuning is often neglected. It thus remains unclear if pre-trained vision models are consistent in their effectiveness under different control policies. To bridge this gap in understanding, we conduct a comprehensive study on 14 pretrained vision models using 3 distinct classes of policy learning methods, including reinforcement learning (RL), imitation learning through behavior cloning (BC), and imitation learning with a visual reward function (VRF). Our study yields a series of intriguing results, including the discovery that the effectiveness of pre-training is highly dependent on the choice of the downstream policy learning algorithm. We show that conventionally accepted evaluation based on RL methods is highly variable and therefore unreliable, and further advocate for using more robust methods like VRF and BC. To facilitate more universal evaluations of pre-trained models and their policy learning methods in the future, we also release a benchmark of 21 tasks across 3 different environments alongside our work."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["robotic manipulation tasks", "ImageNet"], "citation_intent": "method"}
{"citing_id": "2303.07096v1", "cited_id": "1711.06640", "section_title": "Comparisons With State-Of-The-Art Methods", "citation": "Consistent with the performance on VG, PE-Net also achieves competitive results on Models PredCls SGCls SGDet zs-R@50 / 100 zs-R@50 / 100 zs-R@50 / 100 Motifs #REFR 3 Table 3 . Comparison with the state-of-the-art methods on Open-Images V6.", "text_before_citation": ["2, our model outperforms the vanilla Motifs and VCTree by 15.53%, 5.40%, 3.49%, and 15.38%, 4.45%, 2.91% at zs-R@100 on PredCls, SGCls, and SGDet.", "Although TDE significantly improves the zero-shot performance by removing the effect of context bias, Motifs-TDE and VCTree-TDE are still surpassed by our PE-Net with 2.69%, 2.03%, 0.7%, and 3.29%, 2.53%, 0.4% on three tasks, respectively.", "We owe the strength to the Prototype-based Modeling of our PE-Net, which models entity and predicate in the semantic space, significantly improving the model's analogical reasoning capability on unseen relation triplets.", "Open Images.", "To verify the generality of our method on different datasets, we conduct experiments on Open Images and present the results in Tab. 3."], "text_after_citation": ["We adopt the same evaluation metric as in #OTHEREFR .", "The best and second best methods under each setting are marked according to formats.", "Open Images dataset.", "Specifically, our method exceeds the BGNN with a large margin of 2.7% on average at four metrics, and outperforms RU-Net by 1.2%, 2.5%, and 1.4% at wmAP rel , wmAP phr , and score wtd , respectively.", "It powerfully confirms PE-Net's generalization on handling relation recognition under different data distributions."], "citing_paper_content": {"title": "Prototype-Based Embedding Network For Scene Graph Generation", "abstract": "Current Scene Graph Generation (SGG) methods explore contextual information to predict relationships among entity pairs. However, due to the diverse visual appearance of numerous possible subject-object combinations, there is a large intra-class variation within each predicate category, e.g., \"man-eating-pizza, giraffe-eating-leaf\", and the severe inter-class similarity between different classes, e.g., \"man-holding-plate, man-eating-pizza\", in model's latent space. The above challenges prevent current SGG methods from acquiring robust features for reliable relation prediction. In this paper, we claim that the predicate's category-inherent semantics can serve as class-wise prototypes in the semantic space for relieving the challenges. To the end, we propose the Prototype-based Embedding Network (PE-Net), which models entities/predicates with prototype-aligned compact and distinctive representations and thereby establishes matching between entity pairs and predicates in a common embedding space for relation recognition. Moreover, Prototype-guided Learning (PL) is introduced to help PE-Net efficiently learn such entitypredicate matching, and Prototype Regularization (PR) is devised to relieve the ambiguous entity-predicate matching caused by the predicate's semantic overlap. Extensive experiments demonstrate that our method gains superior relation recognition capability on SGG, achieving new state-of-the-art performances on both Visual Genome and Open Images datasets. The codes are available at https://github.com/VL-Group/PENET."}, "cited_paper_content": {"title": "Neural Motifs: Scene Graph Parsing With Global Context", "abstract": "We investigate the problem of producing structured graph representations of visual scenes. Our work analyzes the role of motifs: regularly appearing substructures in scene graphs. We present new quantitative insights on such repeated structures in the Visual Genome dataset. Our analysis shows that object labels are highly predictive of relation labels but not vice-versa. We also find there are recurring patterns even in larger subgraphs: more than 50% of graphs contain motifs involving at least two relations. This analysis leads to a new baseline that is simple, yet strikingly powerful. While hardly considering the overall visual context of an image, it outperforms previous approaches. We then introduce Stacked Motif Networks, a new architecture for encoding global context that is crucial for capturing higher order motifs in scene graphs. Our best model for scene graph detection achieves a 7.3% absolute improvement in recall@50 (41% relative gain) over prior state-of-the-art."}, "keywords": ["PE-Net"], "citation_intent": "result"}
{"citing_id": "2303.10249v1", "cited_id": "1503.03832", "section_title": "Introduction", "citation": "We use a triplet loss #REFR to contrastively optimize the gap between positive and negative pairs based on the cosine distance over the learned deep features.", "text_before_citation": ["Ideally, a machine learning system can make use of all data that is available.", "As a related first step in this direction, we explore the feasibility of predicting information gleaned from 3D geometry using 2D projection images.", "Being able to do so would allow a) pooling datasets that drastically differ in image types or b) relating information from a cheaper 2D screening to more readily interpretable 3D quantities that are difficult for a human observer.", "We propose an image synthesis method for diverse modalities based on multimodal metric learning and k-NN regression.", "To learn the metric, we use image retrieval as the target task, which aims at embedding images such that matching pairs of different modalities are close in the embedding space."], "text_after_citation": ["In contrast to the typical learning process, we carefully design the training scheme to avoid interference when training with longitudinal image data.", "Given the learned embedding, we can synthesize images between diverse image types by k-NN regression through a weighted average based on their distances measured in the embedding space.", "Given a large database, this strategy allows for a quick and simple estimation of one image type from another.", "We use knee osteoarthritis as the driving medical problem and evaluate our proposed approach using the OAI image data.", "Specifically, we predict cartilage thickness maps obtained from 3D MR images using 2D radiographs."], "citing_paper_content": {"title": "Mris: A Multi-Modal Retrieval Approach For Image Synthesis On Diverse Modalities", "abstract": "Multiple imaging modalities are often used for disease diagnosis, prediction, or population-based analyses. However, not all modalities might be available due to cost, different study designs, or changes in imaging technology. If the differences between the types of imaging are small, data harmonization approaches can be used; for larger changes, direct image synthesis approaches have been explored. In this paper, we develop an approach based on multi-modal metric learning to synthesize images of diverse modalities. We use metric learning via multi-modal image retrieval, resulting in embeddings that can relate images of different modalities. Given a large image database, the learned image embeddings allow us to use k-nearest neighbor (k-NN) regression for image synthesis. Our driving medical problem is knee osteoarthritis (KOA), but our developed method is general after proper image alignment. We test our approach by synthesizing cartilage thickness maps obtained from 3D magnetic resonance (MR) images using 2D radiographs. Our experiments show that the proposed method outperforms direct image synthesis and that the synthesized thickness maps retain information relevant to downstream tasks such as progression prediction and Kellgren-Lawrence grading (KLG). Our results suggest that retrieval approaches can be used to obtain high-quality and meaningful image synthesis results given large image databases."}, "cited_paper_content": {"title": "Facenet: A Unified Embedding For Face Recognition And Clustering", "abstract": "Despite significant recent advances in the field of face recognition [10, 14, 15, 17], implementing face verification and recognition efficiently at scale presents serious challenges to current approaches. In this paper we present a system, called FaceNet, that directly learns a mapping from face images to a compact Euclidean space where distances directly correspond to a measure of face similarity. Once this space has been produced, tasks such as face recognition, verification and clustering can be easily implemented using standard techniques with FaceNet embeddings as feature vectors."}, "keywords": ["learned deep features"], "citation_intent": "method"}
{"citing_id": "2304.13013v1", "cited_id": "1804.04235", "section_title": "Preliminaries And Related Work", "citation": "Moreover, in contrast with Shazeer and Stern #REFR we find that an out-of-date second moment estimator is primarily an issue for the (patch) embedding layer, and measure how well loss spikes are predicted by this event.", "text_before_citation": ["In our work we investigate instabilities which arise during CLIP training.", "Unlike the instabilities observed in #OTHEREFR , we find these are not caused by attention entropy collapse.", "Instead, our results indicate that spikes arise when the second moment estimator is out of date for the networks early layers.", "While our analysis and methods build directly on Shazeer and Stern #OTHEREFR (AdaFactor), there are important differences.", "In contrast with Shazeer and Stern #OTHEREFR , who only observe instabilities without warmup, we observe instabilities despite a long warmup period."], "text_after_citation": ["Finally, we note that reserachers have moved away from AdaFactor in its original formulation for large-scale training #OTHEREFR , finding AdaFactor to under-perform AdamW #OTHEREFR .", "We believe this is due to the factored second moment or absence of first moment.", "This is why our focus is AdamW #OTHEREFR which is the de facto standard optimizer for transformers."], "citing_paper_content": {"title": "Stable And Low-Precision Training For Large-Scale Vision-Language Models", "abstract": "We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models. 1) Towards accelerating training, we introduce SwitchBack , a linear layer for int8 quantized training which provides a speed-up of 13-25% while matching the performance of bfloat16 training within 0.1 percentage points for the 1B parameter CLIP ViT-Huge-the largest int8 training to date. Our main focus is int8 as GPU support for float8 is rare, though we also analyze float8 training through simulation. While SwitchBack proves effective for float8, we show that standard techniques are also successful if the network is trained and initialized so that large feature magnitudes are discouraged, which we accomplish via layer-scale initialized with zeros. 2) Towards stable training, we analyze loss spikes and find they consistently occur 1-8 iterations after the squared gradients become underestimated by their AdamW second moment estimator. As a result, we recommend an AdamW-Adafactor hybrid, which we refer to as StableAdamW because it avoids loss spikes when training a CLIP ViT-Huge model and outperforms gradient clipping."}, "cited_paper_content": {"title": "Adafactor: Adaptive Learning Rates With Sublinear Memory Cost", "abstract": "In several recently proposed stochastic optimization methods (e.g. RMSProp, Adam, Adadelta), parameter updates are scaled by the inverse square roots of exponential moving averages of squared past gradients. Maintaining these per-parameter second-moment estimators requires memory equal to the number of parameters. For the case of neural network weight matrices, we propose maintaining only the per-row and per-column sums of these moving averages, and estimating the per-parameter second moments based on these sums. We demonstrate empirically that this method produces similar results to the baseline. Secondly, we show that adaptive methods can produce larger-than-desired updates when the decay rate of the second moment accumulator is too slow. We propose update clipping and a gradually increasing decay rate scheme as remedies. Combining these methods and dropping momentum, we achieve comparable results to the published Adam regime in training the Transformer model on the WMT 2014 English-German machine translation task, while using very little auxiliary storage in the optimizer. Finally, we propose scaling the parameter updates based on the scale of the parameters themselves."}, "keywords": ["loss spikes"], "citation_intent": "result"}
{"citing_id": "2303.12789v1", "cited_id": "2003.08934", "section_title": "Introduction", "citation": "The process is simple: capture a collection of images of a scene from varying viewpoints, reconstruct their camera parameters, and use the posed images to optimize a Neural Radiance Field #REFR .", "text_before_citation": ["With the emergence of efficient neural 3D reconstruction techniques, capturing a realistic digital representation of a real-world 3D scene has never been easier."], "text_after_citation": ["Due to its ease of use, we expect captured 3D content to gradually replace the traditional processes of manually-generated assets.", "Unfortunately, while the pipelines for turning a real scene into a 3D representation are relatively mature and accessible, many of the other necessary tools for the creation of 3D assets (e.g., those needed for editing 3D scenes) remain underdeveloped.", "Traditional processes for editing 3D models involve specialized tools and years of training in order to manually sculpt, extrude, and re-texture a given object.", "This process is made even more involved with the advent of neural representations, which often do not have explicit surfaces.", "This further motivates the need for 3D editing approaches designed for the modern era of 3D representations, particularly approaches that are similarly as accessible as the capture techniques themselves."], "citing_paper_content": {"title": "Instruct-Nerf2Nerf: Editing 3D Scenes With Instructions", "abstract": "NeRF Scene Instruct NeRF2NeRF \"Give him a cowboy hat\" \"Give him a mustache\" \"Make him bald\" \"Turn him into a clown\" \"As a bronze bust\" \"Turn him into Albert Einstein\" \"Turn his face into a skull\" \"Turn him into a Modigliani painting\" \"Turn him into Batman\" Edited NeRF Figure 1: Editing 3D scenes with Instructions. We propose Instruct-NeRF2NeRF, a method for consistent 3D editing of a NeRF scene using text-based instructions. Our method can accomplish a diverse collection of local and global scene edits."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["scene", "Neural Radiance Field"], "citation_intent": "method"}
{"citing_id": "2304.11593v1", "cited_id": "2002.07418", "section_title": "Related Work", "citation": "One of the closest works to ours is #REFR comprised of a two-system, a fuzzy rule controller that takes the represented human knowledge constraints and returns a preferred action and a refined module that tunes the suboptimal knowledge.", "text_before_citation": ["another regularizer), and the design is such that it is equivalent to evaluating some Boolean constraint formula using Weighted Model Counting (WMC) which counts the weights of the solutions to a propositional logic formula #OTHEREFR .", "Similar to #OTHEREFR , #OTHEREFR defines a non-negative loss function using fuzzy logic to incorporate logical constraints.", "This loss measures how far the output of the network is from the nearest satisfying solution. Our approach differs from both of these.", "We compile constraints and add them to the training process of an RL agent; our constraints are motivated by real-world physics, which are crucial for safety-critical domains.", "Unlike #OTHEREFR we do not rely on WMC to evaluate the constraints as they rely on SAT solvers, instead we designed a more suitable metric for constraint evaluation in sequential decision-making tasks."], "text_after_citation": ["The constraints are represented in fuzzy logic and allow for imprecise policy selection.", "During the constraint generation, they assume to have perfect knowledge of the state-space, and the constraints fully capture all aspects of the state-space, which is a strong assumption to hold.", "Our work differs from this: we do not assume to know the full-state space dynamics and do not use constraints to warm-start the policy; further, our constraints are much more general and expressive.", "To illustrate this generality, consider the rule from #OTHEREFR Rule: IF S 1 is M l 1 and S 2 is M l 2 and...", "and S k is M l k THEN Action is a j : where S i are variables that describe different parts of the state, M l i is the fuzzy rules corresponding to each S i and a j is the action taken."], "citing_paper_content": {"title": "System Iii: Learning With Domain Knowledge For Safety Constraints", "abstract": "Reinforcement learning agents naturally learn from extensive exploration. Exploration is costly and can be unsafe in safety-critical domains. This paper proposes a novel framework for incorporating domain knowledge to help guide safe exploration and boost sample efficiency. Previous approaches impose constraints, such as regularisation parameters in neural networks, that rely on large sample sets and often are not suitable for safety-critical domains where agents should almost always avoid unsafe actions. In our approach, called System III, which is inspired by psychologists' notions of the brain's System I and System II, we represent domain expert knowledge of safety in form of first-order logic. We evaluate the satisfaction of these constraints via p-norms in state vector space. In our formulation, constraints are analogous to hazards, objects, and regions of state that have to be avoided during exploration. We evaluated the effectiveness of the proposed method on OpenAI's Gym and Safety-Gym environments. In all tasks, including classic Control and Safety Games, we show that our approach results in safer exploration and sample efficiency."}, "cited_paper_content": {"title": "Kogun: Accelerating Deep Reinforcement Learning Via Integrating Human Suboptimal Knowledge", "abstract": "Reinforcement learning agents usually learn from scratch, which requires a large number of interactions with the environment. This is quite different from the learning process of human. When faced with a new task, human naturally have the common sense and use the prior knowledge to derive an initial policy and guide the learning process afterwards. Although the prior knowledge may be not fully applicable to the new task, the learning process is significantly sped up since the initial policy ensures a quick-start of learning and intermediate guidance allows to avoid unnecessary exploration. Taking this inspiration, we propose knowledge guided policy network (KoGuN), a novel framework that combines human prior suboptimal knowledge with reinforcement learning. Our framework consists of a fuzzy rule controller to represent human knowledge and a refine module to fine-tune suboptimal prior knowledge. The proposed framework is end-to-end and can be combined with existing policy-based reinforcement learning algorithm. We conduct experiments on both discrete and continuous control tasks. The empirical results show that our approach, which combines human suboptimal knowledge and RL, achieves significant improvement on learning efficiency of flat RL algorithms, even with very low-performance human prior knowledge."}, "keywords": ["fuzzy rule controller", "suboptimal knowledge"], "citation_intent": "background"}
{"citing_id": "2304.00933v1", "cited_id": "1711.09601", "section_title": "Improving Representations", "citation": "Here, we revisit MAS #REFR and replay to test how they influence the general representation, see Figure 3b .", "text_before_citation": ["Crucially, we also find that the benefit of dropout is different in nature from that of using augmentations.", "Dropout improves the current task performance, but not the general representation.", "This is also found in #OTHEREFR , where ImageNet models trained without dropout transfer better.", "Considering that dropout also results in faster forgetting (see Figure 4a) , we conclude that optimizing with augmentations, but without dropout is the preferable choice.", "Methods Continual learning methods can prevent forgetting to some extent, but only after a task is learned."], "text_after_citation": ["While MAS remembers early task representations (Section 5.1), we find that its final representation quality is of the same level as that of finetuning.", "This is in line with what is reported in both #OTHEREFR . Instead, Hu et al. report an advantage of MAS #OTHEREFR .", "This is possibly an effect of their trained tasks (ImageNet) overlapping more with their downstream tasks (e.g.", "classifying planes, cars and pets), as we found that MAS can indeed remember some overlapping information.", "Also shown in this plot are two baselines, single task and multi task training, to put these results in perspective."], "citing_paper_content": {"title": "Knowledge Accumulation In Continually Learned Representations And The Issue Of Feature Forgetting", "abstract": "By default, neural networks learn on all training data at once. When such a model is trained on sequential chunks of new data, it tends to catastrophically forget how to handle old data. In this work we investigate how continual learners learn and forget representations. We observe two phenomena: knowledge accumulation, i.e. the improvement of a representation over time, and feature forgetting, i.e. the loss of task-specific representations. To better understand both phenomena, we introduce a new analysis technique called task exclusion comparison. If a model has seen a task and it has not forgotten all the task-specific features, then its representation for that task should be better than that of a model that was trained on similar tasks, but not that exact one. Our image classification experiments show that most task-specific features are quickly forgotten, in contrast to what has been suggested in the past. Further, we demonstrate how some continual learning methods, like replay, and ideas from representation learning affect a continually learned representation. We conclude by observing that representation quality is tightly correlated with continual learning performance."}, "cited_paper_content": {"title": "Memory Aware Synapses: Learning What (Not) To Forget", "abstract": "Humans can learn in a continuous manner. Old rarely utilized knowledge can be overwritten by new incoming information while important, frequently used knowledge is prevented from being erased. In artificial learning systems, lifelong learning so far has focused mainly on accumulating knowledge over tasks and overcoming catastrophic forgetting. In this paper, we argue that, given the limited model capacity and the unlimited new information to be learned, knowledge has to be preserved or erased selectively. Inspired by neuroplasticity, we propose an online method to compute the importance of the parameters of a neural network, based on the data that the network is actively applied to, in an unsupervised manner. After learning a task, whenever a sample is fed to the network, we accumulate an importance measure for each parameter of the network, based on how sensitive the predicted output is to a change in this parameter. When learning a new task, changes to important parameters are penalized. We show that a local version of our method is a direct application of Hebb's rule in identifying the important connections between neurons. We test our method on a sequence of object recognition tasks and on the challenging problem of learning an embedding in a continuous manner. We show state of the art performance and the ability to adapt the importance of the parameters towards what the network needs (not) to forget, which may be different for different test conditions."}, "keywords": ["replay"], "citation_intent": "method"}
{"citing_id": "2304.00472v1", "cited_id": "1810.04805", "section_title": "Background", "citation": "Progress in this field has been driven by two major concepts: the Transformer neural network architecture and the application of transfer learning in training #REFR .", "text_before_citation": ["Our vision is inspired by recent advances in the domain of natural language processing (NLP)."], "text_after_citation": ["The Transformer has now become the standard architecture in NLP.", "One of its benefits is its suitability for parallelization w.r.t.", "previous approaches, which has enabled the creation of massive pre-trained LLMs #OTHEREFR .", "These models are pre-trained on tasks, such as predicting the next word in a sentence, for which large amounts of data are easily accessible.", "Although pre-training is costly, the models can then be adapted to a number of new tasks."], "citing_paper_content": {"title": "Querying Large Language Models With Sql [Vision", "abstract": "In many use-cases, information is stored in text but not available in structured data. However, extracting data from natural language text to precisely fit a schema, and thus enable querying, is a challenging task. With the rise of pre-trained Large Language Models (LLMs), there is now an effective solution to store and use information extracted from massive corpora of text documents. Thus, we envision the use of SQL queries to cover a broad range of data that is not captured by traditional databases by tapping the information in LLMs. To ground this vision, we present Galois, a prototype based on a traditional database architecture, but with new physical operators for querying the underlying LLM. The main idea is to execute some operators of the the query plan with prompts that retrieve data from the LLM. For a large class of SQL queries, querying LLMs returns well structured relations, with encouraging qualitative results. Preliminary experimental results make pre-trained LLMs a promising addition to the field of database systems, introducing a new direction for hybrid query processing. However, we pinpoint several research challenges that must be addressed to build a DBMS that exploits LLMs. While some of these challenges necessitate integrating concepts from the NLP literature, others offer novel research avenues for the DB community."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["Transformer neural network"], "citation_intent": "background"}
{"citing_id": "2305.00805v1", "cited_id": "2001.04721", "section_title": "Related Work", "citation": "Kim, Jeong, and Ko #REFR tried to improve the explainability by simplifying a trained deep random forest model.", "text_before_citation": ["Empirical successes have attracted a line of work to the theoretical analysis of deep forest.", "To start theoretical analysis of deep forests, Lyu, Yang, and Zhou #OTHEREFR prove a margin-based generalization bound for the additive cascade forest.", "For the consistency of deep forests, Arnould, Boyer, and Scornet #OTHEREFR prove a tight bound on the excess risk of two-layer centered random trees, which regards the second layer as a variance reducer.", "Assuming that the raw and new features are separated and independently used in two stages, Lyu, Chen, and Zhou #OTHEREFR prove that new features are easy to cause overfitting risk.", "Although deep forest has been widely used and studied, very little work has been done to help users understand the specific predictions made by deep forest models."], "text_after_citation": ["Although they select the most important paths in the component forests, going through all the selected paths layer by layer is still a big workload, hence the interpretability is limited.", "In this paper, we aim to explain the predictions made by the original trained deep forest, and the explanation is by directly telling how much the considered feature enlarges or reduces the predicted value for a given class.", "Explaining random forests Random forest has long been a successful machine learning method, especially for tabular data #OTHEREFR .", "A wealth of work has focused on explaining how each input feature functions in the model."], "citing_paper_content": {"title": "Interpreting Deep Forest Through Feature Contribution And Mdi Feature Importance", "abstract": "Deep forest is a non-differentiable deep model which has achieved impressive empirical success across a wide variety of applications, especially on categorical/symbolic or mixed modeling tasks. Many of the application fields prefer explainable models, such as random forests with feature contributions that can provide local explanation for each prediction, and Mean Decrease Impurity (MDI) that can provide global feature importance. However, deep forest, as a cascade of random forests, possesses interpretability only at the first layer. From the second layer on, many of the tree splits occur on the new features generated by the previous layer, which makes existing explanatory tools for random forests inapplicable. To disclose the impact of the original features in the deep layers, we design a calculation method with an estimation step followed by a calibration step for each layer, and propose our feature contribution and MDI feature importance calculation tools for deep forest. Experimental results on both simulated data and real world data verify the effectiveness of our methods."}, "cited_paper_content": {"title": "Interpretation And Simplification Of Deep Forest", "abstract": "This paper proposes a new method for interpreting and simplifying a black box model of a deep random forest (RF) using a proposed rule elimination. In deep RF, a large number of decision trees are connected to multiple layers, thereby making an analysis difficult. It has a high performance similar to that of a deep neural network (DNN), but achieves a better generalizability. Therefore, in this study, we consider quantifying the feature contributions and frequency of the fully trained deep RF in the form of a decision rule set. The feature contributions provide a basis for determining how features affect the decision process in a rule set. Model simplification is achieved by eliminating unnecessary rules by measuring the feature contributions. Consequently, the simplified model has fewer parameters and rules than before. Experiment results have shown that a feature contribution analysis allows a black box model to be decomposed for quantitatively interpreting a rule set. The proposed method was successfully applied to various deep RF models and benchmark datasets while maintaining a robust performance despite the elimination of a large number of rules."}, "keywords": ["forest model"], "citation_intent": "method"}
{"citing_id": "2303.09017v1", "cited_id": "1102.0264", "section_title": "A Quantum Switch With Ghz Entangled Control.", "citation": "We compare it to the local fraction for a GHZ experiment with the same measurements (right), as originally shown by Figure 1 At the angular resolution of 100 used by the plots, the minimum causally separable fraction of 41.2% is achieved at 0 = #REFR 25 and 1 = 3 5 .", "text_before_citation": ["The empirical model is causally inseparable for the space right above, with a causally separable fraction of~41.2% over its two causal completions.", "The two completions are induced by the causal orders total ( , , ) \u2228 discrete ( ) \u2228 discrete ( ) and total ( , , ) \u2228 discrete ( ) \u2228 discrete ( ), and their causaltopes are 386-dimensional:", "Causal fraction:~20.6% Causal fraction:~20.6%", "The empirical model becomes causally separable if the no-signalling constraint to Diane and Eve is dropped: the model is 100% supported by the spaces of input histories induced by causal orders total ( , , , , ) and total ( , , , , ), the two causal completions of total ({A, B} , , , ), with each causal completion supporting exactly 50% of the empirical model.", "Below we plot the causally separable fraction (left) and the local fraction (middle) for this empirical model as a function of the 0 and 1 measurement angles used by Charlie, Diane and Eve."], "text_after_citation": ["For comparison, the minimum local fraction for the empirical model is 0%, and it coincides with the minimum local fraction for the corresponding GHZ experiment.", "The local fraction for the empirical model, computed over all 1048576 causal functions (reduced to 32768 compatible with the empirical model support), coincides with the causally separable local fraction, computed over the 458752 causally separable functions (reduced to 16384 compatible with the empirical model support).", "To better understand the differences between the three landscapes above, below we plot the difference between causally separable fraction and local fraction for the empirical model (left), as well as the difference between local fraction for the GHZ experiment and local fraction for the empirical model (right)."], "citing_paper_content": {"title": "The Geometry Of Causality", "abstract": "We provide a unified operational framework for the study of causality, non-locality and contextuality, in a fully device-independent and theory-independent setting. This paper is the final instalment in a trilogy: spaces of input histories, our dynamical generalisation of causal orders, were introduced in \"The Combinatorics of Causality\"; the sheaf-theoretic treatment of causal distributions, known to us as empirical models, was detailed in \"The Topology of Causality\". We define causaltopes-our chosen portmanteau of \"causal polytopes\"-for arbitrary spaces of input histories and arbitrary choices of input contexts. We show that causaltopes are obtained by slicing simpler polytopes of conditional probability distributions with a set of causality equations, which we fully characterise. We provide efficient linear programs to compute the maximal component of an empirical model supported by any given sub-causaltope, as well as the associated causal fraction. We introduce a notion of causal separability relative to arbitrary causal constraints. We provide efficient linear programs to compute the maximal causally separable component of an empirical model, and hence its causally separable fraction, as the component jointly supported by certain sub-causaltopes. We study causal fractions and causal separability for several novel examples, including a selection of quantum switches with entangled or contextual control. In the process, we demonstrate the existence of \"causal contextuality\", a phenomenon where causal inseparability is clearly correlated to, or even directly implied by, non-locality and contextuality."}, "cited_paper_content": {"title": "The Sheaf-Theoretic Structure Of Non-Locality And Contextuality", "abstract": "We use the mathematical language of sheaf theory to give a unified treatment of non-locality and contextuality, in a setting that generalizes the familiar probability tables used in non-locality theory to arbitrary measurement covers; this includes Kochen\u2013Specker configurations and more. We show that contextuality, and non-locality as a special case, correspond exactly to obstructions to the existence of global sections. We describe a linear algebraic approach to computing these obstructions, which allows a systematic treatment of arguments for non-locality and contextuality. We distinguish a proper hierarchy of strengths of no-go theorems, and show that three leading examples\u2014due to Bell, Hardy and Greenberger, Horne and Zeilinger, respectively\u2014occupy successively higher levels of this hierarchy. A general correspondence is shown between the existence of local hidden-variable realizations using negative probabilities, and no-signalling; this is based on a result showing that the linear subspaces generated by the non-contextual and no-signalling models, over an arbitrary measurement cover, coincide. Maximal non-locality is generalized to maximal contextuality, and characterized in purely qualitative terms, as the non-existence of global sections in the support. A general setting is developed for the Kochen\u2013Specker-type results, as generic, model-independent proofs of maximal contextuality, and a new combinatorial condition is given, which generalizes the \u2018parity proofs\u2019 commonly found in the literature. We also show how our abstract setting can be represented in quantum mechanics. This leads to a strengthening of the usual no-signalling theorem, which shows that quantum mechanics obeys no-signalling for arbitrary families of commuting observables, not just those represented on different factors of a tensor product."}, "keywords": ["minimum causally separable"], "citation_intent": "result"}
{"citing_id": "2303.14329v1", "cited_id": "1811.11623", "section_title": "1) Surveillance:", "citation": "It is urgently required that law enforcement and counter-terrorism departments can realize the real-time tracking on the public threats #REFR .", "text_before_citation": ["In major public safety events (e.g., terrorist attacks in public transportation scenarios), law enforcement may want to track down the identified perpetrators #OTHEREFR .", "Across modern cities, a large amount of cameras are installed in public areas around us, including underground transportation networks, ground transportation networks, and air transportation networks."], "text_after_citation": ["By placing cameras in public places such as roads, public transport, retail stores, parks and libraries, relevant departments can help prevent, track and solve crime problems via video analytics technologies #OTHEREFR . Many works focused on the video analytics-based surveillance applications. For example, Yi et al.", "#OTHEREFR highlighted the low-latency video analytics requirements for public safety applications, such as counter-terrorism. Jain et al.", "#OTHEREFR focused on applications based on multiple cameras, like crowd control and spotlight search.", "To sum up, the \"real-time\" characteristic is one of the most essentially focused metrics in surveillance applications, where the edge computing technology can help boost the performance.", "2) Rescue applications: Different from surveillance applications, rescue applications need to not only detect, but also identify and track the corresponding objects."], "citing_paper_content": {"title": "Edge-Based Video Analytics: A Survey", "abstract": "Edge computing has been getting a momentum with ever-increasing data at the edge of the network. In particular, huge amounts of video data and their real-time processing requirements have been increasingly hindering the traditional cloud computing approach due to high bandwidth consumption and high latency. Edge computing in essence aims to overcome this hindrance by processing most video data making use of edge servers, such as small-scale on-premises server clusters, server-grade computing resources at mobile base stations and even mobile devices like smartphones and tablets; hence, the term edge-based video analytics. However, the actual realization of such analytics requires more than the simple, collective use of edge servers. In this paper, we survey state-of-the-art works on edge-based video analytics with respect to applications, architectures, techniques, resource management, security and privacy. We provide a comprehensive and detailed review on what works, what doesn't work and why. These findings give insights and suggestions for next generation edge-based video analytics. We also identify open issues and research directions."}, "cited_paper_content": {"title": "Large Scale Audio-Visual Video Analytics Platform For Forensic Investigations Of Terroristic Attacks", "abstract": "The forensic investigation of a terrorist attack poses a huge challenge to the investigative authorities, as several thousand hours of video footage need to be spotted. To assist law enforcement agencies (LEA) in identifying suspects and securing evidences, we present a platform which fuses information of surveillance cameras and video uploads from eyewitnesses. The platform integrates analytical modules for different input-modalities on a scalable architecture. Videos are analyzed according their acoustic and visual content. Specifically, Audio Event Detection is applied to index the content according to attack-specific acoustic concepts. Audio similarity search is utilized to identify similar video sequences recorded from different perspectives. Visual object detection and tracking are used to index the content according to relevant concepts. The heterogeneous results of the analytical modules are fused into a distributed index of visual and acoustic concepts to facilitate rapid start of investigations, following traits and investigating witness reports."}, "keywords": ["counter-terrorism departments"], "citation_intent": "background"}
{"citing_id": "2303.01589v1", "cited_id": "1801.04381", "section_title": "C. Results On Rocog-V2", "citation": "We fine-tune a MobileNet v2 #REFR using the annotated 70,124 frames extracted from raw video data.", "text_before_citation": ["We implement our methods on the RoCoG-v2 dataset in a preinstalled Ubuntu Linux 18.04 using Robot Operating System 2(ROS2) Galactics."], "text_after_citation": ["The fine-tuned MobileNet v2 #OTHEREFR is operated as the core detector in our auto zoom algorithm.", "We evaluate our method with Mobile Video Networks(MoViNets) #OTHEREFR .", "Specifically, we test our method with MoViNet A0, A2 and A3 on RB5.", "MoViNet A0 is the smallest model in MoViNets and has input size 172\u00d7172. A2 uses deeper network design with input 224\u00d7224. A3 is the larger network, with input 256\u00d7256.", "Also, we choose the stream version of the MoViNets due to lack of support for 3D conv on edge devices, as shown in Table I ."], "citing_paper_content": {"title": "Aztr: Aerial Video Action Recognition With Auto Zoom And Temporal Reasoning", "abstract": "We propose a novel approach for aerial video action recognition. Our method is designed for videos captured using UAVs and can run on edge or mobile devices. We present a learning-based approach that uses customized auto zoom to automatically identify the human target and scale it appropriately. This makes it easier to extract the key features and reduces the computational overhead. We also present an efficient temporal reasoning algorithm to capture the action information along the spatial and temporal domains within a controllable computational cost. Our approach has been implemented and evaluated both on the desktop with high-end GPUs and on the low power Robotics RB5 Platform for robots and drones. In practice, we achieve 6.1-7.4% improvement over SOTA in Top-1 accuracy on the RoCoG-v2 dataset, 8.3-10.4% improvement on the UAV-Human dataset and 3.2% improvement on the Drone Action dataset."}, "cited_paper_content": {"title": "Mobilenetv2: Inverted Residuals And Linear Bottlenecks", "abstract": "In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes. We also describe efficient ways of applying these mobile models to object detection in a novel framework we call SSDLite. Additionally, we demonstrate how to build mobile semantic segmentation models through a reduced form of DeepLabv3 which we call Mobile DeepLabv3. The MobileNetV2 architecture is based on an inverted residual structure where the input and output of the residual block are thin bottleneck layers opposite to traditional residual models which use expanded representations in the input an MobileNetV2 uses lightweight depthwise convolutions to filter features in the intermediate expansion layer. Additionally, we find that it is important to remove non-linearities in the narrow layers in order to maintain representational power. We demonstrate that this improves performance and provide an intuition that led to this design. Finally, our approach allows decoupling of the input/output domains from the expressiveness of the transformation, which provides a convenient framework for further analysis. We measure our performance on Imagenet classification, COCO object detection, VOC image segmentation. We evaluate the trade-offs between accuracy, and number of operations measured by multiply-adds (MAdd), as well as the number of parameters"}, "keywords": ["annotated 70,124 frames", "MobileNet v2"], "citation_intent": "method"}
{"citing_id": "2303.06669v1", "cited_id": "1803.05372", "section_title": "R(\u0391, C ) \u2286 R(\u0391, C), For Any Arc \u0391 \u2208 C \u2229 C .", "citation": "If the insertion of \u03b1 i splits an arc \u03c9 \u2208 C i\u22121 into \u03c9 1 and \u03c9 2 , but \u03c9 2 \u2208 C, then eliminate R(\u03c9 2 , C i ) by calling #REFR .", "text_before_citation": ["Given V l (C i ), 1 \u2264 i \u2264 n, we can delete R(\u03b1, C i ), if R(\u03b1, C i ) \u2286 VR(s \u03b1 , S \u03b1 )", ", where S \u03b1 \u2286 S C is the set of sites that define \u2202R(\u03b1, C i ), in expected time linear on |S \u03b1 |.", "There are two ways to use Proposition 23, if applicable: 1.", "Use it when necessary to maintain the invariant that C i encloses C (by deleting #OTHEREFR any auxiliary arc in C i\u22121 that blocks the insertion of \u03b1 i , thus, eliminating the case C i = C i\u22121 ). 2.", "Eliminate any auxiliary arc at the time of its creation."], "text_after_citation": ["The advantage of the latter is that Voronoi-like cycles become order-independent, therefore, backwards analysis becomes possible to establish the algorithm's time complexity.", "We give the backwards analysis argument on the concrete case of Section 6; the same type of argument, only more technical, can be derived for this abstract formulation as well."], "citing_paper_content": {"title": "Abstract Voronoi-Like Graphs: Extending Delaunay'S Theorem And Applications", "abstract": "Any system of bisectors (in the sense of abstract Voronoi diagrams) defines an arrangement of simple curves in the plane. We define Voronoi-like graphs on such an arrangement, which are graphs whose vertices are locally Voronoi. A vertex v is called locally Voronoi, if v and its incident edges appear in the Voronoi diagram of three sites. In a so-called admissible bisector system, where Voronoi regions are connected and cover the plane, we prove that any Voronoi-like graph is indeed an abstract Voronoi diagram. The result can be seen as an abstract dual version of Delaunay's theorem on (locally) empty circles. Further, we define Voronoi-like cycles in an admissible bisector system, and show that the Voronoi-like graph induced by such a cycle C is a unique tree (or a forest, if C is unbounded). In the special case where C is the boundary of an abstract Voronoi region, the induced Voronoi-like graph can be computed in expected linear time following the technique of [Junginger and Papadopoulou SOCG'18]. Otherwise, within the same time, the algorithm constructs the Voronoi-like graph of a cycle C on the same set (or subset) of sites, which may equal C or be enclosed by C. Overall, the technique computes abstract Voronoi (or Voronoi-like) trees and forests in linear expected time, given the order of their leaves along a Voronoi-like cycle. We show a direct application in updating a constraint Delaunay triangulation in linear expected time, after the insertion of a new segment constraint, simplifying upon the result of [Shewchuk and Brown CGTA 2015]."}, "cited_paper_content": {"title": "Deletion In Abstract Voronoi Diagrams In Expected Linear Time", "abstract": "Updating an abstract Voronoi diagram in linear time, after deletion of one site, has been an open problem for a long time. Similarly for various concrete Voronoi diagrams of generalized sites, other than points. In this paper we present a simple, expected linear-time algorithm to update an abstract Voronoi diagram after deletion. We introduce the concept of a Voronoi-like diagram, a relaxed version of a Voronoi construct that has a structure similar to an abstract Voronoi diagram, without however being one. Voronoi-like diagrams serve as intermediate structures, which are considerably simpler to compute, thus, making an expected linear-time construction possible. We formalize the concept and prove that it is robust under an insertion operation, thus, enabling its use in incremental constructions."}, "keywords": ["arc", "insertion"], "citation_intent": "background"}
{"citing_id": "2303.05938v1", "cited_id": "1711.07566", "section_title": "Representations Of Attention Encoder", "citation": "We obtain the part segmentation mask obtained by rendering the ground truth MANO hand mesh using a differentiable neural renderer #REFR .", "text_before_citation": ["For adaptive global representation learning, we generate heatmaps by adjusting the Gaussian kernel size K according to the bounding box size of the hand in data preparation for supervision (details in Supplementary Material).", "As the first representation of ACR, this map explicitly mitigates inter-dependencies between hands and serves as an attention mask for better global representation learning.", "Part Segmentation map: A p \u2208 R 33\u00d7H\u00d7W is learnt as a probabilistic segmentation volume.", "Each pixel on the volume is a channel of probability logits over 33 classes which consists of 1 background and 16 hand part classes for each hand corresponding to MANO model.", "Thus we have A h p \u2208 R 16\u00d7H\u00d7W ."], "text_after_citation": ["As the second representation of ACR, this map serves as an attention mask for part representation learning.", "Cross-hand Prior map:", "M c \u2208 R 218\u00d7H\u00d7W contains two maps, M h c \u2208 R 109\u00d7H\u00d7W", ".", "It is split into two sets of parameters which are MANO parameter \u03b8 \u2208 R 16\u00d76 , \u03b2 \u2208 R 10 and 3 camera parameters for cross hand inverse feature query."], "citing_paper_content": {"title": "Acr: Attention Collaboration-Based Regressor For Arbitrary Two-Hand Reconstruction", "abstract": "Reconstructing two hands from monocular RGB images is challenging due to frequent occlusion and mutual confusion. Existing methods mainly learn an entangled representation to encode two interacting hands, which are incredibly fragile to impaired interaction, such as truncated hands, separate hands, or external occlusion. This paper presents ACR (Attention Collaborationbased Regressor), which makes the first attempt to reconstruct hands in arbitrary scenarios. To achieve this, ACR explicitly mitigates interdependencies between hands and between parts by leveraging center and part-based attention for feature extraction. However, reducing interdependence helps release the input constraint while weakening the mutual reasoning about reconstructing the interacting hands. Thus, based on center attention, ACR also learns cross-hand prior that handle the interacting hands better. We evaluate our method on various types of hand reconstruction datasets. Our method significantly outperforms the best interacting-hand approaches on the Inter-Hand2.6M dataset while yielding comparable performance with the state-of-the-art single-hand methods on the Frei-Hand dataset. More qualitative results on in-the-wild and hand-object interaction datasets and web images/videos further demonstrate the effectiveness of our approach for arbitrary hand reconstruction. Our code is available at https://github.com/ZhengdiYu/Arbitrary-Hands-3D-Reconstruction."}, "cited_paper_content": {"title": "Neural 3D Mesh Renderer", "abstract": "For modeling the 3D world behind 2D images, which 3D representation is most appropriate? A polygon mesh is a promising candidate for its compactness and geometric properties. However, it is not straightforward to model a polygon mesh from 2D images using neural networks because the conversion from a mesh to an image, or rendering, involves a discrete operation called rasterization, which prevents back-propagation. Therefore, in this work, we propose an approximate gradient for rasterization that enables the integration of rendering into neural networks. Using this renderer, we perform single-image 3D mesh reconstruction with silhouette image supervision and our system outperforms the existing voxel-based approach. Additionally, we perform gradient-based 3D mesh editing operations, such as 2D-to-3D style transfer and 3D DeepDream, with 2D supervision for the first time. These applications demonstrate the potential of the integration of a mesh renderer into neural networks and the effectiveness of our proposed renderer."}, "keywords": ["hand mesh", "differentiable neural renderer"], "citation_intent": "method"}
{"citing_id": "2305.00241v1", "cited_id": "1811.01988", "section_title": "The Big-M Formulation", "citation": "However, it has been observed #REFR ) that this big-M formulation is not strong in the sense that its LP relaxation does not, in general, capture the convex hull of the graph of a given unit; see Figure 9 for an illustration.", "text_before_citation": ["Here, M l i,\u2212 and M l i,+ are data which must satisfy the inequalities", "M l i,\u2212 \u2264 min h l\u22121 \u2208D l\u22121 w l i h l\u22121 + b l i M l i,+ \u2265 max h l\u22121 \u2208D l\u22121 w l i h l\u22121 + b l i .", "This big-M formulation for ReLU-based networks has been used extensively in the literature #OTHEREFR .", "The big-M formulation is compact, with one binary variable and O(1) general inequality constraints for each neuron.", "Applied for each unit in the network, this leads to a MILP formulation with O( l\u2208L n l ) = O(Ln max ) binary variables and general inequality constraints, where n max = max l\u2208L n L ."], "text_after_citation": ["In fact, (Anderson et al., 2019, Example 2) observe that this LP relaxation can be arbitrarily bad, even in fixed input dimension.", "As MILP solvers often bound the objective function between the best feasible point and its tightest optimal continuous relaxation, a weak formulation can negatively impact performance, often substantially.", "It is worth dwelling on where this lack of strength comes from.", "If the input h l\u22121 is one dimensional, the big-M formulation is locally ideal #OTHEREFR : the extreme points of the linear programming relaxation (4a-4d) naturally satisfy the integrality constraints (4e).", "However, this fails to hold in the general Figure 9 : Left: The convex hull of a ReLU neuron (5), and Right: the convex relaxation offered by the big-M formulation (4) Adapted from #OTHEREFR multivariate input case."], "citing_paper_content": {"title": "When Deep Learning Meets Polyhedral Theory: A Survey", "abstract": "In the past decade, deep learning became the prevalent methodology for predictive modeling thanks to the remarkable accuracy of deep neural networks in tasks such as computer vision and natural language processing. Meanwhile, the structure of neural networks converged back to simpler representations based on piecewise constant and piecewise linear functions such as the Rectified Linear Unit (ReLU), which became the most commonly used type of activation function in neural networks. That made certain types of network structure-such as the typical fully-connected feedforward neural network-amenable to analysis through polyhedral theory and to the application of methodologies such as Linear Programming (LP) and Mixed-Integer Linear Programming (MILP) for a variety of purposes. In this paper, we survey the main topics emerging from this fast-paced area of work, which bring a fresh perspective to understanding neural networks in more detail as well as to applying linear optimization techniques to train, verify, and reduce the size of such networks."}, "cited_paper_content": {"title": "Strong Mixed-Integer Programming Formulations For Trained Neural Networks", "abstract": "We present an ideal mixed-integer programming (MIP) formulation for a rectified linear unit (ReLU) appearing in a trained neural network. Our formulation requires a single binary variable and no additional continuous variables beyond the input and output variables of the ReLU. We contrast it with an ideal \u201cextended\u201d formulation with a linear number of additional continuous variables, derived through standard techniques. An apparent drawback of our formulation is that it requires an exponential number of inequality constraints, but we provide a routine to separate the inequalities in linear time. We also prove that these exponentially-many constraints are facet-defining under mild conditions. Finally, we study network verification problems and observe that dynamically separating from the exponential inequalities (1) is much more computationally efficient and scalable than the extended formulation, (2) decreases the solve time of a state-of-the-art MIP solver by a factor of 7 on smaller instances, and (3) nearly matches the dual bounds of a state-of-the-art MIP solver on harder instances, after just a few rounds of separation and in orders of magnitude less time."}, "keywords": ["LP relaxation"], "citation_intent": "background"}
{"citing_id": "2304.05495v1", "cited_id": "1812.11446", "section_title": "B. Pre-Trained Initialization", "citation": "It is observed that when a deeper auxiliary network is used, the final accuracy loss for the entire model is small #REFR .", "text_before_citation": ["Existing research incorporates local error signals to train the device-side model using an auxiliary network #OTHEREFR , #OTHEREFR .", "The auxiliary network consists of a few lightweight hidden layers that map the output of the device-side model to the same dimensions of the ground truth labels.", "Although the local error signals eliminate the need to transfer global gradients from the server to devices, this approach adversely impacts accuracy since the device-side model trained by local error signals may be sub-optimal. We will empirically demonstrate this in Section IV-C.", "A similar accuracy loss is reported in the literature for greedy layer-wise learning.", "In this method, the use of a simple one-hidden layer auxiliary network results in accuracy loss when compared to end-to-end training that uses global loss calculated with the entire model #OTHEREFR ."], "text_after_citation": ["However, adding more layers to the auxiliary network will increase the computational burden on resource-constrained devices, thereby offsetting the performance gains obtained from partitioning the DNN.", "To address the above, ActionFed adopts pre-trained initialization of the device-side model (w C ).", "The device-side model is initialized with pre-trained weights (w C,k ), which are the partial weights of the entire model trained on a large dataset (e.g. ImageNet #OTHEREFR ).", "In addition, we empirically study the impact of local loss on generating activation outputs, which are used to train the server-side model.", "We identify that local loss is not a satisfactory criterion for representing a 'good' activation output for training the server-side model with pretrained initialization (Section IV-C)."], "citing_paper_content": {"title": "Communication Efficient Dnn Partitioning-Based Federated Learning", "abstract": "Efficiently running federated learning (FL) on resource-constrained devices is challenging since they are required to train computationally intensive deep neural networks (DNN) independently. DNN partitioning-based FL (DPFL) has been proposed as one mechanism to accelerate training where the layers of a DNN (or computation) are offloaded from the device to an edge server. However, this creates significant communication overheads since the activation and gradient need to be transferred between the device and the edge server during training. Current techniques reduce the communication introduced by DNN partitioning using local loss-based methods. We demonstrate that these methods adversely impact accuracy and ignore the communication costs incurred when transmitting the activation from the device to the server. This paper proposes ActionFed-a communication efficient framework for DPFL to accelerate training on resource-constrained devices. ActionFed eliminates the transmission of the gradient by developing pretrained initialization of the DNN model on the device for the first time. This reduces the accuracy degradation seen in local loss-based methods. In addition, ActionFed proposes a novel replay buffer mechanism and implements a quantizationbased compression technique to reduce the transmission of the activation. It is experimentally demonstrated that ActionFed can reduce the communication cost by up to 15.77x and accelerates training by up to 3.87x when compared to vanilla DPFL."}, "cited_paper_content": {"title": "Greedy Layerwise Learning Can Scale To Imagenet", "abstract": "Shallow supervised 1-hidden layer neural networks have a number of favorable properties that make them easier to interpret, analyze, and optimize than their deep counterparts, but lack their representational power. Here we use 1-hidden layer learning problems to sequentially build deep networks layer by layer, which can inherit properties from shallow networks. Contrary to previous approaches using shallow networks, we focus on problems where deep learning is reported as critical for success. We thus study CNNs on image classification tasks using the large-scale ImageNet dataset and the CIFAR-10 dataset. Using a simple set of ideas for architecture and training we find that solving sequential 1-hidden-layer auxiliary problems lead to a CNN that exceeds AlexNet performance on ImageNet. Extending this training methodology to construct individual layers by solving 2-and-3-hidden layer auxiliary problems, we obtain an 11-layer network that exceeds several members of the VGG model family on ImageNet, and can train a VGG-11 model to the same accuracy as end-to-end learning. To our knowledge, this is the first competitive alternative to end-to-end training of CNNs that can scale to ImageNet. We illustrate several interesting properties of these models theoretically and conduct a range of experiments to study the properties this training induces on the intermediate layers."}, "keywords": ["deeper auxiliary network"], "citation_intent": "background"}
{"citing_id": "2304.03400v1", "cited_id": "1807.01697", "section_title": "Steganography With Rosteals", "citation": "We use 14 noise sources from ImageNet-C #REFR , a rich and diverse library commonly used for evaluating model robustness.", "text_before_citation": ["The total loss is computed as;", "L = \u03b2L quality + L recovery (4)", "where loss weight \u03b2 controls the trade-off between stego quality and secret recovery. Training for robustness.", "Being robust to noises is a desirable property of both steganography and watermarking.", "We insert a noise model between the image decoder G and the secret decoder D."], "text_after_citation": ["These noises can be split into 3 groups: differentiable including most additive and linear noises (e.g. brightness, saturation, contrast, ...), approximatable with differentiable transforms (e.g. jpeg compression #OTHEREFR ), non-differentiable (e.g. spatter).", "To ensure gradient can be flown back to the preceding layers, we convert any non-differentiable noise n(.) to additive noise via", "n(x) = x + [n(x) \u2212 x] where [.]", "is treated as an additive constant (gradient disconnected from computation graph). Additionally, we apply multiple common noises (e.g.", "contrast, brightness, jpeg compression) concurrently at a certain rate (p = 0.5), along with random individual noises."], "citing_paper_content": {"title": "Rosteals: Robust Steganography Using Autoencoder Latent Space", "abstract": "Data hiding such as steganography and invisible watermarking has important applications in copyright protection, privacy-preserved communication and content provenance. Existing works often fall short in either preserving image quality, or robustness against perturbations or are too complex to train. We propose RoSteALS, a practical steganography technique leveraging frozen pretrained autoencoders to free the payload embedding from learning the distribution of cover images. RoSteALS has a lightweight secret encoder of just 300k parameters, is easy to train, has perfect secret recovery performance and comparable image quality on three benchmarks. Additionally, RoSteALS can be adapted for novel cover-less steganography applications in which the cover image can be sampled from noise or conditioned on text prompts via a denoising diffusion process. Our model and code are available at https://github.com/TuBui/RoSteALS."}, "cited_paper_content": {"title": "Benchmarking Neural Network Robustness To Common Corruptions And Surface Variations", "abstract": "In this paper we establish rigorous benchmarks for image classifier robustness. Our first benchmark, ImageNet-C, standardizes and expands the corruption robustness topic, while showing which classifiers are preferable in safety-critical applications. Then we propose a new dataset called ImageNet-P which enables researchers to benchmark a classifier's robustness to common perturbations. Unlike recent robustness research, this benchmark evaluates performance on common corruptions and perturbations not worst-case adversarial perturbations. We find that there are negligible changes in relative corruption robustness from AlexNet classifiers to ResNet classifiers. Afterward we discover ways to enhance corruption and perturbation robustness. We even find that a bypassed adversarial defense provides substantial common perturbation robustness. Together our benchmarks may aid future work toward networks that robustly generalize."}, "keywords": ["ImageNet-C", "model robustness"], "citation_intent": "method"}
{"citing_id": "2303.11910v2", "cited_id": "1709.06158", "section_title": "A. Data Generation", "citation": "To perform the data generation, we use an opensource tool 2 to convert the 3D mesh semantic labels in Matterport3D #REFR into 194,400 pinhole images with semantic labels.", "text_before_citation": [], "text_after_citation": ["Then, every 18 semantic label pairs are concatenated via a corresponding rotation-translation matrix, yielding 10,800 panoramic semantic ground truth, which is referred to as 360FV-Matterport by us.", "These panoramic semantic images are originally annotated with 40 object categories.", "Because many of them are only a small percentage ( 0.1%), we merges some uncommon classes and maintains the 20 most common object categories: wall, floor, chair, door, table, picture, furniture, objects, window, sofa, bed, sink, stairs, ceiling, toilet, mirror, shower, bathtub, counter, and shelving.", "For another front-view semantic segmentation dataset, Stan-ford2D3D #OTHEREFR , we keep the original object classes: beam, board, bookcase, ceiling, chair, clutter, column, door, floor, sofa, table, wall, window.", "For the presented 360BEV-Stanford dataset, we follow the data split method of Fold-1 of the Stanford2D3D #OTHEREFR dataset."], "citing_paper_content": {"title": "360Bev: Panoramic Semantic Mapping For Indoor Bird'S-Eye View", "abstract": "Seeing only a tiny part of the whole is not knowing the full circumstance. Bird's-eye-view (BEV) perception, a process of obtaining allocentric maps from egocentric views, is restricted when using a narrow Field of View (FoV) alone. In this work, mapping from 360\u00b0panoramas to BEV semantics, the 360BEV task, is established for the first time to achieve holistic representations of indoor scenes in a top-down view. Instead of relying on narrow-FoV image sequences, a panoramic image with depth information is sufficient to generate a holistic BEV semantic map. To benchmark 360BEV, we present two indoor datasets, 360BEV-Matterport and 360BEV-Stanford, both of which include egocentric panoramic images and semantic segmentation labels, as well as allocentric semantic maps. Besides delving deep into different mapping paradigms, we propose a dedicated solution for panoramic semantic mapping, namely 360Mapper. Through extensive experiments, our methods achieve 44.32% and 45.78% in mIoU on both datasets respectively, surpassing previous counterparts with gains of +7.60% and +9.70% in mIoU. 1"}, "cited_paper_content": {"title": "Matterport3D: Learning From Rgb-D Data In Indoor Environments", "abstract": "Access to large, diverse RGB-D datasets is critical for training RGB-D scene understanding algorithms. However, existing datasets still cover only a limited number of views or a restricted scale of spaces. In this paper, we introduce Matterport3D, a large-scale RGB-D dataset containing 10,800 panoramic views from 194,400 RGB-D images of 90 building-scale scenes. Annotations are provided with surface reconstructions, camera poses, and 2D and 3D semantic segmentations. The precise global alignment and comprehensive, diverse panoramic set of views over entire buildings enable a variety of supervised and self-supervised computer vision tasks, including keypoint matching, view overlap prediction, normal prediction from color, semantic segmentation, and region classification."}, "keywords": ["semantic labels", "3D mesh semantic"], "citation_intent": "method"}
{"citing_id": "2304.03400v1", "cited_id": "1512.03385", "section_title": "Steganography With Rosteals", "citation": "We use Resnet50 #REFR as the secret decoder D, replacing the last fully connected layer to output a L-bit secrets.", "text_before_citation": ["In total, our secret encoder F has just over 300K parameters for a 100-bit secret length.", "Interestingly, we empirically verify that conditioning F on both the cover image and secret (e.g. \u03b4 = F (x, s)) is not necessary (c.f. Figure 1 , see also Sup.Mat.).", "The stego image is then constructed asx = G(z + \u03b4) and regulated using a combination of pixel and perceptual losses;", "L M SE = ||\u03b3(x) \u2212 \u03b3(x)|| 2 (2) L quality = L LP IP S (x, x) + \u03b1L M SE (3)", "where \u03b3(.) is a differentiable non-parametric mapping function from RGB to the more perceptually uniform YUV space; L LP IP S refers to the LPIPS loss #OTHEREFR commonly used as evaluator for image quality and \u03b1 is a loss weight constant."], "text_after_citation": ["We use BCE to compute the bit recovery loss between the predicted and groundtruth secret,", "L recovery = L BCE (s,s).", "The total loss is computed as;", "L = \u03b2L quality + L recovery (4)", "where loss weight \u03b2 controls the trade-off between stego quality and secret recovery. Training for robustness."], "citing_paper_content": {"title": "Rosteals: Robust Steganography Using Autoencoder Latent Space", "abstract": "Data hiding such as steganography and invisible watermarking has important applications in copyright protection, privacy-preserved communication and content provenance. Existing works often fall short in either preserving image quality, or robustness against perturbations or are too complex to train. We propose RoSteALS, a practical steganography technique leveraging frozen pretrained autoencoders to free the payload embedding from learning the distribution of cover images. RoSteALS has a lightweight secret encoder of just 300k parameters, is easy to train, has perfect secret recovery performance and comparable image quality on three benchmarks. Additionally, RoSteALS can be adapted for novel cover-less steganography applications in which the cover image can be sampled from noise or conditioned on text prompts via a denoising diffusion process. Our model and code are available at https://github.com/TuBui/RoSteALS."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["secret decoder", "layer"], "citation_intent": "method"}
{"citing_id": "2303.01723v1", "cited_id": "1903.12216", "section_title": "I. Introduction", "citation": "Massive MIMO systems and high frequency communications at millimeter wave (mmWave) and sub-Thz bands are expected to play a key role in future sixth-generation (6G) networks #REFR .", "text_before_citation": [], "text_after_citation": ["These technologies are naturally supportive of each other, as massive MIMO using large transmit and receive antenna arrays facilitates generating highly focused beams that are essential for reliable communications at high frequencies, while short wavelength signaling enables packing MIMO configurations with a massive number of elements in a limited aperture.", "However, implementing such massive MIMO transceivers gives rise to several challenges.", "One of these core challenges is associated with the notable cost and power consumption of RF chains operating at high frequencies, which in conventional fully digital MIMO arrays separately connect each antenna element to the digital signal processing unit.", "Hybrid beamforming is considered to be a leading solution for coping with the above challenge, enabling high frequency massive MIMO communications with a limited number of RF chains #OTHEREFR .", "This is achieved by delegating part of the signal processing to the analog domain, thus, dividing the beamforming task into digital and analog counterparts."], "citing_paper_content": {"title": "Ai-Empowered Hybrid Mimo Beamforming", "abstract": "Hybrid multiple-input multiple-output (MIMO) is an attractive technology for realizing extreme massive MIMO systems envisioned for future wireless communications in a scalable and power-efficient manner. However, the fact that hybrid MIMO systems implement part of their beamforming in analog and part in digital makes the optimization of their beampattern notably more challenging compared with conventional fully digital MIMO. Consequently, recent years have witnessed a growing interest in using data-aided artificial intelligence (AI) tools for hybrid beamforming design. This article reviews candidate strategies to leverage data to improve real-time hybrid beamforming design. We discuss the architectural constraints and characterize the core challenges associated with hybrid beamforming optimization. We then present how these challenges are treated via conventional optimization, and identify different AI-aided design approaches. These can be roughly divided into purely data-driven deep learning models and different forms of deep unfolding techniques for combining AI with classical optimization. We provide a systematic comparative study between existing approaches including both numerical evaluations and qualitative measures. We conclude by presenting future research opportunities associated with the incorporation of AI in hybrid MIMO systems."}, "cited_paper_content": {"title": "Toward 6G Networks: Use Cases And Technologies", "abstract": "Reliable data connectivity is vital for the ever increasingly intelligent, automated, and ubiquitous digital world. Mobile networks are the data highways and, in a fully connected, intelligent digital world, will need to connect everything, including people to vehicles, sensors, data, cloud resources, and even robotic agents. Fifth generation (5G) wireless networks, which are currently being deployed, offer significant advances beyond LTE, but may be unable to meet the full connectivity demands of the future digital society. Therefore, this article discusses technologies that will evolve wireless networks toward a sixth generation (6G) and which we consider as enablers for several potential 6G use cases. We provide a fullstack, system-level perspective on 6G scenarios and requirements, and select 6G technologies that can satisfy them either by improving the 5G design or by introducing completely new communication paradigms."}, "keywords": ["Massive MIMO systems", "future sixth-generation (6G)"], "citation_intent": "background"}
{"citing_id": "2303.06586v1", "cited_id": "1908.10084", "section_title": "Ii. Preliminaries", "citation": "We also leverage a variation of Sentence-BERT #REFR architecture to learn sentence-level representations.", "text_before_citation": ["The unsupervised learning technique enabled the development of more robust NLP systems due to the abundance of textual data.", "We need to optimize the proposed framework using unsupervised representation learning techniques to handle distributional shifts #OTHEREFR , #OTHEREFR in a robust manner so that it can handle new unseen app issues accurately.", "The language models (e.g., GPT-2 #OTHEREFR , BERT #OTHEREFR , T5 #OTHEREFR ) are trained using unsupervised techniques by utilizing vast amounts of text data.", "These models can capture both general semantic and syntactic information effectively, due to the size of the model and training dataset.", "In this work, we employ T5, a pre-trained language model, as the backbone model in our framework."], "text_after_citation": ["Instead of BERT, we use a more powerful pre-trained language model T5 for learning sentence embeddings.", "The T5 model follows transformers #OTHEREFR architecture, contains both encoder and decoder stacks, and has produced state-of-the-art results on several NLP benchmarks, including SuperGLUE. T5 Model.", "Inspired by the idea of casting every NLP task as question answering #OTHEREFR , T5 treats every task as a text-totext problem.", "With a text-to-text framework, virtually every task can be addressed with the same model, objective, training process, and decoding process.", "T5 Adapts the original transformers architecture for multi-task learning, as it has both encoder and decoder stack as opposed to BERT (i.e., only encoder stack is more suitable for natural language understanding tasks) and GPT (i.e., only decoder stack is more appropriate for natural language generation tasks)."], "citing_paper_content": {"title": "Proactive Prioritization Of App Issues Via Contrastive Learning", "abstract": "Mobile app stores produce a tremendous amount of data in the form of user reviews, which is a huge source of user requirements and sentiments; such reviews allow app developers to proactively address issues in their apps. However, only a small number of reviews capture common issues and sentiments which creates a need for automatically identifying prominent reviews. Unfortunately, most existing work in text ranking and popularity prediction focuses on social contexts where other signals are available, which renders such works ineffective in the context of app reviews. In this work, we propose a new framework, PPrior, that enables proactive prioritization of app issues through identifying prominent reviews (ones predicted to receive a large number of votes in a given time window). Predicting highlyvoted reviews is challenging given that, unlike social posts, social network features of users are not available. Moreover, there is an issue of class imbalance, since a large number of user reviews receive little to no votes. PPrior employs a pre-trained T5 model and works in three phases. Phase one adapts the pretrained T5 model to the user reviews data in a self-supervised fashion. In phase two, we leverage contrastive training to learn a generic and task-independent representation of user reviews. Phase three uses radius neighbors classifier to make the final predictions. This phase also uses FAISS index for scalability and efficient search. To conduct extensive experiments, we acquired a large dataset of over 2.1 million user reviews from Google Play. Our experimental results demonstrate the effectiveness of the proposed framework when compared against several state-of-theart approaches. Moreover, the accuracy of PPrior in predicting prominent reviews is comparable to that of experienced app developers. Index Terms-App reviews, app analysis, automatic app issues prioritization."}, "cited_paper_content": {"title": "Sentence-Bert: Sentence Embeddings Using Siamese Bert-Networks", "abstract": "BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019) has set a new state-of-the-art performance on sentence-pair regression tasks like semantic textual similarity (STS). However, it requires that both sentences are fed into the network, which causes a massive computational overhead: Finding the most similar pair in a collection of 10,000 sentences requires about 50 million inference computations (~65 hours) with BERT. The construction of BERT makes it unsuitable for semantic similarity search as well as for unsupervised tasks like clustering. ::: In this publication, we present Sentence-BERT (SBERT), a modification of the pretrained BERT network that use siamese and triplet network structures to derive semantically meaningful sentence embeddings that can be compared using cosine-similarity. This reduces the effort for finding the most similar pair from 65 hours with BERT / RoBERTa to about 5 seconds with SBERT, while maintaining the accuracy from BERT. ::: We evaluate SBERT and SRoBERTa on common STS tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embeddings methods."}, "keywords": ["sentence-level representations", "Sentence-BERT architecture"], "citation_intent": "method"}
{"citing_id": "2304.12387v1", "cited_id": "1801.08914", "section_title": "4.1.", "citation": "Consistent with the results reported in #REFR , the algebraic hybridization solver greatly outperforms the ADS solver on all cases considered.", "text_before_citation": ["In particular, hybrid Gauss-Seidel smoothing is used on the CPU, while 1 -Jacobi smoothing is used on the GPU.", "This appears to have a more significant detrimental effect on the ADS solver; the iteration counts for the saddle-point solver are largely similar to the CPU case.", "The saddle-point solver consistently achieves 9-15\u00d7 speed-up compared with the LOR-ADS solver, with greater speed-up at higher orders.", "At p = 6, the total solve time using the saddle-point solver using the GPU on one node of Lassen is over twice as fast as on the CPU using 4 nodes of Quartz.", "For comparison, we also include CPU results for the matrix-based algebraic hybridization and ADS solvers in Table 4 ; GPU results are not reported because GPU-accelerated versions of these algorithms are not available."], "text_after_citation": ["Similarly, the LOR-ADS solver outperforms the ADS solver for p \u2265 2, as reported in #OTHEREFR , and consequently the saddle-point solver also compares favorably to matrix-based ADS for these test cases. The comparison with algebraic hybridization is more interesting.", "The hybridization solver (which performs a conjugate gradient iteration directly on the hybridized Schur complement system) requires significantly fewer iterations than either ADS or the saddle-point solver.", "For p = 2, hybridization results in the fastest overall runtime, and for p = 3, we obtained roughly equal runtimes with the hybridization and saddle-point solvers.", "However, for p > 3, the time required to assemble and set up the hybridized solver becomes prohibitive, and the saddle-point solver results in significant speed-ups.", "For p = 4, the saddle-point solver is almost twice as fast, and for p = 6, the saddle-point solver results in an almost 30\u00d7 speedup."], "citing_paper_content": {"title": "Matrix-Free Gpu-Accelerated Saddle-Point Solvers For High-Order Problems In H(Div)", "abstract": "This work describes the development of matrix-free GPU-accelerated solvers for high-order finite element problems in H(div). The solvers are applicable to grad-div and Darcy problems in saddle-point formulation, and have applications in radiation diffusion and porous media flow problems, among others. Using the interpolation-histopolation basis (cf. SIAM J. Sci. Comput., 45 (2023), A675-A702), efficient matrix-free preconditioners can be constructed for the (1, 1)-block and Schur complement of the block system. With these approximations, block-preconditioned MINRES converges in a number of iterations that is independent of the mesh size and polynomial degree. The approximate Schur complement takes the form of an M-matrix graph Laplacian, and therefore can be well-preconditioned by highly scalable algebraic multigrid methods. High-performance GPU-accelerated algorithms for all components of the solution algorithm are developed, discussed, and benchmarked. Numerical results are presented on a number of challenging test cases, including the \"crooked pipe\" grad-div problem, the SPE10 reservoir modeling benchmark problem, and a nonlinear radiation diffusion test case."}, "cited_paper_content": {"title": "Algebraic Hybridization And Static Condensation With Application To Scalable H(Div) Preconditioning", "abstract": "We propose a unified algebraic approach for the practical application and preconditioning of static condensation and hybridization, two popular techniques in finite element discretizations. We demo..."}, "keywords": ["algebraic hybridization"], "citation_intent": "result"}
{"citing_id": "2304.08915v1", "cited_id": "1902.03983", "section_title": "Introduction", "citation": "However, this enormous search space includes redundant expressions; the evolutionary process often generates unnecessarily large models with poor interpretability #REFR .", "text_before_citation": ["However, the inefficiency of GP evolutionary search approach highdimensional problems, which causes the above problem from the root, remains to be addressed.", "Another issue is that the interpretability of GP solutions suffers substantially on high-dimensional problems.", "GP often leads to very large expressions or chaining of nonlinear functions, which makes the expressions much harder to analyze #OTHEREFR .", "This problem is related to \"bloating\" in GP (where trees grow unnecessarily large without a corresponding increase in fitness), which is a common side effect of the evolutionary-based search approach in GP.", "The flexible representation of GP means it can theoretically discover the entire set of possible function forms."], "text_after_citation": ["Bloating is a common problem in GP-based SR and has been researched for decades.", "Previous solutions to this problem include incorporating a parsimony measure to the fitness function #OTHEREFR , #OTHEREFR or utilising online program simplification #OTHEREFR . For instance, Kinzett et al.", "#OTHEREFR explored the use of two online program simplification approaches in the evolution process: algebraic simplification #OTHEREFR and numerical simplification.", "This helps to reduce bloating in the evolution of GP.", "However, it needs predefined parameters, i.e., simplification rules and threshold."], "citing_paper_content": {"title": "Differentiable Genetic Programming For High-Dimensional Symbolic Regression", "abstract": "Symbolic regression (SR) is the process of discovering hidden relationships from data with mathematical expressions, which is typically considered an effective way to reach interpretable machine learning. Genetic programming (GP) has been the dominator in solving various SR problems. However, as the scale of SR problems increases, GP often poorly demonstrates and cannot effectively address the real-world high-dimensional problems. This limitation is mainly caused by the stochastic evolutionary nature of traditional GP in constructing the proper trees. In this paper, we propose a differentiable approach named DGP to construct GP trees towards high-dimensional SR for the first time. Specifically, a new data structure called differentiable symbolic tree is proposed to relax the discrete structure to be continuous, thus a gradient-based optimizer can be easily presented for the efficient optimization. In addition, a sampling method is proposed to eliminate the discrepancy caused by the above relaxation for valid symbolic expressions. Furthermore, a diversification mechanism is introduced to promote the optimizer escaping from local optima for globally optimal expressions. With these designs, the proposed DGP method can efficiently search for the GP trees with higher performance, thus being capable of dealing with high-dimensional SR. To demonstrate the effectiveness of DGP, we conducted various experiments against the state of the arts based on both GP and deep neural networks. The experiment results reveal that DGP can outperform these chosen peer competitors on real-world high-dimensional regression benchmarks with dimensions varying from hundreds to thousands. In addition, on the challenging synthetic SR problems, the proposed DGP method can also achieve the best recovery rate even with different noisy levels. It is believed this work can facilitate SR being a powerful alternative to interpretable machine learning for a broader range of real-world problems."}, "cited_paper_content": {"title": "Interaction-Transformation Evolutionary Algorithm For Symbolic Regression", "abstract": "The Interaction-Transformation (IT) is a new representation for Symbolic Regression that restricts the search space into simpler, but expressive, function forms. This representation has the advantage of creating a smoother search space unlike the space generated by Expression Trees, the common representation used in Genetic Programming. This paper introduces an Evolutionary Algorithm capable of evolving a population of IT expressions supported only by the mutation operator. The results show that this representation is capable of finding better approximations to real-world data sets when compared to traditional approaches and a state-of-the-art Genetic Programming algorithm."}, "keywords": ["unnecessarily large models", "evolutionary process"], "citation_intent": "background"}
{"citing_id": "2304.03659v1", "cited_id": "1811.12231", "section_title": "Compositional Evaluation", "citation": "This is similar to findings in #REFR where they found ImageNet trained CNN's are biased towards texture.", "text_before_citation": ["To better understand model failures, we categorized each attribute into 11 categories with results shown in Figure 5 (left).", "For example, \"visibility\" contains attributes like \"dark\" and \"bright\" while \"expression\" contains \"smiley\" and \"sad\". All models struggle when \"visibility\" related compositions are used.", "The highest performance was within the \"material\" category with attributes such as \"cloth\" and \"tin\".", "To compare backbone models, we average the image scores over CLIP backbone architectures in Figure 5 (middle).", "Some noticeable patterns are that the CNN backbone models are better with \"material\", \"pattern\" and \"texture\" related compositions while ViT's are better at \"color\" and \"shape\"."], "text_after_citation": ["Summary (1) Models struggle with compositionality but are better with those most associated with objects such as \"materials\".", "(2) CNN based backbones may be better at recognizing texture and patterns while ViT backbones with color and shape.", "Surprisingly, (3) these models are typically better at matching captions given the image rather than text."], "citing_paper_content": {"title": "Probing Conceptual Understanding Of Large Visual-Language Models", "abstract": "We present a novel framework for probing and improving relational, compositional and contextual understanding of large visual-language models (V+L). While large V+L models have achieved success in various downstream tasks, it is not clear if they have a conceptual grasp of the content. We propose a novel benchmarking dataset for probing three aspects of content understanding. Our probes are grounded in cognitive science and help determine if a V+L model can, for example, determine if snow garnished with a man is implausible, or if it can identify beach furniture by knowing it is located on a beach. We have experimented with 5 well known models, such as CLIP and ViLT, and found that they mostly fail to demonstrate a conceptual understanding. That said, we find interesting insights such as cross-attention helps learning conceptual understandingWe use these insights to propose a new finetuning technique that rewards the three conceptual understanding measures we proposed. We hope that the presented benchmarks will help the community assess and improve the conceptual understanding capabilities of large V+L models. Visual-language (V+L) models are recently and greatly successful deep learning models that learn representations of image and text in a shared space. These representations are useful for downstream tasks like image classification, visual-question answering, image retrieval and more [42, 37, 27, 26, 39, 1]. However, for use in real-world applications, it is also vital that models \"understand\" rather than memorize to perform on more general tasks [24]. While large-language models have been shown to have a moderate amount of \"theory of mind\", as measured by conceptual consistency [29], V+L models have not"}, "cited_paper_content": {"title": "Imagenet-Trained Cnns Are Biased Towards Texture; Increasing Shape Bias Improves Accuracy And Robustness", "abstract": "Convolutional Neural Networks (CNNs) are commonly thought to recognise objects by learning increasingly complex representations of object shapes. Some recent studies suggest a more important role of image textures. We here put these conflicting hypotheses to a quantitative test by evaluating CNNs and human observers on images with a texture-shape cue conflict. We show that ImageNet-trained CNNs are strongly biased towards recognising textures rather than shapes, which is in stark contrast to human behavioural evidence and reveals fundamentally different classification strategies. We then demonstrate that the same standard architecture (ResNet-50) that learns a texture-based representation on ImageNet is able to learn a shape-based representation instead when trained on \"Stylized-ImageNet\", a stylized version of ImageNet. This provides a much better fit for human behavioural performance in our well-controlled psychophysical lab setting (nine experiments totalling 48,560 psychophysical trials across 97 observers) and comes with a number of unexpected emergent benefits such as improved object detection performance and previously unseen robustness towards a wide range of image distortions, highlighting advantages of a shape-based representation."}, "keywords": ["ImageNet"], "citation_intent": "result"}
{"citing_id": "2303.04315v1", "cited_id": "1511.00561", "section_title": "B. Local And Global Context Modeling", "citation": "Loss in localization accuracy was tackled by pioneering works in Deep Semantic Segmentation, such as SegNet #REFR .", "text_before_citation": ["The exploration of larger (ultimately global) contexts, in turn, can be performed by the stacking of convolutional layers so that the receptive fields get larger at each level, as illustrated in Fig 3.", "Depending on the input resolution, though, achieving a large receptive field can result in very deep networks, which incur high computational costs, and can lead to problems such as vanishing or exploding gradients, as well as overfitting issues.", "Given this context, some alternatives are widely adopted in the literature for a good balance of local and global feature exploration.", "1) Fast Downsampling: The first alternative is to perform a fast down-sampling of feature maps, in order to fully extract contextual information with fewer layers.However, this leads to some of the main challenges to highly-accurate semantic segmentation, as stated by Atif et al.", "#OTHEREFR : reduced localization accuracy introduced by pooling operations to incorporate invariance in the network, and reduced feature map resolution caused by strided convolutions."], "text_after_citation": ["They proposed to keep spatial details by preserving the location of the values extracted by the max pooling operation in the encoder stages.", "This information would then be passed to the corresponding decoder layer by means of the so-called Pooling Indices in order to guide the up-sampling stages. #OTHEREFR and #OTHEREFR try a similar approach.", "Excessive downsampling of feature maps leads to the loss of information related to small objects.", "In FCNs, for instance, the output is usually 32 times smaller than the input image #OTHEREFR .", "Throughout the years, various approaches have been proposed in order to reduce loss of spatial details caused by feature downsampling, while guaranteeing the coverage of long-range dependencies."], "citing_paper_content": {"title": "A Threefold Review On Deep Semantic Segmentation: Efficiency-Oriented, Temporal And Depth-Aware Design", "abstract": "Semantic image and video segmentation stand among the most important tasks in computer vision nowadays, since they provide a complete and meaningful representation of the environment by means of a dense classification of the pixels in a given scene. Recently, Deep Learning, and more precisely Convolutional Neural Networks, have boosted semantic segmentation to a new level in terms of performance and generalization capabilities. However, designing Deep Semantic Segmentation models is a complex task, as it may involve application-dependent aspects. Particularly, when considering autonomous driving applications, the robustness-efficiency tradeoff, as well as intrinsic limitations-computational/memory bounds and data-scarcity-and constraints-real-time inference-should be taken into consideration. In this respect, the use of additional data modalities, such as depth perception for reasoning on the geometry of a scene, and temporal cues from videos to explore redundancy and consistency, are promising directions yet not explored to their full potential in the literature. In this paper, we conduct a survey on the most relevant and recent advances in Deep Semantic Segmentation in the context of vision for autonomous vehicles, from three different perspectives: efficiency-oriented model development for real-time operation, RGB-Depth data integration (RGB-D semantic segmentation), and the use of temporal information from videos in temporalaware models. Our main objective is to provide a comprehensive discussion on the main methods, advantages, limitations, results and challenges faced from each perspective, so that the reader can not only get started, but also be up to date in respect to recent advances in this exciting and challenging research field."}, "cited_paper_content": {"title": "Segnet: A Deep Convolutional Encoder-Decoder Architecture For Image Segmentation", "abstract": "We present a novel and practical deep fully convolutional neural network architecture for semantic pixel-wise segmentation termed SegNet. This core trainable segmentation engine consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. The architecture of the encoder network is topologically identical to the 13 convolutional layers in the VGG16 network [1] . The role of the decoder network is to map the low resolution encoder feature maps to full input resolution feature maps for pixel-wise classification. The novelty of SegNet lies is in the manner in which the decoder upsamples its lower resolution input feature map(s). Specifically, the decoder uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling. This eliminates the need for learning to upsample. The upsampled maps are sparse and are then convolved with trainable filters to produce dense feature maps. We compare our proposed architecture with the widely adopted FCN [2] and also with the well known DeepLab-LargeFOV [3] , DeconvNet [4] architectures. This comparison reveals the memory versus accuracy trade-off involved in achieving good segmentation performance. SegNet was primarily motivated by scene understanding applications. Hence, it is designed to be efficient both in terms of memory and computational time during inference. It is also significantly smaller in the number of trainable parameters than other competing architectures and can be trained end-to-end using stochastic gradient descent. We also performed a controlled benchmark of SegNet and other architectures on both road scenes and SUN RGB-D indoor scene segmentation tasks. These quantitative assessments show that SegNet provides good performance with competitive inference time and most efficient inference memory-wise as compared to other architectures. We also provide a Caffe implementation of SegNet and a web demo at http://mi.eng.cam.ac.uk/projects/segnet."}, "keywords": ["Deep Semantic Segmentation"], "citation_intent": "background"}
{"citing_id": "2303.00304v3", "cited_id": "1709.06158", "section_title": "Image-Goal Navigation Results", "citation": "We also have tested the proposed method in MP3D #REFR dataset, and observed similar results with Gibson dataset.", "text_before_citation": ["RNR-Map helps efficient navigation.", "F loc , the success rate and SPL dramatically drop."], "text_after_citation": ["We provide the results in the supplementary material E, and additional ablation study about the submodules of the navigation module is also provided in D.", "F track makes RNR-Map robust to noise.", "Comparing the third row and the last row of Table 2 , we can infer that the pose adjusting function F track helps the agent to find the image goal more effectively, even with the noisy odometry.", "The proposed method shows higher success rates in noisy settings, while SPL values are low.", "We believe this is from the randomness of noises, which helps the local navigation policy to escape from being stuck in a single location."], "citing_paper_content": {"title": "Renderable Neural Radiance Map For Visual Navigation", "abstract": "We propose a novel type of map for visual navigation, a renderable neural radiance map (RNR-Map), which is designed to contain the overall visual information of a 3D environment. The RNR-Map has a grid form and consists of latent codes at each pixel. These latent codes are embedded from image observations, and can be converted to the neural radiance field which enables image rendering given a camera pose. The recorded latent codes implicitly contain visual information about the environment, which makes the RNR-Map visually descriptive. This visual information in RNR-Map can be a useful guideline for visual localization and navigation. We develop localization and navigation frameworks that can effectively utilize the RNR-Map. We evaluate the proposed frameworks on camera tracking, visual localization, and image-goal navigation. Experimental results show that the RNR-Map-based localization framework can find the target location based on a single query image with fast speed and competitive accuracy compared to other baselines. Also, this localization framework is robust to environmental changes, and even finds the most visually similar places when a query image from a different environment is given. The proposed navigation framework outperforms the existing image-goal navigation methods in difficult scenarios, under odometry and actuation noises. The navigation framework shows 65.7% success rate in curved scenarios of the NRNS [23] dataset, which is an improvement of 18.6% over the current state-of-the-art. Project page: https: //rllab-snu.github.io/projects/RNR-Map/"}, "cited_paper_content": {"title": "Matterport3D: Learning From Rgb-D Data In Indoor Environments", "abstract": "Access to large, diverse RGB-D datasets is critical for training RGB-D scene understanding algorithms. However, existing datasets still cover only a limited number of views or a restricted scale of spaces. In this paper, we introduce Matterport3D, a large-scale RGB-D dataset containing 10,800 panoramic views from 194,400 RGB-D images of 90 building-scale scenes. Annotations are provided with surface reconstructions, camera poses, and 2D and 3D semantic segmentations. The precise global alignment and comprehensive, diverse panoramic set of views over entire buildings enable a variety of supervised and self-supervised computer vision tasks, including keypoint matching, view overlap prediction, normal prediction from color, semantic segmentation, and region classification."}, "keywords": ["similar results"], "citation_intent": "result"}
{"citing_id": "2304.12370v1", "cited_id": "1511.04561", "section_title": "The Vicuna Benchmark", "citation": "Format your response as a table containing question ID, score, and reasoning.\" #REFR Under this schema, we obtained the following results: The Vicuna benchmark shows consistent improvement over the base OPT models.", "text_before_citation": ["\"You are a sophisticated and critical judge evaluating the outputs of AI models.", "For each question given above, evaluate the \"answer\" as a response to the prompt.", "Judge each answer on a score of 1-10 based on relevance, coherence, and on its informational value as a response to the question.", "Actions should only be considered useful if their output is also present in the answer.", "Generate both this score and your reasoning for the score."], "text_after_citation": ["The 1.3b OPT model seems to be better, on the surface than the 2.7b model. We don't know why this is.", "2.7b was more curt, more prone to insulting the user, refusing to elaborate on answers.", "It's interesting that the Eluwa 1.3 2-epoch model is doing better here than base OPT 6.7b (a model 5 times its size).", "However, a word of caution: 1.3's verbosity seems to come from regurgitation, and less refined ability to write, and indeed in our testing we found this family of models answering questions with fabricated blogposts, right down to authors' bios and twitter credentials."], "citing_paper_content": {"title": "Better Question-Answering Models On A Budget", "abstract": "Low-rank adaptation (LoRA) and question-answer datasets from large language models have made it much easier for much smaller models to be finetuned to the point where they display sophisticated conversational abilities. In this paper, we present Eluwa, a family of LoRA models that use the Stanford Alpaca dataset and massively improve the capabilities of Facebook's OPT 1.3B, 2.7B and 6.7B models. We benchmark these models in multiple ways, including letting GPT-4 judge their answers to prompts that span general knowledge, writing, programming and other tasks. We show that smaller models here can be fine-tuned to be as performant as models 3x larger-all for as little as 40 USD in compute."}, "cited_paper_content": {"title": "8-Bit Approximations For Parallelism In Deep Learning", "abstract": "The creation of practical deep learning data-products often requires parallelization across processors and computers to make deep learning feasible on large data sets, but bottlenecks in communication bandwidth make it difficult to attain good speedups through parallelism. Here we develop and test 8-bit approximation algorithms which make better use of the available bandwidth by compressing 32-bit gradients and nonlinear activations to 8-bit approximations. We show that these approximations do not decrease predictive performance on MNIST, CIFAR10, and ImageNet for both model and data parallelism and provide a data transfer speedup of 2x relative to 32-bit parallelism. We build a predictive model for speedups based on our experimental data, verify its validity on known speedup data, and show that we can obtain a speedup of 50x and more on a system of 96 GPUs compared to a speedup of 23x for 32-bit. We compare our data types with other methods and show that 8-bit approximations achieve state-of-the-art speedups for model parallelism. Thus 8-bit approximation is an efficient method to parallelize convolutional networks on very large systems of GPUs."}, "keywords": ["response", "base OPT models"], "citation_intent": "result"}
{"citing_id": "2304.00058v1", "cited_id": "1702.02925", "section_title": "Experiments", "citation": "Following the work #REFR , we split all labeled frames into subject-exclusive 3-fold with 12 AUs for both two stages.", "text_before_citation": ["The proposed CLEF is compared with the state-of-theart methods on six popular databases for FER and AUR tasks.", "Furthermore, we conduct ablation studies to verify the component-wise contribution of our method.", "Expert coders select the most expressive 20s of each video clip for AU coding, producing 140,000 labeled frames."], "text_after_citation": ["BP4D+ #OTHEREFR consists of 140 subjects with total of 1.5 M frames in the same laboratory environments.", "For each subject, 20 seconds from 4 activities are annotated, resulting in 192,000 labeled frames.", "First, the 140 subjects are split into four fold, following the same setting in #OTHEREFR .", "In pretraining, we equally sample 480,000 frames from all 1.5 M frames by 10 activity categories.", "In fine-tuning, 12 AUs, the same as in BP4D, are selected for AU recognition."], "citing_paper_content": {"title": "Weakly-Supervised Text-Driven Contrastive Learning For Facial Behavior Understanding", "abstract": "Contrastive learning has shown promising potential for learning robust representations by utilizing unlabeled data. However, constructing effective positive-negative pairs for contrastive learning on facial behavior datasets remains challenging. This is because such pairs inevitably encode the subject-ID information, and the randomly constructed pairs may push similar facial images away due to the limited number of subjects in facial behavior datasets. To address this issue, we propose to utilize activity descriptions, coarse-grained information provided in some datasets, which can provide high-level semantic information about the image sequences but is often neglected in previous studies. More specifically, we introduce a two-stage Contrastive Learning with Text-Embeded framework for Facial behavior understanding (CLEF). The first stage is a weakly-supervised contrastive learning method that learns representations from positive-negative pairs constructed using coarse-grained activity information. The second stage aims to train the recognition of facial expressions or facial action units by maximizing the similarity between image and the corresponding text label names. The proposed CLEF achieves state-of-the-art performance on three inthe-lab datasets for AU recognition and three in-the-wild datasets for facial expression recognition. Recently, many deep learning based approaches [28, 53, 38, 41, 55, 18, 4] have been proposed and achieved state-ofthe-art performance in FER and AUR. A variety of meth"}, "cited_paper_content": {"title": "Eac-Net: A Region-Based Deep Enhancing And Cropping Approach For Facial Action Unit Detection", "abstract": "In this paper, we propose a deep learning basedapproach for facial action unit detection by enhancing andcropping the regions of interest. The approach is implementedby adding two novel nets (layers): the enhancing layers and thecropping layers, to a pretrained CNN model. For the enhancinglayers (the E-Net), we designed an attention map based on faciallandmark features and applied it to a pretrained neural networkto conduct enhanced learning. For the cropping layers (the CNet), we crop facial regions around the detected landmarksand design convolutional layers to learn deeper features foreach facial region. We then fuse the E-Net and the C-Net toobtain our Enhancing and Cropping (EAC) Net, which canlearn both feature enhancing and region cropping functions.Our approach shows significant improvement in performancecompared to the state-of-the-art methods applied to BP4D andDISFA AU datasets."}, "keywords": ["frames"], "citation_intent": "method"}
{"citing_id": "2303.17157v1", "cited_id": "1301.2866", "section_title": "Introduction", "citation": "The first few eigenfunctions corresponding to small eigenvalues (the convergence depends on the decay of the spectral problems #REFR ) are considered as the multiscale basis functions.", "text_before_citation": ["These approaches are usually referred to as multiscale methods in the literature, among which we may mention: Multiscale Finite Element Method #OTHEREFR , the Variational Multiscale Method #OTHEREFR , Mixed Multiscale Finite Element Method #OTHEREFR , Mixed Mortar Multiscale Finite Element Method #OTHEREFR , the two-scale Finite Element Method #OTHEREFR , and the Multiscale Finite Volume method #OTHEREFR .", "The aforementioned methods share model reduction techniques, using different structures to find multiscale solutions, especially in many practical applications such as fluid flow simulations, for instance, #OTHEREFR .", "In particular, we consider a family of an extended version of MsFEM, known as the Generalized Multiscale Finite Element Method (GMsFEM) that was first introduced by #OTHEREFR .", "The main idea of GMsFEM is to construct localized basis functions by solving local spectral problems that are used to approximate the solution on a coarse grid incorporating fine-scale features.", "Following this, we construct an auxiliary space associated with local spectral problems in the coarse grid."], "text_after_citation": ["In this paper, we extend the Constraint Energy Minimization Generalized Multiscale Finite Element Method (CEM-GMsFEM) developed in #OTHEREFR to solve single-phase nonlinear compressible flow.", "The key ideas of the method can be summarized as follows: First, we construct the auxiliary basis functions by solving local spectral problems. Then, by using oversampling techniques and localization (cf.", "#OTHEREFR ), we solve an appropriate energy subject to some constrainted oversampling regions to find the required basis functions.", "Finally, the resulting basis functions are shown to have exponential decay away from the target coarse element, and therefore, they are localizable.", "Then we rigorously analyze the convergence error estimates for the proposed scheme."], "citing_paper_content": {"title": "Convergence Of The Cem-Gmsfem For Compressible Flow In Highly Heterogeneous Media", "abstract": "This paper presents and analyses a Constraint Energy Minimization Generalized Multiscale Finite Element Method (CEM-GMsFEM) for solving single-phase nonlinear compressible flows in highly heterogeneous media. The construction of CEM-GMsFEM hinges on two crucial steps: First, the auxiliary space is constructed by solving local spectral problems, where the basis functions corresponding to small eigenvalues are captured. Then the basis functions are obtained by solving local energy minimization problems over the oversampling domains using the auxiliary space. The basis functions have exponential decay outside the corresponding local oversampling regions. The convergence of the proposed method is provided, and we show that this convergence only depends on the coarse grid size and is independent of the heterogeneities. An online enrichment guided by a posteriori error estimator is developed to enhance computational efficiency. Several numerical experiments on a three-dimensional case to confirm the theoretical findings are presented, illustrating the performance of the method and giving efficient and accurate numerical"}, "cited_paper_content": {"title": "Generalized Multiscale Finite Element Methods (Gmsfem)", "abstract": "In this paper, we propose a general approach called Generalized Multiscale Finite Element Method (GMsFEM) for performing multiscale simulations for problems without scale separation over a complex input space. As in multiscale finite element methods (MsFEMs), the main idea of the proposed approach is to construct a small dimensional local solution space that can be used to generate an efficient and accurate approximation to the multiscale solution with a potentially high dimensional input parameter space. In the proposed approach, we present a general procedure to construct the offline space that is used for a systematic enrichment of the coarse solution space in the online stage. The enrichment in the online stage is performed based on a spectral decomposition of the offline space. In the online stage, for any input parameter, a multiscale space is constructed to solve the global problem on a coarse grid. The online space is constructed via a spectral decomposition of the offline space and by choosing the eigenvectors corresponding to the largest eigenvalues. The computational saving is due to the fact that the construction of the online multiscale space for any input parameter is fast and this space can be re-used for solving the forward problem with any forcing and boundary condition. Compared with the other approaches where global snapshots are used, the local approach that we present in this paper allows us to eliminate unnecessary degrees of freedom on a coarse-grid level. We present various examples in the paper and some numerical results to demonstrate the effectiveness of our method."}, "keywords": ["multiscale basis functions"], "citation_intent": "background"}
{"citing_id": "2303.16464v1", "cited_id": "1707.09564", "section_title": "Adam Optimizer", "citation": "Replacing k by N b in the inequality #REFR leads to the inequality in the proposition.", "text_before_citation": [". From Lemma 5.2, we have:", "\u2206 T \u2264 8 T k 2 \u2022 \u03b7T \u03b3 = 8 k 2 \u2022 \u03b7\u03b3 .", "According to Definition 3.4, we have:", "EQUATION", "The inequality #OTHEREFR has been obtained similar to #OTHEREFR ."], "text_after_citation": ["Theorem 5.5.", "Let (\u0177, y) with the maximum value of L be convex and \u03b3-Lipschitz.", "Assume Adam is run for T iterations with a learning rate \u03b7 and batch size b to obtain f B S ,R .", "Then we have the following upper bound for E(f B S ,R ) with probability at least 1 \u2212 \u03b4:", "EQUATION"], "citing_paper_content": {"title": "Lipschitzness Effect Of A Loss Function On Generalization Performance Of Deep Neural Networks Trained By Adam And Adamw Optimizers", "abstract": "The generalization performance of deep neural networks with regard to the optimization algorithm is one of the major concerns in machine learning. This performance can be affected by various factors. In this paper, we theoretically prove that the Lipschitz constant of a loss function is an important factor to diminish the generalization error of the output model obtained by Adam or AdamW. The results can be used as a guideline for choosing the loss function when the optimization algorithm is Adam or AdamW. In addition, to evaluate the theoretical bound in a practical setting, we choose the human age estimation problem in computer vision. For assessing the generalization better, the training and test datasets are drawn from different distributions. Our experimental evaluation shows that the loss function with lower Lipschitz constant and maximum value improves the generalization of the model trained by Adam or AdamW."}, "cited_paper_content": {"title": "A Pac-Bayesian Approach To Spectrally-Normalized Margin Bounds For Neural Networks", "abstract": "We present a generalization bound for feedforward neural networks in terms of the product of the spectral norm of the layers and the Frobenius norm of the weights. The generalization bound is derived using a PAC-Bayes analysis."}, "keywords": ["inequality"], "citation_intent": "background"}
{"citing_id": "2304.07735v1", "cited_id": "1810.06060", "section_title": "Introduction", "citation": "Split learning #REFR emerges as a potential solution: without sharing its raw data, the edge feeds the private inputs through the first few layers of DNNs and transmits the intermediate features to the cloud to offload partial computation.", "text_before_citation": ["Recent years have witnessed remarkable growth in deep learning applications, as deep neural networks (DNNs) have grown deeper and larger.", "It poses a dilemma for the thin edge device: on one hand, it lacks the computational power to individually train the models; on the other, it would violate data privacy sending all data to an untrusted party, e.g., the cloud, to process."], "text_after_citation": ["However, the vanilla split learning still faces privacy leakage since an adversary could infer the input from the feature #OTHEREFR .", "Hence many works have proposed to hide sensitive information from the features, such as encryption #OTHEREFR , adversarial learning #OTHEREFR , differential privacy [6] , etc.", "However, these works mostly sacrifice either accuracy or efficiency for privacy: crypto-based solutions are often far inefficient; privacy-preserving methods in the plaintext are efficient, but always face the tradeoffs between accuracy and privacy.", "Our work presents the first privacy-preserving split learning paradigm without any accuracy loss in plaintext.", "We discovered many DNN modules have a natural structure that is permutation equivalent, meaning that they can process permuted inputs equivalently to the original inputs."], "citing_paper_content": {"title": "Shuffled Transformer For Privacy-Preserving Split Learning", "abstract": "In conventional split learning, training and testing data often face severe privacy leakage threats. Existing solutions often have to trade learning accuracy for data privacy, or the other way around. We propose a lossless privacy-preserving split learning framework, on the basis of the permutation equivalence properties which are inherent to many neural network modules. We adopt Transformer as the example building block to the framework. It is proved that the Transformer encoder block is permutation equivalent, and thus training/testing could be done equivalently on permuted data. We further introduce shufflingbased privacy guarantee and enhance it by mix-up training. All properties are verified by conducted experiments, which also show strong defence against privacy attacks compared to the stateof-the-art, yet without any accuracy decline."}, "cited_paper_content": {"title": "Distributed Learning Of Deep Neural Network Over Multiple Agents", "abstract": "In domains such as health care and finance, shortage of labeled data and computational resources is a critical issue while developing machine learning algorithms. To address the issue of labeled data scarcity in training and deployment of neural network-based systems, we propose a new technique to train deep neural networks over several data sources. Our method allows for deep neural networks to be trained using data from multiple entities in a distributed fashion. We evaluate our algorithm on existing datasets and show that it obtains performance which is similar to a regular neural network trained on a single machine. We further extend it to incorporate semi-supervised learning when training with few labeled samples, and analyze any security concerns that may arise. Our algorithm paves the way for distributed training of deep neural networks in data sensitive applications when raw data may not be shared directly."}, "keywords": ["private inputs", "Split learning"], "citation_intent": "background"}
{"citing_id": "2303.13496v1", "cited_id": "1911.02116", "section_title": "Dataset", "citation": "We initialize the text encoder from a relatively small XLM-R Base #REFR model with 270M parameters.", "text_before_citation": ["Table 4 shows a comparison with state-of-the-art methods on low-shot IN1k and iNat18, including foundational and self-supervised models.", "Our models show impressive lowshot performance on both IN1k and iNat18, reaching 83.7% and 80.3% top-1 accuracy with only 10 labeled examples per class, and reaching the highest reported performance with just one labeled example per class on IN1k of 62.1%.", "Zero-shot transfer.", "Strong foundational models are expected to also have a good open world understanding of visual concepts.", "To equip our pretrained vision encoders with such capabilities, we utilize LiT #OTHEREFR ."], "text_after_citation": ["Table 5 shows the zero-shot transfer performance of our models on IN1k, and F-101.", "Our ViT-2B attains 80.8% accuracy on IN1k, outperforming a similarly sized OpenCLIP model.", "Our results lag behind other works which pretrain on datasets such as JFT-3B and ALIGN, highlighting the importance of the pretraining Table 6 : Detection and Segmentation results on LVIS (v1 val).", "We compare MAE\u2192WSP with prior work and report the detection (AP box ) and instance segmentation performance (AP mask ).", "Our ViT-H significantly outperforms SWAG which uses the same pretraining and model size, but without pre-pretraining."], "citing_paper_content": {"title": "The Effectiveness Of Mae Pre-Pretraining For Billion-Scale Pretraining", "abstract": "This paper revisits the standard pretrain-then-finetune paradigm used in computer vision for visual recognition tasks. Typically, state-of-the-art foundation models are pretrained using large scale (weakly) supervised datasets with billions of images. We introduce an additional prepretraining stage that is simple and uses the self-supervised MAE technique to initialize the model. While MAE has only been shown to scale with the size of models, we find that it scales with the size of the training dataset as well. Thus, our MAE-based pre-pretraining scales with both model and data size making it applicable for training foundation models. Pre-pretraining consistently improves both the model convergence and the downstream transfer performance across a range of model scales (millions to billions of parameters), and dataset sizes (millions to billions of images). We measure the effectiveness of pre-pretraining on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our largest model achieves new state-of-theart results on iNaturalist-18 (91.3%), 1-shot ImageNet-1k (62.1%), and zero-shot transfer on Food-101 (96.0%). Our study reveals that model initialization plays a significant role, even for web-scale pretraining with billions of images. * Equal technical contribution. \u2020 Project Lead."}, "cited_paper_content": {"title": "Unsupervised Cross-Lingual Representation Learning At Scale", "abstract": "This paper shows that pretraining multilingual language models at scale leads to significant performance gains for a wide range of cross-lingual transfer tasks. We train a Transformer-based masked language model on one hundred languages, using more than two terabytes of filtered CommonCrawl data. Our model, dubbed XLM-R, significantly outperforms multilingual BERT (mBERT) on a variety of cross-lingual benchmarks, including +13.8% average accuracy on XNLI, +12.3% average F1 score on MLQA, and +2.1% average F1 score on NER. XLM-R performs particularly well on low-resource languages, improving 11.8% in XNLI accuracy for Swahili and 9.2% for Urdu over the previous XLM model. We also present a detailed empirical evaluation of the key factors that are required to achieve these gains, including the trade-offs between (1) positive transfer and capacity dilution and (2) the performance of high and low resource languages at scale. Finally, we show, for the first time, the possibility of multilingual modeling without sacrificing per-language performance; XLM-Ris very competitive with strong monolingual models on the GLUE and XNLI benchmarks. We will make XLM-R code, data, and models publicly available."}, "keywords": ["Base model", "text encoder"], "citation_intent": "method"}
{"citing_id": "2303.04388v1", "cited_id": "1802.08129", "section_title": "Introduction", "citation": "Although this representation can explain the prediction somewhat, it can still be abstract to users. To solve this problem, Park et al. #REFR and Wu et al.", "text_before_citation": ["Since these applications are closely related to the safety and property of our life, there are high requirements for the reliability of the VQA methods.", "Although most present methods have demonstrated exemplary performance, they are considered black boxes, as their internal logic and workings are concealed from users, which cannot meet the reliability requirements.", "It is considered that providing a reasonable explanation for the prediction process can become one of the solutions to prove the model's reliability.", "Therefore, new regulations have been created to establish standards for verifying mandatory decisions, which increases the demand for methods' interpretability #OTHEREFR .", "Most conventional interpretable VQA models use a heat map to represent the different attention parts of the model in an image."], "text_after_citation": ["#OTHEREFR have proposed interpretable VQA models to generate explanations with human-readable and detailed natural language sentences.", "These methods require creating another model that generates explanations for the VQA model.", "Due to the generation of explanations and answers being independent, it is difficult to control whether the generated explanation conforms to the logic used by the VQA model.", "Recently, pre-trained large-scale models such as the Bidirectional Encoder Representations from Transformers (BERT) #OTHEREFR , GPT-2 #OTHEREFR , and the Contrastive Language-Image Pre-training (CLIP) #OTHEREFR have been developed rapidly, which makes it possible to generate natural language more correctly.", "Focusing on the deficiency of the previous models, Sammani et al. #OTHEREFR , Marasovic et al. #OTHEREFR , and Kayser et al."], "citing_paper_content": {"title": "Interpretable Visual Question Answering Referring To Outside Knowledge", "abstract": "We present a novel multimodal interpretable VQA model that can answer the question more accurately and generate diverse explanations. Although researchers have proposed several methods that can generate human-readable and fine-grained natural language sentences to explain a model's decision, these methods have focused solely on the information in the image. Ideally, the model should refer to various information inside and outside the image to correctly generate explanations, just as we use background knowledge daily. The proposed method incorporates information from outside knowledge and multiple image captions to increase the diversity of information available to the model. The contribution of this paper is to construct an interpretable visual question answering model using multimodal inputs to improve the rationality of generated results. Experimental results show that our model can outperform state-ofthe-art methods regarding answer accuracy and explanation rationality."}, "cited_paper_content": {"title": "Multimodal Explanations: Justifying Decisions And Pointing To The Evidence", "abstract": "Deep models that are both effective and explainable are desirable in many settings; prior explainable models have been unimodal, offering either image-based visualization of attention weights or text-based generation of post-hoc justifications. We propose a multimodal approach to explanation, and argue that the two modalities provide complementary explanatory strengths. We collect two new datasets to define and evaluate this task, and propose a novel model which can provide joint textual rationale generation and attention visualization. Our datasets define visual and textual justifications of a classification decision for activity recognition tasks (ACT-X) and for visual question answering tasks (VQA-X). We quantitatively show that training with the textual explanations not only yields better textual justification models, but also better localizes the evidence that supports the decision. We also qualitatively show cases where visual explanation is more insightful than textual explanation, and vice versa, supporting our thesis that multimodal explanation models offer significant benefits over unimodal approaches."}, "keywords": ["representation"], "citation_intent": "background"}
{"citing_id": "2303.00313v1", "cited_id": "1906.11081", "section_title": "C. Benchmark Analysis", "citation": "MGCN #REFR , though not showing distinct advantages on other datasets, shows a competitive result as those pre-training techniques on QM7 and QM8 datasets.", "text_before_citation": ["The overall model performance on the MoleculeNet dataset also indicates that self-supervised pre-training tasks result in a performance boost.", "While supervised pre-training tasks introduce the risk of negative transfer #OTHEREFR , well-designed self-supervised pre-trained models show rival performance in learning implicit domain knowledge #OTHEREFR - #OTHEREFR .", "However, some tasks that involve alteration of the molecule structure, such as subgraph removal #OTHEREFR , result in poor performance in datasets sensitive to topology changes, such as the BBBP dataset.", "Moreover, the composition of different augmentation strategies is not bound to perform better since removing various substructures might easily lead to the loss of important structural information.", "Furthermore, models involving 3D information have superior performance on datasets of quantum mechanical properties."], "text_after_citation": ["For retrosynthesis prediction, RetroXpert #OTHEREFR and Reaction-Aware #OTHEREFR combine the information from both graph representation and sequence representation and perform well. Generally speaking, template-free models have better performance. However, LocalRetro #OTHEREFR , a template-based model, outperforms others.", "This may be due to the model design, which not only strictly constrains the templates but also considers global structure.", "For reaction prediction, models such as GTPN #OTHEREFR , Molecular Transformer #OTHEREFR and NERF #OTHEREFR which utilize the state-of-art models in other applications of machine learning also get good results in this task."], "citing_paper_content": {"title": "Deep Learning Methods For Small Molecule Drug Discovery: A Survey", "abstract": "With the development of computer-assisted techniques, research communities including biochemistry and deep learning have been devoted into the drug discovery field for over a decade. Various applications of deep learning have drawn great attention in drug discovery, such as molecule generation, molecular property prediction, retrosynthesis prediction, and reaction prediction. While most existing surveys only focus on one of the applications, limiting the view of researchers in the community. In this paper, we present a comprehensive review on the aforementioned four aspects, and discuss the relationships among different applications. The latest literature and classical benchmarks are presented for better understanding the development of variety of approaches. We commence by summarizing the molecule representation format in these works, followed by an introduction of recent proposed approaches for each of the four tasks. Furthermore, we review a variety of commonly used datasets and evaluation metrics and compare the performance of deep learning-based models. Finally, we conclude by identifying remaining challenges and discussing the future trend for deep learning methods in drug discovery. Impact Statement-As artificial intelligence continues to evolve, an increasing number of deep learning algorithms are proposed to utilize extensive experimental data to accelerate the timeconsuming and costly procedure for drug discovery. Unlike previous surveys that tend to focus on a single aspect of drug discovery, we offer a more comprehensive survey that encompasses four areas of drug discovery utilizing deep learning techniques, including molecule generation, molecular property prediction, retrosynthesis, and reaction prediction. Before introducing the main methods, we start with molecular data representation, which is also crucial for deep learning algorithms. Moreover, we present the recent benchmarks for deep learning-based drug discovery, which cover 9 widely used databases. Index Terms-deep learning, drug discovery."}, "cited_paper_content": {"title": "Molecular Property Prediction: A Multilevel Quantum Interactions Modeling Perspective", "abstract": "Predicting molecular properties (e.g., atomization energy) is an essential issue in quantum chemistry, which could speed up much research progress, such as drug designing and substance discovery. Traditional studies based on density functional theory (DFT) in physics are proved to be time-consuming for predicting large number of molecules. Recently, the machine learning methods, which consider much rule-based information, have also shown potentials for this issue. However, the complex inherent quantum interactions of molecules are still largely underexplored by existing solutions. In this paper, we propose a generalizable and transferable Multilevel Graph Convolutional neural Network (MGCN) for molecular property prediction. Specifically, we represent each molecule as a graph to preserve its internal structure. Moreover, the well-designed hierarchical graph neural network directly extracts features from the conformation and spatial information followed by the multilevel interactions. As a consequence, the multilevel overall representations can be utilized to make the prediction. Extensive experiments on both datasets of equilibrium and off-equilibrium molecules demonstrate the effectiveness of our model. Furthermore, the detailed results also prove that MGCN is generalizable and transferable for the prediction."}, "keywords": ["QM8 datasets"], "citation_intent": "result"}
{"citing_id": "2304.06341v1", "cited_id": "1812.05934", "section_title": "Afl++ Integration", "citation": "In contrast to other analysis tools, this leads to fewer false alarms, e.g., when encountering manual reentrancy locking #REFR .", "text_before_citation": ["For example, the vulnerability depicted in Figure 6 is detected neither by Confuzzius #OTHEREFR nor by Smartian #OTHEREFR .", "However, since EF CF simulates multiple attacker-controlled accounts, it will quickly generate a transaction originating from the first account and leaking the Ether to a second attacker-controlled address.", "Most analysis tools feature explicit Reentrancy bug oracles.", "In contrast, EF CF does not feature an explicit detector for reentrancy but simply generates reentrant transaction sequences that trigger other bug oracles, such as Ether gains.", "In this way, EF CF detects reentrancy bugs, but only if they are actually exploitable."], "text_after_citation": ["By default, EF CF reports unprotected selfdestruct only if the selfdestruct will transfer the remaining Ether of the target contract to the attacker, i.e., the address parameter of the selfdestruct is controlled by the attacker.", "Optionally, EF CF can also report DoS-style unprotected selfdestructs, i.e., if the selfdestruct can be triggered by anyone, but always targets a trusted address such as the owner.", "This style of detection is featured in most other analysis tools.", "With Solidity versions > 0.8, contracts feature automatic integer overflow checking and proper assertion violation reporting.", "EF CF supports this new Solidity exception mechanism to signal errors to the fuzzer."], "citing_paper_content": {"title": "Ef Cf: High Performance Smart Contract Fuzzing For Exploit Generation", "abstract": "Smart contracts are increasingly being used to manage large numbers of high-value cryptocurrency accounts. There is a strong demand for automated, efficient, and comprehensive methods to detect security vulnerabilities in a given contract. While the literature features a plethora of analysis methods for smart contracts, the existing proposals do not address the increasing complexity of contracts. Existing analysis tools suffer from false alarms and missed bugs in today's smart contracts that are increasingly defined by complexity and interdependencies. To scale accurate analysis to modern smart contracts, we introduce EF CF, a high-performance fuzzer for Ethereum smart contracts. In contrast to previous work, EF CF efficiently and accurately models complex smart contract interactions, such as reentrancy and cross-contract interactions, at a very high fuzzing throughput rate. To achieve this, EF CF transpiles smart contract bytecode into native C++ code, thereby enabling the reuse of existing, optimized fuzzing toolchains. Furthermore, EF CF increases fuzzing efficiency by employing a structure-aware mutation engine for smart contract transaction sequences and using a contract's ABI to generate valid transaction inputs. In a comprehensive evaluation, we show that EF CF scales better-without compromising accuracy-to complex contracts compared to state-of-theart approaches, including other fuzzers, symbolic/concolic execution, and hybrid approaches. Moreover, we show that EF CF can automatically generate transaction sequences that exploit reentrancy bugs to steal Ether."}, "cited_paper_content": {"title": "Sereum: Protecting Existing Smart Contracts Against Re-Entrancy Attacks", "abstract": "Recently, a number of existing blockchain systems have witnessed major bugs and vulnerabilities within smart contracts. Although the literature features a number of proposals for securing smart contracts, these proposals mostly focus on proving the correctness or absence of a certain type of vulnerability within a contract, but cannot protect deployed (legacy) contracts from being exploited. In this paper, we address this problem in the context of re-entrancy exploits and propose a novel smart contract security technology, dubbed Sereum (Secure Ethereum), which protects existing, deployed contracts against re-entrancy attacks in a backwards compatible way based on run-time monitoring and validation. Sereum does neither require any modification nor any semantic knowledge of existing contracts. By means of implementation and evaluation using the Ethereum blockchain, we show that Sereum covers the actual execution flow of a smart contract to accurately detect and prevent attacks with a false positive rate as small as 0.06% and with negligible run-time overhead. As a by-product, we develop three advanced re-entrancy attacks to demonstrate the limitations of existing offline vulnerability analysis tools."}, "keywords": ["analysis tools", "fewer false alarms"], "citation_intent": "result"}
{"citing_id": "2304.01200v1", "cited_id": "1405.0312", "section_title": "Experiment 3: Open-World Detection Setting", "citation": "The OWOD problem has recently gained popularity with evaluations being performed on the challenging MS COCO dataset #REFR .", "text_before_citation": ["Lastly, we also validate the generalizability of our feature enrichment and the STO module for open-world object detection (OWOD) in images."], "text_after_citation": ["To this end, we integrate our feature enrichment and STO module (replacing 3D convolutions with 2D) into the recent transformersbased OWOD framework, named OW-DETR #OTHEREFR . Tab.", "3 shows the results on the challenging MS COCO split introduced in #OTHEREFR .", "For a fair comparison, we employ the same self-supervised ResNet50 backbone as in the standard OW-DETR #OTHEREFR .", "We report the results of ORE-EBUI #OTHEREFR and our baseline OW-DETR from #OTHEREFR .", "In the case of Task-1, our approach achieves an impressive performance particularly on the 'Unknown' object class by increasing the U-Recall from 5.7% to 8.8%, while also improving the detection performance on the 'Current Known' classes."], "citing_paper_content": {"title": "Video Instance Segmentation In An Open-World", "abstract": "Existing video instance segmentation (VIS) approaches generally follow a closed-world assumption, where only seen category instances are identified and spatio-temporally segmented at inference. Open-world formulation relaxes the close-world static-learning assumption as follows: (a) first, it distinguishes a set of known categories as well as labels an unknown object as 'unknown' and then (b) it incrementally learns the class of an unknown as and when the corresponding semantic labels become available. We propose the first open-world VIS approach, named OW-VISFormer, that introduces a novel feature enrichment mechanism and a spatio-temporal objectness (STO) module. The feature enrichment mechanism based on a lightweight auxiliary network aims at accurate pixel-level (unknown) object delineation from the background as well as distinguishing category-specific known semantic classes. The STO module strives to generate instance-level pseudolabels by enhancing the foreground activations through a contrastive loss. Moreover, we also introduce an extensive experimental protocol to measure the characteristics of OW-VIS. Our OW-VISFormer performs favorably against a solid baseline in OW-VIS setting. Further, we evaluate our contributions in the standard fully-supervised VIS setting by integrating them into the recent SeqFormer, achieving an absolute gain of 1.6% AP on Youtube-VIS 2019 val. set. Lastly, we show the generalizability of our contributions for the open-world detection (OWOD) setting, outperforming the best existing OWOD method in the literature."}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["dataset"], "citation_intent": "background"}
{"citing_id": "2303.01794v2", "cited_id": "1810.04805", "section_title": "Related Work", "citation": "They fine-tuned mBERT model #REFR in a crosslingual setup and in a multi-task setup, and found that the results are mixed in terms of their benefits.", "text_before_citation": ["There exist several studies on multilingual and/or multi-task learning in the context of media analysis. #OTHEREFR", "(2021) applied cross-lingual training to multilingual false information detection and showed that cross-lingual training results in on par performance to monolingual training. Alam et al.", "(2021a,b) annotated Tweets in multiple languages with multiple choice questions regarding COVID-19 disinformation."], "text_after_citation": ["The SemEval 2023 Task 3 dataset (Piskorski et al., 2023) exhibits significantly different properties to the datasets previous works used and we found that cross-lingual and/or multi-task learning help in this dataset.", "We also carried out more thorough experiments than the previous works, which might have been the key to the improvement.", "There also exist different approaches for multilingual and low-resource settings in broader domains.", "#OTHEREFR proposed a method for training multilingual sentence embeddings and demonstrated its effectiveness in 50+ languages. In a low-resource setup, Heinisch et al.", "(2022) used a data augmentation approach that retrieves additional data from related datasets by automatically labeling them."], "citing_paper_content": {"title": "Hitachi At Semeval-2023 Task 3: Exploring Cross-Lingual Multi-Task Strategies For Genre And Framing Detection In Online News", "abstract": "This paper explains the participation of team Hitachi to SemEval-2023 Task 3 \"Detecting the genre, the framing, and the persuasion techniques in online news in a multilingual setup.\" Based on the multilingual, multi-task nature of the task and the low-resource setting, we investigated different cross-lingual and multi-task strategies for training the pretrained language models. Through extensive experiments, we found that (a) cross-lingual/multi-task training, and (b) collecting an external balanced dataset, can benefit the genre and framing detection. We constructed ensemble models from the results and achieved the highest macro-averaged F1 scores in Italian and Russian genre categorization subtasks. The target articles are those identified to be"}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["crosslingual setup"], "citation_intent": "result"}
{"citing_id": "2304.09540v1", "cited_id": "1909.04559", "section_title": "Approach Using A Modified Wta Assumption", "citation": "With the new WTA assumption, the learning analysis for general hierarchies follows the same pattern as the analysis for tree hierarchies in #REFR , and yields the same time bound.", "text_before_citation": ["Therefore, they satisfy the WTA requirement to be candidate neurons, provided that they have strictly more than o \u2022 k incoming neighbors that fire.", "But we know that, during the learning process, they have k incoming neighbors that fire (in the noise-free case), which is > o \u2022 k by the assumption that o < 1.", "A similar argument holds for the noisy case, based on having at least pk incoming neighbors that fire, which is > o \u2022 k by the assumption that o < p.", "For neurons that were not previously engaged, Lemma 7.4, Property 1, implies that the incoming potential is exactly kw.", "For the single neuron that was previously engaged for c, Lemma 7.4, Property 2 implies that the incoming potential is strictly greater than kw. So the WTA rule selects the previously-engaged neuron."], "text_after_citation": ["Implementing Assumption 7.3 will require some additional mechanism, in addition to the mechanisms that are used to implement the basic WTA satisfying Assumption 7.2.", "Such a mechanism could serve as a pre-processing step, before the basic WTA.", "The new mechanism could allow a layer neuron u to fire (and somehow reflect its incoming potential) exactly if u has strictly more than o \u2022 k incoming neighbors that contribute potential \u2265 w to u. 13", "Alternative modifications to the WTA: The modification in Section 7.2.2 allows the WTA to explicitly reject candidate neurons having too few incoming neighbors whose edge weight is greater than or equal to the initial weight w. There might be other ways to modify the WTA.", "For example, we might assume that the WTA can remember when it has already selected a neuron to represent a particular concept, and not select it again for a different concept."], "citing_paper_content": {"title": "Learning Hierarchically-Structured Concepts Ii: Overlapping Concepts, And Networks With Feedback", "abstract": "We continue our study from [5], of how concepts that have hierarchical structure might be represented in brain-like neural networks, how these representations might be used to recognize the concepts, and how these representations might be learned. In [5], we considered simple tree-structured concepts and feed-forward layered networks. Here we extend the model in two ways: we allow limited overlap between children of different concepts, and we allow networks to include feedback edges. For these more general cases, we describe and analyze algorithms for recognition and algorithms for learning. We also use the abbreviations supp r,f (B, * , t) for S(, t), supp r,f (B, , *) for t S(, t), and supp r,f (B, * , *) for ,t S(, t), Notice that each of these three unions must converge to a finite set since all the sets S(, t) are subsets of the single finite set C of concepts. Now we have two monotonicity results, for r and f : Lemma 2.5. Let C be any concept hierarchy satisfying the limited-overlap property, and let B \u2286 D 0. Consider r, r , where 0 \u2264 r \u2264 r \u2264 1, and arbitrary f. Then supp r ,f (B) \u2286 supp r,f (B). Lemma 2.6. Let C be any concept hierarchy satisfying the limited-overlap property, and let B \u2286 D 0. Consider f, f , where 0 \u2264 f \u2264 f , and arbitrary r \u2208 [0, 1]. Then supp r,f (B) \u2286 supp r,f (B)."}, "cited_paper_content": {"title": "Learning Hierarchically Structured Concepts", "abstract": "We study the question of how concepts that have structure get represented in the brain. Specifically, we introduce a model for hierarchically structured concepts and we show how a biologically plausible neural network can recognize these concepts, and how it can learn them in the first place. Our main goal is to introduce a general framework for these tasks and prove formally how both (recognition and learning) can be achieved. ::: We show that both tasks can be accomplished even in presence of noise. For learning, we analyze Oja's rule formally, a well-known biologically-plausible rule for adjusting the weights of synapses. We complement the learning results with lower bounds asserting that, in order to recognize concepts of a certain hierarchical depth, neural networks must have a corresponding number of layers."}, "keywords": ["tree hierarchies"], "citation_intent": "result"}
{"citing_id": "2304.05398v1", "cited_id": "1905.11261", "section_title": "Convergence Of Fb-Gvi", "citation": "The convergence rates are of the same order as the convergence rates of the deterministic forward-backward algorithm over R d #REFR .", "text_before_citation": ["\u03b7 = 1 \u03b2 and N \u03b2 \u03b1 log W2(p0,\u03c0) \u03b5", ", we obtain the guarantees", "\u03b1 W 2 2 (\u00b5 N ,\u03c0) \u2264 \u03b5 2 , and F(p 2N ) \u2212 F(\u03c0) \u2264 \u03b5 2 .", "Proof. The proof is given in Appendix E.3.", "Theorem 5.2 states the sublinear convergence of FB-GVI for a convex V (in terms of objective gap) and Theorem 5.4 states the linear convergence of FB-GVI for a strongly convex V ."], "text_after_citation": ["Finally, we also extend our results to the non-convex case, where we obtain a stationary point guarantee.", "Theorem 5.5 (Non-convex case, FB-GVI). Suppose that V is \u03b2-smooth, and", "that 0 < \u03b7 \u2264 1 \u03b2 . Let \u2206 := F(p 0 ) \u2212 F(\u03c0).", "Then, min k\u2208{0,...,N \u22121} \u2207 BW F(p k ) 2 p k \u2264 150\u2206 \u03b7N .", "In particular, when \u03b7 = 1 \u03b2 and N \u03b2\u2206 \u03b5 2 , we obtain the guarantee"], "citing_paper_content": {"title": "Forward-Backward Gaussian Variational Inference Via Jko In The Bures-Wasserstein Space", "abstract": "Variational inference (VI) seeks to approximate a target distribution \u03c0 by an element of a tractable family of distributions. Of key interest in statistics and machine learning is Gaussian VI, which approximates \u03c0 by minimizing the Kullback-Leibler (KL) divergence to \u03c0 over the space of Gaussians. In this work, we develop the (Stochastic) Forward-Backward Gaussian Variational Inference (FB-GVI) algorithm to solve Gaussian VI. Our approach exploits the composite structure of the KL divergence, which can be written as the sum of a smooth term (the potential) and a non-smooth term (the entropy) over the Bures-Wasserstein (BW) space of Gaussians endowed with the Wasserstein distance. For our proposed algorithm, we obtain state-of-the-art convergence guarantees when \u03c0 is log-smooth and log-concave, as well as the first convergence guarantees to first-order stationary solutions when \u03c0 is only log-smooth."}, "cited_paper_content": {"title": "A Unified Theory Of Sgd: Variance Reduction, Sampling, Quantization And Coordinate Descent", "abstract": "In this paper we introduce a unified analysis of a large family of variants of proximal stochastic gradient descent ({\\tt SGD}) which so far have required different intuitions, convergence analyses, have different applications, and which have been developed separately in various communities. We show that our framework includes methods with and without the following tricks, and their combinations: variance reduction, importance sampling, mini-batch sampling, quantization, and coordinate sub-sampling. As a by-product, we obtain the first unified theory of {\\tt SGD} and randomized coordinate descent ({\\tt RCD}) methods, the first unified theory of variance reduced and non-variance-reduced {\\tt SGD} methods, and the first unified theory of quantized and non-quantized methods. A key to our approach is a parametric assumption on the iterates and stochastic gradients. In a single theorem we establish a linear convergence result under this assumption and strong-quasi convexity of the loss function. Whenever we recover an existing method as a special case, our theorem gives the best known complexity result. Our approach can be used to motivate the development of new useful methods, and offers pre-proved convergence guarantees. To illustrate the strength of our approach, we develop five new variants of {\\tt SGD}, and through numerical experiments demonstrate some of their properties."}, "keywords": ["deterministic forward-backward algorithm", "convergence rates"], "citation_intent": "result"}
{"citing_id": "2303.11726v1", "cited_id": "1701.01370", "section_title": "B.3 Additional Quantitative Results", "citation": "Table 10 shows the metrics of the fitted mesh on the SURREAL #REFR test set.", "text_before_citation": ["This may be because the arm has larger variations in the dataset.", "Comparison to fitting.", "In order to disentangle the ability of mesh regression from markers using\u00c2 with the ability to detect the virtual markers accurately from images, we first compute the estimation errors of the virtual markers.", "The MPJPE over all the virtual markers is 35.5mm, which demonstrates that these virtual markers can be accurately detected from the images.", "We then fit the mesh model parameters to these virtual markers."], "text_after_citation": ["As we can see, the fitted mesh has a similar error as our regression ones which uses the interpolation matrix\u00c2, which validates the accuracy of the estimated virtual markers.", "Figure 12 shows more qualitative comparisons with Pose2Mesh #OTHEREFR on the SURREAL test set in which has diverse body shapes.", "The skeleton representation used in Pose2Mesh loses the body shape information so the method #OTHEREFR can only recover mean shapes.", "For example, in Figure 12 (d) (e), the estimated meshes of Pose2Mesh tend to have the average body shape and fail to estimate the real body shape, regardless of whether the person is slim or stout.", "This is caused by the limited skeleton representation bottleneck so that the model learns a mean shape for the whole training dataset implicitly."], "citing_paper_content": {"title": "3D Human Mesh Estimation From Virtual Markers", "abstract": "Inspired by the success of volumetric 3D pose estimation, some recent human mesh estimators propose to estimate 3D skeletons as intermediate representations, from which, the dense 3D meshes are regressed by exploiting the mesh topology. However, body shape information is lost in extracting skeletons, leading to mediocre performance. The advanced motion capture systems solve the problem by placing dense physical markers on the body surface, which allows to extract realistic meshes from their non-rigid motions. However, they cannot be applied to wild images without markers. In this work, we present an intermediate representation, named virtual markers, which learns 64 landmark keypoints on the body surface based on the large-scale mocap data in a generative style, mimicking the effects of physical markers. The virtual markers can be accurately detected from wild images and can reconstruct the intact meshes with realistic shapes by simple interpolation. Our approach outperforms the state-of-the-art methods on three datasets. In particular, it surpasses the existing methods by a notable margin on the SURREAL dataset, which has diverse body shapes. Code is available at https: //github.com/ShirleyMaxx/VirtualMarker."}, "cited_paper_content": {"title": "Learning From Synthetic Humans", "abstract": "Estimating human pose, shape, and motion from images and videos are fundamental challenges with many applications. Recent advances in 2D human pose estimation use large amounts of manually-labeled training data for learning convolutional neural networks (CNNs). Such data is time consuming to acquire and difficult to extend. Moreover, manual labeling of 3D pose, depth and motion is impractical. In this work we present SURREAL (Synthetic hUmans foR REAL tasks): a new large-scale dataset with synthetically-generated but realistic images of people rendered from 3D sequences of human motion capture data. We generate more than 6 million frames together with ground truth pose, depth maps, and segmentation masks. We show that CNNs trained on our synthetic dataset allow for accurate human depth estimation and human part segmentation in real RGB images. Our results and the new dataset open up new possibilities for advancing person analysis using cheap and large-scale synthetic data."}, "keywords": ["fitted mesh", "SURREAL test set"], "citation_intent": "background"}
{"citing_id": "2303.03443v1", "cited_id": "1810.01969", "section_title": "Introduction", "citation": "In a recent paper of Guruswami, Nakkiran, and Sudan #REFR , they consider hidden Markov channels, where the states evolve according to some Markov process.", "text_before_citation": ["The problem of designing coding schemes for stochastic channels is hard."], "text_after_citation": ["The \"hidden\" refers to the fact that one cannot determine the state from the output of the channel.", "One important thing to note here is that error introduced by the channel is only dependent on the state of the Markov process.", "In designing encoding and decoding algorithms for the problem of hidden Markov channels, Markovian sources become quite important due to a well-known correspondence relating error-correction for additive Markov chains to compression and decompression algorithms for Markovian sources (when the compression is linear).", "Recall that additive channels are the those which map inputs from some alphabet \u03a3 to outputs over \u03a3, where there is an abelian group defined on \u03a3 (needed for the additive purposes).", "The channel creates an error sequence from the associated Markov source which is independent of the input sequence."], "citing_paper_content": {"title": "Improving The Runtime Of Algorithmic Polarization Of Hidden Markov Models", "abstract": "We improve the runtime of the linear compression scheme for hidden Markov sources presented in a 2018 paper of Guruswami, Nakkiran, and Sudan. Under the previous scheme, compressing a message of length n takes O(n log n) runtime, and decompressing takes O(n 1+\u03b4) runtime for any fixed \u03b4 > 0. We present how to improve the runtime of the decoding scheme to O(n log n) by caching intermediate results to avoid repeating computation."}, "cited_paper_content": {"title": "Algorithmic Polarization For Hidden Markov Models", "abstract": "Using a mild variant of polar codes we design linear compression schemes compressing Hidden Markov sources (where the source is a Markov chain, but whose state is not necessarily observable from its output), and to decode from Hidden Markov channels (where the channel has a state and the error introduced depends on the state). We give the first polynomial time algorithms that manage to compress and decompress (or encode and decode) at input lengths that are polynomial $\\it{both}$ in the gap to capacity and the mixing time of the Markov chain. Prior work achieved capacity only asymptotically in the limit of large lengths, and polynomial bounds were not available with respect to either the gap to capacity or mixing time. Our results operate in the setting where the source (or the channel) is $\\it{known}$. If the source is $\\it{unknown}$ then compression at such short lengths would lead to effective algorithms for learning parity with noise -- thus our results are the first to suggest a separation between the complexity of the problem when the source is known versus when it is unknown."}, "keywords": ["hidden Markov channels"], "citation_intent": "background"}
{"citing_id": "2303.16637v1", "cited_id": "1801.05124", "section_title": "Active Learning", "citation": "The work of #REFR introduced two metrics called \"localization tightness\" and \"localization stability\" to measure the informativeness of object localization.", "text_before_citation": ["(b) Combine uncertainty-based scores and category re-weighting approach to evaluate the informative region score (Sec. 4.3).", "(c) Then, a scale-aware selection method is used to select the diverse regions and allocate labeling budgets for each scale (Sec. 4.4).", "(d) The selected regions are annotated from a human expert, and subsequently updated to the labeled set (Sec. 4.5). The process repeats for a specified number of cycles.", "fication scores as metrics to measure model uncertainty, as seen in works such as #OTHEREFR .", "#OTHEREFR proposed an entropybased Non-Maximum Suppression method that considers both instance-level uncertainty and diversity."], "text_after_citation": ["Furthermore, some researchers have focused on combining the uncertainty of localization with the uncertainty of classification.", "For example, [4] proposed a mixture density network to aggregate these two sources of uncertainty, and #OTHEREFR considered the consistency of localization and predicted class distributions in augmented and unlabeled images.", "Fine-grained active learning strategies query partial labels within specific instances or regions, instead of selecting the entire image.", "For instance, #OTHEREFR proposed a self-supervised sample mining technique that uses pseudolabeling with high consistency samples and an active learning strategy to select low consistency samples within crossimage validation.", "[6] employed the coreset method #OTHEREFR and classification uncertainty to select instance labels."], "citing_paper_content": {"title": "Mural: Multi-Scale Region-Based Active Learning For Object Detection", "abstract": "Obtaining large-scale labeled object detection dataset can be costly and time-consuming, as it involves annotating images with bounding boxes and class labels. Thus, some specialized active learning methods have been proposed to reduce the cost by selecting either coarse-grained samples or fine-grained instances from unlabeled data for labeling. However, the former approaches suffer from redundant labeling, while the latter methods generally lead to training instability and sampling bias. To address these challenges, we propose a novel approach called Multiscale Region-based Active Learning (MuRAL) for object detection. MuRAL identifies informative regions of various scales to reduce annotation costs for well-learned objects and improve training performance. The informative region score is designed to consider both the predicted confidence of instances and the distribution of each object category, enabling our method to focus more on difficult-to-detect classes. Moreover, MuRAL employs a scale-aware selection strategy that ensures diverse regions are selected from different scales for labeling and downstream finetuning, which enhances training stability. Our proposed method surpasses all existing coarse-grained and fine-grained baselines on Cityscapes and MS COCO datasets, and demonstrates significant improvement in difficult category performance."}, "cited_paper_content": {"title": "Localization-Aware Active Learning For Object Detection", "abstract": "Active learning - a class of algorithms that iteratively searches for the most informative samples to include in a training dataset - has been shown to be effective at annotating data for image classification. However, the use of active learning for object detection is still largely unexplored as determining informativeness of an object-location hypothesis is more difficult. In this paper, we address this issue and present two metrics for measuring the informativeness of an object hypothesis, which allow us to leverage active learning to reduce the amount of annotated data needed to achieve a target object detection performance. Our first metric measures 'localization tightness' of an object hypothesis, which is based on the overlapping ratio between the region proposal and the final prediction. Our second metric measures 'localization stability' of an object hypothesis, which is based on the variation of predicted object locations when input images are corrupted by noise. Our experimental results show that by augmenting a conventional active-learning algorithm designed for classification with the proposed metrics, the amount of labeled training data required can be reduced up to 25%. Moreover, on PASCAL 2007 and 2012 datasets our localization-stability method has an average relative improvement of 96.5% and 81.9% over the baseline method using classification only."}, "keywords": ["object localization"], "citation_intent": "background"}
{"citing_id": "2304.02590v1", "cited_id": "1804.00553", "section_title": "Related Work", "citation": "We remark that the notion of robustness studied in #REFR was quite different from that of the previous works as detailed below.", "text_before_citation": ["The two topics, of stable matching and the design of algorithms that produce solutions that are robust to errors, have been studied extensively for decades and there are today several books on each of them, e.g., see #OTHEREFR and #OTHEREFR .", "Yet, there is a paucity of results at the intersection of these two topics.", "Indeed, before the publication of #OTHEREFR , we are aware of only a few previous works [ABG + 16, ABF + 17, CSS21]."], "text_after_citation": ["Aziz et al.", "[ABG + 16] considered the problem of finding stable matching under uncertain linear preferences. They proposed three different uncertainty models:", "1.", "Lottery Model: Each agent has a probability distribution over strict preference lists, independent of other agents.", "2."], "citing_paper_content": {"title": "A Structural And Algorithmic Study Of Stable Matching Lattices Of Multiple Instances", "abstract": "Recently [MV18a] identified and initiated work on the new problem of understanding structural relationships between the lattices of solutions of two \"nearby\" instances of stable matching. They also gave an application of their work to finding a robust stable matching. However, the types of changes they allowed in going from instance A to B were very restricted, namely any one agent executes an upward shift. In this paper, we allow any one agent to permute its preference list arbitrarily. Let M A and M B be the sets of stable matchings of the resulting pair of instances A and B, and let L A and L B be the corresponding lattices of stable matchings. We prove that the matchings in M A \u2229 M B form a sublattice of both L A and L B and those in M A \\ M B form a join semi-sublattice of L A. These properties enable us to obtain a polynomial time algorithm for not only finding a stable matching in M A \u2229 M B , but also for obtaining the partial order, as promised by Birkhoff's Representation Theorem [Bir37], thereby enabling us to generate all matchings in this sublattice. Our algorithm also helps solve a version of the robust stable matching problem. We discuss another potential application, namely obtaining new insights into the incentive compatibility properties of the Gale-Shapley Deferred Acceptance Algorithm."}, "cited_paper_content": {"title": "Finding Stable Matchings That Are Robust To Errors In The Input", "abstract": "We study the problem of finding solutions to the stable matching problem that are robust to errors in the input and we obtain a polynomial time algorithm for a special class of errors. In the process, we also initiate work on a new structural question concerning the stable matching problem, namely finding relationships between the lattices of solutions of two\"nearby\"instances. Our main algorithmic result is the following: We identify a polynomially large class of errors, $D$, that can be introduced in a stable matching instance. Given an instance $A$ of stable matching, let $B$ be the random variable that represents the instance that results after introducing {\\em one} error from $D$, chosen via a given discrete probability distribution. The problem is to find a stable matching for $A$ that maximizes the probability of being stable for $B$ as well. Via new structural properties of the type described in the question stated above, we give a combinatorial polynomial time algorithm for this problem. We also show that the set of robust stable matchings for instance $A$, under probability distribution $p$, forms a sublattice of the lattice of stable matchings for $A$. We give an efficient algorithm for finding a succinct representation for this set; this representation has the property that any member of the set can be efficiently retrieved from it."}, "keywords": ["robustness"], "citation_intent": "result"}
{"citing_id": "2304.13115v1", "cited_id": "1812.07603", "section_title": "Related Work", "citation": "FML #REFR proposes a self-supervised method to learn identity and appearance models from videos.", "text_before_citation": ["Recently, deep learning approaches use the 3DMM fitting as prior and learn additional corrective displacements or normal maps #OTHEREFR , in order to capture missing facial details.", "DECA #OTHEREFR and EMOCA #OTHEREFR first predict the parameters of FLAME #OTHEREFR , and then learn a UV-map of vertex displacements. Tran et al.", "#OTHEREFR borrow details from reference bump maps in order to complete occluded facial regions.", "Several subject-specific methods leverage multi-view images #OTHEREFR or learn implicit representations using multi-layer perceptrons (MLPs) #OTHEREFR . However, these methods cannot generalize to multiple identities.", "Monocular videobased optimization techniques [24-27, 57, 60] leverage the multi-frame consistency to learn facial details."], "text_after_citation": ["However, all these videoonly methods ignore the audio modality that is usually included in most cases.", "In addition, they do not explicitly handle occlusions that frequently appear in talking face videos.", "Audio-Driven 3D Face Animation.", "Another line of work addresses the problem of audio-driven 3D facial animation. Busso et al.", "#OTHEREFR use hidden Markov models (HMMs) to model the temporal relation between prosodic audio features and head motion sequences."], "citing_paper_content": {"title": "Avface: Towards Detailed Audio-Visual 4D Face Reconstruction", "abstract": "Figure 1. Given a monocular talking face video and the corresponding speech segment, AVFace reconstructs the detailed 4D face geometry for any input speaker, and it is robust in cases when either modality is insufficient."}, "cited_paper_content": {"title": "Fml: Face Model Learning From Videos", "abstract": "Monocular image-based 3D reconstruction of faces is a long-standing problem in computer vision. Since image data is a 2D projection of a 3D face, the resulting depth ambiguity makes the problem ill-posed. Most existing methods rely on data-driven priors that are built from limited 3D face scans. In contrast, we propose multi-frame video-based self-supervised training of a deep network that (i) learns a face identity model both in shape and appearance while (ii) jointly learning to reconstruct 3D faces. Our face model is learned using only corpora of in-the-wild video clips collected from the Internet. This virtually endless source of training data enables learning of a highly general 3D face model. In order to achieve this, we propose a novel multi-frame consistency loss that ensures consistent shape and appearance across multiple frames of a subject's face, thus minimizing depth ambiguity. At test time we can use an arbitrary number of frames, so that we can perform both monocular as well as multi-frame reconstruction."}, "keywords": ["appearance models"], "citation_intent": "method"}
{"citing_id": "2303.06266v1", "cited_id": "0907.1061", "section_title": "C. Minimum User Identification Cost", "citation": "Due to the equivalence of active user identification and GT, we use an adaptation of the GT decoding strategy in #REFR to the context of user identification in non-coherent ( , k)-MnAC.", "text_before_citation": ["Proof: Using standard capacity arguments #OTHEREFR along with Lemma 1, it is evident that user identification cost in the non-coherent ( , k)-MnAC must obey n( )", "\u2265 log ( k ) C", "where C is as given in Lemma 1. Thus, in the k = \u0398(1) regime we have", "EQUATION", "Next, we present a decoder as shown in Algorithm 1 which can achieve the equality in #OTHEREFR relying on classical thresholding techniques in information theory #OTHEREFR ."], "text_after_citation": ["The algorithm operates as follows: Given the set of preambles P and the detector outputs Z, Number of threshold tests passed, n t = 0."], "citing_paper_content": {"title": "Non-Coherent Active Device Identification For Massive Random Access", "abstract": "Massive Machine-Type Communications (mMTC) is a key service category in the current generation of wireless networks featuring an extremely high density of energy and resource-limited devices with sparse and sporadic activity patterns. In order to enable random access in such mMTC networks, base station needs to identify the active devices while operating within stringent access delay constraints. In this paper, an energy efficient active device identification protocol is proposed in which active devices transmit On-Off Keying (OOK) modulated preambles jointly and base station employs non-coherent energy detection avoiding channel estimation overheads. The minimum number of channel-uses required by the active user identification protocol is characterized in the asymptotic regime of total number of devices when the number of active devices k scales as k = \u0398(1) along with an achievability scheme relying on the equivalence of activity detection to a group testing problem. Several practical schemes based on Belief Propagation (BP) and Combinatorial Orthogonal Matching Pursuit (COMP) are also proposed. Simulation results show that BP strategies outperform COMP significantly and can operate close to the theoretical achievability bounds. In a partial-recovery setting where few misdetections are allowed, BP continues to perform well."}, "cited_paper_content": {"title": "Boolean Compressed Sensing And Noisy Group Testing", "abstract": "The fundamental task of group testing is to recover a small distinguished subset of items from a large population while efficiently reducing the total number of tests (measurements). The key contribution of this paper is in adopting a new information-theoretic perspective on group testing problems. We formulate the group testing problem as a channel coding/decoding problem and derive a single-letter characterization for the total number of tests used to identify the defective set. Although the focus of this paper is primarily on group testing, our main result is generally applicable to other compressive sensing models."}, "keywords": ["active user identification", "user identification"], "citation_intent": "method"}
{"citing_id": "2303.18232v1", "cited_id": "1608.03983", "section_title": "Conclusion", "citation": "We warm up the training for 4 epochs and then we decay the learning rate using a cosine schedule #REFR .", "text_before_citation": ["We apply the basic data augmentations (only random cropping, flipping and data whitening) to the input images when computing the student features.", "We empirically find out more advanced data augmentations (such as \"rand-m9-n3-mstd0.5\") harm the distillation performance.", "We argue that the advanced data augmentations for the student model's input images enlarge the discrepancy of the student feature and the fixed pre-extracted teacher feature, which lead to the poor performance in vision-language knowledge distillation. We train all models for 100 epochs.", "During the training, we use the Adam optimizer #OTHEREFR with decoupled weight decay regularization [57] .", "We set the initial learning rate as 8 \u00d7 10 \u22124 and weight decay 0.05."], "text_after_citation": ["We remove the weight decay of the weights that are the gains or biases.", "For our model Distill-ViT-B/32, we optimize with the batch size 12288 for images and texts. For Distill-UniCL*, we adopt 8192 as the batch size.", "We grid search for the best hyperparameters \u00b5 vl , \u00b5 p-vl and \u00b5 udist when performing ablation studies on losses in Sec 4.5 (main paper).", "We finally set \u00b5 vl = 100, \u00b5 p-vl = 33.3 and \u00b5 udist = 14.3 for all experiments."], "citing_paper_content": {"title": "Dime-Fm : Distilling Multimodal And Efficient Foundation Models", "abstract": "Large Vision-Language Foundation Models (VLFM), such as CLIP, ALIGN and Florence, are trained on largescale datasets of image-caption pairs and achieve superior transferability and robustness on downstream tasks, but they are difficult to use in many practical applications due to their large size, high latency and fixed architectures. Unfortunately, recent work shows training a small custom VLFM for resource-limited applications is currently very difficult using public and smaller-scale data. In this paper, we introduce a new distillation mechanism (DIME-FM) that allows us to transfer the knowledge contained in large VLFMs to smaller, customized foundation models using a relatively small amount of inexpensive, unpaired images and sentences. We transfer the knowledge from the pre-trained CLIP-ViT-L/14 model to a ViT-B/32 model, with only 40M public images and 28.4M unpaired public sentences. The resulting model \"Distill-ViT-B/32\" rivals the CLIP-ViT-B/32 model pre-trained on its private WiT dataset (400M image-text pairs): Distill-ViT-B/32 achieves similar results in terms of zero-shot and linear-probing performance on both Ima-geNet and the ELEVATER (20 image classification tasks) benchmarks. It also displays comparable robustness when evaluated on five datasets with natural distribution shifts from ImageNet."}, "cited_paper_content": {"title": "Sgdr: Stochastic Gradient Descent With Warm Restarts", "abstract": "Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets, where we demonstrate new state-of-the-art results at 3.14% and 16.21%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at https://github.com/loshchil/SGDR"}, "keywords": ["training", "learning rate"], "citation_intent": "method"}
{"citing_id": "2304.02982v1", "cited_id": "1812.04948", "section_title": "Vipprint Dataset", "citation": "The face images for this dataset are originally from Flickr-Faces-HQ (FFHQ) dataset #REFR which provides a large human face images.", "text_before_citation": ["Also, the fact that there are many large face datasets available that can be beneficial for further studies.", "besides, face is an important human attribute as progresses in AI", "D R A F T 6 E. Nowroozi et al.", "resulted in advanced synthetic human face image generator models.", "This phenomenon makes it more important to have countermeasures to avoid malicious activities in MFs #OTHEREFR ."], "text_after_citation": ["Some sample images of FFHQ dataset are shown in Figure 1 .", "They provided images given from multiple printers as reported in the Table 1 .", "All printers are laser type, and 200 images are used from each.", "They printed 200 images on fifty A4 papers, each paper including 4 images, using landscape orientation.", "They assumed that 200 images may not be enough to deploy them for many deep learning techniques, so they provided another set of images which included Regions of Interest (ROI) extracted from the original images."], "citing_paper_content": {"title": "D R A F T Spritz-Ps: Validation Of Synthetic Face Images Using A Large Dataset Of Printed Documents", "abstract": "The capability of doing effective forensic analysis on printed and scanned (PS) images is essential in many applications. PS documents may be used to conceal the artifacts of images which is due to the synthetic nature of images since these artifacts are typically present in manipulated images and the main artifacts in the synthetic images can be removed after the PS. Due to the appeal of Generative Adversarial Networks (GANs), synthetic face images generated with GANs models are difficult to differentiate from genuine human faces and may be used to create counterfeit identities. Additionally, since GANs models do not account for physiological constraints for generating human faces and their impact on human IRISes, distinguishing genuine from synthetic IRISes in the PS scenario becomes extremely difficult. As a result of the lack of large-scale reference IRIS datasets in the PS scenario, we aim at developing a novel dataset to become a standard for Multimedia Forensics (MFs) investigation which is available at [45]. In this paper, we provide a novel dataset made up of a large number of synthetic and natural printed IRISes taken from VIPPrint Printed and Scanned face images. We extracted irises from face images and it is possible that the model due to eyelid occlusion captured the incomplete irises. To fill the missing pixels of extracted iris, we applied techniques to discover the complex link between the iris images. To highlight the problems involved with the evaluation of the dataset's IRIS images, we conducted a large number of analyses employing Siamese Neural Networks to assess the similarities between genuine and synthetic human IRISes, such as ResNet50, Xception, VGG16, and MobileNet-v2. For instance, using the Xception network, we achieved 56.76% similarity of IRISes for synthetic images and 92.77% similarity of IRISes for real images."}, "cited_paper_content": {"title": "A Style-Based Generator Architecture For Generative Adversarial Networks", "abstract": "We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces."}, "keywords": ["face images"], "citation_intent": "method"}
{"citing_id": "2303.04592v1", "cited_id": "2002.03647", "section_title": "Skill Discovery", "citation": "Consistent with prior work #REFR the skill discovery problem is formalized as a Markov Decision Process (MDP) M = (S, A, P) without external rewards, S and A respectively denote the state and action spaces, and P is the transition function. Skills introduced by Sutton et al.", "text_before_citation": [], "text_after_citation": ["#OTHEREFR are temporally extended actions (sub-behavior), which consist of a sequence of primitive actions.", "We define skills as policies ( | , ) conditioned on a fixed latent variable \u2208 .", "Skill discovery methods aim to learn these latent-conditioned policies by maximising the mutual information between S and .", "Due to symmetry, the corresponding mutual information can be expressed in two forms:", "(S; ) = ( ) \u2212 ( |S) reverse = (S) \u2212 (S| ) forward (1)"], "citing_paper_content": {"title": "Controlled Diversity With Preference : Towards Learning A Diverse Set Of Desired Skills", "abstract": "Autonomously learning diverse behaviors without an extrinsic reward signal has been a problem of interest in reinforcement learning. However, the nature of learning in such mechanisms is unconstrained, often resulting in the accumulation of several unusable, unsafe or misaligned skills. In order to avoid such issues and ensure the discovery of safe and human-aligned skills, it is necessary to incorporate humans into the unsupervised training process, which remains a largely unexplored research area. In this work, we propose Controlled Diversity with Preference (CDP) 1 , a novel, collaborative human-guided mechanism for an agent to learn a set of skills that is diverse as well as desirable. The key principle is to restrict the discovery of skills to those regions that are deemed to be desirable as per a preference model trained using human preference labels on trajectory pairs. We evaluate our approach on 2D navigation and Mujoco environments and demonstrate the ability to discover diverse, yet desirable skills."}, "cited_paper_content": {"title": "Explore, Discover And Learn: Unsupervised Discovery Of State-Covering Skills", "abstract": "Acquiring abilities in the absence of a task-oriented reward function is at the frontier of reinforcement learning research. This problem has been studied through the lens of empowerment, which draws a connection between option discovery and information theory. Information-theoretic skill discovery methods have garnered much interest from the community, but little research has been conducted in understanding their limitations. Through theoretical analysis and empirical evidence, we show that existing algorithms suffer from a common limitation -- they discover options that provide a poor coverage of the state space. In light of this, we propose 'Explore, Discover and Learn' (EDL), an alternative approach to information-theoretic skill discovery. Crucially, EDL optimizes the same information-theoretic objective derived from the empowerment literature, but addresses the optimization problem using different machinery. We perform an extensive evaluation of skill discovery methods on controlled environments and show that EDL offers significant advantages, such as overcoming the coverage problem, reducing the dependence of learned skills on the initial state, and allowing the user to define a prior over which behaviors should be learned. Code is publicly available at https://github.com/victorcampos7/edl."}, "keywords": ["skill discovery problem"], "citation_intent": "method"}
{"citing_id": "2304.10515v1", "cited_id": "1210.3555", "section_title": "Core-Periphery Structure", "citation": "In the brain science field, it has been shown that brain dynamics has a core-periphery organization #REFR .", "text_before_citation": ["Core-periphery structure represents a relationship between nodes in a graph where the core nodes are densely connected with each other while periphery nodes are sparsely connected to the core nodes and among each other #OTHEREFR , #OTHEREFR .", "Core-periphery graph has been applied in a variety of fields, including social network analysis #OTHEREFR , #OTHEREFR , economics #OTHEREFR , biology such as modeling the structure of protein interaction networks #OTHEREFR ."], "text_after_citation": ["The functional brain networks also demonstrate a coreperiphery structure #OTHEREFR .", "A recent study revealed the coreperiphery characteristics of the human brain from a structural perspective #OTHEREFR .", "It is shown that gyri and sulci, two prominent cortical folding patterns, could cooperate as a core-periphery network which improves the efficiency of information transmission in the brain #OTHEREFR ."], "citing_paper_content": {"title": "Cp-Cnn: Core-Periphery Principle Guided Convolutional Neural Network", "abstract": "The evolution of convolutional neural networks (CNNs) can be largely attributed to the design of its architecture, i.e., the network wiring pattern. Neural architecture search (NAS) advances this by automating the search for the optimal network architecture, but the resulting network instance may not generalize well in different tasks. To overcome this, exploring network design principles that are generalizable across tasks is a more practical solution. In this study, We explore a novel brain-inspired design principle based on the core-periphery property of the human brain network to guide the design of CNNs. Our work draws inspiration from recent studies suggesting that artificial and biological neural networks may have common principles in optimizing network architecture. We implement the core-periphery principle in the design of network wiring patterns and the sparsification of the convolution operation. The resulting core-periphery principle guided CNNs (CP-CNNs) are evaluated on three different datasets. The experiments demonstrate the effectiveness and superiority compared to CNNs and ViT-based methods. Overall, our work contributes to the growing field of brain-inspired AI by incorporating insights from the human brain into the design of neural networks."}, "cited_paper_content": {"title": "Task-Based Core-Periphery Organization Of Human Brain Dynamics", "abstract": "As a person learns a new skill, distinct synapses, brain regions, and circuits are engaged and change over time. In this paper, we develop methods to examine patterns of correlated activity across a large set of brain regions. Our goal is to identify properties that enable robust learning of a motor skill. We measure brain activity during motor sequencing and characterize network properties based on coherent activity between brain regions. Using recently developed algorithms to detect time-evolving communities, we find that the complex reconfiguration patterns of the brain's putative functional modules that control learning can be described parsimoniously by the combined presence of a relatively stiff temporal core that is composed primarily of sensorimotor and visual regions whose connectivity changes little in time and a flexible temporal periphery that is composed primarily of multimodal association regions whose connectivity changes frequently. The separation between temporal core and periphery changes over the course of training and, importantly, is a good predictor of individual differences in learning success. The core of dynamically stiff regions exhibits dense connectivity, which is consistent with notions of core-periphery organization established previously in social networks. Our results demonstrate that core-periphery organization provides an insightful way to understand how putative functional modules are linked. This, in turn, enables the prediction of fundamental human capacities, including the production of complex goal-directed behavior."}, "keywords": ["brain science field", "brain dynamics"], "citation_intent": "background"}
{"citing_id": "2304.13761v1", "cited_id": "0811.1790", "section_title": "Regularization Effect On Robustness And Performance", "citation": "The proof when uncertainty set is (16) is similar to proof of Theorem 1 in #REFR .", "text_before_citation": ["or", "EQUATION", "EQUATION", "Proof.", "We show the proof when the uncertainty set is (15)."], "text_after_citation": ["max \u2206\u03a6\u2208U y \u2212 (\u03a6 + \u2206\u03a6)\u03b2 2 \u2264 y \u2212 \u03a6\u03b2 2 + max \u2206\u03a6\u2208U \u2206\u03a6\u03b2 2 \u2264 y \u2212 \u03a6\u03b2 2 + max \u2206\u03a6\u2208U \u2206\u03a6 2 \u03b2 2 = y \u2212 \u03a6\u03b2 2 + c \u03b2 2 2 Now, let u = y \u2212 \u03a6\u03b2 y \u2212 \u03a6\u03b2 2 , and let \u2206\u03a6 * i \u2212c\u03b2 i sgn (\u03b2 i ) u, we have \u2206\u03a6 * 2 = c \u03b2 2 , therefore \u2206\u03a6 * \u2208 U.", "Meanwhile, max \u2206\u03a6\u2208U y \u2212 (\u03a6 + \u2206\u03a6)\u03b2 2 \u2265 max \u2206\u03a6\u2208U y \u2212 (\u03a6 + \u2206\u03a6 * )\u03b2 2 = y \u2212 \u03a6\u03b2 + c \u03b2 2 2 u 2 = y \u2212 \u03a6\u03b2 2 + c \u03b2 2 2 .", "This completes the proof of the theorem.", "Here we do not make a definitive conclusion on which regularization method is better.", "The theoretical results are based on different assumptions that are usually hard to verify in practice."], "citing_paper_content": {"title": "Enhancing Robustness Of Gradient-Boosted Decision Trees Through One-Hot Encoding And Regularization *", "abstract": "Gradient-boosted decision trees (GBDT) are widely used and highly effective machine learning approach for tabular data modeling. However, their complex structure may lead to low robustness against small covariate perturbation in unseen data. In this study, we apply one-hot encoding to convert a GBDT model into a linear framework, through encoding of each tree leaf to one dummy variable. This allows for the use of linear regression techniques, plus a novel risk decomposition for assessing the robustness of a GBDT model against covariate perturbations. We propose to enhance the robustness of GBDT models by refitting their linear regression forms with L 1 or L 2 regularization. Theoretical results are obtained about the effect of regularization on the model performance and robustness. It is demonstrated through numerical experiments that the proposed regularization approach can enhance the robustness of the one-hot-encoded GBDT models. * The views expressed in the paper are those of the authors and do not represent the views of Wells Fargo."}, "cited_paper_content": {"title": "Robust Regression And Lasso", "abstract": "Lasso, or l1 regularized least squares, has been explored extensively for its remarkable sparsity properties. In this paper it is shown that the solution to Lasso, in addition to its sparsity, has robustness properties: it is the solution to a robust optimization problem. This has two important consequences. First, robustness provides a connection of the regularizer to a physical property, namely, protection from noise. This allows a principled selection of the regularizer, and in particular, generalizations of Lasso that also yield convex optimization problems are obtained by considering different uncertainty sets. Second, robustness can itself be used as an avenue for exploring different properties of the solution. In particular, it is shown that robustness of the solution explains why the solution is sparse. The analysis as well as the specific results obtained differ from standard sparsity results, providing different geometric intuition. Furthermore, it is shown that the robust optimization formulation is related to kernel density estimation, and based on this approach, a proof that Lasso is consistent is given, using robustness directly. Finally, a theorem is proved which states that sparsity and algorithmic stability contradict each other, and hence Lasso is not stable."}, "keywords": ["uncertainty"], "citation_intent": "background"}
{"citing_id": "2303.01561v1", "cited_id": "1703.04414", "section_title": "B. Error-Correction Performance", "citation": "The results also agree with #REFR , i.e., they indicate that DSCF\u03c9 outperforms standard SCF.", "text_before_citation": ["The decoders with and without the SRM are depicted in dashed and solid lines, respectively.", "The SCF decoder is in black, DSCF-2 is in light-blue and DSCF-3 is in light-green with unique markers.", "From Fig.", "3 , it can be seen that the SRM does not alter the error-correction performance.", "This is expected and in line with the definition of the mechanism described in Section III-A."], "text_after_citation": ["The DSCF-3 decoder offers the best performance, thus the motivation to reduce the executiontime characteristics of that algorithm. Fig.", "4 show the average execution time of SCF and DSCF-\u03c9 decoders for P(1024, 128) polar code.", "The decoders with and without the SRM are in dashed and solid lines accordingly. The latency of SC decoding is provided for reference.", "From the figures, we observe that using the SRM provides greater gain to the DSCF-2 and DSCF-3 decoders throughout the FER range.", "We explain this by higher number of additional trials performed by the multi bit-flipping decoders in average."], "citing_paper_content": {"title": "Successive-Cancellation Flip Decoding Of Polar Codes With A Simplified Restart Mechanism", "abstract": "Polar codes are a class of error-correcting codes that provably achieve the capacity of practical channels. The successive-cancellation flip (SCF) decoder is a low-complexity decoder that was proposed to improve the performance of the successive-cancellation (SC) decoder as an alternative to the highcomplexity successive-cancellation list (SCL) decoder. The SCF decoder improves the error-correction performance of the SC decoder, but the variable execution time and the high worst-case execution time pose a challenge for the realization of receivers with fixed-time algorithms. The dynamic SCF (DSCF) variation of the SCF decoder further improves the error-correction performance but the challenge of decoding delay remains. In this work, we propose a simplified restart mechanism (SRM) that reduces the execution time of SCF and DSCF decoders through conditional restart of the additional trials from the second half of the codeword. We show that the proposed mechanism is able to improve the execution time characteristics of SCF and DSCF decoders while providing identical error-correction performance. For a DSCF decoder that can flip up to 3 simultaneous bits per decoding trial, the average execution time, the average additional execution time and the execution-time variance are reduced by approximately 31%, 37% and 57%, respectively. For this setup, the mechanism requires approximately 3.9% additional memory."}, "cited_paper_content": {"title": "Dynamic-Scflip Decoding Of Polar Codes", "abstract": "This paper proposes a generalization of the recently introduced Successive Cancellation Flip (SCFlip) decoding of polar codes, characterized by a number of extra decoding attempts, where one or several positions are flipped from the standard Successive Cancellation (SC) decoding. To make such an approach effective, we first introduce the concept of higher-order bit-flips, and propose a new metric to determine the bit-flips that are more likely to correct the trajectory of the SC decoding. We then propose a generalized SCFlip decoding algorithm, referred to as Dynamic-SCFlip (D-SCFlip), which dynamically builds a list of candidate bit-flips, while guaranteeing that extra decoding attempts are performed by decreasing probability of success. Simulation results show that D-SCFlip is an effective alternative to SC-List decoding of polar codes, by providing very good error correcting performance, with an average computation complexity close to the one of the SC decoder."}, "keywords": ["DSCF\u03c9"], "citation_intent": "result"}
{"citing_id": "2304.10321v1", "cited_id": "1904.08779", "section_title": "D. Visualization", "citation": "This is consistent with specaugment #REFR , which converts over-fitting problems into under-fitting problems, and improves model performance through longer training time.", "text_before_citation": ["The attention in the ST model tends to be smooth between input frames.", "However, after applying DropDim, both types of attention become more concentrated, indicating that DropDim can not only help the model to remove interference but also make the model better align with the target.", "Training Process Visualization.", "It can be seen from the network training curve in Fig.", "6 that when DropDim is used, the network can underfit the training loss, which is in sharp contrast with the usual situation where networks tend to overfit to the training data."], "text_after_citation": ["This also reveals to us that the combination of Dropdim and specaugment is not suitable.", "Because although this combination can improve performance, the training time will be further increased.", "Dropdim is more suitable to be combined with Speed Perturbs, because the latter improves the model performance by increasing the amount of data, which is different from Dropdim and specaugment."], "citing_paper_content": {"title": "Dropdim: A Regularization Method For Transformer Networks", "abstract": "We introduce DropDim, a structured dropout method designed for regularizing the self-attention mechanism, which is a key component of the transformer. In contrast to the general dropout method, which randomly drops neurons, DropDim drops part of the embedding dimensions. In this way, the semantic information can be completely discarded. Thus, the excessive co-adapting between different embedding dimensions can be broken, and the self-attention is forced to encode meaningful features with a certain number of embedding dimensions erased. Experiments on a wide range of tasks executed on the MUST-C English-Germany dataset show that DropDim can effectively improve model performance, reduce over-fitting, and show complementary effects with other regularization methods. When combined with label smoothing, the WER can be reduced from 19.1% to 15.1% on the ASR task, and the BLEU value can be increased from 26.90 to 28.38 on the MT task. On the ST task, the model can reach a BLEU score of 22.99, an increase by 1.86 BLEU points compared to the strong baseline."}, "cited_paper_content": {"title": "Specaugment: A Simple Data Augmentation Method For Automatic Speech Recognition", "abstract": "We present SpecAugment, a simple data augmentation method for speech recognition. SpecAugment is applied directly to the feature inputs of a neural network (i.e., filter bank coefficients). The augmentation policy consists of warping the features, masking blocks of frequency channels, and masking blocks of time steps. We apply SpecAugment on Listen, Attend and Spell networks for end-to-end speech recognition tasks. We achieve state-of-the-art performance on the LibriSpeech 960h and Swichboard 300h tasks, outperforming all prior work. On LibriSpeech, we achieve 6.8% WER on test-other without the use of a language model, and 5.8% WER with shallow fusion with a language model. This compares to the previous state-of-the-art hybrid system of 7.5% WER. For Switchboard, we achieve 7.2%/14.6% on the Switchboard/CallHome portion of the Hub5'00 test set without the use of a language model, and 6.8%/14.1% with shallow fusion, which compares to the previous state-of-the-art hybrid system at 8.3%/17.3% WER."}, "keywords": ["over-fitting problems", "specaugment"], "citation_intent": "result"}
{"citing_id": "2303.06234v1", "cited_id": "1505.01462", "section_title": "Estimation Error Lower Bound", "citation": "To construct the set of models, we follow the construction similar to that in the proof of Theorem 1 of #REFR .", "text_before_citation": ["D KL (P a (X) P b (X)) \u2264 \u03be ,", "where P a (X) denotes the distribution over X under a model parametrized by \u03b2 * a . Then the minimax risk is lower bounded as", "inf \u03b2 sup \u03b2 * \u2208B E[ \u03b2 \u2212 \u03b2 * 2 2 ] \u2265 \u03b4 2 2 1 \u2212 \u03be + log 2 log M ,", "where\u03b2 is the output of any statistical estimator.", "Intuitively, the theorem says that if we can construct a set of models where every pair of model is sufficiently different (measured in 2 2 distance) yet they parametrize 'similar' distributions (measured in KL divergence), then any statistical estimator will fail to identify the correct model and thus suffer from a minimum estimation error."], "text_after_citation": ["We first restate a coding theoretic due to #OTHEREFR . #OTHEREFR", "(2015) ) For any \u03b1 \u2208 (0, 1 4 ), there exists a set of M (\u03b1) = \u0398(exp (m)) binary vectors {z 1 , . . . , z M (\u03b1)"], "citing_paper_content": {"title": "Optimal And Private Learning From Human Response Data", "abstract": "Item response theory (IRT) is the study of how people make probabilistic decisions, with diverse applications in education testing, recommendation systems, among others. The Rasch model of binary response data, one of the most fundamental models in IRT, remains an active area of research with important practical significance. Recently, Nguyen and Zhang (2022) proposed a new spectral estimation algorithm that is efficient and accurate. In this work, we extend their results in two important ways. Firstly, we obtain a refined entrywise error bound for the spectral algorithm, complementing the 'average error' 2 bound in their work. Notably, under mild sampling conditions, the spectral algorithm achieves the minimax optimal error bound (modulo a log factor). Building on the refined analysis, we also show that the spectral algorithm enjoys optimal sample complexity for top-K recovery (e.g., identifying the best K items from approval/disapproval response data), explaining the empirical findings in the previous work."}, "cited_paper_content": {"title": "Estimation From Pairwise Comparisons: Sharp Minimax Bounds With Topology Dependence", "abstract": "Data in the form of pairwise comparisons arises in many domains, including preference elicitation, sporting competitions, and peer grading among others. We consider parametric ordinal models for such pairwise comparison data involving a latent vector $w^* \\in \\mathbb{R}^d$ that represents the\"qualities\"of the $d$ items being compared; this class of models includes the two most widely used parametric models--the Bradley-Terry-Luce (BTL) and the Thurstone models. Working within a standard minimax framework, we provide tight upper and lower bounds on the optimal error in estimating the quality score vector $w^*$ under this class of models. The bounds depend on the topology of the comparison graph induced by the subset of pairs being compared via its Laplacian spectrum. Thus, in settings where the subset of pairs may be chosen, our results provide principled guidelines for making this choice. Finally, we compare these error rates to those under cardinal measurement models and show that the error rates in the ordinal and cardinal settings have identical scalings apart from constant pre-factors."}, "keywords": ["models"], "citation_intent": "method"}
{"citing_id": "2303.08577v1", "cited_id": "1912.04958", "section_title": "Stylegan2", "citation": "In particular, we are interested in the second version of the StyleGAN, the StyleGAN2 #REFR , which is a revisiting of the architecture of the StyleGAN synthesis network.", "text_before_citation": ["StyleGANs are a re-design of the GANs generator architecture, which aim is to control the image synthesis process #OTHEREFR .", "The typical architecture of a StyleGAN is shown in Appendix A.1 in Figure 6 , together with the technical details."], "text_after_citation": ["Figure 7 in Appendix A.1 exemplifies the changes made to the original architecture up to the final StyleGAN2 network. Further information is provided in Appendix A.1.", "The StyleGAN2 architecture makes it possible to control image synthesis via scale-specific style modifications.", "In particular, this approach attains layer-wise decomposition of visual properties, allowing StyleGAN to control global aspects of the picture, such as pose, lighting conditions or colour schemes, coherently over the entire image.", "However, while this model successfully disentangles global properties, it is more limited in its ability to perform spatial decomposition.", "It provides no direct means to control the style of localised regions within the generated image."], "citing_paper_content": {"title": "Investigating Gansformer: A Replication Study Of A State-Of-The-Art Image Generation Model", "abstract": "The field of image generation through generative modelling is abundantly discussed nowadays. It can be used for various applications, such as up-scaling existing images, creating non-existing objects, such as interior design scenes, products or even human faces, and achieving transferlearning processes. In this context, Generative Adversarial Networks (GANs) are a class of widely studied machine learning frameworks first appearing in the paper \"Generative adversarial nets\" by Goodfellow et al. [1] that achieve the goal above. In our work, we reproduce and evaluate a novel variation of the original GAN network, the GANformer, proposed in \"Generative Adversarial Transformers\" by Hudson and Zitnick [2]. This project aimed to recreate the methods presented in this paper to reproduce the original results and comment on the authors' claims. Due to resources and time limitations, we had to constrain the network's training times, dataset types, and sizes. Our research successfully recreated both variations of the proposed GANformer model and found differences between the authors' and our results. Moreover, discrepancies between the publication methodology and the one implemented, made available in the code, allowed us to study two undisclosed variations of the presented procedures."}, "cited_paper_content": {"title": "Analyzing And Improving The Image Quality Of Stylegan", "abstract": "The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality."}, "keywords": ["StyleGAN synthesis network"], "citation_intent": "method"}
{"citing_id": "2303.05342v1", "cited_id": "1405.0312", "section_title": "Introduction", "citation": "To acquire such knowledge, we propose to automatically construct a visual relation knowledge graph from tremendous open domain image captions #REFR .", "text_before_citation": ["Rather than representing each relationship statically with few-shot training examples #OTHEREFR or word vectors #OTHEREFR , we propose to generate the knowledge-enhanced representation of relationships dynamically with a pre-trained language model, based on the classes of Subject and Object and a simple prompt template.", "As a result, the semantic diversity of each relationship can be largely covered with the help of the embedded knowledge in the pre-trained language model. Visual Relation Knowledge.", "Though convey a large amount of useful information, knowledge from text suffer from the reporting bias problem #OTHEREFR and misses common relation facts such as letter-painted on-sign and sign-hanging from-pole.", "In particular, the wellknown text-based knowledge graph ConceptNet #OTHEREFR covers only 17% of relation facts in visual relation learning datasets #OTHEREFR .", "However, visual relation knowledge which explicitly represents the common relationships between objects can be leveraged to bridge the information of entities and relationships from the textual knowledge."], "text_after_citation": ["In detail, we adopt a rule-based parser #OTHEREFR to extract relation triplets from captions and then resort to a distributed representation of visual relation knowledge graph with a novel knowledge encoding module.", "Our contributions can be summarized as three-fold:", "\u2022 To the best of our knowledge, this is the first work to combine textual knowledge and visual relation knowledge in VRD.", "We devise a Knowledge-augmented Few-shot VRD (KFV) framework, which incorporates both kinds of knowledge to improve the generalization ability of few-shot VRD.", "\u2022 We present a novel prompt-based method to acquire textual knowledge from pre-trained language models and an efficient way to learn distributed visual relation knowledge graph representation from image captions to augment few-shot VRD."], "citing_paper_content": {"title": "Knowledge-Augmented Few-Shot Visual Relation Detection", "abstract": "Visual Relation Detection (VRD) aims to detect relationships between objects for image understanding. Most existing VRD methods rely on thousands of training samples of each relationship to achieve satisfactory performance. Some recent papers tackle this problem by few-shot learning with elaborately designed pipelines and pre-trained word vectors. However, the performance of existing few-shot VRD models is severely hampered by the poor generalization capability, as they struggle to handle the vast semantic diversity of visual relationships. Nonetheless, humans have the ability to learn new relationships with just few examples based on their knowledge. Inspired by this, we devise a knowledge-augmented, few-shot VRD framework leveraging both textual knowledge and visual relation knowledge to improve the generalization ability of few-shot VRD. The textual knowledge and visual relation knowledge are acquired from a pre-trained language model and an automatically constructed visual relation knowledge graph, respectively. We extensively validate the effectiveness of our framework. Experiments conducted on three benchmarks from the commonly used Visual Genome dataset show that our performance surpasses existing state-of-the-art models with a large improvement."}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["visual relation knowledge"], "citation_intent": "method"}
{"citing_id": "2304.04163v2", "cited_id": "1311.2012", "section_title": "0H", "citation": "According to #REFR , the likelihood function p(y p |x, \u2206\u03c6) = CN (y p ; f p x, \u03c3 2 e ), where f p is the p-th row of measurement matrix F .", "text_before_citation": ["The expectation is taken over the marginal posterior p(x n |y, \u2206\u03c6) given measurements y and off-grid offsets \u2206\u03c6.", "To exploit the inherent statistical structure of the HAP-UAV channel in terms of channel support and value vectors, the marginal posterior distribution can be expressed by Bayes' rule as", "EQUATION", "where \u221d indicates proportionality up to a constant scale factor, x \u2212m means a vector x excluding the element", "x m ."], "text_after_citation": ["The optimal off-grid offsets \u2206\u03c6 can be obtained by a maximum likelihood (ML) approach", "EQUATION", "After obtaining the estimated offsets \u2206\u03c6 * , we can then calculate the corresponding marginal posterior distribution #OTHEREFR .", "It's the production of many distribution functions, each of which is determined by only a subset of variables.", "Such decomposition structures can be calculated by a graphical model, i.e., a factor graph, which can be an undirected bipartite graph used to represent the connection between random variables (called variable nodes) and related probability density functions (pdfs) (called factor nodes)."], "citing_paper_content": {"title": "Energy-Efficient Urllc Service Provision Via A Near-Space Information Network", "abstract": "The integration of a near-space information network (NSIN) with the reconfigurable intelligent surface (RIS) is envisioned to significantly enhance the communication performance of future wireless communication systems by proactively altering wireless channels. This paper investigates the problem of deploying a RIS-integrated NSIN to provide energy-efficient, ultra-reliable and low-latency communications (URLLC) services. We mathematically formulate this problem as a resource optimization problem, aiming to maximize the effective throughput and minimize the system power consumption, subject to URLLC and physical resource constraints. The formulated problem is challenging in terms of accurate channel estimation, RIS phase alignment, theoretical analysis, and effective solution. We propose a joint resource allocation algorithm to handle these challenges. In this algorithm, we develop an accurate channel estimation approach by exploring message passing and optimize phase shifts of RIS reflecting elements to further increase the channel gain. Besides, we derive an analysis-friend expression of decoding error probability and decompose the problem into two-layered optimization problems by analyzing the monotonicity, which makes the formulated problem analytically tractable. Extensive simulations have been conducted to verify the performance of the proposed algorithm. Simulation results show that the proposed algorithm can achieve outstanding channel estimation performance and is more energy-efficient than diverse benchmark algorithms."}, "cited_paper_content": {"title": "Quasi-Static Multiple-Antenna Fading Channels At Finite Blocklength", "abstract": "This paper investigates the maximal achievable rate for a given blocklength and error probability over quasi-static multiple-input multiple-output (MIMO) fading channels, with and without channel state information (CSI) at the transmitter and/or the receiver. The principal finding is that outage capacity, despite being an asymptotic quantity, is a sharp proxy for the finite-blocklength fundamental limits of slow-fading channels. Specifically, the channel dispersion is shown to be zero regardless of whether the fading realizations are available at both transmitter and receiver, at only one of them, or at neither of them. These results follow from analytically tractable converse and achievability bounds. Numerical evaluation of these bounds verifies that zero dispersion may indeed imply fast convergence to the outage capacity as the blocklength increases. In the example of a particular $1 \\times 2$ single-input multiple-output (SIMO) Rician fading channel, the blocklength required to achieve $90\\%$ of capacity is about an order of magnitude smaller compared to the blocklength required for an AWGN channel with the same capacity. For this specific scenario, the coding/decoding schemes adopted in the LTE-Advanced standard are benchmarked against the finite-blocklength achievability and converse bounds."}, "keywords": ["likelihood function", "measurement matrix F"], "citation_intent": "background"}
{"citing_id": "2303.01925v1", "cited_id": "1909.12077", "section_title": "Experimental Setup", "citation": "We compare against two NN models: a standard (non-Hamiltonian) Neural ODE (NODE) (Chen et al., 2018) with a neural time derivative f \u03b8 (x), and a HNN, based on the Unstructured SymODEN model of #REFR with a neural Hamiltonian H \u03b8 (x).", "text_before_citation": ["Bottom row: full cycle estimates, where the baseline is able to fit the vector field, but is oblivious to the energy.", "challenging, as such good initialisations are very important, particularly for the variational parameters of the Hamiltonian GP.", "We can obtain a good initialisation by observing that the derivative function f and H are jointly a GP.", "Using the approximate numerical derivatives of the trajectory data, we can form some data for f , which can then be conditioned on, using equations (10) to (12), to obtain an estimate for the mean function of the Hamiltonian.", "We use this estimate as the initial value for the variational mean m. See appendix B.2 for details. NN baselines."], "text_after_citation": ["Both methods perform poorly when trained on raw trajectories due to vanishing/exploding gradients.", "To allow for a fair comparison, we split the trajectories into smaller sub-trajectories by running a sliding window of fixed length over the data, and form these into batches of training data. GP baselines.", "In order to evaluate the effect of our Hamiltonian prior, we compare against the non-Hamiltonian shooting GPODE baseline of Hegde et al. (2022) .", "We use the same hyperparameter settings and initialisations as in the HGP model, where applicable.", "To initialise the inducing variable distribution means, we follow the same procedure described by Hegde et al. (2022) ."], "citing_paper_content": {"title": "Learning Energy Conserving Dynamics Efficiently With Hamiltonian Gaussian Processes", "abstract": "Hamiltonian mechanics is one of the cornerstones of natural sciences. Recently there has been significant interest in learning Hamiltonian systems in a free-form way directly from trajectory data. Previous methods have tackled the problem of learning from many short, low-noise trajectories, but learning from a small number of long, noisy trajectories, whilst accounting for model uncertainty has not been addressed. In this work, we present a Gaussian process model for Hamiltonian systems with efficient decoupled parameterisation, and introduce an energy-conserving shooting method that allows robust inference from both short and long trajectories. We demonstrate the method's success in learning Hamiltonian systems in various data settings."}, "cited_paper_content": {"title": "Symplectic Ode-Net: Learning Hamiltonian Dynamics With Control", "abstract": "In this paper, we introduce Symplectic ODE-Net (SymODEN), a deep learning framework which can infer the dynamics of a physical system from observed state trajectories. To achieve better generalization with fewer training samples, SymODEN incorporates appropriate inductive bias by designing the associated computation graph in a physics-informed manner. In particular, we enforce Hamiltonian dynamics with control to learn the underlying dynamics in a transparent way which can then be leveraged to draw insight about relevant physical aspects of the system, such as mass and potential energy. In addition, we propose a parametrization which can enforce this Hamiltonian formalism even when the generalized coordinate data is embedded in a high-dimensional space or we can only access velocity data instead of generalized momentum. This framework, by offering interpretable, physically-consistent models for physical systems, opens up new possibilities for synthesizing model-based control strategies."}, "keywords": ["neural Hamiltonian H", "standard (non-Hamiltonian) Neural"], "citation_intent": "method"}
{"citing_id": "2303.13203v1", "cited_id": "1512.04150", "section_title": "Results Obtained Using Gap/Gmp Layer(S)", "citation": "Therefore, GAP is better suited than GMP to extract and retain relevant features. This finding was also observed in #REFR .", "text_before_citation": ["In our proposed learning model, except for the last regular GAP layer, three additional GAP layers were used in each sub-network.", "The place of each GAP/GMP layer governs the level of features to be considered.", "To show the contribution of each GAP/GMP layer, we performed an ablation study.", "As can be seen in Table 3 , the performance achieved by exploiting GAP layers at various levels is almost consistently better than that of GMP, whatever the combination of the levels.", "This is mostly caused by the fact that GMP only maintains one element with the max value in each feature map; as a result, it discards much texture-related information."], "text_after_citation": ["In addition, different GAP and GMP combinations were also evaluated, where GAP and GMP units were mixed. However, no performance improvement was observed.", "Figure 4 : Comparison of the attention maps generated by the final convolutional layer of each model."], "citing_paper_content": {"title": "A Confident Labelling Strategy Based On Deep Learning For Improving Early Detection Of Knee Osteoarthritis", "abstract": "Knee OsteoArthritis (KOA) is a prevalent musculoskeletal disorder that causes decreased mobility in seniors. The diagnosis provided by physicians is subjective, however, as it relies on personal experience and the semi-quantitative Kellgren-Lawrence (KL) scoring system. KOA has been successfully diagnosed by Computer-Aided Diagnostic (CAD) systems that use deep learning techniques like Convolutional Neural Networks (CNN). In this paper, we propose a novel Siamese-based network, and we introduce a new hybrid loss strategy for the early detection of KOA. The model extends the classical Siamese network by integrating a collection of Global Average Pooling (GAP) layers for feature extraction at each level. Then, to improve the classification performance, a novel training strategy that partitions each training batch into low-, medium-and high-confidence subsets, and a specific hybrid loss function are used for each new label attributed to each sample. The final loss function is then derived by combining the latter loss functions with optimized weights. Our test results demonstrate that our proposed approach significantly improves the detection performance."}, "cited_paper_content": {"title": "Learning Deep Features For Discriminative Localization", "abstract": "In this work, we revisit the global average pooling layer proposed in [13], and shed light on how it explicitly enables the convolutional neural network (CNN) to have remarkable localization ability despite being trained on imagelevel labels. While this technique was previously proposed as a means for regularizing training, we find that it actually builds a generic localizable deep representation that exposes the implicit attention of CNNs on an image. Despite the apparent simplicity of global average pooling, we are able to achieve 37.1% top-5 error for object localization on ILSVRC 2014 without training on any bounding box annotation. We demonstrate in a variety of experiments that our network is able to localize the discriminative image regions despite just being trained for solving classification task1."}, "keywords": ["relevant features"], "citation_intent": "result"}
{"citing_id": "2303.07923v1", "cited_id": "1809.05791", "section_title": "Our Problem And Results", "citation": "This is in contrast to the uncapacitated version of the problem, where one can get (1+\u01eb)-approximation in 2 O((k log k)/\u01eb 2 ) \u2022n O #REFR time as mentioned before.", "text_before_citation": ["We also complement our approximability results by hardness bounds.", "The NP-hardness of Capacitated Sum of Radii trivially follows from the NP-hardness of Sum of Radii. We strengthen this bound by showing the following result. Theorem 1.5.", "Capacitated Sum of Radii with uniform capacities cannot be solved in f (k)n o(k) time for any computable function f , unless ETH is false. Moreover, it does not admit any FPTAS, unless P=NP.", "We also show an inapproximability bound in the Euclidean case even with uniform capacities.", "Although the above bound does not eradicate the possibility of obtaining a (1+\u01eb)-approximation in the Euclidean case, it shows that to obtain such an approximation, even when k = 2, one needs n f (\u01eb) time for some non-constant function f that depends on \u01eb."], "text_after_citation": ["As by products of our techniques we have also obtained improved bi-criteria approximations for the uniform-capacitated version of the problem where we are allowed to use (1+\u01eb)U capacity. Theorem 1.7.", "There is a randomized algorithm for Capacitated Sum of Radii with uniform capacities that runs in time 2 O(k log(k/\u01eb)) \u2022 n O #OTHEREFR and returns a solution with constant probability, such that each ball in the solution uses at most (1 + \u01eb)U capacity and the cost of the solution is at most (2 + \u01eb) \u2022 OPT, where OPT is the cost of any optimal solution in which the balls use at most U capacity.", "The above theorem improves the approximation factor in Theorem 1.2. In the Euclidean case, we obtain a similar algorithm. Theorem 1.8.", "There is a randomized algorithm for the Euclidean version of Capacitated Sum of Radii with uniform capacities that runs in time 2 O((k/\u01eb 2 ) log k) \u2022dn #OTHEREFR and returns a solution with constant probability, such that each ball in the solution uses at most (1 + \u01eb)U capacity and the cost of the solution is at most (1 + \u01eb) \u2022 OPT, where OPT is the cost of any optimal solution in which the balls use at most U capacity.", "Note that, by Theorem 1.6, a result as in the above theorem is not possible if we are not allowed to violate the capacity."], "citing_paper_content": {"title": "Fpt Constant-Approximations For Capacitated Clustering To Minimize The Sum Of Cluster Radii *", "abstract": "Clustering with capacity constraints is a fundamental problem that attracted significant attention throughout the years. In this paper, we give the first FPT constant-factor approximation algorithm for the problem of clustering points in a general metric into k clusters to minimize the sum of cluster radii, subject to non-uniform hard capacity constraints (Capacitated Sum of Radii). In particular, we give a (15 + \u01eb)-approximation algorithm that runs in 2 O(k 2 log k) \u2022 n 3 time. When capacities are uniform, we obtain the following improved approximation bounds. \u2022 A (4 + \u01eb)-approximation with running time 2 O(k log(k/\u01eb)) n 3 , which significantly improves over the FPT 28-approximation of Inamdar and Varadarajan [ESA 2020]. \u2022 A (2 + \u01eb)-approximation with running time 2 O(k/\u01eb 2 \u2022log(k/\u01eb)) dn 3 and a (1+\u01eb)-approximation with running time 2 O(kd log((k/\u01eb))) n 3 in the Euclidean space. Here d is the dimension."}, "cited_paper_content": {"title": "Constant Factor Fpt Approximation For Capacitated K-Median", "abstract": "Capacitated k-median is one of the few outstanding optimization problems for which the existence of a polynomial time constant factor approximation algorithm remains an open problem. In a series of recent papers algorithms producing solutions violating either the number of facilities or the capacity by a multiplicative factor were obtained. However, to produce solutions without violations appears to be hard and potentially requires different algorithmic techniques. Notably, if parameterized by the number of facilities $k$, the problem is also $W[2]$ hard, making the existence of an exact FPT algorithm unlikely. In this work we provide an FPT-time constant factor approximation algorithm preserving both cardinality and capacity of the facilities. The algorithm runs in time $2^{\\mathcal{O}(k\\log k)}n^{\\mathcal{O}(1)}$ and achieves an approximation ratio of $7+\\varepsilon$."}, "keywords": ["2 O((k log", "uncapacitated version"], "citation_intent": "result"}
{"citing_id": "2303.08112v1", "cited_id": "1909.03368", "section_title": "The Tuned Lens", "citation": "It ensures that the probes are not incentivized to learn extra information over and above what the model has learned, which can become a problem when training probes with ground truth labels #REFR .", "text_before_citation": ["Loss function.", "We train the translators to minimize KL between the tuned lens logits and the final layer logits:", "EQUATION", "where f > (h ) refers to the rest of the transformer after layer .", "This can be viewed as a distillation loss, using the final layer distribution as a soft label #OTHEREFR ."], "text_after_citation": ["Implementation details.", "When readily available, we train translators on a slice of the validation set used during pretraining, and use a separate slice for evaluation.", "Since BLOOM and GPT-2 do not have publicly available validation sets, we use the Pile validation set (Gao et al., 2020; .", "The OPT validation set is also not publicly available, but a member of the OPT team helped us train a tuned lens on the OPT validation set.", "Documents are concatenated and split into uniform chunks of length 2048."], "citing_paper_content": {"title": "Eliciting Latent Predictions From Transformers With The Tuned Lens", "abstract": "We analyze transformers from the perspective of iterative inference, seeking to understand how model predictions are refined layer by layer. To do so, we train an affine probe for each block in a frozen pretrained model, making it possible to decode every hidden state into a distribution over the vocabulary. Our method, the tuned lens, is a refinement of the earlier \"logit lens\" technique, which yielded useful insights but is often brittle. We test our method on various autoregressive language models with up to 20B parameters, showing it to be more predictive, reliable and unbiased than the logit lens. With causal experiments, we show the tuned lens uses similar features to the model itself. We also find the trajectory of latent predictions can be used to detect malicious inputs with high accuracy. All code needed to reproduce our results can be found at https://github.com/ AlignmentResearch/tuned-lens."}, "cited_paper_content": {"title": "Designing And Interpreting Probes With Control Tasks", "abstract": "Probes, supervised models trained to predict properties (like parts-of-speech) from representations (like ELMo), have achieved high accuracy on a range of linguistic tasks. But does this mean that the representations encode linguistic structure or just that the probe has learned the linguistic task? In this paper, we propose control tasks, which associate word types with random outputs, to complement linguistic tasks. By construction, these tasks can only be learned by the probe itself. So a good probe, (one that reflects the representation), should be selective, achieving high linguistic task accuracy and low control task accuracy. The selectivity of a probe puts linguistic task accuracy in context with the probe's capacity to memorize from word types. We construct control tasks for English part-of-speech tagging and dependency edge prediction, and show that popular probes on ELMo representations are not selective. We also find that dropout, commonly used to control probe complexity, is ineffective for improving selectivity of MLPs, but that other forms of regularization are effective. Finally, we find that while probes on the first layer of ELMo yield slightly better part-of-speech tagging accuracy than the second, probes on the second layer are substantially more selective, which raises the question of which layer better represents parts-of-speech."}, "keywords": ["model", "probes"], "citation_intent": "background"}
{"citing_id": "2303.12756v1", "cited_id": "1905.02249", "section_title": "Experiments On Cifartoy Dataset", "citation": "Moreover, the improvement over the supervised learning with coarse labels (SupCE) on good split (13.98%) is much higher than the bad split (4.41%), which shows the substantial potential of MaskCon in dealing with more realistic coarse-labelled datasets #REFR .", "text_before_citation": ["Specifically, the classes 'airplane', 'truck', 'automobile', and 'ship' are designated as Class A, while 'horse', 'dog', 'bird', and 'cat' are assigned to Class B.", "Note that Class A comprises non-organic objects, while Class B comprises animals.", "Moreover, a bad split of the aforementioned dataset is defined, wherein 'airplane', 'automobile', 'bird', and 'cat' are Class A, and 'horse', 'dog', 'ship', and 'truck' are Class B.", "As shown in Tab.", "12, our method achieves the best performance in both splits."], "text_after_citation": ["In Fig.", "3 we visualize the learned features of all test samples in the good split.", "Clearly, the learned representations with supervised contrastive learning (SupCon) and self-supervised contrastive learning (SelfCon) tend to under-cluster and over-cluster the samples, respectively.", "By contrast, our method gets more compact and clear clusters, in line with the results when trained with fine labels (SupFINE)."], "citing_paper_content": {"title": "Maskcon: Masked Contrastive Learning For Coarse-Labelled Dataset", "abstract": "Deep learning has achieved great success in recent years with the aid of advanced neural network structures and large-scale human-annotated datasets. However, it is often costly and difficult to accurately and efficiently annotate large-scale datasets, especially for some specialized domains where fine-grained labels are required. In this setting, coarse labels are much easier to acquire as they do not require expert knowledge. In this work, we propose a contrastive learning method, called masked contrastive learning (MaskCon) to address the under-explored problem setting, where we learn with a coarse-labelled dataset in order to address a finer labelling problem. More specifically, within the contrastive learning framework, for each sample our method generates soft-labels with the aid of coarse labels against other samples and another augmented view of the sample in question. By contrast to self-supervised contrastive learning where only the sample's augmentations are considered hard positives, and in supervised contrastive learning where only samples with the same coarse labels are considered hard positives, we propose soft labels based on sample distances, that are masked by the coarse labels. This allows us to utilize both inter-sample relations and coarse labels. We demonstrate that our method can obtain as special cases many existing state-of-the-art works and that it provides tighter bounds on the generalization error. Experimentally, our method achieves significant improvement over the current state-of-the-art in various datasets, including CIFAR10, CIFAR100, ImageNet-1K, Standford Online Products and Stanford Cars196 datasets. Code and annotations are available at https://github.com/ MrChenFeng/MaskCon_CVPR2023."}, "cited_paper_content": {"title": "Mixmatch: A Holistic Approach To Semi-Supervised Learning", "abstract": "Semi-supervised learning has proven to be a powerful paradigm for leveraging unlabeled data to mitigate the reliance on large labeled datasets. In this work, we unify the current dominant approaches for semi-supervised learning to produce a new algorithm, MixMatch, that works by guessing low-entropy labels for data-augmented unlabeled examples and mixing labeled and unlabeled data using MixUp. We show that MixMatch obtains state-of-the-art results by a large margin across many datasets and labeled data amounts. For example, on CIFAR-10 with 250 labels, we reduce error rate by a factor of 4 (from 38% to 11%) and by a factor of 2 on STL-10. We also demonstrate how MixMatch can help achieve a dramatically better accuracy-privacy trade-off for differential privacy. Finally, we perform an ablation study to tease apart which components of MixMatch are most important for its success."}, "keywords": ["realistic coarse-labelled datasets"], "citation_intent": "result"}
{"citing_id": "2303.06163v1", "cited_id": "2003.09754", "section_title": "Evaluation Metric", "citation": "Following #REFR , we adopt the two metrics of part accuracy (Part Acc.) and shape chamfer distance (Shape CD) to evaluate the assembled shape structure.", "text_before_citation": ["Shape Evaluation Metric."], "text_after_citation": ["Part accuracy threshold \u03c4 p is chosen to be 0.1.", "EQUATION", "EQUATION", ") Joint Evaluation Metric.", "We propose two metrics for evaluating the second objective of the joint-centric part assembly task, joint accuracy (Joint Acc.) and joint chamfer distance (Joint CD). Specifically, Joint Acc."], "citing_paper_content": {"title": "Category-Level Multi-Part Multi-Joint 3D Shape Assembly", "abstract": "Shape assembly composes complex shapes geometries by arranging simple part geometries and has wide applications in autonomous robotic assembly and CAD modeling. Existing works focus on geometry reasoning and neglect the actual physical assembly process of matching and fitting joints, which are the contact surfaces connecting different parts. In this paper, we consider contacting joints for the task of multi-part assembly. A successful joint-optimized assembly needs to satisfy the bilateral objectives of shape structure and joint alignment. We propose a hierarchical graph learning approach composed of two levels of graph representation learning. The part graph takes part geometries as input to build the desired shape structure. The joint-level graph uses part joints information and focuses on matching and aligning joints. The two kinds of information are combined to achieve the bilateral objectives. Extensive experiments demonstrate that our method outperforms previous methods, achieving better shape structure and higher joint alignment accuracy."}, "cited_paper_content": {"title": "Learning 3D Part Assembly From A Single Image", "abstract": "Autonomous assembly is a crucial capability for robots in many applications. For this task, several problems such as obstacle avoidance, motion planning, and actuator control have been extensively studied in robotics. However, when it comes to task specification, the space of possibilities remains underexplored. Towards this end, we introduce a novel problem, single-image-guided 3D part assembly, along with a learningbased solution. We study this problem in the setting of furniture assembly from a given complete set of parts and a single image depicting the entire assembled object. Multiple challenges exist in this setting, including handling ambiguity among parts (e.g., slats in a chair back and leg stretchers) and 3D pose prediction for parts and part subassemblies, whether visible or occluded. We address these issues by proposing a two-module pipeline that leverages strong 2D-3D correspondences and assembly-oriented graph message-passing to infer part relationships. In experiments with a PartNet-based synthetic benchmark, we demonstrate the effectiveness of our framework as compared with three baseline approaches."}, "keywords": ["assembled shape structure"], "citation_intent": "method"}
{"citing_id": "2303.08777v1", "cited_id": "1310.5796", "section_title": "Results Of Section 4.", "citation": "This appendix is a review of VC theory, except for novel results presented in Section A.2.3 for the case of unbounded loss functions, where we obtain new bounds for relative type I estimation error by extending the results in #REFR .", "text_before_citation": ["and the result is then direct from the proof of Theorem 4.3. But (7.19) is clearly true since", "L(h M ) \u2212 L(h ) L(h M ) > L(M ) =\u21d2 L(h M ) \u2212 L(h ) > L(h M ) L(M ) \u2265 .", "refinement of the bounds.", "Hence, we present results which support those in this paper and outline the main ideas of VC theory, even though are not the tightest available bounds.", "We omit the proofs, and note that refined versions of the results presented here may be found at one or more of the references."], "text_after_citation": ["We start defining the shatter coefficient and VC dimension of a hypotheses space under loss function . Remark A.3.", "If there is no confusion about which loss function we are referring, or when it is not of importance to our argument, we omit and denote the shatter coefficient and VC dimension simply by S(H, N ) and", "d V C (H).", "We note that if the hypotheses in H are binary valued functions and is the simple loss function ((x, y), h) = 1{h(x) = y}, then H = G H, , and its N -th shatter coefficient is actually the maximum number of dichotomies that can be generated by the functions in H with N points.", "A.1. Generalized Glivenko-Cantelli Problems."], "citing_paper_content": {"title": "Distribution-Free Deviation Bounds Of Learning Via Model Selection With Cross-Validation Risk Estimation", "abstract": "Cross-validation techniques for risk estimation and model selection are widely used in statistics and machine learning. However, the understanding of the theoretical properties of learning via model selection with cross-validation risk estimation is quite low in face of its widespread use. In this context, this paper presents learning via model selection with cross-validation risk estimation as a general systematic learning framework within classical statistical learning theory and establishes distribution-free deviation bounds in terms of VC dimension, giving detailed proofs of the results and considering both bounded and unbounded loss functions. We also deduce conditions under which the deviation bounds of learning via model selection are tighter than that of learning via empirical risk minimization in the whole hypotheses space, supporting the better performance of model selection frameworks observed empirically in some instances."}, "cited_paper_content": {"title": "Relative Deviation Learning Bounds And Generalization With Unbounded Loss Functions", "abstract": "We present an extensive analysis of relative deviation bounds, including detailed proofs of two-sided inequalities and their implications. We also give detailed proofs of two-sided generalization bounds that hold in the general case of unbounded loss functions, under the assumption that a moment of the loss is bounded. We then illustrate how to apply these results in a sample application: the analysis of importance weighting."}, "keywords": ["unbounded loss functions"], "citation_intent": "background"}
{"citing_id": "2303.12304v1", "cited_id": "1504.08083", "section_title": "C. Corrective Loss", "citation": "As a result, we add Smooth L1 loss #REFR in addition to the initial IoU loss.", "text_before_citation": ["EQUATION", "where IoU denotes Intersection over Union between prediction results and ground truth. However, IoU loss is not perfect.", "The value of IoU would be zero if is no overlap between the bounding box and the ground truth.", "The gradient of the loss function is now zero, making it impossible to optimize the parameters.", "IoU loss can not satisfy our requirement for regression loss in this situation."], "text_after_citation": ["Smooth L1 loss can evaluate the regression offset of the output by computing the distance between each feature map point and four edges in the ground truth.", "As a result, we merge the Smooth L1 loss and IoU loss to produce the regression branch's loss function. The specific formula is as follows:", "EQUATION", "where d stands for the output regression offset andd represents the target offset. L denotes Smooth L1 loss.", "As we mention above, the optimization between the two branches in the past approach was independent."], "citing_paper_content": {"title": "Siamthn: Siamese Target Highlight Network For Visual Tracking", "abstract": "Siamese network based trackers develop rapidly in the field of visual object tracking in recent years. The majority of siamese network based trackers now in use treat each channel in the feature maps generated by the backbone network equally, making the similarity response map sensitive to background influence and hence challenging to focus on the target region. Additionally, there are no structural links between the classification and regression branches in these trackers, and the two branches are optimized separately during training. Therefore, there is a misalignment between the classification and regression branches, which results in less accurate tracking results. In this paper, a Target Highlight Module is proposed to help the generated similarity response maps to be more focused on the target region. To reduce the misalignment and produce more precise tracking results, we propose a corrective loss to train the model. The two branches of the model are jointly tuned with the use of corrective loss to produce more reliable prediction results. Experiments on 5 challenging benchmark datasets reveal that the method outperforms current models in terms of performance, and runs at 38 fps, proving its effectiveness and efficiency."}, "cited_paper_content": {"title": "Fast R-Cnn", "abstract": "This paper proposes a Fast Region-based Convolutional Network method (Fast R-CNN) for object detection. Fast R-CNN builds on previous work to efficiently classify object proposals using deep convolutional networks. Compared to previous work, Fast R-CNN employs several innovations to improve training and testing speed while also increasing detection accuracy. Fast R-CNN trains the very deep VGG16 network 9x faster than R-CNN, is 213x faster at test-time, and achieves a higher mAP on PASCAL VOC 2012. Compared to SPPnet, Fast R-CNN trains VGG16 3x faster, tests 10x faster, and is more accurate. Fast R-CNN is implemented in Python and C++ (using Caffe) and is available under the open-source MIT License at https://github.com/rbgirshick/fast-rcnn."}, "keywords": ["Smooth L1 loss"], "citation_intent": "method"}
{"citing_id": "2304.00160v2", "cited_id": "1903.03936", "section_title": "Defense Performance With The State-Of-The-Art Model Poisoning Attack", "citation": "To evaluate the defense performance, we evaluate the CosDefense on top of the state-of-the-art model poisoning attack methods, Inner Product Manipulation (IPM) #REFR attack.", "text_before_citation": [], "text_after_citation": ["In the following experiments, we set the attacker control 30% of the participants as malicious nodes, and the attack would launch at round 200.", "The server randomly samples 10 clients from 100 for local model training in each communication round.", "Evaluation Results: Figure 3 shows the accuracy curve over 1000 communication rounds for Krum, Clipping-Median, and CosDefense for MNIST, Fashion-MNIST, and CIFAR-10.", "The detailed numerical results and ablation study related to the non-iid degree are shown in Appendix E. We have three observations from the results.", "(1) As Krum and Clipping-Median use all layers of the updates, the results of CosDefense indicate that the similarity of the last layer between the collected model update and global model is more sensitive to attacks, which could effectively be used to filter out the malicious model updates."], "citing_paper_content": {"title": "Secure Federated Learning Against Model Poi-Soning Attacks Via Client Filtering", "abstract": "Given the distributed nature, detecting and defending against the backdoor attack under federated learning (FL) systems is challenging. In this paper, we observe that the cosine similarity of the last layer's weight between the global model and each local update could be used effectively as an indicator of malicious model updates. Therefore, we propose CosDefense, a cosine-similarity-based attacker detection algorithm. Specifically, under CosDefense, the server calculates the cosine similarity score of the last layer's weight between the global model and each client update, labels malicious clients whose score is much higher than the average, and filters them out of the model aggregation in each round. Compared to existing defense schemes, CosDefense does not require any extra information besides the received model updates to operate and is compatible with client sampling. Experiment results on three real-world datasets demonstrate that CosDefense could provide robust performance under the state-of-the-art FL poisoning attack."}, "cited_paper_content": {"title": "Fall Of Empires: Breaking Byzantine-Tolerant Sgd By Inner Product Manipulation", "abstract": "Recently, new defense techniques have been developed to tolerate Byzantine failures for distributed machine learning. The Byzantine model captures workers that behave arbitrarily, including malicious and compromised workers. In this paper, we break two prevailing Byzantine-tolerant techniques. Specifically we show robust aggregation methods for synchronous SGD -- coordinate-wise median and Krum -- can be broken using new attack strategies based on inner product manipulation. We prove our results theoretically, as well as show empirical validation."}, "keywords": ["(IPM) attack"], "citation_intent": "method"}
{"citing_id": "2303.01297v1", "cited_id": "1805.10000", "section_title": "Related Works", "citation": "The Virtual-Taobao #REFR is a multiagent reinforcement learning system designed to improve search in the social Taobao web site; it makes use of a GAN for simulating internal distributions.", "text_before_citation": ["Trust information is used in #OTHEREFR to make effective recommendations.", "They propose a GAN where the discriminator is an MLP model, and the generator is an LSTM model.", "Finally, CF datasets are usually imbalanced due to their social data collection (e.g: more young than old people).", "To address this limitation, #OTHEREFR proposes a Wasserstein GAN model in the generator, and the PacGAN concept in the discriminator #OTHEREFR , to minimize the mode collapse problem.", "A platform for multi-agent RS simulation is the probabilistic-based RecSim #OTHEREFR , that generates synthetic profiles of users and items, and it uses Markov chains and recurrent neural networks."], "text_after_citation": ["A simple matrix factorization is used #OTHEREFR to inject topic diversification in the recommendation process.", "The DataGenCars #OTHEREFR is a Java-based generator of RS synthetic data; it contains a statistical basement that provides flexibility, but it returns low accuracy compared to deep learning generative models.", "Finally, the SynEvaRec framework #OTHEREFR provides synthetic RS datasets generation using the Synthetic Data Vault (SVD) library.", "This library models multivariate distributions using copula functions; its CTGAN sub-library includes GAN models.", "The main advantage of SynEvaRec, compared to previous frameworks, is that it can take different RS as source; its main drawbacks are the poor quality of the results in most of the cases, and the excessive time consumed to accomplish the training stage."], "citing_paper_content": {"title": "Creating Synthetic Datasets For Collaborative Filtering Recommender Systems Using Generative Adversarial Networks", "abstract": "Research and education in machine learning needs diverse, representative, and open datasets that contain sufficient samples to handle the necessary training, validation, and testing tasks. Currently, the Recommender Systems area includes a large number of subfields in which accuracy and beyond accuracy quality measures are continuously improved. To feed this research variety, it is necessary and convenient to reinforce the existing datasets with synthetic ones. This paper proposes a Generative Adversarial Network (GAN)-based method to generate collaborative filtering datasets in a parameterized way, by selecting their preferred number of users, items, samples, and stochastic variability. This parameterization cannot be made using regular GANs. Our GAN model is fed with dense, short, and continuous embedding representations of items and users, instead of sparse, large, and discrete vectors, to make an accurate and quick learning, compared to the traditional approach based on large and sparse input vectors. The proposed architecture includes a DeepMF model to extract the dense user and item embeddings, as well as a clustering process to convert from the dense GAN generated samples to the discrete and sparse ones, necessary to create each required synthetic dataset. The results of three different source datasets show adequate distributions and expected quality values and evolutions on the generated datasets compared to the source ones. Synthetic datasets and source codes are available to researchers."}, "cited_paper_content": {"title": "Virtual-Taobao: Virtualizing Real-World Online Retail Environment For Reinforcement Learning", "abstract": "Applying reinforcement learning in physical-world tasks is extremely challenging. It is commonly infeasible to sample a large number of trials, as required by current reinforcement learning methods, in a physical environment. This paper reports our project on using reinforcement learning for better commodity search in Taobao, one of the largest online retail platforms and meanwhile a physical environment with a high sampling cost. Instead of training reinforcement learning in Taobao directly, we present our environment-building approach: we build Virtual-Taobao, a simulator learned from historical customer behavior data, and then we train policies in Virtual-Taobao with no physical sampling costs. To improve the simulation precision, we propose GAN-SD (GAN for Simulating Distributions) for customer feature generation with better matched distribution; we propose MAIL (Multiagent Adversarial Imitation Learning) for generating better generalizable customer actions. To further avoid overfitting the imperfection of the simulator, we propose ANC (Action Norm Constraint) strategy to regularize the policy model. In experiments, Virtual-Taobao is trained from hundreds of millions of real Taobao customers\u2019 records. Compared with the real Taobao, Virtual-Taobao faithfully recovers important properties of the real environment. We further show that the policies trained purely in Virtual-Taobao, which has zero physical sampling cost, can have significantly superior real-world performance to the traditional supervised approaches, through online A/B tests. We hope this work may shed some light on applying reinforcement learning in complex physical environments."}, "keywords": ["GAN", "Virtual-Taobao"], "citation_intent": "method"}
{"citing_id": "2303.09057v1", "cited_id": "1910.11480", "section_title": "Experimental Setup", "citation": "We adopt a Parallel WaveGAN vocoder #REFR pretrained on the VCTK dataset to convert log mel-spectrograms to waveforms.", "text_before_citation": ["We split the dataset into ratios of 60%, 20%, and 20% for train, validation, and test set, respectively, considering speakers and utterances.", "For conversion scenarios, we select 20 speakers and generate 600 pairs per Seen-to-Seen (S2S) and Unseen-to-Unseen (U2U) scenarios.", "After downsampling the audio to 16 kHz, we extract CPC, f0, and log mel-spectrogram features based on a frame size of 25ms, hop size of 10ms, and mel bins of 80.", "For training details, we use a batch size of 64, an epoch of 400, and Adam optimizer with a learning rate of 10 \u22125 .", "For model parameters, we take H = 256, C = 512, and L = 6."], "text_after_citation": ["For comparison, benchmark models are reproduced using their official codes.", "They are trained with melspectrogram features except for S2VC which uses CPC features. The conversion results are available on the demo page. #OTHEREFR"], "citing_paper_content": {"title": "Triaan-Vc: Triple Adaptive Attention Normalization For Any-To-Any Voice Conversion", "abstract": "Voice Conversion (VC) must be achieved while maintaining the content of the source speech and representing the characteristics of the target speaker. The existing methods do not simultaneously satisfy the above two aspects of VC, and their conversion outputs suffer from a trade-off problem between maintaining source contents and target characteristics. In this study, we propose Triple Adaptive Attention Normalization VC (TriAAN-VC), comprising an encoder-decoder and an attention-based adaptive normalization block, that can be applied to non-parallel any-to-any VC. The proposed adaptive normalization block extracts target speaker representations and achieves conversion while minimizing the loss of the source content with siamese loss. We evaluated TriAAN-VC on the VCTK dataset in terms of the maintenance of the source content and target speaker similarity. Experimental results for one-shot VC suggest that TriAAN-VC achieves state-of-the-art performance while mitigating the trade-off problem encountered in the existing VC methods."}, "cited_paper_content": {"title": "Parallel Wavegan: A Fast Waveform Generation Model Based On Generative Adversarial Networks With Multi-Resolution Spectrogram", "abstract": "We propose Parallel WaveGAN, a distillation-free, fast, and small-footprint waveform generation method using a generative adversarial network. In the proposed method, a non-autoregressive WaveNet is trained by jointly optimizing multi-resolution spectrogram and adversarial loss functions, which can effectively capture the time-frequency distribution of the realistic speech waveform. As our method does not require density distillation used in the conventional teacher-student framework, the entire model can be easily trained. Furthermore, our model is able to generate high-fidelity speech even with its compact architecture. In particular, the proposed Parallel WaveGAN has only 1.44 M parameters and can generate 24 kHz speech waveform 28.68 times faster than real-time on a single GPU environment. Perceptual listening test results verify that our proposed method achieves 4.16 mean opinion score within a Transformer-based text-to-speech framework, which is comparative to the best distillation-based Parallel WaveNet system."}, "keywords": ["Parallel WaveGAN vocoder"], "citation_intent": "method"}
{"citing_id": "2303.09440v1", "cited_id": "2001.11767", "section_title": "Preprocessing", "citation": "We first attempted to use the 'lungmask' neural network software package to perform the segmentation #REFR .", "text_before_citation": ["Also, the scans with a large number of slices will not be able to fit into memory for the graphics processing units (GPU) available.", "That is why the previous version of Cov3d first preprocessed the scan volumes to preset sizes.", "This usually mean a loss of detail as the volumes were converted to a lower resolution.", "This can cause a problem if the lungs only constitute a small proportion of the volume since a lot of the detail for detecting the effects of COVID-19 will be lost.", "To rectify this, first we segmented the lungs in the volume and cropped the volumes to a bounding box containing the voxels identified as part of the lungs."], "text_after_citation": ["This worked well for slices where the lungs were visible, many of it produced false positive results on slices where no lungs were present.", "Instead we used an automated procedure of 3D image morphological operations, adapted from #OTHEREFR .", "After cropping, the volumes were resized into standard sizes using scikit-image #OTHEREFR . The three standard sizes are described in table 3."], "citing_paper_content": {"title": "Enhanced Detection Of The Presence And Severity Of Covid-19 From Ct Scans Using Lung Segmentation", "abstract": "Improving automated analysis of medical imaging will provide clinicians more options in providing care for patients. The 2023 AI-enabled Medical Image Analysis Workshop and Covid-19 Diagnosis Competition (AI-MIA-COV19D) provides an opportunity to test and refine machine learning methods for detecting the presence and severity of COVID-19 in patients from CT scans. This paper presents version 2 of Cov3d, a deep learning model submitted in the 2022 competition. The model has been improved through a preprocessing step which segments the lungs in the CT scan and crops the input to this region. It results in a validation macro F1 score for predicting the presence of COVID-19 in the CT scans at 92.22% which is significantly above the baseline of 74%. It gives a macro F1 score for predicting the severity of COVID-19 on the validation set for task 2 as 67% which is above the baseline of 38%."}, "cited_paper_content": {"title": "Automatic Lung Segmentation In Routine Imaging Is A Data Diversity Problem, Not A Methodology Problem", "abstract": "Automated segmentation of anatomical structures is a crucial step in many medical image analysis tasks. For lung segmentation, a variety of approaches exist, involving sophisticated pipelines trained and validated on a range of different data sets. However, during translation to clinical routine the applicability of these approaches across diseases remains limited. Here, we show that the accuracy and reliability of lung segmentation algorithms on demanding cases primarily does not depend on methodology, but on the diversity of training data. We compare 4 generic deep learning approaches and 2 published lung segmentation algorithms on routine imaging data with more than 6 different disease patterns and 3 published data sets. We show that a basic approach - U-net - performs either better, or competitively with other approaches on both routine data and published data sets, and outperforms published approaches once trained on a diverse data set covering multiple diseases. Training data composition consistently has a bigger impact than algorithm choice on accuracy across test data sets. We carefully analyse the impact of data diversity, and the specifications of annotations on both training and validation sets to provide a reference for algorithms, training data, and annotation. Results on a seemingly well understood task of lung segmentation suggest the critical importance of training data diversity compared to model choice."}, "keywords": ["'lungmask' neural network"], "citation_intent": "method"}
{"citing_id": "2303.02741v1", "cited_id": "1604.01685", "section_title": "C. Informed Mix", "citation": "To further validate that the class-level ECS is well calibrated, we also show the reliability diagram for some of the representative classes from Cityscapes #REFR in Fig.", "text_before_citation": ["To further validate the property of ECS s c and ECS t c , we show the changes of the two ECSs during training in Fig. 4 .", "Here we smooth the value of raw ECS by the EMA.", "EQUATION", "where \u03c4 is the smoothness weight; j represents the j th iteration during training. As we can see from Fig.", "4 , ECS values of almost-all classes are monotonously increasing during training for both the source domain and target domain."], "text_after_citation": ["5 where we show the relation between the ECS values and the widely-adopted segmentation performance metric mean-Intersection-over-Union (mIoU).", "We can see that in general the mIoU is positively correlated with the value of ECS.", "2) Class Selection Strategy: We propose our Informed Mix (IMix) based on the DACS #OTHEREFR which is a variant of ClassMix #OTHEREFR , and build upon the idea of bridging images across domains.", "DACS is the first to apply the idea of ClassMix #OTHEREFR to domain adaptation.", "In DACS, x a is selected from the source domain while x b is selected from the target domain."], "citing_paper_content": {"title": "Ida: Informed Domain Adaptive Semantic Segmentation", "abstract": "Mixup-based data augmentation has been validated to be a critical stage in the self-training framework for unsupervised domain adaptive semantic segmentation (UDA-SS), which aims to transfer knowledge from a well-annotated (source) domain to an unlabeled (target) domain. Existing self-training methods usually adopt the popular region-based mixup techniques with a random sampling strategy, which unfortunately ignores the dynamic evolution of different semantics across various domains as training proceeds. To improve the UDA-SS performance, we propose an Informed Domain Adaptation (IDA) model, a self-training framework that mixes the data based on class-level segmentation performance, which aims to emphasize small-region semantics during mixup. In our IDA model, the class-level performance is tracked by an expected confidence score (ECS). We then use a dynamic schedule to determine the mixing ratio for data in different domains. Extensive experimental results reveal that our proposed method is able to outperform the state-of-the-art UDA-SS method by a margin of 1.1 mIoU in the adaptation of GTA-V to Cityscapes and of 0.9 mIoU in the adaptation of SYNTHIA to Cityscapes."}, "cited_paper_content": {"title": "The Cityscapes Dataset For Semantic Urban Scene Understanding", "abstract": "Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations, 20 000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark."}, "keywords": ["Cityscapes"], "citation_intent": "method"}
{"citing_id": "2304.03126v3", "cited_id": "1603.06393", "section_title": "Quantitative Evaluation", "citation": "The second one (B 2 ) was another sequence-tosequence model that incorporated a copy mechanism #REFR to deal with unseen queries.", "text_before_citation": ["We evaluated the performance of the decomposition network based on the corpus introduced in Section 4.2.", "In particular, we computed the rate of exactly match between the decomposition results and ground truth, and calculated the averaged SARI scores #OTHEREFR of the decomposition results.", "SARI score is commonly used in text simplification tasks, it evaluates the goodness of words that are added, deleted, and kept in the simplified sentences.", "Two baseline models introduced in #OTHEREFR were used for comparison.", "In particular, the first baseline (B 1 ) was a sequence-to-sequence neural network with a 5-layer LSTM encoder and a cross attention."], "text_after_citation": ["We trained these models based on the entire corpus and tested their performance via a validation set containing 30% of data that were randomly selected from the corpus with the data queries being replaced by similar but different questions. The evaluation results are summarized in Table 2 .", "We measured the effectiveness of the calibration network via success rate and the retain rate.", "Here, the success rate was defined as the percentage of the incorrect decomposition results S d that could be successfully amended by the calibration network via the feedback Sc among all the incorrect results.", "The retain rate was defined as the percentage of correct decomposition results that remain correct after updating the model's parameters via calibrations.", "We performed the experiment based on the above decomposition network D and calculated the success rate and the retain rate based on the dataset introduced in Section 4.3."], "citing_paper_content": {"title": "Datamator: An Intelligent Authoring Tool For Creating Datamations Via Data Query Decomposition", "abstract": "Fig. 1. A datamation generated by Datamator that visualizes the analysis pipeline regarding to the data query \"how many students were born in 2000?\" through an action-oriented unit visualization."}, "cited_paper_content": {"title": "Incorporating Copying Mechanism In Sequence-To-Sequence Learning", "abstract": "We address an important problem in sequence-to-sequence (Seq2Seq) learning referred to as copying, in which certain segments in the input sequence are selectively replicated in the output sequence. A similar phenomenon is observable in human language communication. For example, humans tend to repeat entity names or even long phrases in conversation. The challenge with regard to copying in Seq2Seq is that new machinery is needed to decide when to perform the operation. In this paper, we incorporate copying into neural network-based Seq2Seq learning and propose a new model called CopyNet with encoder-decoder structure. CopyNet can nicely integrate the regular way of word generation in the decoder with the new copying mechanism which can choose sub-sequences in the input sequence and put them at proper places in the output sequence. Our empirical study on both synthetic data sets and real world data sets demonstrates the efficacy of CopyNet. For example, CopyNet can outperform regular RNN-based model with remarkable margins on text summarization tasks."}, "keywords": ["unseen queries", "another sequence-tosequence model"], "citation_intent": "method"}
{"citing_id": "2305.00749v1", "cited_id": "1511.05208", "section_title": "Matrix And Tensor Cur Approximation Methods", "citation": "The DEIM method and leverage score sampling methods are also used in #REFR to sample columns of the unfolding matrices to approximate the factor matrices.", "text_before_citation": ["The sampling methods based on a prior probability distribution are also widely used in the literature using uniform, length-squared, or leverage-score probability distributions, see #OTHEREFR , for an overview on these sampling approaches.", "It has been demonstrated that sampling columns with leverage-score probability distribution can provide approximations with relative error accuracy, which is of more interest in practice #OTHEREFR .", "The MCUR was extended to the tensor case for different types of tensors decompositions.", "For example, the authors in #OTHEREFR proposed to sample some columns of the unfolding matrices randomly to approximate the factor matrices of the Tucker decomposition.", "Also, it is suggested to use the Cross2D method to deterministically sample the fibers instead of random sampling."], "text_after_citation": ["The MCUR is used in #OTHEREFR to compute low rank approximation of unfolding matrices to compute the TT approximation.", "The cross approximation is generalized based on the t-product in #OTHEREFR , where some horizontal and lateral slices are selected.", "Here, the slices are selected based on the tubal leverage scores.", "The uniform sampling without replacements are used in #OTHEREFR for image/video completion and compression.", "However, there are only these works on tubal CUR approximation and this problem has not been investigated extensively."], "citing_paper_content": {"title": "Robust Low-Tubal-Rank Tensor Recovery Using Discrete Empirical Interpolation Method With Optimized Slice/Feature Selection", "abstract": "In this paper, we extend the Discrete Empirical Interpolation Method (DEIM) to the third-order tensor case based on the t-product and use it to select important/significant lateral and horizontal slices/features. The proposed Tubal DEIM (TDEIM) is investigated both theoretically and numerically. The experimental results show that the TDEIM can provide more accurate approximations than the existing methods. An application of the proposed method to the supervised classification task is also presented."}, "cited_paper_content": {"title": "Hoid: Higher Order Interpolatory Decomposition For Tensors Based On Tucker Representation", "abstract": "We derive a CUR-type factorization for tensors in the Tucker format based on interpolatory decomposition, which we will denote as Higher Order Interpolatory Decomposition (HOID). Given a tensor $\\mathcal{X}$, the algorithm provides a set of column vectors $\\{ \\mathbf{C}_n\\}_{n=1}^d$ which are columns extracted from the mode-$n$ tensor unfolding, along with a core tensor $\\mathcal{G}$ and together, they satisfy some error bounds. Compared to the Higher Order SVD (HOSVD) algorithm, the HOID provides a decomposition that preserves certain important features of the original tensor such as sparsity, non-negativity, integer values, etc. Error bounds along with detailed estimates of computational costs are provided. The algorithms proposed in this paper have been validated against carefully chosen numerical examples which highlight the favorable properties of the algorithms. Related methods for subset selection proposed for matrix CUR decomposition, such as Discrete Empirical Interpolation method (DEIM) and leverage score sampling, have also been extended to tensors and are compared against our proposed algorithms."}, "keywords": ["unfolding matrices"], "citation_intent": "method"}
{"citing_id": "2304.01646v2", "cited_id": "1901.03302", "section_title": "Ii. Related Work", "citation": "In contrast to the previously mentioned study, Cruz and Abreu #REFR , and Meneses-Viveros et al.", "text_before_citation": ["Saving energy can also be considered in a broader context, such as that of green Internet of Things, as demonstrated by Arshad et al. #OTHEREFR . Arshad et al.", "outlined multiple challenges and proposed a taxonomy of green IoT techniques.", "This categorization was based on the technologies which were used in various energy-saving IoT models.", "Alongside, the authors evaluated different solutions that can be used to reduce the energy consumption of IoT devices.", "Overall, the paper provides a comprehensive overview of the different opportunities for energy savings in IoT but has no explicit implication on how to save software-based energy on mobile applications."], "text_after_citation": ["#OTHEREFR propose strategies that have direct implications for mobile application developers to utilize the system and optimize battery usage efficiently.", "Specifically, Cruz and Abreu present a comprehensive catalogue of energy patterns that can be applied to mobile applications to enhance energy efficiency.", "This catalogue, comprising 22 design patterns, was developed through a review of commits, issues, and pull requests of over 1700 mobile app repositories and related research.", "The authors assert that some of these patterns may also apply to the IoT domain.", "Meneses-Viveros et al."], "citing_paper_content": {"title": "Energy-Saving Strategies For Mobile Web Apps And Their Measurement: Results From A Decade Of Research", "abstract": "In 2022, over half of the web traffic was accessed through mobile devices. By reducing the energy consumption of mobile web apps, we can not only extend the battery life of our devices, but also make a significant contribution to energy conservation efforts. For example, if we could save only 5% of the energy used by web apps, we estimate that it would be enough to shut down one of the nuclear reactors in Fukushima. This paper presents a comprehensive overview of energy-saving experiments and related approaches for mobile web apps, relevant for researchers and practitioners. To achieve this objective, we conducted a systematic literature review and identified 44 primary studies for inclusion. Through the mapping and analysis of scientific papers, this work contributes: (1) an overview of the energy-draining aspects of mobile web apps, (2) a comprehensive description of the methodology used for the energy-saving experiments, and (3) a categorization and synthesis of various energy-saving approaches."}, "cited_paper_content": {"title": "Catalog Of Energy Patterns For Mobile Applications", "abstract": "Software engineers make use of design patterns for reasons that range from performance to code comprehensibility. Several design patterns capturing the body of knowledge of best practices have been proposed in the past, namely creational, structural and behavioral patterns. However, with the advent of mobile devices, it becomes a necessity a catalog of design patterns for energy efficiency. In this work, we inspect commits, issues and pull requests of 1027 Android and 756 iOS apps to identify common practices when improving energy efficiency. This analysis yielded a catalog, available online, with 22 design patterns related to improving the energy efficiency of mobile apps. We argue that this catalog might be of relevance to other domains such as Cyber-Physical Systems and Internet of Things. As a side contribution, an analysis of the differences between Android and iOS devices shows that the Android community is more energy-aware."}, "keywords": ["previously mentioned study", "contrast"], "citation_intent": "result"}
{"citing_id": "2303.03900v1", "cited_id": "1905.12231", "section_title": "Definition 2.10 (Slater Point). A Slater Point Of The Set", "citation": "Note that the maximum in (9) is indeed attained because problem #REFR as well as the dual of the parametric minimization problem in its objective function are both solvable.", "text_before_citation": ["The inner minimization problem in (12) has a Slater point because of Assumptions 2.9 (ii) and 2.11 (ii).", "By #OTHEREFR Theorem 2] , this minimization problem therefore admits a strong dual of the form", "EQUATION", "which is solvable.", "Substituting this dual maximization problem back into (12) finally yields #OTHEREFR ."], "text_after_citation": ["Step 2.", "Fix now any maximizer ( #OTHEREFR , which exists thanks to the results of Step 1.", "In Step 2 we will show that this maximizer can be used to construct a maximizer of the dual DRO problem (3) (if (3) is solvable) or a sequence of distributions that are feasible and asymptotically optimal in (3) (if 3is not solvable). To this end, define three disjoint index sets", "{ \u22c6 , \u22c6 , \u22c6 } , , { \u22c6 , \u22c6 } ) of problem", "\u2110 + = {\ufe01 \u2208 [ ] : \u22c6 > 0 }\ufe01 , \u2110 0 = {\ufe01 \u2208 [ ] : \u22c6 = 0, \u22c6 = 0 }\ufe01 and \u2110 \u221e = {\ufe01 \u2208 [ ] : \u22c6 = 0, \u22c6 \u0338 = 0 }\ufe01 ,"], "citing_paper_content": {"title": "New Perspectives On Regularization And Computation In Optimal Transport-Based Distributionally Robust Optimization", "abstract": "We study optimal transport-based distributionally robust optimization problems where a fictitious adversary, often envisioned as nature, can choose the distribution of the uncertain problem parameters by reshaping a prescribed reference distribution at a finite transportation cost. In this framework, we show that robustification is intimately related to various forms of variation and Lipschitz regularization even if the transportation cost function fails to be (some power of) a metric. We also derive conditions for the existence and the computability of a Nash equilibrium between the decision-maker and nature, and we demonstrate numerically that nature's Nash strategy can be viewed as a distribution that is supported on remarkably deceptive adversarial samples. Finally, we identify practically relevant classes of optimal transport-based distributionally robust optimization problems that can be addressed with efficient gradient descent algorithms even if the loss function or the transportation cost function are nonconvex (but not both at the same time)."}, "cited_paper_content": {"title": "Multivariate Distributionally Robust Convex Regression Under Absolute Error Loss", "abstract": "This paper proposes a novel non-parametric multidimensional convex regression estimator which is designed to be robust to adversarial perturbations in the empirical measure. We minimize over convex functions the maximum (over Wasserstein perturbations of the empirical measure) of the absolute regression errors. The inner maximization is solved in closed form resulting in a regularization penalty involves the norm of the gradient. We show consistency of our estimator and a rate of convergence of order $ \\widetilde{O}\\left( n^{-1/d}\\right) $, matching the bounds of alternative estimators based on square-loss minimization. Contrary to all of the existing results, our convergence rates hold without imposing compactness on the underlying domain and with no a priori bounds on the underlying convex function or its gradient norm."}, "keywords": ["parametric minimization problem", "objective function"], "citation_intent": "background"}
{"citing_id": "2305.01867v2", "cited_id": "1304.5553", "section_title": "Background", "citation": "The mesh surface contains N t triangles (typically N t \u223c 10 4 ), each described by a triplet t j = [t j,1 , t j,2 , t j,3 ] which references the three vertices #REFR ] from the collection {v n } 1\u2264n\u2264Nv .", "text_before_citation": ["PyCUDA represents a scripting-based approach to GPU run-time code generation.", "Its design philosophy and technical elements have previously been covered by its architect and collaborators in #OTHEREFR and #OTHEREFR .", "In this paper, PyCUDA discussion is centered around practicalities, and it is told from the perspective of solving a specific geometry problem, where the goal is to determine if a set of rays intersect a triangle mesh surface.", "For simplicity, we use the term 'ray' to refer to a line segment specified by a start and end point; this may be written as l i = (r start i , r end i ).", "A typical problem involves a large number of rays, say N r \u2208 [10 6 , 10 8 ], which makes parallel computation highly desirable."], "text_after_citation": ["The Moller-Trumbore algorithm #OTHEREFR provides an efficient solution to this problem.", "However, an exhaustive search [comparing N t triangles with N r rays] is very slow if acceleration structures are not used even when computation is done on a GPU (see Table 1 in #OTHEREFR ).", "v(t j ) \u2261 [v tj,1 , v tj,2 , v tj,", "In the existing CUDA implementation, Morton code is used to encode the location of all mesh triangles.", "These are subsequently reordered to preserve spatial proximity and represented by a bounding volume hierarchy (BVH) using a binary radix tree #OTHEREFR ."], "citing_paper_content": {"title": "An Experience With Pycuda: Refactoring An Existing Implementation Of A Ray-Surface Intersection Algorithm", "abstract": "This article is a sequel to \"GPU implementation of a ray-surface intersection algorithm in CUDA\" (arXiv:2209.02878) [1]. Its main focus is PyCUDA which represents a Python scripting approach to GPU run-time code generation in the Compute Unified Device Architecture (CUDA) framework. It accompanies the open-source code distributed in GitHub which provides a PyCUDA implementation of a GPU-based line-segment, surface-triangle intersection test. The objective is to share a PyCUDA learning experience with people who are new to PyCUDA. Using the existing CUDA code and foundation from [1] as the starting point, we document the key changes made to facilitate a transition to PyCUDA. As the CUDA source for the ray-surface intersection test contains both host and device code and uses multiple kernel functions, these notes offer a substantive example and real-world perspective of what it is like to utilise PyCUDA. It delves into custom data structures such as binary radix tree and highlights some possible pitfalls. The case studies present a debugging strategy which may be used to examine complex C structures (such as a triangle-mesh bounding volume hierarchy) in device memory using standard Python tools without the CUDA-GDB debugger. 1 Existing CUDA implementation: https://github.com/raymondleung8/gpu-ray-surface-intersection-in-cuda."}, "cited_paper_content": {"title": "Gpu Scripting And Code Generation With Pycuda", "abstract": "Publisher Summary A common lament in the field of scientific computing concerns the ever-widening gap between hypothetical machine capabilities and the effort an individual programmer is able to spend to move a computational application towards exploiting a machine to the full extent of its capability. More sophisticated tools, compilers, and libraries are generally hoped to level this field, enabling users to achieve good results even with modest investment. PyCUDA is a contribution to the tools for graphics processing unit (GPU) computing. PyCUDA has a two-fold aim. First, it aims to simplify the usage of existing basic concepts of CUDA C. Importantly, it does not attempt to change or reinvent the basic notions of GPU programming, but instead, as a foundation for further tools, just exposes them as is through a carefully engineered interface. Key features of this interface include generous use of sensible defaults, automatic error checking, and automatic resource management. Second, and strictly on top of the first, basic layer, PyCUDA provides abstractions that support a number of very common usage patterns involving a GPU equivalent of NumPy arrays. Although PyCUDA gives the user access to a compiled language (CUDA C), it attempts to avoid the development iteration penalty commonly associated with compiled languages, instead retaining the satisfaction and immediacy of scripting. PyCUDA and PyOpenCL, as open-source projects, thrive on user feedback, particularly feedback regarding limitations, bugs, or missing features that users encounter."}, "keywords": ["mesh surface"], "citation_intent": "background"}
{"citing_id": "2304.10398v1", "cited_id": "2002.06755", "section_title": "Multi-Label Classification On Graph-Structured Data", "citation": "GCN-LPA #REFR utilizes LPA serves as regularization to assist the GCN in learning proper edge weights that lead to improved classification performance.", "text_before_citation": ["Label propagation with GNNs.", "Prior to the advent of GNNs label propagation (LPA) algorithms constituted popular approaches for the task of node classification. Both LPA and GNNs are based on message passing.", "While GNNs propagate and transform node features, LPA propagates node label information along the edges of the graph to predict the label distribution of the unlabelled nodes.", "A few recent works (Yang et al., 2021; #OTHEREFR have explored the possibilities of combining LPA and GNNs.", "(Yang et al., 2021) employs knowledge distillation in which the trained GNN model (teacher model) is distilled into a combination of GNN and parameterized label propagation module."], "text_after_citation": ["Different from our work both of the above works implicitly assume a multi-class setting.", "Moreover, (Yang et al., 2021) focuses on the interpretable extraction of knowledge from trained GNN and can only discover the patterns learned by the teacher GNN."], "citing_paper_content": {"title": "Multi-Label Node Classification On Graph-Structured Data", "abstract": "Graph Neural Networks (GNNs) have shown state-of-the-art improvements in node classification tasks on graphs. While these improvements have been largely demonstrated in a multi-class classification scenario, a more general and realistic scenario in which each node could have multiple labels has so far received little attention. The first challenge in conducting focused studies on multi-label node classification is the limited number of publicly available multi-label graph datasets. Therefore, as our first contribution, we collect and release three real-world biological datasets and develop a multi-label graph generator to generate datasets with tunable properties. While high label similarity (high homophily) is usually attributed to the success of GNNs, we argue that a multi-label scenario does not follow the usual semantics of homophily and heterophily so far defined for a multi-class scenario. As our second contribution, besides defining homophily for the multi-label scenario, we develop a new approach that dynamically fuses the feature and label correlation information to learn label-informed representations. Finally, we perform a large-scale comparative study with 10 methods and 9 datasets which also showcase the effectiveness of our approach. We release our benchmark at https://anonymous.4open.science/r/LFLF-5D8C/."}, "cited_paper_content": {"title": "Unifying Graph Convolutional Neural Networks And Label Propagation", "abstract": "Label Propagation (LPA) and Graph Convolutional Neural Networks (GCN) are both message passing algorithms on graphs. Both solve the task of node classification but LPA propagates node label information across the edges of the graph, while GCN propagates and transforms node feature information. However, while conceptually similar, theoretical relation between LPA and GCN has not yet been investigated. Here we study the relationship between LPA and GCN in terms of two aspects: (1) feature/label smoothing where we analyze how the feature/label of one node is spread over its neighbors; And, (2) feature/label influence of how much the initial feature/label of one node influences the final feature/label of another node. Based on our theoretical analysis, we propose an end-to-end model that unifies GCN and LPA for node classification. In our unified model, edge weights are learnable, and the LPA serves as regularization to assist the GCN in learning proper edge weights that lead to improved classification performance. Our model can also be seen as learning attention weights based on node labels, which is more task-oriented than existing feature-based attention models. In a number of experiments on real-world graphs, our model shows superiority over state-of-the-art GCN-based methods in terms of node classification accuracy."}, "keywords": ["regularization", "proper edge weights"], "citation_intent": "background"}
{"citing_id": "2303.04859v1", "cited_id": "1708.05622", "section_title": "B. Learning With L 2 -Polynomial Regression", "citation": "However, we note that under special cases (e.g., d k = \u2308n r \u2309) the regression can be done in O(nd k\u03c9(r) ), where \u03c9(r) is a constant given in #REFR .", "text_before_citation": ["The factor d k is because a polynomial of degree up to k is a linear combination of monomials of the form k j=1 X ij , where", "i j \u2208 [d].", "Linear regression can be implemented via Moore-Penrose (generalized) inverse.", "The generalized inverse is computed using classical methods in O(nD 2 + D 3 ).", "Hence, given that n > D, we can perform polynomial regression in O(nD 2 ) = O(nd 2k )."], "text_after_citation": ["Hence, the computational complexity of this algorithm is O(nd \u0398(k) ) as noted in Table I ."], "citing_paper_content": {"title": "Agnostic Pac Learning Of K-Juntas Using L 2 -Polynomial Regression", "abstract": "Many conventional learning algorithms rely on loss functions other than the natural 0-1 loss for computational efficiency and theoretical tractability. Among them are approaches based on absolute loss (L 1 regression) and square loss (L 2 regression). The first is proved to be an agnostic PAC learner for various important concept classes such as juntas, and half-spaces. On the other hand, the second is preferable because of its computational efficiency, which is linear in the sample size. However, PAC learnability is still unknown as guarantees have been proved only under distributional restrictions. The question of whether L 2 regression is an agnostic PAC learner for 0-1 loss has been open since 1993 and yet has to be answered. This paper resolves this problem for the junta class on the Boolean cube-proving agnostic PAC learning of k-juntas using L 2 polynomial regression. Moreover, we present a new PAC learning algorithm based on the Boolean Fourier expansion with lower computational complexity. Fourier-based algorithms, such as Linial et al. (1993), have been used under distributional restrictions, such as uniform distribution. We show that with an appropriate change, one can apply those algorithms in agnostic settings without any distributional assumption. We prove our results by connecting the PAC learning with 0-1 loss to the minimum mean square estimation (MMSE) problem. We derive an elegant upper bound on the 0-1 loss in terms of the MMSE error. Based on that, we show that the sign of the MMSE is a PAC learner for any concept class containing it."}, "cited_paper_content": {"title": "Improved Rectangular Matrix Multiplication Using Powers Of The Coppersmith-Winograd Tensor", "abstract": "In the past few years, successive improvements of the asymptotic complexity of square matrix multiplication have been obtained by developing novel methods to analyze the powers of the Coppersmith-Winograd tensor, a basic construction introduced thirty years ago. In this paper we show how to generalize this approach to make progress on the complexity of rectangular matrix multiplication as well, by developing a framework to analyze powers of tensors in an asymmetric way. By applying this methodology to the fourth power of the Coppersmith-Winograd tensor, we succeed in improving the complexity of rectangular matrix multiplication. Let $\\alpha$ denote the maximum value such that the product of an $n\\times n^\\alpha$ matrix by an $n^\\alpha\\times n$ matrix can be computed with $O(n^{2+\\epsilon})$ arithmetic operations for any $\\epsilon>0$. By analyzing the fourth power of the Coppersmith-Winograd tensor using our methods, we obtain the new lower bound $\\alpha>0.31389$, which improves the previous lower bound $\\alpha>0.30298$ obtained five years ago by Le Gall (FOCS'12) from the analysis of the second power of the Coppersmith-Winograd tensor. More generally, we give faster algorithms computing the product of an $n\\times n^k$ matrix by an $n^k\\times n$ matrix for any value $k\\neq 1$. (In the case $k=1$, we recover the bounds recently obtained for square matrix multiplication). These improvements immediately lead to improvements in the complexity of a multitude of fundamental problems for which the bottleneck is rectangular matrix multiplication, such as computing the all-pair shortest paths in directed graphs with bounded weights."}, "keywords": ["regression", "k\u03c9(r"], "citation_intent": "background"}
{"citing_id": "2303.07991v1", "cited_id": "1810.04805", "section_title": "Introduction", "citation": "Therefore, models such as BERT #REFR commonly limit the maximum sequence length to 512 tokens.", "text_before_citation": ["Transformer-based architectures #OTHEREFR have become ubiquitous in natural language processing research.", "A key attribute to their success is the multi-head self-attention mechanism #OTHEREFR .", "However, its computational and memory requirements grow quadratically with input sequence length."], "text_after_citation": ["Longer documents are truncated #OTHEREFR or staggered position embeddings are used .", "This limitation motivated long-text transfomers, such as Big Bird #OTHEREFR and Longformer #OTHEREFR , which reduce the complexity of self-attention through the use of sparse attention and improve the performance of transformers on long documents.", "The task of rationale extraction focuses on selecting a subset of input as a justification for the model's output.", "In particular, our work focuses on extracting token-level rationales from long documents.", "The models presented are supervised using document labels only, with no token-level annotations used during training."], "citing_paper_content": {"title": "Finding The Needle In A Haystack: Unsupervised Rationale Extraction From Long Text Classifiers", "abstract": "Long-sequence transformers are designed to improve the representation of longer texts by language models and their performance on downstream document-level tasks. However, not much is understood about the quality of token-level predictions in long-form models. We investigate the performance of such architectures in the context of document classification with unsupervised rationale extraction. We find standard soft attention methods to perform significantly worse when combined with the Longformer language model. We propose a compositional soft attention architecture that applies RoBERTa sentence-wise to extract plausible rationales at the tokenlevel. We find this method to significantly outperform Longformer-driven baselines on sentiment classification datasets, while also exhibiting significantly lower runtimes."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["maximum sequence length", "BERT"], "citation_intent": "background"}
{"citing_id": "2304.12958v1", "cited_id": "1803.09956", "section_title": "A. Decomposed Reward Functions & Primitive Actions", "citation": "Further, following #REFR , each action a is parameterized as a motion primitive \u03b6 that is executed at a 3D position projected by a pixel P of an orthographic RGB-D heightmap of the scene, i.e., a = ( \u03b6, P ) .", "text_before_citation": ["Under the property-guided Q-Map learning framework, individual sub-rewards (components) r k account for properties that are semantic (high-level) and task-relevant.", "As an example, criteria for a flying agent to land on appropriate locations could consist of 1) being on flat surfaces and 2) being away from crowds.", "Each criterion can be represented as a sub-reward which serves as an indicator function I(C), denoting whether a task-specific property (C) was identified."], "text_after_citation": ["This approach enables the use of property-linked rewards rather than low-level feature rewards, such as rewarding the agent for being in close proximity of targets, where the action is the low-level visuomotor control.", "Although sub-rewards are created manually based on task criteria, we argue that the decomposition of the overall task reward could be automated via extensions to approaches like #OTHEREFR , which extracts reward channels from agent interaction data, and we defer this to future work.", "Additionally, learning property representations is similar to discovering concepts in images #OTHEREFR ."], "citing_paper_content": {"title": "A Closer Look At Reward Decomposition For High-Level Robotic Explanations", "abstract": "Explaining the behavior of intelligent agents such as robots to humans is challenging due to their incomprehensible proprioceptive states, variational intermediate goals, and resultant unpredictability. Moreover, one-step explanations for reinforcement learning agents can be ambiguous as they fail to account for the agent's future behavior at each transition, adding to the complexity of explaining robot actions. By leveraging abstracted actions that map to task-specific primitives, we avoid explanations on the movement level. Our proposed framework combines reward decomposition (RD) with abstracted action spaces into an explainable learning framework, allowing for non-ambiguous and high-level explanations based on object properties in the task. We demonstrate the effectiveness of our framework through quantitative and qualitative analysis of two robot scenarios, showcasing visual and textual explanations, from output artifacts of RD explanation, that are easy for humans to comprehend. Additionally, we demonstrate the versatility of integrating these artifacts with large language models for reasoning and interactive querying."}, "cited_paper_content": {"title": "Learning Synergies Between Pushing And Grasping With Self-Supervised Deep Reinforcement Learning", "abstract": "Skilled robotic manipulation benefits from complex synergies between non-prehensile (e.g. pushing) and prehensile (e.g. grasping) actions: pushing can help rearrange cluttered objects to make space for arms and fingers; likewise, grasping can help displace objects to make pushing movements more precise and collision-free. In this work, we demonstrate that it is possible to discover and learn these synergies from scratch through model-free deep reinforcement learning. Our method involves training two fully convolutional networks that map from visual observations to actions: one infers the utility of pushes for a dense pixel-wise sampling of end effector orientations and locations, while the other does the same for grasping. Both networks are trained jointly in a Q-learning framework and are entirely self-supervised by trial and error, where rewards are provided from successful grasps. In this way, our policy learns pushing motions that enable future grasps, while learning grasps that can leverage past pushes. During picking experiments in both simulation and real-world scenarios, we find that our system quickly learns complex behaviors amid challenging cases of clutter, and achieves better grasping success rates and picking efficiencies than baseline alternatives after only a few hours of training. We further demonstrate that our method is capable of generalizing to novel objects. Qualitative results (videos), code, pre-trained models, and simulation environments are available at http://vpg.cs.princeton.edu"}, "keywords": ["action", "pixel P"], "citation_intent": "method"}
{"citing_id": "2304.07853v1", "cited_id": "1712.02478", "section_title": "Related Work", "citation": "Stacked Conditional GAN #REFR is another technique for shadow detection that uses GAN. Nevertheless, this approach needs shadow-free photos to work.", "text_before_citation": ["For the SBU dataset, their method produced good results, although the identification process was computationally expensive during testing. .", "For shadow detection, our suggested method likewise makes use of adversarial training, but it differs significantly from scGAN in key ways.", "While our approach utilizes a generator to produce enhanced training images in the RGB space, scGAN applies the generator to generate a binary shadow mask that is conditioned on the input image.", "In contrast, scGAN employs the discriminator as a regulatory mechanism to encourage global consistency, it is more prominently used in our method to classify shadow pixels.", "Unlike scGAN, our approach achieves real-time shadow detection without requiring output averaging or postprocessing steps."], "text_after_citation": ["A different recent method suggests using contextual information to improve shadow recognition.", "By leveraging various spatial-directional RNN , our method incorporates both local and global information to enhance shadow detection performance on standard benchmarks.", "Nonetheless, to further refine the results, a post-processing step utilizing a Conditional Random Field (CRF) is required.", "In accordance with current trends in data augmentation, we provide a technique to enhance shadow detection using enhanced training samples. Zhang et al.", "#OTHEREFR proposed a simple data augmentation strategy that enhances the network's generalization and resilience to adversarial examples by augmenting the dataset with linear combinations of pairs of examples and their corresponding labels. Shrivastava et al."], "citing_paper_content": {"title": "Ai Driven Shadow Model Detection In Agropv Farms", "abstract": "Agro-photovoltaic (APV) is a burgeoning farming practice that combines agricultural activities and solar photovoltaic projects (PV) within the same area. This market is projected to experience significant growth in the next few years, with investments of approximately 9 billion US dollars expected in 2030, accounting for over 5% of the total photovoltaic investment. In certain countries, APV may even become the primary market for photovoltaic. The identification of shadows is a crucial aspect of APV research because shadows significantly impact several phenomena in the APV environment, including plant growth, microclimate, and evapotranspiration. Shadow detection can be highly beneficial, enabling a better understanding of the environment and ensuring a fair distribution of solar energy between PV and agricultural activities. In this study, we focus on using state-of-the-art CNN and GAN-based neural networks to identify shadows in agro-PV farms. CNN-based techniques can categorize pixels as either shadow or non-shadow by extracting features from input images using convolutional layers, while GANbased networks can produce synthetic shadow images and differentiate between fabricated and genuine images. Our case studies demonstrate the effectiveness of these techniques in detecting shadows in agro-PV farms, but we also acknowledge the challenges that need to be addressed, such as the impact of partial shadowing from moving objects like animals or people, as well as the need for real-time monitoring. We suggest that future research should focus on developing more sophisticated neural network-based shadow detection algorithms and integrating shadow detection with other control systems for agro-PV farms. Overall, this study emphasizes how important shadow detection is for increasing the productivity and profitability of agro-PV farms while offering environmental and social benefits that support the plants, soil, and farmers."}, "cited_paper_content": {"title": "Stacked Conditional Generative Adversarial Networks For Jointly Learning Shadow Detection And Shadow Removal", "abstract": "Understanding shadows from a single image spontaneously derives into two types of task in previous studies, containing shadow detection and shadow removal. In this paper, we present a multi-task perspective, which is not embraced by any existing work, to jointly learn both detection and removal in an end-to-end fashion that aims at enjoying the mutually improved benefits from each other. Our framework is based on a novel STacked Conditional Generative Adversarial Network (ST-CGAN), which is composed of two stacked CGANs, each with a generator and a discriminator. Specifically, a shadow image is fed into the first generator which produces a shadow detection mask. That shadow image, concatenated with its predicted mask, goes through the second generator in order to recover its shadow-free image consequently. In addition, the two corresponding discriminators are very likely to model higher level relationships and global scene characteristics for the detected shadow region and reconstruction via removing shadows, respectively. More importantly, for multi-task learning, our design of stacked paradigm provides a novel view which is notably different from the commonly used one as the multi-branch version. To fully evaluate the performance of our proposed framework, we construct the first large-scale benchmark with 1870 image triplets (shadow image, shadow mask image, and shadow-free image) under 135 scenes. Extensive experimental results consistently show the advantages of ST-CGAN over several representative state-of-the-art methods on two large-scale publicly available datasets and our newly released one."}, "keywords": ["shadow detection", "Stacked Conditional GAN"], "citation_intent": "method"}
{"citing_id": "2304.09012v1", "cited_id": "1810.04805", "section_title": "Object/Relation Predictor", "citation": "In order to capture conceptually diverse embedding and exploiting the co-occurrence among objects, predicates and parents, we follow the technique used in BERT #REFR and mask randomly words from the input that must be predicted by the object/relation predictor.", "text_before_citation": ["Instead of using directly the sequence S as input in the Object/Relation Predictor, we embed s 1:T into e P 1:T , which allows to take into consideration different features: word embedding e w 1:T that allows to identify the class of the object (e.g. \"button\" or \"container\") or the relation (e.g.", "\"inside\" or \"right), object ID embedding e o 1:T to differentiate instances of the same object, relationship ID embedding e s 1:T to separate each relationship, type of word embedding e t 1:T to identify parts of a relationship (subject = 1, predicate = 2, object = 3) for example the part of the the relationship \"button inside container\" are \"subject predicate object\" and instead of writting the plain text, we rather associate an ID to each part of the relationship to differentiate them, and parent ID embedding e p 1:T , which is a feature that allows for each of the component to know its parent.", "Those features are concatenated to form the input embedding for the object/relation predictor.", "It is given by The object/relation predictor learns to produce three different outputs: (1) the contextualized feature vectors f 1:T , (2) the size vectors s 1:T , and (3) the position vectors p 1:T .", "The contextualized feature vectors f 1:T describe objects, relations and their context with features from the input."], "text_after_citation": ["The size vectors s 1:T are predictions of the bounding boxes size for each object made by the object/relation predictor.", "It is used as indicator later in the GUI layout generator to generate final bounding boxes size.", "The position vectors p 1:T are predictions of bounding boxes position for each object.", "Finally, in order to compute the objective function to train this part of the network, we need to predict\u00ea P 1:T from the features f 1:T using a single linear layer.", "Indeed, since there is not ground truth for f 1:T , we predict\u00ea P 1:T from f 1:T and match it with e P 1:T to minimize the reconstruction error."], "citing_paper_content": {"title": "Guilget: Gui Layout Generation With Transformer", "abstract": "Sketching out Graphical User Interface (GUI) layout is part of the pipeline of designing a GUI and a crucial task for the success of a software application. Arranging all components inside a GUI layout manually is a time-consuming task. In order to assist designers, we developed a method named GUILGET to automatically generate GUI layouts from positional constraints represented as GUI arrangement graphs (GUI-AGs). The goal is to support the initial step of GUI design by producing realistic and diverse GUI layouts. The existing image layout generation techniques often cannot incorporate GUI design constraints. Thus, GUILGET needs to adapt existing techniques to generate GUI layouts that obey to constraints specific to GUI designs. GUILGET is based on transformers in order to capture the semantic in relationships between elements from GUI-AG. Moreover, the model learns constraints through the minimization of losses responsible for placing each component inside its parent layout, for not letting components overlap if they are inside the same parent, and for component alignment. Our experiments, which are conducted on the CLAY dataset, reveal that our model has the best understanding of relationships from GUI-AG and has the best performances in most of evaluation metrics. Therefore, our work contributes to improved GUI layout generation by proposing a novel method that effectively accounts for the constraints on GUI elements and paves the road for a more efficient GUI design pipeline."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["object/relation predictor", "BERT"], "citation_intent": "method"}
{"citing_id": "2305.03052v1", "cited_id": "1602.00763", "section_title": "Related Work", "citation": "As objects only need to be localized when they are visible, occlusions can be handled by simple re-association, but it has been shown that maintaining a hypothesis about the location of invisible objects can help reduce the number of identity switches #REFR .", "text_before_citation": ["Notably, most prior datasets and methods focus on localizing occluded objects with a bounding box.", "In contrast, we focus on a more precise video object segmentation setting.", "Moreover, rather than attempting to perfectly localize the invisible instance, which is not always possible in practice, we extend the setting of the problem to segmenting the occluder instead in ambiguous scenarios.", "Finally, we introduce a clear distinction between containment and occlusion at the output level.", "Object permanence methods in computer vision were mostly studied in the context of multi-object tracking -the task of localizing all the objects from a pre-defined vocabulary with bounding boxes and associating them over time based on identity #OTHEREFR ."], "text_after_citation": ["To this end, most approaches rely on a simple constant velocity heuristic #OTHEREFR , which propagates the last observed location of an object with a linear motion model.", "It is, however, only robust when the camera is static and the object velocity does not change significantly during the occlusion (e.g. because it is short).", "More complex, heuristicbased methods include #OTHEREFR , which localize invisible objects by modeling inter-occlusion relationships, and #OTHEREFR which capitalizes on the correlation between the motion of visible and invisible instances.", "More recently, several learning-based methods for localizing invisible objects have been proposed.", "#OTHEREFR takes precomputed bounded boxes for visible objects as input and passes them through a recurrent network #OTHEREFR that is trained to predict the bounding box for the occluded target."], "citing_paper_content": {"title": "Tracking Through Containers And Occluders In The Wild", "abstract": "Tracking objects with persistence in cluttered and dynamic environments remains a difficult challenge for computer vision systems. In this paper, we introduce TCOW, a new benchmark and model for visual tracking through heavy occlusion and containment. We set up a task where the goal is to, given a video sequence, segment both the projected extent of the target object, as well as the surrounding container or occluder whenever one exists. To study this task, we create a mixture of synthetic and annotated real datasets to support both supervised learning and structured evaluation of model performance under various forms of task variation, such as moving or nested containment. We evaluate two recent transformer-based video models and find that while they can be surprisingly capable of tracking targets under certain settings of task variation, there remains a considerable performance gap before we can claim a tracking model to have acquired a true notion of object permanence."}, "cited_paper_content": {"title": "Simple Online And Realtime Tracking", "abstract": "This paper explores a pragmatic approach to multiple object tracking where the main focus is to associate objects efficiently for online and realtime applications. To this end, detection quality is identified as a key factor influencing tracking performance, where changing the detector can improve tracking by up to 18.9%. Despite only using a rudimentary combination of familiar techniques such as the Kalman Filter and Hungarian algorithm for the tracking components, this approach achieves an accuracy comparable to state-of-the-art online trackers. Furthermore, due to the simplicity of our tracking method, the tracker updates at a rate of 260 Hz which is over 20x faster than other state-of-the-art trackers."}, "keywords": ["objects"], "citation_intent": "background"}
{"citing_id": "2305.01083v1", "cited_id": "1803.05652", "section_title": "Pr[Dec", "citation": "First, the constructions of #REFR utilize a public random seed for a collisionresistant hash function, so their crLDC definition is quantified over the randomness of the seed generation algorithm.", "text_before_citation": ["We require that for all adversaries PPT adversaries A there exists a negligible function \u03b5 L (\u2022) such that for all \u03bb \u2208 N and all x \u2208 \u03a3 k 1 , we have Pr[Limit(A(y), \u03c1, \u03b4, x, y, \u03bb) = 1] \u2a7d \u03b5 L (\u03bb), where the probability is taken over A and y = Enc \u03bb (x).", "If dist is the normalized Hamming distance HAM, we say the code is a Hamming crLDC; if dist is the normalized edit distance ED, we say the code is a InsDel crLDC.", "Here, \u2113 is the locality, \u03c1 is the error-rate, p is the success probability, and a function is negligible if it is o(x \u2212c ) for all constants c > 0.", "If q 2 = 2, we say that C is a family of binary crLDCs, and if q 1 = q 2 we simply write C \u03bb [K, k, q 1 ].", "Definition 1 closely follows the crLDC definition of Blocki et al. #OTHEREFR with a few modifications."], "text_after_citation": ["Our constructions do not require a public random seed so we omit this algorithm from our definition and instead quantify the security of our crLDC over a code family {C \u03bb } \u03bb\u2208N .", "This quantification also captures the notion of asymptotic security when interacting with PPT adversaries, which differs from standard (r)LDC definitions that consider information-theoretic adversaries.", "Moreover, #OTHEREFR requires the public random seed to be generated in an honest (i.e., trusted) way, and our definition and constructions circumvent this requirement.", "Second, we slightly strengthen the security definition by tweaking the predicate Fool: in Definition 1, the adversary wins if there exists an index i (not necessarily known by the adversary) such that the probability the decoder outputs correctly on input i is less than p.", "In contrast, #OTHEREFR requires the adversary to output corrupt codeword y \u2032 and a target index i such that the probability the decoder outputs correctly on index i is less than p."], "citing_paper_content": {"title": "Computationally Relaxed Locally Decodable Codes, Revisited *", "abstract": "We revisit computationally relaxed locally decodable codes (crLDCs) (Blocki et al., Trans. Inf. Theory '21) and give two new constructions. Our first construction is a Hamming crLDC that is conceptually simpler than prior constructions, leveraging digital signature schemes and an appropriately chosen Hamming code. Our second construction is an extension of our Hamming crLDC to handle insertion-deletion (InsDel) errors, yielding an InsDel crLDC. This extension crucially relies on the noisy binary search techniques of Block et al. (FSTTCS '20) to handle InsDel errors. Both crLDC constructions have binary codeword alphabets, are resilient to a constant fraction of Hamming and InsDel errors, respectively, and under suitable parameter choices have poly-logarithmic locality and encoding length linear in the message length and polynomial in the security parameter. These parameters compare favorably to prior constructions in the poly-logarithmic locality regime."}, "cited_paper_content": {"title": "Relaxed Locally Correctable Codes In Computationally Bounded Channels", "abstract": "Error-correcting codes that admit local decoding and correcting algorithms have been the focus of much recent research due to their numerous theoretical and practical applications. The goal is to obtain the best possible tradeoffs between the number of queries the algorithm makes to its oracle (the locality of the task), and the amount of redundancy in the encoding (the information rate). ::: In Hamming's classical adversarial channel model, the current tradeoffs are dramatic, allowing either small locality, but superpolynomial blocklength, or small blocklength, but high locality. However, in the computationally bounded, adversarial channel model, proposed by Lipton (STACS 1994), constructions of locally decodable codes suddenly exhibit small locality and small blocklength. The first such constructions are due to Ostrovsky, Pandey and Sahai (ICALP 2007) who build private locally decodable codes under the assumption that one-way functions exist, and in the setting where the sender and receiver share a private key. ::: We study variants of locally decodable and locally correctable codes in computationally bounded, adversarial channels, under the much weaker assumption that collision-resistant hash functions exist, and with no public-key or private-key cryptographic setup. Specifically, we provide constructions of relaxed locally correctable and relaxed locally decodable codes over the binary alphabet, with constant information rate, and poly-logarithmic locality. Our constructions compare favorably with existing schemes built under much stronger cryptographic assumptions, and with their classical analogues in the computationally unbounded, Hamming channel. ::: Our constructions crucially employ collision-resistant hash functions and local expander graphs, extending ideas from recent cryptographic constructions of memory-hard functions."}, "keywords": ["collisionresistant hash function"], "citation_intent": "background"}
{"citing_id": "2304.05648v1", "cited_id": "1806.10061", "section_title": "B. Signal Model", "citation": "Besides, the statistical channel inversion power control #REFR is implemented, i.e., the transmit power for each user is chosen such that = , where is the target received signal strength at the BS.", "text_before_citation": ["Since the number of users can be much larger than the pilot length, assigning orthogonal pilots to all the users is impractical.", "Therefore, a unique pilot sequence \u221a p is allocated to each user, where", "p ,1 , \u2022 \u2022 \u2022 , , T and", ", \u223c CN (0, 1/ ).", "Despite being non-orthogonal, the set of pilot sequences achieves asymptotic orthogonality when \u2192 \u221e #OTHEREFR ."], "text_after_citation": ["Then, the received pilot signal Y ( ) \u2208 C \u00d7 at the BS in the -th transmission block can be expressed as follows:", "EQUATION", "where A short data packet with payload bits needs to be transmitted for each active user, which is encoded as a codeword of symbols, denoted as s .", "We further assume each symbol in a codeword has unit power.", "Therefore, the received data signal at the BS in the -th transmission block Y ( ) \u2208 C \u00d7 can be expressed as follows:"], "citing_paper_content": {"title": "Grant-Free Massive Random Access With Retransmission: Receiver Optimization And Performance Analysis", "abstract": "There is an increasing demand of massive machine-type communication (mMTC) to provide scalable access for a large number of devices, which has prompted extensive investigation on grant-free massive random access (RA) in 5G and beyond wireless networks. Although many efficient signal processing algorithms have been developed, the limited radio resource for pilot transmission in grant-free massive RA systems makes accurate user activity detection and channel estimation challenging, which thereby compromises the communication reliability. In this paper, we adopt retransmission as a means to improve the quality of service (QoS) for grant-free massive RA. Specifically, by jointly leveraging the user activity correlation between adjacent transmission blocks and the historical channel estimation results, we first develop an activity-correlation-aware receiver for grant-free massive RA systems with retransmission based on the correlated approximate message passing (AMP) algorithm. Then, we analyze the performance of the proposed receiver, including the user activity detection, channel estimation, and data error, by resorting to the state evolution of the correlated AMP algorithm and the random matrix theory (RMT). Our analysis admits a tight closed-form approximation for frame error rate (FER) evaluation. Simulation results corroborate our theoretical analysis and demonstrate the effectiveness of X. Bian and J. Zhang are with the Department of Electronic and Computer Engineering,"}, "cited_paper_content": {"title": "Grant-Free Massive Mtc-Enabled Massive Mimo: A Compressive Sensing Approach", "abstract": "A key challenge of massive MTC (mMTC), is the joint detection of device activity and decoding of data. The sparse characteristics of mMTC makes compressed sensing (CS) approaches a promising solution to the device detection problem. However, utilizing CS-based approaches for device detection along with channel estimation, and using the acquired estimates for coherent data transmission is suboptimal, especially when the goal is to convey only a few bits of data. First, we focus on the coherent transmission and demonstrate that it is possible to obtain more accurate channel state information by combining conventional estimators with CS-based techniques. Moreover, we illustrate that even simple power control techniques can enhance the device detection performance in mMTC setups. Second, we devise a new non-coherent transmission scheme for mMTC and specifically for grant-free random access. We design an algorithm that jointly detects device activity along with embedded information bits. The approach leverages elements from the approximate message passing (AMP) algorithm, and exploits the structured sparsity introduced by the non-coherent transmission scheme. Our analysis reveals that the proposed approach has superior performance compared with application of the original AMP approach."}, "keywords": ["statistical channel inversion"], "citation_intent": "method"}
{"citing_id": "2303.10983v1", "cited_id": "1902.00132", "section_title": "Related Work", "citation": "Compared with QPPNet #REFR that first proposes the plan-structured network, the first and second differences as compared to TLSTM-Cost also apply.", "text_before_citation": ["Also, in terms of performance, FasCo demands much fewer training samples than TLSTMCost, which reduces the training cost significantly.", "In addition, FasCo does not need to learn the representation of predicates, which means it can be applied to any type of SQL query.", "In contrast, TLSTMCost cannot be applied on complex SQL, such as some queries in TPC-H.", "Recently there are some techniques that aim to improve the representations of predicates #OTHEREFR .", "However, these embedding models are still plagued by complex predicates, such as the \"Like\" predicate, and they cannot effectively reduce the training cost on string predicates."], "text_after_citation": ["Our FasCo inherits the basic tree-based framework but our techniques including featurization and loss function along the training and inference process are surprisingly effective.", "Also, QPPNet learns a model for each operator, which can lead to a long training time since the independent model cannot share information.", "In contrast, FasCo learns all the operations in one model.", "In addition, FasCo adopts the cardinality calibration which significantly improves the accuracy while QPPNet uses the original plan cardinalities.", "There are other problem settings around the cost model, such as the end-to-end optimizer."], "citing_paper_content": {"title": "Less Is More: Towards Lightweight Cost Estimator For Database Systems [Data Management For Data Science", "abstract": "We present FasCo, a simple yet effective learning-based estimator for the cost of executing a database query plan. FasCo uses significantly shorter training time and a lower inference cost than the state-of-the-art approaches, while achieving higher estimation accuracy. The effectiveness of FasCo comes from embedding abundant explicit execution-plan-based features and incorporating a novel technique called cardinality calibration. Extensive experimental results show that FasCo achieves orders of magnitude higher efficiency than the state-of-the-art methods: on the JOB-M benchmark dataset, it cuts off training plans by 98%, reducing training time from more than two days to about eight minutes while entailing better accuracy. Furthermore, in dynamic environments, FasCo can maintain satisfactory accuracy even without retraining, narrowing the gap between learning-based estimators and real systems. CCS CONCEPTS \u2022 Information systems \u2192 Data management systems; \u2022 Computing methodologies \u2192 Machine learning."}, "cited_paper_content": {"title": "Plan-Structured Deep Neural Network Models For Query Performance Prediction", "abstract": "Query performance prediction, the task of predicting the latency of a query, is one of the most challenging problem in database management systems. Existing approaches rely on features and performance models engineered by human experts, but often fail to capture the complex interactions between query operators and input relations, and generally do not adapt naturally to workload characteristics and patterns in query execution plans. In this paper, we argue that deep learning can be applied to the query performance prediction problem, and we introduce a novel neural network architecture for the task: a plan-structured neural network. Our approach eliminates the need for human-crafted feature selection and automatically discovers complex performance models both at the operator and query plan level. Our novel neural network architecture can match the structure of any optimizer-selected query execution plan and predict its latency with high accuracy. We also propose a number of optimizations that reduce training overhead without sacrificing effectiveness. We evaluated our techniques on various workloads and we demonstrate that our plan-structured neural network can outperform the state-of-the-art in query performance prediction."}, "keywords": ["plan-structured network"], "citation_intent": "result"}
{"citing_id": "2303.10310v1", "cited_id": "1801.01973", "section_title": "Related Work", "citation": "The IS does not capture intra-class diversity and lacks correlation with the human judgment of image quality and may produce inconsistent results when compared to human evaluations #REFR .", "text_before_citation": ["Multiple efforts were spent to improve and build upon the pix2pix model and overcome its limitations. #OTHEREFR The IS in Eq.", "(1) is a measure of the quality and diversity of generated images, based on the KL divergence between the predicted and true class distributions of a pre-trained Inception network #OTHEREFR .", "where E x\u223cpgen is the expectation over the images sampled from the generator, KL refers to the Kullback-Leibler divergence, p(y) is the marginal distribution of class labels, and (p(y | x) represents the conditional distribution of class labels y given an input image x.", "EQUATION", "Although used commonly, The IS has some limitations."], "text_after_citation": ["Moreover, it is insensitive to the prior distribution over labels (hence is biased towards ImageNet dataset and Inception model), and is very sensitive to model parameters and implementations #OTHEREFR ).", "The IS also requires a large sample size to be reliable.", "Similar to IS, the FID depends on the Inception model to generate its value, the difference, however, is that the FID calculates the Wasserstein-2 (a.k.a Fr\u00e9chet) distance between multivariate Gaussians fitted to the embedding space of the Inception-v3 network of generated and real images #OTHEREFR . The FID is Denoted as:", "EQUATION", "In this equation, \u00b5 g and \u00b5 r represent the means of the feature maps of the generated images and real images, respectively, and C g and C r represent the covariance matrices of the feature maps of the generated images and real images, respectively."], "citing_paper_content": {"title": "Pseudo Supervised Metrics: Evaluating Unsupervised Image To Image Translation Models In Unsupervised Cross-Domain Classification Frameworks", "abstract": "Highlights \u2022 New metric for evaluating UI2I translation models using pseudo supervised metrics that is designed specifically to support cross-domain classification. \u2022 The metric not only outperforms unsupervised metrics such as the FID, but is also highly correlated with the true supervised metrics, robust, and explainable. \u2022 We demonstrate that this metric can be used for future research in this field by applying it to a critical real-world problem (the boiling crisis problem)."}, "cited_paper_content": {"title": "A Note On The Inception Score", "abstract": "Deep generative models are powerful tools that have produced impressive results in recent years. These advances have been for the most part empirically driven, making it essential that we use high quality evaluation metrics. In this paper, we provide new insights into the Inception Score, a recently proposed and widely used evaluation metric for generative models, and demonstrate that it fails to provide useful guidance when comparing models. We discuss both suboptimalities of the metric itself and issues with its application. Finally, we call for researchers to be more systematic and careful when evaluating and comparing generative models, as the advancement of the field depends upon it."}, "keywords": ["intra-class diversity", "human evaluations"], "citation_intent": "result"}
{"citing_id": "2304.14329v1", "cited_id": "1301.7375", "section_title": "Setup", "citation": "In classical transduction #REFR , given data points {x i } l i=1 and their labels {y i } l i=1 , the objective is making predictions for points {x i } l+k i=l+1 .", "text_before_citation": ["I.e., when there exists some X \u2282 X such that P x\u223cDtest [x \u2208 X ] > 0, but P x\u223cDtrain [x \u2208 X ] = 0 (see Fig 2b) .", "We term the problem of achieving low risk on such a D test as OOS extrapolation.", "Out-of-Combination (OOC). This is a special case of OOS.", "Let X = X 1 \u00d7 X 2 be the product of two spaces.", "Let D train,X1 , D train,X2 denote the marginal distributions of Transduction."], "text_after_citation": ["In this paper we refer to predictors of x that are functions of the labeled data points {x i } l i=1 as transductive predictors.", "x 1 \u2208 X 1 , x 2 \u2208 X 2 under D train ,"], "citing_paper_content": {"title": "", "abstract": "Machine learning systems, especially with overparameterized deep neural networks, can generalize to novel test instances drawn from the same distribution as the training data. However, they fare poorly when evaluated on out-of-support test points. In this work, we tackle the problem of developing machine learning systems that retain the power of overparameterized function approximators while enabling extrapolation to out-of-support test points when possible. This is accomplished by noting that under certain conditions, a \"transductive\" reparameterization can convert an out-of-support extrapolation problem into a problem of within-support combinatorial generalization. We propose a simple strategy based on bilinear embeddings to enable this type of combinatorial generalization, thereby addressing the out-of-support extrapolation problem under certain conditions. We instantiate a simple, practical algorithm applicable to various supervised learning and imitation learning tasks."}, "cited_paper_content": {"title": "Learning By Transduction", "abstract": "We describe a method for predicting a classification of an object given classifications of the objects in the training set, assuming that the pairs object/classification are generated by an i.i.d. process from a continuous probability distribution. Our method is a modification of Vapnik's support-vector machine; its main novelty is that it gives not only the prediction itself but also a practicable measure of the evidence found in support of that prediction. We also describe a procedure for assigning degrees of confidence to predictions made by the support vector machine. Some experimental results are presented, and possible extensions of the algorithms are discussed."}, "keywords": ["predictions"], "citation_intent": "background"}
{"citing_id": "2303.11150v1", "cited_id": "cond-mat/9910332", "section_title": "C.2 Pull Protocol", "citation": "We show below that the similar values of p k lead to the same log 2 n \u00b1 O #REFR time it takes to inform a constant fraction of the nodes.", "text_before_citation": ["Interestingly, the expansion phase of the pull protocol (when viewed from a distance) resembles the expansion phase of the push protocol-the probability that an uninformed node becomes informed in a round starting with k informed nodes is p k = k n and thus, for small k, very close to the k n \u2212 \u0398( k 2 n 2 ) probability of the push protocol. Nevertheless, the precise processes are very different.", "For example, in the push protocol we almost surely observe a perfect doubling of the number of informed nodes as long as o( \u221a n) nodes are informed.", "For the pull protocol, the number of newly informed nodes in the first round is binomially distributed with parameters n \u2212 1 and 1 n , so the probability for a perfect doubling is asymptotically equal to 1 e .", "For this reason, the existing analyses of the push protocol cannot easily be transferred to the pull protocol.", "This is different for our method, which ignored many details of the process and only relies on the rough characteristics p k and c k of the process."], "text_after_citation": ["From that point on, the double exponential shrinking conditions are obvious, leading to a double logarithmic remaining time."], "citing_paper_content": {"title": "Randomized Rumor Spreading Revisited", "abstract": "We develop a simple and generic method to analyze randomized rumor spreading processes in fully connected networks. In contrast to all previous works, which heavily exploit the precise definition of the process under investigation, we only need to understand the probability and the covariance of the events that uninformed nodes become informed. This universality allows us to easily analyze the classic push, pull, and push-pull protocols both in their pure version and in several variations such as messages failing with constant probability or nodes calling a random number of others each round. Some dynamic models can be analyzed as well, e.g., when the network is a G(n, p) random graph sampled independently each round [Clementi et al. (ESA 2013)]. Despite this generality, our method determines the expected rumor spreading time precisely apart from additive constants, which is more precise than almost all previous works. We also prove tail bounds showing that a deviation from the expectation by more than an additive number of r rounds occurs with probability at most exp(\u2212\u2126(r)). We further use our method to discuss the common assumption that nodes can answer any number of incoming calls. We observe that the restriction that only one call can be answered leads to a significant increase of the runtime of the push-pull protocol. In particular, the double logarithmic end phase of the process now takes logarithmic time. This also increases the message complexity from the asymptotically optimal \u0398(n log log n) [Karp, Shenker, Schindelhauer, V\u00f6cking (FOCS 2000)] to \u0398(n log n). We propose a simple variation of the push-pull protocol that reverts back to the double logarithmic end phase and thus to the \u0398(n log log n) message complexity."}, "cited_paper_content": {"title": "Emergence Of Scaling In Random Networks", "abstract": "Systems as diverse as genetic networks or the World Wide Web are best described as networks with complex topology. A common property of many large networks is that the vertex connectivities follow a scale-free power-law distribution. This feature was found to be a consequence of two generic mechanisms: (i) networks expand continuously by the addition of new vertices, and (ii) new vertices attach preferentially to sites that are already well connected. A model based on these two ingredients reproduces the observed stationary scale-free distributions, which indicates that the development of large networks is governed by robust self-organizing phenomena that go beyond the particulars of the individual systems."}, "keywords": ["nodes"], "citation_intent": "background"}
{"citing_id": "2304.04029v1", "cited_id": "1804.06876", "section_title": "Introduction", "citation": "It is heavily one-sided #REFR , as will be observed with the results in this work.", "text_before_citation": ["2 In this work, we define social bias as the unbalanced disposition (or prejudice) in favor of or against a thing, person or group, relative to another, in a way that is deemed as unfair #OTHEREFR .", "3 This is harmful bias and it is related to fairness.", "In some quarters, bias also involves overgeneralization #OTHEREFR , fulfilling characteristic 2 in the next paragraph.", "As a motivation, we address the challenge of estimating bias in text data from some of the many axes (or dimensions) of bias (e.g. race and gender).", "Social bias in text usually has some of the following characteristics: 3 1."], "text_after_citation": ["2.", "It uses extreme or inappropriate language #OTHEREFR .", "This forms the basis of the assumption (for some of the samples) in the two datasets used to create the new multi-axes bias dataset (MAB), as discussed in Section 3.", "3.", "It is based on unsupported or unsubstantiated claims, such as stereotypes #OTHEREFR ."], "citing_paper_content": {"title": "Bipol: A Novel Multi-Axes Bias Evaluation Metric With Explainability For Nlp", "abstract": "We introduce bipol, a new metric with explainability, for estimating social bias in text data. Harmful bias is prevalent in many online sources of data that are used for training machine learning (ML) models. In a step to address this challenge we create a novel metric that involves a two-step process: corpuslevel evaluation based on model classification and sentence-level evaluation based on (sensitive) term frequency (TF). After creating new models to detect bias along multiple axes using SotA architectures, we evaluate two popular NLP datasets (COPA and SQuADv2). As additional contribution, we created a large dataset (with almost 2 million labelled samples) for training models in bias detection and make it publicly available. We also make public our codes. 1"}, "cited_paper_content": {"title": "Gender Bias In Coreference Resolution: Evaluation And Debiasing Methods", "abstract": "We introduce a new benchmark, WinoBias, for coreference resolution focused on gender bias. Our corpus contains Winograd-schema style sentences with entities corresponding to people referred by their occupation (e.g. the nurse, the doctor, the carpenter). We demonstrate that a rule-based, a feature-rich, and a neural coreference system all link gendered pronouns to pro-stereotypical entities with higher accuracy than anti-stereotypical entities, by an average difference of 21.1 in F1 score. Finally, we demonstrate a data-augmentation approach that, in combination with existing word-embedding debiasing techniques, removes the bias demonstrated by these systems in WinoBias without significantly affecting their performance on existing coreference benchmark datasets. Our dataset and code are available at http://winobias.org."}, "keywords": ["results", "work"], "citation_intent": "result"}
{"citing_id": "2304.07728v1", "cited_id": "1706.03762", "section_title": "I. Introduction", "citation": "For heterogeneous data fusion between LiDAR and IMU, we introduce the Transformer #REFR encoder architecture.", "text_before_citation": ["Although learning-based frameworks offer several advantages such as non-calibration and initialization, robust feature extraction, and reduced time consumption, How should we perform fusion among different modalities in a supervised sensor fusion odometry estimation task? is still one of the key issues remains, which we introduce and comprehensively evaluate in this study.", "In this work, we focus on the fusion module in the learningbased multi-modal fusion field, which is also a general component of tasks that uses multiple modalities as input.", "In contrast to operate simple element-wise summation and concatenation, the proposed fusion strategy generates the adaptive weights or attentions between different modalities.", "The overview of the proposed TransFusionOdom is shown in Fig. 1 .", "To achieve homogeneous data fusion between LiDAR's vertex and normal estimation, we implement the Soft Mask Attentional Fusion (SMAF) #OTHEREFR ."], "text_after_citation": ["This multi-attention fusion approach is designed to avoid the overfitting problem, which is prone to occur if we overly increase the complexity of Transformer-based fusion network #OTHEREFR .", "The main contributions of this article can be summarized as follows:", "1) An end-to-end Transformer-based supervised multimodal fusion odometry estimation network (TransFu-sionOdom) is developed, which achieves the competitive or even superior performance on KITTI dataset.", "The multi-attention fusion module is proposed to achieve the adaptive weights learning for the fusion between homogeneous and heterogeneous modalities.", "2) A general visualization approach is demonstrated the interactions between two modalities inside the Transformer architecture, which enhances the interpretability of the proposed learning-based solution."], "citing_paper_content": {"title": "Transfusionodom: Interpretable Transformer-Based Lidar-Inertial Fusion Odometry Estimation", "abstract": "Multi-modal fusion of sensors is a commonly used approach to enhance the performance of odometry estimation, which is also a fundamental module for mobile robots. However, the question of how to perform fusion among different modalities in a supervised sensor fusion odometry estimation task? is still one of challenging issues remains. Some simple operations, such as element-wise summation and concatenation, are not capable of assigning adaptive attentional weights to incorporate different modalities efficiently, which make it difficult to achieve competitive odometry results. Recently, the Transformer architecture has shown potential for multi-modal fusion tasks, particularly in the domains of vision with language. In this work, we propose an end-to-end supervised Transformer-based LiDAR-Inertial fusion framework (namely TransFusionOdom) for odometry estimation. The multi-attention fusion module demonstrates different fusion approaches for homogeneous and heterogeneous modalities to address the overfitting problem that can arise from blindly increasing the complexity of the model. Additionally, to interpret the learning process of the Transformer-based multi-modal interactions, a general visualization approach is introduced to illustrate the interactions between modalities. Moreover, exhaustive ablation studies evaluate different multi-modal fusion strategies to verify the performance of the proposed fusion strategy. A synthetic multi-modal dataset is made public to validate the generalization ability of the proposed fusion strategy, which also works for other combinations of different modalities. The quantitative and qualitative odometry evaluations on the KITTI dataset verify the proposed TransFusionOdom could achieve superior performance compared with other related works."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["LiDAR", "Transformer encoder architecture"], "citation_intent": "method"}
{"citing_id": "2303.01475v1", "cited_id": "1905.13277", "section_title": "Using Mixup Only In The Early Stage Of Training", "citation": "We note that our results here can be regarded as a complement to #REFR , where the authors show that regularization techniques only matter during the early phase of learning.", "text_before_citation": ["Such a conclusion implies that turning of Mixup after a certain number of epochs and returning to standard ERM training may prevent the training from overfitting the noises induced by Mixup.", "We now present results obtained from such a training scheme on both CIFAR10 and SVHN in Figure 6 .", "Results in Figure 6 clearly indicate that switching from Mixup to ERM at an appropriate time will successfully avoid the generalization degradation.", "Figure 6 also suggests that switching Mixup to ERM too early may not boost the model performance.", "In addition, if the switch is too late, memorization of noisy data may already taken effect, which impact generalization negatively."], "text_after_citation": [], "citing_paper_content": {"title": "Over-Training With Mixup May Hurt Generalization", "abstract": "Mixup, which creates synthetic training instances by linearly interpolating random sample pairs, is a simple and yet effective regularization technique to boost the performance of deep models trained with SGD. In this work, we report a previously unobserved phenomenon in Mixup training: on a number of standard datasets, the performance of Mixup-trained models starts to decay after training for a large number of epochs, giving rise to a U-shaped generalization curve. This behavior is further aggravated when the size of original dataset is reduced. To help understand such a behavior of Mixup, we show theoretically that Mixup training may introduce undesired data-dependent label noises to the synthesized data. Via analyzing a least-square regression problem with a random feature model, we explain why noisy labels may cause the U-shaped curve to occur: Mixup improves generalization through fitting the clean patterns at the early training stage, but as training progresses, Mixup becomes over-fitting to the noise in the synthetic data. Extensive experiments are performed on a variety of benchmark datasets, validating this explanation."}, "cited_paper_content": {"title": "Time Matters In Regularizing Deep Networks: Weight Decay And Data Augmentation Affect Early Learning Dynamics, Matter Little Near Convergence", "abstract": "Regularization is typically understood as improving generalization by altering the landscape of local extrema to which the model eventually converges. Deep neural networks (DNNs), however, challenge this view: We show that removing regularization after an initial transient period has little effect on generalization, even if the final loss landscape is the same as if there had been no regularization. In some cases, generalization even improves after interrupting regularization. Conversely, if regularization is applied only after the initial transient, it has no effect on the final solution, whose generalization gap is as bad as if regularization never happened. This suggests that what matters for training deep networks is not just whether or how, but when to regularize. The phenomena we observe are manifest in different datasets (CIFAR-10, CIFAR-100), different architectures (ResNet-18, All-CNN), different regularization methods (weight decay, data augmentation), different learning rate schedules (exponential, piece-wise constant). They collectively suggest that there is a ``critical period'' for regularizing deep networks that is decisive of the final performance. More analysis should, therefore, focus on the transient rather than asymptotic behavior of learning."}, "keywords": ["regularization techniques"], "citation_intent": "result"}
{"citing_id": "2303.06870v1", "cited_id": "1903.05134", "section_title": "C. More Analysis", "citation": "Lemma C.1 s = 3 is the theoretical minimum number of samples for US-Net #REFR .", "text_before_citation": [], "text_after_citation": ["Proof.", "First, from #OTHEREFR we know sandwich rules: Performances at all widths are bounded by the performance of the model at the smallest and the largest width.", "In other words, optimizing the lower and upper bounds of performance can implicitly optimize all sub-networks in a US-Net.", "To optimize for arbitrary widths, we need at least one randomly sampled width per iteration, except for the largest and smallest sub-networks.", "In conclusion, s = 3 is the theoretical minimum number of samples for US-Net. \u25a1"], "citing_paper_content": {"title": "Three Guidelines You Should Know For Universally Slimmable Self-Supervised Learning", "abstract": "We propose universally slimmable self-supervised learning (dubbed as US3L) to achieve better accuracy-efficiency trade-offs for deploying self-supervised models across different devices. We observe that direct adaptation of selfsupervised learning (SSL) to universally slimmable networks misbehaves as the training process frequently collapses. We then discover that temporal consistent guidance is the key to the success of SSL for universally slimmable networks, and we propose three guidelines for the loss design to ensure this temporal consistency from a unified gradient perspective. Moreover, we propose dynamic sampling and group regularization strategies to simultaneously improve training efficiency and accuracy. Our US3L method has been empirically validated on both convolutional neural networks and vision transformers. With only once training and one copy of weights, our method outperforms various state-ofthe-art methods (individually trained or not) on benchmarks including recognition, object detection and instance segmentation. Our code is available at https://github.com/ megvii-research/US3L-CVPR2023."}, "cited_paper_content": {"title": "Universally Slimmable Networks And Improved Training Techniques", "abstract": "Slimmable networks are a family of neural networks that can instantly adjust the runtime width. The width can be chosen from a predefined widths set to adaptively optimize accuracy-efficiency trade-offs at runtime. In this work, we propose a systematic approach to train universally slimmable networks (US-Nets), extending slimmable networks to execute at arbitrary width, and generalizing to networks both with and without batch normalization layers. We further propose two improved training techniques for US-Nets, named the sandwich rule and inplace distillation, to enhance training process and boost testing accuracy. We show improved performance of universally slimmable MobileNet v1 and MobileNet v2 on ImageNet classification task, compared with individually trained ones and 4-switch slimmable network baselines. We also evaluate the proposed US-Nets and improved training techniques on tasks of image super-resolution and deep reinforcement learning. Extensive ablation experiments on these representative tasks demonstrate the effectiveness of our proposed methods. Our discovery opens up the possibility to directly evaluate FLOPs-Accuracy spectrum of network architectures. Code and models are available at: this https URL"}, "keywords": ["US-Net"], "citation_intent": "background"}
{"citing_id": "2303.15042v1", "cited_id": "1811.06713", "section_title": "Parameter Optimization", "citation": "To estimate the unknown parameters in the testing phase, we follow the MCEM optimization scheme by #REFR .", "text_before_citation": [], "text_after_citation": ["At the Expectation step, the complete-data log-likelihood is approximated by averaging over R samples:", "EQUATION", "R samples of the latent variable are drawn using the Metropolis-Hastings algorithm with a Gaussian as a symmetric proposal distribution. \u03b8 * is an initialization of the parameters.", "At the Maximization step, we minimize the loss function, i.e., the negative log-likelihood \u2212RQ(\u03b8; \u03b8 * ), with respect to the unknown parameters \u03b8 ={HE,WB,HB,R S,f ,R B,f ,gt} using the auxiliary function technique.", "For this, equation #OTHEREFR can be viewed as the superposition of a convex function (the first term) and a concave function (the second term), where the former can be bounded using the Jensen's trace inequality and the latter can be bounded using a first-order Taylor expansion #OTHEREFR Appendix A] ."], "citing_paper_content": {"title": "Partially Adaptive Multichannel Joint Reduction Of Ego-Noise And Environmental Noise", "abstract": "Human-robot interaction relies on a noise-robust audio processing module capable of estimating target speech from audio recordings impacted by environmental noise, as well as self-induced noise, so-called ego-noise. While external ambient noise sources vary from environment to environment, ego-noise is mainly caused by the internal motors and joints of a robot. Egonoise and environmental noise reduction are often decoupled, i.e., ego-noise reduction is performed without considering environmental noise. Recently, a variational autoencoder (VAE)-based speech model has been combined with a fully adaptive non-negative matrix factorization (NMF) noise model to recover clean speech under different environmental noise disturbances. However, its enhancement performance is limited in adverse acoustic scenarios involving, e.g. ego-noise. In this paper, we propose a multichannel partially adaptive scheme to jointly model ego-noise and environmental noise utilizing the VAE-NMF framework, where we take advantage of spatially and spectrally structured characteristics of ego-noise by pre-training the ego-noise model, while retaining the ability to adapt to unknown environmental noise. Experimental results show that our proposed approach outperforms the methods based on a completely fixed scheme and a fully adaptive scheme when ego-noise and environmental noise are present simultaneously."}, "cited_paper_content": {"title": "Semi-Supervised Multichannel Speech Enhancement With Variational Autoencoders And Non-Negative Matrix Factorization", "abstract": "In this paper we address speaker-independent multichannel speech enhancement in unknown noisy environments. Our work is based on a well-established multichannel local Gaussian modeling framework. We propose to use a neural network for modeling the speech spectro-temporal content. The parameters of this supervised model are learned using the framework of variational autoencoders. The noisy recording environment is supposed to be unknown, so the noise spectro-temporal modeling remains unsupervised and is based on non-negative matrix factorization (NMF). We develop a Monte Carlo expectation-maximization algorithm and we experimentally show that the proposed approach outperforms its NMF-based counterpart, where speech is modeled using supervised NMF."}, "keywords": ["MCEM optimization scheme"], "citation_intent": "method"}
{"citing_id": "2303.17987v1", "cited_id": "1907.09693", "section_title": "Big Data Of Fl4M", "citation": "FL, a new generation of technology, is applied in finance, healthcare, and security to address isolated data islands and privacy protection issues #REFR .", "text_before_citation": ["\u2022 Methods.", "In compliance with data security management, each participant in the federated model is able to use its own data for joint training with others in a safe and fair manner without causing data leakage.", "In the metaverse, each user can be abstracted as a data owner.", "Utilizing all the potential value that may be found in data and fully optimizing business technology and services are both possible with the help of big data and FL technology.", "\u2022 Applications."], "text_after_citation": ["FATE is the first industrial-level federated learning open-source framework developed by WeBank, which can build federated models of general business scenarios."], "citing_paper_content": {"title": "Federated Learning For Metaverse: A Survey", "abstract": "The metaverse, which is at the stage of innovation and exploration, faces the dilemma of data collection and the problem of private data leakage in the process of development. This can seriously hinder the widespread deployment of the metaverse. Fortunately, federated learning (FL) is a solution to the above problems. FL is a distributed machine learning paradigm with privacy-preserving features designed for a large number of edge devices. Federated learning for metaverse (FL4M) will be a powerful tool. Because FL allows edge devices to participate in training tasks locally using their own data, computational power, and model-building capabilities. Applying FL to the metaverse not only protects the data privacy of participants but also reduces the need for high computing power and high memory on servers. Until now, there have been many studies about FL and the metaverse, respectively. In this paper, we review some of the early advances of FL4M, which will be a research direction with unlimited development potential. We first introduce the concepts of metaverse and FL, respectively. Besides, we discuss the convergence of key metaverse technologies and FL in detail, such as big data, communication technology, the Internet of Things, edge computing, blockchain, and extended reality. Finally, we discuss some key challenges and promising directions of FL4M in detail. In summary, we hope that our up-to-date brief survey can help people better understand FL4M and build a fair, open, and secure metaverse. CCS CONCEPTS \u2022 Computing methodologies \u2192 Federated learning."}, "cited_paper_content": {"title": "A Survey On Federated Learning Systems: Vision, Hype And Reality For Data Privacy And Protection", "abstract": "Federated learning has been a hot research topic in enabling the collaborative training of machine learning models among different organizations under the privacy restrictions. As researchers try to support more machine learning models with different privacy-preserving approaches, there is a requirement in developing systems and infrastructures to ease the development of various federated learning algorithms. Similar to deep learning systems such as PyTorch and TensorFlow that boost the development of deep learning, federated learning systems (FLSs) are equivalently important, and face challenges from various aspects such as effectiveness, efficiency, and privacy. In this survey, we conduct a comprehensive review on federated learning systems. To achieve smooth flow and guide future research, we introduce the definition of federated learning systems and analyze the system components. Moreover, we provide a thorough categorization for federated learning systems according to six different aspects, including data distribution, machine learning model, privacy mechanism, communication architecture, scale of federation and motivation of federation. The categorization can help the design of federated learning systems as shown in our case studies. By systematically summarizing the existing federated learning systems, we present the design factors, case studies, and future research opportunities."}, "keywords": ["privacy protection issues"], "citation_intent": "background"}
{"citing_id": "2303.01758v1", "cited_id": "1601.08188", "section_title": "Introduction", "citation": "For example, by applying a method known as lip reading, images of the mouth of the speaker or the entire face are captured by a camera, and the content of the utterance is estimated from those images #REFR .", "text_before_citation": ["However, as for the speech interface, two challenges must be overcome. First, using the interface in public places presents limitations.", "In addition to being an annoyance to the surrounding people, disclosing personal information or secret information by uttering it in public is risky in terms of information security.", "Next, it cannot be used in a noisy environment, because the accuracy of speech recognition may be declined.", "These issues are particularly acute when trying to use a speech interface to interact with wearables or mobile computers.", "To overcome these challenges, research on \"silent-speech recognition\" has been conducted #OTHEREFR ."], "text_after_citation": ["If the user could simply mouth the utterance without actually voicing it, it would be possible to use such a voice interaction in public places.", "However, with the camera method, it is necessary to install a camera in front of the face, and its form factor renders it unsuitable for wearables or mobile applications.", "Other approaches of the studies on \"non-audible murmur\" (NAM) #OTHEREFR attempts to recognize utterances with a microphone or an accelerometer worn on the skin or throat of the user.", "In this case, the user speaks with articulated respiratory sound without vocal-fold vibration (namely, whispering).", "However, to be recognized accurately by the system, a user's whispering voice tends to be noticeable by other people nearby."], "citing_paper_content": {"title": "Sottovoce: An Ultrasound Imaging-Based Silent Speech Interaction Using Deep Neural Networks", "abstract": "Computing methodologies \u2192 Neural networks; \u2022 Humancentered computing \u2192 Interface design prototyping; Mobile devices; Sound-based input / output; \u2022 Social and professional topics \u2192 People with disabilities."}, "cited_paper_content": {"title": "Lipreading With Long Short-Term Memory", "abstract": "Lipreading, i.e. speech recognition from visual-only recordings of a speaker's face, can be achieved with a processing pipeline based solely on neural networks, yielding significantly better accuracy than conventional methods. Feed-forward and recurrent neural network layers (namely Long Short-Term Memory; LSTM) are stacked to form a single structure which is trained by back-propagating error gradients through all the layers. The performance of such a stacked network was experimentally evaluated and compared to a standard Support Vector Machine classifier using conventional computer vision features (Eigenlips and Histograms of Oriented Gradients). The evaluation was performed on data from 19 speakers of the publicly available GRID corpus. With 51 different words to classify, we report a best word accuracy on held-out evaluation speakers of 79.6% using the end-to-end neural network-based solution (11.6% improvement over the best feature-based solution evaluated)."}, "keywords": ["lip reading"], "citation_intent": "method"}
{"citing_id": "2305.02535v1", "cited_id": "cs/0310022", "section_title": "Results And Paper Organization", "citation": "We take a step towards explaining this observation via a smoothed-analysis result #REFR Teng, 2004, Sankar et al., 2006] .", "text_before_citation": ["While the above results focus on the single vector Krylov method, our bounds naturally generalize to other small block sizes between 1 and k (e.g. 4, 10, or k \u2212 1).", "For general small block size b, we show that dependence on g min can be replaced with a dependence on the smallest \"b th order\" gap g min,b := min i\u2208{1,...,k\u2212b}", "\u03c3 i \u2212\u03c3 i+b \u03c3 i .", "Removing the gap dependence, Section 5.", "While a dependence on singular value gaps is unavoidable for small block methods in the worst-case, the parameter seems to rarely have an impact in practice."], "text_after_citation": ["Specifically, we leverage work in random matrix theory on eigenvalue repulsion, which shows that small spectral gaps in a matrix are brittle: adding a tiny amount of random noise to any matrix ensures that its singular value gaps are at worst inverse polynomial in the problem parameters.", "Using this fact, we present bounds that replace the dependence on log(1/g min ) in our prior results with a dependence on log( d\u03ba k \u03b4\u03b5 ) for randomly perturbed matrices, where \u03ba k = \u03c3 1 /\u03c3 k measures the conditioning of the top k singular values.", "From an algorithm design perspective, the log(1/g min ) can be removed even in the worst case by explicitly adding a random diagonal perturbation to A."], "citing_paper_content": {"title": "On The Unreasonable Effectiveness Of Single Vector Krylov Methods For Low-Rank Approximation", "abstract": "Krylov subspace methods are a ubiquitous tool for computing near-optimal rank k approximations of large matrices. While \"large block\" Krylov methods with block size at least k give the best known theoretical guarantees, block size one (a single vector) or a small constant is often preferred in practice. Despite their popularity, we lack theoretical bounds on the performance of such \"small block\" Krylov methods for low-rank approximation. We address this gap between theory and practice by proving that small block Krylov methods essentially match all known low-rank approximation guarantees for large block methods. Via a black-box reduction we show, for example, that the standard single vector Krylov method run for t iterations obtains the same spectral norm and Frobenius norm error bounds as a Krylov method with block size \u2265 k run for O(t/) iterations, up to a logarithmic dependence on the smallest gap between sequential singular values. That is, for a given number of matrix-vector products, single vector methods are essentially as effective as any choice of large block size. By combining our result with tail-bounds on eigenvalue gaps in random matrices, we prove that the dependence on the smallest singular value gap can be eliminated if the input matrix is perturbed by a small random matrix. Further, we show that single vector methods match the more complex algorithm of [Bakshi et al. '22], which combines the results of multiple block sizes to achieve an improved algorithm for Schatten p-norm low-rank approximation."}, "cited_paper_content": {"title": "Smoothed Analysis Of The Condition Numbers And Growth Factors Of Matrices", "abstract": "Let $\\orig{A}$ be any matrix and let $A$ be a slight random perturbation of $\\orig{A}$. We prove that it is unlikely that $A$ has large condition number. Using this result, we prove it is unlikely that $A$ has large growth factor under Gaussian elimination without pivoting. By combining these results, we bound the smoothed precision needed by Gaussian elimination without pivoting. Our results improve the average-case analysis of Gaussian elimination without pivoting performed by Yeung and Chan (SIAM J. Matrix Anal. Appl., 1997)."}, "keywords": ["smoothed-analysis result Teng"], "citation_intent": "background"}
{"citing_id": "2303.07030v1", "cited_id": "2001.02209", "section_title": "Semantics And Differentiability", "citation": "We can refine our semantics of Logical SDQLite by interpreting our types as generalizations of Cartesian spaces, called diffeological spaces [Iglesias-Zemmour 2013] , and our functions as morphisms of diffeological spaces, which can be thought of as generalized smooth functions between these spaces, closely following #REFR .", "text_before_citation": ["The correctness of the rewrite rules from Figure 9 is immediate, as they can all be seen as special cases of optimizations given in related work #OTHEREFR", "2022; , and they rely on the semi-ring structure of our dictionaries.", "Smooth semantics.", "The limitation of the non-smooth semantics is that it does not inform us of the differentiability of our programs.", "A function : R \u2192 R is smooth if it is differentiable and its derivative is smooth."], "text_after_citation": ["More precisely, a diffeological space is a pair ( , P ) of a set and sets of functions P \u2286 for all subsets of Euclidean spaces R , called plots.", "Each such set of plots P must contain all constant functions, be closed under pre-composition by smooth functions, and be closed under countable glueing.", "The latter means that if ( : \u2286 R \u2192 ) \u2208N are a countable family of plots such that for each , \u2208 N, and agree on \u2229 , then : \u2192 , \u2208 \u21a6 \u2192 ( ) is a plot.", "Morphisms : ( , P ) \u2192 ( , P ) between diffeological spaces are functions : \u2192 such that for all \u2208 P , \u2022 \u2208 P .", "N or B with constant functions as plots form diffeological spaces."], "citing_paper_content": {"title": "\u2207Sd: Differentiable Programming For Sparse Tensors", "abstract": "Sparse tensors are prevalent in many data-intensive applications, yet existing differentiable programming frameworks are tailored towards dense tensors. This presents a significant challenge for efficiently computing gradients through sparse tensor operations, as their irregular sparsity patterns can result in substantial memory and computational overheads. In this work, we introduce a novel framework that enables the efficient and automatic differentiation of sparse tensors, addressing this fundamental issue. Our experiments demonstrate the effectiveness of the proposed framework in terms of performance and scalability, outperforming stateof-the-art frameworks across a range of synthetic and real-world datasets. Our approach offers a promising direction for enabling efficient and scalable differentiable programming with sparse tensors, which has significant implications for numerous applications in machine learning, natural language processing, and scientific computing."}, "cited_paper_content": {"title": "Correctness Of Automatic Differentiation Via Diffeologies And Categorical Gluing", "abstract": "We present semantic correctness proofs of Automatic Differentiation (AD). We consider a forward-mode AD method on a higher order language with algebraic data types, and we characterise it as the unique structure preserving macro given a choice of derivatives for basic operations. We describe a rich semantics for differentiable programming, based on diffeological spaces. We show that it interprets our language, and we phrase what it means for the AD method to be correct with respect to this semantics. We show that our characterisation of AD gives rise to an elegant semantic proof of its correctness based on a gluing construction on diffeological spaces. We explain how this is, in essence, a logical relations argument. Finally, we sketch how the analysis extends to other AD methods by considering a continuation-based method."}, "keywords": ["generalized smooth functions", "diffeological spaces"], "citation_intent": "background"}
{"citing_id": "2303.06628v1", "cited_id": "1803.05407", "section_title": "Weight Ensemble In Parameter Space", "citation": "The method is related to Stochastic Weight Averaging (SWA) #REFR , but we do not use a modified learning rate schedule here because instead of getting better generalization ability, WE aim to give an improved learning-forgetting tradeoff. As shown in Fig.", "text_before_citation": ["3) , the model performs better on new tasks while losing accuracy on previous ones.", "The models among training compose a sequence of different learning-forgetting tradeoffs.", "Instead of interpolating only between the initial and the final model, our method weight ensemble (WE) averages the weights in the sequence during the training time:", "EQUATION", "where model weight sampling happens every I iteration."], "text_after_citation": ["3 , WE achieves better performance on downstream tasks than WiSE-FT.", "In addition, while WiSE-FT is sensitive to different values of \u03b1, our method is much more robust under different hyperparameter (I) choices."], "citing_paper_content": {"title": "Preventing Zero-Shot Transfer Degradation In Continual Learning Of Vision-Language Models", "abstract": "Continual learning (CL) can help pre-trained visionlanguage models efficiently adapt to new or under-trained data distributions without retraining. Nevertheless, during the continual training of the Contrastive Language-Image Pre-training (CLIP) model, we observe that the model's zero-shot transfer ability significantly degrades due to catastrophic forgetting. Existing CL methods can mitigate forgetting by replaying previous data. However, since the CLIP dataset is private, replay methods cannot access the pre-training dataset. In addition, replaying data of previously learned downstream tasks can enhance their performance but comes at the cost of sacrificing zero-shot performance. To address this challenge, we propose a novel method ZSCL to prevent zero-shot transfer degradation in the continual learning of vision-language models in both feature and parameter space. In the feature space, a reference dataset is introduced for distillation between the current and initial models. The reference dataset should have semantic diversity but no need to be labeled, seen in pre-training, or matched image-text pairs. In parameter space, we prevent a large parameter shift by averaging weights during the training. We propose a more challenging Multi-domain Task Incremental Learning (MTIL) benchmark to evaluate different methods, where tasks are from various domains instead of class-separated in a single dataset. Our method outperforms other methods in the traditional class-incremental learning setting and the MTIL by 9.7% average score. Our code locates at https: //github.com/Thunderbeee/ZSCL."}, "cited_paper_content": {"title": "Averaging Weights Leads To Wider Optima And Better Generalization", "abstract": "Deep neural networks are typically trained by optimizing a loss function with an SGD variant, in conjunction with a decaying learning rate, until convergence. We show that simple averaging of multiple points along the trajectory of SGD, with a cyclical or constant learning rate, leads to better generalization than conventional training. We also show that this Stochastic Weight Averaging (SWA) procedure finds much broader optima than SGD, and approximates the recent Fast Geometric Ensembling (FGE) approach with a single model. Using SWA we achieve notable improvement in test accuracy over conventional SGD training on a range of state-of-the-art residual networks, PyramidNets, DenseNets, and Shake-Shake networks on CIFAR-10, CIFAR-100, and ImageNet. In short, SWA is extremely easy to implement, improves generalization, and has almost no computational overhead."}, "keywords": ["improved learning-forgetting tradeoff", "Stochastic Weight Averaging"], "citation_intent": "method"}
{"citing_id": "2304.09987v1", "cited_id": "2003.06620", "section_title": "Method", "citation": "In many scenarios, an approximation of the scene geometry is either given, e.g., when using SfM to compute the input camera poses, or can be computed, e.g., via MVS or single-view depth predictions #REFR .", "text_before_citation": ["A common strategy in the literature is to represent the scene explicitly through a voxel volume.", "In contrast to this uniform subdivision, we investigate using an adaptive subdivision of the scene."], "text_after_citation": ["This allows us to compute an adaptive subdivision of the scene via Delaunay triangulation #OTHEREFR of such a point cloud.", "This results in a set of non-overlapping tetrahedra, where smaller tetrahedra are created close to the surface of the scene (c.f . Fig. 2 ).", "In the following, we explain how this adaptive subdivision of the scene can be used instead of voxels for volume rendering and neural rendering."], "citing_paper_content": {"title": "Tetra-Nerf: Representing Neural Radiance Fields Using Tetrahedra", "abstract": "2) scene representation: dense tetrahedra 3) barycetric interpolation of features 4) shallow MLP predicts density and colour Figure 1. The input to Tetra-NeRF is a point cloud which is triangulated to get a set of tetrahedra used to represent the radiance field. Rays are sampled, and the field is queried. The barycentric interpolation is used to interpolate tetrahedra vertices, and the resulting features are passed to a shallow MLP to get the density and colours for volumetric rendering."}, "cited_paper_content": {"title": "Monocular Depth Estimation Based On Deep Learning: An Overview", "abstract": "Depth information is important for autonomous systems to perceive environments and estimate their own state. Traditional depth estimation methods, like structure from motion and stereo vision matching, are built on feature correspondences of multiple viewpoints. Meanwhile, the predicted depth maps are sparse. Inferring depth information from a single image (monocular depth estimation) is an ill-posed problem. With the rapid development of deep neural networks, monocular depth estimation based on deep learning has been widely studied recently and achieved promising performance in accuracy. Meanwhile, dense depth maps are estimated from single images by deep neural networks in an end-to-end manner. In order to improve the accuracy of depth estimation, different kinds of network frameworks, loss functions and training strategies are proposed subsequently. Therefore, we survey the current monocular depth estimation methods based on deep learning in this review. Initially, we conclude several widely used datasets and evaluation indicators in deep learning-based depth estimation. Furthermore, we review some representative existing methods according to different training manners: supervised, unsupervised and semi-supervised. Finally, we discuss the challenges and provide some ideas for future researches in monocular depth estimation."}, "keywords": ["single-view depth predictions"], "citation_intent": "method"}
{"citing_id": "2305.01795v1", "cited_id": "1809.00812", "section_title": "B.2 Details Of Module Outputs", "citation": "C Dataset Details C.1 RECIPEPLAN Data Repurpose RECIPEQA was proposed in #REFR that provide four tasks (Textual Cloze, Visual Cloze, Visual Ordering, Visual Coherence) for multimodal machine comprehension of cooking recipes. This dataset contains question-answer pairs generated from copyrightfree recipes.", "text_before_citation": ["We showcase all the details of the outputs of each module for the example task \"How to make a candy bouquet\" in Figure 8 .", "In addition, we showcase examples on \"How to make peppermint patty pudding shot\" and \"How to make bartending for beginners simple margarita\" in Figure 9 ."], "text_after_citation": ["Each of them is under a license, which is provided in each data JSON file.", "We collect RECIPEPLAN by repurposing the test dataset from RECIPEQA that relates to the Visual Ordering task as the sequence generation task for the multimodal procedural planning evaluation testbed.", "We use recipe instructions as textual plan reference and their paired images as visual plan reference.", "Dataset Statistics We visualize two examples of our repurposed RECIPEPLAN for multimodal procedural planning in Figure 10 .", "We also show the word-cloud distribution of task name and textual plan reference in Figure 12 ."], "citing_paper_content": {"title": "Multimodal Procedural Planning Via Dual Text-Image Prompting", "abstract": "How to make traditional szechuan chicken? Step 5: Add other ingredients like sichuan peppercorn powder and chili bean paste, then stir fry. Step 6: Pour in remaining marinade mixture into the same wok and stir fry until combined. Step 7: Return chicken pieces into the wok and stir fry until all liquid forms a glossy glaze. Step 8: Transfer Szechuan Chicken onto serving plate garnished with green & serve immediately! Step 1: Gather the ingredients such as chicken, garlic, ginger, and light soy sauce. Step 2: Marinade the chicken with a teaspoon of light soy sauce and cooking wine. Step 3: Heat up a wok over medium-high heat until it starts to smoke and add oil. Step 4: Add in marinated chicken pieces and stir fry until they turn golden brown. Figure 1: Our dual Text-Image Prompting (TIP) model generates coherent and authentic multimodal procedural plans with multiple steps towards a high-level goal, providing useful guidelines in task completion."}, "cited_paper_content": {"title": "Recipeqa: A Challenge Dataset For Multimodal Comprehension Of Cooking Recipes", "abstract": "Understanding and reasoning about cooking recipes is a fruitful research direction towards enabling machines to interpret procedural text. In this work, we introduce RecipeQA, a dataset for multimodal comprehension of cooking recipes. It comprises of approximately 20K instructional recipes with multiple modalities such as titles, descriptions and aligned set of images. With over 36K automatically generated question-answer pairs, we design a set of comprehension and reasoning tasks that require joint understanding of images and text, capturing the temporal flow of events and making sense of procedural knowledge. Our preliminary results indicate that RecipeQA will serve as a challenging test bed and an ideal benchmark for evaluating machine comprehension systems. The data and leaderboard are available at this http URL"}, "keywords": ["multimodal machine comprehension"], "citation_intent": "method"}
{"citing_id": "2303.14371v1", "cited_id": "1505.04597", "section_title": "Registration-And Uncertainty-Based Framework For Subject-Level One-Shot White Matter Tract Segmentation", "citation": "Pseudo subjects and corresponding pseudo labels are generated by deep-learning-based registration using 3D U-Net #REFR .", "text_before_citation": ["Our RPA module based on more advanced deep-learning-based registration methods #OTHEREFR can potentially obtain the high-quality pseudo dataset.", "We modified registration methods in #OTHEREFR , which are originally for MRI data, to be used for our peak subjects.", "Let {x, l} be the only labeled subject and its corresponding tract segmentation label, and y (i) be a set of unlabeled subjects on a spatial domain R 3 .", "Pseudo peak subjects and their corresponding labels are synthesized with the RPA module.", "Specifically, a spatial transform model is learned to register the labeled subject to unlabeled subjects."], "text_after_citation": ["Set g(x, y (i) ) = u (i) is the spatial transform model, where \u03b8 are the model parameters, and the output of the model u is the voxel-wise displacement field.", "The deformation function \u03d5 (i) = id + u (i) , where id is the identity transform #OTHEREFR .", "Therefore, the pseudo subject x (i) and the corresponding pseudo label l (i) are as followed:", "EQUATION", "EQUATION"], "citing_paper_content": {"title": "A Registration-And Uncertainty-Based Framework For White Matter Tract Segmentation With Only One Annotated Subject", "abstract": "White matter (WM) tract segmentation based on diffusion magnetic resonance imaging (dMRI) plays an important role in the analysis of human health and brain diseases. However, the annotation of WM tracts is time-consuming and needs experienced neuroanatomists. In this study, to explore tract segmentation in the challenging setting of minimal annotations, we propose a novel framework utilizing only one annotated subject (subject-level one-shot) for tract segmentation. Our method is constructed by proposed registration-based peak augmentation (RPA) and uncertaintybased refining (URe) modules. RPA module synthesizes pseudo subjects and their corresponding labels to improve the tract segmentation performance. The proposed URe module alleviates the negative influence of the low-confidence voxels on pseudo subjects. Experimental results show that our method outperforms other state-of-the-art methods by a large margin, and our proposed modules are effective. Overall, our method achieves accurate whole-brain tract segmentation with only one annotated subject. Our code is available at https://github.com/HaoXu0507/ ISBI2023-One-Shot-WM-Tract-Segmentation."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["deep-learning-based registration"], "citation_intent": "method"}
{"citing_id": "2305.00537v1", "cited_id": "1710.00935", "section_title": "Ii. Classical Nn Based Methods", "citation": "Hence, for this class of models, the word 'interpretability' refers to the ability to clarify and extract knowledge representations in different layers of NN-based models as defined in #REFR .", "text_before_citation": ["In the past several years, classical NN-based models have achieved great success and outperform humans in numerous difficult tasks, such as image and visual classification, natural language processing and recognition, and board games.", "However, the black-box nature of the classical models presents a real challenge to explain the underlying mechanisms and behaviors of the networks.", "The popular solution to address this problem is to explore interpretability via explaining the inside of the black-box."], "text_after_citation": ["In this section, the most studied methods in this class, FFNN based and DL based, are surveyed."], "citing_paper_content": {"title": "Interpretability Of Machine Learning: Recent Advances And Future Prospects", "abstract": "The proliferation of machine learning (ML) has drawn unprecedented interest in the study of various multimedia contents such as text, image, audio and video, among others. Consequently, understanding and learning ML-based representations have taken center stage in knowledge discovery in intelligent multimedia research and applications. Nevertheless, the black-box nature of contemporary ML, especially in deep neural networks (DNNs), has posed a primary challenge for ML-based representation learning. To address this black-box problem, the studies on interpretability of ML have attracted tremendous interests in recent years. This paper presents a survey on recent advances and future prospects on interpretability of ML, with several application examples pertinent to multimedia computing, including text-image cross-modal representation learning, face recognition, and the recognition of objects. It is evidently shown that the study of interpretability of ML promises an important research direction, one which is worth further investment in."}, "cited_paper_content": {"title": "Interpretable Convolutional Neural Networks", "abstract": "This paper proposes a method to modify a traditional convolutional neural network (CNN) into an interpretable CNN, in order to clarify knowledge representations in high conv-layers of the CNN. In an interpretable CNN, each filter in a high conv-layer represents a specific object part. Our interpretable CNNs use the same training data as ordinary CNNs without a need for any annotations of object parts or textures for supervision. The interpretable CNN automatically assigns each filter in a high conv-layer with an object part during the learning process. We can apply our method to different types of CNNs with various structures. The explicit knowledge representation in an interpretable CNN can help people understand the logic inside a CNN, i.e. what patterns are memorized by the CNN for prediction. Experiments have shown that filters in an interpretable CNN are more semantically meaningful than those in a traditional CNN. The code is available at https://github.com/zqs1022/interpretableCNN."}, "keywords": ["interpretability"], "citation_intent": "background"}
{"citing_id": "2304.13242v1", "cited_id": "2002.11477", "section_title": "A. Partial World State Representation", "citation": "The coefficients a 0 , a 1 , and a 2 are derived by satisfying boundary conditions #REFR . Fig. 2 shows visual examples of a sample augmentation.", "text_before_citation": ["We leverage geometric data augmentation #OTHEREFR to improve model generalization performance by learning geometric invariance.", "Each sample is augmented by random rotation and translation, and a polynomial warping function applied on the dense maps and observed trajectories", "a 0 (\u03be ) 2 + a 1 (\u03be ) + a 2 = \u03be (1)", "where \u03be is a substitute for spatial coordinates i and j, and \u03be denotes warped coordinates.", "We create dense warp maps by using the inverse function of (1) to map each warped coordinate \u03be to an original coordinate \u03be."], "text_after_citation": [], "citing_paper_content": {"title": "Learning To Predict Navigational Patterns From Partial Observations", "abstract": "Human beings cooperatively navigate ruleconstrained environments by adhering to mutually known navigational patterns, which may be represented as directional pathways or road lanes. Inferring these navigational patterns from incompletely observed environments is required for intelligent mobile robots operating in unmapped locations. However, algorithmically defining these navigational patterns is nontrivial. This paper presents the first self-supervised learning (SSL) method for learning to infer navigational patterns in realworld environments from partial observations only. We explain how geometric data augmentation, predictive world modeling, and an information-theoretic regularizer enables our model to predict an unbiased local directional soft lane probability (DSLP) field in the limit of infinite data. We demonstrate how to infer global navigational patterns by fitting a maximum likelihood graph to the DSLP field. Experiments show that our SSL model outperforms two SOTA supervised lane graph prediction models on the nuScenes dataset. We propose our SSL method as a scalable and interpretable continual learning paradigm for navigation by perception. Code released upon publication."}, "cited_paper_content": {"title": "Learning A Directional Soft Lane Affordance Model For Road Scenes Using Self-Supervision", "abstract": "Humans navigate complex environments in an organized yet flexible manner, adapting to the context and implicit social rules. Understanding these naturally learned patterns of behavior is essential for applications such as autonomous vehicles. However, algorithmically defining these implicit rules of human behavior remains difficult. This work proposes a novel self-supervised method for training a probabilistic network model to estimate the regions humans are most likely to drive in as well as a multimodal representation of the inferred direction of travel at each point. The model is trained on individual human trajectories conditioned on a representation of the driving environment. The model is shown to successfully generalize to new road scenes, demonstrating potential for real-world application as a prior for socially acceptable driving behavior in challenging or ambiguous scenarios which are poorly handled by explicit traffic rules."}, "keywords": ["sample augmentation"], "citation_intent": "background"}
{"citing_id": "2303.17166v1", "cited_id": "1903.01292", "section_title": "Results Of Training The Distortion Estimator", "citation": "By contrast, our distortion estimator has two regressors for the focal length f and distortion coefficient k1. The input image is generated from #REFR .", "text_before_citation": ["We report the performance of our distortion estimator to describe the difference between Wakai's method #OTHEREFR and the distortion estimator.", "Our distortion estimator is composed of Wakai's calibration network #OTHEREFR without the tilt Figure 10 : Network architectures of (a) Wakai's calibration network #OTHEREFR and (b) our distortion network.", "Wakai's calibration network predicts extrinsics (tilt \u03b8 and roll \u03c8 angles) and intrinsics (focal length f and a distortion coefficient k1)."], "text_after_citation": ["and roll angle regressors, as shown in Figure 10 .", "We optimized the distortion estimator after pretraining on Wakai's calibration network.", "The distortion estimator achieved slight improvements in the focal length f and distortion coefficient k 1 because the number of predicted camera parameters was reduced by two, that is, the tilt and roll angles, as shown in Table 12 ."], "citing_paper_content": {"title": "Deep Single Image Camera Calibration By Heatmap Regression To Recover Fisheye Images Under Manhattan World Assumption Without Ambiguity", "abstract": "In orthogonal world coordinates, a Manhattan world lying along cuboid buildings is widely useful for various computer vision tasks. However, the Manhattan world has much room for improvement because the origin of pan angles from an image is arbitrary, that is, four-fold rotational symmetric ambiguity of pan angles. To address this problem, we propose a definition for the pan-angle origin based on the directions of the roads with respect to a camera and the direction of travel. We propose a learning-based calibration method that uses heatmap regression to remove the ambiguity by each direction of labeled image coordinates, similar to pose estimation keypoints. Simultaneously, our two-branched network recovers the rotation and removes fisheye distortion from a general scene image. To alleviate the lack of vanishing points in images, we introduce auxiliary diagonal points that have the optimal 3D arrangement of spatial uniformity. Extensive experiments demonstrated that our method outperforms conventional methods on large-scale datasets and with off-the-shelf cameras."}, "cited_paper_content": {"title": "The Streetlearn Environment And Dataset", "abstract": "Navigation is a rich and well-grounded problem domain that drives progress in many different areas of research: perception, planning, memory, exploration, and optimisation in particular. Historically these challenges have been separately considered and solutions built that rely on stationary datasets - for example, recorded trajectories through an environment. These datasets cannot be used for decision-making and reinforcement learning, however, and in general the perspective of navigation as an interactive learning task, where the actions and behaviours of a learning agent are learned simultaneously with the perception and planning, is relatively unsupported. Thus, existing navigation benchmarks generally rely on static datasets (Geiger et al., 2013; Kendall et al., 2015) or simulators (Beattie et al., 2016; Shah et al., 2018). To support and validate research in end-to-end navigation, we present StreetLearn: an interactive, first-person, partially-observed visual environment that uses Google Street View for its photographic content and broad coverage, and give performance baselines for a challenging goal-driven navigation task. The environment code, baseline agent code, and the dataset are available at this http URL"}, "keywords": ["distortion estimator", "input image"], "citation_intent": "method"}
{"citing_id": "2303.13386v1", "cited_id": "1808.06752", "section_title": "Baselines", "citation": "For MedNLI, we compare with the best transfer learning results so far, ESIM (MultiNLI) which was trained on MultiNLI datasets #REFR .", "text_before_citation": ["Models only evaluated on NLI do not require summarisation data, hence \"\u2212\".", "\"\u223c\" indicates that BERT and CXR-BERT were fine-tuned on MedNLI, a 'neardomain' NLI dataset.", "BERT -Miura et al. \u223c / \u2212 CXR-BERT -Boecking et al. \u223c / \u2212 ESIM -Chen et al.", "/ \u2212 T0 & T0++ / GPT-3 / GPT-3-{NLI, GW} / CXR-BERT-NLI / \u2212 PubMedBERT-NLI / \u2212 SciFive large -{NLI, GW} / T5 large -MLM\u2192Task / DOT5 /", "NLI."], "text_after_citation": ["For radiology summarisation, to our knowledge, we are the first to report results on direct transfer from general-domain summarisation.", "For biomedical summarisation, since we use a new dataset (PubMed ShortSum), there is no prior comparison.", "Large language models T0 #OTHEREFR and GPT-3 (Brown et al., 2020) are massively pretrained language models that can be used off-theshelf for zero-shot or few-shot inference.", "T0 is pretrained with multiple tasks including generaldomain summarisation datasets (but not NLI), and shows strong transfer ability #OTHEREFR .", "T0 can be seen as a strong general-domain summarisation model and also strong zero-shot domain transfer baseline on summarisation."], "citing_paper_content": {"title": "Compositional Zero-Shot Domain Transfer With Text-To-Text Models", "abstract": "Label scarcity is a bottleneck for improving task performance in specialised domains. We propose a novel compositional transfer learning framework (DOT5 1) for zeroshot domain transfer. Without access to in-domain labels, DOT5 jointly learns domain knowledge (from masked language modelling of unlabelled in-domain free text) and task knowledge (from task training on more readily available general-domain data) in a multi-task manner. To improve the transferability of task training, we design a strategy named NLGU: we simultaneously train natural language generation (NLG) for indomain label-to-data generation which enables data augmentation for self-finetuning and natural language understanding (NLU) for label prediction. We evaluate DOT5 on the biomedical domain and the resource-lean subdomain of radiology, focusing on natural language inference, text summarisation and embedding learning. DOT5 demonstrates the effectiveness of compositional transfer learning through multi-task learning. In particular, DOT5 outperforms the current stateof-the-art in zero-shot transfer by over 7 absolute points in accuracy on RadNLI. We validate DOT5 with ablations and a case study demonstrating its ability to solve challenging NLI examples requiring in-domain expertise. * Work done at Microsoft Health Futures. 1 DOT5 (read as \"dot five\"): Domain Compositional ZerOshot T5."}, "cited_paper_content": {"title": "Lessons From Natural Language Inference In The Clinical Domain", "abstract": "State of the art models using deep neural networks have become very good in learning an accurate mapping from inputs to outputs. However, they still lack generalization capabilities in conditions that differ from the ones encountered during training. This is even more challenging in specialized, and knowledge intensive domains, where training data is limited. To address this gap, we introduce MedNLI - a dataset annotated by doctors, performing a natural language inference task (NLI), grounded in the medical history of patients. We present strategies to: 1) leverage transfer learning using datasets from the open domain, (e.g. SNLI) and 2) incorporate domain knowledge from external data and lexical sources (e.g. medical terminologies). Our results demonstrate performance gains using both strategies."}, "keywords": ["MultiNLI datasets"], "citation_intent": "result"}
{"citing_id": "2303.05937v1", "cited_id": "1702.04405", "section_title": "Implementation", "citation": "Single-view planar estimation results on ScanNet #REFR by measuring the plane recall with a fixed Intersection over Union threshold 0.5 and a varying depth error threshold (from 0 to 0.8m).", "text_before_citation": ["The model is trained on 4 NVIDIA-V100 GPUs for a total of 100k steps.", "Training.", "Before our main training, we add a bootstrap training phase for 50k steps, where we initialize the alpha prediction with the segmentation by turning off L rgb and enforcing a segmentation loss on the alpha channel.", "To train the model with multi-view input, we set T = 2 and the trained model can be applied for T \u2265 1.", "We cancel L pln for non-planar regions in the second input view. Figure 5 ."], "text_after_citation": [], "citing_paper_content": {"title": "Structural Multiplane Image: Bridging Neural View Synthesis And 3D Reconstruction", "abstract": "The Multiplane Image (MPI), containing a set of frontoparallel RGB\u03b1 layers, is an effective and efficient representation for view synthesis from sparse inputs. Yet, its fixed structure limits the performance, especially for surfaces imaged at oblique angles. We introduce the Structural MPI (S-MPI), where the plane structure approximates 3D scenes concisely. Conveying RGB\u03b1 contexts with geometricallyfaithful structures, the S-MPI directly bridges view synthesis and 3D reconstruction. It can not only overcome the critical limitations of MPI, i.e., discretization artifacts from sloped surfaces and abuse of redundant layers, and can also acquire planar 3D reconstruction. Despite the intuition and demand of applying S-MPI, great challenges are introduced, e.g., high-fidelity approximation for both RGB\u03b1 layers and plane poses, multi-view consistency, non-planar regions modeling, and efficient rendering with intersected planes. Accordingly, we propose a transformer-based network based on a segmentation model [4]. It predicts compact and expressive S-MPI layers with their corresponding masks, poses, and RGB\u03b1 contexts. Non-planar regions are inclusively handled as a special case in our unified framework. Multi-view consistency is ensured by sharing global proxy embeddings, which encode plane-level features covering the complete 3D scenes with aligned coordinates. Intensive experiments show that our method outperforms both previous state-of-the-art MPI-based view synthesis methods and planar reconstruction methods."}, "cited_paper_content": {"title": "Scannet: Richly-Annotated 3D Reconstructions Of Indoor Scenes", "abstract": "A key requirement for leveraging supervised deep learning methods is the availability of large, labeled datasets. Unfortunately, in the context of RGB-D scene understanding, very little data is available \u2013 current datasets cover a small range of scene views and have limited semantic annotations. To address this issue, we introduce ScanNet, an RGB-D video dataset containing 2.5M views in 1513 scenes annotated with 3D camera poses, surface reconstructions, and semantic segmentations. To collect this data, we designed an easy-to-use and scalable RGB-D capture system that includes automated surface reconstruction and crowdsourced semantic annotation. We show that using this data helps achieve state-of-the-art performance on several 3D scene understanding tasks, including 3D object classification, semantic voxel labeling, and CAD model retrieval."}, "keywords": ["Single-view planar estimation"], "citation_intent": "method"}
{"citing_id": "2304.08304v1", "cited_id": "1910.06528", "section_title": "B. Image And Lidar Pipeline", "citation": "Following MVF #REFR , we use dynamic voxelization, which establishes the bi-directional mapping between points and voxels without sampling operation.", "text_before_citation": ["For images captured by a monocular camera, we train a semantic segmentation backbone, i.e., LR-ASPP #OTHEREFR , to generate image feature maps F img \u2208 R H\u00d7W \u00d7C .", "Each channel in the image feature map represents the confidence score for a specific category, including background.", "Notice that the parameters of these layers are frozen during training.", "For LiDAR point clouds, voxelization can divide points into an evenly spaced grid of voxels.", "The size of a voxel is predefined as (l, w, h), representing length along X, Y, Z axes in 3D space."], "text_after_citation": ["For Multi-scale Voxelization Fusion Framework, we define different voxel sizes and apply voxelization within every scale.", "Point p i is assigned to voxels {v j0 , . . .", ", v js } based on their relative spatial coordinates in different scales."], "citing_paper_content": {"title": "Sdvrf: Sparse-To-Dense Voxel Region Fusion For Multi-Modal 3D Object Detection", "abstract": "In the perception task of autonomous driving, multimodal methods have become a trend due to the complementary characteristics of LiDAR point clouds and image data. However, the performance of previous methods is usually limited by the sparsity of the point cloud or the noise problem caused by the misalignment between LiDAR and the camera. To solve these two problems, we present a new concept, Voxel Region (VR), which is obtained by projecting the sparse local point clouds in each voxel dynamically. And we propose a novel fusion method, named Sparse-to-Dense Voxel Region Fusion (SDVRF). Specifically, more pixels of the image feature map inside the VR are gathered to supplement the voxel feature extracted from sparse points and achieve denser fusion. Meanwhile, different from prior methods, which project the size-fixed grids, our strategy of generating dynamic regions achieves better alignment and avoids introducing too much background noise. Furthermore, we propose a multiscale fusion framework to extract more contextual information and capture the features of objects of different sizes. Experiments on the KITTI dataset show that our method improves the performance of different baselines, especially on classes of small size, including Pedestrian and Cyclist."}, "cited_paper_content": {"title": "End-To-End Multi-View Fusion For 3D Object Detection In Lidar Point Clouds", "abstract": "Recent work on 3D object detection advocates point cloud voxelization in birds-eye view, where objects preserve their physical dimensions and are naturally separable. When represented in this view, however, point clouds are sparse and have highly variable point density, which may cause detectors difficulties in detecting distant or small objects (pedestrians, traffic signs, etc.). On the other hand, perspective view provides dense observations, which could allow more favorable feature encoding for such cases. In this paper, we aim to synergize the birds-eye view and the perspective view and propose a novel end-to-end multi-view fusion (MVF) algorithm, which can effectively learn to utilize the complementary information from both. Specifically, we introduce dynamic voxelization, which has four merits compared to existing voxelization methods, i) removing the need of pre-allocating a tensor with fixed size; ii) overcoming the information loss due to stochastic point/voxel dropout; iii) yielding deterministic voxel embeddings and more stable detection outcomes; iv) establishing the bi-directional relationship between points and voxels, which potentially lays a natural foundation for cross-view feature fusion. By employing dynamic voxelization, the proposed feature fusion architecture enables each point to learn to fuse context information from different views. MVF operates on points and can be naturally extended to other approaches using LiDAR point clouds. We evaluate our MVF model extensively on the newly released Waymo Open Dataset and on the KITTI dataset and demonstrate that it significantly improves detection accuracy over the comparable single-view PointPillars baseline."}, "keywords": ["voxels"], "citation_intent": "method"}
{"citing_id": "2304.08307v1", "cited_id": "1505.04597", "section_title": "D. Architecture And Training Of Neural Network", "citation": "The U-net architecture was used #REFR because of the ability to extract features from the input data at different spatial resolutions which are later used in the decoder part of the network to form a prediction.", "text_before_citation": ["All calculations were performed on a DGX station equipped with Tesla V100 GPU cards (Nvidia, Santa Clara, CA, US). The PyTorch DL framework #OTHEREFR was used."], "text_after_citation": ["The network had 4 levels, with the encoder part at each level containing the two 3D convolution layers each followed by the leakyRelu activation.", "The spatial resolution was decreased with a max-pooling layer by a factor of two.", "The bottom of the U-net consists of two 3D convolution and one 3D transposed convolution layers, each followed by leakyRelu activation.", "The decoder part, at each level, consists of two 3D transposed convolution each followed by leaky Relu.", "The spatial resolution was increase by a trilinear interpolation by factor of 2."], "citing_paper_content": {"title": "Predicting Dynamic, Motion-Related Changes In B 0 Field In The Brain At A 7 T Mri Using A Subject-Specific Fine-Tuned U-Net", "abstract": "Subject movement during the magnetic resonance examination is inevitable and causes not only image artefacts but also deteriorates the homogeneity of the main magnetic field (B0), which is a prerequisite for high quality data. Thus, characterization of changes to B0, e.g. induced by patient movement, is important for MR applications that are prone to B0 inhomogeneities. We propose a deep learning based method to predict such changes within the brain from the change of the head position to facilitate retrospective or even real-time correction. A 3D U-net was trained on in vivo brain 7 T MRI data. The input consisted of B0 maps and anatomical images at an initial position, and anatomical images at a different head position (obtained by applying a rigid-body transformation on the initial anatomical image). The output consisted of B0 maps at the new head positions. We further fine-tuned the network weights to each subject by measuring a limited number of head positions of the given subject, and trained the U-net with these data. Our approach was compared to established dynamic B0 field mapping via interleaved navigators, which suffer from limited spatial resolution and the need for undesirable sequence modifications. Qualitative and quantitative comparison showed similar performance between an interleaved navigator-equivalent method and proposed method. We therefore conclude that it is feasible to predict B0 maps from rigid subject movement and, when combined with external tracking hardware, this information could be used to improve the quality of magnetic resonance acquisitions without the use of navigators."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["prediction", "U-net architecture"], "citation_intent": "method"}
{"citing_id": "2303.04980v1", "cited_id": "1911.07140", "section_title": "Random Gradient-Free Optimization", "citation": "Huang & Zhang #REFR generated the step using a generator network. These methods properly solved score-based attack using RGF.", "text_before_citation": ["While the attacker cannot know the architecture or the gradient of the victim model (black-box attack), the perturbation cannot be updated through gradient-based optimization.", "Some researchers solved this problem by using the RGF algorithm proposed by Nesterov #OTHEREFR .", "Most RGFbased adversarial attack aims to find ways to find good random steps. M.Chenget al.", "#OTHEREFR proposed binary searchbased RGF optimization to build a perturbation efficiently. S.Chenget al.", "#OTHEREFR sampled random steps on a hypersphere and adopted the momentum optimization method proposed by Polyak #OTHEREFR ."], "text_after_citation": ["So far, RGF is not appropriate for decision-based attacks because it cannot evaluate the direction properly, especially for universal attacks."], "citing_paper_content": {"title": "Decision-Badge: Decision-Based Adversarial Batch Attack With Directional Gradient Estimation", "abstract": "The vulnerability of deep neural networks to adversarial examples has led to the rise in the use of adversarial attacks. While various decision-based and universal attack methods have been proposed, none have attempted to create a decision-based universal adversarial attack. This research proposes Decision-BADGE, which uses random gradientfree optimization and batch attack to generate universal adversarial perturbations for decision-based attacks. Multiple adversarial examples are combined to optimize a single universal perturbation, and the accuracy metric is reformulated into a continuous Hamming distance form. The effectiveness of accuracy metric as a loss function is demonstrated and mathematically proven. The combination of Decision-BADGE and the accuracy loss function performs better than both score-based image-dependent attack and white-box universal attack methods in terms of attack time efficiency. The research also shows that Decision-BADGE can successfully deceive unseen victims and accurately target specific classes. The code is available at our GitHub 1 repository."}, "cited_paper_content": {"title": "Black-Box Adversarial Attack With Transferable Model-Based Embedding", "abstract": "We present a new method for black-box adversarial attack. Unlike previous methods that combined transfer-based and scored-based methods by using the gradient or initialization of a surrogate white-box model, this new method tries to learn a low-dimensional embedding using a pretrained model, and then performs efficient search within the embedding space to attack an unknown target network. The method produces adversarial perturbations with high level semantic patterns that are easily transferable. We show that this approach can greatly improve the query efficiency of black-box adversarial attack across different target network architectures. We evaluate our approach on MNIST, ImageNet and Google Cloud Vision API, resulting in a significant reduction on the number of queries. We also attack adversarially defended networks on CIFAR10 and ImageNet, where our method not only reduces the number of queries, but also improves the attack success rate."}, "keywords": ["score-based attack"], "citation_intent": "method"}
{"citing_id": "2303.13755v1", "cited_id": "1706.03762", "section_title": "Introduction", "citation": "Rather, it is desirable to design sparse attention algorithms that mimic full self attention's instancedependent nature #REFR , thereby capturing the variable distribution of semantic information in the input image content.", "text_before_citation": ["For example, local and strided fixed attention patterns are used #OTHEREFR , in combination with randomlyinitialized connectivities #OTHEREFR .", "However, such fixed attention patterns limit the the capacity of the self-attention connections to a fixed subset of tokens (Fig. 1) .", "These attention patterns' fixed nature is less effective compared with the direct communication between tokens in full self attention.", "For example, Swin transformer #OTHEREFR has a limited receptive field at shallower layers and needs many layers to model long-range dependencies.", "And BigBird #OTHEREFR needs to combine multiple fixed attention patterns to achieve good performance."], "text_after_citation": ["To address aforementioned challenges, we propose a method called Sparsifiner that learns to compute sparse connectivity patterns over attention that are both instancedependent and unstructured.", "The instance-dependent nature of the attention pattern allows each token to use its limited attention budget of nonzero elements more efficiently compared to fixed sparse attention patterns.", "For example, in attention heads that attend to semantic rather than positional content #OTHEREFR , tokens containing similar semantic information should be considered to have high connectivity scores despite their spatial distance.", "Similarly, nearby tokens with irrelevant semantic relation should have lower connectivity scores despite their spatial proximity.", "Furthermore, Sparsifiner improves attention pattern flexibility compared to token pruning by pruning individual connectivities, instead of entire rows and columns of the attention matrix."], "citing_paper_content": {"title": "Sparsifiner: Learning Sparse Instance-Dependent Attention For Efficient Vision Transformers", "abstract": "Vision Transformers (ViT) have shown their competitive advantages performance-wise compared to convolutional neural networks (CNNs) though they often come with high computational costs. To this end, previous methods explore different attention patterns by limiting a fixed number of spatially nearby tokens to accelerate the ViT's multi-head selfattention (MHSA) operations. However, such structured attention patterns limit the token-to-token connections to their spatial relevance, which disregards learned semantic connections from a full attention mask. In this work, we propose a novel approach to learn instance-dependent attention patterns, by devising a lightweight connectivity predictor module to estimate the connectivity score of each pair of tokens. Intuitively, two tokens have high connectivity scores if the features are considered relevant either spatially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature and therefore provide the opportunity to accelerate the network via sparse computations. Equipped with the learned unstructured attention pattern, sparse attention ViT (Sparsifiner) produces a superior Pareto-optimal trade-off between FLOPs and top-1 accuracy on ImageNet compared to token sparsity. Our method reduces 48% \u223c 69% FLOPs of MHSA while the accuracy drop is within 0.4%. We also show that combining attention and token sparsity reduces ViT FLOPs by over 60%."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["sparse attention algorithms"], "citation_intent": "background"}
{"citing_id": "2304.02993v1", "cited_id": "1707.08150", "section_title": "I. Introduction", "citation": "This is because, unlike mobile/service robots, manipulator robots are expected to physically interact with their environment through precise and fine-grained actions, such as reaching for an object, forming stable grasp or contacts with it, and delivering it to the target location #REFR .", "text_before_citation": ["This aspect serves as a key differentiator between human-human and human-robot workers.", "In recent years, researchers have explored the use of verbal communication for coordination between mobile robots and humans, as evidenced by studies such as #OTHEREFR .", "One such application is in tour guiding, where verbal communication is primarily used to provide the necessary information to the human. Additionally, Zhou et al.", "(2021) #OTHEREFR proposed a novel approach for training a reinforcement learning agent using natural language goal-conditioned adversarial learning algorithm and a variational goal generator to sample natural language goals from trajectories of given demonstrations.", "Coordinating task execution between a manipulating robot and a human using natural language processing (NLP) is a arXiv:2304.02993v1 [cs.RO] 6 Apr 2023 more challenging task than using NLP for mobile/service robots."], "text_after_citation": ["This paper introduces one of the first frameworks for verbally commanding a cobot to perform pick-and-place tasks. Pick-and-place is a common yet straightforward task to automate.", "The proposed framework comprises three key components: natural language processing, semantic segmentation, and autonomous grasping.", "The autonomous grasping module generates grasping poses for a given scene, allowing a human operator to command the robot to execute a desired grasp.", "The contribution of this paper is manifold: (i) We created a dictionary of the common tasks involved in robotic manipulation in our NLP component.", "(ii) Our framework includes a translator-middleware capable of translating natural language instructions into low-level joint space or operational space robot commands."], "citing_paper_content": {"title": "Natural Language Robot Programming: Nlp Integrated With Autonomous Robotic Grasping", "abstract": "In this paper, we present a grammar-based natural language framework for robot programming, specifically for pick-and-place tasks. Our approach uses a custom dictionary of action words, designed to store together words that share meaning, allowing for easy expansion of the vocabulary by adding more action words from a lexical database. We validate our Natural Language Robot Programming (NLRP) framework through simulation and real-world experimentation, using a Franka Panda robotic arm equipped with a calibrated camerain-hand and a microphone. Participants were asked to complete a pick-and-place task using verbal commands, which were converted into text using Google's Speech-to-Text API and processed through the NLRP framework to obtain joint space trajectories for the robot. Our results indicate that our approach has a high system usability score. The framework's dictionary can be easily extended without relying on transfer learning or large data sets. In the future, we plan to compare the presented framework with different approaches of humanassisted pick-and-place tasks via a comprehensive user study."}, "cited_paper_content": {"title": "Safe Robotic Grasping: Minimum Impact-Force Grasp Selection", "abstract": "This paper addresses the problem of selecting from a choice of possible grasps, so that impact forces will be minimised if a collision occurs while the robot is moving the grasped object along a post-grasp trajectory. Such considerations are important for safety in human-robot interaction, where even a certified\"human-safe\"(e.g. compliant) arm may become hazardous once it grasps and begins moving an object, which may have significant mass, sharp edges or other dangers. Additionally, minimising collision forces is critical to preserving the longevity of robots which operate in uncertain and hazardous environments, e.g. robots deployed for nuclear decommissioning, where removing a damaged robot from a contaminated zone for repairs may be extremely difficult and costly. Also, unwanted collisions between a robot and critical infrastructure (e.g. pipework) in such high-consequence environments can be disastrous. In this paper, we investigate how the safety of the post-grasp motion can be considered during the pre-grasp approach phase, so that the selected grasp is optimal in terms applying minimum impact forces if a collision occurs during a desired post-grasp manipulation. We build on the methods of augmented robot-object dynamics models and\"effective mass\"and propose a method for combining these concepts with modern grasp and trajectory planners, to enable the robot to achieve a grasp which maximises the safety of the post-grasp trajectory, by minimising potential collision forces. We demonstrate the effectiveness of our approach through several experiments with both simulated and real robots."}, "keywords": ["mobile/service robots", "stable grasp"], "citation_intent": "background"}
{"citing_id": "2304.07063v1", "cited_id": "1806.07297", "section_title": "The Gap Between Tree-Form And Efo 1 Queries", "citation": "The gap is filled by considering those two properties with one assumption, which is a common practice #REFR .", "text_before_citation": ["Property 1 (Negation among Existential Variables).", "There is a negated atomic formula of the form \u00acr(x 1 , x 2 ) or \u00acr(x 1 , y) where x 1 , x 2 are existential variables and y is the free variable.", "Property 2 (Existential Leaves).", "If we expand the query graph by using the free variable as the root, then the leaf node of the query graph can be an existential variable rather than a constant entity."], "text_after_citation": ["Theorem 3.", "With assumption 1, any EFO 1 query with a simple acyclic query graph that does not have Property 1 and 2 is a TF query.", "The proof of Theorem 3 is detailed in Appendix A."], "citing_paper_content": {"title": "On Existential First Order Queries Inference On Knowledge Graphs", "abstract": "Reasoning on knowledge graphs is a challenging task because it utilizes observed information to predict the missing one. Specifically, answering first-order logic formulas is of particular interest because of its clear syntax and semantics. Recently, the query embedding method has been proposed which learns the embedding of a set of entities and treats logic operations as set operations. Though there has been much research following the same methodology, it lacks a systematic inspection from the standpoint of logic. In this paper, we characterize the scope of queries investigated previously and precisely identify the gap between it and the whole family of existential formulas. Moreover, we develop a new dataset containing ten new formulas and discuss the new challenges coming simultaneously. Finally, we propose a new search algorithm from fuzzy logic theory which is capable of solving new formulas and outperforming the previous methods in existing formulas. Preprint. Under review."}, "cited_paper_content": {"title": "Canonical Tensor Decomposition For Knowledge Base Completion", "abstract": "The problem of Knowledge Base Completion can be framed as a 3rd-order binary tensor completion problem. In this light, the Canonical Tensor Decomposition (CP) (Hitchcock, 1927) seems like a natural solution; however, current implementations of CP on standard Knowledge Base Completion benchmarks are lagging behind their competitors. In this work, we attempt to understand the limits of CP for knowledge base completion. First, we motivate and test a novel regularizer, based on tensor nuclear p-norms. Then, we present a refor-mulation of the problem that makes it invariant to arbitrary choices in the inclusion of predicates or their reciprocals in the dataset. These two methods combined allow us to beat the current state of the art on several datasets with a CP decomposition , and obtain even better results using the more advanced ComplEx model."}, "keywords": ["one assumption"], "citation_intent": "background"}
{"citing_id": "2303.04212v1", "cited_id": "1011.0686", "section_title": "V. Conclusions", "citation": "One solution is to follow a paradigm like DAGGER #REFR to keep exploring the environment and collect new states and labels to update our CBF critics and world model.", "text_before_citation": ["We apply our method to two simulated domains and show that our method outperforms existing classical and learningbased safe navigation approaches.", "Furthermore, we show that ConBaT can be quickly adapted to new safety constraints from limited demonstrations.", "This paradigm of learning safety constraints implicitly from minimal labels reduces training effort in creating safe agents and makes it easier to adapt to new definitions of safety.", "Though empirically ConBaT can learn safer policies and quickly adapts to new safety concepts, there are some limitations.", "Our method might not guarantee safety if ConBaT makes wrong CBF score predictions due to out-of-distribution data, or if the online action optimization falls into a local minimum."], "text_after_citation": ["The runtime for ConBaT in evaluation is also 0.2 \u223c 1.0x higher than other learning-based methods due to the back-propagation step in online optimization. We discuss different ConBaT architectures in Sec.", "III-E to weigh the trade-off between the safety performance and the runtime.", "Further runtime improvement can be made by faster back-propagation algorithms optimized for GPU hardware. Currently, our framework was only applied in simulated environments.", "In the future, we wish to extend our approach to image-based control in 3D and toward more complex robotic platforms in simulation and the real world."], "citing_paper_content": {"title": "Conbat: Control Barrier Transformer For Safe Policy Learning", "abstract": "Large-scale self-supervised models have recently revolutionized our ability to perform a variety of tasks within the vision and language domains. However, using such models for autonomous systems is challenging because of safety requirements: besides executing correct actions, an autonomous agent must also avoid the high cost and potentially fatal critical mistakes. Traditionally, self-supervised training mainly focuses on imitating previously observed behaviors, and the training demonstrations carry no notion of which behaviors should be explicitly avoided. In this work, we propose Control Barrier Transformer (ConBaT), an approach that learns safe behaviors from demonstrations in a self-supervised fashion. ConBaT is inspired by the concept of control barrier functions in control theory and uses a causal transformer that learns to predict safe robot actions autoregressively using a critic that requires minimal safety data labeling. During deployment, we employ a lightweight online optimization to find actions that ensure future states lie within the learned safe set. We apply our approach to different simulated control tasks and show that our method results in safer control policies compared to other classical and learning-based methods such as imitation learning, reinforcement learning, and model predictive control."}, "cited_paper_content": {"title": "A Reduction Of Imitation Learning And Structured Prediction To No-Regret Online Learning", "abstract": "Sequential prediction problems such as imitation learning, where future observations depend on previous predictions (actions), violate the common i.i.d. assumptions made in statistical learning. This leads to poor performance in theory and often in practice. Some recent approaches provide stronger guarantees in this setting, but remain somewhat unsatisfactory as they train either non-stationary or stochastic policies and require a large number of iterations. In this paper, we propose a new iterative algorithm, which trains a stationary deterministic policy, that can be seen as a no regret algorithm in an online learning setting. We show that any such no regret algorithm, combined with additional reduction assumptions, must find a policy with good performance under the distribution of observations it induces in such sequential settings. We demonstrate that this new approach outperforms previous approaches on two challenging imitation learning problems and a benchmark sequence labeling problem."}, "keywords": ["labels"], "citation_intent": "method"}
{"citing_id": "2303.11242v1", "cited_id": "1512.03385", "section_title": "Discussion For Dp With Sam In Fl", "citation": "These experiments are conducted on CIFAR-10 with ResNet-18 #REFR and Dirichlet \u03b1 = 0.6. The norm of local update.", "text_before_citation": ["In this subsection, we empirically discuss how SAM mitigates the negative impacts of DP from the aspects of the norm of local update and the visualization of loss landscape and contour.", "Meanwhile, we also observe the training performance with SAM under different privacy budgets \u03f5 compared with all baselines."], "text_after_citation": ["To validate the theoretical results for mitigating the adverse impacts of the norm of local updates, we conduct experiments on DP-FedSAM and DP-FedAvg with clipping threshold C = 0.2 as shown in Figure  3 .", "We show the norm \u2206 t i distribution and average norm \u2206 t of local updates before clipping during the communication rounds.", "In contrast to DP-FedAvg, most of the norm is distributed at the smaller value in our scheme in Figure 3 (a) , which means that the clipping operation drops less information.", "Meanwhile, the on-average norm \u2206 t is smaller than DP-FedAvg as shown in Figure 3 (b) .", "These observations are also consistent with our theoretical results in Section 5. Performance under different privacy budgets \u03f5."], "citing_paper_content": {"title": "Make Landscape Flatter In Differentially Private Federated Learning", "abstract": "To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), clientlevel Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with R\u00e9nyi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["norm", "ResNet-18"], "citation_intent": "method"}
{"citing_id": "2303.13123v1", "cited_id": "1703.04977", "section_title": "How Do Aleatoric And Epistemic Uncertainty Dif-", "citation": "This is not the case for models with Dropout as an epistemic component, confirming the findings by #REFR .", "text_before_citation": ["What does this tell us? We see that the aleatoric and epistemic components of the LSN clearly capture different aspects of the model's uncertainty.", "To better understand where this difference comes from, Figure 5 shows epistemic and aleatoric variance maps of sigmoids for a given ISIC test set sample.", "While the epistemic component assigns low variance to this in-distribution sample, the aleatoric component expresses the variation of plausible predictions for the image.", "We find this behavior attractive: In the Bayesian framework, the aleatoric component should play the role of the predictor, as it imitates the distribution of annotations Figure 4 : Precision-Recall plots for the ISIC test dataset show that aleatoric and epistemic variance in the LSN models (first row) are not the same, opposed to the SSN + Dropout models.", "The better performance of the aleatoric component (higher AUC) indicates that the model learns to separate data variation from model uncertainty."], "text_after_citation": [], "citing_paper_content": {"title": "Laplacian Segmentation Networks: Improved Epistemic Uncertainty From Spatial Aleatoric Uncertainty", "abstract": "Out of distribution (OOD) medical images are frequently encountered, e.g. because of site-or scanner differences, or image corruption. OOD images come with a risk of incorrect image segmentation, potentially negatively affecting downstream diagnoses or treatment. To ensure robustness to such incorrect segmentations, we propose Laplacian Segmentation Networks (LSN) that jointly model epistemic (model) and aleatoric (data) uncertainty in image segmentation. We capture data uncertainty with a spatially correlated logit distribution. For model uncertainty, we propose the first Laplace approximation of the weight posterior that scales to large neural networks with skip connections that have high-dimensional outputs. Empirically, we demonstrate that modelling spatial pixel correlation allows the Laplacian Segmentation Network to successfully assign high epistemic uncertainty to out-of-distribution objects appearing within images."}, "cited_paper_content": {"title": "What Uncertainties Do We Need In Bayesian Deep Learning For Computer Vision?", "abstract": "There are two major types of uncertainty one can model. Aleatoric uncertainty captures noise inherent in the observations. On the other hand, epistemic uncertainty accounts for uncertainty in the model - uncertainty which can be explained away given enough data. Traditionally it has been difficult to model epistemic uncertainty in computer vision, but with new Bayesian deep learning tools this is now possible. We study the benefits of modeling epistemic vs. aleatoric uncertainty in Bayesian deep learning models for vision tasks. For this we present a Bayesian deep learning framework combining input-dependent aleatoric uncertainty together with epistemic uncertainty. We study models under the framework with per-pixel semantic segmentation and depth regression tasks. Further, our explicit uncertainty formulation leads to new loss functions for these tasks, which can be interpreted as learned attenuation. This makes the loss more robust to noisy data, also giving new state-of-the-art results on segmentation and depth regression benchmarks."}, "keywords": ["Dropout", "epistemic component"], "citation_intent": "result"}
{"citing_id": "2303.09716v1", "cited_id": "1602.01783", "section_title": "A. Introduction", "citation": "Multi-agent reinforcement learning algorithms have contributed to many successes in machine learning, including games such as chess and Go Silver et al. (2016) ; #REFR ; Silver et al.", "text_before_citation": [], "text_after_citation": ["(2017b,a) , automated warehouses, robotic arms with multiple arms #OTHEREFR , and autonomous traffic control #OTHEREFR ; #OTHEREFR ; see #OTHEREFR ; ; for surveys.", "Multi-agent RL can refer to one of many scenarios: (i) where a team of agents work towards a common goal with all the agents have the same information or not #OTHEREFR", "(2022) , (ii) non-cooperative games where there are multiple agents with their own objectives #OTHEREFR", "(2019) , and (iii) zero-sum games where there are only two players with opposing objectives.", "We can further categorize problems as infinite-horizon, discounted reward/cost, finite-horizon, simultaneous move or turn-based games."], "citing_paper_content": {"title": "A New Policy Iteration Algorithm For Reinforcement Learning In Zero-Sum Markov Games", "abstract": "Many model-based reinforcement learning (RL) algorithms can be viewed as having two phases that are iteratively implemented: a learning phase where the model is approximately learned and a planning phase where the learned model is used to derive a policy. In the case of standard MDPs, the learning problem can be solved using either value iteration or policy iteration. However, in the case of zero-sum Markov games, there is no efficient policy iteration algorithm; e.g., it has been shown in Hansen et al. (2013) that one has to solve \u2126(1/(1 \u2212 \u03b1)) MDPs to implement the only known convergent version of policy iteration. Another algorithm for Markov zero-sum games, called naive policy iteration, is easy to implement but is only provably convergent under very restrictive assumptions. Prior attempts to fix naive policy iteration algorithm have several limitations. Here, we show that a simple variant of naive policy iteration for games converges, and converges exponentially fast. The only addition we propose to naive policy iteration is the use of lookahead in the policy improvement phase. This is appealing because lookahead is anyway often used in RL for games. We further show that lookahead can be implemented efficiently in linear Markov games, which are the counterpart of the linear MDPs and have been the subject of much attention recently. We then consider multi-agent reinforcement learning which uses our algorithm in the planning phases, and provide sample and time complexity bounds for such an algorithm."}, "cited_paper_content": {"title": "Asynchronous Methods For Deep Reinforcement Learning", "abstract": "We propose a conceptually simple and lightweight framework for deep reinforcement learning that uses asynchronous gradient descent for optimization of deep neural network controllers. We present asynchronous variants of four standard reinforcement learning algorithms and show that parallel actor-learners have a stabilizing effect on training allowing all four methods to successfully train neural network controllers. The best performing method, an asynchronous variant of actor-critic, surpasses the current state-of-the-art on the Atari domain while training for half the time on a single multi-core CPU instead of a GPU. Furthermore, we show that asynchronous actor-critic succeeds on a wide variety of continuous motor control problems as well as on a new task of navigating random 3D mazes using a visual input."}, "keywords": ["Multi-agent reinforcement learning"], "citation_intent": "background"}
{"citing_id": "2304.07537v1", "cited_id": "1602.05629", "section_title": "Fedxgbllr", "citation": "For the federated training of the one-layer 1D CNN after round 1 (line 8), the protocol follows the standard FL algorithm, and we use FedAvg #REFR .", "text_before_citation": ["The global federated XGBoost model consists of all clients' locally trained XGBoost tree ensembles and the globally trained one-layer 1D CNN. The detailed procedure is shown in Algorithm. 1.", "At round 0 (line 1 to 7), each client first trains its local XGBoost tree ensemble.", "The server then conducts tree ensemble aggregation and CNN initialization.", "After receiving the aggregated tree ensemble, all clients calculate the prediction outcomes given the aggregated tree ensemble on their local data samples. The calculated prediction outcomes are inputs of the CNN.", "It is worth noticing that the clients only build XGBoost models at round 0, and the aggregated tree ensemble is fixed after round 0."], "text_after_citation": ["In FedXGBllr, the number of communication rounds is equal to the FL training rounds ( ) because we send the trees (at round 0) and CNN's model parameters (after round 1)."], "citing_paper_content": {"title": "Gradient-Less Federated Gradient Boosting Trees With Learnable Learning Rates", "abstract": "The privacy-sensitive nature of decentralized datasets and the robustness of eXtreme Gradient Boosting (XGBoost) on tabular data raise the needs to train XGBoost in the context of federated learning (FL). Existing works on federated XGBoost in the horizontal setting rely on the sharing of gradients, which induce per-node level communication frequency and serious privacy concerns. To alleviate these problems, we develop an innovative framework for horizontal federated XGBoost which does not depend on the sharing of gradients and simultaneously boosts privacy and communication efficiency by making the learning rates of the aggregated tree ensembles learnable. We conduct extensive evaluations on various classification and regression datasets, showing our approach achieves performance comparable to the stateof-the-art method and effectively improves communication efficiency by lowering both communication rounds and communication overhead by factors ranging from 25x to 700x."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["federated training"], "citation_intent": "method"}
{"citing_id": "2303.05499v1", "cited_id": "1707.07998", "section_title": "Introduction", "citation": "In this work, we aim to develop a strong system to detect arbitrary objects specified by human language inputs, which we name as open-set object detection #REFR .", "text_before_citation": ["Understanding novel concepts is a fundamental capability of visual intelligence."], "text_after_citation": ["The task has wide applications for its great potential as a generic object detector.", "For example, we can cooperate it with generative models for image editing (as shown in Fig. 1 (b) ).", "The key to open-set detection is introducing language for unseen object generalization #OTHEREFR .", "For example, GLIP #OTHEREFR reformulates object detection as a phrase grounding task and introduces contrastive training between object regions and language phrases.", "It shows a great flexibility for heterogeneous datasets and remarkable performance on both closed-set and open-set detection."], "citing_paper_content": {"title": "Grounding Dino: Marrying Dino With Grounded Pre-Training For Open-Set Object Detection", "abstract": "Standard Object Detection COCO pre-defined categories Zero-Shot Transfer to Novel Categories worldcup Human-input novel categories ear, lion, bench The left lion The bottom man with his head up Referring Object Detection (Referring Expression Comprehension) Human-input reference sentences bench person (c) Application: Image Editing Collaborate with stable diffusion. Prompt (modify background): All people around the world cheer with a worldcup."}, "cited_paper_content": {"title": "Bottom-Up And Top-Down Attention For Image Captioning And Visual Question Answering", "abstract": "Top-down visual attention mechanisms have been used extensively in image captioning and visual question answering (VQA) to enable deeper image understanding through fine-grained analysis and even multiple steps of reasoning. In this work, we propose a combined bottom-up and top-down attention mechanism that enables attention to be calculated at the level of objects and other salient image regions. This is the natural basis for attention to be considered. Within our approach, the bottom-up mechanism (based on Faster R-CNN) proposes image regions, each with an associated feature vector, while the top-down mechanism determines feature weightings. Applying this approach to image captioning, our results on the MSCOCO test server establish a new state-of-the-art for the task, achieving CIDEr / SPICE / BLEU-4 scores of 117.9, 21.5 and 36.9, respectively. Demonstrating the broad applicability of the method, applying the same approach to VQA we obtain first place in the 2017 VQA Challenge."}, "keywords": ["open-set object detection", "human language inputs"], "citation_intent": "method"}
{"citing_id": "2303.10067v1", "cited_id": "1702.02287", "section_title": "Results", "citation": "We picked this name for comparison because of two reasons; 1) the number of authors sharing this name is among the top five as shown in Table 3 and 2) All methods cited in #REFR could not achieve a good result.", "text_before_citation": ["For example, 1) two authors with the same common name authoring a single publication.", "2) more than one author with the same common atomic name variate authoring a single publication, 3) number of authors with the same full name, 4) the uncertainty of the accuracy of the dataset, etc.", "Although the comparison is difficult and cannot be completely fair, we compare WhoIs to other state-of-the-art approaches, whose results are reported in #OTHEREFR .", "These results are obtained on a collection from CiteSeerX 16 that contains records of authors with the name / atomic name variate 'Y Chen'.", "This collection consists of 848 complete documents authored by 71 distinct authors."], "text_after_citation": ["We applied WhoIs on this collection by randomly splitting the records into 70% for training, 15% for validation and 15% for testing. The results are shown in Table 5 .", "Note that in our collection, we consider way more records and distinct authors (see Table 3 ) and we use only reference attributes (i.e. co-authors, title and source).", "As the results presented in Table 5 show, WhoIs outperforms other methods in resolving the disambiguation of the author name 'Y Chen' on the CiteSeerX dataset, which is a relatively small dataset and does not really reflect the performance of all presented approaches in real scenarios.", "The disparity between the results shown in Table 4 and Table 5 demonstrates that the existing benchmark datasets are manually prepared for the sake of accuracy.", "However, this leads to covering a very small portion of records whose authors share similar Table 5 Comparison between WhoIs and other baseline methods on CiteSeerX dataset in terms of Macro F1 score as reported in #OTHEREFR ."], "citing_paper_content": {"title": "Deep Author Name Disambiguation Using Dblp Data", "abstract": "In the academic world, the number of scientists grows every year and so does the number of authors sharing the same names. Consequently, it challenging to assign newly published papers to their respective authors. Therefore, Author Name Ambiguity (ANA) is considered a critical open problem in digital libraries. This paper proposes an Author Name Disambiguation (AND) approach that links author names to their real-world entities by leveraging their co-authors and domain of research. To this end, we use data collected from the DBLP repository that contains more than 5 million bibliographic records authored by around 2.6 million co-authors. Our approach first groups authors who share the same last names and same first name initials. The author within each group is identified by capturing the relation with his/her co-authors and area of research, represented by the titles of the validated publications of the corresponding author. To this end, we train a neural network model that learns from the representations of the co-authors and titles. We validated the effectiveness of our approach by conducting extensive experiments on a large dataset."}, "cited_paper_content": {"title": "Name Disambiguation In Anonymized Graphs Using Network Embedding", "abstract": "In real-world, our DNA is unique but many people share names. This phenomenon often causes erroneous aggregation of documents of multiple persons who are namesake of one another. Such mistakes deteriorate the performance of document retrieval, web search, and more seriously, cause improper attribution of credit or blame in digital forensic. To resolve this issue, the name disambiguation task is designed which aims to partition the documents associated with a name reference such that each partition contains documents pertaining to a unique real-life person. Existing solutions to this task substantially rely on feature engineering, such as biographical feature extraction, or construction of auxiliary features from Wikipedia. However, for many scenarios, such features may be costly to obtain or unavailable due to the risk of privacy violation. In this work, we propose a novel name disambiguation method. Our proposed method is non-intrusive of privacy because instead of using attributes pertaining to a real-life person, our method leverages only relational data in the form of anonymized graphs. In the methodological aspect, the proposed method uses a novel representation learning model to embed each document in a low dimensional vector space where name disambiguation can be solved by a hierarchical agglomerative clustering algorithm. Our experimental results demonstrate that the proposed method is significantly better than the existing name disambiguation methods working in a similar setting."}, "keywords": ["authors", "name"], "citation_intent": "method"}
{"citing_id": "2304.09176v1", "cited_id": "1703.02091", "section_title": "C. Enhancing Personalized Ranking Via Gauc Optimization", "citation": "In practice, we find that the GAUC #REFR metric is more consistent with the online effect.", "text_before_citation": ["The above section details how to construct paired samples within a batch, which do not meet the requirement for personalized recommendation."], "text_after_citation": ["Consequently, it is a natural idea to add the GAUC metric to the objective when optimizing the model.", "Considering the original calculation of the GAUC metric, the samples are firstly divided into several groups.", "In this context, the groups are partitioned by the user ID.", "Then the AUC metric is calculated within each group respectively, and the GAUC metric is derived by averaging the AUC metrics from all groups with weight.", "The weight is optionally proportional to the impression or click times, here we set the weight to be 1 for all users."], "citing_paper_content": {"title": "Enhancing Personalized Ranking With Differentiable Group Auc Optimization", "abstract": "AUC is a common metric for evaluating the performance of a classifier. However, most classifiers are trained with cross entropy, and it does not optimize the AUC metric directly, which leaves a gap between the training and evaluation stage. In this paper, we propose the PDAOM loss, a Personalized and Differentiable AUC Optimization method with Maximum violation, which can be directly applied when training a binary classifier and optimized with gradient-based methods. Specifically, we construct the pairwise exponential loss with difficult pair of positive and negative samples within sub-batches grouped by user ID, aiming to guide the classifier to pay attention to the relation between hard-distinguished pairs of opposite samples from the perspective of independent users. Compared to the origin form of pairwise exponential loss, the proposed PDAOM loss not only improves the AUC and GAUC metrics in the offline evaluation, but also reduces the computation complexity of the training objective. Furthermore, online evaluation of the PDAOM loss on the 'Guess What You Like' feed recommendation application in Meituan manifests 1.40% increase in click count and 0.65% increase in order count compared to the baseline model, which is a significant improvement in this well-developed online life service recommendation system."}, "cited_paper_content": {"title": "Optimized Cost Per Click In Taobao Display Advertising", "abstract": "Taobao, as the largest online retail platform in the world, provides billions of online display advertising impressions for millions of advertisers every day. For commercial purposes, the advertisers bid for specific spots and target crowds to compete for business traffic. The platform chooses the most suitable ads to display in tens of milliseconds. Common pricing methods include cost per mille (CPM) and cost per click (CPC). Traditional advertising systems target certain traits of users and ad placements with fixed bids, essentially regarded as coarse-grained matching of bid and traffic quality. However, the fixed bids set by the advertisers competing for different quality requests cannot fully optimize the advertisers' key requirements. Moreover, the platform has to be responsible for the business revenue and user experience. Thus, we proposed a bid optimizing strategy called optimized cost per click (OCPC) which automatically adjusts the bid to achieve finer matching of bid and traffic quality of page view (PV) request granularity. Our approach optimizes advertisers' demands, platform business revenue and user experience and as a whole improves traffic allocation efficiency. We have validated our approach in Taobao display advertising system in production. The online A/B test shows our algorithm yields substantially better results than previous fixed bid manner."}, "keywords": ["online effect"], "citation_intent": "background"}
{"citing_id": "2304.00202v1", "cited_id": "2002.11569", "section_title": "Results On Cifar-10.", "citation": "Compared with state-of-the-art multi-step PGD-AT #REFR , the proposed FGSM-PGK can obtain better adversarial robustness under all attack scenarios with less training cost.", "text_before_citation": ["We use ResNet18 as the backbone.", "The comparison experiment results on CIFAR-10 are shown in Table 2 .", "It can be observed that the proposed FGSM-PGK not only prevents catastrophic overfitting but also achieves excellent adversarial robustness performance improvement."], "text_after_citation": ["In particular, under the strongest attack method (AA), PGD-AT achieves the performance of about 48.7%, but the proposed FGSM-PGK achieves the performance of about 49.5%, which indicates that single-step adversarial training with priorguided knowledge has the potential to outperform multistep adversarial training.", "More importantly, our FGSM-PGK is about 3.6 times faster than PGD-AT.", "Compared with previous fast adversarial training methods, our FGSM-PGK obtains the best adversarial robustness under all attack scenarios on the best and last checkpoints.", "For example, under the PGD-50 attack, the most robust one among the previous fast adversarial training methods achieves an accuracy of about 53%, but our FGSM-PGK achieves an accuracy of about 55%.", "In terms of training efficiency, our proposed FGSM-PGK only takes 1.2 hours to train, which is faster than previous fast adversarial training methods."], "citing_paper_content": {"title": "Improving Fast Adversarial Training With Prior-Guided Knowledge", "abstract": "Fast adversarial training (FAT) is an efficient method to improve robustness. However, the original FAT suffers from catastrophic overfitting, which dramatically and suddenly reduces robustness after a few training epochs. Although various FAT variants have been proposed to prevent overfitting, they require high training costs. In this paper, we investigate the relationship between adversarial example quality and catastrophic overfitting by comparing the training processes of standard adversarial training and FAT. We find that catastrophic overfitting occurs when the attack success rate of adversarial examples becomes worse. Based on this observation, we propose a positive prior-guided adversarial initialization to prevent overfitting by improving adversarial example quality without extra training costs. This initialization is generated by using high-quality adversarial perturbations from the historical training process. We provide theoretical analysis for the proposed initialization and propose a prior-guided regularization method that boosts the smoothness of the loss function. Additionally, we design a prior-guided ensemble FAT method that averages the different model weights of historical models using different decay rates. Our proposed method, called FGSM-PGK, assembles the prior-guided knowledge, i.e., the prior-guided initialization and model weights, acquired during the historical training process. Evaluations of four datasets demonstrate the superiority of the proposed method."}, "cited_paper_content": {"title": "Overfitting In Adversarially Robust Deep Learning", "abstract": "It is common practice in deep learning to use overparameterized networks and train for as long as possible; there are numerous studies that show, both theoretically and empirically, that such practices surprisingly do not unduly harm the generalization performance of the classifier. In this paper, we empirically study this phenomenon in the setting of adversarially trained deep networks, which are trained to minimize the loss under worst-case adversarial perturbations. We find that overfitting to the training set does in fact harm robust performance to a very large degree in adversarially robust training across multiple datasets (SVHN, CIFAR-10, CIFAR-100, and ImageNet) and perturbation models ($\\ell_\\infty$ and $\\ell_2$). Based upon this observed effect, we show that the performance gains of virtually all recent algorithmic improvements upon adversarial training can be matched by simply using early stopping. We also show that effects such as the double descent curve do still occur in adversarially trained models, yet fail to explain the observed overfitting. Finally, we study several classical and modern deep learning remedies for overfitting, including regularization and data augmentation, and find that no approach in isolation improves significantly upon the gains achieved by early stopping. All code for reproducing the experiments as well as pretrained model weights and training logs can be found at this https URL."}, "keywords": ["better adversarial robustness"], "citation_intent": "method"}
{"citing_id": "2303.01074v1", "cited_id": "1906.11110", "section_title": "Background", "citation": "We refer the reader to #REFR for a more formal treatment of these games, which we omit for brevity.", "text_before_citation": ["= \u2211\ufe01 =1 ( \u22121 , ).", "We say the algorithm is a regret minimizer, if the external regret", "ext = max \u2208 {1,..., } =1 ( , ) grows sublinearly in ,", "where is a Kronecker delta.", "We will use the online algorithm in two-player zero-sum matrix games and sequential imperfect-information games."], "text_after_citation": ["We employ the online algorithms at each state a player \u2208 {1, 2} can't distinguish (i.e.", "at each infostate), as in prior work on counterfactual regret minimization #OTHEREFR .", "In a sequential game, the players take turns making decisions at each infostate, until they arrive at a terminal state where the player receives its reward ( ) \u2208 R.", "We denote the maximum difference in rewards as \u0394 max = max ( )\u2212min ( ). The expected reward (in the whole game) is", "( ) = E \u223c ( ), where = ( 1 , 2 )"], "citing_paper_content": {"title": "Learning Not To Regret", "abstract": "Regret minimization is a key component of many algorithms for finding Nash equilibria in imperfect-information games. To scale to games that cannot fit in memory, we can use search with value functions. However, calling the value functions repeatedly in search can be expensive. Therefore, it is desirable to minimize regret in the search tree as fast as possible. We propose to accelerate the regret minimization by introducing a general \"learning not to regret\" framework, where we meta-learn the regret minimizer. The resulting algorithm is guaranteed to minimize regret in arbitrary settings and is (meta)-learned to converge fast on a selected distribution of games. Our experiments show that meta-learned algorithms converge substantially faster than prior regret minimization algorithms."}, "cited_paper_content": {"title": "Rethinking Formal Models Of Partially Observable Multiagent Decision Making", "abstract": "Multiagent decision-making problems in partially observable environments are usually modeled as either extensive-form games (EFGs) within the game theory community or partially observable stochastic games (POSGs) within the reinforcement learning community. While most practical problems can be modeled in both formalisms, the communities using these models are mostly distinct with little sharing of ideas or advances. The last decade has seen dramatic progress in algorithms for EFGs, mainly driven by the challenge problem of poker. We have seen computational techniques achieving super-human performance, some variants of poker are essentially solved, and there are now sound local search algorithms which were previously thought impossible. While the advances have garnered attention, the fundamental advances are not yet understood outside the EFG community. This can be largely explained by the starkly different formalisms between the game theory and reinforcement learning communities and, further, by the unsuitability of the original EFG formalism to make the ideas simple and clear. This paper aims to address these hindrances, by advocating a new unifying formalism, a variant of POSGs, which we call Factored-Observation Games (FOGs). We prove that any timeable perfect-recall EFG can be efficiently modeled as a FOG as well as relating FOGs to other existing formalisms. Additionally, a FOG explicitly identifies the public and private components of observations, which is fundamental to the recent EFG breakthroughs. We conclude by presenting the two building-blocks of these breakthroughs --- counterfactual regret minimization and public state decomposition --- in the new formalism, illustrating our goal of a simpler path for sharing recent advances between game theory and reinforcement learning community."}, "keywords": ["games"], "citation_intent": "background"}
{"citing_id": "2303.15724v1", "cited_id": "1706.03762", "section_title": "Sdm-Unips", "citation": "Given aggregated features at different pixels in the same sample set, we apply another na\u00efve Transformer #REFR to perform non-local interactions.", "text_before_citation": ["To address these issues, we draw inspiration from recent Transformers on 3-D points #OTHEREFR and apply a Transformer on a fixed number (m) of pixel samples (e.g., m = 2048) from random locations in the input coordinate system. We term this the pixel-sampling Transformer.", "Unlike image-based approaches, pixel-sampling Transformer's memory consumption is constant per sample set, scaling to arbitrary image sizes.", "Moreover, by applying the Transformer to a randomly sampled set of locations, local interactions that may lead to over-smoothing of feature maps (e.g., in CNNs) are almost entirely eliminated.", "Concretely, given m random pixels x i=1,...,m from the masked region of the input coordinate system, we interpolate features at those pixels as F 1,...,K (S(x i )), where S is the bilinear interpolation operator.", "Then, interpolated features are concatenated with corresponding raw observations I 1,...,K (x i ) and aggregated to A(x i ) with pooling by multihead attention (PMA) #OTHEREFR , as in #OTHEREFR ."], "text_after_citation": ["Since the goal of this process is to consider surface-level interactions based on physical attributes, pixel coordinate information is unnecessary.", "Thus, we don't apply position embeddings to samples, unlike most existing visual Transformer models (e.g.", "#OTHEREFR ), allowing the samples to propagate their aggregated features without location information.", "After the non-local interaction, we apply a two-layer MLP to predict surface normals at sampled locations.", "Finally, surface normals for each set are merged to obtain the final surface normal map at the input image resolution."], "citing_paper_content": {"title": "Scalable, Detailed And Mask-Free Universal Photometric Stereo", "abstract": "Input image (1st out of 9) UniPS [Ikehata2022] 3D Scanner (EinScan-SE) INPUT: Multiple images under varying unknown lighting conditions 1st 2nd 23rd Figure 1. Given multiple images under unknown spatially-varying illuminations, our method can recover the detailed surface normal map of non-convex, non-Lambertian surfaces (Left). Our method even surpasses the level of detail provided by consumer 3-D scanners (Right)."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["aggregated features", "another na\u00efve Transformer"], "citation_intent": "method"}
{"citing_id": "2304.08673v1", "cited_id": "1803.11188", "section_title": "Conclusions", "citation": "In practice, as we do in this paper, we can use sliced Wasserstein distance #REFR to compute Equation 14, avoiding the min-max formulation by directly approximating the Wasserstein distance.", "text_before_citation": ["Thus, if \u03b8 \u03c6 are the parameters of \u03c6, we can re-write the IPM loss in Equation 14as follows", "EQUATION", "This gives us the following min \u2212 max algorithm to solve Equation (6) \u03b8NF = arg min", "EQUATION", "where the loss function is given by Equation (8)."], "text_after_citation": ["We now discuss our framework in the general case when source and target spaces have different dimensions or are high dimensional.", "In such a case, we can project the two data spaces on to a latent space of the same dimension and consider the diffeomorphic pushforward between the latent spaces in exactly the same fashion we have done so far.", "Let \u2126 s \u2282 R ds and \u2126 t \u2282 R dt . Consider the latent spaces", "\u2126 L,s , \u2126 L,t \u2282 R d where d \u2264 min(d s , d t ).", "Associated with each latent space is a pair of encoder and decoder maps, that is enc s : \u2126 s \u2192 \u2126 L,s and enc t : \u2126 t \u2192 \u2126 L,t respectively (with the corresponding decoder maps)."], "citing_paper_content": {"title": "Semi-Supervised Learning Of Pushforwards For Domain Translation & Adaptation *", "abstract": "Given two probability densities on related data spaces, we seek a map pushing one density to the other while satisfying application-dependent constraints. For maps to have utility in a broad application space (including domain translation, domain adaptation, and generative modeling), the map must be available to apply on out-of-sample data points and should correspond to a probabilistic model over the two spaces. Unfortunately, existing approaches, which are primarily based on optimal transport, do not address these needs. In this paper, we introduce a novel pushforward map learning algorithm that utilizes normalizing flows to parameterize the map. We first re-formulate the classical optimal transport problem to be map-focused and propose a learning algorithm to select from all possible maps under the constraint that the map minimizes a probability distance and application-specific regularizers; thus, our method can be seen as solving a modified optimal transport problem. Once the map is learned, it can be used to map samples from a source domain to a target domain. In addition, because the map is parameterized as a composition of normalizing flows, it models the empirical distributions over the two data spaces and allows both sampling and likelihood evaluation for both data sets. We compare our method (parOT) to related optimal transport approaches in the context of domain adaptation and domain translation on benchmark data sets. Finally, to illustrate the impact of our work on applied problems, we apply parOT to a real scientific application: spectral calibration for high-dimensional measurements from two vastly different environments."}, "cited_paper_content": {"title": "Generative Modeling Using The Sliced Wasserstein Distance", "abstract": "Generative Adversarial Nets (GANs) are very successful at modeling distributions from given samples, even in the high-dimensional case. However, their formulation is also known to be hard to optimize and often not stable. While this is particularly true for early GAN formulations, there has been significant empirically motivated and theoretically founded progress to improve stability, for instance, by using the Wasserstein distance rather than the Jenson-Shannon divergence. Here, we consider an alternative formulation for generative modeling based on random projections which, in its simplest form, results in a single objective rather than a saddle-point formulation. By augmenting this approach with a discriminator we improve its accuracy. We found our approach to be significantly more stable compared to even the improved Wasserstein GAN. Further, unlike the traditional GAN loss, the loss formulated in our method is a good measure of the actual distance between the distributions and, for the first time for GAN training, we are able to show estimates for the same."}, "keywords": ["Wasserstein distance", "sliced Wasserstein distance"], "citation_intent": "method"}
{"citing_id": "2303.02622v1", "cited_id": "1503.02531", "section_title": "Continual Learning Based On Expansion", "citation": "After that, the compression is done using the knowledge distillation approach #REFR and soft labels; thus, the network reshapes to its original architecture.", "text_before_citation": ["In this method, for expanding the network, an LSTM network is used via reinforcement learning and policy gradient to predict the optimal number of nodes that should be added to each layer, with respect to both the detection rate and size of the network.", "Although the approaches mentioned above try to expand the network as efficiently as possible, the network's size will still grow after each task, which is considered a drawback in the long run.", "An approach for fully compressing the network after the expansion is proposed in regularize, expand, and compress (REC) framework #OTHEREFR .", "Similar to RCL, REC exploits reinforcement learning (AutoML #OTHEREFR ) to expand the network.", "The whole network is trained on the new task with regularizations based on multi-task learning and the Fisher information matrix."], "text_after_citation": [], "citing_paper_content": {"title": "A Multi-Agent Adaptive Deep Learning Framework For Online Intrusion Detection", "abstract": "The network security analyzers use intrusion detection systems (IDSes) to distinguish malicious traffic from benign ones. The deep learning-based (DL-based) IDSes are proposed to auto-extract high-level features and eliminate the time-consuming and costly signature extraction process. However, this new generation of IDSes still suffers from a number of challenges. One of the main issues of an IDS is facing traffic concept drift which manifests itself as new (i.e., zero-day) attacks, in addition to the changing behavior of benign users/applications. Furthermore, a practical DL-based IDS needs to be conformed to a distributed architecture to handle big data challenges. In this paper, we propose a framework for adapting DL-based models to the changing attack/benign traffic behaviors, considering a more practical scenario (i.e., online adaptable IDSes). This framework employs continual deep anomaly detectors in addition to the federated learning approach to solve the above-mentioned challenges. Furthermore, the proposed framework implements sequential packet labeling for each flow, which provides an attack probability score for the flow by gradually observing each flow packet and updating its estimation. We evaluate the proposed framework by employing different deep models (including CNN-based and LSTM-based) over the CIC-IDS2017 and CSE-CIC-IDS2018 datasets. Through extensive evaluations and experiments, we show that the proposed distributed framework is well adapted to the traffic concept drift. More precisely, our results indicate that the CNN-based models are well suited for continually adapting to the traffic concept drift (i.e., achieving an average detection rate of above 95% while needing just 128 new flows for the updating phase), and the LSTM-based models are a good candidate for sequential packet labeling in practical online IDSes (i.e., detecting intrusions by just observing their first 15 packets)."}, "cited_paper_content": {"title": "Distilling The Knowledge In A Neural Network", "abstract": "A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel."}, "keywords": ["knowledge distillation approach"], "citation_intent": "method"}
{"citing_id": "2303.10251v1", "cited_id": "1904.06830", "section_title": "Training Data", "citation": "ContactDB #REFR ) is a collection of grasping and manipulation data for applications in human-robot interaction.", "text_before_citation": [], "text_after_citation": ["The data were collected by thermally imaging 3D printed household objects after they were grasped by human subjects.", "We study 8 objects, each handled by a single participant.", "The data take the form of triangle meshes with a single feature (intensity) at each vertex, indicating thermal energy transferred to the object.", "As in #OTHEREFR , we process the contact maps by passing the intensities through a sigmoid function, assigning 0.999 probability to the maximum intensity and 0.001 to the minimum.", "As in #OTHEREFR , we average the corresponding probabilities on each face to obtain an unnormalized probability distribution over faces."], "citing_paper_content": {"title": "Conformal Generative Modeling On Triangulated Surfaces", "abstract": "We propose conformal generative modeling, a framework for generative modeling on 2D surfaces approximated by discrete triangle meshes. Our approach leverages advances in discrete conformal geometry to develop a map from a source triangle mesh to a target triangle mesh of a simple manifold such as a sphere. After accounting for errors due to the mesh discretization, we can use any generative modeling approach developed for simple manifolds as a plug-and-play subroutine. We demonstrate our framework on multiple complicated manifolds and multiple generative modeling subroutines, where we show that our approach can learn good estimates of distributions on meshes from samples, and can also learn simultaneously from multiple distinct meshes of the same underlying manifold."}, "cited_paper_content": {"title": "Contactdb: Analyzing And Predicting Grasp Contact Via Thermal Imaging", "abstract": "Grasping and manipulating objects is an important human skill. Since hand-object contact is fundamental to grasping, capturing it can lead to important insights. However, observing contact through external sensors is challenging because of occlusion and the complexity of the human hand. We present ContactDB, a novel dataset of contact maps for household objects that captures the rich hand-object contact that occurs during grasping, enabled by use of a thermal camera. Participants in our study grasped 3D printed objects with a post-grasp functional intent. ContactDB includes 3750 3D meshes of 50 household objects textured with contact maps and 375K frames of synchronized RGB-D+thermal images. To the best of our knowledge, this is the first large-scale dataset that records detailed contact maps for human grasps. Analysis of this data shows the influence of functional intent and object size on grasping, the tendency to touch/avoid \u2018active areas\u2019, and the high frequency of palm and proximal finger contact. Finally, we train state-of-the art image translation and 3D convolution algorithms to predict diverse contact patterns from object shape. Data, code and models are available at https://contactdb.cc.gatech.edu."}, "keywords": ["grasping"], "citation_intent": "background"}
{"citing_id": "2303.07850v1", "cited_id": "1802.05098", "section_title": "Grid-Worlds", "citation": "LOLA do not achieve significant results in MH, and similar results also appear in #REFR .", "text_before_citation": ["4 is the evaluation results of ARSP and other baselines in MH.", "Global rewards are the summation of all agents' returns in the environment.", "In MH, ARSP agents can rapidly discover the high payoff NE where two agents stay together and wait for the monster.", "Notably, ARSP agents can stably converge to the coordination strategy in less than 25k time steps, which shows the superiority of ARSP in sample efficiency.", "However, most other baseline agents can only converge to some guaranteed but lower payoff NEs, i.e., avoid stag and eat apples alone."], "text_after_citation": ["We suspect that LOLA can not scale to high dimensional settings.", "R-IQN performs better than other baselines in MH and the second row of Fig 6 shows that R-IQN agents can achieve one or two co-operations in one episode, demonstrating that the risk-seeking policy plays an important role in discovering coordination strategies.", "Similar results are also achieved in the Escalation environment, where ARSP significantly outperforms other baselines both in asymptotic performance and sample efficiency.", "Although with the increase of cooperation times, the punishment suffered by agents for betrayal is also increasing in the Escalation, agents can still achieve stable cooperation shown in"], "citing_paper_content": {"title": "Learning Adaptable Risk-Sensitive Policies To Coordinate In Multi-Agent General-Sum Games", "abstract": "In general-sum games, the interaction of selfinterested learning agents commonly leads to socially worse outcomes, such as defect-defect in the iterated stag hunt (ISH). Previous works address this challenge by sharing rewards or shaping their opponents' learning process, which require too strong assumptions. In this paper, we demonstrate that agents trained to optimize expected returns are more likely to choose a safe action that leads to guaranteed but lower rewards. However, there typically exists a risky action that leads to higher rewards in the long run only if agents cooperate, e.g., cooperate-cooperate in ISH. To overcome this, we propose using action value distribution to characterize the decision's risk and corresponding potential payoffs. Specifically, we present Adaptable Risk-Sensitive Policy (ARSP). ARSP learns the distributions over agent's return and estimates a dynamic risk-seeking bonus to discover risky coordination strategies. Furthermore, to avoid overfitting training opponents, ARSP learns an auxiliary opponent modeling task to infer opponents' types and dynamically alter corresponding strategies during execution. Empirically, agents trained via ARSP can achieve stable coordination during training without accessing opponent's rewards or learning process, and can adapt to noncooperative opponents during execution. To the best of our knowledge, it is the first method to learn coordination strategies between agents both in iterated prisoner's dilemma (IPD) and iterated stag hunt (ISH) without shaping opponents or rewards, and can adapt to opponents with distinct strategies during execution. Furthermore, we show that ARSP can be scaled to high-dimensional settings."}, "cited_paper_content": {"title": "Dice: The Infinitely Differentiable Monte-Carlo Estimator", "abstract": "The score function estimator is widely used for estimating gradients of stochastic objectives in Stochastic Computation Graphs (SCG), eg. in reinforcement learning and meta-learning. While deriving the first-order gradient estimators by differentiating a surrogate loss (SL) objective is computationally and conceptually simple, using the same approach for higher-order gradients is more challenging. Firstly, analytically deriving and implementing such estimators is laborious and not compliant with automatic differentiation. Secondly, repeatedly applying SL to construct new objectives for each order gradient involves increasingly cumbersome graph manipulations. Lastly, to match the first-order gradient under differentiation, SL treats part of the cost as a fixed sample, which we show leads to missing and wrong terms for higher-order gradient estimators. To address all these shortcomings in a unified way, we introduce DiCE, which provides a single objective that can be differentiated repeatedly, generating correct gradient estimators of any order in SCGs. Unlike SL, DiCE relies on automatic differentiation for performing the requisite graph manipulations. We verify the correctness of DiCE both through a proof and through numerical evaluation of the DiCE gradient estimates. We also use DiCE to propose and evaluate a novel approach for multi-agent learning. Our code is available at this https URL"}, "keywords": ["MH", "similar results"], "citation_intent": "result"}
{"citing_id": "2303.14084v1", "cited_id": "1711.06940", "section_title": "Our Contributions", "citation": "Our bounds are comparable to those for non-private synthetic control (e.g., #REFR ), and in Section 5.1.1, we explicitly show that the cost of privacy in synthetic control is O(1/ ).", "text_before_citation": ["For privacy (Theorems 3.1 and 3.4), although our algorithmic techniques rely on existing approaches, prior results on privacy do not apply in our setting.", "DP methods add noise that scales with the sensitivity of the function being computed, which is defined as the maximum change in the function's output that can be caused by changing a single donor's data.", "However, synthetic control performs a regression in a vertical way, treating each time point, rather than each donor's data, as one sample -thus, the transposed setting changes the definition of neighboring databases, completely altering the impact of a single donor's data.", "The majority of our privacy analysis is devoted to computing sensitivity of this new method.", "We also provide accuracy guarantees for each algorithm (Theorems 3.2 and 3.5), bounding the root mean squared error (RMSE) of the algorithm's output compared to the post-intervention target signal."], "text_after_citation": ["That is, the RMSE of our algorithm relative to a non-private version is only greater by a factor of O(1/ ), which is unavoidable in most analysis tasks.", "To better interpret our bounds in terms of natural problem parameters such as number of samples and length of observations, we also provide Corollaries 5.5 and 7.2, which give explicit closed-form upper bounds on the RMSE under mild assumptions on the underlying data distribution."], "citing_paper_content": {"title": "Differentially Private Synthetic Control", "abstract": "Synthetic control is a causal inference tool used to estimate the treatment effects of an intervention by creating synthetic counterfactual data. This approach combines measurements from other similar observations (i.e., donor pool) to predict a counterfactual time series of interest (i.e., target unit) by analyzing the relationship between the target and the donor pool before the intervention. As synthetic control tools are increasingly applied to sensitive or proprietary data, formal privacy protections are often required. In this work, we provide the first algorithms for differentially private synthetic control with explicit error bounds. Our approach builds upon tools from non-private synthetic control and differentially private empirical risk minimization. We provide upper and lower bounds on the sensitivity of the synthetic control query and provide explicit error bounds on the accuracy of our private synthetic control algorithms. We show that our algorithms produce accurate predictions for the target unit, and that the cost of privacy is small. Finally, we empirically evaluate the performance of our algorithm, and show favorable performance in a variety of parameter regimes, as well as providing guidance to practitioners for hyperparameter tuning."}, "cited_paper_content": {"title": "Robust Synthetic Control", "abstract": "We present a robust generalization of the synthetic control method for comparative case studies. Like the classical method cf. Abadie and Gardeazabal (2003), we present an algorithm to estimate the unobservable counterfactual of a treatment unit. A distinguishing feature of our algorithm is that of de-noising the data matrix via singular value thresholding, which renders our approach robust in multiple facets: it automatically identifies a good subset of donors for the synthetic control, overcomes the challenges of missing data, and continues to work well in settings where covariate information may not be provided. We posit that the setting can be viewed as an instance of the Latent Variable Model and provide the first finite sample analysis (coupled with asymptotic results) for the estimation of the counterfactual. Our algorithm accurately imputes missing entries and filters corrupted observations in producing a consistent estimator of the underlying signal matrix, provided p = \u03a9(T-1+\u03b6) for some \u03b6 > 0; here, p is the fraction of observed data and T is the time interval of interest. Under the same proportion of observations, we demonstrate that the mean-squared error in our counterfactual estimation scales as O(\u03c32/p + 1/\u221aT), where \u03c32 is the variance of the inherent noise. Additionally, we introduce a Bayesian framework to quantify the estimation uncertainty. Our experiments, using both synthetic and real-world datasets, demonstrate that our robust generalization yields an improvement over the classical synthetic control method."}, "keywords": ["non-private synthetic control", "synthetic control"], "citation_intent": "result"}
{"citing_id": "2304.01143v1", "cited_id": "1904.05160", "section_title": "Proposed Long-Tail Video Benchmarks", "citation": "We call these curated versions SSv2-LT and VideoLT-LT, and resample these following the recipe used in #REFR for ImageNet-LT and Places-LT (sampling from a Pareto distribution with \u03b1 = 6).", "text_before_citation": ["Having identified weaknesses in current benchmarks used for long-tail video recognition, we first propose to use EPIC-KITCHENS-100 as it is naturally-collected and satisfies the long-tail properties (as defined in Sec. 2).", "We also propose to resample public video datasets, so their properties are in line with curated long-tail image datasets.", "SSv2 #OTHEREFR is chosen as it is widely considered to be a good test of temporal understanding and has previously been re-purposed for evaluating few-shot works #OTHEREFR . Similarly, VideoLT #OTHEREFR targets fine-grained classes."], "text_after_citation": ["Table 1 demonstrates these curated versions match the desired properties as visualised in Before proceeding to the method, ablations are first performed at a dataset level, where different curated versions of SSv2-LT are compared to demonstrate the impact on longtail properties and the effect of few-shot classes.", "Full implementation details of models and metrics will be given in Sec.", "6, but for these ablations it suffices to say that Motionformer #OTHEREFR is trained with cross-entropy, reporting average class accuracy over the test set, as well as over few-shot, tail and head classes."], "citing_paper_content": {"title": "Use Your Head: Improving Long-Tail Video Recognition", "abstract": "This paper presents an investigation into long-tail video recognition. We demonstrate that, unlike naturallycollected video datasets and existing long-tail image benchmarks, current video benchmarks fall short on multiple long-tailed properties. Most critically, they lack few-shot classes in their tails. In response, we propose new video benchmarks that better assess long-tail recognition, by sampling subsets from two datasets: SSv2 and VideoLT. We then propose a method, Long-Tail Mixed Reconstruction (LMR), which reduces overfitting to instances from few-shot classes by reconstructing them as weighted combinations of samples from head classes. LMR then employs label mixing to learn robust decision boundaries. It achieves state-of-the-art average class accuracy on EPIC-KITCHENS and the proposed SSv2-LT and VideoLT-LT."}, "cited_paper_content": {"title": "Large-Scale Long-Tailed Recognition In An Open World", "abstract": "Real world data often have a long-tailed and open-ended distribution. A practical recognition system must classify among majority and minority classes, generalize from a few known instances, and acknowledge novelty upon a never seen instance. We define Open Long-Tailed Recognition (OLTR) as learning from such naturally distributed data and optimizing the classification accuracy over a balanced test set which include head, tail, and open classes. OLTR must handle imbalanced classification, few-shot learning, and open-set recognition in one integrated algorithm, whereas existing classification approaches focus only on one aspect and deliver poorly over the entire class spectrum. The key challenges are how to share visual knowledge between head and tail classes and how to reduce confusion between tail and open classes. We develop an integrated OLTR algorithm that maps an image to a feature space such that visual concepts can easily relate to each other based on a learned metric that respects the closed-world classification while acknowledging the novelty of the open world. Our so-called dynamic meta-embedding combines a direct image feature and an associated memory feature, with the feature norm indicating the familiarity to known classes. On three large-scale OLTR datasets we curate from object-centric ImageNet, scene-centric Places, and face-centric MS1M data, our method consistently outperforms the state-of-the-art. Our code, datasets, and models enable future OLTR research and are publicly available at \\url{https://liuziwei7.github.io/projects/LongTail.html}."}, "keywords": ["VideoLT", "ImageNet-LT"], "citation_intent": "method"}
{"citing_id": "2303.02063v1", "cited_id": "1708.00588", "section_title": "C. Training Paradigms", "citation": "The former corresponds to solely infer time-dependent traffic flow fields, and the latter corresponds to system identification of traffic states #REFR .", "text_before_citation": ["Once the physics model is selected, we need to determine the sequence of parameter calibration prior to, or during the training of PUNN."], "text_after_citation": ["1) Sequential training Sequential training aims to first calibrate parameters of the PICG (i.e., parameter calibration) and then encode the known physics into the PUNN for training.", "Parameter calibration has been extensively studied in TSE using non-linear programming #OTHEREFR , genetic algorithm #OTHEREFR , least-squares fitting #OTHEREFR - #OTHEREFR , and kernel smoothing #OTHEREFR .", "The physics based parameters include \u03c1 max , u max and other nominal parameters. Tab. I summarizes the existing methods for model discovery.", "Sequential training is the default paradigm in most PIDLrelated studies, with a focus on how to make the training robust and stable, when large-scale NNs and complicated physicsinformed loss functions are involved. A growing amount of #OTHEREFR .", "The adaptive activation function has also been used in DeLISA (deep learning based iteration scheme approximation), which adopts the implicit multistep method and Runge-Kutta method for time iteration scheme to construct the learning loss when training the PUNN #OTHEREFR ."], "citing_paper_content": {"title": "Physics-Informed Deep Learning For Traffic State Estimation: A Survey And The Outlook", "abstract": "For its robust predictive power (compared to pure physics-based models) and sample-efficient training (compared to pure deep learning models), physics-informed deep learning (PIDL), a paradigm hybridizing physics-based models and deep neural networks (DNN), has been booming in science and engineering fields. One key challenge of applying PIDL to various domains and problems lies in the design of a computational graph that integrates physics and DNNs. In other words, how physics are encoded into DNNs and how the physics and data components are represented. In this paper, we provide a variety of architecture designs of PIDL computational graphs and how these structures are customized to traffic state estimation (TSE), a central problem in transportation engineering. When observation data, problem type, and goal vary, we demonstrate potential architectures of PIDL computational graphs and compare these variants using the same real-world dataset. Impact Statement-Despite benefits of physics-informed deep learning (PIDL) for robust and efficient pattern learning, how to inject physics into neural networks (NN) remains underexplored and varies case by case and across domains. We will summarize and establish a systematic design pipeline for hybrid computational graphs that facilitates the integration of physics and NNs. The insights into the modular design of the hybrid PIDL paradigm as well as the established visualization tool will not only be useful to guide transportation researchers to pursue PIDL, but can also facilitate researchers at large to better understand and decompose a PIDL problem when applied to their own application domains. It will hopefully open up the conversation across domains about establishing a unified PIDL paradigm."}, "cited_paper_content": {"title": "Hidden Physics Models: Machine Learning Of Nonlinear Partial Differential Equations", "abstract": "While there is currently a lot of enthusiasm about\"big data\", useful data is usually\"small\"and expensive to acquire. In this paper, we present a new paradigm of learning partial differential equations from {\\em small} data. In particular, we introduce \\emph{hidden physics models}, which are essentially data-efficient learning machines capable of leveraging the underlying laws of physics, expressed by time dependent and nonlinear partial differential equations, to extract patterns from high-dimensional data generated from experiments. The proposed methodology may be applied to the problem of learning, system identification, or data-driven discovery of partial differential equations. Our framework relies on Gaussian processes, a powerful tool for probabilistic inference over functions, that enables us to strike a balance between model complexity and data fitting. The effectiveness of the proposed approach is demonstrated through a variety of canonical problems, spanning a number of scientific domains, including the Navier-Stokes, Schr\\\"odinger, Kuramoto-Sivashinsky, and time dependent linear fractional equations. The methodology provides a promising new direction for harnessing the long-standing developments of classical methods in applied mathematics and mathematical physics to design learning machines with the ability to operate in complex domains without requiring large quantities of data."}, "keywords": ["time-dependent traffic flow", "system identification"], "citation_intent": "background"}
{"citing_id": "2303.14690v3", "cited_id": "1909.03292", "section_title": "Piston Design", "citation": "For achieving leak-proof optimized pistons, the user may refer to #REFR for recommendations on choosing etaf and betaf.", "text_before_citation": ["15b , c that are obtained by the following call TOPress (300 , 100 , 0.2 , 3 , 2.4 , 0.1 , 8 , lst , 150) with lst=1 and lst=0, respectively.", "We consider the full model to solve this problem to check any deviation from the symmetric nature of the problem.", "One notes that the optimized design retains the symmetric nature ( Fig.", "15b ) and qualitatively resembles the features of the piston structures reported in the previous literature.", "One may also solve this problem using only the symmetric half of the domain."], "text_after_citation": ["Different optimized material layouts are obtained with lst=1 and lst=0."], "citing_paper_content": {"title": "Topress: A Matlab Implementation For Topology Optimization Ofstructures Subjected To Design-Dependent Pressure Loads", "abstract": "In a topology optimization setting, design-dependent fluidic pressure loads pose several challenges as their direction, magnitude, and location alter with topology evolution. This paper offers a compact 100-line MATLAB code, TOPress, for topology optimization of structures subjected to fluidic pressure loads using the method of moving asymptotes. The code is intended for pedagogical purposes and aims to ease the beginners' and students' learning toward the topology optimization with designdependent fluidic pressure loads. TOPress is developed per the approach first reported in Kumar et al. (Struct Multidisc Optim 61(4):1637-1655, 2020). The Darcy law, in conjunction with the drainage term, is used to model the applied pressure load. The consistent nodal loads are determined from the obtained pressure field. The employed approach facilitates inexpensive computation of the load sensitivities using the adjoint-variable method. Compliance minimization subject to volume constraint optimization problems are solved. The success and efficacy of the code are demonstrated by solving benchmark numerical examples involving pressure loads, wherein the importance of load sensitivities is also demonstrated. TOPress contains six main parts, is described in detail, and is extended to solve different problems. Steps to include a projection filter are provided to achieve loadbearing designs close to 0-1. The code is provided in Appendix B and can also be downloaded along with its extensions from https://github.com/PrabhatIn/TOPress."}, "cited_paper_content": {"title": "Topology Optimization Of Fluidic Pressure Loaded Structures And Compliant Mechanisms Using The Darcy Method", "abstract": "In various applications, design problems involving structures and compliant mechanisms experience fluidic pressure loads. During topology optimization of such design problems, these loads adapt their direction and location with the evolution of the design, which poses various challenges. A new density-based topology optimization approach using Darcy\u2019s law in conjunction with a drainage term is presented to provide a continuous and consistent treatment of design-dependent fluidic pressure loads. The porosity of each finite element and its drainage term are related to its density variable using a Heaviside function, yielding a smooth transition between the solid and void phases. A design-dependent pressure field is established using Darcy\u2019s law and the associated PDE is solved using the finite element method. Further, the obtained pressure field is used to determine the consistent nodal loads. The approach provides a computationally inexpensive evaluation of load sensitivities using the adjoint-variable method. To show the efficacy and robustness of the proposed method, numerical examples related to fluidic pressure-loaded stiff structures and small-deformation compliant mechanisms are solved. For the structures, compliance is minimized, whereas for the mechanisms, a multi-criteria objective is minimized with given resource constraints."}, "keywords": ["leak-proof optimized pistons"], "citation_intent": "background"}
{"citing_id": "2303.11530v1", "cited_id": "2003.08515", "section_title": "Datasets And Pre-Training", "citation": "We begin our experiments by rendering synthetic models from the PartNet-Mobility dataset #REFR in various articulation states since this enables us to obtain sufficient annotations to train 2D segmentation networks and thus enable transfer learning applications.", "text_before_citation": ["Note that such manual annotations are used only to evaluate our active learning framework.", "From Table 1 , we observe that the majority of data samples in OPDReal belong to the Storage category (91.67%), with the rest distributed among the remaining categories.", "The difference in distribution between the largest and the second largest category is 87.74%, and between the largest and smallest categories is 91.14%.", "With such data skewness towards one category, models trained on OPDReal will likely overfit to the dominant category.Our dataset, on the other hand, contains smaller variations in data distributions across the six object categories, where the difference in data distribution between the largest and second largest categories is 5.88%, and between the largest and smallest categories is 29.47%.", "Pre-training."], "text_after_citation": ["Our synthetic dataset contains around 32K articulation images, with equal data samples for each object category.", "We use a 90%-10% train-test split to train on this synthetic dataset, and use this trained model for fine-tuning on real images.", "To this end, we use all data samples from both datasets, OPDReal and ours, with 80%-20% train-test split.", "We implement our network in PyTorch on a Nvidia RTX 2080 Ti GPU. All images are resized to 256\u00d7256 for training.", "During pre-training on the PartNet-Mobility dataset, we use the Adam optimizer with a learning rate (lr) of 2.5e-4, and train for 2K epochs."], "citing_paper_content": {"title": "Coarse-To-Fine Active Segmentation Of Interactable Parts In Real Scene Images", "abstract": "OPDRCNN-C Mask2Former ours w/o AL GT ours with AL input Figure 1: 2D segmentation results on interactable object parts in real-world indoor scene images. We show part segmentation masks obtained using similar methods, viz. OPD [15] and Mask2Former [7]. We also contrast the results of our proposed pose-aware masked-attention model, with and without Active Learning (AL) framework."}, "cited_paper_content": {"title": "Sapien: A Simulated Part-Based Interactive Environment", "abstract": "Building home assistant robots has long been a pursuit for vision and robotics researchers. To achieve this task, a simulated environment with physically realistic simulation, sufficient articulated objects, and transferability to the real robot is indispensable. Existing environments achieve these requirements for robotics simulation with different levels of simplification and focus. We take one step further in constructing an environment that supports household tasks for training robot learning algorithm. Our work, SAPIEN, is a realistic and physics-rich simulated environment that hosts a large-scale set for articulated objects. Our SAPIEN enables various robotic vision and interaction tasks that require detailed part-level understanding.We evaluate state-of-the-art vision algorithms for part detection and motion attribute recognition as well as demonstrate robotic interaction tasks using heuristic approaches and reinforcement learning algorithms. We hope that our SAPIEN can open a lot of research directions yet to be explored, including learning cognition through interaction, part motion discovery, and construction of robotics-ready simulated game environment."}, "keywords": ["2D segmentation networks"], "citation_intent": "method"}
{"citing_id": "2304.10001v1", "cited_id": "1801.04264", "section_title": "Introduction", "citation": "In this work, we propose to use the weakly supervised anomaly detection #REFR to detect baby cries on audio signals.", "text_before_citation": ["On the other hand, the complexity of the CNN may be high, preventing it be used in common embedded devices, like low-cost IP cameras or tablets.", "This paper will address this issue by using a super lightweight CNN.", "Most, if not all, existing CNN methods use supervised learning. Therefore, frame-level annotation is needed.", "This annotation is very time-consuming and prone to human mistakes.", "In the few datasets available online, some of the audio files [8, 27, #OTHEREFR are trimmed and annotated, and some others #OTHEREFR are untrimmed with only audio-level annotations."], "text_after_citation": ["This weak supervision only requires weak annotation, i.e., if there is a cry in the audio file without frame-level annotation, therefore making the data annotation a lot easier.", "Based on this weakly supervised anomaly detection and a well-known pretrained VGGish network #OTHEREFR for audio feature extraction, we design a data mining technique to obtain frame-level datasets for supervised CNN classification.", "We use this dataset to train a simple, lightweight CNN, which runs very fast on embedded devices.", "This CNN, as the feature extractor in an anomaly detection framework, gives better performance than the CNN alone.", "The contribution of this paper is three-fold,"], "citing_paper_content": {"title": "Weakly Supervised Detection Of Baby Cry", "abstract": "Detection of baby cries is an important part of baby monitoring and health care. Almost all existing methods use supervised SVM, CNN, or their varieties. In this work, we propose to use weakly supervised anomaly detection to detect a baby cry. In this weak supervision, we only need weak annotation if there is a cry in an audio file. We design a data mining technique using the pretrained VGGish feature extractor and an anomaly detection network on long untrimmed audio files. The obtained datasets are used to trained a simple CNN feature network for cry/non-cry classification. This CNN is then used as a feature extractor in an anomaly detection framework to achieve better cry detection performance."}, "cited_paper_content": {"title": "Real-World Anomaly Detection In Surveillance Videos", "abstract": "Surveillance videos are able to capture a variety of realistic anomalies. In this paper, we propose to learn anomalies by exploiting both normal and anomalous videos. To avoid annotating the anomalous segments or clips in training videos, which is very time consuming, we propose to learn anomaly through the deep multiple instance ranking framework by leveraging weakly labeled training videos, i.e. the training labels (anomalous or normal) are at video-level instead of clip-level. In our approach, we consider normal and anomalous videos as bags and video segments as instances in multiple instance learning (MIL), and automatically learn a deep anomaly ranking model that predicts high anomaly scores for anomalous video segments. Furthermore, we introduce sparsity and temporal smoothness constraints in the ranking loss function to better localize anomaly during training. We also introduce a new large-scale first of its kind dataset of 128 hours of videos. It consists of 1900 long and untrimmed real-world surveillance videos, with 13 realistic anomalies such as fighting, road accident, burglary, robbery, etc. as well as normal activities. This dataset can be used for two tasks. First, general anomaly detection considering all anomalies in one group and all normal activities in another group. Second, for recognizing each of 13 anomalous activities. Our experimental results show that our MIL method for anomaly detection achieves significant improvement on anomaly detection performance as compared to the state-of-the-art approaches. We provide the results of several recent deep learning baselines on anomalous activity recognition. The low recognition performance of these baselines reveals that our dataset is very challenging and opens more opportunities for future work."}, "keywords": ["baby cries", "weakly supervised anomaly"], "citation_intent": "method"}
{"citing_id": "2303.08463v1", "cited_id": "1801.04264", "section_title": "A. Annotation", "citation": "The UCF-Crime #REFR dataset is a large-scale video anomaly detection dataset, which contains 13 anomaly classes.", "text_before_citation": [], "text_after_citation": ["This dataset consists of not only abnormal videos but also normal videos where no abnormal events occur.", "UCF-Crime only has video-level abnormal-or-not labels, i.e., whether the current video has abnormal events or not, without the specific class of the abnormal events.", "When we use UCF-Crime for TAL, we only care about abnormal events. The normal events act as the background.", "Therefore, to better match the temporal action localization task, we remove videos which only contain normal events to construct our UCF-Crime-TAL dataset.", "Then we annotate the UCF-Crime-TAL dataset with dense labels at the frame level. Given a video, we first locate the abnormal events."], "citing_paper_content": {"title": "Co-Occurrence Matters: Learning Action Relation For Temporal Action Localization", "abstract": "Temporal action localization (TAL) is a prevailing task due to its great application potential. Existing works in this field mainly suffer from two weaknesses: (1) They often neglect the multi-label case and only focus on temporal modeling. (2) They ignore the semantic information in class labels and only use the visual information. To solve these problems, we propose a novel Co-Occurrence Relation Module (CORM) that explicitly models the co-occurrence relationship between actions. Besides the visual information, it further utilizes the semantic embeddings of class labels to model the co-occurrence relationship. The CORM works in a plug-and-play manner and can be easily incorporated with the existing sequence models. By considering both visual and semantic co-occurrence, our method achieves high multi-label relationship modeling capacity. Meanwhile, existing datasets in TAL always focus on low-semantic atomic actions. Thus we construct a challenging multi-label dataset UCF-Crime-TAL that focuses on high-semantic actions by annotating the UCF-Crime dataset at frame level and considering the semantic overlap of different events. Extensive experiments on two commonly used TAL datasets, i.e., MultiTHUMOS and TSU, and our newly proposed UCF-Crime-TAL demenstrate the effectiveness of the proposed CORM, which achieves state-of-theart performance on these datasets."}, "cited_paper_content": {"title": "Real-World Anomaly Detection In Surveillance Videos", "abstract": "Surveillance videos are able to capture a variety of realistic anomalies. In this paper, we propose to learn anomalies by exploiting both normal and anomalous videos. To avoid annotating the anomalous segments or clips in training videos, which is very time consuming, we propose to learn anomaly through the deep multiple instance ranking framework by leveraging weakly labeled training videos, i.e. the training labels (anomalous or normal) are at video-level instead of clip-level. In our approach, we consider normal and anomalous videos as bags and video segments as instances in multiple instance learning (MIL), and automatically learn a deep anomaly ranking model that predicts high anomaly scores for anomalous video segments. Furthermore, we introduce sparsity and temporal smoothness constraints in the ranking loss function to better localize anomaly during training. We also introduce a new large-scale first of its kind dataset of 128 hours of videos. It consists of 1900 long and untrimmed real-world surveillance videos, with 13 realistic anomalies such as fighting, road accident, burglary, robbery, etc. as well as normal activities. This dataset can be used for two tasks. First, general anomaly detection considering all anomalies in one group and all normal activities in another group. Second, for recognizing each of 13 anomalous activities. Our experimental results show that our MIL method for anomaly detection achieves significant improvement on anomaly detection performance as compared to the state-of-the-art approaches. We provide the results of several recent deep learning baselines on anomalous activity recognition. The low recognition performance of these baselines reveals that our dataset is very challenging and opens more opportunities for future work."}, "keywords": ["UCF-Crime dataset"], "citation_intent": "background"}
{"citing_id": "2304.01518v1", "cited_id": "1807.01613", "section_title": "Background", "citation": "NPs are stochastic processes using neural networks to capture the ground truth stochastic processes that generate the given data #REFR .", "text_before_citation": ["In a regression problem, y C i is a continuous vector, whereas in a classification problem, y C i can be a one-hot encoded vector. A target set is defined as", "T = {x T i } N T i=1", "where the number of target points is N T .", "We let a mapping function f : X \u2192 Y be sampled from a probability distribution over functions and a stochastic process defined as a probability measure of such sampled functions.", "We assume that there exists a finite set of all possible functions P(X ) where f \u2208 P(X )."], "text_after_citation": ["NPs learn the distribution of functions and provide uncertainty of target samples, preserving the property of GPs.", "At the same time, NPs exploit function approximation of neural networks in a more efficient manner than GPs.", "The learning objective of NPs is to maximise the likelihood of p(f (T )|C, T ) (Garnelo et al., 2018b; a; #OTHEREFR ."], "citing_paper_content": {"title": "Multimodal Neural Processes For Uncertainty Estimation", "abstract": "Neural processes (NPs) have brought the representation power of parametric deep neural networks and the reliable uncertainty estimation of non-parametric Gaussian processes together. Although recent development of NPs has shown success in both regression and classification, how to adapt NPs to multimodal data has not be carefully studied. For the first time, we propose a new model of NP family for multimodal uncertainty estimation, namely Multimodal Neural Processes. In a holistic and principled way, we develop a dynamic context memory updated by the classification error, a multimodal Bayesian aggregation mechanism to aggregate multimodal representations, and a new attention mechanism for calibrated predictions. In extensive empirical evaluation, our method achieves the stateof-the-art multimodal uncertainty estimation performance, showing its appealing ability of being robust against noisy samples and reliable in outof-domain detection."}, "cited_paper_content": {"title": "Conditional Neural Processes", "abstract": "Deep neural networks excel at function approximation, yet they are typically trained from scratch for each new function. On the other hand, Bayesian methods, such as Gaussian Processes (GPs), exploit prior knowledge to quickly infer the shape of a new function at test time. Yet GPs are computationally expensive, and it can be hard to design appropriate priors. In this paper we propose a family of neural models, Conditional Neural Processes (CNPs), that combine the benefits of both. CNPs are inspired by the flexibility of stochastic processes such as GPs, but are structured as neural networks and trained via gradient descent. CNPs make accurate predictions after observing only a handful of training data points, yet scale to complex functions and large datasets. We demonstrate the performance and versatility of the approach on a range of canonical machine learning tasks, including regression, classification and image completion."}, "keywords": ["ground truth stochastic", "neural networks"], "citation_intent": "method"}
{"citing_id": "2304.07250v1", "cited_id": "1804.05625", "section_title": "Rpr From Optical Flow", "citation": "On the other hand, VI-DSO #REFR ) employs a combined energy functional to jointly estimate camera poses and sparse scene geometry by minimizing the visual and inertial measurement errors.", "text_before_citation": ["Nevertheless, Transformers are not adequate for integrating absolute and relative poses with their low-dimensional information.", "Also, the bidirectional properties #OTHEREFR of LSTMs are unfeasible for pose modeling.", "#OTHEREFR fuse an absolute pose (from GNSS data) with a relative pose (from vehicle odometry).", "The following two works integrate relative poses from inertial data with absolute poses from visual data.", "VINet #OTHEREFR ) utilizes a concatenation approach to merge relative features extracted by an inertial LSTM encoder with absolute features extracted by a visual encoder."], "text_after_citation": [], "citing_paper_content": {"title": "Fusing Structure From Motion And Simulation-Augmented Pose Regression From Optical Flow For Challenging Indoor Environments \u22c6", "abstract": "The localization of objects is a crucial task in various applications such as robotics, virtual and augmented reality, and the transportation of goods in warehouses. Recent advances in deep learning have enabled the localization using monocular visual cameras. While structure from motion (SfM) predicts the absolute pose from a point cloud, absolute pose regression (APR) methods learn a semantic understanding of the environment through neural networks. However, both fields face challenges caused by the environment such as motion blur, lighting changes, repetitive patterns, and featureless structures. This study aims to address these challenges by incorporating additional information and regularizing the absolute pose using relative pose regression (RPR) methods. The optical flow between consecutive images is computed using the Lucas-Kanade algorithm, and the relative pose is predicted using an auxiliary small recurrent convolutional network. The fusion of absolute and relative poses is a complex task due to the mismatch between the global and local coordinate systems. State-of-the-art methods fusing absolute and relative poses use pose graph optimization (PGO) to regularize the absolute pose predictions using relative poses. In this work, we propose recurrent fusion networks to optimally align absolute and relative pose predictions to improve the absolute pose prediction. We evaluate eight different recurrent units and construct a simulation environment to pre-train the APR and RPR networks for better generalized training. Additionally, we record a large database of different scenarios in a challenging large-scale indoor environment that mimics a warehouse with transportation robots. We conduct hyperparameter searches and experiments to show the effectiveness of our recurrent fusion method compared to PGO. \u22c6 Evaluating the performance of absolute pose regression methods, i.e., SfM and PoseNet, in challenging scenarios such as repetitive patterns, motion blur, structure-less surfaces, and difficult lighting conditions. Fusing absolute and relative poses using either PGO or recurrent networks."}, "cited_paper_content": {"title": "Direct Sparse Visual-Inertial Odometry Using Dynamic Marginalization", "abstract": "We present VI-DSO, a novel approach for visual-inertial odometry, which jointly estimates camera poses and sparse scene geometry by minimizing photometric and IMU measurement errors in a combined energy functional. The visual part of the system performs a bundle-adjustment like optimization on a sparse set of points, but unlike key-point based systems it directly minimizes a photometric error. This makes it possible for the system to track not only corners, but any pixels with large enough intensity gradients. IMU information is accumulated between several frames using measurement preintegration, and is inserted into the optimization as an additional constraint between keyframes. We explicitly include scale and gravity direction into our model and jointly optimize them together with other variables such as poses. As the scale is often not immediately observable using IMU data this allows us to initialize our visual-inertial system with an arbitrary scale instead of having to delay the initialization until everything is observable. We perform partial marginalization of old variables so that updates can be computed in a reasonable time. In order to keep the system consistent we propose a novel strategy which we call\"dynamic marginalization\". This technique allows us to use partial marginalization even in cases where the initial scale estimate is far from the optimum. We evaluate our method on the challenging EuRoC dataset, showing that VI-DSO outperforms the state of the art."}, "keywords": ["camera poses", "inertial measurement errors"], "citation_intent": "background"}
{"citing_id": "2303.10128v1", "cited_id": "1303.6175", "section_title": "Direct Evidence Of Compression", "citation": "The ample evidence of compression even when orthographic word lengths are replaced by word durations in human languages provides stronger support for compression as a universal principle of the organization of languages with respect to previous research relying on word length in characters #REFR .", "text_before_citation": ["We have found that word lengths are shorter than expected by chance ( < ) in all languages in every collection (Figure 2 ).", "Such a systematic finding is unlikely to be accidental and strongly indicates that compression is acting in all languages in our sample.", "Crucially, the finding holds independently of how word length is measured."], "text_after_citation": ["It could be argued that these findings constitute evidence of compression in ensembles of language but not in individual languages.", "The reason is that < does not imply that the difference between the actual word length and the random baseline is statistically significant for a single language.", "However, we have shown that the Pearson correlation is indeed a linear function of and (Appendix A) and thus is significantly small in every language where the law of abbreviation has been confirmed using a Pearson correlation test.", "Finally, the direct correspondence we have established between the average length of types ( ) and the random baseline sheds new light on previous research.", "For instance, it has been shown that < in Chinese characters in six time periods spanning two millennia (Chen et al., 2015, Fig."], "citing_paper_content": {"title": "Direct And Indirect Evidence Of Compression Of Word Lengths. Zipf'S Law Of Abbreviation Revisited", "abstract": "Direct and indirect evidence of compression of word lengths. Zipf's law of abbreviation revisited. Sonia Petrini 1 (0000-0002-0514-6223), Antoni Casas-i-Mu\u00f1oz 2 (0000-0001-5690-316X), Jordi Cluet-i-Martinell 2 (0000-0003-4188-6728), Mengxue Wang 2 (0000-0002-8262-9333), Christian Bentz 3 (0000-0001-6570-9326), Ramon Ferrer-i-Cancho 1* (0000-0002-7820-923X)"}, "cited_paper_content": {"title": "Compression As A Universal Principle Of Animal Behavior", "abstract": "A key aim in biology and psychology is to identify fundamental principles underpinning the behavior of animals, including humans. Analyses of human language and the behavior of a range of non-human animal species have provided evidence for a common pattern underlying diverse behavioral phenomena: Words follow Zipf\u2019s law of brevity (the tendency of more frequently used words to be shorter), and conformity to this general pattern has been seen in the behavior of a number of other animals. It has been argued that the presence of this law is a sign of efficient coding in the information theoretic sense. However, no strong direct connection has been demonstrated between the law and compression, the information theoretic principle of minimizing the expected length of a code. Here, we show that minimizing the expected code length implies that the length of a word cannot increase as its frequency increases. Furthermore, we show that the mean code length or duration is significantly small in human language, and also in the behavior of other species in all cases where agreement with the law of brevity has been found. We argue that compression is a general principle of animal behavior that reflects selection for efficiency of coding."}, "keywords": ["orthographic word lengths", "word length"], "citation_intent": "result"}
{"citing_id": "2303.04894v1", "cited_id": "1709.06428", "section_title": "A. Framework A) Robot:", "citation": "This differs from our previous work #REFR where we assign stationary sensors to localize targets.", "text_before_citation": ["where z j i,t denotes the measurement of target j by robot i's sensor at time t.", "v j i,t (x i,t , y j,t ) denotes the Gaussian measurement noise with zero mean and covariance", "R(x i,t , y j,t ), i.e., v j i,t (x i,t , y j,t ) \u223c N (0, R(x i,t , y j,t )).", "Here, both the sensing model h j i (x i,t , y j,t ) and measurement noise v j i,t (x i,t , y j,t ) depend on the state of robot i and target j.", "d) Objective: The objective is to maximize the tracking quality of the targets by assigning robots and their selected actions to the targets."], "text_after_citation": ["Here, considering that robots are moving by executing actions, we assign robot-actions (per robot an action) to track the targets.", "That way, we denote \u03c6(j) as the set of robotactions assigned to target j and \u03c6 \u22121 (i k ) as the set of targets assigned to robot-action i k (i.e., robot i with action k).", "We use \u03c6 l (j) to denote the l th robot-action assigned to target j. We order the assigned robot-actions such that", "\u03c6 1 (j) \u2264 \u03c6 2 (j) \u2264 \u2022 \u2022 \u2022 \u2264 \u03c6 L (j).", "Denote the corresponding robot for each robot-action \u03c6 l (j) as i(\u03c6 l (j))."], "citing_paper_content": {"title": "Assignment Algorithms For Multi-Robot Multi-Target Tracking With Sufficient And Limited Sensing Capability", "abstract": "We study the problem of assigning robots with actions to track targets. The objective is to optimize the robot team's tracking quality which can be defined as the reduction in the uncertainty of the targets' states. Specifically, we consider two assignment problems given the different sensing capabilities of the robots. In the first assignment problem, a single robot is sufficient to track a target. To this end, we present a greedy algorithm (Algorithm 1) that assigns a robot with its action to each target. We prove that the greedy algorithm has a 1/2approximation bound and runs in polynomial time. Then, we study the second assignment problem where two robots are necessary to track a target. We design another greedy algorithm (Algorithm 2) that assigns a pair of robots with their actions to each target. We prove that the greedy algorithm achieves a 1/3-approximation bound and has a polynomial running time. Moreover, we illustrate the performance of the two greedy algorithms in the ROS-Gazebo environment where the tracking patterns of one robot following one target using Algorithm 1 and two robots following one target using Algorithm 2 are clearly observed. Further, we conduct extensive comparisons to demonstrate that the two greedy algorithms perform close to their optimal counterparts and much better than their respective (1/2 and 1/3) approximation bounds."}, "cited_paper_content": {"title": "Sensor Assignment Algorithms To Improve Observability While Tracking Targets", "abstract": "In this paper, we study two sensor assignment problems for multitarget tracking with the goal of improving the observability of the underlying estimator. We consider various measures of the observability matrix as the assignment value function. We first study the general version where the sensors must form teams to track individual targets. If the value function is monotonically increasing and submodular, then a greedy algorithm yields a $1/2$ \u2013approximation. We then study a restricted version where exactly two sensors must be assigned to each target. We present a $1/3$ \u2013approximation algorithm for this problem, which holds for arbitrary value functions (not necessarily submodular or monotone). In addition to approximation algorithms, we also present various properties of observability measures. We show that the inverse of the condition number of the observability matrix is neither monotone nor submodular, but present other measures that are. Specifically, we show that the trace and rank of the symmetric observability matrix are monotone and submodular and the log determinant of the symmetric observability matrix is monotone and submodular when the matrix is nonsingular. If the target's motion model is not known, the inverse cannot be computed exactly. Instead, we present a lower bound for distance sensors. In addition to theoretical results, we evaluate our results empirically through simulations."}, "keywords": ["stationary sensors"], "citation_intent": "result"}
{"citing_id": "2303.03898v1", "cited_id": "1804.02767", "section_title": "D. Locanet", "citation": "The proposed network architecture is largely similar to the object detector YOLOv3 #REFR with changes in the output.", "text_before_citation": ["We summarize Locanet, a framework that estimates the 3D positions of neighboring robots using a deep neural network #OTHEREFR ."], "text_after_citation": ["In particular, rather than predicting bounding boxes for each robot, the network outputs for each robot the pixel coordinates of the robot's center and its distance to the camera directly.", "In order to support multi-robot detection, the neural network outputs 28 \u00d7 40 grid cells with two channels each. The first channel is the confidence map.", "Grid cell locations refer to the center of the robot in the image plane, where a confidence score above a threshold indicates that a robot was detected.", "A score close to zero signals that no robot was detected in this part of the image.", "The second channel is the depth map containing predicted distances between the neighboring robot and the camera. This channel is only valid for high confidence scores."], "citing_paper_content": {"title": "A Dataset And Comparative Study For Vision-Based Relative Position Estimation Of Multirotor Teams Flying In Close Proximity", "abstract": "Multirotor teams are useful for inspection, delivery, and construction tasks, in which they might be required to fly very close to each other. In such close-proximity cases, nonlinear aerodynamic effects can cause catastrophic crashes, necessitating each robots' awareness of the surroundings. Existing approaches rely on expensive or heavy perception sensors. Instead, we propose to use the often ignored yaw degree-offreedom of multirotors to spin a single, cheap and lightweight monocular camera at a high angular rate for omnidirectional awareness. We provide a dataset collected with real-world physical flights as well as with 3D-rendered scenes and compare two existing learning-based methods in different settings with respect to success rate, relative position estimation, and downwash prediction accuracy. As application we demonstrate that our proposed spinning camera is capable of predicting the presence of aerodynamic downwash in a challenging swapping task."}, "cited_paper_content": {"title": "Yolov3: An Incremental Improvement", "abstract": "We present some updates to YOLO! We made a bunch of little design changes to make it better. We also trained this new network that's pretty swell. It's a little bigger than last time but more accurate. It's still fast though, don't worry. At 320x320 YOLOv3 runs in 22 ms at 28.2 mAP, as accurate as SSD but three times faster. When we look at the old .5 IOU mAP detection metric YOLOv3 is quite good. It achieves 57.9 mAP@50 in 51 ms on a Titan X, compared to 57.5 mAP@50 in 198 ms by RetinaNet, similar performance but 3.8x faster. As always, all the code is online at this https URL"}, "keywords": ["object detector"], "citation_intent": "method"}
{"citing_id": "2304.02786v1", "cited_id": "1708.06733", "section_title": "Effectiveness.", "citation": "For example, for the backdoored models produced via patch attack #REFR , the inverted trigger share the same location (i.e., right upper corner) and the color (i.e., yellow) with the injected trigger.", "text_before_citation": ["These results indicate UNICORN is effective for inverting backdoor triggers injected in different spaces.", "We also show the visualization of the inverted triggers as well as the ground-truth triggers in Fig. 3 .", "The first row shows the original image and the samples pasted with different injected triggers.", "The dataset and the model used here are ImageNet and VGG16.", "As can be observed, UNICORN can effectively invert the trigger that is similar to injected triggers."], "text_after_citation": ["For the attack in filter space, feature space and numerical space, our method also successfully inverted the global perturbations.", "The inverted triggers might be not equal to the injected triggers.", "This is because existing backdoor attacks are inaccurate #OTHEREFR .", "In other words, a trigger that is similar but not equal to the injected one is also able to activate the malicious backdoor behaviors of the model.", "We observed that pixel space attacks are more accurate than the attacks in other spaces."], "citing_paper_content": {"title": "Unicorn: A Unified Backdoor Trigger Inver-Sion Framework", "abstract": "The backdoor attack, where the adversary uses inputs stamped with triggers (e.g., a patch) to activate pre-planted malicious behaviors, is a severe threat to Deep Neural Network (DNN) models. Trigger inversion is an effective way of identifying backdoor models and understanding embedded adversarial behaviors. A challenge of trigger inversion is that there are many ways of constructing the trigger. Existing methods cannot generalize to various types of triggers by making certain assumptions or attack-specific constraints. The fundamental reason is that existing work does not consider the trigger's design space in their formulation of the inversion problem. This work formally defines and analyzes the triggers injected in different spaces and the inversion problem. Then, it proposes a unified framework to invert backdoor triggers based on the formalization of triggers and the identified inner behaviors of backdoor models from our analysis. Our prototype UNICORN is general and effective in inverting backdoor triggers in DNNs. The code can be found at https://github.com/ RU-System-Software-and-Security/UNICORN."}, "cited_paper_content": {"title": "Badnets: Identifying Vulnerabilities In The Machine Learning Model Supply Chain", "abstract": "Deep learning-based techniques have achieved state-of-the-art performance on a wide variety of recognition and classification tasks. However, these networks are typically computationally expensive to train, requiring weeks of computation on many GPUs; as a result, many users outsource the training procedure to the cloud or rely on pre-trained models that are then fine-tuned for a specific task. In this paper we show that outsourced training introduces new security risks: an adversary can create a maliciously trained network (a backdoored neural network, or a \\emph{BadNet}) that has state-of-the-art performance on the user's training and validation samples, but behaves badly on specific attacker-chosen inputs. We first explore the properties of BadNets in a toy example, by creating a backdoored handwritten digit classifier. Next, we demonstrate backdoors in a more realistic scenario by creating a U.S. street sign classifier that identifies stop signs as speed limits when a special sticker is added to the stop sign; we then show in addition that the backdoor in our US street sign detector can persist even if the network is later retrained for another task and cause a drop in accuracy of {25}\\% on average when the backdoor trigger is present. These results demonstrate that backdoors in neural networks are both powerful and---because the behavior of neural networks is difficult to explicate---stealthy. This work provides motivation for further research into techniques for verifying and inspecting neural networks, just as we have developed tools for verifying and debugging software."}, "keywords": ["backdoored models"], "citation_intent": "background"}
{"citing_id": "2303.07012v1", "cited_id": "1801.03924", "section_title": "Quantitative Evaluation Metrics & User Study", "citation": "We use the learned perceptual image patch similarity (LPIPS) metric #REFR to measure the diversity of generated images.", "text_before_citation": ["The lower the FID, the better the quality of the generated images. NDB and JSD.", "We use the number of statistically different bins (NDB) and the Jensen-Shannon divergence (JSD) #OTHEREFR to evaluate the authenticity of the generated images.", "Compared with FID, the NDB and JSD metrics are directly applied to image pixels and do not rely on the learned representation.", "This makes the metric more sensitive to pixel-level differences in images.", "We set the number of bins for NDB to 50. LPIPS."], "text_after_citation": ["We generate 1,000 samples for each class and compute the LPIPS distance between pairs of samples.", "The average LPIPS distance for all the classes is used as the final LPIPS value. User Study. We conduct a user study based on pairwise comparisons.", "Given the POC image groups generated by AGTGAN and other models, each subject needs to answer the question \"Which POC image group is more realistic?\" with a real POC image as a reference."], "citing_paper_content": {"title": "Agtgan: Unpaired Image Translation For Photographic Ancient Character Generation", "abstract": "The study of ancient writings has great value for archaeology and philology. Essential forms of material are photographic characters, but manual photographic character recognition is extremely timeconsuming and expertise-dependent. Automatic classification is therefore greatly desired. However, the current performance is limited due to the lack of annotated data. Data generation is an inexpensive but useful solution to data scarcity. Nevertheless, the diverse glyph shapes and complex background textures of photographic ancient characters make the generation task difficult, leading to unsatisfactory results of existing methods. To this end, we propose an unsupervised generative adversarial network called AGTGAN in this paper. By explicitly modeling global and local glyph shape style, followed by a stroke-aware texture transfer and an associate adversarial learning mechanism, our method can generate characters with diverse glyphs and realistic textures. We evaluate our method on photographic ancient character datasets, e.g., OBC306 and CSDD. Our method outperforms other state-of-the-art methods in terms of various metrics and performs much better in terms of the diversity and authenticity of generated samples. With our generated images, experiments on the largest photographic oracle bone character dataset show that our method can achieve a significant increase in classification accuracy, up to 16.34%. The source code is"}, "cited_paper_content": {"title": "The Unreasonable Effectiveness Of Deep Features As A Perceptual Metric", "abstract": "While it is nearly effortless for humans to quickly assess the perceptual similarity between two images, the underlying processes are thought to be quite complex. Despite this, the most widely used perceptual metrics today, such as PSNR and SSIM, are simple, shallow functions, and fail to account for many nuances of human perception. Recently, the deep learning community has found that features of the VGG network trained on the ImageNet classification task has been remarkably useful as a training loss for image synthesis. But how perceptual are these so-called\"perceptual losses\"? What elements are critical for their success? To answer these questions, we introduce a new Full Reference Image Quality Assessment (FR-IQA) dataset of perceptual human judgments, orders of magnitude larger than previous datasets. We systematically evaluate deep features across different architectures and tasks and compare them with classic metrics. We find that deep features outperform all previous metrics by huge margins. More surprisingly, this result is not restricted to ImageNet-trained VGG features, but holds across different deep architectures and levels of supervision (supervised, self-supervised, or even unsupervised). Our results suggest that perceptual similarity is an emergent property shared across deep visual representations."}, "keywords": ["generated images", "learned perceptual image"], "citation_intent": "method"}
{"citing_id": "2305.00923v1", "cited_id": "1803.02155", "section_title": "Multi-Head Self-Attention Layer", "citation": "The elements in the output sequence z are calculated as the weighted sum of linearly transformed input elements #REFR .", "text_before_citation": ["Attention is the technique that has the capability of selecting and focusing on only the most pertinent information.", "In self-attention essentially the inputs are able to interact with each other to identify what parts to pay more attention to.", "This mechanism employs similar pixels and disregards the ones having no correlation to the other pixels in a feature map #OTHEREFR .", "This works in a way that each attention head presides over an input sequence, x = (x1, x2, \u2026, xn) where xi \u2208 \u211d dx , and outputs another sequence z = (z1, z2, . . ., zn) where zi \u2208 \u211d dz .", "Here dx and dz refer to the embedding size of input and output sequence respectively."], "text_after_citation": ["The weight coefficient \u03b1ij is calculated with the help of a softmax function as follows:", "zi = \u2211 =1 ij(xjW V ) -(1)", "\u03b1ij = exp ( ) \u2211 exp (e ik ) =1", "Here eij is calculated by implementing a scaled dot product to compare two input values: In multi-headed self-attention, individual attention heads are generally concatenated and run parallelly with the application of a linear transformation.", "This permits the model to together pay attention to information from multiple representation subspaces simultaneously (Vaswani et al., 2017) ."], "citing_paper_content": {"title": "Early Detection Of Alzheimer'S Disease Using Bottleneck Transformers", "abstract": "Early detection of Alzheimer's Disease (AD) and its prodromal state, Mild Cognitive Impairment (MCI), is crucial for providing suitable treatment and preventing the disease from progressing. It can also aid researchers and clinicians to identify early biomarkers and minister new treatments that have been a subject of extensive research. The application of deep learning techniques on structural Magnetic Resonance Imaging (MRI) has shown promising results in diagnosing the disease. In this research, we intend to introduce a novel approach of using an ensemble of the self-attention-based Bottleneck Transformers with a sharpness aware minimizer for early detection of Alzheimer's Disease. The proposed approach has been tested on the widely accepted ADNI dataset and evaluated using accuracy, precision, recall, F1 score, and ROC-AUC score as the performance metrics."}, "cited_paper_content": {"title": "Self-Attention With Relative Position Representations", "abstract": "Relying entirely on an attention mechanism, the Transformer introduced by Vaswani et al. (2017) achieves state-of-the-art results for machine translation. In contrast to recurrent and convolutional neural networks, it does not explicitly model relative or absolute position information in its structure. Instead, it requires adding representations of absolute positions to its inputs. In this work we present an alternative approach, extending the self-attention mechanism to efficiently consider representations of the relative positions, or distances between sequence elements. On the WMT 2014 English-to-German and English-to-French translation tasks, this approach yields improvements of 1.3 BLEU and 0.3 BLEU over absolute position representations, respectively. Notably, we observe that combining relative and absolute position representations yields no further improvement in translation quality. We describe an efficient implementation of our method and cast it as an instance of relation-aware self-attention mechanisms that can generalize to arbitrary graph-labeled inputs."}, "keywords": ["output sequence"], "citation_intent": "background"}
{"citing_id": "2303.05416v1", "cited_id": "1905.03079", "section_title": "Related Work", "citation": "The model produces superior results in comparison to Cudeiro et al. #REFR and Richard et al.", "text_before_citation": ["#OTHEREFR aims for audio-driven animation that can capture variations in multiple speakers including a much larger dataset of 250 subjects.", "They address the problem of lack of upper face motions using a categorical latent space that disentangles audio-correlated and audio-uncorrelated information based on a cross-modality loss. Fan et al.", "#OTHEREFR proposes an audio and text-driven facial animation method that incorporates the large language model GPT-2 #OTHEREFR to encode the textual information.", "The authors found that combined audio and text input yielded better results than audio-only or text-only model. However, the results still have problems regarding accurate lip-sync.", "Most closely related to our work is Face-Former #OTHEREFR which uses a self-supervised pretrained speech model that addresses the scarcity of available data in existing audio-visual datasets."], "text_after_citation": ["#OTHEREFR using a modified version of transformers to handle longer sequences of data.", "However, none of these methods can handle arbitrary variations in speech input while producing accurate lower and upper facial animation for multiple identities and emotions."], "citing_paper_content": {"title": "Facexhubert: Text-Less Speech-Driven E(X)Pressive 3D Facial Animation Synthesis Using Self-Supervised Speech Representation Learning", "abstract": "This paper presents FaceXHuBERT, a text-less speechdriven 3D facial animation generation method that allows to capture personalized and subtle cues in speech (e.g. identity, emotion and hesitation). It is also very robust to background noise and can handle audio recorded in a variety of situations (e.g. multiple people speaking). Recent approaches employ end-to-end deep learning taking into account both audio and text as input to generate facial animation for the whole face. However, scarcity of publicly available expressive audio-3D facial animation datasets poses a major bottleneck. The resulting animations still have issues regarding accurate lip-synching, expressivity, person-specific information and generalizability. We effectively employ self-supervised pretrained HuBERT model in the training process that allows us to incorporate both lexical and non-lexical information in the audio without using a large lexicon. Additionally, guiding the training with a binary emotion condition and speaker identity distinguishes the tiniest subtle facial motion. We carried out extensive objective and subjective evaluation in comparison to ground-truth and state-of-the-art work. A perceptual user study demonstrates that our approach produces superior results with respect to the realism of the animation 78% of the time in comparison to the state-of-the-art. In addition, our method is 4 times faster eliminating the use of complex sequential models such as transformers. We strongly recommend watching the supplementary video before reading the paper. We also provide the implementation and evaluation codes with a GitHub repository link."}, "cited_paper_content": {"title": "Capture, Learning, And Synthesis Of 3D Speaking Styles", "abstract": "Audio-driven 3D facial animation has been widely explored, but achieving realistic, human-like performance is still unsolved. This is due to the lack of available 3D datasets, models, and standard evaluation metrics. To address this, we introduce a unique 4D face dataset with about 29 minutes of 4D scans captured at 60 fps and synchronized audio from 12 speakers. We then train a neural network on our dataset that factors identity from facial motion. The learned model, VOCA (Voice Operated Character Animation) takes any speech signal as input\u2014even speech in languages other than English\u2014and realistically animates a wide range of adult faces. Conditioning on subject labels during training allows the model to learn a variety of realistic speaking styles. VOCA also provides animator controls to alter speaking style, identity-dependent facial shape, and pose (i.e. head, jaw, and eyeball rotations) during animation. To our knowledge, VOCA is the only realistic 3D facial animation model that is readily applicable to unseen subjects without retargeting. This makes VOCA suitable for tasks like in-game video, virtual reality avatars, or any scenario in which the speaker, speech, or language is not known in advance. We make the dataset and model available for research purposes at http://voca.is.tue.mpg.de."}, "keywords": ["model"], "citation_intent": "result"}
{"citing_id": "2304.14546v1", "cited_id": "1907.09448", "section_title": "A. Related Work", "citation": "An extension to the quasi-static Rayleigh fading channel has been addressed in #REFR , where a scheme based on low-density parity check (LDPC) codes using a belief propagation decoder has been proposed.", "text_before_citation": ["Since its initial conceptualization in #OTHEREFR , the URA paradigm has attracted considerable interest from the research community, giving rise to many candidate solutions tailored to various communication scenarios.", "Earlier works such as #OTHEREFR focus on information-theoretic limits and propose coding schemes for the additive white Gaussian noise (AWGN) channel setting.", "Later on, several practical schemes based on compressed sensing have been proposed for the same setting #OTHEREFR , #OTHEREFR ."], "text_after_citation": ["More recently the focus has shifted towards coding schemes for the massive MIMO (mMIMO) regime.", "Initially, the URA problem on a Rayleigh block-fading channel in a mMIMO setting was formulated in #OTHEREFR .", "The authors combined the coded compressed-sensig (CCS) paradigm with an outer tree code #OTHEREFR , and a covariance-based activity detection (AD) algorithm, showing the capability of achieving sum spectral efficiencies that scale with the coherence blocklength.", "In the operating regime where the coherence blocklength exceeds the number of active users, the state-of-the-art approaches for URA leverage the large blocklength to device two-step approaches, in which communication resources are split between user activity detection and channel estimation on one hand, and coherent detection on the other hand.", "Specifically, #OTHEREFR proposed the first pilot-aided communication scheme that employs multi-measurement-vector approximate message passing (MMV-AMP) for user activity detection and channel estimation, in combination with maximum ratio combining (MRC) outer Polar channel coding."], "citing_paper_content": {"title": "Bisparcs For Unsourced Random Access In Massive Mimo", "abstract": "This paper considers the massive MIMO unsourced random access problem in a quasi-static Rayleigh fading setting. The proposed coding scheme is based on a concatenation of a \"conventional\" channel code (such as, e.g., LDPC) serving as an outer code, and a sparse regression code (SPARC) serving as an inner code. The scheme combines channel estimation, singleuser decoding, and successive interference cancellation in a novel way. The receiver performs joint channel estimation and SPARC decoding via an instance of a bilinear generalized approximate message passing (BiGAMP) based algorithm, which leverages the intrinsic bilinear structure that arises in the considered communication regime. The detection step is followed by a per-user soft-input-soft-output (SISO) decoding of the outer channel code in combination with a successive interference cancellation (SIC) step. We show via numerical simulation that the resulting scheme achieves stat-of-the-art performance in the massive connectivity setting, while attaining comparatively low implementation complexity. Index Terms-Internet of Things (IoT), Unsourced Random Access, bilinear generalized approximate message passing (BiGAMP), massive MIMO."}, "cited_paper_content": {"title": "Energy Efficient Coded Random Access For The Wireless Uplink", "abstract": "We discuss the problem of designing channel access architectures for enabling fast, low-latency, grant-free and uncoordinated uplink for densely packed wireless nodes. Specifically, we study random-access codes, previously introduced for the AWGN multiple-access channel (MAC) by Polyanskiy'2017, in the practically more relevant case of users subject to Rayleigh fading, when channel gains are unknown to the decoder. We propose a random coding achievability bound, which we analyze both non-asymptotically (at finite blocklength) and asymptotically. As a candidate practical solution, we propose an explicit sparse-graph based coding scheme together with an alternating belief-propagation decoder. The latter's performance is found to be surprisingly close to the finite-blocklength bounds. Our main findings are twofold. First, just like in the AWGN MAC we see that jointly decoding large number of users leads to a surprising phase transition effect, where at spectral efficiencies below a critical threshold (5-15 bps/Hz depending on reliability) a perfect multi-user interference cancellation is possible. Second, while the presence of Rayleigh fading significantly increases the minimal required energy-per-bit $E_b/N_0$ (from about 0-2 dB to about 8-11 dB), the inherent randomization introduced by the channel makes it much easier to attain the optimal performance via iterative schemes. ::: In all, it is hoped that a principled definition of the random-access model together with our information-theoretic analysis will open the road towards unified benchmarking and comparison performance of various random-access solutions, such as the currently discussed candidates (MUSA, SCMA, RSMA) for the 5G/6G."}, "keywords": ["quasi-static Rayleigh fading"], "citation_intent": "method"}
{"citing_id": "2303.07502v1", "cited_id": "1905.01102", "section_title": "Learning The Parameters", "citation": "Continuing the previous work and with the same belief of developing the meta-learner without forgetting base classes, #REFR proposed a denoising auto-encoder network (DAE) in the form of graph neural network (GNN), which is responsible for regularizing the weight generating meta-model.", "text_before_citation": ["[82] proposed a dynamic algorithm by introducing an attention-based classification weight generator and a cosinesimilarity based ConvNet recognition model.", "The former is responsible for producing weights vectors for unseen classes by seeing a few examples from those classes.", "The authors argued that it is not feasible for a ConvNet model with a typical linear layer classifier to handle classification weights of both base and novel classes.", "They address this issue by proposing the second component that provides prediction by computing the cosine similarity between the generated weights and samples' representations.", "The combination of these two components allowed the authors to achieve fast adaptation for novel classes without forgetting the base classes."], "text_after_citation": ["Basically, DAE takes contaminated classification weights with Gaussian noise and generates classification weights based on the conditional distribution of seen data from both novel and base classes.", "Results showed that this model outperforms their previous model in #OTHEREFR .", "#OTHEREFR argued that embedding vectors are comparable with the weights of the last linear layer of the ConvNet classifier.", "Accordingly, instead of using two different components of meta-learner and base learner, they proposed a new process called weight imprinting that basically sets final layer weights based on normalized embeddings of the novel class's training samples.", "Figure 19 shows the overall flow of their model, in which an embedding function first trains on base classes then tunes on novel classes where imprint weights are used in the final layers."], "citing_paper_content": {"title": "Meta-Learning Approaches For Few-Shot Learning: A Survey Of Recent Advances", "abstract": "Despite its astounding success in learning deeper multi-dimensional data, the performance of deep learning declines on new unseen tasks mainly due to its focus on same-distribution prediction. Moreover, deep learning is notorious for poor generalization from few samples. Meta-learning is a promising approach that addresses these issues by adapting to new tasks with few-shot datasets. This survey first briefly introduces meta-learning and then investigates state-of-the-art meta-learning methods and recent advances in: (i) metric-based, (ii) memory-based, (iii), and learning-based methods. Finally, current challenges and insights for future researches are discussed."}, "cited_paper_content": {"title": "Generating Classification Weights With Gnn Denoising Autoencoders For Few-Shot Learning", "abstract": "Given an initial recognition model already trained on a set of base classes, the goal of this work is to develop a meta-model for few-shot learning. The meta-model, given as input some novel classes with few training examples per class, must properly adapt the existing recognition model into a new model that can correctly classify in a unified way both the novel and the base classes. To accomplish this goal it must learn to output the appropriate classification weight vectors for those two types of classes. To build our meta-model we make use of two main innovations: we propose the use of a Denoising Autoencoder network (DAE) that (during training) takes as input a set of classification weights corrupted with Gaussian noise and learns to reconstruct the target-discriminative classification weights. In this case, the injected noise on the classification weights serves the role of regularizing the weight generating meta-model. Furthermore, in order to capture the co-dependencies between different classes in a given task instance of our meta-model, we propose to implement the DAE model as a Graph Neural Network (GNN). In order to verify the efficacy of our approach, we extensively evaluate it on ImageNet based few-shot benchmarks and we report strong results that surpass prior approaches. The code and models of our paper will be published on: this https URL"}, "keywords": ["meta-learner", "graph neural network"], "citation_intent": "method"}
{"citing_id": "2304.03468v2", "cited_id": "2003.07743", "section_title": "Dataset", "citation": "We first adopted one of the most representative benchmark datasets DBP15K(EN-FR) #REFR , which is widely adopted in previous works.", "text_before_citation": [], "text_after_citation": ["To measure the real effect of the existing EA methods on HHKGs, we also conducted extensive experiments on our new proposed datasets: ICEWS-WIKI and ICEWS-YAGO.", "The statistics of these selected datasets are summarized in Table 1 ."], "citing_paper_content": {"title": "Rethinking Gnn-Based Entity Alignment On Heterogeneous Knowledge Graphs: New Datasets And A New Method", "abstract": "The development of knowledge graph (KG) applications has led to a rising need for entity alignment (EA) between heterogeneous KGs that are extracted from various sources. Recently, graph neural networks (GNNs) have been widely adopted in EA tasks due to GNNs' impressive ability to capture structure information. However, we have observed that the oversimplified settings of the existing common EA datasets are distant from real-world scenarios, which obstructs a full understanding of the advancements achieved by recent methods. This phenomenon makes us ponder: Do existing GNN-based EA methods really make great progress? In this paper, to study the performance of EA methods in realistic settings, we focus on the alignment of highly heterogeneous KGs (HHKGs) (e.g., event KGs and general KGs) which are different with regard to the scale and structure, and share fewer overlapping entities. First, we sweep the unreasonable settings, and propose two new HHKG datasets that closely mimic real-world EA scenarios. Then, based on the proposed datasets, we conduct extensive experiments to evaluate previous representative EA methods, and reveal interesting findings about the progress of GNN-based EA methods. We find that the structural information becomes difficult to exploit but still valuable in aligning HHKGs. This phenomenon leads to inferior performance of existing EA methods, especially GNNbased methods. Our findings shed light on the potential problems resulting from an impulsive application of GNN-based methods as a panacea for all EA datasets. Finally, we introduce a simple but effective method: Simple-HHEA, which comprehensively utilizes entity name, structure, and temporal information. Experiment results show Simple-HHEA outperforms previous models on HHKG datasets."}, "cited_paper_content": {"title": "A Benchmarking Study Of Embedding-Based Entity Alignment For Knowledge Graphs", "abstract": "Entity alignment seeks to find entities in different knowledge graphs (KGs) that refer to the same real-world object. Recent advancement in KG embedding impels the advent of embedding-based entity alignment, which encodes entities in a continuous embedding space and measures entity similarities based on the learned embeddings. In this paper, we conduct a comprehensive experimental study of this emerging field. This study surveys 23 recent embedding-based entity alignment approaches and categorizes them based on their techniques and characteristics. We further observe that current approaches use different datasets in evaluation, and the degree distributions of entities in these datasets are inconsistent with real KGs. Hence, we propose a new KG sampling algorithm, with which we generate a set of dedicated benchmark datasets with various heterogeneity and distributions for a realistic evaluation. This study also produces an open-source library, which includes 12 representative embedding-based entity alignment approaches. We extensively evaluate these approaches on the generated datasets, to understand their strengths and limitations. Additionally, for several directions that have not been explored in current approaches, we perform exploratory experiments and report our preliminary findings for future studies. The benchmark datasets, open-source library and experimental results are all accessible online and will be duly maintained."}, "keywords": ["representative benchmark datasets"], "citation_intent": "method"}
{"citing_id": "2303.12419v1", "cited_id": "1412.6596", "section_title": "Learning With Noisy Labels", "citation": "Bootstrapping #REFR combines the noisy label and the DNN prediction as a soft pseudo-label to replace the original label.", "text_before_citation": ["Supervised training of deep learning models requires precisely labeled datasets and noisy labels can significantly degrade the generalization of models.", "Therefore, various methods were proposed to improve robustness against noisy labels.", "The typical algorithms for combating noisy labels including adding regularization #OTHEREFR , designing robust loss functions #OTHEREFR , selecting possible clean samples #OTHEREFR and correcting the labels #OTHEREFR .", "We mainly introduce the label correction approaches which are most related to this work.", "Specifically, label correction algorithms aim to identify and correct suspicious labels in an iterative framework."], "text_after_citation": ["Furthermore, Dynamic bootstrapping #OTHEREFR use a Beta-Mixture-Model to combine label dynamically for every sample.", "Differently, Joint Optimization #OTHEREFR trained their model with a large learning rate for several epochs firstly, and then generated pseudo-labels by averaging the predictions of that epochs.", "Recently, SELFIE #OTHEREFR and Ada-Corr #OTHEREFR selectively combine the labels of noisy data for corresponding to true labels with a high probability.", "However, almost all existing noisy label studies, focusing on the classification task, cannot be adapted to cross-modal matching directly in which the noisy label refers to the alignment errors between data pairs rather than the categorical labeling errors.", "The label of cross-modal matching can be consider as y i \u2208 {0, 1}, where data pair is positively correlated (y i = 1) or not (y i = 0)."], "citing_paper_content": {"title": "Bicro: Noisy Correspondence Rectification For Multi-Modality Data Via Bi-Directional Cross-Modal Similarity Consistency", "abstract": "As one of the most fundamental techniques in multimodal learning, cross-modal matching aims to project various sensory modalities into a shared feature space. To achieve this, massive and correctly aligned data pairs are required for model training. However, unlike unimodal datasets, multimodal datasets are extremely harder to collect and annotate precisely. As an alternative, the co-occurred data pairs (e.g., image-text pairs) collected from the Internet have been widely exploited in the area. Unfortunately, the cheaply collected dataset unavoidably contains many mismatched data pairs, which have been proven to be harmful to the model's performance. To address this, we propose a general framework called BiCro (Bidirectional Cross-modal similarity consistency), which can be easily integrated into existing cross-modal matching models and improve their robustness against noisy data. Specifically, BiCro aims to estimate soft labels for noisy data pairs to reflect their true correspondence degree. The basic idea of BiCro is motivated by that-taking image-text matching as an example-similar images should have similar textual descriptions and vice versa. Then the consistency of these two similarities can be recast as the estimated soft labels to train the matching model. The experiments on three popular cross-modal matching datasets demonstrate that our method significantly improves the noise-robustness of various matching models, and surpass the state-of-the-art by a clear margin.The code is available at https://github.com/xu5zhao/BiCro."}, "cited_paper_content": {"title": "Training Deep Neural Networks On Noisy Labels With Bootstrapping", "abstract": "Current state-of-the-art deep learning systems for visual object recognition and detection use purely supervised training with regularization such as dropout to avoid overfitting. The performance depends critically on the amount of labeled examples, and in current practice the labels are assumed to be unambiguous and accurate. However, this assumption often does not hold; e.g. in recognition, class labels may be missing; in detection, objects in the image may not be localized; and in general, the labeling may be subjective. In this work we propose a generic way to handle noisy and incomplete labeling by augmenting the prediction objective with a notion of consistency. We consider a prediction consistent if the same prediction is made given similar percepts, where the notion of similarity is between deep network features computed from the input data. In experiments we demonstrate that our approach yields substantial robustness to label noise on several datasets. On MNIST handwritten digits, we show that our model is robust to label corruption. On the Toronto Face Database, we show that our model handles well the case of subjective labels in emotion recognition, achieving state-of-the- art results, and can also benefit from unlabeled face images with no modification to our method. On the ILSVRC2014 detection challenge data, we show that our approach extends to very deep networks, high resolution images and structured outputs, and results in improved scalable detection."}, "keywords": ["DNN prediction", "noisy label"], "citation_intent": "background"}
{"citing_id": "2303.14351v1", "cited_id": "2002.08811", "section_title": "I. Introduction", "citation": "With increasing demands and diverse services, several other projects like SAT5G and SATis5 are involving in NTN technologies, especially in LEOs #REFR .", "text_before_citation": ["Low Earth orbit (LEO) satellite constellation has been developed rapidly, bringing a large number of prospects in sixthgeneration (6G) non-terrestrial network (NTN) communications #OTHEREFR - #OTHEREFR .", "Lots of global organization and industry, such as SpaceX, Amazon, OneWeb, and TeleSAT have announced investment for extensive services and advanced researches to provide globalcoverage applications #OTHEREFR - #OTHEREFR .", "The NTN specification is specified by 3rd Generation Partnership Project (3GPP) since Release 15, aiming for global-coverage and throughput-aware services #OTHEREFR ."], "text_after_citation": ["LEO satellites possess several benefits compared to the conventional satellite communications of medium Earth orbit (MEO), high Earth orbit (HEO) and geostationary orbit (GEO) #OTHEREFR .", "Comparatively, lower LEO height allows lower latency, alleviated pathloss and lower deployment cost.", "To conquer the huge pathloss compared to terrestrial channels, LEOs have to conduct multiple narrow beamforming to concentrate power with high beam-gain to the desired users in several directions #OTHEREFR .", "Moreover, massive LEOs are deployed and controlled by the ground gateway to support seamless coverage not covered by terrestrial infrastructure #OTHEREFR , #OTHEREFR .", "Nevertheless, under such dense LEO networks, limited resources should be appropriately assigned to prevent severe backdrops of inter-LEO, inter-beam and Doppler-induced interferences from high-mobility #OTHEREFR - #OTHEREFR , which becomes compellingly imperative."], "citing_paper_content": {"title": "Hierarchical Multi-Agent Multi-Armed Bandit For Resource Allocation In Multi-Leo Satellite Constellation Networks", "abstract": "Low Earth orbit (LEO) satellite constellation is capable of providing global coverage area with high-rate services in the next sixth-generation (6G) non-terrestrial network (NTN). Due to limited onboard resources of operating power, beams, and channels, resilient and efficient resource management has become compellingly imperative under complex interference cases. However, different from conventional terrestrial base stations, LEO is deployed at considerable height and under high mobility, inducing substantially long delay and interference during transmission. As a result, acquiring the accurate channel state information between LEOs and ground users is challenging. Therefore, we construct a framework with a twoway transmission under unknown channel information and no data collected at long-delay ground gateway. In this paper, we propose hierarchical multi-agent multi-armed bandit resource allocation for LEO constellation (mmRAL) by appropriately assigning available radio resources. LEOs are considered as collaborative multiple macro-agents attempting unknown trials of various actions of microagents of respective resources, asymptotically achieving suitable allocation with only throughput information. In simulations, we evaluate mmRAL in various cases of LEO deployment, serving numbers of users and LEOs, hardware cost and outage probability. Benefited by efficient and resilient allocation, the proposed mmRAL system is capable of operating in homogeneous or heterogeneous orbital planes or constellations, achieving the highest throughput performance compared to the existing benchmarks in open literature."}, "cited_paper_content": {"title": "Satellite Communications In The New Space Era: A Survey And Future Challenges", "abstract": "Satellite communications have recently entered a period of renewed interest motivated by technological advances and nurtured through private investment and ventures. The present survey aims at capturing the state of the art in SatComs, while highlighting the most promising open research topics. Firstly, the main innovation drivers are motivated, such as new constellation types, on-board processing capabilities, nonterrestrial networks and space-based data collection/processing. Secondly, the most promising applications are described i.e. 5G integration, space communications, Earth observation, aeronautical and maritime tracking and communication. Subsequently, an in-depth literature review is provided across five axes: i) system aspects, ii) air interface, iii) medium access, iv) networking, v) testbeds & prototyping. Finally, a number of future challenges and the respective open research topics are described."}, "keywords": ["LEOs", "SAT5G"], "citation_intent": "background"}
{"citing_id": "2304.02733v1", "cited_id": "1703.00443", "section_title": "Problem Formulation And Approach", "citation": "Our approach to solve the above problem is based on an end-to-end trainable framework that includes a differentiable optimization layer #REFR formulated by the proposed att-CLF based QP.", "text_before_citation": ["Problem setup.", "Given (i) a sensor measurement z z z \u2208 R d (e.g., front-view RGB images of the vehicle in this work), where d \u2208 N is its dimension, (ii) vehicle dynamics (1), (iii) a nominal controller h (x x x, z z z) = u u u (such as a model predictive controller), and (iv) a learning-based end-to-end perception-to-control model h(z z z|\u03b8 ) = u u u, parametrized by \u03b8 , our goal is to find optimal parameters:", "EQUATION", "where l is a similarity measure (or a loss function).", "We also wish to equip the learning-based endto-end controller h with stability properties which we will describe in the following. Approach."], "text_after_citation": ["The proposed att-CLF is observation dependent, and thus, it is adaptive to changing environments.", "The model takes front-view RGB images z z z as inputs whose features are extracted by CNN.", "In order to enforce temporal coherency and prevent the model from being susceptible to sudden input change (e.g., affected by sunlight), we employ LSTM after the CNN.", "The outputs of the LSTM are used to construct the att-CLF, as well as to determine other optimization parameters in the att-CLF based QP.", "The outputs of the QP are the controls u u u of the vehicle."], "citing_paper_content": {"title": "Learning Stability Attention In Vision-Based End-To-End Driving Policies", "abstract": "Modern end-to-end learning systems can learn to explicitly infer control from perception. However, it is difficult to guarantee stability and robustness for these systems since they are often exposed to unstructured, high-dimensional, and complex observation spaces (e.g., autonomous driving from a stream of pixel inputs). We propose to leverage control Lyapunov functions (CLFs) to equip end-to-end vision-based policies with stability properties and introduce stability attention in CLFs (att-CLFs) to tackle environmental changes and improve learning flexibility. We also present an uncertainty propagation technique that is tightly integrated into att-CLFs. We demonstrate the effectiveness of att-CLFs via comparison with classical CLFs, model predictive control, and vanilla end-to-end learning in a photo-realistic simulator and on a real full-scale autonomous vehicle."}, "cited_paper_content": {"title": "Optnet: Differentiable Optimization As A Layer In Neural Networks", "abstract": "This paper presents OptNet, a network architecture that integrates optimization problems (here, specifically in the form of quadratic programs) as individual layers in larger end-to-end trainable deep networks. These layers allow complex dependencies between the hidden states to be captured that traditional convolutional and fully-connected layers are not able to capture. In this paper, we develop the foundations for such an architecture: we derive the equations to perform exact differentiation through these layers and with respect to layer parameters; we develop a highly efficient solver for these layers that exploits fast GPU-based batch solves within a primal-dual interior point method, and which provides backpropagation gradients with virtually no additional cost on top of the solve; and we highlight the application of these approaches in several problems. In one particularly standout example, we show that the method is capable of learning to play Sudoku given just input and output games, with no a priori information about the rules of the game; this task is virtually impossible for other neural network architectures that we have experimented with, and highlights the representation capabilities of our approach."}, "keywords": ["differentiable optimization layer"], "citation_intent": "method"}
{"citing_id": "2303.05892v1", "cited_id": "1504.00325", "section_title": "Main Results", "citation": "While C-OVD is not our primary concern, we include the corresponding methods for reference and report their performance when only COCO Captions #REFR is available.", "text_before_citation": ["For completeness, we include two related benchmarks: Zero-Shot Detection (ZSD) #OTHEREFR and Weakly Supervised Detection (WSD) #OTHEREFR .", "Our method mainly focuses on the V-OVD benchmark and the G-OVD benchmark. As shown in tab.", "2, our OADP framework achieves 30.0 mAP N 50 on the V-OVD benchmark.", "RegionCLIP* #OTHEREFR uses CLIP #OTHEREFR as the pretrained weight, thus adhering to the V-OVD benchmark.", "Some V-OVD methods #OTHEREFR are not included because they rely on large-scale detection and image-text datasets and cannot be compared fairly."], "text_after_citation": ["Under such constraint, the performance of the caption-based methods is relatively lower than the V-OVD methods, even if additional caption data is used.", "Moreover, our OADP framework is perpendicular to the caption-based methods and has the potential to achieve higher performance using captions.", "For the G-OVD benchmark, we generate pseudo labels for novel categories as described in sec.", "#OTHEREFR Most of the methods under the OV-LVIS setting adhere to the V-OVD benchmark, so we conduct experiments on the V-OVD benchmark only.", "For the object detection task, our OADP framework achieves 21.9 AP r , surpassing DetPro #OTHEREFR by 1.1 AP r ."], "citing_paper_content": {"title": "Object-Aware Distillation Pyramid For Open-Vocabulary Object Detection", "abstract": "Open-vocabulary object detection aims to provide object detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbitrary text queries. Previous methods adopt knowledge distillation to extract knowledge from Pretrained Vision-and-Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE) module and a Distillation Pyramid (DP) mechanism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and complete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to current methods. Especially on the MS-COCO dataset, our OADP framework reaches 35.6 mAP N 50 , surpassing the current state-of-the-art method by 3.3 mAP N 50. Code is released at https://github.com/LutingWang/OADP."}, "cited_paper_content": {"title": "Microsoft Coco Captions: Data Collection And Evaluation Server", "abstract": "In this paper we describe the Microsoft COCO Caption dataset and evaluation server. When completed, the dataset will contain over one and a half million captions describing over 330,000 images. For the training and validation images, five independent human generated captions will be provided. To ensure consistency in evaluation of automatic caption generation algorithms, an evaluation server is used. The evaluation server receives candidate captions and scores them using several popular metrics, including BLEU, METEOR, ROUGE and CIDEr. Instructions for using the evaluation server are provided."}, "keywords": ["COCO Captions"], "citation_intent": "method"}
{"citing_id": "2304.04537v1", "cited_id": "1901.09237", "section_title": "Related Work", "citation": "They first divided an image into 64 \u00d7 64 patches, predicted fake patches by their model, and calculated the threshold by dividing the number of fake patches to the total number of patches #REFR . In 2020, Jane et al.", "text_before_citation": ["For analyzing image textures and recording long-term information about image textures, introduced the Gram-Net network, derived by adding 6 Gram blocks to the ResNet18 architecture.", "Each Gram block in the ResNet18 architecture contained one convolution layer, one Gram matrix #OTHEREFR layer to extract image texture properties, two convolution layers, batch normalization, Relu, and one pooling layer #OTHEREFR .", "Some works #OTHEREFR used facial patches for diagnosis. In 2018, Jain et al.", "introduced an architecture consisting of 5 convolutional layers and 2 fully connected layers, and a wide residual connection in order to add the output of the second layer to the fifth layer.", "A thresholdbased approach and prediction based on threshold were used to classify images as real or fake."], "text_after_citation": ["#OTHEREFR introduced a three-level patch-based approach for fake detection.In the first level, they divided an image into 64 \u00d7 64 blocks, and obtained a vector for each image.", "They also trained a SVM with RBF kernel with these vectors to determine whether an image was real or altered.", "On the second level, weights of the previous layer were frozen, and if an image has been altered, the network detected that images have been retouched or generated by a GAN.", "Finally, in the third level, when an image was produced by a GAN, weights of the previous level were frozen, and at this level, the type of GANs was detected.", "Bharati et al, in 2016, proposed considering four facial patches, including the right periocular, left periocular, nose, and mouth. They trained a Deep Boltzmann machine for each patch."], "citing_paper_content": {"title": "Deepfake Detection Of Occluded Images Using A Patch-Based Approach", "abstract": "DeepFake involves the use of deep learning and artificial intelligence techniques to produce or change video and image contents typically generated by GANs. Moreover, it can be misused and leads to fictitious news, ethical and financial crimes, and also affects the performance of facial recognition systems. Thus, detection of real or fake images is significant specially to authenticate originality of people's images or videos. One of the most important challenges in this topic is obstruction that decreases the system precision. In this study, we present a deep learning approach using the entire face and face patches to distinguish real/fake images in the presence of obstruction with a three-path decision: first entire-face reasoning, second a decision based on the concatenation of feature vectors of face patches, and third a majority vote decision based on these features. To test our approach, new datasets including real and fake images are created. For producing fake images, StyleGAN and StyleGAN2 are trained by FFHQ images and also StarGAN and PGGAN are trained by CelebA images. The CelebA and FFHQ datasets are used as real images. The proposed approach reaches higher results in early epochs than other methods and increases the SoTA results by 0.4%-7.9% in the different built data-sets. Also, we have shown in experimental results that weighing the patches may improve accuracy."}, "cited_paper_content": {"title": "On Detecting Gans And Retouching Based Synthetic Alterations", "abstract": "Digitally retouching images has become a popular trend, with people posting altered images on social media and even magazines posting flawless facial images of celebrities. Further, with advancements in Generative Adversarial Networks (GANs), now changing attributes and retouching have become very easy. Such synthetic alterations have adverse effect on face recognition algorithms. While researchers have proposed to detect image tampering, detecting GANs generated images has still not been explored. This paper proposes a supervised deep learning algorithm using Convolutional Neural Networks (CNNs) to detect synthetically altered images. The algorithm yields an accuracy of 99.65% on detecting retouching on the ND-IIITD dataset. It outperforms the previous state of the art which reported an accuracy of 87% on the database. For distinguishing between real images and images generated using GANs, the proposed algorithm yields an accuracy of 99.83%."}, "keywords": ["fake patches"], "citation_intent": "background"}
{"citing_id": "2304.04326v1", "cited_id": "1610.02136", "section_title": "Ood Detector Calibration", "citation": "To address this issue, we use an Out-of-Distribution (OoD) detector #REFR to identify a subset of D P that maps very close to the private training set D i T . Figure 3 .", "text_before_citation": ["In traditional distillation methods, the labels generated by the teacher model on the public dataset are used to train the student model #OTHEREFR .", "However, all the public dataset samples are used irrespective of their similarity to the local dataset."], "text_after_citation": ["The OoD detector used in our IDKD framework for decentralized training.", "After local convergence, a calibration dataset (can be the public dataset or a separate dataset) and the private dataset are used as OoD data and ID data respectively. This is used to identify the optimal threshold.", "OoD detector identifies the samples in the public set D P that look like In-Distribution (ID) data to the local model on each node.", "Among different OoD detectors proposed in literature #OTHEREFR , we choose the maximum softmax probability (MSP) detector #OTHEREFR because of its simplicity and low computational overhead. Figure 3 visualizes the MSP detector.", "The samples with confidence (i.e., softmax probability) greater than a set threshold t are classified as ID by the MSP detector The hyperparameter t i.e, threshold is tuned based on the True Positive Rate (TPR) of the OoD detector."], "citing_paper_content": {"title": "Homogenizing Non-Iid Datasets Via In-Distribution Knowledge Distillation For Decentralized Learning", "abstract": "Decentralized learning enables serverless training of deep neural networks (DNNs) in a distributed manner on multiple nodes. This allows for the use of large datasets, as well as the ability to train with a wide variety of data sources. However, one of the key challenges with decentralized learning is heterogeneity in the data distribution across the nodes. In this paper, we propose In-Distribution Knowledge Distillation (IDKD) to address the challenge of heterogeneous data distribution. The goal of IDKD is to homogenize the data distribution across the nodes. While such data homogenization can be achieved by exchanging data among the nodes sacrificing privacy, IDKD achieves the same objective using a common public dataset across nodes without breaking the privacy constraint. This public dataset is different from the training dataset and is used to distill the knowledge from each node and communicate it to its neighbors through the generated labels. With traditional knowledge distillation, the generalization of the distilled model is reduced because all the public dataset samples are used irrespective of their similarity to the local dataset. Thus, we introduce an Out-of-Distribution (OoD) detector at each node to label a subset of the public dataset that maps close to the local training data distribution. Finally, only labels corresponding to these subsets are exchanged among the nodes and with appropriate label averaging each node is finetuned on these data subsets along with its local data. Our experiments on multiple image classification datasets and graph topologies show that the proposed IDKD scheme is more effective than traditional knowledge distillation and achieves state-of-the-art generalization performance on heterogeneously distributed data with minimal communication overhead. Figure 1. Overview of decentralized training. Here networks are connected in a ring topology. Each node has its own private dataset. This dataset is not shared between the nodes, rather model weights/gradients are shared between the neighbors during training. However, such a setup struggles to converge when private datasets are heterogeneous i.e. non-identical and independently distributed (non-IID)."}, "cited_paper_content": {"title": "A Baseline For Detecting Misclassified And Out-Of-Distribution Examples In Neural Networks", "abstract": "We consider the two related problems of detecting if an example is misclassified or out-of-distribution. We present a simple baseline that utilizes probabilities from softmax distributions. Correctly classified examples tend to have greater maximum softmax probabilities than erroneously classified and out-of-distribution examples, allowing for their detection. We assess performance by defining several tasks in computer vision, natural language processing, and automatic speech recognition, showing the effectiveness of this baseline across all. We then show the baseline can sometimes be surpassed, demonstrating the room for future research on these underexplored detection tasks."}, "keywords": ["private training", "Distribution"], "citation_intent": "method"}
{"citing_id": "2304.09084v1", "cited_id": "1812.03871", "section_title": "Convergence", "citation": "But since the delay is bounded and assumptions on the loss function hold, the sequence produced by our Distributed Gradient update rule converges to the unique minimizer. (Theorem 1 #REFR ).", "text_before_citation": ["Each of them will participate in the update of the global parameters, by adding his computed loss, we can conclude that the final loss of our algorithm will be :", "L (w) = \u2211 M i=1 L i (w)", ", with L i (w) the total loss created by the DO i .", "We know that the time taken by the DO to compute the gradient, depend on a finite block of interaction, so we know that there exist D < \u221e, such that for any DO i , the delay d i k < D.", "Assume that each L i is differentiable and that \u2207L i is 1 L \u2212 cococercive."], "text_after_citation": ["Moreover, in our experiments, we will use the same loss function as SAROS, we can conclude that both algorithms should converge to the same minimizer."], "citing_paper_content": {"title": "Drift: A Federated Recommender System With Implicit Feedback On The Items", "abstract": "Nowadays there are more and more items available online, this makes it hard for users to find items that they like. Recommender systems aim to find the item who best suits the user, using his historical interactions. Depending on the context, these interactions may be more or less sensitive and collecting them brings an important problem concerning the users' privacy. Federated systems have shown that it is possible to make accurate and efficient recommendations without storing users' personal information. However, these systems use instantaneous feedback from the user. In this report, we propose DRIFT, a federated architecture for recommender systems, using implicit feedback. Our learning model is based on a recent algorithm for recommendation with implicit feedbacks SAROS [4]. We aim to make recommendations as precise as SAROS, without compromising the users' privacy. In this report we show that thanks to our experiments, but also thanks to a theoretical analysis on the convergence. We have shown also that the computation time has a linear complexity with respect to the number of interactions made. Finally, we have shown that our algorithm is secure, and participants in our federated system cannot guess the interactions made by the user, except DOs that have the item involved in the interaction. Acknowledgement I would like to express my sincere gratitude to Marta Soare, Radu Ciucanu, and Massih-Reza Amini for their invaluable assistance throughout the internship. I would also like to express my sincere thanks to Aleksandra Burashnikova, for all the help she gave me in understanding her algorithm."}, "cited_paper_content": {"title": "Asynchronous Distributed Learning With Sparse Communications And Identification", "abstract": "In this paper, we present an asynchronous optimization algorithm for distributed learning, that efficiently reduces the communications between a master and working machines by randomly sparsifying the local updates. This sparsification allows to lift the communication bottleneck often present in distributed learning setups where computations are performed by workers on local data while a master machine coordinates their updates to optimize a global loss. We prove that despite its sparse asynchronous communications, our algorithm allows for a fixed stepsize and benefits from a linear convergence rate in the strongly convex case. Moreover, for $\\ell_1$-regularized problems, this algorithm identifies near-optimal sparsity patterns, so that all communications eventually become sparse. We furthermore leverage on this identification to improve our sparsification technique. We illustrate on real and synthetic data that this algorithm converges faster in terms of data exchanges."}, "keywords": ["loss function hold", "Distributed Gradient"], "citation_intent": "background"}
{"citing_id": "2303.05246v1", "cited_id": "1802.00420", "section_title": "Introduction", "citation": "They suggest that adaptive ODE solvers, used to solve the underlying initial value problems, cause gradient obfuscation #REFR , which, in turn, causes weaker adversarial attacks to fail.", "text_before_citation": ["As standard neural networks suffer from severely reduced accuracies when trained for robustness, novel architectures with inherent robustness properties have recently received increasing attention #OTHEREFR .", "Neural Ordinary Differential Equations One particularly interesting such architecture are neural ODEs (NODEs) .", "Built around solving initial value problems with learned dynamics, they are uniquely suited to time-series-based problems #OTHEREFR but have also been successfully applied to image classification .", "More importantly, NODEs have been observed to exhibit inherent robustness properties against adversarial attacks #OTHEREFR Kang et al., 2021; #OTHEREFR .", "However, recently found that this robustness is greatly diminished against stronger attacks."], "text_after_citation": ["This highlights the need for formal robustness guarantees to rigorously evaluate the true robustness of a model or architecture.", "Robustness Verification For standard neural networks, many robustness verification methods have been proposed #OTHEREFR .", "One particularly successful class of such methods #OTHEREFR propagates convex shapes through the neural network that capture the reachable sets of every neuron's values and uses them to check"], "citing_paper_content": {"title": "Efficient Certified Training And Robustness Verification Of Neural Odes", "abstract": "Neural Ordinary Differential Equations (NODEs) are a novel neural architecture, built around initial value problems with learned dynamics which are solved during inference. Thought to be inherently more robust against adversarial perturbations, they were recently shown to be vulnerable to strong adversarial attacks, highlighting the need for formal guarantees. However, despite significant progress in robustness verification for standard feed-forward architectures, the verification of high dimensional NODEs remains an open problem. In this work, we address this challenge and propose GAINS, an analysis framework for NODEs combining three key ideas: (i) a novel class of ODE solvers, based on variable but discrete time steps, (ii) an efficient graph representation of solver trajectories, and (iii) a novel abstraction algorithm operating on this graph representation. Together, these advances enable the efficient analysis and certified training of high-dimensional NODEs, by reducing the runtime from an intractable O(exp(d) + exp(T)) to O(d + T 2 log 2 T) in the dimensionality d and integration time T. In an extensive evaluation on computer vision (MNIST and FMNIST) and time-series forecasting (PHYSIO-NET) problems, we demonstrate the effectiveness of both our certified training and verification methods."}, "cited_paper_content": {"title": "Obfuscated Gradients Give A False Sense Of Security: Circumventing Defenses To Adversarial Examples", "abstract": "We identify obfuscated gradients as a phenomenon that leads to a false sense of security in defenses against adversarial examples. While defenses that cause obfuscated gradients appear to defeat optimization-based attacks, we find defenses relying on this effect can be circumvented. ::: For each of the three types of obfuscated gradients we discover, we describe indicators of defenses exhibiting this effect and develop attack techniques to overcome it. In a case study, examining all defenses accepted to ICLR 2018, we find obfuscated gradients are a common occurrence, with 7 of 8 defenses relying on obfuscated gradients. Using our new attack techniques, we successfully circumvent all 7 of them."}, "keywords": ["adaptive ODE solvers", "gradient obfuscation"], "citation_intent": "background"}
{"citing_id": "2304.04827v1", "cited_id": "2002.11776", "section_title": "Applying Ordinal Motifs To Data Sets", "citation": "We demonstrate the applicability of ordinal motifs on real-world data using a medium sized formal context: the spices planner data set #REFR .", "text_before_citation": [], "text_after_citation": ["This context contains 56 meals (objects) and 37 spices and food categories (attributes).", "The incidence encodes that a spice is recommended to be used to cook a meal or a meal belongs to a food category. The context has 531 formal concepts.", "We conduct our experiment on the dual context, i.e., K d := (M, G, I d ), to derive ordinal motifs within the spices and food categories.", "For our application we employ the standard scales #OTHEREFR , as they are the most commonly used.", "We recall their definitions, where [n] := {1,...,n} and K 1 | K 2 is the context apposition operator."], "citing_paper_content": {"title": "Ordinal Motifs In Lattices", "abstract": "Lattices are a commonly used structure for the representation and analysis of relational and ontological knowledge. In particular, the analysis of these requires a decomposition of a large and high-dimensional lattice into a set of understandably large parts. With the present work we propose /ordinal motifs/ as analytical units of meaning. We study these ordinal substructures (or standard scales) through order-embeddings and (full) scale-measures of formal contexts from the field of formal concept analysis. We show that the underlying decision problems are NP-complete and provide results on how one can incrementally identify ordinal motifs to save computational effort. Accompanying our theoretical results, we demonstrate how ordinal motifs can be leveraged to achieve textual explanations based on principles from human computer interaction."}, "cited_paper_content": {"title": "Knowledge Cores In Large Formal Contexts", "abstract": "Knowledge computation tasks are often infeasible for large data sets. This is in particular true when deriving knowledge bases in formal concept analysis (FCA). Hence, it is essential to come up with techniques to cope with this problem. Many successful methods are based on random processes to reduce the size of the investigated data set. This, however, makes them hardly interpretable with respect to the discovered knowledge. Other approaches restrict themselves to highly supported subsets and omit rare and interesting patterns. An essentially different approach is used in network science, called $k$-cores. These are able to reflect rare patterns if they are well connected in the data set. In this work, we study $k$-cores in the realm of FCA by exploiting the natural correspondence to bi-partite graphs. This structurally motivated approach leads to a comprehensible extraction of knowledge cores from large formal contexts data sets."}, "keywords": ["ordinal motifs", "real-world data"], "citation_intent": "method"}
{"citing_id": "2303.10554v1", "cited_id": "1009.3780", "section_title": "Metric Regularity On Riemannian Manifolds", "citation": "Then #REFR implies that g(p \u2032\u2032 ) \u2212 g(p \u2032 ) e \u2264 \u01ebd(p \u2032 , p \u2032\u2032 ) for all p \u2032 , p \u2032\u2032 \u2208 B \u03b4 (p).", "text_before_citation": ["f (p \u2032 ) \u2212 f (p \u2032\u2032 ) \u2212 Df (p)[exp \u22121 p p \u2032 \u2212 exp \u22121 p p \u2032\u2032 ] e \u2264 1 0 Df (\u03b3 p \u2032\u2032 p \u2032 (t))[\u03b3 p \u2032\u2032 p \u2032 (t)] \u2212 Df (p)[exp \u22121 p p \u2032 \u2212 exp \u22121 p p \u2032\u2032 ] e dt, p \u2032 , p \u2032\u2032 \u2208 B \u03b4 (p). (15)", "On the other hand, from Proposition 4 with p = p \u2032\u2032 and q = p \u2032 we hav\u0117", "\u03b3 p \u2032\u2032 p \u2032 (t) = exp \u22121 \u03b3 p \u2032\u2032 p \u2032 (t) p \u2032 \u2212 exp \u22121 \u03b3 p \u2032\u2032 p \u2032 (t) p \u2032\u2032 , t \u2208 [0, 1].", "Taking into account this equality and #OTHEREFR", "with p = \u03b3 p \u2032\u2032 p \u2032 (t), (15) yields f (p \u2032 ) \u2212 f (p \u2032\u2032 ) \u2212 Df (p)(exp \u22121 p p \u2032 \u2212 exp \u22121 p p \u2032\u2032 ) e \u2264 \u01ebd(p \u2032 , p \u2032\u2032 ), p \u2032 , p \u2032\u2032 \u2208 B \u03b4 (p)."], "text_after_citation": ["Since \u01eb > 0 was chosen arbitrarily we can conclude that lip(g;p) = 0.", "Finally, to obtain the desired equality just use Lemma 1 with g as in #OTHEREFR and \u03a6 = f + F ."], "citing_paper_content": {"title": "Inexact Newton Methods For Solving Generalized Equations On Riemannian Manifolds", "abstract": "The local convergence of an inexact Newton method is studied for solving generalized equations on Riemannian manifolds by using the metric regularity property which is explored as well. Under suitable conditions and without any additional geometric assumptions, local convergence results with linear and quadratic rate and a semi-local convergence result are obtained for the proposed method. Finally, the theory can be applied to problems of finding a singularity of the sum of two vector fields."}, "cited_paper_content": {"title": "Algorithms For The Split Variational Inequality Problem", "abstract": "We propose a prototypical Split Inverse Problem (SIP) and a new variational problem, called the Split Variational Inequality Problem (SVIP), which is a SIP. It entails finding a solution of one inverse problem (e.g., a Variational Inequality Problem (VIP)), the image of which under a given bounded linear transformation is a solution of another inverse problem such as a VIP. We construct iterative algorithms that solve such problems, under reasonable conditions, in Hilbert space and then discuss special cases, some of which are new even in Euclidean space."}, "keywords": ["\u2212 g(p \u2032", "e \u2264"], "citation_intent": "background"}
{"citing_id": "2303.15919v1", "cited_id": "1603.07285", "section_title": "Given An Input Feature Map", "citation": "The input feature map is always implicitly padded on each side. We refer to Dumoulin and Visin #REFR for illustrations.", "text_before_citation": ["However, it remains essential to predetermine the structure of hyperbolic feature maps.", "Extension to the transposed setting The transposed convolutional layer is usually used in encoderdecoder architectures for up-sampling.", "It is straightforward to generalize to hyperbolic space, as it can be defined using a standard convolutional layer.", "A convolutional layer carries out a transposed convolution when the correct local connectivity is established by inserting zeros at certain positions.", "Specifically, when stride s > 1, then s \u2212 1 zero vectors are inserted between the features."], "text_after_citation": ["Under this relationship, the Lorentz transposed convolutional layer is a Lorentz convolutional layer with changed connectivity through origin padding."], "citing_paper_content": {"title": "Hyperbolic Geometry In Computer Vision: A Novel Framework For Convolutional Neural Networks", "abstract": "Real-world visual data exhibit intrinsic hierarchical structures that can be represented effectively in hyperbolic spaces. Hyperbolic neural networks (HNNs) are a promising approach for learning feature representations in such spaces. However, current methods in computer vision rely on Euclidean backbones and only project features to the hyperbolic space in the task heads, limiting their ability to fully leverage the benefits of hyperbolic geometry. To address this, we present HCNN, the first fully hyperbolic convolutional neural network (CNN) designed for computer vision tasks. Based on the Lorentz model, we generalize fundamental components of CNNs and propose novel formulations of the convolutional layer, batch normalization, and multinomial logistic regression (MLR). Experimentation on standard vision tasks demonstrates the effectiveness of our HCNN framework and the Lorentz model in both hybrid and fully hyperbolic settings. Overall, we aim to pave the way for future research in hyperbolic computer vision by offering a new paradigm for interpreting and analyzing visual data. Our code is publicly available at https://github.com/kschwethelm/HyperbolicCV. * denotes equal contribution Preprint. Under review."}, "cited_paper_content": {"title": "A Guide To Convolution Arithmetic For Deep Learning", "abstract": "We introduce a guide to help deep learning practitioners understand and manipulate convolutional neural network architectures. The guide clarifies the relationship between various properties (input shape, kernel shape, zero padding, strides and output shape) of convolutional, pooling and transposed convolutional layers, as well as the relationship between convolutional and transposed convolutional layers. Relationships are derived for various cases, and are illustrated in order to make them intuitive."}, "keywords": ["input feature map"], "citation_intent": "background"}
{"citing_id": "2303.15965v1", "cited_id": "2002.08546", "section_title": "Related Work", "citation": "DL-based UDA approaches can broadly be split into three categories #REFR : discrepancybased, reconstruction-based, and adversarial.", "text_before_citation": ["Unsupervised Domain Adaptation (UDA): UDA aims to exploit the knowledge learned from a source dataset to help to create a discriminative model for a related but unlabelled target dataset #OTHEREFR ."], "text_after_citation": ["Discrepancy based approaches aim to minimise a divergence criterion, which measures the distance between the source and target data distributions encoded in a learned feature space #OTHEREFR .", "Reconstruction-based approaches, instead, use reconstruction as a proxy task to enable the learning of a shared representation for both image domains #OTHEREFR .", "Finally, adversarial approaches deploy a discriminator that aims to identify the source of the data; the model is trained both to do the task and to trick the discriminator, creating domain invariant features #OTHEREFR .", "These methods all assume simultaneous access to the source and target data, which poses data privacy challenges.", "Federated Learning and Domain Adaptation: Federated learning (FL) has been proposed as a method to train models on distributed data #OTHEREFR ."], "citing_paper_content": {"title": "Sfharmony: Source Free Domain Adaptation For Distributed Neuroimaging Analysis", "abstract": "To represent the biological variability of clinical neuroimaging populations, it is vital to be able to combine data across scanners and studies. However, different MRI scanners produce images with different characteristics, resulting in a domain shift known as the 'harmonisation problem'. Additionally, neuroimaging data is inherently personal in nature, leading to data privacy concerns when sharing the data. To overcome these barriers, we propose an Unsupervised Source-Free Domain Adaptation (SFDA) method, SFHarmony. Through modelling the imaging features as a Gaussian Mixture Model and minimising an adapted Bhattacharyya distance between the source and target features, we can create a model that performs well for the target data whilst having a shared feature representation across the data domains, without needing access to the source data for adaptation or target labels. We demonstrate the performance of our method on simulated and real domain shifts, showing that the approach is applicable to classification, segmentation and regression tasks, requiring no changes to the algorithm. Our method outperforms existing SFDA approaches across a range of realistic data scenarios, demonstrating the potential utility of our approach for MRI harmonisation and general SFDA problems. Our code is available at https://github.com/ nkdinsdale/SFHarmony."}, "cited_paper_content": {"title": "Do We Really Need To Access The Source Data? Source Hypothesis Transfer For Unsupervised Domain Adaptation", "abstract": "Unsupervised domain adaptation (UDA) aims to leverage the knowledge learned from a labeled source dataset to solve similar tasks in a new unlabeled domain. Prior UDA methods typically require to access the source data when learning to adapt the model, making them risky and inefficient for decentralized private data. In this work we tackle a novel setting where only a trained source model is available and investigate how we can effectively utilize such a model without source data to solve UDA problems. To this end, we propose a simple yet generic representation learning framework, named \\emph{Source HypOthesis Transfer} (SHOT). Specifically, SHOT freezes the classifier module (hypothesis) of the source model and learns the target-specific feature extraction module by exploiting both information maximization and self-supervised pseudo-labeling to implicitly align representations from the target domains to the source hypothesis. In this way, the learned target model can directly predict the labels of target data. We further investigate several techniques to refine the network architecture to parameterize the source model for better transfer performance. To verify its versatility, we evaluate SHOT in a variety of adaptation cases including closed-set, partial-set, and open-set domain adaptation. Experiments indicate that SHOT yields state-of-the-art results among multiple domain adaptation benchmarks."}, "keywords": ["DL-based UDA approaches"], "citation_intent": "background"}
{"citing_id": "2303.05009v1", "cited_id": "cond-mat/0501335", "section_title": "Ii. Background", "citation": "The sequential algorithm for constructing PMFGs #REFR first sorts all of the edges by their edge weights.", "text_before_citation": ["We briefly describe the sequential construction of PMFGs and TMFGs, and introduce the notation and primitives that we use.", "A planar graph can be drawn on the plane such that no edges cross each other.", "Both PMFGs and TMFGs are maximal planar subgraphs of a complete undirected, edge-weighted graph (\"maximal\" means that no additional edge can be added to the vertex set without violating planarity).", "They give an approximation to the NP-hard Weighted Maximum Planar Graph problem, which requires the sum of the edge weights kept to be maximized #OTHEREFR .", "The complete edge-weighted graph input can also be viewed as a similarity matrix S, where S[i, j] is the weight of edge (i, j) in the graph. Planar Maximally Filtered Graph (PMFG)."], "text_after_citation": ["It then starts with an empty graph G and considers each edge in the sorted order, from highest to lowest weight.", "An edge is added to the graph if and only if it does not violate the planarity of G, which is checked by running a planarity testing algorithm.", "This process is computationally expensive due to the need to run a planarity testing algorithm \u0398(n 2 ) times. Triangulated Maximally Filtered Graph (TMFG).", "TM-FGs #OTHEREFR have been proposed to improve the efficiency of PMFG construction.", "Instead of considering edges one by one, the original sequential TMFG algorithm adds one vertex x to the vertex set of the subgraph under construction, as well as three edges (each from x to one of the three vertices of a triangular face in a planar embedding of the subgraph) on every iteration, and eliminates the vertex from consideration in future iterations."], "citing_paper_content": {"title": "Parallel Filtered Graphs For Hierarchical Clustering", "abstract": "Given all pairwise weights (distances) among a set of objects, filtered graphs provide a sparse representation by only keeping an important subset of weights. Such graphs can be passed to graph clustering algorithms to generate hierarchical clusters. In particular, the directed bubble hierarchical tree (DBHT) algorithm on filtered graphs has been shown to produce good hierarchical clusters for time series data. We propose a new parallel algorithm for constructing triangulated maximally filtered graphs (TMFG), which produces valid inputs for DBHT, and a scalable parallel algorithm for generating DBHTs that is optimized for TMFG inputs. In addition to parallelizing the original TMFG construction, which has limited parallelism, we also design a new algorithm that inserts multiple vertices on each round to enable more parallelism. We show that the graphs generated by our new algorithm have similar quality compared to the original TMFGs, while being much faster to generate. Our new parallel algorithms for TMFGs and DBHTs are 136-2483x faster than state-of-the-art implementations, while achieving up to 41.56x self-relative speedup on 48 cores with hyper-threading, and achieve better clustering results compared to the standard average-linkage and complete-linkage hierarchical clustering algorithms. We show that on a data set, our algorithms produce clusters that align well with human experts' classification."}, "cited_paper_content": {"title": "A Tool For Filtering Information In Complex Systems", "abstract": "We introduce a technique to filter out complex data sets by extracting a subgraph of representative links. Such a filtering can be tuned up to any desired level by controlling the genus of the resulting graph. We show that this technique is especially suitable for correlation-based graphs, giving filtered graphs that preserve the hierarchical organization of the minimum spanning tree but containing a larger amount of information in their internal structure. In particular in the case of planar filtered graphs (genus equal to 0), triangular loops and four-element cliques are formed. The application of this filtering procedure to 100 stocks in the U.S. equity markets shows that such loops and cliques have important and significant relationships with the market structure and properties."}, "keywords": ["sequential algorithm", "edges"], "citation_intent": "method"}
{"citing_id": "2303.00295v1", "cited_id": "1912.01100", "section_title": "Vi. Conclusions And Future Works", "citation": "Our previous experiences on continual learning make us confident that this target is feasible: in particular, latent replay techniques #REFR seem to be very appropriate in this context because they allow to update the model by storing only a small subset of latent representation of past data.", "text_before_citation": ["In this work we introduced a region prediction technique, where map clustering is used to partition nodes into regions and a deep model is trained to predict the region that the robot is currently navigating.", "This approach enables an efficient implementation of place recognition for realtime loop closure detection.", "Our experiments validate the effectiveness of the idea but leave some issues open.", "In fact, while the proposed technique could be already useful in some practical scenario where the predictor is retrained off-line on the whole accumulated map knowledge, its natural evolution is toward an on-line continual learning implementation as required in lifelong SLAM.", "To this purpose, new regions and new instances (e.g., day/night variations) of already known regions need to be incrementally learned by the model without forgetting the old knowledge (a scenario known as NIC in the continual learning terminology #OTHEREFR )."], "text_after_citation": [], "citing_paper_content": {"title": "Region Prediction For Efficient Robot Localization On Large Maps", "abstract": "Recognizing already explored places (a.k.a. place recognition) is a fundamental task in Simultaneous Localization and Mapping (SLAM) to enable robot relocalization and loop closure detection. In topological SLAM the recognition takes place by comparing a signature (or feature vector) associated to the current node with the signatures of the nodes in the known map. However, as the number of nodes increases, matching the current node signature against all the existing ones becomes inefficient and thwarts real-time navigation. In this paper we propose a novel approach to pre-select a subset of map nodes for place recognition. The map nodes are clustered during exploration and each cluster is associated with a region. The region labels become the prediction targets of a deep neural network and, during navigation, only the nodes associated with the regions predicted with high probability are considered for matching. While the proposed technique can be integrated in different SLAM approaches, in this work we describe an effective integration with RTAB-Map (a popular framework for real-time topological SLAM) which allowed us to design and run several experiments to demonstrate its effectiveness."}, "cited_paper_content": {"title": "Latent Replay For Real-Time Continual Learning", "abstract": "Training deep networks on light computational devices is nowadays very challenging. Continual learning techniques, where complex models are incrementally trained on small batches of new data, can make the learning problem tractable even for CPU-only edge devices. However, a number of practical problems need to be solved: catastrophic forgetting before anything else. In this paper we introduce an original technique named ``Latent Replay'' where, instead of storing a portion of past data in the input space, we store activations volumes at some intermediate layer. This can significantly reduce the computation and storage required by native rehearsal. To keep the representation stable and the stored activations valid we propose to slow-down learning at all the layers below the latent replay one, leaving the layers above free to learn at full pace. In our experiments we show that Latent Replay, combined with existing continual learning techniques, achieves state-of-the-art accuracy on a difficult benchmark such as CORe50 NICv2 with nearly 400 small and highly non-i.i.d. batches. Finally, we demonstrate the feasibility of nearly real-time continual learning on the edge through the porting of the proposed technique on a smartphone device."}, "keywords": ["continual learning", "latent replay techniques"], "citation_intent": "background"}
{"citing_id": "2303.04939v1", "cited_id": "1505.04597", "section_title": "A. U-Net-Based Encoder", "citation": "The first branch of the downsampling path in the proposed architecture consists of a U-Net #REFR encoder as shown in Figure 2 .", "text_before_citation": [], "text_after_citation": ["This branch contains 5 steps, each of them consisting of a 3 \u00d7 3 convolution operation, followed by batch normalization.", "For incorporating non-linearity into the model, each output of the batch normalization is passed through a Rectified Linear Unit (ReLU) activation followed by a 2 \u00d7 2 MaxPooling operation.", "This downsampling path progressively extracts image features and increases these feature dimensions in every layer.", "Thus, ultimately a high-dimensional feature representation with high semantic information is obtained at the final layer.", "Skip connections are introduced, enabling the fine-grained information from the encoder to be recovered in the upsampling half of the network."], "citing_paper_content": {"title": "Ut-Net: Combining U-Net And Transformer For Joint Optic Disc And Cup Segmentation And Glaucoma Detection", "abstract": "Glaucoma is a chronic visual disease that may cause permanent irreversible blindness. Measurement of the cup-to-disc ratio (CDR) plays a pivotal role in the detection of glaucoma in its early stage, preventing visual disparities. Therefore, accurate and automatic segmentation of optic disc (OD) and optic cup (OC) from retinal fundus images is a fundamental requirement. Existing CNNbased segmentation frameworks resort to building deep encoders with aggressive downsampling layers, which suffer from a general limitation on modeling explicit long-range dependency. To this end, in this paper, we propose a new segmentation pipeline, called UT-Net, availing the advantages of UNet and transformer both in its encoding layer, followed by an attention-gated bilinear fusion scheme. In addition to this, we incorporate Multi-Head Contextual attention to enhance the regular self-attention used in traditional vision transformers. Thus low-level features along with global dependencies are captured in a shallow manner. Besides, we extract context information at multiple encoding layers for better exploration of receptive fields, and to aid the model to learn deep hierarchical representations. Finally, an enhanced mixing loss is proposed to tightly supervise the overall learning process. The proposed model has been implemented for joint OD and OC segmentation on three publicly available datasets: DRISHTI-GS, RIM-ONE R3, and REFUGE. Additionally, to validate our proposal, we have performed exhaustive experimentation on Glaucoma detection from all three datasets by measuring the Cup to Disc Ratio (CDR) value. Experimental results demonstrate the superiority of UT-Net as compared to the state-of-the-art methods."}, "cited_paper_content": {"title": "U-Net: Convolutional Networks For Biomedical Image Segmentation", "abstract": "There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ."}, "keywords": ["U-Net encoder"], "citation_intent": "method"}
{"citing_id": "2303.14978v1", "cited_id": "1701.04128", "section_title": "Transformer-Cnn Mixture Blocks", "citation": "To explore how the TCM block aggregates local and non-local information, we present effective receptive fields (ERF) #REFR of our model.", "text_before_citation": ["Based on the proposed TCM Block, the main path (g a and g s ) is designed as Fig. 2 shows.", "The Residual Block with stride (RBS), Residual Block Upsampling (RBU) and subpixel conv3x3 are proposed by #OTHEREFR .", "The detailed architectures of these three modules are reported in Supplementary.", "In our framework, except for the last layer of g a and g s , we attach a TCM block after each RBS/RBU to obtain non-local and local information.", "Also, we add TCM blocks into the hyper-prior path by re-designing the hyper-prior encoder h a and hyper-prior decoder h s as Fig. 4 shows."], "text_after_citation": ["Besides, the effective receptive fields of the transformer-based model #OTHEREFR and the CNNbased model #OTHEREFR are used to make comparisons.", "ERF is defined as absolute gradients of a pixel in the output (i.e., | dxp dx |).", "Here, we calculate gradients of the analysis point p = (70, 700) of kodim04 in Kodak dataset.", "In order to estimate the importance of information at different distances, we clip gradients with two thresholds t (i.e., 0.01, 0.0001) to visualize them.", "Note that the clip operation means we reduce the gradient values larger than the threshold to the threshold, and the gradient values smaller than the threshold remain unchanged. This can help our visualization has higher visibility precision. The visualization is shown in Fig. 3 ."], "citing_paper_content": {"title": "Learned Image Compression With Mixed Transformer-Cnn Architectures", "abstract": "Learned image compression (LIC) methods have exhibited promising progress and superior rate-distortion performance compared with classical image compression standards. Most existing LIC methods are Convolutional Neural Networks-based (CNN-based) or Transformer-based, which have different advantages. Exploiting both advantages is a point worth exploring, which has two challenges: 1) how to effectively fuse the two methods? 2) how to achieve higher performance with a suitable complexity? In this paper, we propose an efficient parallel Transformer-CNN Mixture (TCM) block with a controllable complexity to incorporate the local modeling ability of CNN and the non-local modeling ability of transformers to improve the overall architecture of image compression models. Besides, inspired by the recent progress of entropy estimation models and attention modules, we propose a channel-wise entropy model with parameter-efficient swin-transformer-based attention (SWAtten) modules by using channel squeezing. Experimental results demonstrate our proposed method achieves state-of-the-art rate-distortion performances on three different resolution datasets (i.e., Kodak, Tecnick, CLIC Professional Validation) compared to existing LIC methods. The code is at https://github.com/jmliu206/ LIC_TCM ."}, "cited_paper_content": {"title": "Understanding The Effective Receptive Field In Deep Convolutional Neural Networks", "abstract": "We study characteristics of receptive fields of units in deep convolutional networks. The receptive field size is a crucial issue in many visual tasks, as the output must respond to large enough areas in the image to capture information about large objects. We introduce the notion of an effective receptive field size, and show that it both has a Gaussian distribution and only occupies a fraction of the full theoretical receptive field size. We analyze the effective receptive field in several architecture designs, and the effect of sub-sampling, skip connections, dropout and nonlinear activations on it. This leads to suggestions for ways to address its tendency to be too small."}, "keywords": ["effective receptive fields"], "citation_intent": "method"}
{"citing_id": "2303.02693v1", "cited_id": "1712.04851", "section_title": "Introduction", "citation": "However, as discussed in #REFR , the information in spatial and temporal domains is different in natural video data.", "text_before_citation": ["However, searching for a single 3D architecture requires days on multiple GPUs or TPUs, as training and evaluation of the accuracy indicator are required in the process, making the automatic 3D CNN design process time-consuming and/or hardware-dependent.", "To tackle the above issues, we study how to automatically generate (or design) efficient and expressive 3D CNNs with limited computations.", "Recently, training-free technologies have been introduced by some approaches #OTHEREFR , in which kernel spectrum analysis or forward inference are adopted to measure the expressiveness of spatial 2D CNNs.", "Inspired by the training-free concept and information theory, we suggest that a deep network can be regarded as an information system, and measuring the expressiveness of the network can be considered equivalent to analyzing how much information it can capture.", "Therefore, based on the Maximum Entropy Principle #OTHEREFR , the probability distribution of the system that best represents the current state of knowledge is the one with the highest entropy."], "text_after_citation": ["The spatial dimension is usually limited to some local properties, like connectivity #OTHEREFR , while the temporal dimension usually contains more drastic variations with more complex information.", "To address the spatio-temporal discrepancy in video data, we conduct a kernel selection experiment and observe that different 3D kernel selections in different stages have different effects on performance, and the focus of 3D CNNs changes from spatial information to spatio-temporal information, as the network depth increases.", "We thus consider that the design of 3D CNN architecture should focus on spatial-temporal aggregation depth-wisely.", "The above analysis has motivated us to propose a training-free NAS approach to obtain optimal architectures, i.e., entropy-based 3D CNNs (E3D family).", "Concretely, we first formulate a 3D CNN-based architecture as an information system whose expressiveness can be measured by the value of its differential entropy."], "citing_paper_content": {"title": "Maximizing Spatio-Temporal Entropy Of Deep 3D Cnns For Efficient Video Recognition", "abstract": "3D convolution neural networks (CNNs) have been the prevailing option for video recognition. To capture the temporal information, 3D convolutions are computed along the sequences, leading to cubically growing and expensive computations. To reduce the computational cost, previous methods resort to manually designed 3D/2D CNN structures with approximations or automatic search, which sacrifice the modeling ability or make training time-consuming. In this work, we propose to automatically design efficient 3D CNN architectures via a novel training-free neural architecture search approach tailored for 3D CNNs considering the model complexity. To measure the expressiveness of 3D CNNs efficiently, we formulate a 3D CNN as an information system and derive an analytic entropy score, based on the Maximum Entropy Principle. Specifically, we propose a spatio-temporal entropy score (STEntr-Score) with a refinement factor to handle the discrepancy of visual information in spatial and temporal dimensions, through dynamically leveraging the correlation between the feature map size and kernel size depth-wisely. Highly efficient and expressive 3D CNN architectures, i.e., entropy-based 3D CNNs (E3D family), can then be efficiently searched by maximizing the STEntr-Score under a given computational budget, via an evolutionary algorithm without training the network parameters. Extensive experiments on Something-Something V1&V2 and Kinetics400 demonstrate that the E3D family achieves state-of-the-art performance with higher computational efficiency. Code is available at https://github.com/alibaba/ lightweight-neural-architecture-search."}, "cited_paper_content": {"title": "Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-Offs In Video Classification", "abstract": "Despite the steady progress in video analysis led by the adoption of convolutional neural networks (CNNs), the relative improvement has been less drastic as that in 2D static image classification. Three main challenges exist including spatial (image) feature representation, temporal information representation, and model/computation complexity. It was recently shown by Carreira and Zisserman that 3D CNNs, inflated from 2D networks and pretrained on ImageNet, could be a promising way for spatial and temporal representation learning. However, as for model/computation complexity, 3D CNNs are much more expensive than 2D CNNs and prone to overfit. We seek a balance between speed and accuracy by building an effective and efficient video classification system through systematic exploration of critical network design choices. In particular, we show that it is possible to replace many of the 3D convolutions by low-cost 2D convolutions. Rather surprisingly, best result (in both speed and accuracy) is achieved when replacing the 3D convolutions at the bottom of the network, suggesting that temporal representation learning on high-level \u201csemantic\u201d features is more useful. Our conclusion generalizes to datasets with very different properties. When combined with several other cost-effective designs including separable spatial/temporal convolution and feature gating, our system results in an effective video classification system that that produces very competitive results on several action classification benchmarks (Kinetics, Something-something, UCF101 and HMDB), as well as two action detection (localization) benchmarks (JHMDB and UCF101-24)."}, "keywords": ["natural video data"], "citation_intent": "background"}
{"citing_id": "2304.00177v1", "cited_id": "1706.03762", "section_title": "Ii. Related Work", "citation": "The trend started with the introduction of ViT (Vision Transformers), which globally models non-overlapping spatial relationships in image patches using the standard Transformers encoder #REFR .", "text_before_citation": ["These outputs are combined to create beat-to-beat predictions of the EF and to predict the presence of heart failure.", "This study uses echocardiography video dataset from EchoNet-Dynamic #OTHEREFR .", "In other research #OTHEREFR which can also perform the task of predicting EF values by utilizing the Transformers based architecture that capable to process the videos of arbitrary duration.", "The method uses a Transformers architecture based on the Residual Auto-Encoder (ResAE) Network and a BERT model adapted for token classification.", "Based on Reynaud et al #OTHEREFR , it can be concluded that the backbone architecture modeling in computer vision (CV) has begun to shift to the use of the Transformers architecture."], "text_after_citation": ["For this research, we use Video Swin Transformers [16] , which completely follows the original Swin Transformers hierarchical structure #OTHEREFR .", "However, we extend the local attention computation scope from the spatial domain to the spatio-temporal domain.", "The adaptation process was carried out in the 3D patch partition section and replaced the local window self-attention module into a 3D shifted window based on multi-head self-attention (MSA) and shifted window multi-head self-attention (SW-MSA) in the Transformers Block section.", "Video Swin Transformers can do the video-recognition tasks that contains an inductive bias towards spatio-temporal locality."], "citing_paper_content": {"title": "Hierarchical Vision Transformers For Cardiac Ejection Fraction Estimation", "abstract": "The left ventricular of ejection fraction is one of the most important metric of cardiac function. It is used by cardiologist to identify patients who are eligible for lifeprolonging therapies. However, the assessment of ejection fraction suffers from inter-observer variability. To overcome this challenge, we propose a deep learning approach, based on hierarchical vision Transformers, to estimate the ejection fraction from echocardiogram videos. The proposed method can estimate ejection fraction without the need for left ventrice segmentation first, make it more efficient than other methods. We evaluated our method on EchoNet-Dynamic dataset resulting 5.59, 7.59 and 0.59 for MAE, RMSE and R 2 respectivelly. This results are better compared to the state-of-the-art method, Ultrasound Video Transformer (UVT)."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["Vision Transformers", "standard Transformers encoder"], "citation_intent": "method"}
{"citing_id": "2303.13233v1", "cited_id": "1602.07332", "section_title": "Conclusions", "citation": "Table 5 and 7 show the coarse-grained predicates and fine-grained predicates with the number of training instances for each predicate in the Visual Genome dataset #REFR .", "text_before_citation": ["In this work, we propose an automatic boosting framework CaCao that exploits linguistic knowledge from pretrained language models to enrich existing datasets in a lowresource way.", "We tackle the long-tail issue of SGG with the help of abundant informative predicates from CaCao and generalize to open-world predicate learning with the entangled cross-modal prompt design based on VL models.", "Our extensive experiments on three datasets illustrate the significant improvement of our CaCao on fine-grained scene graph generation and open-world generalization capability."], "text_after_citation": ["Table 6 and 8 show the coarse-grained predicates and finegrained predicates with the number of training instances for each predicate after cross-modal boosting by our CaCao.", "We can observe that CaCao increases dataset scale, especially the tail predicates, which significantly alleviates the long-tail distribution problem in SGG.", "For large-scale benchmark SGG, GQA #OTHEREFR contains 113K images and over 3.8M relation annotations.", "In order to ensure the quality and representativeness of the dataset, we perform a manual cleaning process to remove annotations that had poor quality or ambiguous meanings following prior works #OTHEREFR .", "We finally select the top 200 object classes and top 100 predicate classes as the GQA-200 split like VG-50 to explore the generalization ability of CaCao in large-scale SGG."], "citing_paper_content": {"title": "Visually-Prompted Language Model For Fine-Grained Scene Graph Generation In An Open World", "abstract": "Scene Graph Generation (SGG) aims to extract <subject, predicate, object> relationships in images for vision understanding. Although recent works have made steady progress on SGG, they still suffer long-tail distribution issues that tail-predicates are more costly to train and hard to distinguish due to a small amount of annotated data compared to frequent predicates. Existing re-balancing strategies try to haddle it via prior rules but are still confined to pre-defined conditions, which are not scalable for various models and datasets. In this paper, we propose a Crossmodal prediCate boosting (CaCao) framework, where a visually-prompted language model is learned to generate diverse fine-grained predicates in a low-resource way. The proposed CaCao can be applied in a plug-and-play fashion and automatically strengthen existing SGG to tackle the long-tailed problem. Based on that, we further introduce a novel Entangled cross-modal prompt approach for open-world predicate scene graph generation (Epic), where models can generalize to unseen predicates in a zeroshot manner. Comprehensive experiments on three benchmark datasets show that CaCao consistently boosts the performance of multiple scene graph generation models in a model-agnostic way. Moreover, our Epic achieves competitive performance on open-world predicate prediction."}, "cited_paper_content": {"title": "Visual Genome: Connecting Language And Vision Using Crowdsourced Dense Image Annotations", "abstract": "Despite progress in perceptual tasks such as image classification, computers still perform poorly on cognitive tasks such as image description and question answering. Cognition is core to tasks that involve not just recognizing, but reasoning about our visual world. However, models used to tackle the rich content in images for cognitive tasks are still being trained using the same datasets designed for perceptual tasks. To achieve success at cognitive tasks, models need to understand the interactions and relationships between objects in an image. When asked \"What vehicle is the person riding?\", computers will need to identify the objects in an image as well as the relationships riding(man, carriage) and pulling(horse, carriage) to answer correctly that \"the person is riding a horse-drawn carriage.\" In this paper, we present the Visual Genome dataset to enable the modeling of such relationships. We collect dense annotations of objects, attributes, and relationships within each image to learn these models. Specifically, our dataset contains over 108K images where each image has an average of $$35$$35 objects, $$26$$26 attributes, and $$21$$21 pairwise relationships between objects. We canonicalize the objects, attributes, relationships, and noun phrases in region descriptions and questions answer pairs to WordNet synsets. Together, these annotations represent the densest and largest dataset of image descriptions, objects, attributes, relationships, and question answer pairs."}, "keywords": ["fine-grained predicates", "Visual Genome"], "citation_intent": "background"}
{"citing_id": "2303.11114v1", "cited_id": "1802.00420", "section_title": "B.5. Robustness Benchmarks", "citation": "B.1 ; it could be due to a strong obfuscated gradient effect #REFR that cannot be detected by a naive straight-through estimator.", "text_before_citation": ["We also evaluate the nonquantized version of the quantizer (i.e., omitting the vector quantization process, but using the extracted feature by the encoder directly to the ViT input), but we empirically observe that attacking the non-quantized version cannot drop the performance at all.", "Instead, we use the STE, also used during the training as well as the previous extensive robustness study #OTHEREFR .", "We compare the attacked accuracies of DeiT and SeiT by varying \u03b5 (a control parameter for the attack intensity) from 0 to 8 in Fig. B.1 .", "We observe that SeiT shows almost neglectable performance drops even under the strongest attack (showing 73.95 for \u03b5 = 8 where 73.98 for \u03b5 = 0), where DeiT shows 2.8% top-1 accuracy.", "However, we should be careful to interpret Fig."], "text_after_citation": ["Moreover, our method could be vulnerable to the codebook attack by changing the token indices directly, not by perturbing the pixels.", "However, as an efficient and natural ad- versarial attack on discrete domains is still an open problem #OTHEREFR (e.g., altering indices as imperceptible to humans but sensitive to machines -only a small index change can make a huge semantic gap, such as replacing \"huge\" in the previous sentence to \"neglectable\"), we leave the investigation of advanced adversarial attack methods for SeiT beyond straight-through estimator as future work.", "https://github.com/CompVis/taming-transformers"], "citing_paper_content": {"title": "Seit: Storage-Efficient Vision Training With Tokens Using 1% Of Pixel Storage", "abstract": "We need billion-scale images to achieve more generalizable and groundbreaking vision models, as well as massive dataset storage to ship the images (e.g., the LAION-4B dataset needs 240TB storage space). However, it has become challenging to deal with unlimited dataset storage with limited storage infrastructure. A number of storageefficient training methods have been proposed to tackle the problem, but they are rarely scalable or suffer from severe damage to performance. In this paper, we propose a storage-efficient training strategy for vision classifiers for large-scale datasets (e.g., ImageNet) that only uses 1024 tokens per instance without using the raw level pixels; our token storage only needs <1% of the original JPEG-compressed raw pixels. We also propose token augmentations and a Stem-adaptor module to make our approach able to use the same architecture as pixel-based approaches with only minimal modifications on the stem layer and the carefully tuned optimization settings. Our experimental results on ImageNet-1k show that our method significantly outperforms other storage-efficient training methods with a large gap. We further show the effectiveness of our method in other practical scenarios, storage-efficient pre-training, and continual learning. Code is available at https://github.com/naver-ai/seit Adjusting image resolution Adjusting JPEG quality C-score based sampling Uniform random sampling Original DeiT SeiT SeiT (pixel fine-tuned) Storage (GB, log scale) Top 1 Accuracy (%)"}, "cited_paper_content": {"title": "Obfuscated Gradients Give A False Sense Of Security: Circumventing Defenses To Adversarial Examples", "abstract": "We identify obfuscated gradients as a phenomenon that leads to a false sense of security in defenses against adversarial examples. While defenses that cause obfuscated gradients appear to defeat optimization-based attacks, we find defenses relying on this effect can be circumvented. ::: For each of the three types of obfuscated gradients we discover, we describe indicators of defenses exhibiting this effect and develop attack techniques to overcome it. In a case study, examining all defenses accepted to ICLR 2018, we find obfuscated gradients are a common occurrence, with 7 of 8 defenses relying on obfuscated gradients. Using our new attack techniques, we successfully circumvent all 7 of them."}, "keywords": ["strong obfuscated gradient"], "citation_intent": "background"}
{"citing_id": "2303.13825v1", "cited_id": "2003.08934", "section_title": "Neural Radiance Fields (Nerf)", "citation": "As the most popular implicit representation in neural rendering, Neural Radiance Fields (NeRF) #REFR has exhibited stunning results in various tasks since its first introduction.", "text_before_citation": ["Recent years have witnessed the rapid development of neural implicit representations #OTHEREFR in 3D modeling and image synthesis.", "Compared with classic discrete counterparts such as meshes, point clouds, and voxels, neural implicit representations model the scene with neural networks, which are spatially continuous and indicate a higher fidelity and flexibility."], "text_after_citation": ["The original NeRF overfits on one static scene by design, therefore it cannot model time-varying contents.", "Many efforts have been made to adapt NeRF to dynamic scenes.", "Some works condition the NeRF with local #OTHEREFR or global scene representations #OTHEREFR ] to implicitly provide generalizability for it.", "As the pioneers, D-NeRF #OTHEREFR and Nerfies #OTHEREFR use an explicit deformation field to bend straight rays passing through varying targets into a common canonical scene, where a conventional NeRF is optimized.", "Such a pipeline is adopted by many follow-up works #OTHEREFR ."], "citing_paper_content": {"title": "Handnerf: Neural Radiance Fields For Animatable Interacting Hands", "abstract": "Multi-view images Unified modeling and rendering for photo-realistic single / interacting hands under novel views / poses"}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["neural rendering"], "citation_intent": "background"}
{"citing_id": "2304.10415v1", "cited_id": "1912.07849", "section_title": "Challenge Results", "citation": "The top 11 of them produced PSNR scores higher than the baseline method LF-InterNet #REFR .", "text_before_citation": ["Among the 148 registered participants, 12 teams have successfully participated the final test phase and submitted their results, codes, and factsheets."], "text_after_citation": ["Table 1 reports the PSNR and SSIM scores achieved by these methods on both test and validation sets, together with their major details.", "We briefly describe these solutions in Section 4, and introduce the corresponding team members in Appendix 6.", "It can be observed from Table 1 that all these methods surpass the state-of-the-art method DistgSSR #OTHEREFR , and 9 of them surpass the recent top-performing method EPIT #OTHEREFR .", "Note that, the winner solution proposed by the OpenMeow achieves around 1 dB improvement in PSNR over Dist-gSSR #OTHEREFR on both test and validation sets, which significantly push the state-of-the-art of LF image SR to a new height.", "Moreover, the accuracy of the top 2 methods are very close with a minor PSNR difference of 0.02 dB on the test set."], "citing_paper_content": {"title": "Ntire 2023 Challenge On Light Field Image Super-Resolution: Dataset, Methods And Results", "abstract": "In this report, we summarize the first NTIRE challenge on light field (LF) image super-resolution (SR), which aims at super-resolving LF images under the standard bicubic degradation with a magnification factor of 4. This challenge develops a new LF dataset called NTIRE-2023 for validation and test, and provides a toolbox called BasicLFSR to facilitate model development. Compared with single image SR, the major challenge of LF image SR lies in how to exploit complementary angular information from plenty of views with varying disparities. In total, 148 participants have registered the challenge, and 11 teams have successfully submitted results with PSNR scores higher than the baseline method LF-InterNet [1]. These newly developed methods have set new state-of-the-art in LF image SR, e.g., the winning method achieves around 1 dB PSNR improvement over the existing state-of-the-art method DistgSSR [2]. We report the solutions proposed by the participants, and summarize their common trends and useful tricks. We hope this challenge can stimulate future research and inspire new ideas in LF image SR."}, "cited_paper_content": {"title": "Spatial-Angular Interaction For Light Field Image Super-Resolution", "abstract": "Light field (LF) cameras record both intensity and directions of light rays, and capture scenes from a number of viewpoints. Both information within each perspective (i.e., spatial information) and among different perspectives (i.e., angular information) is beneficial to image super-resolution (SR). In this paper, we propose a spatial-angular interactive network (namely, LF-InterNet) for LF image SR. In our method, spatial and angular features are separately extracted from the input LF using two specifically designed convolutions. These extracted features are then repetitively interacted to incorporate both spatial and angular information. Finally, the interacted spatial and angular features are fused to super-resolve each sub-aperture image. Experiments on 6 public LF datasets have demonstrated the superiority of our method. As compared to existing LF and single image SR methods, our method can recover much more details, and achieves significant improvements over the state-of-the-arts in terms of PSNR and SSIM."}, "keywords": ["PSNR scores", "LF-InterNet"], "citation_intent": "method"}
{"citing_id": "2304.09355v2", "cited_id": "physics/0004057", "section_title": "Introduction", "citation": "We discuss the application of the information bottleneck principle #REFR , connections between information theory and generalization, and recent information-theoretic learning algorithms.", "text_before_citation": ["In this paper, we aim to achieve two objectives.", "First, we propose a unified framework that synthesizes existing research on self-supervised and semi-supervised learning from an information-theoretic standpoint.", "This framework allows us to present and compare current methods, analyze their assumptions and difficulties, and discuss the optimal representation for neural networks in general and self-supervised networks in particular.", "Second, we explore different methods and estimators for optimizing information-theoretic quantities in deep neural networks and investigate how recent models optimize various theoretical-information terms.", "By reviewing the literature on various aspects of information-theoretic learning, we provide a comprehensive understanding of the interplay between information theory, self-supervised learning, and deep neural networks."], "text_after_citation": ["Furthermore, we examine how the information-theoretic perspective can offer insights into the design of better self-supervised learning algorithms and the potential benefits of using information theory in SSL across a wide range of applications.", "In addition to the main structure of the paper, we dedicate a section to the challenges and opportunities in extending the information-theoretic perspective to other learning paradigms, such as energy-based models.", "We highlight the potential advantages of incorporating these extensions into self-supervised learning algorithms and discuss the technical and conceptual challenges that need to be addressed.", "The structure of the paper is as follows.", "Section 2 introduces the key concepts in supervised, semi-supervised, and self-supervised learning, information theory, and representation learning."], "citing_paper_content": {"title": "To Compress Or Not To Compress -Self-Supervised Learning And Information Theory: A Review", "abstract": "Deep neural networks have demonstrated remarkable performance in supervised learning tasks but require large amounts of labeled data. Self-supervised learning offers an alternative paradigm, enabling the model to learn from data without explicit labels. Information theory has been instrumental in understanding and optimizing deep neural networks. Specifically, the information bottleneck principle has been applied to optimize the trade-off between compression and relevant information preservation in supervised settings. However, the optimal information objective in self-supervised learning remains unclear. In this paper, we review various approaches to self-supervised learning from an information-theoretic standpoint and present a unified framework that formalizes the self-supervised informationtheoretic learning problem. We integrate existing research into a coherent framework, examine recent self-supervised methods, and identify research opportunities and challenges. Moreover, we discuss empirical measurement of information-theoretic quantities and their estimators. This paper offers a comprehensive review of the intersection between information theory, self-supervised learning, and deep neural networks."}, "cited_paper_content": {"title": "The Information Bottleneck Method", "abstract": "We define the relevant information in a signal $x\\in X$ as being the information that this signal provides about another signal $y\\in \\Y$. Examples include the information that face images provide about the names of the people portrayed, or the information that speech sounds provide about the words spoken. Understanding the signal $x$ requires more than just predicting $y$, it also requires specifying which features of $\\X$ play a role in the prediction. We formalize this problem as that of finding a short code for $\\X$ that preserves the maximum information about $\\Y$. That is, we squeeze the information that $\\X$ provides about $\\Y$ through a `bottleneck' formed by a limited set of codewords $\\tX$. This constrained optimization problem can be seen as a generalization of rate distortion theory in which the distortion measure $d(x,\\x)$ emerges from the joint statistics of $\\X$ and $\\Y$. This approach yields an exact set of self consistent equations for the coding rules $X \\to \\tX$ and $\\tX \\to \\Y$. Solutions to these equations can be found by a convergent re-estimation method that generalizes the Blahut-Arimoto algorithm. Our variational principle provides a surprisingly rich framework for discussing a variety of problems in signal processing and learning, as will be described in detail elsewhere."}, "keywords": ["information bottleneck principle"], "citation_intent": "background"}
{"citing_id": "2303.06664v1", "cited_id": "1607.04311", "section_title": "Related Work", "citation": "Carlini and Wagner #REFR demonstrated that defensive distillation is no more robust to adversity than unprotected neural networks.", "text_before_citation": ["This method, while showing good prospects in some areas of research, is subject to the problem of blind spots #OTHEREFR , which means that if adversarial examples of a certain attack family are not included in the learning phase, the classifier may fail to be robust to these adversarial examples.", "Consequently, this implies that adversarial training could not be as effective against zero-day attacks, which are one of the key issues facing the intrusion detection sector.", "Moreover, the IDS may perform worse in its original classification job as a result of being retrained with adversarial instances #OTHEREFR .", "Defensive distillation, which was originally used to lower the dimensionality of DNNs, is another adversarial defense presented in the literature #OTHEREFR . Papernot et al.", "#OTHEREFR suggest a distillation-based defense with the goal of smoothing the model's decision surface."], "text_after_citation": ["As such, this defense is broken using the C&W attack.", "To defend against adversarial assaults, one might discard any features that could be modified by the attacker without affecting the logic of network communication #OTHEREFR . However, Apruzzese et al.", "#OTHEREFR demonstrated that doing so has a negative influence on detector performance in non-adversarial conditions, for example, by generating more false alarms, which is extremely undesired for current cyber security platforms.", "This decrease in performance is due to the deleted features having a major influence on the underlying processes of the baseline detectors.", "Anomaly detection is an interesting alternative defensive mechanism against evasion attacks."], "citing_paper_content": {"title": "Adv-Bot: Realistic Adversarial Botnet Attacks Against Network Intrusion Detection Systems", "abstract": "Due to the numerous advantages of machine learning (ML) algorithms, many applications now incorporate them. However, many studies in the field of image classification have shown that MLs can be fooled by a variety of adversarial attacks. These attacks take advantage of ML algorithms' inherent vulnerability. This raises many questions in the cybersecurity field, where a growing number of researchers are recently investigating the feasibility of such attacks against machine learning-based security systems, such as intrusion detection systems. The majority of this research demonstrates that it is possible to fool a model using features extracted from a raw data source, but it does not take into account the real implementation of such attacks, i.e., the reverse transformation from theory to practice. The real implementation of these adversarial attacks would be influenced by various constraints that would make their execution more difficult. As a result, the purpose of this study was to investigate the actual feasibility of adversarial attacks, specifically evasion attacks, against network-based intrusion detection systems (NIDS), demonstrating that it is entirely possible to fool these ML-based IDSs using our proposed adversarial algorithm while assuming as many constraints as possible in a black-box setting. In addition, since it is critical to design defense mechanisms to protect ML-based IDSs against such attacks, a defensive scheme is presented. Realistic botnet traffic traces are used to assess this work. Our goal is to create adversarial botnet traffic that can avoid detection while still performing all of its intended malicious functionality."}, "cited_paper_content": {"title": "Defensive Distillation Is Not Robust To Adversarial Examples", "abstract": "We show that defensive distillation is not secure: it is no more resistant to targeted misclassification attacks than unprotected neural networks."}, "keywords": ["unprotected neural networks", "defensive distillation"], "citation_intent": "background"}
{"citing_id": "2304.14326v1", "cited_id": "1905.07773", "section_title": "Uc-O-Gdps Algorithm", "citation": "The next result, which directly follows from #REFR , shows that the cumulative error due to the estimation of the transition probabilities grows sublinearly during time. Lemma 2.", "text_before_citation": ["The confidence set P i is updated at every epoch i as, \u2200 (x, a) \u2208 X \u00d7 A:", "EQUATION", "where \u01eb i (x, a) is defined as:", "\u01eb i (x, a) = 2|X k(x)+1 | ln T |X||A| \u03b4 max {1, N i (x, a)} ,", "and k(x) denotes the index of the layer to which x belongs and \u03b4 \u2208 (0, 1) is the given confidence."], "text_after_citation": ["When the confidence set P is updated as in Equation 10, with probability at least 1 \u2212 2\u03b4, it holds:", "T t=1 ||q t \u2212 q t || 1 \u2264 E q \u03b4 , where E q \u03b4 \u2264\u00d5( \u221a T ).", "Initialization.", "UC-O-GDPS.INIT procedure (Line 1 of Algorithm 2) initializes the epoch index as i = 1 and confidence set P 1 as the set of all possible transition functions.", "For all k \u2208 [0..L \u2212 1] and all (x, a, x \u2032 ) \u2208 X k \u00d7 A\u00d7 X k+1 , the counters are initialized as"], "citing_paper_content": {"title": "A Best-Of-Both-Worlds Algorithm For Constrained Mdps With Long-Term Constraints", "abstract": "We study online learning in episodic constrained Markov decision processes (CMDPs), where the goal of the learner is to collect as much reward as possible over the episodes, while guaranteeing that some long-term constraints are satisfied during the learning process. Rewards and constraints can be selected either stochastically or adversarially, and the transition function is not known to the learner. While online learning in classical (unconstrained) MDPs has received considerable attention over the last years, the setting of CMDPs is still largely unexplored. This is surprising, since in real-world applications, such as, e.g., autonomous driving, automated bidding, and recommender systems, there are usually additional constraints and specifications that an agent has to obey during the learning process. In this paper, we provide the first best-of-both-worlds algorithm for CMDPs with longterm constraints. Our algorithm is capable of handling settings in which rewards and constraints are selected either stochastically or adversarially, without requiring any knowledge of the underling process. Moreover, our algorithm matches state-of-the-art regret and constraint violation bounds for settings in which constraints are selected stochastically, while it is the first to provide guarantees in the case in which they are chosen adversarially."}, "cited_paper_content": {"title": "Online Convex Optimization In Adversarial Markov Decision Processes", "abstract": "We consider online learning in episodic loop-free Markov decision processes (MDPs), where the loss function can change arbitrarily between episodes, and the transition function is not known to the learner. We show $\\tilde{O}(L|X|\\sqrt{|A|T})$ regret bound, where $T$ is the number of episodes, $X$ is the state space, $A$ is the action space, and $L$ is the length of each episode. Our online algorithm is implemented using entropic regularization methodology, which allows to extend the original adversarial MDP model to handle convex performance criteria (different ways to aggregate the losses of a single episode) , as well as improve previous regret bounds."}, "keywords": ["cumulative error"], "citation_intent": "background"}
{"citing_id": "2304.12616v1", "cited_id": "1703.01780", "section_title": "5:", "citation": "Specifically, in this paper, we adopt the traditional teacher-student network #REFR to compose our model.", "text_before_citation": ["Update the teacher network's parameters from student network's parameters by the exponential moving average (EMA) strategy. their temporal context.", "We find that simply applying consistency would cause the model to over-focus on the most discriminative parts of the video and affect the completeness of localization.", "Hence we boost the semantic consistency constraint in a bidirectional manner to suppress co-scene actions while maintaining the stability and integrity of positive actions prediction.", "Furthermore, considering that different transformations of the same action video will produce different predictions, combining multiple transforms of the same video to generate a comprehensive T-CAM always covers a more comprehensive part of actions than a single one.", "Hence we also utilize the comprehensive background suppressed T-CAM to ensuring the integrity of positive actions."], "text_after_citation": ["The model proposed in this paper contains an original branch and an augmentation branch, both of which contain a teacher network and a student network.", "Specifically, in this paper, the original branch includes a teacher-student network.", "The student network of the original branch takes an original video feature X as input while the inputs of the teacher network of the original branch are multiple video features transformed by Intra-TCA in III-C1c.", "In the teacher network of the original branch, K different inputs including I 1 (X), ..., I K (X) are fed into the same modules to get multiple background-suppressed T-CAMS T 1 , ...,S T K , where I() means the proposed intra-video temporal context augmentation strategy.", "Then the comprehensive T-CAM generation module would generate comprehensive background suppressed T-CAM,"], "citing_paper_content": {"title": "> Ieee Transactions On Neural Networks And Learning Systems <", "abstract": "Weakly Supervised Temporal Action Localization (WTAL) aims to classify and localize temporal boundaries of actions for the video, given only video-level category labels in the training datasets. Due to the lack of boundary information during training, existing approaches formulate WTAL as a classification problem, i.e., generating the temporal class activation map (T-CAM) for localization. However, with only classification loss, the model would be sub-optimized, i.e., the action-related scenes are enough to distinguish different class labels. Regarding other actions in the action-related scene (i.e., the scene same as positive actions) as co-scene actions, this sub-optimized model would misclassify the co-scene actions as positive actions. To address this misclassification, we propose a simple yet efficient method, named bidirectional semantic consistency constraint (Bi-SCC), to discriminate the positive actions from co-scene actions. The proposed Bi-SCC firstly adopts a temporal context augmentation to generate an augmented video that breaks the correlation between positive actions and their co-scene actions in the intervideo; Then, a semantic consistency constraint (SCC) is used to enforce the predictions of the original video and augmented video to be consistent, hence suppressing the co-scene actions. However, we find that this augmented video would destroy the original temporal context. Simply applying the consistency constraint would affect the completeness of localized positive actions. Hence, we boost the SCC in a bidirectional way to suppress co-scene actions while ensuring the integrity of positive actions, by cross-supervising the original and augmented videos. Finally, our proposed Bi-SCC can be applied to current WTAL approaches, and improve their performance. Experimental results show that our approach outperforms the state-of-the-art methods on THUMOS14 and ActivityNet. The code is available at https://github.com/lgzlIlIlI/BiSCC."}, "cited_paper_content": {"title": "Weight-Averaged Consistency Targets Improve Semi-Supervised Deep Learning Results", "abstract": "The recently proposed Temporal Ensembling has achieved state-of-the-art results in several semi-supervised learning benchmarks. It maintains an exponential moving average of label predictions on each training example, and penalizes predictions that are inconsistent with this target. However, because the targets change only once per epoch, Temporal Ensembling becomes unwieldy when learning large datasets. To overcome this problem, we propose Mean Teacher, a method that averages model weights instead of label predictions. As an additional benefit, Mean Teacher improves test accuracy and enables training with fewer labels than Temporal Ensembling. Without changing the network architecture, Mean Teacher achieves an error rate of 4.35% on SVHN with 250 labels, outperforming Temporal Ensembling trained with 1000 labels. We also show that a good network architecture is crucial to performance. Combining Mean Teacher and Residual Networks, we improve the state of the art on CIFAR-10 with 4000 labels from 10.55% to 6.28%, and on ImageNet 2012 with 10% of the labels from 35.24% to 9.11%."}, "keywords": ["model", "traditional teacher-student network"], "citation_intent": "method"}
{"citing_id": "2304.10756v1", "cited_id": "1703.01780", "section_title": "Benchmark For Semi-Supervised Multi-Modal Segmentation", "citation": "We compare our proposed M3L semi-supervised framework against supervised-only and mean teacher #REFR to prove its efficacy.", "text_before_citation": ["We create three different semi-supervised configurations using 0.1% (49), 0.2% (98), and 1% (491) of the data as labeled and the rest as unlabeled (49159, 49101, 48708) .", "SUN RGBD is a challenging dataset with 37 classes and has 5049 test samples.", "It originally had 5285 training samples which we split into 90/10 train/val (4757 / 528) samples.", "For semi-supervised training, we create three configurations and treat 6.25% (297), 12.5% (594), and 25% (1189) of the data as labeled and the rest as unlabeled (4460, 4163, 3568) .", "Baselines."], "text_after_citation": ["We also compare the performance of Linear Fusion with supervised-only transformer-based fusion approaches presented in the literature #OTHEREFR .", "\u2022 URN #OTHEREFR : We choose a straightforward fusion mechanism that closely resembles the approach presented in Unified Representation Network #OTHEREFR , but we use Segformer as the base architecture.", "Different encoders are used for the two modalities and the encoded representations are fused (averaged) before passing them to a single decoder.", "Since two encoders are trained, URN has \u223c 2\u00d7 the number of parameters than other methods described below.", "\u2022 TF #OTHEREFR : The state-of-the-art Token Fusion #OTHEREFR framework."], "citing_paper_content": {"title": "Missing Modality Robustness In Semi-Supervised Multi-Modal Semantic Segmentation", "abstract": "Using multiple spatial modalities has been proven helpful in improving semantic segmentation performance. However, there are several real-world challenges that have yet to be addressed: (a) improving label efficiency and (b) enhancing robustness in realistic scenarios where modalities are missing at the test time. To address these challenges, we first propose a simple yet efficient multi-modal fusion mechanism Linear Fusion, that performs better than the stateof-the-art multi-modal models even with limited supervision. Second, we propose M3L: Multi-modal Teacher for Masked Modality Learning, a semi-supervised framework that not only improves the multi-modal performance but also makes the model robust to the realistic missing modality scenario using unlabeled data. We create the first benchmark for semi-supervised multi-modal semantic segmentation and also report the robustness to missing modalities. Our proposal shows an absolute improvement of up to 10% on robust mIoU above the most competitive baselines. Our code is available at https:"}, "cited_paper_content": {"title": "Weight-Averaged Consistency Targets Improve Semi-Supervised Deep Learning Results", "abstract": "The recently proposed Temporal Ensembling has achieved state-of-the-art results in several semi-supervised learning benchmarks. It maintains an exponential moving average of label predictions on each training example, and penalizes predictions that are inconsistent with this target. However, because the targets change only once per epoch, Temporal Ensembling becomes unwieldy when learning large datasets. To overcome this problem, we propose Mean Teacher, a method that averages model weights instead of label predictions. As an additional benefit, Mean Teacher improves test accuracy and enables training with fewer labels than Temporal Ensembling. Without changing the network architecture, Mean Teacher achieves an error rate of 4.35% on SVHN with 250 labels, outperforming Temporal Ensembling trained with 1000 labels. We also show that a good network architecture is crucial to performance. Combining Mean Teacher and Residual Networks, we improve the state of the art on CIFAR-10 with 4000 labels from 10.55% to 6.28%, and on ImageNet 2012 with 10% of the labels from 35.24% to 9.11%."}, "keywords": ["M3L semi-supervised framework", "supervised-only"], "citation_intent": "method"}
{"citing_id": "2304.09540v1", "cited_id": "1909.04559", "section_title": "Introduction", "citation": "These learning algorithms are adapted from the noise-free learning algorithm in #REFR , and work for both tree hierarchies and general concept hierarchies.", "text_before_citation": ["Section 6 contains algorithms for robust recognition in networks with feedback, for both tree hierarchies and general hierarchies.", "For tree hierarchies in networks with feedback, we show that recognition requires only enough time for two passes: an upward pass to recognize whatever can be recognized based on lower-level information, and a downward pass to recognize concepts based on a combination of higher-level and lower-level information.", "On the other hand, for general concept hierarchies in networks with feedback, we may need much more time-enough for many passes, both upward and downward.", "We again describe how the recognition results extend to settings in which weights are approximate, and in which weights and thresholds are scaled.", "In Section 7, we describe noise-free learning algorithms in feed-forward networks, which produce edge weights for the upward edges that suffice to support robust recognition."], "text_after_citation": ["The extension to general hierarchies requires us to reconsider the use of the Winner-Take-All (WTA) module in the algorithm of #OTHEREFR , since the previous version does not work with overlap; we present a new version of the WTA mechanism.", "As before, the weight adjustments are based on Oja's rule #OTHEREFR .", "We show that our new learning algorithms can be viewed as producing approximate, scaled weights as described in Section 5, which can be used to decompose the correctness proof for the learning algorithms. Finally, we briefly discuss extensions to noisy learning.", "In Section 8, we extend the learning algorithms for feed-forward networks to accommodate feedback.", "Here we simply separate the learning mechanisms for the weights of the upward and downward edges, using a different rule for learning the weights of the downward edges."], "citing_paper_content": {"title": "Learning Hierarchically-Structured Concepts Ii: Overlapping Concepts, And Networks With Feedback", "abstract": "We continue our study from [5], of how concepts that have hierarchical structure might be represented in brain-like neural networks, how these representations might be used to recognize the concepts, and how these representations might be learned. In [5], we considered simple tree-structured concepts and feed-forward layered networks. Here we extend the model in two ways: we allow limited overlap between children of different concepts, and we allow networks to include feedback edges. For these more general cases, we describe and analyze algorithms for recognition and algorithms for learning. We also use the abbreviations supp r,f (B, * , t) for S(, t), supp r,f (B, , *) for t S(, t), and supp r,f (B, * , *) for ,t S(, t), Notice that each of these three unions must converge to a finite set since all the sets S(, t) are subsets of the single finite set C of concepts. Now we have two monotonicity results, for r and f : Lemma 2.5. Let C be any concept hierarchy satisfying the limited-overlap property, and let B \u2286 D 0. Consider r, r , where 0 \u2264 r \u2264 r \u2264 1, and arbitrary f. Then supp r ,f (B) \u2286 supp r,f (B). Lemma 2.6. Let C be any concept hierarchy satisfying the limited-overlap property, and let B \u2286 D 0. Consider f, f , where 0 \u2264 f \u2264 f , and arbitrary r \u2208 [0, 1]. Then supp r,f (B) \u2286 supp r,f (B)."}, "cited_paper_content": {"title": "Learning Hierarchically Structured Concepts", "abstract": "We study the question of how concepts that have structure get represented in the brain. Specifically, we introduce a model for hierarchically structured concepts and we show how a biologically plausible neural network can recognize these concepts, and how it can learn them in the first place. Our main goal is to introduce a general framework for these tasks and prove formally how both (recognition and learning) can be achieved. ::: We show that both tasks can be accomplished even in presence of noise. For learning, we analyze Oja's rule formally, a well-known biologically-plausible rule for adjusting the weights of synapses. We complement the learning results with lower bounds asserting that, in order to recognize concepts of a certain hierarchical depth, neural networks must have a corresponding number of layers."}, "keywords": ["learning algorithms", "general concept hierarchies"], "citation_intent": "method"}
{"citing_id": "2303.04712v1", "cited_id": "1403.6652", "section_title": "Language-Specific Deepwalk Embeddings", "citation": "To create language-specific embeddings, we utilize Deep-Walk #REFR , and follow a uniform random sampling approach: the next entity to visit in the random walk is chosen uniformly from all neighbors of the current entity.", "text_before_citation": ["To leverage the information of entities and the structure of the language-specific knowledge graph efficiently, we propose a language-specific embedding technique that learns continuous vector representation of entities representing their relations in a language .", "This technique maps the entities to low-dimensional vectors, which are similar in cases where two entities appear close to each other in the language-specific context of ."], "text_after_citation": ["After creating a set of random walks following the Deep-Walk approach, we train a Word2Vec model.", "The resulting language-specific embeddings are utilized for (i) creating a candidate set of events relevant to the query entity and (ii) for measuring the language-specific relevance between the query entity and the event."], "citing_paper_content": {"title": "Laser: Language-Specific Event Recommendation", "abstract": "While societal events often impact people worldwide, a significant fraction of events has a local focus that primarily affects specific language communities. Examples include national elections, the development of the Coronavirus pandemic in different countries, and local film festivals such as the C\u00e9sar Awards in France and the Moscow International Film Festival in Russia. However, existing entity recommendation approaches do not sufficiently address the language context of recommendation. This article introduces the novel task of language-specific event recommendation, which aims to recommend events relevant to the user query in the language-specific context. This task can support essential information retrieval activities, including web navigation and exploratory search, considering the language context of user information needs. We propose LaSER, a novel approach toward language-specific event recommendation. LaSER blends the language-specific latent representations (embeddings) of entities and events and spatio-temporal event features in a learning to rank model. This model is trained on publicly available Wikipedia Clickstream data. The results of our user study demonstrate that LaSER outperforms state-of-the-art recommendation baselines by up to 33 percentage points in MAP@5 concerning the language-specific relevance of recommended events."}, "cited_paper_content": {"title": "Deepwalk: Online Learning Of Social Representations", "abstract": "We present DeepWalk, a novel approach for learning latent representations of vertices in a network. These latent representations encode social relations in a continuous vector space, which is easily exploited by statistical models. DeepWalk generalizes recent advancements in language modeling and unsupervised feature learning (or deep learning) from sequences of words to graphs. DeepWalk uses local information obtained from truncated random walks to learn latent representations by treating walks as the equivalent of sentences. We demonstrate DeepWalk's latent representations on several multi-label network classification tasks for social networks such as BlogCatalog, Flickr, and YouTube. Our results show that DeepWalk outperforms challenging baselines which are allowed a global view of the network, especially in the presence of missing information. DeepWalk's representations can provide $F_1$ scores up to 10% higher than competing methods when labeled data is sparse. In some experiments, DeepWalk's representations are able to outperform all baseline methods while using 60% less training data. DeepWalk is also scalable. It is an online learning algorithm which builds useful incremental results, and is trivially parallelizable. These qualities make it suitable for a broad class of real world applications such as network classification, and anomaly detection."}, "keywords": ["language-specific embeddings"], "citation_intent": "method"}
{"citing_id": "2304.14106v1", "cited_id": "1805.12471", "section_title": "A.1 Source Benchmark Statistics", "citation": "The CoLa task, which is based on The Corpus of Linguistic Acceptability #REFR , is a linguistic acceptability task.", "text_before_citation": ["\u2022 Aggression and AggressionPer.", "The Aggression and AggressionPer tasks, which are based on the Wikipedia Talk Labels dataset #OTHEREFR , are offensive language detection tasks.", "These tasks are presented in the form of binary classification, with the task being to classify text as either aggressive or non-aggressive.", "Specifically, AggressionPer uses the personalized variant by providing user annotations with prompts.", "\u2022 CoLa."], "text_after_citation": ["This task is presented in the form of binary classification, with the task being to classify text as either grammatically correct or not.", "\u2022 ColBERT.", "The ColBERT task, which is based on the ColBERT dataset #OTHEREFR , is a humor recognition task.", "This task is presented in the form of binary classification, with the task being to classify text as either humor or not humor.", "\u2022 Sarcasm The Sarcasm task, which is based on the Sarcasmania dataset #OTHEREFR , is sarcasm classification task."], "citing_paper_content": {"title": "Chatlog: Recording And Analyzing Chatgpt Across Time", "abstract": "While there are abundant researches about evaluating ChatGPT on natural language understanding and generation tasks, few studies have investigated how ChatGPT's behavior changes over time. In this paper, we collect a coarse-to-fine temporal dataset called ChatLog, consisting of two parts that update monthly and daily: ChatLog-Monthly is a dataset of 38,730 question-answer pairs collected every month including questions from both the reasoning and classification tasks. ChatLog-Daily, on the other hand, consists of ChatGPT's responses to 1000 identical questions for long-form generation every day. We conduct comprehensive automatic and human evaluation to provide the evidence for the existence of ChatGPT evolving patterns. We further analyze the unchanged characteristics of ChatGPT over time by extracting its knowledge and linguistic features. We find some stable features to improve the robustness of a RoBERTa-based detector on new versions of ChatGPT. We will continuously maintain our project at https://github.com/ THU-KEG/ChatLog."}, "cited_paper_content": {"title": "Neural Network Acceptability Judgments", "abstract": "This paper investigates the ability of artificial neural networks to judge the grammatical acceptability of a sentence, with the goal of testing their linguistic competence. We introduce the Corpus..."}, "keywords": ["linguistic acceptability task"], "citation_intent": "background"}
{"citing_id": "2304.10196v1", "cited_id": "1806.09031", "section_title": "I. Introduction", "citation": "There is a standard and very flexible technique for incorporating equality, and other \"enrichments\" of logics #REFR .", "text_before_citation": ["We introduce the categorical structure abstracting logical equivalence with respect to the studied logical fragment, establish the corresponding categorical FVM theorem and illustrate this with concrete examples.", "The examples in Section IV include a new refinement of a result by Dawar, Severini, and Zapata #OTHEREFR , showing that equivalence in 3variable counting logic without equality and with restricted conjunctions implies cospectrality.", "Section VI develops FVM theorems for products of structures, uniformly in the logic of interest, greatly generalising Mostowski's original result.", "Unlike our methods, which are parametric in the choice of game comonad, a typical game argument cannot be used to prove such a result, as the game is tied to a specific choice of logic.", "In fact, the game comonads E k and P k naturally deal with first-order logic without equality."], "text_after_citation": ["This is introduced and related to FVM theorems in Section VII, and illustrated with further examples, including the handling of weak-bisimulation."], "citing_paper_content": {"title": "A Categorical Account Of Composition Methods In Logic", "abstract": "We present a categorical theory of the composition methods in finite model theory-a key technique enabling modular reasoning about complex structures by building them out of simpler components. The crucial results required by the composition methods are Feferman-Vaught-Mostowski (FVM) type theorems, which characterize how logical equivalence behaves under composition and transformation of models. Our results are developed by extending the recently introduced game comonad semantics for model comparison games. This level of abstraction allow us to give conditions yielding FVM type results in a uniform way. Our theorems are parametric in the classes of models, logics and operations involved. Furthermore, they naturally account for the positive existential fragment, and extensions with counting quantifiers of these logics. We also reveal surprising connections between FVM type theorems, and classical concepts in the theory of monads. We illustrate our methods by recovering many classical theorems of practical interest, including a refinement of a previous result by Dawar, Severini, and Zapata concerning the 3-variable counting logic and cospectrality. To highlight the importance of our techniques being parametric in the logic of interest, we prove a family of FVM theorems for products of structures, uniformly in the logic in question, which cannot be done using specific game arguments."}, "cited_paper_content": {"title": "Relating Structure And Power: Comonadic Semantics For Computational Resources", "abstract": "Combinatorial games are widely used in finite model theory, constraint satisfaction, modal logic and concurrency theory to characterize logical equivalences between structures. In particular, Ehrenfeucht-Fraisse games, pebble games, and bisimulation games play a central role. We show how each of these types of games can be described in terms of an indexed family of comonads on the category of relational structures and homomorphisms. The index k is a resource parameter which bounds the degree of access to the underlying structure. The coKleisli categories for these comonads can be used to give syntax-free characterizations of a wide range of important logical equivalences. Moreover, the coalgebras for these indexed comonads can be used to characterize key combinatorial parameters: tree-depth for the Ehrenfeucht-Fraisse comonad, tree-width for the pebbling comonad, and synchronization-tree depth for the modal unfolding comonad. These results pave the way for systematic connections between two major branches of the field of logic in computer science which hitherto have been almost disjoint: categorical semantics, and finite and algorithmic model theory."}, "keywords": ["logics"], "citation_intent": "method"}
{"citing_id": "2305.01320v1", "cited_id": "1606.03135", "section_title": "Reconstruction Functions", "citation": "In this case, the reconstructions are symmetric and do not require more computation time than the Taylor series-based first-order reconstructions in equation #REFR .", "text_before_citation": ["\u03bb i\u2192 (x) = \u03bb(x) \u2212 1 2 \u2207 i \u03bb, x \u2212 x i (25b)", "yields first-order accurate reconstruction functions.", "Empirically, these lead to worse results than the Pythagorean means, possibly due to the asymmetry of the reconstructions \u03bb ij = \u03bb ji , while requiring additional computational steps. A higher-order scheme follows from a one-dimensional Hermite interpolation", "EQUATION", "with the directional vector d ij = x j \u2212 x i ."], "text_after_citation": [], "citing_paper_content": {"title": "Higher-Order Gfdm For Linear Elliptic Operators *", "abstract": "We present a novel approach of discretizing diffusion operators of the form \u2207 \u2022 (\u03bb\u2207u) in the context of meshfree generalized finite difference methods. Our ansatz uses properties of derived operators and combines the discrete Laplace operator with reconstruction functions approximating the diffusion coefficient \u03bb. Provided that the reconstructions are of a sufficiently high order, we prove that the order of accuracy of the discrete Laplace operator transfers to the derived diffusion operator. We show that the new discrete diffusion operator inherits the diagonal dominance property of the discrete Laplace operator and fulfills enrichment properties. Our numerical results for elliptic and parabolic partial differential equations show that even loworder reconstructions preserve the order of the underlying discrete Laplace operator for sufficiently smooth diffusion coefficients. In experiments, we demonstrate the applicability of the new discrete diffusion operator to interface problems with point clouds not aligning to the interface and numerically prove first-order convergence."}, "cited_paper_content": {"title": "The Overlapped Radial Basis Function-Finite Difference (Rbf-Fd) Method For The Numerical Solution Of Pdes", "abstract": "We present a generalization of the RBF-FD method that allows full control of the overlap between RBF-FD stencils. We accomplish this by introducing a continuous overlap parameter $\\delta \\in [0,1]$ such that $\\delta=1$ recovers the standard RBF-FD method and $\\delta=0$ results in a full decoupling of the RBF-FD stencils. We show with a simple example that global interpolation with both RBFs and RBFs augmented with polynomials is superior to polynomial least squares in confining Runge oscillations to the boundary, thereby justifying retaining more RBF-FD weights from each RBF-FD stencil than just the center; this also provides greater insight into the success of the recent polyharmonic spline RBF-FD method with high-degree polynomials appended. We demonstrate the efficiency of the \\emph{overlapped RBF-FD method} by solving parabolic differential equations in 2D and 3D. Our results show that the new method can achieve as high as a 20x speedup over the standard RBF-FD method in the task of forming differentiation matrices. Further, the overlapped RBF-FD method allows for stencil sizes up to 6x larger than those usable in the standard RBF-FD method for a lower total computational cost."}, "keywords": ["Taylor series-based first-order"], "citation_intent": "background"}
{"citing_id": "2304.06136v1", "cited_id": "1706.03762", "section_title": "Background", "citation": "Similar to traditional PLMs like BERT, these recent LLMs also follow the Transformer architecture #REFR , but they differ from each other in training objectives, data sources, optimization details, etc.", "text_before_citation": ["In #OTHEREFR , a largescale corpus of agricultural literature with more than 300 million tokens is used to fine-tune the generic BERT into AgriBERT.", "Such domain-specific language models could be potentially used to assist many agricultural applications, such as agricultural marketing, agricultural social media analytics, etc.", "Large Language Models (LLMs).", "LLMs usually refer to PLMs with more than billions of model parameters, such as the GPT (Generative Pre-trained Transformer) models (GPT-1 #OTHEREFR , GPT-2 #OTHEREFR , GPT-3 #OTHEREFR , Instruct-GPT #OTHEREFR , ChatGPT, and GPT-4 [125] ), LLaMA (Large Language Model Meta AI) #OTHEREFR , and BLOOM #OTHEREFR .", "For instance, GPT-3 and BLOOM have 175 billion parameters, and LLaMA has 65 billion parameters."], "text_after_citation": ["For example, BERT only adopts the encoder of Transformer as it focuses on learning language representations, while GPT models adopt the decoder of Transformer, as their major objective is to generate realistic text.", "Owing to their huge model sizes and effective incorporation of human feedback (e.g., prompts), the latest LLMs like ChatGPT and GPT-4 have not only achieved amazing performance in traditional NLP tasks like language understanding, language generation, and question answering, but also demonstrated remarkable capabilities in zero-shot learning and reasoning #OTHEREFR .", "For instance, ChatGPT has shown a very promising capability in inductive reasoning and deductive reasoning #OTHEREFR .", "GPT-4 is a large multimodal model that can deal with both language and visual inputs.", "In addition, LLMs are also effective tools for data augmentation #OTHEREFR , which is very helpful for certain real-world scenarios with limited data."], "citing_paper_content": {"title": "Agi For Agriculture", "abstract": "Artificial General Intelligence (AGI) is poised to revolutionize a variety of sectors, including healthcare, finance, transportation, and education. Within healthcare, AGI is being utilized to analyze clinical medical notes, recognize patterns in patient data, and aid in patient management. Agriculture is another critical sector that impacts the lives of individuals worldwide. It serves as a foundation for providing food, fiber, and fuel, yet faces several challenges, such as climate change, soil degradation, water scarcity, and food security. AGI has the potential to tackle these issues by enhancing crop yields, reducing waste, and promoting sustainable farming practices. It can also help farmers make informed decisions by leveraging real-time data, leading to more efficient and effective farm management. This paper delves into the potential future applications of AGI in agriculture, such as agriculture image processing, natural language processing (NLP), robotics, knowledge graphs, and infrastructure, and their impact on precision livestock and precision crops. By leveraging the power of AGI, these emerging technologies can provide farmers with actionable insights, allowing for optimized decisionmaking and increased productivity. The transformative potential of AGI in agriculture is vast, and this paper aims to highlight its potential to revolutionize the industry."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["BERT", "Transformer architecture"], "citation_intent": "background"}
{"citing_id": "2305.02909v1", "cited_id": "1908.09492", "section_title": "B. Implementation Details", "citation": "Furthermore, we adopt the ground truth sampling strategy of #REFR which randomly takes boxes and their points from a database and places them in the input point cloud.", "text_before_citation": ["The point cloud of a non keyframe collected at timestamp t \u2212 k (k = 1, . . .", ", 9) are mapped from the LiDAR frame at this time step to the world frame W using ego vehicle poses and LiDAR calibration.", "The ground truth of rectification transformation in (2) requires knowing objects' poses in the world frame W at the keyframe timestamp W T O(t) and non keyframe timestamp W T O(t\u2212k) .", "Since NuScenes only provides objects' poses in keyframes, we obtain their poses in non keyframes by linearly interpolating annotations of two keyframes that are respectively prior and successor to each non keyframe.", "To improve the generalization of our model, the following geometric transformations are applied to point clouds and ground truth: random flip along the x-and y-axis of the world frame W, global scaling with a factor sampled from U [0.95,1.05] , and global rotation by an angle sampled from U [\u2212\u03c0/8,\u03c0/8] around the z-axis of the world frame W. Here, U denotes the uniform distribution."], "text_after_citation": ["Notably, we introduce a modification to this sampling strategy by complementing each sampled ground truth with points in its trajectory spanning from the current time step to 10 steps in the past.", "This modification is similar to the \"Sequence GtAug\" of #OTHEREFR .", "Its motivation is to increase the number of moving objects in each point cloud, thus increasing the amount of supervision on the Object Motion Estimation.", "A comparison between the regular ground truth sampling and our modified version is shown in Fig. 3 .", "In our experiments, input point clouds are limited to the range of [\u221251.2, 51.2] \u00d7 [\u221251.2, 51.2] \u00d7 [\u22125.0, 3.0] meters along X-, Y-, and Z-axis of the world frame W and the voxel size is set to (0.1, 0.1, 0.2). We use OpenPCDet #OTHEREFR for our implementation."], "citing_paper_content": {"title": "Aligning Bird-Eye View Representation Of Point Cloud Sequences Using Scene Flow", "abstract": "Low-resolution point clouds are challenging for object detection methods due to their sparsity. Densifying the present point cloud by concatenating it with its predecessors is a popular solution to this challenge. Such concatenation is possible thanks to the removal of ego vehicle motion using its odometry. This method is called Ego Motion Compensation (EMC). Thanks to the added points, EMC significantly improves the performance of single-frame detectors. However, it suffers from the shadow effect that manifests in dynamic objects' points scattering along their trajectories. This effect results in a misalignment between feature maps and objects' locations, thus limiting performance improvement to stationary and slow-moving objects only. Scene flow allows aligning point clouds in 3D space, thus naturally resolving the misalignment in feature spaces. By observing that scene flow computation shares several components with 3D object detection pipelines, we develop a plug-in module that enables single-frame detectors to compute scene flow to rectify their Bird-Eye View representation. Experiments on the NuScenes dataset show that our module leads to a significant increase (up to 16%) in the Average Precision of large vehicles, which interestingly demonstrates the most severe shadow effect. The code is published at https://github.com/quan-dao/pc-corrector. Index Terms-object detection, deep learning, scene flow, perception, LiDAR \u2022 Scene flow pipelines require points' features which can be computed by first converting input point clouds to BEV representation, then interpolating using points' projection"}, "cited_paper_content": {"title": "Class-Balanced Grouping And Sampling For Point Cloud 3D Object Detection", "abstract": "This report presents our method which wins the nuScenes3D Detection Challenge [17] held in Workshop on Autonomous Driving(WAD, CVPR 2019). Generally, we utilize sparse 3D convolution to extract rich semantic features, which are then fed into a class-balanced multi-head network to perform 3D object detection. To handle the severe class imbalance problem inherent in the autonomous driving scenarios, we design a class-balanced sampling and augmentation strategy to generate a more balanced data distribution. Furthermore, we propose a balanced group-ing head to boost the performance for the categories withsimilar shapes. Based on the Challenge results, our methodoutperforms the PointPillars [14] baseline by a large mar-gin across all metrics, achieving state-of-the-art detection performance on the nuScenes dataset. Code will be released at CBGS."}, "keywords": ["input point cloud"], "citation_intent": "method"}
{"citing_id": "2305.01267v1", "cited_id": "1602.05629", "section_title": "Introduction", "citation": "Recently, federated learning (FL) #REFR has been widely studied as a privacypreserving distributed training paradigm, where clients cooperatively train a machine learning (ML) model under the coordination of a central server. FL training consists of multiple communication rounds.", "text_before_citation": [], "text_after_citation": ["In each communication round, the server first broadcasts a global model to the clients.", "Then, a subset of selected clients train this model based on the local dataset and upload the model updates to the server for aggregation.", "Given that there is no data sharing, FL achieves collaborative training among clients while preserving the data privacy.", "However, recent studies #OTHEREFR demonstrate that FL is vulnerable to model attacks due to the data and device heterogeneity of clients.", "Backdoor attack #OTHEREFR ) misleads an ML model to misclassify the data with specific triggers into a certain label."], "citing_paper_content": {"title": "Dabs: Data-Agnostic Backdoor Attack At The Server In Federated Learning", "abstract": "Federated learning (FL) attempts to train a global model by aggregating local models from distributed devices under the coordination of a central server. However, the existence of a large number of heterogeneous devices makes FL vulnerable to various attacks, especially the stealthy backdoor attack. Backdoor attack aims to trick a neural network to misclassify data to a target label by injecting specific triggers while keeping correct predictions on original training data. Existing works focus on client-side attacks which try to poison the global model by modifying the local datasets. In this work, we propose a new attack model for FL, namely Data-Agnostic Backdoor attack at the Server (DABS), where the server directly modifies the global model to backdoor an FL system. Extensive simulation results show that this attack scheme achieves a higher attack success rate compared with baseline methods while maintaining normal accuracy on the clean data."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["federated learning", "privacypreserving distributed training"], "citation_intent": "background"}
{"citing_id": "2304.10136v1", "cited_id": "1905.03696", "section_title": "Features Under Quantitative Approaches", "citation": "The HAWQ algorithm #REFR utilizes the Hessian eigenvalues as the sensitivity measurement matrix for hierarchical features in the network and adopts mixed-precision quantization for network layers with divergent sensitivities.", "text_before_citation": ["EQUATION", "where (r j , r j + 1] represents the range of real values depending on the quantization bits k, i.e., j = 0, ..., 2 k \u2212 1.", "The quantization process map this range (r j , r j + 1] to the value q j .", "However, directly using a low-bit precision to quantize a model will decrease the model accuracy.", "A promising approach to achieve a trade-off between model size and accuracy is to mix different precisions to quantize models."], "text_after_citation": ["This motivates us to consider sensitivity when modifying network features.", "In other words, changing parts of the model at different levels will have different effects on how the model works.", "To measure network layers, we adopt the average Hessian trace H proposed by the HAWQ-V2 algorithm as the sensitivity matrix. We have the following Lemma: Lemma 1 #OTHEREFR .", "Suppose that the model is twice differentiable and converges to a local minimum when we quantize two layers with the same perturbation, we have:", "EQUATION"], "citing_paper_content": {"title": "Diversifying The High-Level Features For Better Adversarial Transferability", "abstract": "Given the great threat of adversarial attacks against Deep Neural Networks (DNNs), numerous works have been proposed to boost transferability to attack real-world applications. However, existing attacks often utilize advanced gradient calculation or input transformation but ignore the white-box model. Inspired by the fact that DNNs are overparameterized for superior performance, we propose diversifying the high-level features (DHF) for more transferable adversarial examples. In particular, DHF perturbs the highlevel features by randomly transforming the high-level features and mixing them with the feature of benign samples when calculating the gradient at each iteration. Due to the redundancy of parameters, such transformation does not affect the classification performance but helps identify the invariant features across different models, leading to much better transferability. Empirical evaluations on ImageNet dataset show that DHF could effectively improve the transferability of existing momentum-based attacks. Incorporated into the input transformation-based attacks, DHF generates more transferable adversarial examples and outperforms the baselines with a clear margin when attacking several defense models, showing its generalization to various attacks and high effectiveness for boosting transferability. Recent studies have shown that Deep Neural Networks (DNNs) are vulnerable to adversarial examples [46], i.e., the inputs with indistinguishable perturbation can mislead the DNNs. Such vulnerability brings a significant threat to the security of widely deployed DNNs in the physical world, e.g."}, "cited_paper_content": {"title": "Hawq: Hessian Aware Quantization Of Neural Networks With Mixed-Precision", "abstract": "Model size and inference speed/power have become a major challenge in the deployment of Neural Networks for many applications. A promising approach to address these problems is quantization. However, uniformly quantizing a model to ultra low precision leads to significant accuracy degradation. A novel solution for this is to use mixed-precision quantization, as some parts of the network may allow lower precision as compared to other layers. However, there is no systematic way to determine the precision of different layers. A brute force approach is not feasible for deep networks, as the search space for mixed-precision is exponential in the number of layers. Another challenge is a similar factorial complexity for determining block-wise fine-tuning order when quantizing the model to a target precision. Here, we introduce Hessian AWare Quantization (HAWQ), a novel second-order quantization method to address these problems. HAWQ allows for the automatic selection of the relative quantization precision of each layer, based on the layer's Hessian spectrum. Moreover, HAWQ provides a deterministic fine-tuning order for quantizing layers, based on second-order information. We show the results of our method on Cifar-10 using ResNet20, and on ImageNet using Inception-V3, ResNet50 and SqueezeNext models. Comparing HAWQ with state-of-the-art shows that we can achieve similar/better accuracy with $8\\times$ activation compression ratio on ResNet20, as compared to DNAS~\\cite{wu2018mixed}, and up to $1\\%$ higher accuracy with up to $14\\%$ smaller models on ResNet50 and Inception-V3, compared to recently proposed methods of RVQuant~\\cite{park2018value} and HAQ~\\cite{wang2018haq}. Furthermore, we show that we can quantize SqueezeNext to just 1MB model size while achieving above $68\\%$ top1 accuracy on ImageNet."}, "keywords": ["mixed-precision quantization"], "citation_intent": "method"}
{"citing_id": "2304.02714v1", "cited_id": "1812.04948", "section_title": "Vi. Conclusion And Discussion", "citation": "For example, we may use a generator architecture based on a styletransfer network which improves the generated sample quality #REFR .", "text_before_citation": ["In our whistle extraction experiments, we did not use this extended idea, because the appearance of our foreground object (the whistles) is relatively simple, i.e., the variance of appearance is mainly rooted in the whistle contour shape and whistle magnitude.", "Therefore, we directly add whistle signals to the background using the third GAN in our framework.", "Our framework can be readily extended to extract calls of other whale species (e.g, baleen whales) and to other similar tasks (e.g., semantic image segmentation).", "Though it may not affect the main contributions of this work, our data generation method can be improved in three aspects in the future.", "Firstly, we may use improved generative neural network architecture and training strategies."], "text_after_citation": ["The discriminator augmentation mechanism proposed in #OTHEREFR may help stabilize training in limited data regimes.", "We may also explore generating larger patches of high quality with the method in #OTHEREFR .", "Secondly, we may use real data in the data generation process to enrich the data variance.", "The real background and annotated whistle contours can be used as the input data of our GAN in the third stage, and we can generate whistle signsals of novel shapes on real background or generate whistle signals of annotated contour shapes on GANgenerated background. Thirdly, we may improve the sample selection method.", "In this paper, we use a simple yet effective pixel-wise entropy method to select whistle contour of good quality."], "citing_paper_content": {"title": "Learning Stage-Wise Gans For Whistle Extraction In Time-Frequency Spectrograms", "abstract": "Whistle contour extraction aims to derive animal whistles from time-frequency spectrograms as polylines. For toothed whales, whistle extraction results can serve as the basis for analyzing animal abundance, species identity, and social activities. During the last few decades, as long-term recording systems have become affordable, automated whistle extraction algorithms were proposed to process large volumes of recording data. Recently, a deep learning-based method demonstrated superior performance in extracting whistles under varying noise conditions. However, training such networks requires a large amount of labor-intensive annotation, which is not available for many species. To overcome this limitation, we present a framework of stage-wise generative adversarial networks (GANs), which compile new whistle data suitable for deep model training via three stages: generation of background noise in the spectrogram, generation of whistle contours, and generation of whistle signals. By separating the generation of different components in the samples, our framework composes visually promising whistle data and labels even when few expert annotated data are available. Regardless of the amount of human-annotated data, the proposed data augmentation framework leads to a consistent improvement in performance of the whistle extraction model, with a maximum increase of 1.69 in the whistle extraction mean F1-score. Our stage-wise GAN also surpasses one single GAN in improving whistle extraction models with augmented data."}, "cited_paper_content": {"title": "A Style-Based Generator Architecture For Generative Adversarial Networks", "abstract": "We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces."}, "keywords": ["styletransfer network"], "citation_intent": "method"}
{"citing_id": "2303.03811v1", "cited_id": "1801.01290", "section_title": "A. Policy Optimization", "citation": "We adopt soft-actor critic (SAC) #REFR as our policy optimization algorithm, which demonstrates impressive performance for continuous tasks.", "text_before_citation": ["\u03c0 * = arg max \u03c0 E \u221e t=0 r (s t , a t ) .", "(1)", "In maximum entropy reinforcement learning, the expected entropy term is augmented #OTHEREFR , which helps the agents explore more widely and give up on clearly unpromising avenues.", "In this case, the optimal policy \u03c0 * with parameter \u03b1 that determines the relative importance of the expected entropy term can be written as:", "\u03c0 * = arg max \u03c0 E \u221e t=0 [r (s t , a t ) + \u03b1H (\u03c0 ( \u2022 |s t ))] . (2)"], "text_after_citation": ["In the policy improvement step, the policy parameters can be directly updated by minimizing the following:", "EQUATION", "where \u03c6 represents parameters of policy network \u03c0 \u03c6 (a t |s t ) and \u03b8 are parameters of soft Q-function Q \u03b8 (s t , a t ), D is the distribution of states and actions sampled from interaction experience, and Z \u03b8 (s t ) is the partition function that normalizes the distribution.", "For our offline scenario, D is the distribution of the origin offline dataset combined with a replay buffer which contains simulated trajectories generated by the learned model.", "Finally, actions a t are generated by re-parameterizing the policy using noise vector t :"], "citing_paper_content": {"title": "Entropy: Environment Transformer And Offline Policy Optimization", "abstract": "Model-based methods provide an effective approach to offline reinforcement learning (RL). They learn an environmental dynamics model from interaction experiences and then perform policy optimization based on the learned model. However, previous model-based offline RL methods lack long-term prediction capability, resulting in large errors when generating multi-step trajectories. We address this issue by developing a sequence modeling architecture, Environment Transformer, which can generate reliable long-horizon trajectories based on offline datasets. We then propose a novel model-based offline RL algorithm, ENTROPY, that learns the dynamics model and reward function by ENvironment TRansformer and performs Offline PolicY optimization. We evaluate the proposed method on MuJoCo continuous control RL environments. Results show that ENTROPY performs comparably or better than the state-of-the-art model-based and model-free offline RL methods and demonstrates more powerful long-term trajectory prediction capability compared to existing model-based offline methods."}, "cited_paper_content": {"title": "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning With A Stochastic Actor", "abstract": "Model-free deep reinforcement learning (RL) algorithms have been demonstrated on a range of challenging decision making and control tasks. However, these methods typically suffer from two major challenges: very high sample complexity and brittle convergence properties, which necessitate meticulous hyperparameter tuning. Both of these challenges severely limit the applicability of such methods to complex, real-world domains. In this paper, we propose soft actor-critic, an off-policy actor-critic deep RL algorithm based on the maximum entropy reinforcement learning framework. In this framework, the actor aims to maximize expected reward while also maximizing entropy - that is, succeed at the task while acting as randomly as possible. Prior deep RL methods based on this framework have been formulated as Q-learning methods. By combining off-policy updates with a stable stochastic actor-critic formulation, our method achieves state-of-the-art performance on a range of continuous control benchmark tasks, outperforming prior on-policy and off-policy methods. Furthermore, we demonstrate that, in contrast to other off-policy algorithms, our approach is very stable, achieving very similar performance across different random seeds."}, "keywords": ["policy optimization algorithm", "soft-actor critic"], "citation_intent": "method"}
{"citing_id": "2305.02459v1", "cited_id": "1706.04599", "section_title": "Model Update", "citation": "This is contrastive to the prior literature revealing that poor calibration of large language models often renders the models to rarely be uncertain of their outcomes #REFR .", "text_before_citation": ["The domain transfer from both tasks (or a combination of them) gives the active annotation a head-start for initial sample selection.", "Table 3 shows the overlap of samples picked out in each iteration from the same larger data pool for the model at iteration 0 (base model).", "RAN-DOM has the lowest overlaps with all the other strategies.", "We also find that there is a significant overlap (> 80%) in the samples between ENTROPY or CAL, the state-of-the-art approach, and PRC.", "CAL has a higher overlap with ENTROPY rather than CORESET, showing that samples deemed to be both highly informative and contrastive by the model are also usually likely to be dissonant."], "text_after_citation": ["All strategies except RANDOM have > 55% overlap with each other.", "This implies that diversity-and uncertainty-based methods are not as different from each other as they theoretically are and inclined to pick similar samples -hinting that a lot of diversity-based sampling measures mostly pick highly informative samples as well.", "Furthermore, PRC tends to choose samples that the \"state-of-the-art\" model also picks in rare-case scenarios, indicating that it could be a computationally inexpensive alternative.", "Table 4 shows the results averaged over two rounds of active annotation and learning for five strategies with two types of model updates.", "While the performance for dissonance class across all strategies do not seem to boost much in a single round of active learning (since adding 300 new annotations adds only between 10-30 dissonance examples in each round), Figure 3 shows that the CM approach always performs better than IT."], "citing_paper_content": {"title": "Transfer And Active Learning For Dissonance Detection: Addressing The Rare-Class Challenge", "abstract": "While transformer-based systems have enabled greater accuracies with fewer training examples, data acquisition obstacles still persist for rare-class tasks-when the class label is very infrequent (e.g. < 5% of samples). Active learning has in general been proposed to alleviate such challenges, but choice of selection strategy, the criteria by which rareclass examples are chosen, has not been systematically evaluated. Further, transformers enable iterative transfer-learning approaches. We propose and investigate transfer-and active learning solutions to the rare class problem of dissonance detection through utilizing models trained on closely related tasks and the evaluation of acquisition strategies, including a proposed probability-of-rare-class (PRC) approach. We perform these experiments for a specific rare class problem: collecting language samples of cognitive dissonance from social media. We find that PRC is a simple and effective strategy to guide annotations and ultimately improve model accuracy while transferlearning in a specific order can improve the cold-start performance of the learner but does not benefit iterations of active learning."}, "cited_paper_content": {"title": "On Calibration Of Modern Neural Networks", "abstract": "Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions."}, "keywords": ["large language models", "poor calibration"], "citation_intent": "result"}
{"citing_id": "2304.05634v1", "cited_id": "1810.02508", "section_title": "Related Work", "citation": "The Multimodal EmotionLines Dataset (MELD) #REFR is an example of Emotion Recognition in Conversations (ERC) and attempts to estimate the emotion for every dialog utterance in TV episodes from Friends.", "text_before_citation": ["Different from the above, our work focuses on multi-label emotions and mental states recognition in movies exploiting multimodal context both at the scene-and character-level.", "Multimodal datasets for emotion recognition have seen recent adoption.", "Acted Facial Expressions in the Wild #OTHEREFR aims to predict emotions from faces, but does not provide any context.", "The Stanford Emotional Narratives Dataset #OTHEREFR contains participant shared narratives of positive/negative events in their lives.", "While multimodal, these are quite different from edited movies and stories that are our focus."], "text_after_citation": ["Different from MELD, we operate at the time-scale of a cohesive story unit, a movie scene.", "Finally, closest to our work, Annotated Creative Commons Emotional DatabasE (LIRIS-ACCEDE) #OTHEREFR obtains emotion annotations for short movie clips.", "However, the clips are quite small (8-12s) and annotations are obtained in the continuous valence-arousal space.", "Different from the above works, we also aim to predict character-level mental states and demonstrate that video and dialog context helps for such labels.", "Multimodal emotion recognition methods."], "citing_paper_content": {"title": "How You Feelin'? Learning Emotions And Mental States In Movie Scenes", "abstract": "Movie story analysis requires understanding characters' emotions and mental states. Towards this goal, we formulate emotion understanding as predicting a diverse and multi-label set of emotions at the level of a movie scene and for each character. We propose EmoTx, a multimodal Transformer-based architecture that ingests videos, multiple characters, and dialog utterances to make joint predictions. By leveraging annotations from the MovieGraphs dataset [74], we aim to predict classic emotions (e.g. happy, angry) and other mental states (e.g. honest, helpful). We conduct experiments on the most frequently occurring 10 and 25 labels, and a mapping that clusters 181 labels to 26. Ablation studies and comparison against adapted stateof-the-art emotion recognition approaches shows the effectiveness of EmoTx. Analyzing EmoTx's self-attention scores reveals that expressive emotions often look at character tokens while other mental states rely on video and dialog cues."}, "cited_paper_content": {"title": "Meld: A Multimodal Multi-Party Dataset For Emotion Recognition In Conversations", "abstract": "Emotion recognition in conversations is a challenging Artificial Intelligence (AI) task. Recently, it has gained popularity due to its potential applications in many interesting AI tasks such as empathetic dialogue generation, user behavior understanding, and so on. To the best of our knowledge, there is no multimodal multi-party conversational dataset available, which contains more than two speakers in a dialogue. In this work, we propose the Multimodal EmotionLines Dataset (MELD), which we created by enhancing and extending the previously introduced EmotionLines dataset. MELD contains 13,708 utterances from 1433 dialogues of Friends TV series. MELD is superior to other conversational emotion recognition datasets SEMAINE and IEMOCAP as it consists of multiparty conversations and number of utterances in MELD is almost twice as these two datasets. Every utterance in MELD is associated with an emotion and a sentiment label. Utterances in MELD are multimodal encompassing audio and visual modalities along with the text. We have also addressed several shortcomings in EmotionLines and proposed a strong multimodal baseline. The baseline results show that both contextual and multimodal information play important role in emotion recognition in conversations."}, "keywords": ["Multimodal EmotionLines"], "citation_intent": "background"}
{"citing_id": "2304.01343v1", "cited_id": "1505.05116", "section_title": "Related Literature", "citation": "For example, we refer to the study by Esfahani and Kuhn #REFR , where the ambiguity set is formed by a ball in the space of (multivariate and non-discrete) probability distributions with respect to the Wasserstein metric.", "text_before_citation": ["As the first example, we refer to the study by Delage and Ye #OTHEREFR , who consider convex optimization problems with random parameters in the objective function.", "The ambiguity set in #OTHEREFR is constructed by leveraging a given convex support and confidence sets for the mean and the covariance matrix of the uncertain problem parameters. Then it is", "shown that for several particular forms of the loss function (x, c), the proposed distributionally robust formulation can be solved efficiently by leveraging strong duality results for moment problems #OTHEREFR .", "Despite the fact that moment-based ambiguity sets are rather standard in the DRO literature; see, e.g., #OTHEREFR , it is demonstrated in #OTHEREFR that using the second-order moment constraints with a linear mixed-integer set of feasible decisions X results in a non-linear single-level MIP reformulation, which can only be approximated by a sequence of semi-definite programming relaxations.", "In this regard, several related studies focus on data-driven DRO problems, which admit linear MIP reformulations."], "text_after_citation": ["The center of the ball is at the uniform distribution on the training samples and the radius can be viewed as a decreasing function in the sample size.", "The authors demonstrate that, if the Wasserstein metric is defined in terms of l 1 -norm or l \u221e -norm, then under some mild assumptions the associated worst-case expectation problem admits an equivalent dual linear programming reformulation.", "Furthermore, whenever the set X is linear mixed-integer, the associated DRO problem can be recast as a linear MIP problem; see, e.g., #OTHEREFR .", "On the other hand, Bertsimas et al. [6] analyze a data-driven robust optimization approach.", "Instead of the stochastic programming formulation (1), the authors in #OTHEREFR consider an associated robust formulation of the form:"], "citing_paper_content": {"title": "Distributionally Robust Mixed-Integer Programming With Wasserstein Metric: On The Value Of Uncertain Data", "abstract": "This study addresses a class of linear mixed-integer programming (MIP) problems that involve uncertainty in the objective function coefficients. The coefficients are assumed to form a random vector, which probability distribution can only be observed through a finite training data set. Unlike most of the related studies in the literature, we also consider uncertainty in the underlying data set. The data uncertainty is described by a set of linear constraints for each random sample, and the uncertainty in the distribution (for a fixed realization of data) is defined using a type-1 Wasserstein ball centered at the empirical distribution of the data. The overall problem is formulated as a three-level distributionally robust optimization (DRO) problem. We prove that for a class of bi-affine loss functions the three-level problem admits a linear MIP reformulation. Furthermore, it turns out that in several important particular cases the three-level problem can be solved reasonably fast by leveraging the nominal MIP problem. Finally, we conduct a computational study, where the out-of-sample performance of our model and computational complexity of the proposed MIP reformulation are explored numerically for several application domains."}, "cited_paper_content": {"title": "Data-Driven Distributionally Robust Optimization Using The Wasserstein Metric: Performance Guarantees And Tractable Reformulations", "abstract": "We consider stochastic programs where the distribution of the uncertain parameters is only observable through a finite training dataset. Using the Wasserstein metric, we construct a ball in the space of (multivariate and non-discrete) probability distributions centered at the uniform distribution on the training samples, and we seek decisions that perform best in view of the worst-case distribution within this Wasserstein ball. The state-of-the-art methods for solving the resulting distributionally robust optimization problems rely on global optimization techniques, which quickly become computationally excruciating. In this paper we demonstrate that, under mild assumptions, the distributionally robust optimization problems over Wasserstein balls can in fact be reformulated as finite convex programs\u2014in many interesting cases even as tractable linear programs. Leveraging recent measure concentration results, we also show that their solutions enjoy powerful finite-sample performance guarantees. Our theoretical results are exemplified in mean-risk portfolio optimization as well as uncertainty quantification."}, "keywords": ["Wasserstein metric"], "citation_intent": "background"}
{"citing_id": "2304.06619v1", "cited_id": "2003.03901", "section_title": "Faster-Ilod", "citation": "Faster-ILOD #REFR adopts a similar approach to continual object detection than ILOD, using KD as the approach to limit forgetting.", "text_before_citation": [], "text_after_citation": ["However, rather than only distilling knowledge from the last layer, namely the class logits and box coordinates, the authors add a KD loss between the feature maps of the teacher and student networks, and another for the RPN.", "This time, a L1 loss is used for distillation of feature maps activations as follows:", "EQUATION", "where H, W and D are the height, width and depth of the feature maps, respectively.", "Notice that KD is only applied if the activation of the previous model is larger than the one of the current model."], "citing_paper_content": {"title": "Class-Incremental Learning Of Plant And Disease Detection: Growing Branches With Knowledge Distillation", "abstract": "This paper investigates the problem of class-incremental object detection for agricultural applications where a model needs to learn new plant species and diseases incrementally without forgetting the previously learned ones. We adapt two public datasets to include new categories over time, simulating a more realistic and dynamic scenario. We then compare three class-incremental learning methods that leverage different forms of knowledge distillation to mitigate catastrophic forgetting. Our experiments show that all three methods suffer from catastrophic forgetting, but the recent Dynamic Y-KD approach, which additionally uses a dynamic architecture that grows new branches to learn new tasks, outperforms ILOD and Faster-ILOD in most scenarios both on new and old classes. These results highlight the challenges and opportunities of continual object detection for agricultural applications. In particular, the large intra-class and small interclass variability that is typical of plant images exacerbate the difficulty of learning new categories without interfering with previous knowledge. We publicly release our code to encourage future work. 1"}, "cited_paper_content": {"title": "Faster Ilod: Incremental Learning For Object Detectors Based On Faster Rcnn", "abstract": "The human vision and perception system is inherently incremental where new knowledge is continually learned over time whilst existing knowledge is retained. On the other hand, deep learning networks are ill-equipped for incremental learning. When a well-trained network is adapted to new categories, its performance on the old categories will dramatically degrade. To address this problem, incremental learning methods have been explored to preserve the old knowledge of deep learning models. However, the state-of-the-art incremental object detector employs an external fixed region proposal method that increases overall computation time and reduces accuracy compared to object detectors such as Faster RCNN that use trainable Region Proposal Networks (RPNs). The purpose of this paper is to design an efficient end-to-end incremental object detector using knowledge distillation for object detectors based on RPNs. We first evaluate and analyze the performance of RPN-based detector with classic distillation towards incremental detection tasks. Then, we introduce multi-network adaptive distillation that properly retains knowledge from the old categories when fine-turning the model for new task. Experiments on the benchmark datasets, PASCAL VOC and COCO, demonstrate that the proposed incremental detector is more accurate as well as being 13 times faster than the baseline detector."}, "keywords": ["continual object detection"], "citation_intent": "method"}
{"citing_id": "2303.16342v1", "cited_id": "1904.07750", "section_title": "Audio Separation Model", "citation": "Prior work #REFR has demonstrated the advantages of performing audio-visual separation at the object level for videos with multiple objects.", "text_before_citation": ["To predict the separated audio source, each element of the mask is multiplied with the corresponding location in the input spectrogram: A S =M A S , where denotes the Hadamard product. Video and Multiple Instance Learning for mask prediction.", "Since language annotations are not used during training, we train only with unlabeled videos.", "During training, as well as inference on audio-visual separation, we use the CLIP visual encoder to extract a set of language-grounded spatiotemporal region representations f V \u2208 R T \u00d7H\u00d7W \u00d7D .", "Similarly, we can perform the same mask inference operation using videos by using the representation of each spatiotemporal region representation interchangeably in place of the language representation.", "This interchangeability is due to the fact that the CLIP representations project the video to the joint vision-language embedding space."], "text_after_citation": ["In the absence of object detectors and labels, we propose an MIL formulation, where all spatiotemporal regions in the input video are treated as positive candidates.", "Specifically, given a spatiotemporal grid of region representations extracted from the visual encoder as defined in this section, we predict a spectrogram mask M j for the j-th region.", "Finally, we aggregate them over all regions to obtain a spectrogram mask for the entire video:", "EQUATION", "Intuitively, this formulation encourages the model to learn to identify the salient regions with sounding objects for audio separation."], "citing_paper_content": {"title": "Language-Guided Audio-Visual Source Separation Via Trimodal Consistency", "abstract": "We propose a self-supervised approach for learning to perform audio source separation in videos based on natural language queries, using only unlabeled video and audio pairs as training data. A key challenge in this task is learning to associate the linguistic description of a soundemitting object to its visual features and the corresponding components of the audio waveform, all without access to annotations during training. To overcome this challenge, we adapt off-the-shelf vision-language foundation models to provide pseudo-target supervision via two novel loss functions and encourage a stronger alignment between the audio, visual and natural language modalities. During inference, our approach can separate sounds given text, video and audio input, or given text and audio input alone. We demonstrate the effectiveness of our self-supervised approach on three audiovisual separation datasets, including MUSIC, SOLOS and AudioSet, where we outperform state-of-the-art strongly supervised approaches despite not using object detectors or text labels during training. Our project page including publicly available code can be found at https://cs-people.bu.edu/rxtan/projects/VAST."}, "cited_paper_content": {"title": "Co-Separating Sounds Of Visual Objects", "abstract": "Learning how objects sound from video is challenging, since they often heavily overlap in a single audio channel. Current methods for visually-guided audio source separation sidestep the issue by training with artificially mixed video clips, but this puts unwieldy restrictions on training data collection and may even prevent learning the properties of \"true\" mixed sounds. We introduce a co-separation training paradigm that permits learning object-level sounds from unlabeled multi-source videos. Our novel training objective requires that the deep neural network's separated audio for similar-looking objects be consistently identifiable, while simultaneously reproducing accurate video-level audio tracks for each source training pair. Our approach disentangles sounds in realistic test videos, even in cases where an object was not observed individually during training. We obtain state-of-the-art results on visually-guided audio source separation and audio denoising for the MUSIC, AudioSet, and AV-Bench datasets."}, "keywords": ["audio-visual separation"], "citation_intent": "background"}
{"citing_id": "2304.02976v1", "cited_id": "1711.08135", "section_title": "Definition 2.", "citation": "While in this paper our focus is on the contractivity and dissipativity properties of the continuous-time model (1), we remark that integration methods that preserve contractivity have been proposed in #REFR .", "text_before_citation": ["To summarize, our goal is to learn the parameters \u03b8 of a dynamical system (1) -i.e., a Neural ODE -that optimize a given cost as per (2), with the hard constraint that either (3) or (6) (or both) hold.", "This new requirement (i.e., the hard constraint) must be satisfied for all the parameters \u03b8 we optimize over.", "In other words, we consider a form of failsafe learning, in the sense that the property of the model to be contracting or dissipative must be guaranteed during and after parameter optimization. Remark 1.", "For solving problem (2), one has to select an integration scheme for (1), which allows performing forward propagation and gradient computations.", "The properties of the continuous-time model (1) may not coincide with those of the discretized one, in general."], "text_after_citation": ["Tailoring these integration methods to IQC properties, or developing new ones, is an interesting direction for future work.", "Incremental Property Q R S s \u2206 (\u2206u, \u2206y) L 2 -gain bound (\u03b3 \u2265 0) \u2212 1 \u03b3 I \u03b3I 0 \u03b3|\u2206u| 2 \u2212 1 \u03b3 |\u2206y| 2 Passivity 0 0 1 2 I \u2206u \u2206y Input Passivity (\u03bd \u2265 0) 0 \u22122\u03bdI I \u2206u \u2206y \u2212 \u03bd|\u2206u| 2 Output Passivity (\u03b5 \u2265 0) \u22122\u03b5I 0 I \u2206u \u2206y \u2212 \u03b5|\u2206y| 2", "III.", "CONTRACTIVITY AND DISSIPATIVITY OF NODERENS In this section, we present and analyze novel DNN structures that arise from appropriately combining Neural ODEs #OTHEREFR with RENs #OTHEREFR .", "The starting idea is to define the functions f \u03b8 and g \u03b8 in (1) through the model utilized for the layer equation in #OTHEREFR ."], "citing_paper_content": {"title": "Unconstrained Parametrization Of Dissipative And Contracting Neural Ordinary Differential Equations", "abstract": "In this work, we introduce and study a class of Deep Neural Networks (DNNs) in continuous-time. The proposed architecture stems from the combination of Neural Ordinary Differential Equations (Neural ODEs) with the model structure of recently introduced Recurrent Equilibrium Networks (RENs). We show how to endow our proposed NodeRENs with contractivity and dissipativity-crucial properties for robust learning and control. Most importantly, as for RENs, we derive parametrizations of contractive and dissipative NodeRENs which are unconstrained, hence enabling their learning for a large number of parameters. We validate the properties of NodeRENs, including the possibility of handling irregularly sampled data, in a case study in nonlinear system identification."}, "cited_paper_content": {"title": "Contracting Nonlinear Observers: Convex Optimization And Learning From Data", "abstract": "A new approach to design of nonlinear observers (state estimators) is proposed. The main idea is to (i) construct a convex set of dynamical systems which are contracting observers for a particular system, and (ii) optimize over this set for one which minimizes a bound on state-estimation error on a simulated noisy data set. We construct convex sets of continuous-time and discrete-time observers, as well as contracting sampled-data observers for continuous-time systems. Convex bounds for learning are constructed using Lagrangian relaxation. The utility of the proposed methods are verified using numerical simulation."}, "keywords": ["continuous-time model"], "citation_intent": "method"}
{"citing_id": "2303.03225v1", "cited_id": "2002.06078", "section_title": "Theorem 1.", "citation": "It was shown in #REFR that every cograph can be odd coloured using at most three colours, and that graphs of treewidth at most k can be odd coloured using at most k + 1 colours.", "text_before_citation": ["One way of Conjecture 3 is to consider that subdivided cliques appear to be essentially the graphs that require most colours to be odd coloured.", "More specifically, consider the following family of graphs B = {G obtained from a graph G by adding, for every pair of vertices u, v \u2208 V (G), a vertex w uv and edges uw uv and vw uv , and G has even order}.", "Note that subdivided cliques of even order are exactly those graphs in B where graph G is edgeless, and that the graphs in B have \u03c7 odd (G ) = |V (G)| \u2208 \u0398( |V (G )|).", "A question closely related to Conjecture 3 is whether if a class of graphs G does not contain arbitrarily large graphs of B as induced subgraphs, then G has odd chromatic number O( \u221a n), i.e., they satisfy Conjecture 3.", "This question was already answered positively for some graph classes. In fact, the bounds provided were constant."], "text_after_citation": ["In fact, those results can easily be extended to all graphs admitting a join, and H-minor free graphs, respectively. Using a similar argument, Aashtab et al.", "#OTHEREFR showed that planar graphs are 4-odd colourable, and this is tight due to subdivided K 4 being planar and 4-odd colourable, as explained above.", "They also proved that subcubic graphs are 4-odd colourable, which is again tight due to subdivided K 4 , and conjecture that this result can be generalized to all graphs, i.e., \u03c7 odd (G) \u2264 \u2206 + 1, where \u2206 denotes the maximum degree of G.", "Observe that none of those graph classes contain arbitrarily large graphs from B as induced subgraphs.", "On the negative side, bipartite graphs and split graphs contain arbitrarily large graphs from B, and therefore the bound of Conjecture 3 is best possible."], "citing_paper_content": {"title": "Odd Chromatic Number Of Graph Classes", "abstract": "A graph is called odd (respectively, even) if every vertex has odd (respectively, even) degree. Gallai proved that every graph can be partitioned into two even induced subgraphs, or into an odd and an even induced subgraph. We refer to a partition into odd subgraphs as an odd colouring of G. Scott [Graphs and Combinatorics, 2001] proved that a graph admits an odd colouring if and only if it has an even number of vertices. We say that a graph G is k-odd colourable if it can be partitioned into at most k odd induced subgraphs. We initiate the systematic study of odd colouring and odd chromatic number of graph classes. In particular, we consider for a number of classes whether they have bounded odd chromatic number. Our main results are that interval graphs, graphs of bounded modular-width and graphs of bounded maximum degree all have bounded odd chromatic number. 2012 ACM Subject Classification Mathematics of computing \u2192 Graph coloring Keywords and phrases graph classes, vertex partition problem, odd colouring, colouring variant, upper bounds"}, "cited_paper_content": {"title": "On The Complexity Of Finding Large Odd Induced Subgraphs And Odd Colorings", "abstract": "We study the complexity of the problems of finding, given a graph $G$, a largest induced subgraph of $G$ with all degrees odd (called an odd subgraph), and the smallest number of odd subgraphs that partition $V(G)$. We call these parameters ${\\sf mos}(G)$ and $\\chi_{\\sf odd}(G)$, respectively. We prove that deciding whether $\\chi_{\\sf odd}(G) \\leq q$ is polynomial-time solvable if $q \\leq 2$, and NP-complete otherwise. We provide algorithms in time $2^{O({\\sf rw})} \\cdot n^{O(1)}$ and $2^{O(q \\cdot {\\sf rw})} \\cdot n^{O(1)}$ to compute ${\\sf mos}(G)$ and to decide whether $\\chi_{\\sf odd}(G) \\leq q$ on $n$-vertex graphs of rank-width at most ${\\sf rw}$, respectively, and we prove that the dependency on rank-width is asymptotically optimal under the ETH. Finally, we give some tight bounds for these parameters on restricted graph classes or in relation to other parameters."}, "keywords": ["three colours", "every cograph"], "citation_intent": "background"}
{"citing_id": "2303.12148v1", "cited_id": "1412.6980", "section_title": "Experiments", "citation": "We train NPP with ADAM #REFR and a batch size of 2, for a maximum of 60 epochs.", "text_before_citation": ["Training details.", "We created a large-scale dataset of 3D T1-weighted (T1w) brain MRI volumes by aggregating 7 datasets: GSP #OTHEREFR , ADNI #OTHEREFR , OASIS #OTHEREFR , ADHD #OTHEREFR , ABIDE #OTHEREFR , MCIC #OTHEREFR , and COBRE #OTHEREFR . The whole training set contains 10,083 scans.", "As ground-truth target images, we used FreeSurfer generated skull-stripped, intensity-normalized and affine-registered (to so-called MNI atlas coordinates) images."], "text_after_citation": ["The initial learning rate is 1e-4 and then decreases by half after 30 epochs.", "We use random gamma transformation as a data augmentation technique, with parameter log gamma (-0.3,0.3).", "We randomly sampled \u03bb from a log-uniform distribution on (\u22123, 1) for each mini-batch.", "Architecture Details.", "f \u03b8 is a U-Net-style architecture containing an encoder and decoder with skip connections in between."], "citing_paper_content": {"title": "Neural Pre-Processing: A Learning Framework For End-To-End Brain Mri Pre-Processing", "abstract": "Head MRI pre-processing involves converting raw images to an intensity-normalized, skull-stripped brain in a standard coordinate space. In this paper, we propose an end-to-end weakly supervised learning approach, called Neural Pre-processing (NPP), for solving all three sub-tasks simultaneously via a neural network, trained on a large dataset without individual sub-task supervision. Because the overall objective is highly under-constrained, we explicitly disentangle geometricpreserving intensity mapping (skull-stripping and intensity normalization) and spatial transformation (spatial normalization). Quantitative results show that our model outperforms state-of-the-art methods which tackle only a single sub-task. Our ablation experiments demonstrate the importance of the architecture design we chose for NPP. Furthermore, NPP affords the user the flexibility to control each of these tasks at inference time. The code and model are freely-available at https: //github.com/Novestars/Neural-Pre-processing."}, "cited_paper_content": {"title": "Adam: A Method For Stochastic Optimization", "abstract": "We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm."}, "keywords": ["ADAM"], "citation_intent": "method"}
{"citing_id": "2303.01123v1", "cited_id": "1810.01619", "section_title": "A. Lidar Bias Estimation", "citation": "The error achieves 16cm for incidence angles around 80\u00b0which confirms the results described in #REFR for another lidar sensor.", "text_before_citation": ["The goal is estimation of the difference between the measured board pose with respect to its ground truth for different orientations.", "The truth board pose is provided by the theodolite sensor that tracks board's top-left corner.", "Knowing the board's dimensions and the fact that its base is fixed, the full pose of the board is given.", "As the metric, we calculate the point-to-point distance between the board's measured point cloud and the point cloud sampled from its mesh (1000 points are sampled uniformly from the board's cuboid).", "The metric's values for different board angles are given in the Figure 5 ."], "text_after_citation": ["As dark surfaces have weaker reflective qualities, it was observed that the error is larger when the black side of the board was measured.", "The metric values for higher angles are not provided as there are almost no points measured by the lidar of the board oriented almost perpendicular.", "Note, that we did not observe dependence of the ICP point-to-point distance on incidence angles when the white-side target was placed reasonably close to the lidar (for 5.3m distance)."], "citing_paper_content": {"title": "Self-Supervised Depth Correction Of Lidar Measurements From Map Consistency Loss", "abstract": "Depth perception is considered an invaluable source of information in the context of 3D mapping and various robotics applications. However, point cloud maps acquired using consumer-level light detection and ranging sensors (lidars) still suffer from bias related to local surface properties such as measuring beam-to-surface incidence angle, distance, texture, reflectance, or illumination conditions. This fact has recently motivated researchers to exploit traditional filters, as well as the deep learning paradigm, in order to suppress the aforementioned depth sensors error while preserving geometric and map consistency details. Despite the effort, depth correction of lidar measurements is still an open challenge mainly due to the lack of clean 3D data that could be used as ground truth. In this paper, we introduce two novel point cloud map consistency losses, which facilitate self-supervised learning on real data of lidar depth correction models. Specifically, the models exploit multiple point cloud measurements of the same scene from different viewpoints in order to learn to reduce the bias based on the constructed map consistency signal. Complementary to the removal of the bias from the measurements, we demonstrate that the depth correction models help to reduce localization drift. Additionally, we release a data set that contains point cloud data captured in an indoor corridor environment with precise localization and ground truth mapping information 1 ."}, "cited_paper_content": {"title": "Lidar Measurement Bias Estimation Via Return Waveform Modelling In A Context Of 3D Mapping", "abstract": "In a context of 3D mapping, it is very important to obtain accurate measurements from sensors. In particular, Light Detection And Ranging (LIDAR) measurements are typically treated as a zero-mean Gaussian distribution. We show that this assumption leads to predictable localisation drifts, especially when a bias related to measuring obstacles with high incidence angles is not taken into consideration. Moreover, we present a way to physically understand and model this bias, which generalizes to multiple sensors. Using an experimental setup, we measured the bias of the Sick LMS151, Velodyne HDL-32E, and Robosense RS-LiDAR-16 as a function of depth and incidence angle, and showed that the bias can reach 20 cm for high incidence angles. We then used our model to remove the bias from the measurements, leading to more accurate maps and a reduced localisation drift."}, "keywords": ["another lidar sensor"], "citation_intent": "result"}
{"citing_id": "2303.06873v1", "cited_id": "1610.02501", "section_title": "Analysis And Discussion.", "citation": "Is IBMIL just post-processing? Since our proposed IB-MIL shares some commonalities with the embedding-based MIL #REFR , one may ask: Do the improvements only come from exploring the bag relations? To answer this question, we make minor modifications on ABMIL.", "text_before_citation": ["We will explore the way of incorporating bag labels in future work.", "On the other hand, the unsupervised fashion makes our scheme compatible with large-scale unlabelled datasets.", "We explore more unlabelled bags via combining the bags of TCGA and Camelyon16, and constituting the confounder set via the non-parametric aggregators. From Fig.", "5b , we observe a clear improvement on AUC under both max-and mean-poolings.", "That indicates, with more bags, our implementation can achieve better approximation of confounders."], "text_after_citation": ["Instead of interventional training with confounders, we obtain the confounder dictionary via the class-specific K-means and treat it as a KNN classifier for evaluation.", "As can be seen, it brings limited improvements or even degrades the performance, verifying the improvements comes from interventional training, which is not just post-processing."], "citing_paper_content": {"title": "Interventional Bag Multi-Instance Learning On Whole-Slide Pathological Images", "abstract": "Multi-instance learning (MIL) is an effective paradigm for whole-slide pathological images (WSIs) classification to handle the gigapixel resolution and slide-level label. Prevailing MIL methods primarily focus on improving the feature extractor and aggregator. However, one deficiency of these methods is that the bag contextual prior may trick the model into capturing spurious correlations between bags and labels. This deficiency is a confounder that limits the performance of existing MIL methods. In this paper, we propose a novel scheme, Interventional Bag Multi-Instance Learning (IBMIL), to achieve deconfounded bag-level prediction. Unlike traditional likelihood-based strategies, the proposed scheme is based on the backdoor adjustment to achieve the interventional training, thus is capable of suppressing the bias caused by the bag contextual prior. Note that the principle of IBMIL is orthogonal to existing bag MIL methods. Therefore, IBMIL is able to bring consistent performance boosting to existing schemes, achieving new state-of-the-art performance. Code is available at https://github.com/HHHedo/IBMIL."}, "cited_paper_content": {"title": "Revisiting Multiple Instance Neural Networks", "abstract": "Recently neural networks and multiple instance learning are both attractive topics in Artificial Intelligence related research fields. Deep neural networks have achieved great success in supervised learning problems, and multiple instance learning as a typical weakly-supervised learning method is effective for many applications in computer vision, biometrics, nature language processing, etc. In this paper, we revisit the problem of solving multiple instance learning problems using neural networks. Neural networks are appealing for solving multiple instance learning problem. The multiple instance neural networks perform multiple instance learning in an end-to-end way, which take a bag with various number of instances as input and directly output bag label. All of the parameters in a multiple instance network are able to be optimized via back-propagation. We propose a new multiple instance neural network to learn bag representations, which is different from the existing multiple instance neural networks that focus on estimating instance label. In addition, recent tricks developed in deep learning have been studied in multiple instance networks, we find deep supervision is effective for boosting bag classification accuracy. In the experiments, the proposed multiple instance networks achieve state-of-the-art or competitive performance on several MIL benchmarks. Moreover, it is extremely fast for both testing and training, e.g., it takes only 0.0003 second to predict a bag and a few seconds to train on a MIL datasets on a moderate CPU."}, "keywords": ["bag relations"], "citation_intent": "method"}
{"citing_id": "2304.02767v1", "cited_id": "1512.03385", "section_title": "Ablation Studies", "citation": "We also experimented with different types of feature extractors for SFG module, and observed that a Resnet18 or Resnet50 #REFR is more effective than a 2 linear layer feature extractor as shown in Table 2 .", "text_before_citation": ["Using CH 4 candidates feature from SFG improves the bounding box detection performance by 0.14 mAP and mask prediction by 0.09 mIOU.", "This demonstrate that guiding queries with CH 4 candidates feature generated by SFG produces better embeddings as compared to random queries.", "Along with this, we explored the provision of CH 4 candidates feature at 2 places, (i) at input level concatenating it with f comb ; and (ii) as input to query refiner.", "We see an improvement of 0.09 mAP and 0.08 mIOU when SFG module output is passed to query refiner.", "We hypothesize that this is because on concatenating with input features, the CH 4 candidates feature information gets lost, while as cross-attention with queries reduces the search space for decoder and generate better embeddings."], "text_after_citation": ["Spectral Linear Filter: We experimented with SLF for computing covariance (Cov) using different subset of columns in the input hyperspectral image.", "We observed that the SLF is most effective when covariance is computed class-wise based on land cover.", "Class-wise Cov ensures that the radiance absorption by ground terrain is same for all the pixels while computing CH 4 enhancement. As can be seen in Fig.", "4 , SLF amplifies CH 4 candidate detection and reduces false positives.", "SLF leads to a 0.03 mAP improved in detection compared to traditional filters. The prediction from MM is shown row-1 of Fig. 5 ."], "citing_paper_content": {"title": "Methanemapper: Spectral Absorption Aware Hyperspectral Transformer For Methane Detection", "abstract": "Methane (CH 4) is the chief contributor to global climate change. Recent Airborne Visible-Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) has been very useful in quantitative mapping of methane emissions. Existing methods for analyzing this data are sensitive to local terrain conditions, often require manual inspection from domain experts, prone to significant error and hence are not scalable. To address these challenges, we propose a novel end-to-end spectral absorption wavelength aware transformer network, MethaneMapper, to detect and quantify the emissions. MethaneMapper introduces two novel modules that help to locate the most relevant methane plume regions in the spectral domain and uses them to localize these accurately. Thorough evaluation shows that MethaneMapper achieves 0.63 mAP in detection and reduces the model size (by 5\u00d7) compared to the current state of the art. In addition, we also introduce a large-scale dataset of methane plume segmentation mask for over 1200 AVIRIS-NG flight lines from 2015-2022. It contains over 4000 methane plume sites. Our dataset will provide researchers the opportunity to develop and advance new methods for tackling this challenging greenhouse gas detection problem with significant broader social impact. Dataset and source code link 1 ."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["feature extractors", "Resnet18"], "citation_intent": "method"}
{"citing_id": "2303.17679v1", "cited_id": "1012.0006", "section_title": "Parallel Active Block Scheduling", "citation": "Thus, active block pairs of different rounds are stored interleaved in and the end of a round does not induce a synchronization point as in the original algorithm #REFR .", "text_before_citation": ["Threads that are not involved in scheduling can join parallel flow computations. In a parameter study [62, p.", "108] , we found that = 1 offers a good trade-off between conflicting searches and scalability.", "Initially, we push all pairs of adjacent blocks into a concurrent FIFO queue .", "The threads then poll from and if a search finds an improvement on a block pair ( , ), we mark both as active using a seperate bitset for each round.", "If either or becomes active, we push all adjacent blocks into if they are not contained yet."], "text_after_citation": ["A round ends when all of its block pairs have been processed and all prior rounds have ended.", "If the relative improvement at the end of a round is less than 0.1%, we immediately terminate the algorithm.", "t s t s \u221e \u221e \u221e \u221e \u221e \u221e \u03c9(e) ein eout V 1 V 2 B 1 B 2", "Apply Moves.", "Since concurrently scheduled flow computations can operate on overlapping regions, there are three conflict types that can occur when applying a sequence of node moves to the global partition \u03a0: balance constraint violations, \u0394 exp \u2260 \u0394 (i.e., the expected does not match the actual connectivity reduction), and nodes in may already be moved by other searches."], "citing_paper_content": {"title": "Scalable High-Quality Hypergraph Partitioning", "abstract": "Balanced hypergraph partitioning is an NP-hard problem with many applications, e.g., optimizing communication in distributed data placement problems. The goal is to place all nodes across k different blocks of bounded size, such that hyperedges span as few parts as possible. This problem is well-studied in sequential and distributed settings, but not in shared-memory. We close this gap by devising efficient and scalable shared-memory algorithms for all components employed in the best sequential solvers without compromises with regards to solution quality. This work presents the scalable and high-quality hypergraph partitioning framework Mt-KaHyPar. Its most important components are parallel improvement algorithms based on the FM algorithm and maximum flows, as well as a parallel clustering algorithm for coarsening-which are used in a multilevel scheme with log() levels. As additional components, we parallelize the-level partitioning scheme, devise a deterministic version of our algorithm, and present optimizations for plain graphs. We evaluate our solver on more than 800 graphs and hypergraphs, and compare it with 25 different algorithms from the literature. Our fastest configuration outperforms almost all existing hypergraph partitioners with regards to both solution quality and running time. Our highest-quality configuration achieves the same solution quality as the best sequential partitioner KaHyPar, while being an order of magnitude faster with ten threads. Thus, two of our configurations occupy all fronts of the Pareto curve for hypergraph partitioning. Furthermore, our solvers exhibit good speedups, e.g., 29.6x in the geometric mean on 64 cores (deterministic), 22.3x (log()-level), and 25.9x (-level)."}, "cited_paper_content": {"title": "Engineering Multilevel Graph Partitioning Algorithms", "abstract": "We present a multi-level graph partitioning algorithm using novel local improvement algorithms and global search strategies transferred from multigrid linear solvers. Local improvement algorithms are based on max-flow min-cut computations and more localized FM searches. By combining these techniques, we obtain an algorithm that is fast on the one hand and on the other hand is able to improve the best known partitioning results for many inputs. For example, in Walshaw's well known benchmark tables we achieve 317 improvements for the tables at 1%, 3% and 5% imbalance. Moreover, in 118 out of the 295 remaining cases we have been able to reproduce the best cut in this benchmark."}, "keywords": ["active block pairs", "original algorithm"], "citation_intent": "background"}
{"citing_id": "2303.12027v1", "cited_id": "1506.01497", "section_title": "Semantics-Guided Temporal Modeling", "citation": "After the tracking process of each frame, we perform region of interest (RoI) pooling #REFR on h t according to the predicted box to obtain the target region features which are further flattened to obtain the historical target patch embeddings.", "text_before_citation": ["More concretely, grounding is performed on a standalone image but tracking is executed on consecutive video which means there exist historical appearances of the Figure 3 . Architecture of the proposed semantic-guided temporal modeling module.", "target for exploitation.", "To improve the adaptability to the variations of the target using the temporal information, we propose a Semantics-Guided Temporal Modeling (SGTM) module.", "It is designed to learn the temporal clue about recent target appearances from historical target states with semantics from the natural language as guidance.", "Figure 3 shows the architecture of the SGTM module, which mainly consists of a transformer encoder and decoder."], "text_after_citation": ["Besides, we also take into account the global semantic representation h 1 l of the Natural Language (NL) outputted by the multi-source relation modeling module, which is called the NL CLS token denoted as h CLS l .", "We directly concatenate the NL CLS token with the historical target patch embeddings and feed them into the encoder.", "Herein, the encoder is used to enhance the token embeddings corresponding to the target region and suppress those corresponding to the noise region with the guidance of NL CLS.", "After that, we utilize a decoder to compute cross-attention between enhanced target patch embeddings and a temporal query which is a learnable vector.", "The temporal query can attend to the language context and historical target embeddings, thus learning a robust historical target representation for future tracking."], "citing_paper_content": {"title": "Joint Visual Grounding And Tracking With Natural Language Specification", "abstract": "Tracking by natural language specification aims to locate the referred target in a sequence based on the natural language description. Existing algorithms solve this issue in two steps, visual grounding and tracking, and accordingly deploy the separated grounding model and tracking model to implement these two steps, respectively. Such a separated framework overlooks the link between visual grounding and tracking, which is that the natural language descriptions provide global semantic cues for localizing the target for both two steps. Besides, the separated framework can hardly be trained end-to-end. To handle these issues, we propose a joint visual grounding and tracking framework, which reformulates grounding and tracking as a unified task: localizing the referred target based on the given visual-language references. Specifically, we propose a multi-source relation modeling module to effectively build the relation between the visual-language references and the test image. In addition, we design a temporal modeling module to provide a temporal clue with the guidance of the global semantic information for our model, which effectively improves the adaptability to the appearance variations of the target. Extensive experimental results on TNL2K, LaSOT, OTB99, and RefCOCOg demonstrate that our method performs favorably against state-of-the-art algorithms for both tracking and grounding. Code is available at https://github.com/lizhou-cs/JointNLT."}, "cited_paper_content": {"title": "Faster R-Cnn: Towards Real-Time Object Detection With Region Proposal Networks", "abstract": "State-of-the-art object detection networks depend on region proposal algorithms to hypothesize object locations. Advances like SPPnet [1] and Fast R-CNN [2] have reduced the running time of these detection networks, exposing region proposal computation as a bottleneck. In this work, we introduce a Region Proposal Network (RPN) that shares full-image convolutional features with the detection network, thus enabling nearly cost-free region proposals. An RPN is a fully convolutional network that simultaneously predicts object bounds and objectness scores at each position. The RPN is trained end-to-end to generate high-quality region proposals, which are used by Fast R-CNN for detection. We further merge RPN and Fast R-CNN into a single network by sharing their convolutional features\u2014using the recently popular terminology of neural networks with \u2019attention\u2019 mechanisms, the RPN component tells the unified network where to look. For the very deep VGG-16 model [3] , our detection system has a frame rate of 5 fps ( including all steps ) on a GPU, while achieving state-of-the-art object detection accuracy on PASCAL VOC 2007, 2012, and MS COCO datasets with only 300 proposals per image. In ILSVRC and COCO 2015 competitions, Faster R-CNN and RPN are the foundations of the 1st-place winning entries in several tracks. Code has been made publicly available."}, "keywords": ["target region features"], "citation_intent": "method"}
{"citing_id": "2303.05823v1", "cited_id": "1911.06016", "section_title": "Linearly Implicit Methods For Semilinear Evolution Pdes", "citation": "Before adapting the definition of the linearly implicit methods from #REFR to the PDE context at hand in 2.1.2, we recall some basic notations for classical Runge-Kutta methods applied to PDE problems and introduce their (linear) stability function in Section 2.1.1.", "text_before_citation": ["In this section, we introduce the linearly implicit methods from #OTHEREFR adapted to the PDE setting of the equation 1as well as the associated functional framework.", "2.1. Description of the linearly implicit methods.", "Linearly implicit method from #OTHEREFR applied to a semilinear PDE of the form (1) rely on a classical Runge-Kutta method, referred to as the \"underlying Runger-Kutta method\" in the following, as well as on additional variables that are explicitly updated at every time step."], "text_after_citation": ["2.1.1. The underlying Runge-Kutta method.", "We fix a number of stages s P N \u2039 and we consider throughout this paper a Runge-Kutta method with s P N \u2039 stages, and real coefficients pa ij q 1\u010fi,j\u010fs , pb i q 1\u010fi\u010fs , pc i q 1\u010fi\u010fs .", "We group these coefficients in a square matrix A of size s with coefficients pa i,j q 1\u010fi,j\u010fs , a column vector b of size s with coefficients pb i q 1\u010fi\u010fs , and a column vector c of size s with coefficients pc i q 1\u010fi\u010fs .", "We moreover denote by 1 the column vector of size s with all components equal to 1.", "We recall that the (linear) stability function of the method is defined for \u03bb P C by #OTHEREFR Rp\u03bbq \" 1`\u03bbb t pI\u00b4\u03bbAq\u00b411,"], "citing_paper_content": {"title": "High Order Linearly Implicit Methods For Semilinear Evolution Pdes", "abstract": "This paper considers the numerical integration of semilinear evolution PDEs using the high order linearly implicit methods developped in [21] in the ODE setting. These methods use a collocation Runge-Kutta method as a basis, and additional variables that are updated explicitly and make the implicit part of the collocation Runge-Kutta method only linearly implicit. In this paper, we introduce several notions of stability for the underlying Runge-Kutta methods as well as for the explicit step on the additional variables necessary to fit the context of evolution PDE. We prove a main theorem about the high order of convergence of these linearly implicit methods in this PDE setting, using the stability hypotheses introduced before. We use nonlinear Schr\u00f6dinger equations and heat equations as main examples but our results extend beyond these two classes of evolution PDEs. We illustrate our main result numerically in dimensions 1 and 2, and we compare the efficiency of the linearly implicit methods with other methods from the litterature. We also illustrate numerically the necessity of the stability conditions of our main result."}, "cited_paper_content": {"title": "High Order Linearly Implicit Methods For Evolution Equations: How To Solve An Ode By Inverting Only Linear Systems", "abstract": "This paper introduces a new class of numerical methods for the time integration of evolution equations set as Cauchy problems of ODEs or PDEs. The systematic design of these methods mixes the Runge-Kutta collocation formalism with collocation techniques, in such a way that the methods are linearly implicit and have high order. The fact that these methods are implicit allows to avoid CFL conditions when the large systems to integrate come from the space discretization of evolution PDEs. Moreover, these methods are expected to be efficient since they only require to solve one linear system of equations at each time step, and efficient techniques from the literature can be used to do so. After the introduction of the methods, we set suitable definitions of consistency and stability for these methods. This allows for a proof that arbitrarily high order linearly implicit methods exist and converge when applied to ODEs. Eventually, we perform numerical experiments on ODEs and PDEs that illustrate our theoretical results for ODEs, and compare our methods with standard methods for several evolution PDEs."}, "keywords": ["classical Runge-Kutta methods", "linearly implicit methods"], "citation_intent": "method"}
{"citing_id": "2304.03054v3", "cited_id": "1901.07555", "section_title": "Federated Recommendation Framework", "citation": "Let U and V denote the sets of benign users (clients) #REFR and items, respectively. |U| and |V | are users' and items' sizes.", "text_before_citation": ["Following #OTHEREFR , we employ the most commonly used federated recommendation framework proposed by #OTHEREFR as our base FedRec framework.", "Most existing FedRecs #OTHEREFR are based on this framework."], "text_after_citation": ["In FedRec, each user is a client who manages its' local training dataset D .", "D consists of many user-item interactions ( , , ), where is a binary rating denoting whether user has interacted with item .", "i.e., = 1 means has interacted with , while = 0 indicates no interaction between and .", "V + and V \u2212 are used to denote the sets of interacted items and non-interacted items for user .", "The FedRec aims to predict^between and each non-interacted item and recommend the top-K ones with the highest predicted scores."], "citing_paper_content": {"title": "Manipulating Federated Recommender Systems: Poisoning With Synthetic Users And Its Countermeasures", "abstract": "Federated Recommender Systems (FedRecs) are considered privacypreserving techniques to collaboratively learn a recommendation model without sharing user data. Since all participants can directly influence the systems by uploading gradients, FedRecs are vulnerable to poisoning attacks of malicious clients. However, most existing poisoning attacks on FedRecs are either based on some prior knowledge or with less effectiveness. To reveal the real vulnerability of FedRecs, in this paper, we present a new poisoning attack method to manipulate target items' ranks and exposure rates effectively in the top-recommendation without relying on any prior knowledge. Specifically, our attack manipulates target items' exposure rate by a group of synthetic malicious users who upload poisoned gradients considering target items' alternative products. We conduct extensive experiments with two widely used FedRecs (Fed-NCF and Fed-LightGCN) on two real-world recommendation datasets. The experimental results show that our attack can significantly improve the exposure rate of unpopular target items with extremely fewer malicious users and fewer global epochs than state-of-the-art attacks. In addition to disclosing the security hole, we design a novel countermeasure for poisoning attacks on FedRecs. Specifically, we propose a hierarchical gradient clipping with sparsified updating to defend against existing poisoning attacks. The empirical results demonstrate that the proposed defending mechanism improves the robustness of FedRecs. CCS CONCEPTS \u2022 Information systems \u2192 Recommender systems; \u2022 Security and privacy \u2192 Web application security."}, "cited_paper_content": {"title": "Managing Popularity Bias In Recommender Systems With Personalized Re-Ranking", "abstract": "Many recommender systems suffer from popularity bias: popular items are recommended frequently while less popular, niche products, are recommended rarely or not at all. However, recommending the ignored products in the `long tail' is critical for businesses as they are less likely to be discovered. In this paper, we introduce a personalized diversification re-ranking approach to increase the representation of less popular items in recommendations while maintaining acceptable recommendation accuracy. Our approach is a post-processing step that can be applied to the output of any recommender system. We show that our approach is capable of managing popularity bias more effectively, compared with an existing method based on regularization. We also examine both new and existing metrics to measure the coverage of long-tail items in the recommendation."}, "keywords": ["benign users", "items"], "citation_intent": "background"}
{"citing_id": "2305.03034v1", "cited_id": "1703.01780", "section_title": "Related Work", "citation": "However, our work is the first to analyze the synergy between Mean Teacher #REFR and contrastive learning.", "text_before_citation": ["To align contrastive pre-training with downstream tasks other than image classification (e.g., object detection, semantic segmentation), more fine-grained approaches have been proposed based on masks #OTHEREFR , objects #OTHEREFR , or regions #OTHEREFR .", "Our object-level contrastive learning strategy is inspired by this line of research.", "Instead of applying contrastive learning in pre-training visual backbones, we study how to improve domain adaptive object detectors using noisy pseudo-labels and object-level contrast.", "Technically, we construct contrastive pairs using the predicted classes in pseudo-labels and optimize multi-scale features, both of which are different from typical objectlevel contrastive learning.", "Recently, contrastive learning is explored in teacher-student learning for detection #OTHEREFR ."], "text_after_citation": ["Moreover, we present a simple and general framework CMT that does not rely on negative sample mining or selection."], "citing_paper_content": {"title": "Contrastive Mean Teacher For Domain Adaptive Object Detectors", "abstract": "Object detectors often suffer from the domain gap between training (source domain) and real-world applications (target domain). Mean-teacher self-training is a powerful paradigm in unsupervised domain adaptation for object detection, but it struggles with low-quality pseudo-labels. In this work, we identify the intriguing alignment and synergy between mean-teacher self-training and contrastive learning. Motivated by this, we propose Contrastive Mean Teacher (CMT)-a unified, general-purpose framework with the two paradigms naturally integrated to maximize beneficial learning signals. Instead of using pseudo-labels solely for final predictions, our strategy extracts objectlevel features using pseudo-labels and optimizes them via contrastive learning, without requiring labels in the target domain. When combined with recent mean-teacher selftraining methods, CMT leads to new state-of-the-art targetdomain performance: 51.9% mAP on Foggy Cityscapes, outperforming the previously best by 2.1% mAP. Notably, CMT can stabilize performance and provide more significant gains as pseudo-label noise increases."}, "cited_paper_content": {"title": "Weight-Averaged Consistency Targets Improve Semi-Supervised Deep Learning Results", "abstract": "The recently proposed Temporal Ensembling has achieved state-of-the-art results in several semi-supervised learning benchmarks. It maintains an exponential moving average of label predictions on each training example, and penalizes predictions that are inconsistent with this target. However, because the targets change only once per epoch, Temporal Ensembling becomes unwieldy when learning large datasets. To overcome this problem, we propose Mean Teacher, a method that averages model weights instead of label predictions. As an additional benefit, Mean Teacher improves test accuracy and enables training with fewer labels than Temporal Ensembling. Without changing the network architecture, Mean Teacher achieves an error rate of 4.35% on SVHN with 250 labels, outperforming Temporal Ensembling trained with 1000 labels. We also show that a good network architecture is crucial to performance. Combining Mean Teacher and Residual Networks, we improve the state of the art on CIFAR-10 with 4000 labels from 10.55% to 6.28%, and on ImageNet 2012 with 10% of the labels from 35.24% to 9.11%."}, "keywords": ["contrastive learning"], "citation_intent": "background"}
{"citing_id": "2303.11833v1", "cited_id": "1811.12823", "section_title": "Fragment Set Configuration", "citation": "For the supplementary experiment aimed at discovering five target-hitting molecules (Supplementary Note 1), another fragment set containing 2,547 BRICS fragments that appeared more than 100 times in the training set of the MOSES database #REFR was used.", "text_before_citation": ["Two types of fragment sets were used in this study.", "One set contained 2,207 BRICS fragments that appeared more than 150 times in the training set #OTHEREFR curated from the ChEMBL database #OTHEREFR (release version 25).", "This fragment set was used for the three major experiments: the discovery of seven targethitting molecules, protein docking materials, and HIV inhibitors."], "text_after_citation": ["The detailed reasoning for using these fragment sets is described in Supplementary Note 2."], "citing_paper_content": {"title": "Materials Discovery With Extreme Properties Via Ai-Driven Combinatorial Chemistry", "abstract": "The goal of most materials discovery is to discover materials that are superior to those currently known. Fundamentally, this is close to extrapolation, which is a weak point for most machine learning models that learn the probability distribution of data. Herein, we develop AIdriven combinatorial chemistry, which is a rule-based inverse molecular designer that does not rely on data. Since our model has the potential to generate all possible molecular structures that can be obtained from combinations of molecular fragments, unknown materials with superior properties can be discovered. We theoretically and empirically demonstrate that our model is more suitable for discovering better materials than probability distribution-learning models. In an experiment aimed at discovering molecules that hit seven target properties, our model discovered 1,315 of all target-hitting molecules and 7,629 of five target-hitting molecules out of 100,000 trials, whereas the probability distributionlearning models failed. To illustrate the performance in actual problems, 1"}, "cited_paper_content": {"title": "Molecular Sets (Moses): A Benchmarking Platform For Molecular Generation Models", "abstract": "Deep generative models such as generative adversarial networks, variational autoencoders, and autoregressive models are rapidly growing in popularity for the discovery of new molecules and materials. In this work, we introduce MOlecular SEtS (MOSES), a benchmarking platform to support research on machine learning for drug discovery. MOSES implements several popular molecular generation models and includes a set of metrics that evaluate the diversity and quality of generated molecules. MOSES is meant to standardize the research on molecular generation and facilitate the sharing and comparison of new models. Additionally, we provide a large-scale comparison of existing state of the art models and elaborate on current challenges for generative models that might prove fertile ground for new research. Our platform and source code are freely available at this https URL."}, "keywords": ["five target-hitting molecules"], "citation_intent": "method"}
{"citing_id": "2304.07428v1", "cited_id": "1605.09557", "section_title": "Definition 8 (Valid Control Refinement):", "citation": "The next theorem states that similar to the (\u03b5, \u03b4)approximate simulation relation defined in #REFR , there always exists at least one valid control refinement for our newly defined SSR. Proof.", "text_before_citation": ["upper-bounds the sub-probability coupling v, namel\u0233", "EQUATION", "for all measurable sets A \u2282X \u00d7 X, all (x k , x k ) \u2208 R, and all\u00fb k \u2208\u00db.", "Note that for a model class M(\u03b8) that is in relation with a model M the right-hand side of (12) can depend on \u03b8 and the left-hand side can depend on \u03b8 only via the dynamics of M(\u03b8) represented by the kernel t, i.e.,", "(x + ,x + )\u2208A s(dx + |x k , x k , x k+1 ,\u00fb k )t(dx + |x k , i(x k , x k ,\u00fb k ); \u03b8) \u2265 v(A|x k , x k ,\u00fb k ; \u03b8)."], "text_after_citation": ["If M \u03b4 \u03b5 M(\u03b8), then this implies that there also exists a probabilistic kernelv(\u2022|\u2022; \u03b8) satisfying the conditions for an (\u03b5, \u03b4)-approximate probabilistic simulation relation (cf. #OTHEREFR , Def. 9).", "Furthermore, we can then compute the state mapping s as the conditional kernel obtained fromv(", "\u2022|\u2022; \u03b8), that is, s(dx k+1 |x k , x k , x k+1 ,\u00fb k ) :=v(dx k+1 |x k , x k , x k+1 ,\u00fb k ; \u03b8),", "where the right term is conditioned on x k+1 such that", "v(dx k+1 \u00d7 dx k+1 |x k , x k , x k+1 ,\u00fb k ; \u03b8) =v(dx k+1 |x k , x k , x k+1 ,\u00fb k ; \u03b8)t(dx k+1 |x k , i(\u2022); \u03b8)."], "citing_paper_content": {"title": "Bayesian Approach To Temporal Logic Control Of Uncertain Systems", "abstract": "This paper addresses the problem of data-driven computation of controllers that are correct by design for safetycritical systems and can provably satisfy (complex) functional requirements. With focus on continuous-state uncertain systems, we propose a two-stage approach that decomposes the problem into a learning stage and a robust formal controller synthesis stage. The first stage utilizes available Bayesian regression results to compute robust credible sets for the true parameters of the system. For the second stage, we introduce methods for systems subject to both stochastic and parametric uncertainties. We provide for the first time simulation relations for enabling correct-by-design control refinement that are founded on coupling uncertainties of stochastic systems via subprobability measures. The presented relations are essential for constructing abstract models that are related to not only one model but to a set of parameterized models. The results are demonstrated on a linear model and the nonlinear model of the Van der Pol Oscillator."}, "cited_paper_content": {"title": "Verification Of General Markov Decision Processes By Approximate Similarity Relations And Policy Refinement", "abstract": "In this work we introduce new approximate similarity relations that are shown to be key for policy (or control) synthesis over general Markov decision processes. The models of interest are discrete-time Markov decision processes, endowed with uncountably-infinite state spaces and metric output (or observation) spaces. The new relations, underpinned by the use of metrics, allow in particular for a useful trade-off between deviations over probability distributions on states, and distances between model outputs. We show that the new probabilistic similarity relations, inspired by a notion of simulation developed for finite-state models, can be effectively employed over general Markov decision processes for verification purposes, and specifically for control refinement from abstract models."}, "keywords": ["control refinement"], "citation_intent": "background"}
{"citing_id": "2304.05346v2", "cited_id": "1702.05743", "section_title": "Lagrange Elementary For Optimization", "citation": "As a result of the branch dynamics of the complex system and the stochastic algorithm that is still undetermined, it is challenging to manage the collective dynamics #REFR .", "text_before_citation": ["To implement the evolutionary Lagrange technique, a method like Lin et al. (2003) integrated into the evolutionary min-max algorithm #OTHEREFR .", "This EA combines self-adaptation for penalty parameters to achieve global integration, allowing for the use of smaller penalty parameters without impacting the final search results.", "Besides, the Lagrange interpolating polynomial is the solitary polynomial of the lowest degree that interpolates a given collection of data in numerical analysis and data mining.", "For mathematicians, the foundations of mathematics are a dynamic complex system.", "A method for solving discrete optimization problems relied on the interpretation of calculus algorithms."], "text_after_citation": ["EA seems to be a useful tool for optimizing complex and non-convex optimization problems.", "The earlier algorithms, such as Suma (2021), employing EA-based network reconstruction methods in the proposed new algorithm #OTHEREFR , have concentrated on the broad applicability of EA-based methodologies and their enhanced efficiency in solving reconstruction situations in complicated applications.", "Also, both the objective and constraint functions are non-convex and incorporate expectations over random states to find a stationary point for a broad non-convex stochastic optimization problem. In addition, Liu et al.", "(2019) presented a constrained stochastic successive convex approximation algorithm #OTHEREFR .", "In several deliberations, it was demonstrated that using a hybrid algorithm and Lagrange technique to solve limited optimization problems."], "citing_paper_content": {"title": "Leo: Lagrange Elementary Optimization", "abstract": "Global optimization problems are frequently solved using the practical and efficient method of evolutionary sophistication. But as the original problem becomes more complex, so does its efficacy and expandability. Thus, the purpose of this research is to introduce the Lagrange Elementary Optimization (Leo) as an evolutionary method, which is self-adaptive inspired by the remarkable accuracy of vaccinations using the albumin quotient of human blood. They develop intelligent agents using their fitness function value after gene crossing. These genes direct the search agents during both exploration and exploitation. The main objective of the Leo algorithm is presented in this paper along with the inspiration and motivation for the concept. To demonstrate its precision, the proposed algorithm is validated against a variety of test functions, including 19 traditional benchmark functions and the CECC06 2019 test functions. The results of Leo for 19 classic benchmark test functions are evaluated against DA, PSO, and GA separately, and then two other recent algorithms such as FDO and LPB are also included in the evaluation. In addition, the Leo is tested by ten functions on CECC06 2019 with DA, WOA, SSA, FDO, LPB, and FOX algorithms distinctly. The cumulative outcomes demonstrate Leo's capacity to increase the starting population and move toward the global optimum. Different standard measurements are used to verify and prove the stability of Leo in both the exploration and exploitation phases. Moreover, Statistical analysis supports the findings results of the proposed research. Finally, novel applications in the real world are introduced to demonstrate the practicality of Leo."}, "cited_paper_content": {"title": "Dr$^{2}$-Net: Deep Residual Reconstruction Network For Image Compressive Sensing", "abstract": "Most traditional algorithms for compressive sensing image reconstruction suffer from the intensive computation. Recently, deep learning-based reconstruction algorithms have been reported, which dramatically reduce the time complexity than iterative reconstruction algorithms. In this paper, we propose a novel Deep Residual Reconstruction Network (DR2-Net) to reconstruct the image from its Compressively Sensed (CS) measurement. The DR2-Net is proposed based on two observations: (1) linear mapping could reconstruct a high-quality preliminary image, and (2) residual learning could further improve the reconstruction quality. Accordingly, DR2-Net consists of two components, i.e., linear mapping network and residual network, respectively. Specifically, the fully-connected layer in neural network implements the linear mapping network. We then expand the linear mapping network to DR2-Net by adding several residual learning blocks to enhance the preliminary image. Extensive experiments demonstrate that the DR2-Net outperforms traditional iterative methods and recent deep learning-based methods by large margins at measurement rates 0.01, 0.04, 0.1, and 0.25, respectively. The code of DR2-Net has been released on: https://github.com/coldrainyht/caffe_dr2 ."}, "keywords": ["stochastic algorithm"], "citation_intent": "background"}
{"citing_id": "2303.06138v1", "cited_id": "2003.08934", "section_title": "Related Work", "citation": "To address this issue, Neural Radiance Fields (NeRFs) #REFR directly optimizes a continuous radiance field representation using a multi-layer perceptron.", "text_before_citation": ["Neural appearance reconstruction.", "Traditional methods use Structure-From-Motion #OTHEREFR and bundle adjustment #OTHEREFR to reconstruct colored point clouds.", "More recently, a number of learning-based novel view synthesis methods have been proposed, but they require 3D geometry as inputs #OTHEREFR .", "Recently, neural volume rendering approaches have been used to reconstruct scenes from only images #OTHEREFR .", "However, the rendering resolution of these methods is limited by the time and computational complexity of the discretely sampled volumes."], "text_after_citation": ["This allows for synthesizing novel views of realistic images at an unprecedented level of fidelity.", "Various extensions of NeRF have also been proposed to improve efficiency #OTHEREFR Yu et al., 2021; #OTHEREFR , generalization #OTHEREFR Yu et al., 2022; #OTHEREFR , quality #OTHEREFR , compositionality #OTHEREFR etc.", "While these neural methods produce high-quality novel views of a scene, they do not support relighting. In contrast, our approach aims at relightable appearance reconstruction.", "Learning-based relighting.", "Learning-based methods for relighting without explicit geometry have been proposed #OTHEREFR , but they are not 3D-aware."], "citing_paper_content": {"title": "Learning Object-Centric Neural Scattering Functions For Free-Viewpoint Relighting And Scene Composition", "abstract": "Photorealistic object appearance modeling from 2D images is a constant topic in vision and graphics. While neural implicit methods (such as Neural Radiance Fields) have shown high-fidelity view synthesis results, they cannot relight the captured objects. More recent neural inverse rendering approaches have enabled object relighting, but they represent surface properties as simple BRDFs, and therefore cannot handle translucent objects. We propose Object-Centric Neural Scattering Functions (OSFs) for learning to reconstruct object appearance from only images. OSFs not only support free-viewpoint object relighting, but also can model both opaque and translucent objects. While accurately modeling subsurface light transport for translucent objects can be highly complex and even intractable for neural methods, OSFs learn to approximate the radiance transfer from a distant light to an outgoing direction at any spatial location. This approximation avoids explicitly modeling complex subsurface scattering, making learning a neural implicit model tractable. Experiments on real and synthetic data show that OSFs accurately reconstruct appearances for both opaque and translucent objects, allowing faithful free-viewpoint relighting as well as scene composition. In our supplementary material, we include a video for an overview."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["Neural Radiance Fields"], "citation_intent": "method"}
{"citing_id": "2303.16058v1", "cited_id": "1608.00859", "section_title": "B. More Implementation Details B.1. Model Architecture And Training Details", "citation": "Since we the Kinetics pre-trained models adopt sparse sampling #REFR , we use a frame span of 300 for action detection, which is the default frame number of Kinetics videos. More details are listed in Table 23 .", "text_before_citation": ["We adopt the Stage-1 pre-trained video encoder and add an extra classification layer for finetuning.", "Detailed hyperparameters for different datasets are shown in Table 22 .", "In our experiments, we have tried to fine-tune the Stage-2 pre-trained video encoder, but the results on Kinetics are similar.", "Action Detection.", "Following VideoMAE #OTHEREFR and ST-MAE #OTHEREFR , we add ROIAlign with MaxPooling to generate the regions of interest."], "text_after_citation": ["Video-text retrieval.", "For fine-tuning, we adopt the same architecture as in Stage 2, but we only apply VTC and VTM losses.", "For all datasets, we sparsely sample 12 frames for both training and testing. More details are listed in Table 26 .", "For a fair comparison, we follow Singularity #OTHEREFR to apply flip augmentation for SSV2 retrieval, which may harm the performance of this temporal-related dataset. Video question-answering.", "Following the previous works #OTHEREFR , we formulate this task as text generation instead of classification."], "citing_paper_content": {"title": "Unmasked Teacher: Towards Training-Efficient Video Foundation Models", "abstract": "Video Foundation Models (VFMs) have received limited exploration due to high computational costs and data scarcity. Previous VFMs rely on Image Foundation Models (IFMs), which face challenges in transferring to the video domain. Although VideoMAE has trained a robust ViT from limited data, its low-level reconstruction poses convergence difficulties and conflicts with high-level cross-modal alignment. This paper proposes a training-efficient method for temporal-sensitive VFMs that integrates the benefits of existing methods. To increase data efficiency, we mask out most of the low-semantics video tokens, but selectively align the unmasked tokens with IFM, which serves as the UnMasked Teacher (UMT). By providing semantic guidance, our method enables faster convergence and multimodal friendliness. With a progressive pre-training framework, our model can handle various tasks including scenerelated, temporal-related, and complex video-language understanding. Using only public sources for pre-training in 6 days on 32 A100 GPUs, our scratch-built ViT-L/16 achieves state-of-the-art performances on various video tasks. The code and models will be released at https://github. com/OpenGVLab/unmasked_teacher."}, "cited_paper_content": {"title": "Temporal Segment Networks: Towards Good Practices For Deep Action Recognition", "abstract": "Deep convolutional networks have achieved great success for visual recognition in still images. However, for action recognition in videos, the advantage over traditional methods is not so evident. This paper aims to discover the principles to design effective ConvNet architectures for action recognition in videos and learn these models given limited training samples. Our first contribution is temporal segment network (TSN), a novel framework for video-based action recognition. which is based on the idea of long-range temporal structure modeling. It combines a sparse temporal sampling strategy and video-level supervision to enable efficient and effective learning using the whole action video. The other contribution is our study on a series of good practices in learning ConvNets on video data with the help of temporal segment network. Our approach obtains the state-the-of-art performance on the datasets of HMDB51 (\\( 69.4\\,\\% \\)) and UCF101 (\\( 94.2\\,\\% \\)). We also visualize the learned ConvNet models, which qualitatively demonstrates the effectiveness of temporal segment network and the proposed good practices (Models and code at https://github.com/yjxiong/temporal-segment-networks)."}, "keywords": ["action detection"], "citation_intent": "method"}
{"citing_id": "2303.01716v1", "cited_id": "1202.6409", "section_title": "Preliminaries", "citation": "Theorem 1 is consistent with the result in #REFR . given by C = {000, 112, 220, 332}.", "text_before_citation": ["Z i = \u230a m 2 \u230b j=1 y x j u\u2208Ci\u222aC \u2032 i \u03c0(i) a=1 v\u2208Eija \u03c7(u i \u2022 v i ) = \u230a m 2 \u230b j=1 y x j u\u2208Ci\u222aC \u2032 i \u03c0(i) a=1 v i \u2208V i ,w L (v ia )=j; w L (v i b )<j for b<a;w L (v i b )\u2264j for b>a \u03c7(u i \u2022 v i ) = 2 \u230a m 2 \u230b j=1 y x j u\u2208Ci\u222aC \u2032 i \u03c0(i) a=1 cos 2\u03c0u ia j m b<a 1 + 2 j\u22121 t=1 cos 2\u03c0u i b t m b>a 1 + 2 j t=1 cos 2\u03c0u i b t m = 2 \u230a m 2 \u230b j=1 y x j |C i | (2j + 1) \u03c0(i) \u2212 (2j \u2212 1) \u03c0(i) 2 + u\u2208C \u2032 i \u03c0(i) a=1 cos 2\u03c0u ia j m b<a 1 + 2 j\u22121 t=1 cos 2\u03c0u i b t m b>a 1 + 2 j t=1 cos 2\u03c0u i b t m .", "The result then follows from (3.1) and (32).", "The case for m even can be proved in a similar way.", "Remark 3.1.", "Note that when we consider the pomset block metric over Z s 2 and Z s 3 , the pomset block metric will coincide with poset block metric."], "text_after_citation": [], "citing_paper_content": {"title": "Macwilliams Type Identities For Linear Block Codes On Certain Pomsets", "abstract": "Pomset block metric is a generalization of pomset metric. In this paper, we define weight enumerator of linear block codes in pomset metric over Z m and establish MacWilliams type identities for linear block codes with respect to certain pomsets. The relation between weight enumerators of two linear pomset block codes and their direct sum is also investigated."}, "cited_paper_content": {"title": "Classification Of Poset-Block Spaces Admitting Macwilliams-Type Identity", "abstract": "In this work we prove that a poset-block space admits a MacWilliams-type identity if and only if the poset is hierarchical and at any level of the poset, all the blocks have the same dimension. When the poset-block admits the MacWilliams-type identity we explicit the relation between the weight enumerators of a code and its dual."}, "keywords": ["result"], "citation_intent": "result"}
{"citing_id": "2305.01278v1", "cited_id": "1607.08822", "section_title": "Opt125M Opt350M Opt1.3B Opt2.7B", "citation": "The SPICE #REFR metric on COCO caption is used to evaluate the VPG's visual perception ability, whereas SPICE is mainly designed for checking the objects, attributes, and relationships in the generated captions. As shown in Fig.", "text_before_citation": ["We notice an interesting phenomenon that when transferring to a given LLM tgt , both the convergence rate and optimal performance are roughly inversely proportional to the size of LLM src .", "For example, as shown in Table 3 , the OPT 125M\u21922.7B and OPT 350M\u21922.7B have much higher speed-up rate than OPT 1.3B\u21922.7B on all of the datasets. Meanwhile, as demonstrated in Fig.", "7 , the optimal performance of OPT 125M\u21922.7B and OPT 350M\u21922.7B are better than OPT 1.3B\u21922.7B on 3 VQA tasks.", "We hypothesize that training VPG on larger LLM will have a worse influence on VPG's existing finegrained perception ability, due to the enlarging embedding dimensions.", "To validate our hypothesis, we fix the VPG weight and only tune linear projectors to test VPGs trained with different LLM src through cross-size transfer."], "text_after_citation": ["8 , for each row, given the LLM tar , the performance of VPG trained on smaller LLM src can outperform the larger ones for most conditions, which indicates a better visual perception ability of VPG trained on smaller LLM src .", "Therefore, adapting a VPG from a smaller LLM which is less affected, is helpful to take fewer steps to reach the TFS's best performance and achieve even better performance.", "As we illustrated before, FlanT5 large training is extremely unstable.", "Even when the learning rate is adjusted to one-tenth of its original value, the model does not converge or shows a very slow convergence rate after 4 epochs.", "However, we find that by lowering the learning rate for stage-2 training, our VPGTrans can achieve stable training on FlanT5 large ."], "citing_paper_content": {"title": "Transfer Visual Prompt Generator Across Llms", "abstract": "While developing a new vision-language LLM (VL-LLM) by pre-training on tremendous image-text pairs from scratch can be exceedingly resource-consuming, connecting an existing LLM with a comparatively lightweight visual prompt generator (VPG) becomes a feasible paradigm. However, further tuning the VPG part of the VL-LLM still suffers from indispensable computational costs, i.e., requiring thousands of GPU hours and millions of training data. One alternative solution is to transfer an existing VPG from any existing VL-LLMs for the target VL-LLM."}, "cited_paper_content": {"title": "Spice: Semantic Propositional Image Caption Evaluation", "abstract": "There is considerable interest in the task of automatically generating image captions. However, evaluation is challenging. Existing automatic evaluation metrics are primarily sensitive to n-gram overlap, which is neither necessary nor sufficient for the task of simulating human judgment. We hypothesize that semantic propositional content is an important component of human caption evaluation, and propose a new automated caption evaluation metric defined over scene graphs coined SPICE. Extensive evaluations across a range of models and datasets indicate that SPICE captures human judgments over model-generated captions better than other automatic metrics (e.g., system-level correlation of 0.88 with human judgments on the MS COCO dataset, versus 0.43 for CIDEr and 0.53 for METEOR). Furthermore, SPICE can answer questions such as which caption-generator best understands colors? and can caption-generators count?"}, "keywords": ["VPG's visual perception", "generated captions"], "citation_intent": "method"}
{"citing_id": "2303.07226v1", "cited_id": "1808.06226", "section_title": "Vl-Moe Architecture", "citation": "The text data is tokenized using a SentencePiece #REFR tokenizer with a vocabulary size of 64k.", "text_before_citation": ["We randomly initialize the model parameters using the method described in BEiT #OTHEREFR .", "The image resolution is set to 224\u00d7224, and the patch size is 16\u00d716. The maximum sequence length for text is 128.", "We use a batch size of 6, 144 and train the model from scratch for 200k steps, which is equivalent to 40 epochs of the image-text pairs.", "Each batch contains 2, 048 images, 2, 048 texts, and 2, 048 image-text pairs.", "We perform image augmentation using random resized cropping, horizontal flipping, and color jittering, following the same method as BEiT #OTHEREFR ."], "text_after_citation": ["We use the Adam optimizer #OTHEREFR with \u03b2 1 = 0.9 and \u03b2 2 = 0.999 to optimize the model.", "The peak learning rate is 2e-3, and we use linear warmup for the first 10, 000 steps and cosine learning rate decay.", "The weight decay is 0.05, and we disable dropout and use stochastic depth #OTHEREFR with a rate of 0.1.", "The three pretrain losses are equally weighted as in BEIT-3 #OTHEREFR .", "MoE Setting."], "citing_paper_content": {"title": "Scaling Vision-Language Models With Sparse Mixture Of Experts", "abstract": "The field of natural language processing (NLP) has made significant strides in recent years, particularly in the development of large-scale vision-language models (VLMs). These models aim to bridge the gap between text and visual information, enabling a more comprehensive understanding of multimedia data. However, as these models become larger and more complex, they also become more challenging to train and deploy. One approach to addressing this challenge is the use of sparsely-gated mixture-of-experts (MoE) techniques, which divide the model into smaller, specialized sub-models that can jointly solve a task. In this paper, we explore the effectiveness of MoE in scaling vision-language models, demonstrating its potential to achieve state-of-theart performance on a range of benchmarks over dense models of equivalent computational cost. Our research offers valuable insights into stabilizing the training of MoE models, understanding the impact of MoE on model interpretability, and balancing the trade-offs between compute performance when scaling VLMs. We hope our work will inspire further research into the use of MoE for scaling large-scale visionlanguage models and other multimodal machine learning applications."}, "cited_paper_content": {"title": "Sentencepiece: A Simple And Language Independent Subword Tokenizer And Detokenizer For Neural Text Processing", "abstract": "This paper describes SentencePiece, a language-independent subword tokenizer and detokenizer designed for Neural-based text processing, including Neural Machine Translation. It provides open-source C++ and Python implementations for subword units. While existing subword segmentation tools assume that the input is pre-tokenized into word sequences, SentencePiece can train subword models directly from raw sentences, which allows us to make a purely end-to-end and language independent system. We perform a validation experiment of NMT on English-Japanese machine translation, and find that it is possible to achieve comparable accuracy to direct subword training from raw sentences. We also compare the performance of subword training and segmentation with various configurations. SentencePiece is available under the Apache 2 license at https://github.com/google/sentencepiece."}, "keywords": ["vocabulary size", "SentencePiece tokenizer"], "citation_intent": "method"}
{"citing_id": "2304.07645v1", "cited_id": "1502.01852", "section_title": "Experimental Setup", "citation": "Hypernetwork weights are initialized using Kaiming initialization #REFR on fan out mode and biases are initialized to zero.", "text_before_citation": ["Model.", "We implement the hypernetwork as a neural network with fully connected layers and LeakyReLU activations #OTHEREFR on all but the last layer which has linear output.", "Hypernetwork inputs \u03b3 are sampled uniformly in the range [0,1]."], "text_after_citation": ["Unless specified otherwise, the hypernetwork is modeled with two hidden layers with 16 and 128 neurons respectively.", "We treat the hypernetwork output as a set of separate fully connected layers each one corresponding to a different weight tensor of the primary network.", "We initialize each one of these layers independently, only considering their respective input and output dimensions.", "In a similar fashion, we initialize the parameters \u03b8 0 as separate subsets to take into account the corresponding weight tensor dimensions of the primary network.", "The choice of primary network varies among the considered task, full details are included in Appendix A."], "citing_paper_content": {"title": "Non-Proportional Parametrizations For Stable Hypernetwork Learning", "abstract": "Hypernetworks are neural networks that generate the parameters of another neural network. In many scenarios current hypernetwork training strategies are unstable, and convergence is often far slower than for non-hypernetwork models. We show that this problem is linked to an issue that arises when using common choices of hypernetwork architecture and initialization. We demonstrate analytically and experimentally how this numerical issue can lead to an instability during training that slows, and sometimes even prevents, convergence. We also demonstrate that popular deep learning normalization strategies fail to address these issues. We then propose a solution to the problem based on a revised hypernetwork formulation that uses non-proportional additive parametrizations. We test the proposed reparametrization on several tasks, and demonstrate that it consistently leads to more stable training, achieving faster convergence."}, "cited_paper_content": {"title": "Delving Deep Into Rectifiers: Surpassing Human-Level Performance On Imagenet Classification", "abstract": "Rectified activation units (rectifiers) are essential for state-of-the-art neural networks. In this work, we study rectifier neural networks for image classification from two aspects. First, we propose a Parametric Rectified Linear Unit (PReLU) that generalizes the traditional rectified unit. PReLU improves model fitting with nearly zero extra computational cost and little overfitting risk. Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities. This method enables us to train extremely deep rectified models directly from scratch and to investigate deeper or wider network architectures. Based on the learnable activation and advanced initialization, we achieve 4.94% top-5 test error on the ImageNet 2012 classification dataset. This is a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66% [33]). To our knowledge, our result is the first to surpass the reported human-level performance (5.1%, [26]) on this dataset."}, "keywords": ["Hypernetwork weights"], "citation_intent": "method"}
{"citing_id": "2303.04428v1", "cited_id": "1312.6533", "section_title": "Vi. Validation", "citation": "The joint velocities thereby exhibit a bang-bang type of behavior (generating joint velocities at the limits) for as fast as possible task convergence #REFR .", "text_before_citation": ["At around 36 s the augmentation of the level 8 position task stops as the right hand reached its prescribed position, see Fig. 20 .", "The right hand then follows a target swinging from the front [3, \u22121.5, 2] m to the back [\u22123, \u22121.5, 0] m of the robot (for reference, HRP-2Kai is 1.71 m tall with a 2.11 m arm span).", "The target is always out of reach but numerically stable robot control is obtained thanks to the switch to the Newton's method.", "As can be seen from the middle and bottom graph of Fig.", "22 , the joint velocities are numerically stable and do not show signs of high frequency oscillations as for example observed in Fig. 14 for the GN algorithm."], "text_after_citation": ["During the stretch motions, the CoM is well outside the support polygon of the feet (approximately at [\u00b10.1, \u00b10.2, \u00b1\u221e] m) which stresses the importance of the left hand supporting contact (see Fig. 19 ).", "Figure 23 shows the norms of the tracking error of the right hand during the stretch motions.", "During forward stretches LexDynAH performs better than WLS (Err WLS \u2212Err LexDynAH > 0).", "Especially during the second forward stretch (130s+) LexDynAH gets over 0.05 m closer to the target than WLS due to fully stretching its arm leading to a kinematic singularity.", "LexDynAH has a higher error norm than WLS during the backward stretch (see Fig. 23 , 100 s to 117 s)."], "citing_paper_content": {"title": "The Hierarchical Newton'S Method For Numerically Stable Prioritized Dynamic Control", "abstract": "This work links optimization approaches from hierarchical least-squares programming to instantaneous prioritized whole-body robot control. Concretely, we formulate the hierarchical Newton's method which solves prioritized non-linear least-squares problems in a numerically stable fashion even in the presence of kinematic and algorithmic singularities of the approximated kinematic constraints. These results are then transferred to control problems which exhibit the additional variability of time. This is necessary in order to formulate acceleration based controllers and to incorporate the second order dynamics. However, we show that the Newton's method without complicated adaptations is not appropriate in the acceleration domain. We therefore formulate a velocity based controller which exhibits second order proportional derivative convergence characteristics. Our developments are verified in toy robot control scenarios as well as in complex robot experiments which stress the importance of prioritized control and its singularity resolution."}, "cited_paper_content": {"title": "A General, Fast, And Robust Implementation Of The Time-Optimal Path Parameterization Algorithm", "abstract": "Finding the time-optimal parameterization of a given path subject to kinodynamic constraints is an essential component in many robotic theories and applications. The objective of this paper is to provide a general, fast, and robust implementation of this component. For this, we give a complete solution to the issue of dynamic singularities, which are the main cause of failure in existing implementations. We then present an open-source implementation of the algorithm in C++/Python and demonstrate its robustness and speed in various robotics settings."}, "keywords": ["joint velocities"], "citation_intent": "background"}
{"citing_id": "2303.18193v1", "cited_id": "1903.08527", "section_title": "Sample", "citation": "We obtain the camera pose for the input image using off-the-shelf face pose estimator #REFR .", "text_before_citation": ["because of inconsistency in the 2D super-resolution module.", "In contrast, our method retains the quality in out-of-training camera distribution.", "In Fig.", "9 , we provide the monocular fitting results to a given input image.", "Given an input image, we use our trained model to fit to the input image and show reconstruction and novel view renderings."], "text_after_citation": ["Then we optimize the latent vector w for about 1200 iterations using reconstruction loss and simultaneously optimize the generator parameter and latent vector for another 800 iterations to get the better fitting. In the same Fig.", "9 , we provide a comparison with Authentic Volumetric Avatars(AVA) #OTHEREFR .", "While AVA is not a generative model, it can generate a volumetric avatar using a phone scan, we thus qualitatively compare with this method as well.", "Please note AVA is not a main comparison, as this method targets expression retargeting, while we don't.", "#OTHEREFR heavily relies on the tracked mesh to place the primitives and suffers in the case of long hair, which results in truncated results."], "citing_paper_content": {"title": "Gvp: Generative Volumetric Primitives", "abstract": "Advances in 3D-aware generative models have pushed the boundary of image synthesis with explicit camera control. To achieve high-resolution image synthesis, several attempts have been made to design efficient generators, such as hybrid architectures with both 3D and 2D components. However, such a design compromises multiview consistency, and the design of a pure 3D generator with high resolution is still an open problem. In this work, we present Generative Volumetric Primitives (GVP), the first pure 3D generative model that can sample and render 512resolution images in real-time. GVP jointly models a number of volumetric primitives and their spatial information, both of which can be efficiently generated via a 2D convolutional network. The mixture of these primitives naturally captures the sparsity and correspondence in the 3D volume. The training of such a generator with high degree of freedom is made possible through a knowledge distillation technique. Experiments on several datasets demonstrate superior efficiency and 3D consistency of GVP over the state-ofthe-art."}, "cited_paper_content": {"title": "Accurate 3D Face Reconstruction With Weakly-Supervised Learning: From Single Image To Image Set", "abstract": "Recently, deep learning based 3D face reconstruction methods have shown promising results in both quality and efficiency.However, training deep neural networks typically requires a large volume of data, whereas face images with ground-truth 3D face shapes are scarce. In this paper, we propose a novel deep 3D face reconstruction approach that 1) leverages a robust, hybrid loss function for weakly-supervised learning which takes into account both low-level and perception-level information for supervision, and 2) performs multi-image face reconstruction by exploiting complementary information from different images for shape aggregation. Our method is fast, accurate, and robust to occlusion and large pose. We provide comprehensive experiments on three datasets, systematically comparing our method with fifteen recent methods and demonstrating its state-of-the-art performance."}, "keywords": ["camera", "estimator"], "citation_intent": "method"}
{"citing_id": "2304.09371v1", "cited_id": "1906.05797", "section_title": "A. Dataset Review", "citation": "Replica Dataset (https://github.com/facebookresearch/Repl ica-Dataset) is a dataset of high-quality 3D reconstructions of indoor environments, including RGB-D images, semantic annotations, and camera poses #REFR .", "text_before_citation": ["Freiburg Forest Dataset (http://deepscene.cs.uni-freiburg.de/) contains RGB, depth, and thermal images captured in a forest environment using a custom sensor setup #OTHEREFR .", "3D60: Indoor Scene Understanding in 3D (https://vcl3d.gith ub.io/3D60/) is a dataset containing 60 indoor scenes with 360\u00b0panoramas, depth maps, and semantic annotations #OTHEREFR .", "The dataset has high-quality images and depth maps, suitable for evaluating 3D reconstruction, scene understanding, and navigation tasks, diverse set of scenes with varying object types and room layouts.", "Taskonomy Dataset (http://taskonomy.stanford.edu/) is a large scale dataset that contains 4 million images from 20,000 scenes #OTHEREFR .", "The dataset contains 3D point clouds, surface normals, and semantic annotations."], "text_after_citation": ["SCARED dataset (The stereo correspondence and reconstruction of endoscopic data) (https://endovissub2019-scared.gran d-challenge.org/) is a medical dataset captured using a da Vinci Xi surgical robot.", "The dataset consists of 7 training datasets and 2 test datasets, and each of the dataset corresponds to a single porcine subject, in which an endoscope was moved using the da Vinci Xi to observe scenes, and the camera's pose relative to the fixed robot base was recorded based on robot kinematic status #OTHEREFR .", "EndoSLAM dataset (https://github.com/CapsuleEndoscope/En doSLAM) is a medical dataset that consists of both ex-vivo and synthetically generated data.", "The ex-vivo part of the dataset includes standard as well as capsule endoscopy recordings. The dataset is divided into 35 sub-datasets.", "Specifically, 18, 5 and 12 sub-datasets exist for colon, small intestine and stomach respectively #OTHEREFR ."], "citing_paper_content": {"title": "3 Dimensional Dense Reconstruction: A Review Of Algorithms And Dataset", "abstract": "3D dense reconstruction refers to the process of obtaining the complete shape and texture features of 3D objects from 2D planar images. 3D reconstruction is an important and extensively studied problem, but it is far from being solved. This work systematically introduces classical methods of 3D dense reconstruction based on geometric and optical models, as well as methods based on deep learning. It also introduces datasets for deep learning and the performance and advantages and disadvantages demonstrated by deep learning methods on these datasets."}, "cited_paper_content": {"title": "The Replica Dataset: A Digital Replica Of Indoor Spaces", "abstract": "We introduce Replica, a dataset of 18 highly photo-realistic 3D indoor scene reconstructions at room and building scale. Each scene consists of a dense mesh, high-resolution high-dynamic-range (HDR) textures, per-primitive semantic class and instance information, and planar mirror and glass reflectors. The goal of Replica is to enable machine learning (ML) research that relies on visually, geometrically, and semantically realistic generative models of the world - for instance, egocentric computer vision, semantic segmentation in 2D and 3D, geometric inference, and the development of embodied agents (virtual robots) performing navigation, instruction following, and question answering. Due to the high level of realism of the renderings from Replica, there is hope that ML systems trained on Replica may transfer directly to real world image and video data. Together with the data, we are releasing a minimal C++ SDK as a starting point for working with the Replica dataset. In addition, Replica is `Habitat-compatible', i.e. can be natively used with AI Habitat for training and testing embodied agents."}, "keywords": ["high-quality 3D reconstructions", "Replica Dataset"], "citation_intent": "background"}
{"citing_id": "2303.12891v1", "cited_id": "1904.01352", "section_title": "I. Introduction", "citation": "CFS-BA has performed well for CIC-IDS2017, but does not have CSE-CIC-IDS2018 results and lacks results for DNN and RF models #REFR .", "text_before_citation": ["This methodology performs much faster than any embedded method, and can give more intuitive reasoning into the decision making process, leading us to choose it.", "The only issue is that we cannot explore every possible subset, and as such need to use an optimizer method to find the closest subset to the global maxima score.", "Bat Algorithm (BA) was proposed by Yang in 2010 as a metaheuristic optimization algorithm to explore a space mimicking the behavior of microbats using echolocation #OTHEREFR .", "Most metaheuristic algorithms have two phases; exploration attempts to discover the widest range of the feature space, and exploitation aims to discover the best local solution within a given region.", "Unlike most modern metaheuristic algorithms, BA gradually switches between exploration and exploitation using a self-tuning hyperparameter #OTHEREFR ."], "text_after_citation": ["The Aquila Optimizer was introduced as a faster and potentially more efficient metaheuristic algorithm than prior methodology.", "While slower to converge onto a maxima, it has been shown to outperform BA when used in feature selection on cybersecurity benchmark datasets #OTHEREFR .", "Because of its promising nature, the algorithm will be used and compared to BA with CFS scoring.", "AO has performed well for CIC-IDS2017, but used a different fitness than CFS and does not have CSE-CIC-IDS2018 results #OTHEREFR .", "The paper is organized as follows: our motivation for the particular fields, tools, and algorithms used are in section I, particular mathematical and technical details regarding the implementation of our research are in section II, information specific to our dataset and its distribution and processing are in section III, experimental results are in section IV, interpretation of results and limitations of the study are in section V, and a summary of the work's main achievements as well as future work are in section VI."], "citing_paper_content": {"title": "Feature Reduction Method Comparison Towards Explainability And Efficiency In Cybersecurity Intrusion Detection Systems", "abstract": "In the realm of cybersecurity, intrusion detection systems (IDS) detect and prevent attacks based on collected computer and network data. In recent research, IDS models have been constructed using machine learning (ML) and deep learning (DL) methods such as Random Forest (RF) and deep neural networks (DNN). Feature selection (FS) can be used to construct faster, more interpretable, and more accurate models. We look at three different FS techniques; RF information gain (RF-IG), correlation feature selection using the Bat Algorithm (CFS-BA), and CFS using the Aquila Optimizer (CFS-AO). Our results show CFS-BA to be the most efficient of the FS methods, building in 55% of the time of the best RF-IG model while achieving 99.99% of its accuracy. This reinforces prior contributions attesting to CFS-BA's accuracy while building upon the relationship between subset size, CFS score, and RF-IG score in final results."}, "cited_paper_content": {"title": "Building An Efficient Intrusion Detection System Based On Feature Selection And Ensemble Classifier", "abstract": "Since Internet is so popular and prevailing in human life, countering cyber threats, especially attack detection, is a challenging area of research in the field of cyber security. Intrusion detection systems (IDSs) are essential entities in a network topology aiming to safeguard the integrity and availability of sensitive assets in the protected systems. Although many supervised and unsupervised learning approaches from the field of machine learning and pattern recognition have been used to increase the efficacy of IDSs, it is still a problem to deal with lots of redundant and irrelevant features in high-dimension datasets for network anomaly detection. To this end, we propose a novel methodology combining the benefits of correlation-based feature selection(CFS) and bat algorithm(BA) with an ensemble classifier based on C4.5, Random Forest(RF), and Forest by Penalizing Attributes(Forest PA), which can be able to classify both common and rare types of attacks with high accuracy and efficiency. The experimental results, using a novel intrusion detection dataset, namely CIC-IDS2017, reveal that our CFS-BA-Ensemble method is able to contribute more critical features and significantly outperforms individual approaches, achieving high accuracy and low false alarm rate. Moreover, compared with the majority of the existing state-of-the-art and legacy techniques, our approach exhibits better performance under several classification metrics in the context of classification accuracy, f-measure, attack detection rate, and false alarm rate."}, "keywords": ["CIC-IDS2017"], "citation_intent": "result"}
{"citing_id": "2304.11084v1", "cited_id": "1712.09381", "section_title": ". E X P E R I M E N T S", "citation": "The implementation used the Python library Ray RLLib #REFR for implementations of the reinforcement learning algorithms.", "text_before_citation": ["Three experiments were performed to evaluate the simulator and defender agent: A comparison between the RL policy and heuristic policies at different levels of IDS accuracy, a comparison between RL policies trained with different attackers and an analysis of how the policy training is affected as the graph size increases.", "RL policy training and evaluation was performed on a Google Cloud virtual machine equipped with an Nvidia V100 GPU, 12 CPU cores, and 30 GB of RAM."], "text_after_citation": ["All policies trained with RL were run for 500 PPO policy iterations. All policies were evaluated by running 500 episodes.", "Each policy took roughly 20 minutes to train and evaluate.", "Experiments were performed three times with different seeds for the random number generator.", "The neural networks used two hidden layers with 128 nodes.", "All TTC values except those explicitly set to 0 were initialized at the start of each episode by sampling from the exponential distribution f (x; \u03b2) = 1 \u03b2 e \u2212 1 \u03b2 x , where \u03b2 is the mean TTC value assigned to each attack step."], "citing_paper_content": {"title": "Learning Automated Defense Strategies Using Graph-Based Cyber Attack Simulations", "abstract": "We implemented and evaluated an automated cyber defense agent. The agent takes security alerts as input and uses reinforcement learning to learn a policy for executing predefined defensive measures. The defender policies were trained in an environment intended to simulate a cyber attack. In the simulation, an attacking agent attempts to capture targets in the environment, while the defender attempts to protect them by enabling defenses. The environment was modeled using attack graphs based on the Meta Attack Language language. We assumed that defensive measures have downtime costs, meaning that the defender agent was penalized for using them. We also assumed that the environment was equipped with an imperfect intrusion detection system that occasionally produces erroneous alerts based on the environment state. To evaluate the setup, we trained the defensive agent with different volumes of intrusion detection system noise. We also trained agents with different attacker strategies and graph sizes. In experiments, the defensive agent using policies trained with reinforcement learning outperformed agents using heuristic policies. Experiments also demonstrated that the policies could generalize across different attacker strategies. However, the performance of the learned policies decreased as the attack graphs increased in size."}, "cited_paper_content": {"title": "Rllib: Abstractions For Distributed Reinforcement Learning", "abstract": "Reinforcement learning (RL) algorithms involve the deep nesting of highly irregular computation patterns, each of which typically exhibits opportunities for distributed computation. We argue for distributing RL components in a composable way by adapting algorithms for top-down hierarchical control, thereby encapsulating parallelism and resource requirements within short-running compute tasks. We demonstrate the benefits of this principle through RLlib: a library that provides scalable software primitives for RL. These primitives enable a broad range of algorithms to be implemented with high performance, scalability, and substantial code reuse. RLlib is available at this https URL."}, "keywords": ["reinforcement learning algorithms"], "citation_intent": "method"}
{"citing_id": "2305.01151v1", "cited_id": "1907.07804", "section_title": "Transformer", "citation": "We refer the reader to #REFR for further details and discussion of gated multi-head attention blocks.", "text_before_citation": ["This first attention block's output form the 'queries', and along with the 'keys' and 'values' from the spatial cache, make up the inputs to the gated multi-head attention block #OTHEREFR .", "In this way, temporal features can attend to spatial features and learn complementary cross-modality information.", "Furthermore, with gated multi-head attention, temporal attention scores respectively scale the corresponding attention scores of spatial cache elements.", "For instance, if an image receives high attention in the first, temporal attention block, its corresponding subpixels will receive higher attention in this gated, spatial attention block.", "This is a mechanism to ensure high temporal attention translates to high spatial attention."], "text_after_citation": ["Finally, the output of this gated, multi-head attention block (after another residual addition and layer normalization) form the inputs for two separate, feed forward heads: one for the policy and one for the classifier."], "citing_paper_content": {"title": "Early Classifying Multimodal Sequences", "abstract": "Often pieces of information are received sequentially over time. When did one collect enough such pieces to classify? Trading wait time for decision certainty leads to early classification problems that have recently gained attention as a means of adapting classification to more dynamic environments. However, so far results have been limited to unimodal sequences. In this pilot study, we expand into early classifying multimodal sequences by combining existing methods. We show our new method yields experimental AUC advantages of up to 8.7%."}, "cited_paper_content": {"title": "Omninet: A Unified Architecture For Multi-Modal Multi-Task Learning", "abstract": "Transformer is a popularly used neural network architecture, especially for language understanding. We introduce an extended and unified architecture which can be used for tasks involving a variety of modalities like image, text, videos, etc. We propose a spatio-temporal cache mechanism that enables learning spatial dimension of the input in addition to the hidden states corresponding to the temporal input sequence. The proposed architecture further enables a single model to support tasks with multiple input modalities as well as asynchronous multi-task learning, thus we refer to it as OmniNet. For example, a single instance of OmniNet can concurrently learn to perform the tasks of part-of-speech tagging, image captioning, visual question answering and video activity recognition. We demonstrate that training these four tasks together results in about three times compressed model while retaining the performance in comparison to training them individually. We also show that using this neural network pre-trained on some modalities assists in learning an unseen task. This illustrates the generalization capacity of the self-attention mechanism on the spatio-temporal cache present in OmniNet."}, "keywords": ["gated multi-head attention"], "citation_intent": "background"}
{"citing_id": "2304.04811v1", "cited_id": "1810.04805", "section_title": "Model Training And Performance", "citation": "The classification model is a fine tuned COVID-Twitter Pre-Trained (M\u00fcller, Salath\u00e9, and Kummervold 2020) BERT Large #REFR .", "text_before_citation": ["The training process of the model is based on Jiang et al.", "2021, and we will briefly explain the key settings in this section.", "For more detailed training settings, please refer to the paper."], "text_after_citation": ["It contains 24 transformer layers, and in our experiment, only the parameters in the last transformer encoding layer are unlocked for fine-tuning, while the remaining BERT weights remain frozen.", "Similar to the BERT next sentence prediction task, the input to the model is constructed as ' We conduct five-folds 'Leave-Claim-Out Cross Validation' described in #OTHEREFR to compare enrichment performance.", "Note that the five-folds are only conducted on the original Jiang et al. (2021) annotated dataset.", "The enriched data is only added as supplemental training data, therefore, the total number of test samples is the same as in the experiment described in #OTHEREFR .", "The purpose of Leave-Claim-Out Cross Validation is to test the classification performance for unseen misinformation."], "citing_paper_content": {"title": "A Large-Scale Comparative Study Of Accurate Covid-19 Information Versus Misinformation", "abstract": "The COVID-19 pandemic led to an infodemic where an overwhelming amount of COVID-19 related content was being disseminated at high velocity through social media. This made it challenging for citizens to differentiate between accurate and inaccurate information about COVID-19. This motivated us to carry out a comparative study of the characteristics of COVID-19 misinformation versus those of accurate COVID-19 information through a large-scale computational analysis of over 242 million tweets. The study makes comparisons alongside four key aspects: 1) the distribution of topics, 2) the live status of tweets, 3) language analysis and 4) the spreading power over time. An added contribution of this study is the creation of a COVID-19 misinformation classification dataset. Finally, we demonstrate that this new dataset helps improve misinformation classification by more than 9% based on average F1 measure."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["fine tuned COVID-Twitter", "Kummervold 2020) BERT"], "citation_intent": "method"}
{"citing_id": "2303.07909v1", "cited_id": "1907.05600", "section_title": "Development Before Ddpm", "citation": "With the gradient of log probability density as score function #REFR , SGM generates the samples towards decreasing noise levels and trains the model by estimating the score functions for noisy data distribution.", "text_before_citation": ["Experimental results on multiple datasets show the effectiveness of DPM when estimating complex data distribution.", "DPM #OTHEREFR can be viewed as the foundation of DDPM #OTHEREFR , while DDPM #OTHEREFR optimizes DPM #OTHEREFR with improved implementations.", "Score-based Generative model(SGM).", "Techniques for improving score-based generative models have also been investigated in #OTHEREFR , #OTHEREFR .", "SGM #OTHEREFR proposes to perturb the data with random Gaussian noise of various magnitudes."], "text_after_citation": ["Despite different motivations, SGM shares a similar optimization objective with DDPM during training, which is also discussed in #OTHEREFR that the DDPM under a certain parameterization is equivalent to SGM during training.", "A improved variant of SGM is investigated in #OTHEREFR for generalizing to high-resolution images."], "citing_paper_content": {"title": "Text-To-Image Diffusion Model In Generative Ai: A Survey", "abstract": "This survey reviews text-to-image diffusion models in the context that diffusion models have emerged to be popular for a wide range of generative tasks. As a self-contained work, this survey starts with a brief introduction of how a basic diffusion model works for image synthesis, followed by how condition or guidance improves learning. Based on that, we present a review of state-of-the-art methods on text-conditioned image synthesis, i.e. text-to-image. We further summarize applications beyond text-to-image generation: text-guided creative generation and text-guided image editing. Beyond the progress made so far, we discuss existing challenges and promising future directions."}, "cited_paper_content": {"title": "Generative Modeling By Estimating Gradients Of The Data Distribution", "abstract": "We introduce a new generative model where samples are produced via Langevin dynamics using gradients of the data distribution estimated with score matching. Because gradients can be ill-defined and hard to estimate when the data resides on low-dimensional manifolds, we perturb the data with different levels of Gaussian noise, and jointly estimate the corresponding scores, i.e., the vector fields of gradients of the perturbed data distribution for all noise levels. For sampling, we propose an annealed Langevin dynamics where we use gradients corresponding to gradually decreasing noise levels as the sampling process gets closer to the data manifold. Our framework allows flexible model architectures, requires no sampling during training or the use of adversarial methods, and provides a learning objective that can be used for principled model comparisons. Our models produce samples comparable to GANs on MNIST, CelebA and CIFAR-10 datasets, achieving a new state-of-the-art inception score of 8.87 on CIFAR-10. Additionally, we demonstrate that our models learn effective representations via image inpainting experiments."}, "keywords": ["gradient"], "citation_intent": "method"}
{"citing_id": "2304.08645v1", "cited_id": "1911.10194", "section_title": "I. Introduction", "citation": "To address these shortcomings, in this work, we introduce Probabilistic Panoptic-DeepLab (ProPanDL), a family of networks that extend the Panoptic-DeepLab architecture #REFR to produce pixel-wise predictive uncertainty in panoptic segmentation.", "text_before_citation": ["To date, there has been little research exploring the application of uncertainty estimation to the panoptic segmentation task. Uncertainty-aware panoptic segmentation was introduced by Sirohi et al.", "#OTHEREFR , who introduce a network architecture capable of producing pixel-wise uncertainties as well two metrics, pECE and uECE, for evaluating the task. However, their approach is not without limitations.", "First, while panoptic segmentation has two aspects (semantic segmentation and instance discrimination), their proposed metrics evaluate only the total uncertainty rather than the two components independently.", "Second, their EvPSNet segmentation network is capable of only producing scalar uncertainty scores in the range [0, 1], rather than full distributions.", "This limitation is reflected in the uECE and pECE metrics, which evaluate only real-valued uncertainty rather than full distributions."], "text_after_citation": ["To our knowledge, this work is the first to predict full output distributions for either instance or panoptic segmentation.", "ProPanDL uses separate semantic and spatial branches to jointly estimate semantic segmentation, instance discrimination, and their associated pixel-wise uncertainties.", "Semantic segmentation uncertainty is represented as a calibrated categorical distribution, produced via Temperature Scaling #OTHEREFR or Evidential Deep Learning #OTHEREFR .", "For instance discrimination, we apply Variance Networks #OTHEREFR and Sam-pleNet #OTHEREFR to estimate the distribution of pixels relative to their corresponding instance center.", "In addition, we propose to decouple the semantic and spatial uncertainty evaluation in pECE."], "citing_paper_content": {"title": "Propandl: A Modular Architecture For Uncertainty-Aware Panoptic Segmentation", "abstract": "We introduce ProPanDL, a family of networks capable of uncertainty-aware panoptic segmentation. Unlike existing segmentation methods, ProPanDL is capable of estimating full probability distributions for both the semantic and spatial aspects of panoptic segmentation. We implement and evaluate ProPanDL variants capable of estimating both parametric (Variance Network) and parameter-free (SampleNet) distributions quantifying pixel-wise spatial uncertainty. We couple these approaches with two methods (Temperature Scaling and Evidential Deep Learning) for semantic uncertainty estimation. To evaluate the uncertainty-aware panoptic segmentation task, we address limitations with existing approaches by proposing new metrics that enable separate evaluation of spatial and semantic uncertainty. We additionally propose the use of the energy score, a proper scoring rule, for more robust evaluation of spatial output distributions. Using these metrics, we conduct an extensive evaluation of ProPanDL variants. Our results demonstrate that ProPanDL is capable of estimating well-calibrated and meaningful output distributions while still retaining strong performance on the base panoptic segmentation task."}, "cited_paper_content": {"title": "Panoptic-Deeplab: A Simple, Strong, And Fast Baseline For Bottom-Up Panoptic Segmentation", "abstract": "In this work, we introduce Panoptic-DeepLab, a simple, strong, and fast system for panoptic segmentation, aiming to establish a solid baseline for bottom-up methods that can achieve comparable performance of two-stage methods while yielding fast inference speed. In particular, PanopticDeepLab adopts the dual-ASPP and dual-decoder structures specific to semantic, and instance segmentation, respectively. The semantic segmentation branch is the same as the typical design of any semantic segmentation model (e.g., DeepLab), while the instance segmentation branch is class-agnostic, involving a simple instance center regression. As a result, our single Panoptic-DeepLab simultaneously ranks first at all three Cityscapes benchmarks, setting the new state-of-art of 84.2% mIoU, 39.0% AP, and 65.5% PQ on test set. Additionally, equipped with MobileNetV3, Panoptic-DeepLab runs nearly in real-time with a single 1025 x 2049 image (15.8 frames per second), while achieving a competitive performance on Cityscapes (54.1 PQ% on test set). On Mapillary Vistas test set, our ensemble of six models attains 42.7% PQ, outperforming the challenge winner in 2018 by a healthy margin of 1.5%. Finally, our Panoptic-DeepLab also performs on par with several topdown approaches on the challenging COCO dataset. For the first time, we demonstrate a bottom-up approach could deliver state-of-the-art results on panoptic segmentation."}, "keywords": ["panoptic segmentation"], "citation_intent": "method"}
{"citing_id": "2303.10875v1", "cited_id": "1903.02428", "section_title": "A. Experimental Setup", "citation": "In addition, all the experimental and profiling results are developed on PyTorch Geometric (PyG) framework #REFR , taking the average results of 10 runs. HGNAS settings.", "text_before_citation": ["Baselines and datasets.", "For evaluating HGNAS, we consider three baselines: the popular point cloud processing model DGCNN #OTHEREFR , and two manually optimized methods on the DGCNN architecture #OTHEREFR .", "Our experiments are conducted on the public benchmark Model-Net40 #OTHEREFR for classification task with 1024 points, and following the default hyperparameter settings in #OTHEREFR ."], "text_after_citation": ["We assign 12 positions for the GNN supernet to cover DGCNN architectures.", "During the design space exploration, the max iteration is set to 1000, while the population size of EA is set as 20.", "Searching and training of HGNAS are both conducted on Nvidia V100 GPU.", "During function search and operation search, the number of GNN supernet training epochs is set as 50 and 500, respectively.", "Predictor settings."], "citing_paper_content": {"title": "Hardware-Aware Graph Neural Network Automated Design For Edge Computing Platforms", "abstract": "Graph neural networks (GNNs) have emerged as a popular strategy for handling non-Euclidean data due to their state-of-the-art performance. However, most of the current GNN model designs mainly focus on task accuracy, lacking in considering hardware resources limitation and real-time requirements of edge application scenarios. Comprehensive profiling of typical GNN models indicates that their execution characteristics are significantly affected across different computing platforms, which demands hardware awareness for efficient GNN designs. In this work, HGNAS is proposed as the first Hardware-aware Graph Neural Architecture Search framework targeting resource constraint edge devices. By decoupling the GNN paradigm, HGNAS constructs a fine-grained design space and leverages an efficient multi-stage search strategy to explore optimal architectures within a few GPU hours. Moreover, HGNAS achieves hardware awareness during the GNN architecture design by leveraging a hardware performance predictor, which could balance the GNN model accuracy and efficiency corresponding to the characteristics of targeted devices. Experimental results show that HGNAS can achieve about 10.6\u00d7 speedup and 88.2% peak memory reduction with a negligible accuracy loss compared to DGCNN on various edge devices, including Nvidia RTX3080, Jetson TX2, Intel i7-8700K and Raspberry Pi 3B+."}, "cited_paper_content": {"title": "Fast Graph Representation Learning With Pytorch Geometric", "abstract": "We introduce PyTorch Geometric, a library for deep learning on irregularly structured input data such as graphs, point clouds and manifolds, built upon PyTorch. In addition to general graph data structures and processing methods, it contains a variety of recently published methods from the domains of relational learning and 3D data processing. PyTorch Geometric achieves high data throughput by leveraging sparse GPU acceleration, by providing dedicated CUDA kernels and by introducing efficient mini-batch handling for input examples of different size. In this work, we present the library in detail and perform a comprehensive comparative study of the implemented methods in homogeneous evaluation scenarios."}, "keywords": ["PyG"], "citation_intent": "method"}
{"citing_id": "2304.07931v4", "cited_id": "1503.07241", "section_title": "Improving Graphicionado", "citation": "Then, the vertex, whether or not it is active, processes its incoming properties, reduces them to a single new value, and applies this value to its property #REFR .", "text_before_citation": ["In this section, we demonstrate both the generality of TeAAL as a tool for modeling a broader class of accelerators and the value of TeAAL when proposing and evaluating new designs.", "As an example, we model Graphicionado #OTHEREFR , which accelerates graph algorithms written in the vertexcentric programming paradigm, and use TeAAL to propose two optimizations.", "Vertex-centric programming describes graph algorithms from the point of view of a single vertex.", "During each iteration, a vertex, if active, sends its property to all of its destination vertices."], "text_after_citation": ["For some algorithms (e.g., PageRank #OTHEREFR ), all vertices are active every iteration, while for other algorithms (e.g., breadth-first search (BFS)), only the vertices whose properties are updated are active in the next iteration #OTHEREFR .", "In this evaluation, we will focus on the latter, more complex, class of algorithms.", "Figure 11a shows a cascade of Einsums representing Graphicionado #OTHEREFR .", "We omit the rest of the TeAAL specification for space. Graphicionado divides its evaluation into two stages.", "During the processing phase, the active vertices A0 are used to select the edges that need to be processed SO (Line 2), the weights of those edges are combined with the source vertex properties (Line 3), and reduced into R (implicit in Line 3)."], "citing_paper_content": {"title": "Teaal: A Declarative Framework For Modeling Sparse Tensor Accelerators", "abstract": "Over the past few years, the explosion in sparse tensor algebra workloads has led to a corresponding rise in domain-specific accelerators to service them. Due to the irregularity present in sparse tensors, these accelerators employ a wide variety of novel solutions to achieve good performance. At the same time, prior work on design-flexible sparse accelerator modeling does not express this full range of design features, making it difficult to understand the impact of each design choice and compare or extend the state-of-the-art. To address this, we propose TeAAL: a language and compiler for the concise and precise specification and evaluation of sparse tensor algebra architectures. We use TeAAL to represent and evaluate four disparate state-ofthe-art accelerators-ExTensor, Gamma, OuterSPACE, and SIGMA-and verify that it reproduces their performance with high accuracy. Finally, we demonstrate the potential of TeAAL as a tool for designing new accelerators by showing how it can be used to speed up Graphicionado-by 38\u00d7 on BFS and 4.3\u00d7 on SSSP."}, "cited_paper_content": {"title": "Graphmat: High Performance Graph Analytics Made Productive", "abstract": "Given the growing importance of large-scale graph analytics, there is a need to improve the performance of graph analysis frameworks without compromising on productivity. GraphMat is our solution to bridge this gap between a user-friendly graph analytics framework and native, hand-optimized code. GraphMat functions by taking vertex programs and mapping them to high performance sparse matrix operations in the backend. We thus get the productivity benefits of a vertex programming framework without sacrificing performance. GraphMat is a single-node multicore graph framework written in C++ which has enabled us to write a diverse set of graph algorithms with the same effort compared to other vertex programming frameworks. GraphMat performs 1.1-7X faster than high performance frameworks such as GraphLab, CombBLAS and Galois. GraphMat also matches the performance of MapGraph, a GPU-based graph framework, despite running on a CPU platform with significantly lower compute and bandwidth resources. It achieves better multicore scalability (13-15X on 24 cores) than other frameworks and is 1.2X off native, hand-optimized code on a variety of graph algorithms. Since GraphMat performance depends mainly on a few scalable and well-understood sparse matrix operations, GraphMat can naturally benefit from the trend of increasing parallelism in future hardware."}, "keywords": ["vertex"], "citation_intent": "background"}
{"citing_id": "2303.00848v3", "cited_id": "1205.2629", "section_title": "Relationship Between Kl Divergence And Fisher Divergence", "citation": "See Appendix D for a derivation and a comparison with a similar result by #REFR .", "text_before_citation": ["Interestingly, if we make the assumption that the score network encodes a conservative vector field, then the time derivative of the KL divergence equals a scaled Fisher divergence:", "EQUATION"], "text_after_citation": [], "citing_paper_content": {"title": "Understanding The Diffusion Objective As A Weighted Integral Of Elbos", "abstract": "Diffusion models in the literature are optimized with various objectives that are special cases of a weighted loss, where the weighting function specifies the weight per noise level. Uniform weighting corresponds to maximizing the ELBO, a principled approximation of maximum likelihood. In current practice diffusion models are optimized with non-uniform weighting due to better results in terms of sample quality. In this work we expose a direct relationship between the weighted loss (with any weighting) and the ELBO objective. We show that the weighted loss can be written as a weighted integral of ELBOs, with one ELBO per noise level. If the weighting function is monotonic, then the weighted loss is a likelihood-based objective: it maximizes the ELBO under simple data augmentation, namely Gaussian noise perturbation. Our main contribution is a deeper theoretical understanding of the diffusion objective, but we also performed some experiments comparing monotonic with non-monotonic weightings, finding that monotonic weighting performs competitively with the best published results."}, "cited_paper_content": {"title": "Interpretation And Generalization Of Score Matching", "abstract": "Score matching is a recently developed parameter learning method that is particularly effective to complicated high dimensional density models with intractable partition functions. In this paper, we study two issues that have not been completely resolved for score matching. First, we provide a formal link between maximum likelihood and score matching. Our analysis shows that score matching finds model parameters that are more robust with noisy training data. Second, we develop a generalization of score matching. Based on this generalization, we further demonstrate an extension of score matching to models of discrete data."}, "keywords": ["derivation", "similar result"], "citation_intent": "result"}
{"citing_id": "2304.00740v1", "cited_id": "1810.04805", "section_title": "Related Work", "citation": "Editing LLMs In the past, LLMs have been predominantly adapted to new tasks and knowledge through finetuning #REFR .", "text_before_citation": ["Probing representations of individual situations Neural LMs have also been shown to build representations of context-dependent knowledge.", "show that they track aspects of entity state over a discourse, and this state can be extracted from LM representations of contextualized entity tokens.", "Furthermore, many LMs have been (indirectly) evaluated on their ability to track context-dependent knowledge by having their performance measured on downstream reading comprehension tasks in wich the LM is expected to answer questions about facts within a discourse.", "Reading comprehension datasets such as CoQA #OTHEREFR , RACE #OTHEREFR , and SQuAD #OTHEREFR are now part of the standard evaluation suite for new LMs; and most modern LMs perform well #OTHEREFR . However, generating does not always imply knowing.", "Datasets contain spurious correlations #OTHEREFR , and LMs are sensitive to the phrasing of prompts and questions #OTHEREFR ."], "text_after_citation": ["Recently, with very large LMs, new classes of adaptation methods have been introduced, which generally fall into one of the following two categories:", "(1) Prompt design approaches prepend a textual prompt to each example specifying the adaptation target #OTHEREFR .", "Though these techniques have generally been explored in the context of teaching LMs new tasks, they have also been used to imbue LMs with new knowledge.", "(2) Prefix-tuning approaches prepend continuous learned tokens ahead of each example.", "The learned tokens specify a task for the LM similar to how a textual prompt would #OTHEREFR ."], "citing_paper_content": {"title": "Measuring And Manipulating Knowledge Representations In Language Models", "abstract": "Neural language models (LMs) represent facts about the world described by text. Sometimes these facts derive from training data (in most LMs, a representation of the word banana encodes the fact that bananas are fruits). Sometimes facts derive from input text itself (a representation of the sentence I poured out the bottle encodes the fact that the bottle became empty). Tools for inspecting and modifying LM fact representations would be useful almost everywhere LMs are used: making it possible to update them when the world changes, to localize and remove sources of bias, and to identify errors in generated text. We describe REMEDI, an approach for querying and modifying factual knowledge in LMs. REMEDI learns a map from textual queries to fact encodings in an LM's internal representation system. These encodings can be used as knowledge editors: by adding them to LM hidden representations, we can modify downstream generation to be consistent with new facts. REMEDI encodings can also be used as model probes: by comparing them to LM representations, we can ascertain what properties LMs attribute to mentioned entities, and predict when they will generate outputs that conflict with background knowledge or input text. REMEDI thus links work on probing, prompting, and model editing, and offers steps toward general tools for fine-grained inspection and control of knowledge in LMs."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["knowledge", "new tasks"], "citation_intent": "background"}
{"citing_id": "2304.14837v1", "cited_id": "1911.11763", "section_title": "Visual Localization", "citation": "Our IMP obtains similar performance to Super-Glue* #REFR and outperforms SGMNet and ClusterGNN especially for night images.", "text_before_citation": ["As most filtering methods don't provide results on visual localization task, we only show results of OANet #OTHEREFR and AdaLAM #OTHEREFR .", "OANet and AdaLAM give close accuracy to MNN at error thresholds of 0.25m, 2 \u2022 and 0.5m, 5 \u2022 , but much better performance at threshold of 5m, 10 \u2022 , because Table 3 .", "Results on Aachen v1.0 (top) and v1.1 (bottom) dataset #OTHEREFR . The best and second best results are highlighted.", "test images with larger viewpoint changes are more sensitive to outliers which can be partially filtered by OANet and AdaLAM."], "text_after_citation": ["Note that our models are only trained on the MegaDepth dataset #OTHEREFR while SuperGlue* is additionally pretrained on Oxford and Pairs dataset #OTHEREFR . EIMP slightly outperforms IMP and SuperGlue*.", "That is because in the long-term large-scale localization task, query and reference images usually have larger viewpoint and illumination changes with more keypoints without true matches.", "EIMP effectively discards these keypoints, guaranteeing the quality of matches and thus improving the localization accuracy."], "citing_paper_content": {"title": "Imp: Iterative Matching And Pose Estimation With Adaptive Pooling", "abstract": "Previous methods solve feature matching and pose estimation using a two-stage process by first finding matches and then estimating the pose. As they ignore the geometric relationships between the two tasks, they focus on either improving the quality of matches or filtering potential outliers, leading to limited efficiency or accuracy. In contrast, we propose an iterative matching and pose estimation framework (IMP) leveraging the geometric connections between the two tasks: a few good matches are enough for a roughly accurate pose estimation; a roughly accurate pose can be used to guide the matching by providing geometric constraints. To this end, we implement a geometry-aware recurrent attention-based module which jointly outputs sparse matches and camera poses. Specifically, for each iteration, we first implicitly embed geometric information into the module via a pose-consistency loss, allowing it to predict geometry-aware matches progressively. Second, we introduce an efficient IMP, called EIMP, to dynamically discard keypoints without potential matches, avoiding redundant updating and significantly reducing the quadratic time complexity of attention computation in transformers. Experiments on YFCC100m, Scannet, and Aachen Day-Night datasets demonstrate that the proposed method outperforms previous approaches in terms of accuracy and efficiency."}, "cited_paper_content": {"title": "Superglue: Learning Feature Matching With Graph Neural Networks", "abstract": "This paper introduces SuperGlue, a neural network that matches two sets of local features by jointly finding correspondences and rejecting non-matchable points. Assignments are estimated by solving a differentiable optimal transport problem, whose costs are predicted by a graph neural network. We introduce a flexible context aggregation mechanism based on attention, enabling SuperGlue to reason about the underlying 3D scene and feature assignments jointly. Compared to traditional, hand-designed heuristics, our technique learns priors over geometric transformations and regularities of the 3D world through end-to-end training from image pairs. SuperGlue outperforms other learned approaches and achieves state-of-the-art results on the task of pose estimation in challenging real-world indoor and outdoor environments. The proposed method performs matching in real-time on a modern GPU and can be readily integrated into modern SfM or SLAM systems. The code and trained weights are publicly available at https://github.com/magicleap/SuperGluePretrainedNetwork."}, "keywords": ["night images", "SGMNet"], "citation_intent": "result"}
{"citing_id": "2304.02994v2", "cited_id": "1511.06181", "section_title": "Vision-Based Trajectory Analysis In Sports", "citation": "In #REFR , Mixed Integer Programming was used to track, under motion blur conditions, the position of volleyball, basketball, and soccer balls, with the goal of computing their motion trajectories.", "text_before_citation": ["Computer vision-based techniques have been successfully used in different sports disciplines to reconstruct the trajectory of various elements #OTHEREFR . Chen et al.", "developed an algorithm to compute the 3D trajectory of a volleyball ball in a static-camera video by transforming the output of the ball's visual tracker with prior knowledge of the volleyball court #OTHEREFR . In similar capturing settings, Calandre et al.", "reconstructed the trajectory of table-tennis  balls with a visual detection and tracking module, a registration between image and table coordinates , and a kinematic ball motion model #OTHEREFR .", "In #OTHEREFR , the tracking-by-deblatting paradigm was introduced for better localize of -hence compute better trajectories -fast-moving objects such as tabletennis or tennis balls in static camera videos."], "text_after_citation": ["In basketball analytics, different solutions have been proposed to reconstruct the trajectories of moving players #OTHEREFR .", "All these research endeavors presented a pipeline composed of a visual tracking algorithm to get player-specific localizations across video frames, and the application of a homography transformation between frame and court coordinates.", "Their differences lie in the tracking principles used, such as histogram-based condensation #OTHEREFR , Cam-Shift #OTHEREFR , deformable part models and kalman filters #OTHEREFR , or deep learning methods #OTHEREFR , and in the homography estimation, based either on manual annotation #OTHEREFR or on the matching of court lines and boundaries #OTHEREFR .", "Similar approaches have been also exploited in computer visionbased applications for soccer #OTHEREFR and ice-hockey #OTHEREFR .", "The methodology described in this paper is similar to #OTHEREFR in the design of a system to track an athlete (in our case a skier), and to compute key-point correspondence from which to estimate a homography."], "citing_paper_content": {"title": "Visualizing Skiers' Trajectories In Monocular Videos", "abstract": "Figure 1. Visualization of skiers' trajectories in monocular videos. We present SkiTraVis, an algorithm to compute and visualize the trajectory performed by an alpine skier in videos acquired by a single monocular camera. SkiTraVis takes in input a streaming video capturing a skier and visualizes its past trajectory in each frame. In this picture, the first and third rows show the input frames, while the second and fourth the trajectory computed by SkiTraVis. The interval in seconds from the start of the video is also reported."}, "cited_paper_content": {"title": "What Players Do With The Ball: A Physically Constrained Interaction Modeling", "abstract": "Tracking the ball is critical for video-based analysis of team sports. However, it is difficult, especially in lowresolution images, due to the small size of the ball, its speed that creates motion blur, and its often being occluded by players. In this paper, we propose a generic and principled approach to modeling the interaction between the ball and the players while also imposing appropriate physical constraints on the ball's trajectory. We show that our approach, formulated in terms of a Mixed Integer Program, is more robust and more accurate than several state-of-the-art approaches on real-life volleyball, basketball, and soccer sequences."}, "keywords": ["motion trajectories", "soccer balls"], "citation_intent": "method"}
{"citing_id": "2303.02698v2", "cited_id": "1601.07648", "section_title": "Introduction", "citation": "The experimental results support our idea that the new algorithm is more robust to noise and point discrepancy in point clouds than #REFR .", "text_before_citation": ["This paper is especially related to #OTHEREFR which proposed a two-stage procedure to recover correspondences between unlabelled, affinely transformed point sets.", "The first step finds the affine transformation and the second the matching between the point sets.", "Specifically, we adopt their first step of using the Grassmannian which is invariant under linear (affine) transformations of the point sets.", "As noted in #OTHEREFR it is provably better, in certain cases, to use an indefinite relaxation of the Quadratic Assignment Problem (QAP), used to find the matching, which almost always discovers the optimal permutation, while the common convex relaxation almost always fails.", "These theoretical results bring us to propose a relaxed quadratic programming approach to find the correct permutation (match) between the point sets."], "text_after_citation": [], "citing_paper_content": {"title": "Robust Affine Feature Matching Via Quadratic Assignment On Grassmannians", "abstract": "GraNNI (Grassmannians for Nearest Neighbours Identification) a new algorithm to solve the problem of affine registration is proposed. The algorithm is based on the Grassmannian of k-dimensional planes in R n and minimizing the Frobenius norm between the two elements of the Grassmannian. The Quadratic Assignment Problem (QAP) is used to find the matching. The results of the experiments show that the algorithm is more robust to noise and point discrepancy in point clouds than previous approaches."}, "cited_paper_content": {"title": "A Grassmannian Graph Approach To Affine Invariant Feature Matching", "abstract": "In this work, we present a novel and practical approach to address one of the longstanding problems in computer vision: 2D and 3D affine invariant feature matching. Our Grassmannian Graph (GrassGraph) framework employs a two stage procedure that is capable of robustly recovering correspondences between two unorganized, affinely related feature (point) sets. The first stage maps the feature sets to an affine invariant Grassmannian representation, where the features are mapped into the same subspace. It turns out that coordinate representations extracted from the Grassmannian differ by an arbitrary orthonormal matrix. In the second stage, by approximating the Laplace-Beltrami operator (LBO) on these coordinates, this extra orthonormal factor is nullified, providing true affine-invariant coordinates which we then utilize to recover correspondences via simple nearest neighbor relations. The resulting GrassGraph algorithm is empirically shown to work well in non-ideal scenarios with noise, outliers, and occlusions. Our validation benchmarks use an unprecedented 440,000+ experimental trials performed on 2D and 3D datasets, with a variety of parameter settings and competing methods. State-of-the-art performance in the majority of these extensive evaluations confirm the utility of our method."}, "keywords": ["point clouds"], "citation_intent": "result"}
{"citing_id": "2303.11096v1", "cited_id": "2003.03041", "section_title": "B. Performance Comparison", "citation": "Interestingly, statistical beamforming is shown to be very effective in the case of sparse channels #REFR , which supports the observed behavior here.", "text_before_citation": ["We observe that our proposed DNN-based technique outperforms all rival methods (except ZF with perfect CSI).", "In particular, we see a significant performance advantage in comparison to the DNN-based method in #OTHEREFR , especially for small feedback sizes.", "This should be mainly attributed to the fact that our #OTHEREFR ACS method in #OTHEREFR MRT with perfect CSI ZF with perfect CSI Significant Sum-Rate gain for small B #OTHEREFR ACS method in #OTHEREFR MRT with perfect CSI Gain due to sum-rate maximization Gain due to pre-beamforming proposed scheme exploits channel statistics knowledge at the BS.", "Even when there is practically no feedback (B \u00d1 0), our scheme is able to achieve a relatively large sum-rate because one component of the designed precoder in #OTHEREFR , namely the pre-beamforming matrix B depends only on channel statistics and not the feedback.", "In this case, our DNN is essentially performing a kind of statistical beamforming with (almost) no CSI."], "text_after_citation": ["The proposed method also outperforms MRT, which is due to the use of ZF precoding in our architecture.", "The advantage with respect to the case with no pre-beamforming matrix (the red curve) is rather small.", "This is due to the fact that the channels are sparse and L \u0103 \u03b2, and therefore with sufficient feedback no beam-selection is necessary.", "Here, the optimized beam-selection vector is \u03bb \u00ab 1, which is very close to no pre-beamforming with B \" F H .", "Finally, we see a similar advantage in comparison to the ACS method in #OTHEREFR , resulting from the direct maximization of sumrate rather than multiplexing gain."], "citing_paper_content": {"title": "Deep-Learning Aided Channel Training And Precoding In Fdd Massive Mimo With Channel Statistics Knowledge", "abstract": "We propose a method for channel training and precoding in FDD massive MIMO based on deep neural networks (DNNs), exploiting Downlink (DL) channel covariance knowledge. The DNN is optimized to maximize the DL multiuser sum-rate, by producing a pre-beamforming matrix based on user channel covariances that maps the original channel vectors to \"effective channels\". Measurements of these effective channels are received at the users via common pilot transmission and sent back to the base station (BS) through analog feedback without further processing. The BS estimates the effective channels from received feedback and constructs a linear precoder by concatenating the optimized pre-beamforming matrix with a zero-forcing precoder over the effective channels. We show that the proposed method yields significantly higher sum-rates than the state-of-the-art DNN-based channel training and precoding scheme, especially in scenarios with small pilot and feedback size relative to the channel coherence block length. Unlike many works in the literature, our proposition does not involve deployment of a DNN at the user side, which typically comes at a high computational cost and parameter-transmission overhead on the system, and is therefore considerably more practical. Index Terms-FDD massive MIMO, channel statistics knowledge, analog feedback, DNN-based training and precoding."}, "cited_paper_content": {"title": "Statistical Beamforming For Fdd Downlink Massive Mimo Via Spatial Information Extraction And Beam Selection", "abstract": "In this paper, we study the beamforming design problem in frequency-division duplexing (FDD) downlink massive MIMO systems, where instantaneous channel state information (CSI) is assumed to be unavailable at the base station (BS). We propose to extract the information of the angle-of-departures (AoDs) and the corresponding large-scale fading coefficients (a.k.a. spatial information) of the downlink channel from the uplink channel estimation procedure, based on which a novel downlink beamforming design is presented. By separating the subpaths for different users based on the spatial information and the hidden sparsity of the physical channel, we construct near-orthogonal virtual channels in the beamforming design. Furthermore, we derive a sum-rate expression and its approximations for the proposed system. Based on these closed-form rate expressions, we develop two low-complexity beam selection schemes and carry out asymptotic analysis to provide valuable insights on the system design. Numerical results demonstrate a significant performance improvement of our proposed algorithm over the state-of-the-art beamforming approach."}, "keywords": ["sparse channels", "statistical beamforming"], "citation_intent": "result"}
{"citing_id": "2303.12768v1", "cited_id": "1505.05599", "section_title": "Linear-Size Additive Spanners In Sub-Quadratic Time", "citation": "In this subsection, we prove the following theorem, which shows that an additive spanner with the current best error (as in #REFR ) can be computed in subquadratic time.", "text_before_citation": ["In this section we provide the proof of Theorem 1.2.", "We will first show in Section 6.1 a simple subquadratic time algorithm for computing an +O(n 3/7+\u03b5 )-additive spanner.", "Then we will slightly modify it to achieve better additive error bound in Section 6.2.", "6.1 A subquadratic time algorithm for O(n 3/7+\u03b5 )-additive spanner"], "text_after_citation": ["Theorem 6.1.", "There is an algorithm, that, given any undirected unweighted graph on n vertices and m edges, and any parameter \u03b5 > 0, computes an spanner on 2 O(1/\u03b5) \u2022 n edges with +\u00d5(n 3/7+\u03b5 ) error, in time\u00d5 m + 2 O(1/\u03b5) \u2022 n 13/7 .", "We start with the following lemma for a preliminary sparsification of the input graph.", "Lemma 6.2.", "There is an algorithm, that given any graph G and parameter 1 < d < n, computes in\u00d5(m) time an +\u00d5(n/d) additive spanner G \u2032 of G with |E(G \u2032 )| \u2264 O(nd) edges."], "citing_paper_content": {"title": "Almost-Optimal Sublinear Additive Spanners", "abstract": "Given an undirected unweighted graph G = (V, E) on n vertices and m edges, a subgraph H \u2286 G is a spanner of G with stretch function f : R + \u2192 R + , iff for every pair s, t of vertices in V , dist H (s, t) \u2264 f (dist G (s, t)). When f (d) = d + o(d), H is called a sublinear additive spanner ; when f (d) = d + o(n), H is called an additive spanner, and f (d) \u2212 d is usually called the additive stretch of H. As our primary result, we show that for any constant \u03b4 > 0 and constant integer k \u2265 2, every graph on n vertices has a sublinear additive spanner with stretch function f (d) = d + O(d 1\u22121/k) and O n 1+ 1+\u03b4 2 k+1 \u22121 edges. When k = 2, this improves upon the previous spanner construction with stretch function f (d) = d + O(d 1/2) and\u00d5(n 1+3/17) edges [Chechik, 2013]; for any constant integer k \u2265 3, this improves upon the previous spanner construction with stretch function f (d) = d + O(d 1\u22121/k) and O n"}, "cited_paper_content": {"title": "Better Distance Preservers And Additive Spanners", "abstract": "We make improvements to the upper bounds on several popular types of distance preserving graph sketches. These sketches are all various restrictions of the {\\em additive pairwise spanner} problem, in which one is given an undirected unweighted graph $G$, a set of node pairs $P$, and an error allowance $+\\beta$, and one must construct a sparse subgraph $H$ satisfying $\\delta_H(u, v) \\le \\delta_G(u, v) + \\beta$ for all $(u, v) \\in P$. The first part of our paper concerns {\\em pairwise distance preservers}, which make the restriction $\\beta=0$ (i.e. distances must be preserved {\\em exactly}). Our main result here is an upper bound of $|H| = O(n^{2/3}|P|^{2/3} + n|P|^{1/3})$ when $G$ is undirected and unweighted. This improves on existing bounds whenever $|P| = \\omega(n^{3/4})$, and it is the first such improvement in the last ten years. We then devise a new application of distance preservers to graph clustering algorithms, and we apply this algorithm to {\\em subset spanners}, which require $P = S \\times S$ for some node subset $S$, and {\\em (standard) spanners}, which require $P = V \\times V$. For both of these objects, our construction generalizes the best known bounds when the error allowance is constant, and we obtain the strongest polynomial error/sparsity tradeoff that has yet been reported (in fact, for subset spanners, ours is the {\\em first} nontrivial construction that enjoys improved sparsity from a polynomial error allowance). We leave open a conjecture that $O(n^{2/3}|P|^{2/3} + n)$ pairwise distance preservers are possible for undirected unweighted graphs. Resolving this conjecture in the affirmative would improve and simplify our upper bounds for all the graph sketches mentioned above."}, "keywords": ["additive spanner"], "citation_intent": "background"}
{"citing_id": "2303.02449v1", "cited_id": "1506.04579", "section_title": "Experimental Setting", "citation": "The learning rate is initialized to 1.0 for GLM and 0.01 for SLM, decreased at each iteration with polynomial decay #REFR , and further decreased by 1/10 at the ResNet50 backbone.", "text_before_citation": ["We use an ImageNet #OTHEREFR pretrained ResNet50 #OTHEREFR as the CAM backbone in GSLM and the classification layer is a 1 \u00d7 1 convolution with output channels adapted to the number of categories in the dataset.", "In Seed Reactivation, we adopt \u03c8 k for activation values limitation as in Eq. 4, where k is set to 6. The balance factor \u03b1 in Eq. (6) is set to 0.5.", "The thresholds \u03b8 f g and \u03b8 bg in Coarse Generation are set to 0.30 and 0.05, respectively.", "The number of iterations of SLM is set to 3 for GSLM.", "We adopt stochastic gradient descent (SGD) for optimization, with different optimization settings for GLM and SLM."], "text_after_citation": ["The batch size is 16, the number of training epochs is 5, and the weight decay is 0.001.", "The remaining optimization settings are the same as in #OTHEREFR .", "Moreover, we build a stronger baseline GSLM + by replacing GLM in GSLM with IRN #OTHEREFR and keeping SLM the same but with a single iteration. All other settings are the same as GSLM."], "citing_paper_content": {"title": "Exploit Cam By Itself: Complementary Learning System For Weakly Supervised Semantic Segmentation", "abstract": "Weakly Supervised Semantic Segmentation (WSSS) with image-level labels has long been suffering from fragmentary object regions led by Class Activation Map (CAM), which is incapable of generating fine-grained masks for semantic segmentation. To guide CAM to find more non-discriminating object patterns, this paper turns to an interesting working mechanism in agent learning named Complementary Learning System (CLS). CLS holds that the neocortex builds a sensation of general knowledge, while the hippocampus specially learns specific details, completing the learned patterns. Motivated by this simple but effective learning pattern, we propose a General-Specific Learning Mechanism (GSLM) to explicitly drive a coarse-grained CAM to a fine-grained pseudo mask. Specifically, GSLM develops a General Learning Module (GLM) and a Specific Learning Module (SLM). The GLM is trained with image-level supervision to extract coarse and general localization representations from CAM. Based on the general knowledge in the GLM, the SLM progressively exploits the specific spatial knowledge from the localization representations, expanding the CAM in an explicit way. To this end, we propose the Seed Reactivation to help SLM reactivate nondiscriminating regions by setting a boundary for activation values, which successively identifies more regions of CAM. Without extra refinement processes, our method is able to achieve breakthrough improvements for CAM of over 20.0% mIoU on PASCAL VOC 2012 and 10.0% mIoU on MS COCO 2014 datasets, representing a new state-of-the-art among existing WSSS methods."}, "cited_paper_content": {"title": "Parsenet: Looking Wider To See Better", "abstract": "We present a technique for adding global context to deep convolutional networks for semantic segmentation. The approach is simple, using the average feature for a layer to augment the features at each location. In addition, we study several idiosyncrasies of training, significantly increasing the performance of baseline networks (e.g. from FCN). When we add our proposed global feature, and a technique for learning normalization parameters, accuracy increases consistently even over our improved versions of the baselines. Our proposed approach, ParseNet, achieves state-of-the-art performance on SiftFlow and PASCAL-Context with small additional computational cost over baselines, and near current state-of-the-art performance on PASCAL VOC 2012 semantic segmentation with a simple approach. Code is available at this https URL ."}, "keywords": ["ResNet50 backbone"], "citation_intent": "method"}
{"citing_id": "2303.18094v1", "cited_id": "2003.06917", "section_title": "B. Normal Driving Maneuvers", "citation": "For the yaw rate estimation, the errors of the presented state of the art approaches are similar except for the GRU-based #REFR approach that performs the worst.", "text_before_citation": ["The acceleration plot shows that the maximum acceleration is a max y = 2.10 m/s 2 indicating low dynamic driving.", "The MAE of the different observers for the considered maneuver are shown in Table III .", "The proposed approach is able to perform best among the considered state of the art observers for the three variables.", "It can be seen that for the longitudinal velocity estimation, the KalmanNet based approach #OTHEREFR ranks second with an estimation error that is close to our approach; the EKF based on the dynamic bicycle model #OTHEREFR shows the worst performance with an error 3 times higher than the proposed approach.", "For the lateral velocity estimation, the GRU-based approach #OTHEREFR ranks second; the KalmanNet based approach #OTHEREFR performs the worst with an error 5 times higher than the proposed approach."], "text_after_citation": ["To further inspect the presented results, we plot the estimations by our proposed approach and the next best performing approach of each of the variables.", "We plot them in a subset of the maneuver sequence summarized in Fig. 6 .", "This subset includes the highest lateral accelerations in the dataset.", "All approaches are compared to the iTraceRT sensor values for reference.", "Higher accelerations are associated with more dynamic maneuvers and thus more challenging estimations. The plot is shown in Fig. 7 ."], "citing_paper_content": {"title": "Robust Lstm-Based Vehicle Velocity Observer For Regular And Near-Limits Applications", "abstract": "Accurate velocity estimation is key to vehicle control. While the literature describes how model-based and learningbased observers are able to estimate a vehicle's velocity in normal driving conditions, the challenge remains to estimate the velocity in near-limits maneuvers while using only conventional in-car sensors. In this paper, we introduce a novel neural network architecture based on Long Short-Term Memory (LSTM) networks to accurately estimate the vehicle's velocity in different driving conditions, including maneuvers at the limits of handling. The approach has been tested on real vehicle data and it provides more accurate estimations than state-of-the-art model-based and learning-based methods, for both regular and near-limits driving scenarios. Our approach is robust since the performance of the state-of-the-art observers deteriorates with higher dynamics, while our method adapts to different maneuvers, providing accurate estimations even at the vehicle's limits of handling."}, "cited_paper_content": {"title": "End-To-End Velocity Estimation For Autonomous Racing", "abstract": "Velocity estimation plays a central role in driverless vehicles, but standard and affordable methods struggle to cope with extreme scenarios like aggressive maneuvers due to the presence of high sideslip. To solve this, autonomous race cars are usually equipped with expensive external velocity sensors. In this paper, we present an end-to-end recurrent neural network that takes available raw sensors as input (IMU, wheel odometry, and motor currents) and outputs velocity estimates. The results are compared to two state-of-the-art Kalman filters, which respectively include and exclude expensive velocity sensors. All methods have been extensively tested on a formula student driverless race car with very high sideslip (10{\\deg} at the rear axle) and slip ratio (~20%), operating close to the limits of handling. The proposed network is able to estimate lateral velocity up to 15x better than the Kalman filter with the equivalent sensor input and matches (0.06 m/s RMSE) the Kalman filter with the expensive velocity sensor setup."}, "keywords": ["yaw rate estimation"], "citation_intent": "result"}
{"citing_id": "2304.11435v1", "cited_id": "1410.6072", "section_title": "Convergence Analysis", "citation": "Specifically, W (v) and V (v) are iteratively updated based on gradient descent method, it is necessary to prove that their corresponding minimization functions Eq. #REFR and Eq.", "text_before_citation": ["Herein, we analyze the convergence property for the proposed HLRCF."], "text_after_citation": ["(29) will gradually decrease and eventually stabalize as the number of iterations increases.", "First, we introduce an auxiliary function Z(\u2022), whose definition is given as Definition 1. The complete proof is provided in the Appendix A. Briefly, we can derive the following inequality chain", "EQUATION", "EQUATION", "The above inequalities indicate that the value of the objective function decreases with each iteration, thereby ensuring that the optimization of both W (v) and V (v) occurs in the direction of convergence."], "citing_paper_content": {"title": "Hyper-Laplacian Regularized Concept Factorization In Low-Rank Tensor Space For Multi-View Clustering", "abstract": "Tensor-oriented multi-view subspace clustering has achieved significant strides in assessing highorder correlations and improving clustering analysis of multi-view data. Nevertheless, most of existing investigations are typically hampered by the two flaws. First, self-representation based tensor subspace learning usually induces high time and space complexity, and is limited in perceiving nonlinear local structure in the embedding space. Second, the tensor singular value decomposition (t-SVD) model redistributes each singular value equally without considering the diverse importance among them. To well cope with the issues, we propose a hyper-Laplacian regularized concept factorization (HLRCF) in low-rank tensor space for multi-view clustering. Specifically, we adopt the concept factorization to explore the latent cluster-wise representation of each view. Further, the hypergraph Laplacian regularization endows the model with the capability of extracting the nonlinear local structures in the latent space. Considering that different tensor singular values associate structural information with unequal importance, we develop a self-weighted tensor Schatten p-norm to constrain the tensor comprised of all cluster-wise representations. Notably, the tensor with smaller size greatly decreases the time and space complexity in the low-rank optimization. Finally, experimental results on eight benchmark datasets exhibit that HLRCF outperforms other multi-view methods, showingcasing its superior performance."}, "cited_paper_content": {"title": "Nuclear Norm Of Higher-Order Tensors", "abstract": "We establish several mathematical and computational properties of the nuclear norm for higher-order tensors. We show that like tensor rank, tensor nuclear norm is dependent on the choice of base field --- the value of the nuclear norm of a real 3-tensor depends on whether we regard it as a real 3-tensor or a complex 3-tensor with real entries. We show that every tensor has a nuclear norm attaining decomposition and every symmetric tensor has a symmetric nuclear norm attaining decomposition. There is a corresponding notion of nuclear rank that, unlike tensor rank, is upper semicontinuous. We establish an analogue of Banach's theorem for tensor spectral norm and Comon's conjecture for tensor rank --- for a symmetric tensor, its symmetric nuclear norm always equals its nuclear norm. We show that computing tensor nuclear norm is NP-hard in several sense. Deciding weak membership in the nuclear norm unit ball of 3-tensors is NP-hard, as is finding an $\\varepsilon$-approximation of nuclear norm for 3-tensors. In addition, the problem of computing spectral or nuclear norm of a 4-tensor is NP-hard, even if we restrict the 4-tensor to be bi-Hermitian, bisymmetric, positive semidefinite, nonnegative valued, or all of the above. We discuss some simple polynomial-time approximation bounds. As an aside, we show that the nuclear $(p,q)$-norm of a matrix is NP-hard in general but can be computed in polynomial-time if $p=1$, $q = 1$, or $p=q=2$, with closed-form expressions for the nuclear $(1,q)$- and $(p,1)$-norms."}, "keywords": ["gradient descent method"], "citation_intent": "method"}
{"citing_id": "2303.01414v1", "cited_id": "1711.00103", "section_title": "Our Results", "citation": "These two results make up one ( 3 2 + )-approximation that improves on the best known result by Jansen and Land #REFR in multiple ways.", "text_before_citation": ["The resulting knapsack instance will only have a small amount of different item sizes.", "We then apply an algorithm introduced by Axiotis and Tzamos #OTHEREFR that works well on such instances.", "Thanks to our rounding we will be able to do the pre-processing of their algorithm in linear time as well. Theorem 2.", "For moldable job scheduling there exists an algorithm that for instances with m \u2264 16n and for any > 0 yields a #OTHEREFR 2 + approximation in time:", "O(n log 2 ( 1 + log( m) ) + n log( 1 )log( m))"], "text_after_citation": ["For large m we manage to reduce the dependency on m even further.", "When m is small we improve on their running time by reducing the dependency on by a factor of 1 and several polylogarithmic factors.", "We also argue that our algorithm is overall simpler compared to theirs, as we do not require to solve knapsack with compressible items in a complicated manner.", "Instead our algorithm merely constructs the modified knapsack instance and delegates to a simple and elegant algorithm from Axiotis and Tzamos #OTHEREFR ."], "citing_paper_content": {"title": "Improved Algorithms For Monotone Moldable Job Scheduling Using Compression And Convolution", "abstract": "In the moldable job scheduling problem one has to assign a set of n jobs to m machines, in order to minimize the time it takes to process all jobs. Each job is moldable, so it can be assigned not only to one but any number of the equal machines. We assume that the work of each job is monotone and that jobs can be placed noncontiguously. In this work we present a (3 2 +)-approximation algorithm with a worst-case runtime of O(n log 2 (1 + log(m)) + n log(1)log(m)) when m \u2264 16n. This is an improvement over the best known algorithm of the same quality by a factor of 1 and several logarithmic dependencies. We complement this result with an improved FPTAS with running time O(n log 2 (1 + log(m))) for instances with many machines m > 8 n. This yields a 3 2-approximation with runtime O(n log 2 (log m)) when m > 16n. We achieve these results through one new core observation: In an approximation setting one does not need to consider all m possible allotments for each job. We will show that we can reduce the number of relevant allotments for each job from m to O(1 + log(m)). Using this observation immediately yields the improved FPTAS. For the other result we use a reduction to the knapsack problem first introduced by Mouni\u00e9, Rapine and Trystram. We use the reduced number of machines to give a new elaborate rounding scheme and define a modified version of this this knapsack instance. This in turn allows for the application of a convolution based algorithm by Axiotis and Tzamos. We further back our theoretical results through a practical implementation and compare our algorithm to the previously known best result."}, "cited_paper_content": {"title": "Scheduling Monotone Moldable Jobs In Linear Time", "abstract": "A moldable job is a job that can be executed on an arbitrary number of processors, and whose processing time depends on the number of processors allotted to it. A moldable job is monotone if its work doesn't decrease for an increasing number of allotted processors. We consider the problem of scheduling monotone moldable jobs to minimize the makespan. We argue that for certain compact input encodings a polynomial algorithm has a running time polynomial in n and log(m), where n is the number of jobs and m is the number of machines. We describe how monotony of jobs can be used to counteract the increased problem complexity that arises from compact encodings, and give tight bounds on the approximability of the problem with compact encoding: it is NP-hard to solve optimally, but admits a PTAS. The main focus of this work are efficient approximation algorithms. We describe different techniques to exploit the monotony of the jobs for better running times, and present a (3/2+{\\epsilon})-approximate algorithm whose running time is polynomial in log(m) and 1/{\\epsilon}, and only linear in the number n of jobs."}, "keywords": [")-approximation"], "citation_intent": "result"}
{"citing_id": "2304.02480v1", "cited_id": "1606.03476", "section_title": "A. Imitation Learning (Il)", "citation": "Instead of assuming linear rewards, GAIL #REFR leverages DNNs to represent the reward function (also called the discriminator) and shows dramatic performance improvements.", "text_before_citation": ["A basic difference between BC and DAGGER is that the former learns with only expert data, while the latter needs expert policies and on-line interactions.", "IRL deals with the IL problem from the aspect of RL #OTHEREFR .", "There are two alternate steps in IRL: (1) learn a reward function that can assign high rewards to expert data and low rewards to agent's experiences; (2) conduct RL with the learned reward function in step #OTHEREFR to improve the performance of agents.", "Abbeel and Ng #OTHEREFR present apprenticeship learning, which employs linear functions to approximate the reward function.", "Maximum entropy IRL #OTHEREFR uses probabilistic models to obtain a reward function with less ambiguity based on maximum entropy and maintains performance guarantees."], "text_after_citation": ["For more recent advances in IRL, please refer to #OTHEREFR - #OTHEREFR ."], "citing_paper_content": {"title": "Quantum Imitation Learning", "abstract": "Despite remarkable successes in solving various complex decision-making tasks, training an imitation learning (IL) algorithm with deep neural networks (DNNs) suffers from the high computation burden. In this work, we propose quantum imitation learning (QIL) with a hope to utilize quantum advantage to speed up IL. Concretely, we develop two QIL algorithms, quantum behavioural cloning (Q-BC) and quantum generative adversarial imitation learning (Q-GAIL). Q-BC is trained with a negative log-likelihood loss in an off-line manner that suits extensive expert data cases, whereas Q-GAIL works in an inverse reinforcement learning scheme, which is on-line and on-policy that is suitable for limited expert data cases. For both QIL algorithms, we adopt variational quantum circuits (VQCs) in place of DNNs for representing policies, which are modified with data re-uploading and scaling parameters to enhance the expressivity. We first encode classical data into quantum states as inputs, then perform VQCs, and finally measure quantum outputs to obtain control signals of agents. Experiment results demonstrate that both Q-BC and Q-GAIL can achieve comparable performance compared to classical counterparts, with the potential of quantum speed-up. To our knowledge, we are the first to propose the concept of QIL and conduct pilot studies, which paves the way for the quantum era. Index Terms-Quantum imitation learning (QIL), variational quantum circuits (VQCs), imitation learning (IL), behavioural cloning (BC), inverse reinforcement learning (IRL)."}, "cited_paper_content": {"title": "Generative Adversarial Imitation Learning", "abstract": "Consider learning a policy from example expert behavior, without interaction with the expert or access to reinforcement signal. One approach is to recover the expert's cost function with inverse reinforcement learning, then extract a policy from that cost function with reinforcement learning. This approach is indirect and can be slow. We propose a new general framework for directly extracting a policy from data, as if it were obtained by reinforcement learning following inverse reinforcement learning. We show that a certain instantiation of our framework draws an analogy between imitation learning and generative adversarial networks, from which we derive a model-free imitation learning algorithm that obtains significant performance gains over existing model-free methods in imitating complex behaviors in large, high-dimensional environments."}, "keywords": ["reward function"], "citation_intent": "background"}
{"citing_id": "2303.11570v1", "cited_id": "1610.05820", "section_title": "Experimental Settings", "citation": "For privacy guarantee, we construct a simple yet general membership inference attack (MIA) based on #REFR using the output of the unlearned model and test the attack success rate (ASR).", "text_before_citation": ["For the fine-tune process in Boundary Unlearning we use a learning rate of 10 \u22125 for 10 epochs.", "Note that we use the same forgetting class as the forgetting data for all unlearning methods and the rest of classes are the remaining data. Metrics.", "We verify the unlearning performance on both utility and privacy guarantees.", "For utility guarantee, we utilize four accuracy metrics: accuracy on the remaining training data D r , forgetting training data D f , remaining testing data D rt , forgetting testing data D f t .", "The unlearned model is expected to get close accuracy with the retrained model."], "text_after_citation": ["MIA aims to infer whether a data record was used to train a target ML model or not #OTHEREFR , and its ASR is widely used as an evaluation metric to verify the privacy guarantee of an unlearning method #OTHEREFR .", "Ideally, a forgetting procedure should have the same attack success as a retrained model."], "citing_paper_content": {"title": "Boundary Unlearning", "abstract": "The practical needs of the \"right to be forgotten\" and poisoned data removal call for efficient machine unlearning techniques, which enable machine learning models to unlearn, or to forget a fraction of training data and its lineage. Recent studies on machine unlearning for deep neural networks (DNNs) attempt to destroy the influence of the forgetting data by scrubbing the model parameters. However, it is prohibitively expensive due to the large dimension of the parameter space. In this paper, we refocus our attention from the parameter space to the decision space of the DNN model, and propose Boundary Unlearning, a rapid yet effective way to unlearn an entire class from a trained DNN model. The key idea is to shift the decision boundary of the original DNN model to imitate the decision behavior of the model retrained from scratch. We develop two novel boundary shift methods, namely Boundary Shrink and Boundary Expanding, both of which can rapidly achieve the utility and privacy guarantees. We extensively evaluate Boundary Unlearning on CIFAR-10 and Vggface2 datasets, and the results show that Boundary Unlearning can effectively forget the forgetting class on image classification and face recognition tasks, with an expected speed-up of 17\u00d7 and 19\u00d7, respectively, compared with retraining from the scratch."}, "cited_paper_content": {"title": "Membership Inference Attacks Against Machine Learning Models", "abstract": "We quantitatively investigate how machine learning models leak information about the individual data records on which they were trained. We focus on the basic membership inference attack: given a data record and black-box access to a model, determine if the record was in the model's training dataset. To perform membership inference against a target model, we make adversarial use of machine learning and train our own inference model to recognize differences in the target model's predictions on the inputs that it trained on versus the inputs that it did not train on. We empirically evaluate our inference techniques on classification models trained by commercial \"machine learning as a service\" providers such as Google and Amazon. Using realistic datasets and classification tasks, including a hospital discharge dataset whose membership is sensitive from the privacy perspective, we show that these models can be vulnerable to membership inference attacks. We then investigate the factors that influence this leakage and evaluate mitigation strategies."}, "keywords": ["unlearned model", "membership inference attack"], "citation_intent": "method"}
{"citing_id": "2303.14608v1", "cited_id": "1905.04899", "section_title": "Human-Model Alignment", "citation": "To our surprise, some data augmentation strategies that showed better performance on WSOL (Cutout and CutMix #REFR ), showed the opposite results in our human-model alignment experiments.", "text_before_citation": ["Then, a plot representing EHR score at each threshold is obtained. Finally, the area under the curve (AUC) is computed. Let S c,\u03bb", "(i,j) = (L c (i,j) > \u03bb)", "where \u03bb denotes a threshold.", "EQUATION", "Experimental results on human-model alignment are given in Tab. 1."], "text_after_citation": ["Attribution maps calculated from baseline ResNet-50 scored the highest alignment with human-drawn bounding boxes.", "This does not indicate that the mixed sample data augmentation techniques degrade localization ability.", "Rather, it shows that the features used importantly for the prediction of models are less consistent with that of humans when the model is trained with the augmentation strategy.", "CutMix and Salien-cyMix consistently score last or second last except for the IBA-EHR experiment, where the score difference between the four augmentation methods is marginal.", "Qualitative analysis at Fig."], "citing_paper_content": {"title": "Analyzing Effects Of Mixed Sample Data Augmentation On Model Interpretability", "abstract": "Data augmentation strategies are actively used when training deep neural networks (DNNs). Recent studies suggest that they are effective at various tasks. However, the effect of data augmentation on DNNs' interpretability is not yet widely investigated. In this paper, we explore the relationship between interpretability and data augmentation strategy in which models are trained with different data augmentation methods and are evaluated in terms of interpretability. To quantify the interpretability, we devise three evaluation methods based on alignment with humans, faithfulness to the model, and the number of humanrecognizable concepts in the model. Comprehensive experiments show that models trained with mixed sample data augmentation show lower interpretability, especially for CutMix and SaliencyMix augmentations. This new finding suggests that it is important to carefully adopt mixed sample data augmentation due to the impact on model interpretability, especially in mission-critical applications."}, "cited_paper_content": {"title": "Cutmix: Regularization Strategy To Train Strong Classifiers With Localizable Features", "abstract": "Regional dropout strategies have been proposed to enhance the performance of convolutional neural network classifiers. They have proved to be effective for guiding the model to attend on less discriminative parts of objects (e.g. leg as opposed to head of a person), thereby letting the network generalize better and have better object localization capabilities. On the other hand, current methods for regional dropout remove informative pixels on training images by overlaying a patch of either black pixels or random noise. Such removal is not desirable because it leads to information loss and inefficiency during training. We therefore propose the CutMix augmentation strategy: patches are cut and pasted among training images where the ground truth labels are also mixed proportionally to the area of the patches. By making efficient use of training pixels and retaining the regularization effect of regional dropout, CutMix consistently outperforms the state-of-the-art augmentation strategies on CIFAR and ImageNet classification tasks, as well as on the ImageNet weakly-supervised localization task. Moreover, unlike previous augmentation methods, our CutMix-trained ImageNet classifier, when used as a pretrained model, results in consistent performance gains in Pascal detection and MS-COCO image captioning benchmarks. We also show that CutMix improves the model robustness against input corruptions and its out-of-distribution detection performances. Source code and pretrained models are available at this https URL ."}, "keywords": ["data augmentation strategies"], "citation_intent": "result"}
{"citing_id": "2304.06713v1", "cited_id": "1712.06349", "section_title": "Introduction", "citation": "Understanding the quantum query complexity of Boolean functions f : D \u2192 {\u22121, 1}, where D is a subset of {\u22121, 1} n , has been a crucial task of quantum information science #REFR .", "text_before_citation": [], "text_after_citation": ["Query complexity is a model of computing where rigorous upper and lower bounds can be proven, which allows one to compare the power of quantum computers with classical ones.", "Many celebrated quantum algorithms show an advantage in terms of query complexity, for example in unstructured search #OTHEREFR , period finding #OTHEREFR , Simon's problem #OTHEREFR , NAND-tree evaluation #OTHEREFR and element distinctness #OTHEREFR .", "However, these advantages are limited to be polynomial in the case of total functions (those with D = {\u22121, 1} n ), while they can be exponential for highly structured problems (informally, this means that |D| = o(2 n )), such as for Simon's problem #OTHEREFR , period finding #OTHEREFR or k-fold forrelation #OTHEREFR .", "It is widely believed that a lot of structure is necessary for superpolynomial speedups 1 .", "The following folklore conjecture, which has circulated since the late 90s, but was first formally posed by Aaronson and Ambainis #OTHEREFR , formalizes this idea."], "citing_paper_content": {"title": "Influences Of Fourier Completely Bounded Polynomials And Classical Simulation Of Quantum Algorithms", "abstract": "We give a new presentation of the main result of Arunachalam, Bri\u00ebt and Palazuelos (SICOMP'19) and show that quantum query algorithms are characterized by a new class of polynomials which we call Fourier completely bounded polynomials. We conjecture that all such polynomials have an influential variable. This conjecture is weaker than the famous Aaronson-Ambainis (AA) conjecture (Theory of Computing'14), but has the same implications for classical simulation of quantum query algorithms. We prove a new case of the AA conjecture by showing that it holds for homogeneous Fourier completely bounded polynomials. This implies that if the output of d-query quantum algorithm is a homogeneous polynomial p of degree 2d, then it has a variable with influence at least Var[p] 2. In addition, we give an alternative proof of the results of Bansal, Sinha and de Wolf (CCC'22 and QIP'23) showing that block-multilinear completely bounded polynomials have influential variables. Our proof is simpler, obtains better constants and does not use randomness. This research was supported by the European Union's Horizon 2020 research and innovation programme under the Marie Sk lodowska-Curie grant agreement no. 945045, and by the NWO Gravitation project NETWORKS under grant no. 024.002.003. 1 Recently, Yamakawa and Zhandry showed that superpolynomial speedups can be attained in unstructured search problems. However, that does not contradict the believe that structure is needed to achieve superpolynomial speedups in decision problems, which are those modeled by Boolean functions [YZ22]."}, "cited_paper_content": {"title": "Understanding Quantum Algorithms Via Query Complexity", "abstract": "Query complexity is a model of computation in which we have to compute a function $f(x_1, \\ldots, x_N)$ of variables $x_i$ which can be accessed via queries. The complexity of an algorithm is measured by the number of queries that it makes. Query complexity is widely used for studying quantum algorithms, for two reasons. First, it includes many of the known quantum algorithms (including Grover's quantum search and a key subroutine of Shor's factoring algorithm). Second, one can prove lower bounds on the query complexity, bounding the possible quantum advantage. In the last few years, there have been major advances on several longstanding problems in the query complexity. In this talk, we survey these results and related work, including: ::: - the biggest quantum-vs-classical gap for partial functions (a problem solvable with 1 query quantumly but requiring $\\Omega(\\sqrt{N})$ queries classically); ::: - the biggest quantum-vs-determistic and quantum-vs-probabilistic gaps for total functions (for example, a problem solvable with $M$ queries quantumly but requiring $\\tilde{\\Omega}(M^{2.5})$ queries probabilistically); ::: - the biggest probabilistic-vs-deterministic gap for total functions (a problem solvable with $M$ queries probabilistically but requiring $\\tilde{\\Omega}(M^{2})$ queries deterministically); ::: - the bounds on the gap that can be achieved for subclasses of functions (for example, symmetric functions); ::: - the connections between query algorithms and approximations by low-degree polynomials."}, "keywords": ["quantum query"], "citation_intent": "background"}
{"citing_id": "2303.15655v1", "cited_id": "1703.06103", "section_title": "Link Prediction", "citation": "In terms of MRR and Hits@1, HIE is worse than that of several state-of-the-art models, while it obtains a 30% higher Hits@1 than R-GCN #REFR which uses more additional neighbor information.", "text_before_citation": ["Due to the massive inversion relation patterns existing in WN18 where the ratio reaches 94% #OTHEREFR .", "To meet the inversion equation, the transformation matrix in HIE will be an identity matrix.", "That's to say, more information on distance measurement is omitted.", "Thus, the proposed HIE does not achieve best scores of MRR and Hits@1 on WN18.", "\u2022 On WN18, the proposed HIE outperforms the compared algorithms and achieves the best MR, Hits@3 and Hits@10."], "text_after_citation": ["\u2022 On YAGO3-10, in terms of MRR, Hits@3 and Hits@10, HIE achieves the best score.", "As for the MR and Hits@1, HIE is slightly worse than the state-of-the-art RotatE and InteractE in the corresponding aspects.", "Notably, HIE obtains a 7% higher MRR than that of neural networkbased model M-DCN, 59% higher MRR than that of semantic matching-based model DistMult, and 9% higher than MRR that of translation based model RotatE.", "Strikingly, our proposed HIE achieves better performance scores on Hits@3 and Hits@10 than that of ConvE, RotatE and DistMult.", "From Table. 5, we can conclude that:"], "citing_paper_content": {"title": "Joint Embedding In Hierarchical Distance And Semantic Representation Learning For Link Prediction", "abstract": "The link prediction task aims to predict missing links in the knowledge graph and is essential for the downstream application. Existing well-known models deal with this task by mainly focusing on representing knowledge graph triplets in the distance space or semantic space. However, they can not fully capture the information of head and tail entities, nor even make good use of hierarchical level information. Thus, in this paper, we propose a novel knowledge graph embedding model for the link prediction task, namely, HIE, which models each triplet (h, r, t) into distance measurement space and semantic measurement space simultaneously. Moreover, to leverage the hierarchical information of entities and relations, HIE is introduced into hierarchical-aware space for better representation learning. Specifically, we apply distance transformation operation on the head entity in distance space to obtain the tail entity instead of translation-based or rotation-based approaches. Experimental results of HIE on four real-world datasets show that HIE outperforms several existing state-of-the-art knowledge graph embedding methods on link prediction task and deals with complex relations accurately."}, "cited_paper_content": {"title": "Modeling Relational Data With Graph Convolutional Networks", "abstract": "Knowledge bases play a crucial role in many applications, for example question answering and information retrieval. Despite the great effort invested in creating and maintaining them, even the largest representatives (e.g., Yago, DBPedia or Wikidata) are highly incomplete. We introduce relational graph convolutional networks (R-GCNs) and apply them to two standard knowledge base completion tasks: link prediction (recovery of missing facts, i.e.~subject-predicate-object triples) and entity classification (recovery of missing attributes of entities). R-GCNs are a generalization of graph convolutional networks, a recent class of neural networks operating on graphs, and are developed specifically to deal with highly multi-relational data, characteristic of realistic knowledge bases. Our methods achieve competitive performance on standard benchmarks for both tasks, demonstrating especially promising results on the challenging FB15k-237 subset of Freebase."}, "keywords": ["MRR", "R-GCN"], "citation_intent": "result"}
{"citing_id": "2303.09032v1", "cited_id": "1611.04717", "section_title": "Introduction", "citation": "We build COE on commonly used value decomposition methods, and used the hash-based counting technique #REFR to enable counting the visitations in continuous state-action domains.", "text_before_citation": ["At each node of the tree, the preceding agent's taken action can be considered as the parent node of the current agent.", "Then we can perform optimism-based exploration by computing the upper confidence bounds of each action for the current agent, conditioned on its parent node's visitation count.", "We disable exploration after training to obtain decentralized agents for execution.", "We first review the basic background in the MARL setting and the UCT algorithm.", "Then, we describe how the UCT exploration can be applied to the MARL setting to encourage cooperative exploration."], "text_after_citation": ["Our empirical results on various benchmark domains show that our method is more effective than well-known baselines in exploration-challenging tasks, and matches baseline performance in general MARL tasks."], "citing_paper_content": {"title": "Conditionally Optimistic Exploration For Cooperative Deep Multi-Agent Reinforcement Learning", "abstract": "Efficient exploration is critical in cooperative deep Multi-Agent Reinforcement Learning (MARL). In this paper, we propose an exploration method that efficiently encourages cooperative exploration based on the idea of the theoretically justified tree search algorithm UCT (Upper Confidence bounds applied to Trees). The high-level intuition is that to perform optimism-based exploration, agents would achieve cooperative strategies if each agent's optimism estimate captures a structured dependency relationship with other agents. At each node (i.e., action) of the search tree, UCT performs optimismbased exploration using a bonus derived by conditioning on the visitation count of its parent node. We provide a perspective to view MARL as tree search iterations and develop a method called Conditionally Optimistic Exploration (COE). We assume agents take actions following a sequential order, and consider nodes at the same depth of the search tree as actions of one individual agent. COE computes each agent's state-action value estimate with an optimistic bonus derived from the visitation count of the state and joint actions taken by agents up to the current agent. COE is adaptable to any value decomposition method for centralized training with decentralized execution. Experiments across various cooperative MARL benchmarks show that COE outperforms current stateof-the-art exploration methods on hard-exploration tasks. Preprint. Under review."}, "cited_paper_content": {"title": "Exploration: A Study Of Count-Based Exploration For Deep Reinforcement Learning", "abstract": "Count-based exploration algorithms are known to perform near-optimally when used in conjunction with tabular reinforcement learning (RL) methods for solving small discrete Markov decision processes (MDPs). It is generally thought that count-based methods cannot be applied in high-dimensional state spaces, since most states will only occur once. ::: Recent deep RL exploration strategies are able to deal with high-dimensional continuous state spaces through complex heuristics, often relying on optimism in the face of uncertainty or intrinsic motivation. ::: ::: In this work, we describe a surprising finding: a simple generalization of the classic count-based approach can reach near state-of-the-art performance on various high-dimensional and/or continuous deep RL benchmarks. States are mapped to hash codes, which allows to count their occurrences with a hash table. These counts are then used to compute a reward bonus according to the classic count-based exploration theory. We find that simple hash functions can achieve surprisingly good results on many challenging tasks. Furthermore, we show that a domain-dependent learned hash code may further improve these results. ::: ::: Detailed analysis reveals important aspects of a good hash function: 1) having appropriate granularity and 2) encoding information relevant to solving the MDP. This exploration strategy achieves near state-of-the-art performance on both continuous control tasks and Atari 2600 games, hence providing a simple yet powerful baseline for solving MDPs that require considerable exploration."}, "keywords": ["continuous state-action domains", "hash-based counting technique"], "citation_intent": "method"}
{"citing_id": "2303.04437v1", "cited_id": "1802.01933", "section_title": "Introduction", "citation": "Current workhorses to address transparency concerns in machine learning include black-box explanation and transparent design techniques #REFR .", "text_before_citation": ["The ever-increasing integration of machine learning models in high-stakes decision-making contexts (e.g., healthcare, justice, finance) has fostered a growing demand for transparency in recent years."], "text_after_citation": ["Black-box explanation techniques aim at explaining complex machine learning models in a post-hoc fashion with global explanations such as Trepan #OTHEREFR and BETA #OTHEREFR or local explanations such as LIME (Ribeiro et al., 2016) and SHAP (Lundberg and Lee, 2017).", "On the other hand, transparent design concerns the development of inherently interpretable models such as rule lists #OTHEREFR , rule sets #OTHEREFR , decision trees #OTHEREFR , and scoring systems #OTHEREFR .", "However, both black-box explanations and transparent design face performance and trustworthiness challenges that can prevent their wide adoption.", "On the one hand, while inherently interpretable models can be more easily understood and adopted by non-domain experts, their out-of-the-box performance can be worst than non-transparent models.", "Moreover, training such models to optimality is often NP-hard due to their discrete nature."], "citing_paper_content": {"title": "Learning Hybrid Interpretable Models: Theory, Taxonomy, And Methods", "abstract": "A hybrid model involves the cooperation of an interpretable model and a complex black box. At inference, any input of the hybrid model is assigned to either its interpretable or complex component based on a gating mechanism. The advantages of such models over classical ones are twofold: 1) They grant users precise control over the level of transparency of the system and 2) They can potentially perform better than a standalone black box since redirecting some of the inputs to an interpretable model implicitly acts as regularization. Still, despite their high potential, hybrid models remain under-studied in the interpretability/explainability literature. In this paper, we remedy this fact by presenting a thorough investigation of such models from three perspectives: Theory, Taxonomy, and Methods. First, we explore the theory behind the generalization of hybrid models from the Probably-Approximately-Correct (PAC) perspective. A consequence of our PAC guarantee is the existence of a sweet spot for the optimal transparency of the system. When such a sweet spot is attained, a hybrid model can potentially perform better than a standalone black box. Secondly, we provide a general taxonomy for the different ways of training hybrid models: the Post-Black-Box and Pre-Black-Box paradigms. These approaches differ in the order in which the interpretable and complex components are trained. We show where the state-of-the-art hybrid models Hybrid-Rule-Set and Companion-Rule-List fall in this taxonomy. Thirdly, we implement the two paradigms in a single method: HybridCORELS, which extends the CORELS algorithm to hybrid modeling. By leveraging CORELS, Hy-bridCORELS provides a certificate of optimality of its interpretable component and precise control over transparency. We finally show empirically that HybridCORELS is competitive with existing hybrid models, and performs just as well as a standalone black box (or even better) while being partly transparent."}, "cited_paper_content": {"title": "A Survey Of Methods For Explaining Black Box Models", "abstract": "In recent years, many accurate decision support systems have been constructed as black boxes, that is as systems that hide their internal logic to the user. This lack of explanation constitutes both a practical and an ethical issue. The literature reports many approaches aimed at overcoming this crucial weakness, sometimes at the cost of sacrificing accuracy for interpretability. The applications in which black box decision systems can be used are various, and each approach is typically developed to provide a solution for a specific problem and, as a consequence, it explicitly or implicitly delineates its own definition of interpretability and explanation. The aim of this article is to provide a classification of the main problems addressed in the literature with respect to the notion of explanation and the type of black box system. Given a problem definition, a black box type, and a desired explanation, this survey should help the researcher to find the proposals more useful for his own work. The proposed classification of approaches to open black box models should also be useful for putting the many research open questions in perspective."}, "keywords": ["black-box explanation"], "citation_intent": "background"}
{"citing_id": "2304.08384v1", "cited_id": "1803.04189", "section_title": "Introduction", "citation": "Noise2Noise #REFR is the first to use pairs of two different noisy images constructed from the same clean image to train a denoising model. Strictly speaking, Noise2Noise is not an unsupervised learning method. Collecting noisy pairs is still difficult.", "text_before_citation": ["Image denoising #OTHEREFR has been studied for many years.", "Suppose x is a clean image, y is a noisy image of x, and p (y | x) is the noise model.", "Supervised learning methods try to train a model representing a mapping from y to x.", "Due to the difficulty of collecting paired clean and noisy images in practice, methods in unsupervised learning manner are focus of research."], "text_after_citation": ["Despite all this, many other methods #OTHEREFR are inspired from Noise2Noise or borrow the idea behind of it.", "These methods achieve good performance in some simple noise models.", "The main drawback of current methods is the constraint of application.", "Once the noise model is complicated, they either are not applicable or have poor performance.", "Nois-ier2Noise #OTHEREFR can only handle additive noise and pepper noise. Recorrupted-to-Recorrupted [17] is limited to Gaussian noise."], "citing_paper_content": {"title": "Unsupervised Image Denoising With Score Function", "abstract": "Though achieving excellent performance in some cases, current unsupervised learning methods for single image denoising usually have constraints in applications. In this paper, we propose a new approach which is more general and applicable to complicated noise models. Utilizing the property of score function, the gradient of logarithmic probability, we define a solving system for denoising. Once the score function of noisy images has been estimated, the denoised result can be obtained through the solving system. Our approach can be applied to multiple noise models, such as the mixture of multiplicative and additive noise combined with structured correlation. Experimental results show that our method is comparable when the noise model is simple, and has good performance in complicated cases where other methods are not applicable or perform poorly."}, "cited_paper_content": {"title": "Noise2Noise: Learning Image Restoration Without Clean Data", "abstract": "We apply basic statistical reasoning to signal reconstruction by machine learning -- learning to map corrupted observations to clean signals -- with a simple and powerful conclusion: it is possible to learn to restore images by only looking at corrupted examples, at performance at and sometimes exceeding training using clean data, without explicit image priors or likelihood models of the corruption. In practice, we show that a single model learns photographic noise removal, denoising synthetic Monte Carlo images, and reconstruction of undersampled MRI scans -- all corrupted by different processes -- based on noisy data only."}, "keywords": ["denoising model"], "citation_intent": "background"}
{"citing_id": "2304.01575v1", "cited_id": "1912.09893", "section_title": "Experimental Results", "citation": "First, it demonstrates the soundness of the GNN architecture used in the first experiment, which achieves results comparable to those of models carefully optimized on the benchmark datasets #REFR .", "text_before_citation": ["Since ASAPool and SAGPool implement an additional MP operation internally, we used only 1 GIN layer before them, rather than 2 as for the other pooling methods.", "Finally, Fig.", "3 shows the average accuracy and the average run-time obtained on the seven benchmark datasets (the detailed results are in Appendix B.3).", "These benchmarks are not designed to test the expressive power and, thus, a GNN equipped with a non-expressive pooling operator could achieve good performance.", "For example, this happens in those datasets where all the necessary information is captured by the first two GIN layers that come before pooling or in datasets where only a small part of the graph determines the class. Nevertheless, this second experiment serves two purposes."], "text_after_citation": ["Second, and most importantly, it shows that the performances on the benchmark datasets and EXPWL1 are aligned; this underlines the relevance of our theoretical result on the expressivity in practical applications.", "It is worth noting that on the benchmark datasets, it was not possible to obtain a pooling ratio of 0.1 for both Graclus and ECPool.", "Using a pooling ratio of 0.5 gives Graclus and ECPool an advantage over other methods, which makes the comparison not completely fair and shows an important limitation of these methods.", "As a concluding remark, we comment on the training time of the dense and sparse pooling methods.", "A popular argument in favor of sparse pooling methods is their computational advantage compared to the dense ones. Our results show that this is not the case."], "citing_paper_content": {"title": "The Expressive Power Of Pooling In Graph Neural Networks", "abstract": "In Graph Neural Networks (GNNs), hierarchical pooling operators generate a coarser representation of the input data by creating local summaries of the graph structure and its vertex features. Considerable attention has been devoted to studying the expressive power of message-passing (MP) layers in GNNs, while a study on how pooling operators affect the expressivity of a GNN is still lacking. Additionally, despite the recent advances in the design of effective pooling operators, there is not a principled criterion to compare them. Our work aims to fill this gap by providing sufficient conditions for a pooling operator to fully preserve the expressive power of the MP layers before it. These conditions serve as a universal and theoretically-grounded criterion for choosing among existing pooling operators or designing new ones. Based on our theoretical findings, we reviewed several existing pooling operators and identified those that fail to satisfy the expressiveness assumptions. Finally, we introduced an experimental setup to empirically measure the expressive power of a GNN equipped with pooling layers, in terms of its capability to perform a graph isomorphism test. * Equal contribution Preprint. Under review."}, "cited_paper_content": {"title": "A Fair Comparison Of Graph Neural Networks For Graph Classification", "abstract": "Experimental reproducibility and replicability are critical topics in machine learning. Authors have often raised concerns about their lack in scientific publications to improve the quality of the field. Recently, the graph representation learning field has attracted the attention of a wide research community, which resulted in a large stream of works. As such, several Graph Neural Network models have been developed to effectively tackle graph classification. However, experimental procedures often lack rigorousness and are hardly reproducible. Motivated by this, we provide an overview of common practices that should be avoided to fairly compare with the state of the art. To counter this troubling trend, we ran more than 47000 experiments in a controlled and uniform framework to re-evaluate five popular models across nine common benchmarks. Moreover, by comparing GNNs with structure-agnostic baselines we provide convincing evidence that, on some datasets, structural information has not been exploited yet. We believe that this work can contribute to the development of the graph learning field, by providing a much needed grounding for rigorous evaluations of graph classification models."}, "keywords": ["GNN architecture"], "citation_intent": "result"}
{"citing_id": "2304.14795v1", "cited_id": "2001.07685", "section_title": "A. Consistency-Based Regularization", "citation": "Inspired by FixMatch #REFR , a newly recent SOTA algorithm for semi-supervised image classification, we incorporate pseudo-labeling into consistencybased regularization and then can measure this similarity directly with cross-entropy.", "text_before_citation": ["They usually have different confidence levels even when predicted to be an identical class.", "Hence, we can combinep t andp t\u22121 using a coefficient \u03ba to form a final perturbed output:", "EQUATION", "How to Enforce Consistency Another important aspect of consistency-based regularization is to enforce consistency, which encourages a model to have close predictions on an unlabeled sample and its perturbed versions.", "Such expected similarity can be computed using some popular distance measures like mean squared error (MSE), Kullback-Leiber (KL) divergence, and Jensen-Shannon (JS) divergence, which are not dependent on any true label."], "text_after_citation": ["Pseudo-labeling is not competitive against other modern semi-supervised learning algorithms on its own, but using it as a part of these modern ones could usually produce better results.", "Given a signal sample x \u2208 U and its corresponding model prediction after being perturbed, i.e.,p, we can convert p into a hard label\u0177 (known as \"one-hot\" encoding), and it is retained only whenp enjoys high confidence:", "EQUATION", "where c \u2208 {1, . . . , C} represents a possible class.", "The predicted class probability ofp on c is \u2126 c (p), and \u03c4 is a preset threshold."], "citing_paper_content": {"title": "Semi-Supervised Rf Fingerprinting With Consistency-Based Regularization", "abstract": "As a promising non-password authentication technology, radio frequency (RF) fingerprinting can greatly improve wireless security. Recent work has shown that RF fingerprinting based on deep learning can significantly outperform conventional approaches. The superiority, however, is mainly attributed to supervised learning using a large amount of labeled data, and it significantly degrades if only limited labeled data is available, making many existing algorithms lack practicability. Considering that it is often easier to obtain enough unlabeled data in practice with minimal resources, we leverage deep semi-supervised learning for RF fingerprinting, which largely relies on a composite data augmentation scheme designed for radio signals, combined with two popular techniques: consistency-based regularization and pseudo-labeling. Experimental results on both simulated and realworld datasets demonstrate that our proposed method for semisupervised RF fingerprinting is far superior to other competing ones, and it can achieve remarkable performance almost close to that of fully supervised learning with a very limited number of examples."}, "cited_paper_content": {"title": "Fixmatch: Simplifying Semi-Supervised Learning With Consistency And Confidence", "abstract": "Semi-supervised learning (SSL) provides an effective means of leveraging unlabeled data to improve a model's performance. In this paper, we demonstrate the power of a simple combination of two common SSL methods: consistency regularization and pseudo-labeling. Our algorithm, FixMatch, first generates pseudo-labels using the model's predictions on weakly-augmented unlabeled images. For a given image, the pseudo-label is only retained if the model produces a high-confidence prediction. The model is then trained to predict the pseudo-label when fed a strongly-augmented version of the same image. Despite its simplicity, we show that FixMatch achieves state-of-the-art performance across a variety of standard semi-supervised learning benchmarks, including 94.93% accuracy on CIFAR-10 with 250 labels and 88.61% accuracy with 40 -- just 4 labels per class. Since FixMatch bears many similarities to existing SSL methods that achieve worse performance, we carry out an extensive ablation study to tease apart the experimental factors that are most important to FixMatch's success. We make our code available at this https URL."}, "keywords": ["semi-supervised image classification"], "citation_intent": "method"}
{"citing_id": "2303.07929v1", "cited_id": "1708.09687", "section_title": "Datasets And Metrics", "citation": "MegaAge-Asian #REFR is newly facial age dataset consisting of 40,000 Asian faces with ages ranging from 0 to 70 years old.", "text_before_citation": ["FG-Net #OTHEREFR contains 1002 facial images from 82 subjects, where the age ranges from 0 to 69.", "In experiment, we adopt the setup of paper #OTHEREFR , which uses leaveone person-out (LOPO) cross-validation.", "Hence, the average performance over 82 splits is reported, which makes the time of each training longer, and the MAEs fluctuate greatly in each split.", "IMDB-Wiki #OTHEREFR consists of 523051 facial images of celebrities, crawled from IMDB and Wikipedia, and age ranges from 0 to 100.", "Since there is much noise in this dataset, we selected about 300,000 images for training, where all non-face and severely occluded images were removed."], "text_after_citation": ["It helps to increase the diversity of human races and improve the generalization ability of the model.", "Followed by the setting of #OTHEREFR , 3,945 images were reserved for testing, and the remains are treated as the training set.", "This dataset applies the cumulative accuracy (CA) #OTHEREFR as the evaluation metric.", "Metrics.", "We take the Mean Absolute Error (MAE) in the experiments to calculate the discrepancy between the estimated age and the ground truth."], "citing_paper_content": {"title": "Daa: A Delta Age Adain Operation For Age Estimation Via Binary Code Transformer", "abstract": "Naked eye recognition of age is usually based on comparison with the age of others. However, this idea is ignored by computer tasks because it is difficult to obtain representative contrast images of each age. Inspired by the transfer learning, we designed the Delta Age AdaIN (DAA) operation to obtain the feature difference with each age, which obtains the style map of each age through the learned values representing the mean and standard deviation. We let the input of transfer learning as the binary code of age natural number to obtain continuous age feature information. The learned two groups of values in Binary code mapping are corresponding to the mean and standard deviation of the comparison ages. In summary, our method consists of four parts: FaceEncoder, DAA operation, Binary code mapping, and AgeDecoder modules. After getting the delta age via AgeDecoder, we take the average value of all comparison ages and delta ages as the predicted age. Compared with state-of-the-art methods, our method achieves better performance with fewer parameters on multiple facial age datasets."}, "cited_paper_content": {"title": "Quantifying Facial Age By Posterior Of Age Comparisons", "abstract": "We introduce a novel approach for annotating large quantity of in-the-wild facial images with high-quality posterior age distribution as labels. Each posterior provides a probability distribution of estimated ages for a face. Our approach is motivated by observations that it is easier to distinguish who is the older of two people than to determine the person's actual age. Given a reference database with samples of known ages and a dataset to label, we can transfer reliable annotations from the former to the latter via human-in-the-loop comparisons. We show an effective way to transform such comparisons to posterior via fully-connected and SoftMax layers, so as to permit end-to-end training in a deep network. Thanks to the efficient and effective annotation approach, we collect a new large-scale facial age dataset, dubbed `MegaAge', which consists of 41,941 images. Data can be downloaded from our project page mmlab.ie.cuhk.edu.hk/projects/MegaAge and github.com/zyx2012/Age_estimation_BMVC2017. With the dataset, we train a network that jointly performs ordinal hyperplane classification and posterior distribution learning. Our approach achieves state-of-the-art results on popular benchmarks such as MORPH2, Adience, and the newly proposed MegaAge."}, "keywords": ["newly facial age"], "citation_intent": "background"}
{"citing_id": "2303.13909v1", "cited_id": "1406.2661", "section_title": "Overview", "citation": "A GAN #REFR is a framework that trains a generator G and discriminator D in a two-player min-max game.", "text_before_citation": [], "text_after_citation": ["In the context of speech synthesis, a generator is identical to a speech synthesizer that synthesizes a speech waveform x from the input data s, where s indicates the intermediate representation (e.g., mel spectrogram) in a twostage approach and text in an end-to-end TTS system.", "In GAN-based speech synthesis, the generator and discriminator are trained with two GAN-related losses: adversarial and feature-matching losses."], "citing_paper_content": {"title": "Wave-U-Net Discriminator: Fast And Lightweight Discriminator For Generative Adversarial Network-Based Speech Synthesis", "abstract": "In speech synthesis, a generative adversarial network (GAN), training a generator (speech synthesizer) and a discriminator in a minmax game, is widely used to improve speech quality. An ensemble of discriminators is commonly used in recent neural vocoders (e.g., HiFi-GAN) and end-to-end text-to-speech (TTS) systems (e.g., VITS) to scrutinize waveforms from multiple perspectives. Such discriminators allow synthesized speech to adequately approach real speech; however, they require an increase in the model size and computation time according to the increase in the number of discriminators. Alternatively, this study proposes a Wave-U-Net discriminator, which is a single but expressive discriminator with Wave-U-Net architecture. This discriminator is unique; it can assess a waveform in a sample-wise manner with the same resolution as the input signal, while extracting multilevel features via an encoder and decoder with skip connections. This architecture provides a generator with sufficiently rich information for the synthesized speech to be closely matched to the real speech. During the experiments, the proposed ideas were applied to a representative neural vocoder (HiFi-GAN) and an end-to-end TTS system (VITS). The results demonstrate that the proposed models can achieve comparable speech quality with a 2.31 times faster and 14.5 times more lightweight discriminator when used in HiFi-GAN and a 1.90 times faster and 9.62 times more lightweight discriminator when used in VITS. 1"}, "cited_paper_content": {"title": "Generative Adversarial Nets", "abstract": "We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to \u00bd everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples."}, "keywords": ["GAN"], "citation_intent": "background"}
{"citing_id": "2303.12080v1", "cited_id": "1810.04805", "section_title": "Conclusion", "citation": "Finally, we also utilize an advanced model, BERT-base #REFR , to extract word representations by averaging the outputs of the last layer.", "text_before_citation": ["We adopt fastText #OTHEREFR as our default gloss feature extractor.", "Here we investigate other alternatives as shown in Table 15 .", "Word2vec #OTHEREFR and GloVe #OTHEREFR are widely-adopted in NLP community.", "They perform comparably to each other that GloVe achieves better results on the top-1 accuracy while word2vec is superior regarding to the top-5 accuracy.", "As an improvement of word2vec, fastText leads to better results on all of the four metrics."], "text_after_citation": ["However, it performs worse than all the other methods since it is not dedicated to word representation learning."], "citing_paper_content": {"title": "Natural Language-Assisted Sign Language Recognition", "abstract": "Sign languages are visual languages which convey information by signers' handshape, facial expression, body movement, and so forth. Due to the inherent restriction of combinations of these visual ingredients, there exist a significant number of visually indistinguishable signs (VISigns) in sign languages, which limits the recognition capacity of vision neural networks. To mitigate the problem, we propose the Natural Language-Assisted Sign Language Recognition (NLA-SLR) framework, which exploits semantic information contained in glosses (sign labels). First, for VISigns with similar semantic meanings, we propose language-aware label smoothing by generating soft labels for each training sign whose smoothing weights are computed from the normalized semantic similarities among the glosses to ease training. Second, for VISigns with distinct semantic meanings, we present an inter-modality mixup technique which blends vision and gloss features to further maximize the separability of different signs under the supervision of blended labels. Besides, we also introduce a novel backbone, video-keypoint network, which not only models both RGB videos and human body keypoints but also derives knowledge from sign videos of different temporal receptive fields. Empirically, our method achieves state-ofthe-art performance on three widely-adopted benchmarks: MSASL, WLASL, and NMFs-CSL. Codes are available at https://github.com/FangyunWei/SLRT."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["word representations", "BERT-base"], "citation_intent": "method"}
{"citing_id": "2303.07211v1", "cited_id": "0903.0544", "section_title": "Iii. A New Randomized Algorithm For Frameproof", "citation": "Our main tool is the celebrated algorithmic version of the Lov\u00e1sz Local Lemma for the symmetric case as given in #REFR .", "text_before_citation": ["We first need to recall the following well-known facts on binomial coefficients, where for positive integers c \u2264 b \u2264 a we have", "EQUATION", "EQUATION", "and the following useful inequality, proved in #OTHEREFR , for positive integers", "EQUATION"], "text_after_citation": ["Lemma III.3. [13] Let E 1 , E 2 , .", ". .", ", E m be events in a probability space, where each event E i is mutually independent of the set of all the other events E j except for at most D, and", "Pr(E i ) \u2264 P for all 1 \u2264 i \u2264 m.", "If eP D \u2264 1, then Pr(\u2229 m i=1 E i ) > 0."], "citing_paper_content": {"title": "Bounds And Algorithms For Frameproof Codes And Related Combinatorial Structures", "abstract": "In this paper, we study upper bounds on the minimum length of frameproof codes introduced by Boneh and Shaw [3] to protect copyrighted materials. A q-ary (k, n)-frameproof code of length t is a t\u00d7n matrix having entries in {0, 1,. .. , q\u22121} and with the property that for any column c and any other k columns, there exists a row where the symbols of the k columns are all different from the corresponding symbol (in the same row) of the column c. In this paper, we show the existence of q-ary (k, n)-frameproof codes of length t = O(k 2 q log n) for q \u2264 k, using the Lov\u00e1sz Local Lemma, and of length t = O(k log(q/k) log(n/k)) for q > k using the expurgation method. Remarkably, for the practical case of q \u2264 k our findings give codes whose length almost matches the lower bound \u2126(k 2 q log k log n) on the length of any q-ary (k, n)-frameproof code and, more importantly, allow us to derive an algorithm of complexity O(tn 2) for the construction of such codes."}, "cited_paper_content": {"title": "A Constructive Proof Of The General Lov\u00e1sz Local Lemma", "abstract": "The Lovasz Local Lemma [EL75] is a powerful tool to non-constructively prove the existence of combinatorial objects meeting a prescribed collection of criteria. In his breakthrough paper [Bec91], Beck demonstrated that a constructive variant can be given under certain more restrictive conditions. Simplifications of his procedure and relaxations of its restrictions were subsequently exhibited in several publications [Alo91, MR98, CS00, Mos06, Sri08, Mos08]. In [Mos09], a constructive proof was presented that works under negligible restrictions, formulated in terms of the Bounded Occurrence Satisfiability problem. In the present paper, we reformulate and improve upon these findings so as to directly apply to almost all known applications of the general Local Lemma."}, "keywords": ["celebrated algorithmic version", "Lov\u00e1sz Local Lemma"], "citation_intent": "background"}
{"citing_id": "2303.10449v1", "cited_id": "1512.03385", "section_title": "E. Detailed Results And More Architectures", "citation": "The comparison results on WideResNet-28 have the same trend as that on ResNet-18 #REFR architecture.", "text_before_citation": ["Tab. 7 and Tab. 9 show the detailed results among all datasets.", "Our method obtains consistently better results across all OOD detection metrics and all datasets.", "Compared with other methods using extra OOD training data (MCD [32], OE #OTHEREFR , UDG #OTHEREFR ), our method boosts the OOD detection performance meanwhile maximally maintaining the ID classification performance and achieves the best results on ACC.", "Following #OTHEREFR we also adopt another network architecture of WideResNet-28 #OTHEREFR to do experiments in Tab. 8 and Tab. 10 and compare the performance."], "text_after_citation": ["Our proposed method combining ET with L rep has advantages on almost all metrics, showing that our method enhances both the OOD detection and the ID classification ability.", "Notably, the previous state-of-the-art approaches generally performed well on SVHN and Texture datasets, but in Tiny-ImageNet, LSUN, and Places365 suffered a defeat.", "It can be explained that the images in the first two datasets have relatively flat backgrounds, which are quite different in style from those in CIFAR10/100, and the resulting covariate shifts make it easier for the model to identify them as OOD examples.", "Our proposed method especially achieves performance advantages on Tiny-ImageNet, LSUN, and Places365, but its performance on CIFAR-100/10 (the styles between them are very similar) of the CIFAR-10/100 benchmark is still not outstanding.", "This indicates that more exploration is needed to overcome the interference of covariate shifts in OOD detection tasks. Table 7 . Detailed results on CIFAR-10 benchmark using ResNet-18."], "citing_paper_content": {"title": "Uncertainty-Aware Optimal Transport For Semantically Coherent Out-Of-Distribution Detection", "abstract": "Semantically coherent out-of-distribution (SCOOD) detection aims to discern outliers from the intended data distribution with access to unlabeled extra set. The coexistence of in-distribution and out-of-distribution samples will exacerbate the model overfitting when no distinction is made. To address this problem, we propose a novel uncertainty-aware optimal transport scheme. Our scheme consists of an energy-based transport (ET) mechanism that estimates the fluctuating cost of uncertainty to promote the assignment of semantic-agnostic representation, and an inter-cluster extension strategy that enhances the discrimination of semantic property among different clusters by widening the corresponding margin distance. Furthermore, a T-energy score is presented to mitigate the magnitude gap between the parallel transport and classifier branches. Extensive experiments on two standard SCOOD benchmarks demonstrate the above-par OOD detection performance, outperforming the state-of-the-art methods by a margin of 27.69% and 34.4% on FPR@95, respectively. Code is available at https://github.com/LuFan31/ET-OOD. * Co-first Author. \u2020Corresponding Author. \u2021Work done during an internship at Ant Group."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["WideResNet-28", "ResNet-18 architecture"], "citation_intent": "result"}
{"citing_id": "2303.05682v1", "cited_id": "1804.04310", "section_title": "Related Work", "citation": "The motivation for this paper is the work in #REFR which proposes a matrix completion approach for the Euclidean distance geometry problem (abbreviated EDG hereafter).", "text_before_citation": [], "text_after_citation": ["The problem is concerned with finding the configuration of points given partial information in terms of pairwise inter-point distances; it is of fundamental importance in many applications #OTHEREFR .", "Consider a set of n points P = {p 1 , p 2 , ..., p n } T \u2208 R n\u00d7r .", "Let D \u2208 R n\u00d7n denote the squared distance matrix with entries D i,j , and let \u2126 \u2282 {(i, j) | i, j = 1, ..., n, i < j}, |\u2126| = m, be the set that consists of all the indices for which we know the corresponding entries of the distance matrix.", "Given k > 0 and partial exact distance information, the EDG problem is to find a configuration of points", "P = {p 1 , p 2 , ..., p n } T \u2208 R n\u00d7k such that ||p i \u2212 p j || 2 2 = D i,j for (i, j) \u2208 \u2126."], "citing_paper_content": {"title": "A Dual Basis Approach To Multidimensional Scaling: Spectral Analysis And Graph Regularity", "abstract": "Classical multidimensional scaling (CMDS) is a technique that aims to embed a set of objects in a Euclidean space given their pairwise Euclidean distance matrix. The main part of CMDS is based on double centering a squared distance matrix and employing a truncated eigendecomposition to recover the point coordinates. A central result in CMDS connects the squared Euclidean matrix to a Gram matrix derived from the set of points. In this paper, we study a dual basis approach to classical multidimensional scaling. We give an explicit formula for the dual basis and fully characterize the spectrum of an essential matrix in the dual basis framework. We make connections to a related problem in metric nearness."}, "cited_paper_content": {"title": "Exact Reconstruction Of Euclidean Distance Geometry Problem Using Low-Rank Matrix Completion", "abstract": "The Euclidean distance geometry problem arises in a wide variety of applications, from determining molecular conformations in computational chemistry to localization in sensor networks. When the distance information is incomplete, the problem can be formulated as a nuclear norm minimization problem. In this paper, this minimization program is recast as a matrix completion problem of a low-rank $r$ Gram matrix with respect to a suitable basis. The well-known restricted isometry property cannot be satisfied in this scenario. Instead, a dual basis approach is introduced to theoretically analyze the reconstruction problem. If the Gram matrix satisfies certain coherence conditions with parameter $\\nu $ , the main result shows that the underlying configuration of $n$ points can be recovered with very high probability from $O(nr\\nu \\log ^{2}(n))$ uniformly random samples. Computationally, simple and fast algorithms are designed to solve the Euclidean distance geometry problem. Numerical tests on different 3-D data and protein molecules validate the effectiveness and efficiency of the proposed algorithms."}, "keywords": ["matrix completion approach"], "citation_intent": "background"}
{"citing_id": "2303.16704v1", "cited_id": "1812.02274", "section_title": "Generating Differentially Private Synthetic Data", "citation": "In #REFR , the authors adopted a so-called variational autoencoder, DP-VAE, which assumes that the mapping from real data to the Gaussian distribution can be efficiently learned.", "text_before_citation": ["Although DP-based generative Artificial Neural Networks (ANNs) have been quite extensively researched in the major field of data science and machine learning, they have not been used in the context of process mining.", "Thus, we mainly focus on some of the work outside the domain of process mining."], "text_after_citation": ["A different direction was then chosen by #OTHEREFR , where the authors used a Wasserstein GAN (WGAN) to generate differentially private mixed-type synthetic outputs employing a Wasserstein-distance-based loss function.", "Finally, in #OTHEREFR , the concepts of WGAN and DP-VAE were combined to first learn a private data encoding and then generate respective encoded data.", "We adapted this principle for our work to cope with the large dimensionality of event data.", "Research in non-private generative models for process mining, primarily focuses on exploiting ANNs and GANs to predict the next state of processes such as #OTHEREFR , and #OTHEREFR .", "Note that the approach in #OTHEREFR only provides synthetic event data without any privacy guarantees."], "citing_paper_content": {"title": "Travag: Differentially Private Trace Variant Generation Using Gans", "abstract": "Process mining is rapidly growing in the industry. Consequently, privacy concerns regarding sensitive and private information included in event data, used by process mining algorithms, are becoming increasingly relevant. State-of-the-art research mainly focuses on providing privacy guarantees, e.g., differential privacy, for trace variants that are used by the main process mining techniques, e.g., process discovery. However, privacy preservation techniques for releasing trace variants still do not fulfill all the requirements of industry-scale usage. Moreover, providing privacy guarantees when there exists a high rate of infrequent trace variants is still a challenge. In this paper, we introduce TraVaG as a new approach for releasing differentially private trace variants based on Generative Adversarial Networks (GANs) that provides industry-scale benefits and enhances the level of privacy guarantees when there exists a high ratio of infrequent variants. Moreover, TraVaG overcomes shortcomings of conventional privacy preservation techniques such as bounding the length of variants and introducing fake variants. Experimental results on real-life event data show that our approach outperforms stateof-the-art techniques in terms of privacy guarantees, plain data utility preservation, and result utility preservation."}, "cited_paper_content": {"title": "Differentially Private Data Generative Models", "abstract": "Deep neural networks (DNNs) have recently been widely adopted in various applications, and such success is largely due to a combination of algorithmic breakthroughs, computation resource improvements, and access to a large amount of data. However, the large-scale data collections required for deep learning often contain sensitive information, therefore raising many privacy concerns. Prior research has shown several successful attacks in inferring sensitive training data information, such as model inversion, membership inference, and generative adversarial networks (GAN) based leakage attacks against collaborative deep learning. In this paper, to enable learning efficiency as well as to generate data with privacy guarantees and high utility, we propose a differentially private autoencoder-based generative model (DP-AuGM) and a differentially private variational autoencoder-based generative model (DP-VaeGM). We evaluate the robustness of two proposed models. We show that DP-AuGM can effectively defend against the model inversion, membership inference, and GAN-based attacks. We also show that DP-VaeGM is robust against the membership inference attack. We conjecture that the key to defend against the model inversion and GAN-based attacks is not due to differential privacy but the perturbation of training data. Finally, we demonstrate that both DP-AuGM and DP-VaeGM can be easily integrated with real-world machine learning applications, such as machine learning as a service and federated learning, which are otherwise threatened by the membership inference attack and the GAN-based attack, respectively."}, "keywords": ["so-called variational autoencoder"], "citation_intent": "method"}
{"citing_id": "2304.11633v1", "cited_id": "1706.04599", "section_title": "Calibration", "citation": "Our observations reveal that all the models exhibit high confidence levels in their predictions, this is consistent with previous research on large models #REFR .", "text_before_citation": ["In this subsection, we first investigate the level of confidence for both the correct and incorrect samples.", "Confidence is typically described in terms of a probability value, indicating the likelihood of belonging to a specific category.", "To obtain prediction probabilities from ChatGPT, we ask it to output the probability (Confidence_Standard and Confidence_Open), as discussed in \u00a7 3.3.", "Our aim is to investigate whether ChatGPT can provide a reasonable prediction confidence scores for its predictions, thus reducing the risk of misinterpretation.", "In Table 6 , we present the confidence scores of correct and incorrect predictions from different models, referred to correct confidence and incorrect confidence, respectively."], "text_after_citation": ["Although ChatGPT performs worse than its BERT-based counterparts in Standard-IE setting, it displays overconfidence in both correct and incorrect predictions. Consequently, this overconfidence could lead to misguidance of users.", "Furthermore, we note a significant confidence gap between correct and incorrect predictions, indicating the need for careful evaluation when ChatGPT's prediction has relatively low confidence.", "We then focus on calibration, a critical property of LLMs as it could estimate the predictive uncertainty for the secure application of LLMs.", "A well-calibrated model not only produces accurate predictions but also provides reliable and informative uncertainty estimates, necessary for sound decision-making.", "In this research, we evaluate the calibration using the Expected Calibration Error (ECE) metric which measures the deviation be- tween predicted confidence and accuracy."], "citing_paper_content": {"title": "Evaluating Chatgpt'S Information Extraction Capabilities: An Assessment Of Performance, Explainability, Calibration, And Faithfulness", "abstract": "The capability of Large Language Models (LLMs) like ChatGPT to comprehend user intent and provide reasonable responses has made them extremely popular lately. In this paper, we focus on assessing the overall ability of ChatGPT using 7 fine-grained information extraction (IE) tasks. Specially, we present the systematically analysis by measuring ChatGPT's performance, explainability, calibration, and faithfulness, and resulting in 15 keys from either the ChatGPT or domain experts. Our findings reveal that Chat-GPT's performance in Standard-IE setting is poor, but it surprisingly exhibits excellent performance in the OpenIE setting, as evidenced by human evaluation. In addition, our research indicates that ChatGPT provides high-quality and trustworthy explanations for its decisions. However, there is an issue of ChatGPT being overconfident in its predictions, which resulting in low calibration. Furthermore, ChatGPT demonstrates a high level of faithfulness to the original text in the majority of cases. We manually annotate and release the test sets of 7 finegrained IE tasks contains 14 datasets to further promote the research. The datasets and code are available at this url. 1"}, "cited_paper_content": {"title": "On Calibration Of Modern Neural Networks", "abstract": "Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions."}, "keywords": ["large models", "predictions"], "citation_intent": "result"}
{"citing_id": "2304.14082v2", "cited_id": "1911.11134", "section_title": "Image Classification", "citation": "Though most results match previous work, we observe a significant improvement for the accuracy achieved by the SET algorithm compared to the implementation done in #REFR .", "text_before_citation": ["Instead, we aim to provide some baseline results using different training recipes and architectures.", "For all experiments, we use the default hyper-parameters provided by the Scenic library.", "ResNet-50 is a popular architecture in sparsity literature #OTHEREFR .", "We train 80% sparse ResNet-50 models on ImageNet to reproduce previous results reported in the literature #OTHEREFR .", "We use uniform sparsity across layers and leave the first convolutional layer dense as recommended by #OTHEREFR ."], "text_after_citation": ["We use ERK sparsity distribution #OTHEREFR in our ViT experiments.", "Sparse vision transformers achieve better generalization even for the shorter, 90 epoch, training runs (ViT-B16).", "STE obtains the best results and exceeds the baseline performance by 0.2%.", "Interestingly, when we increase the number of training epochs to 300 (ViT-B16+), this gap widens and sparse ViT-B/16 trained with STE obtains 1.2% higher accuracy, despite having worse (higher) training loss.", "Dynamic sparse training methods (RigL and SET) performs poorly in shorter training runs, however with extended training, they achieve almost 5% higher accuracy and exceed the dense baseline."], "citing_paper_content": {"title": "A Concise Library For Sparsity Research", "abstract": "This paper introduces JaxPruner, an open-source JAX-based pruning and sparse training library for machine learning research. JaxPruner aims to accelerate research on sparse neural networks by providing concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. Algorithms implemented in JaxPruner use a common API and work seamlessly with the popular optimization library Optax, which, in turn, enables easy integration with existing JAX based libraries. We demonstrate this ease of integration by providing examples in four different codebases: Scenic, t5x, Dopamine and FedJAX and provide baseline experiments on popular benchmarks. Jaxpruner is hosted at github.com/google-research/jaxpruner."}, "cited_paper_content": {"title": "Rigging The Lottery: Making All Tickets Winners", "abstract": "Sparse neural networks have been shown to be more parameter and compute efficient compared to dense networks and in some cases are used to decrease wall clock inference times. There is a large body of work on training dense networks to yield sparse networks for inference. This limits the size of the largest trainable sparse model to that of the largest trainable dense model. In this paper we introduce a method to train sparse neural networks with a fixed parameter count and a fixed computational cost throughout training, without sacrificing accuracy relative to existing dense-to-sparse training methods. Our method updates the topology of the network during training by using parameter magnitudes and infrequent gradient calculations. We show that this approach requires fewer floating-point operations (FLOPs) to achieve a given level of accuracy compared to prior techniques. Importantly, by adjusting the topology it can start from any initialization - not just \"lucky\" ones. We demonstrate state-of-the-art sparse training results with ResNet-50, MobileNet v1 and MobileNet v2 on the ImageNet-2012 dataset, WideResNets on the CIFAR-10 dataset and RNNs on the WikiText-103 dataset. Finally, we provide some insights into why allowing the topology to change during the optimization can overcome local minima encountered when the topology remains static."}, "keywords": ["accuracy"], "citation_intent": "result"}
{"citing_id": "2303.17543v1", "cited_id": "1906.09494", "section_title": "8:", "citation": "An error occurs if the active user set fails at least one of the tests in #REFR or an incorrect set passes all the tests.", "text_before_citation": ["i.e., I 1 := A \u2229\u00c3 and I 0 :=\u00c3 \\ A.", "Then, we perform 2 k \u2212 1 threshold tests as below.", "EQUATION", "for predetermined thresholds \u03b7 s where s := |I 0 | \u2208 {1, 2, . . . , k}.", "We declare\u00c3 as the active user set if it is the only set passing all threshold tests."], "text_after_citation": ["As shown in #OTHEREFR , using a set of thresholds given by \u03b7 s = log", "\u03c1 k( k s )( \u2212k s )", ", we can upper bound P", "( ) e as P ( ) e \u2264 k s=1 k s P log p z | x(I 0 ), x(I 1 ) p (z | x(I 1 )) \u2264 log \u03c1 k k s \u2212k s + \u03c1.", "Since preamble design uses i.i.d Bernoulli random variables, using concentration inequalities for the k = \u0398(1) regime, we can achieve P"], "citing_paper_content": {"title": "Active User Identification In Fast Fading Massive Random Access Channels", "abstract": "Reliable and prompt identification of active users is critical for enabling random access in massive machine-tomachine type networks which typically operate within stringent access delay and energy constraints. In this paper, an energy efficient active user identification protocol is envisioned in which the active users simultaneously transmit On-Off Keying (OOK) modulated preambles whereas the base station uses non-coherent detection to avoid the channel estimation overheads. The minimum number of channel-uses required for active user identification in the asymptotic regime of total number of users when the number of active devices k scales as k = \u0398(1) is characterized along with an achievability scheme relying on the equivalence of activity detection to a group testing problem. A practical scheme for active user identification based on a belief propagation strategy is also proposed and its performance is compared against the theoretical bounds."}, "cited_paper_content": {"title": "Multi-Cell Sparse Activity Detection For Massive Random Access: Massive Mimo Versus Cooperative Mimo", "abstract": "This paper considers sparse device activity detection for cellular machine-type communications with non-orthogonal signatures using the approximate message passing algorithm. This paper compares two network architectures, massive multiple-input-multiple-output (MIMO) and cooperative MIMO, in terms of their effectiveness in overcoming inter-cell interference. In the massive MIMO architecture, each base station (BS) detects only the users from its own cell while treating inter-cell interference as noise. In the cooperative MIMO architecture, each BS detects the users from neighboring cells as well; the detection results are then forwarded in the form of a log-likelihood ratio (LLR) to a central unit where final decisions are made. This paper analytically characterizes the probabilities of false alarm and missed detection for both architectures. The numerical results validate the analytic characterization and show that as the number of antennas increases, a massive MIMO system effectively drives the detection error to zero, while as the cooperation size increases, the cooperative MIMO architecture mainly improves the cell-edge user performance. Moreover, this paper studies the effect of LLR quantization to account for the finite-capacity fronthaul. The numerical simulations of a practical scenario suggest that in specific case cooperating three BSs in a cooperative MIMO system achieves about the same cell-edge detection reliability as a non-cooperative massive MIMO system with four times the number of antennas per BS."}, "keywords": ["active user", "error"], "citation_intent": "background"}
{"citing_id": "2303.14517v1", "cited_id": "1908.10084", "section_title": "Related Works", "citation": "One of the BERT variants that is suitable for sentence embedding is Sentence BERT #REFR .", "text_before_citation": ["Many research focuses on generating high-resolution images and semantically correct with the given text description.", "But research on the text encoder and other languages is necessary.", "CNN-LSTM has become popular text embedding since Reed et al. 2016 use that method.", "Then #OTHEREFR", "2018 improve the text encoder with word-level attention to improve the generated image details. And then Tsue et al. 2020 use BERT as text encoder and word embedding."], "text_after_citation": ["This variant uses a siamese architecture network to produce text embedding that can then be compared with cosine similarity to find sentence with similar meanings."], "citing_paper_content": {"title": "Indonesian Text-To-Image Synthesis With Sentence-Bert And Fastgan", "abstract": "Currently, text-to-image synthesis uses text encoder and image generator architecture. Research on this topic is challenging. This is because of the domain gap between natural language and vision. Nowadays, most research on this topic only focuses on producing a photo-realistic image, but the other domain, in this case, is the language, which is less concentrated. A lot of the current research uses English as the input text. Besides, there are many languages around the world. Bahasa Indonesia, as the official language of Indonesia, is quite popular. This language has been taught in Philipines, Australia, and Japan. Translating or recreating a new dataset into another language with good quality will cost a lot. Research on this domain is necessary because we need to examine how the image generator performs in other languages besides generating photo-realistic images. To achieve this, we translate the CUB dataset into Bahasa using google translate and manually by humans. We use Sentence BERT as the text encoder and FastGAN as the image generator. FastGAN uses lots of skip excitation modules and auto-encoder to generate an image with resolution 512 \u00d7 512 \u00d7 3, which is twice as bigger as the current state-of-the-art model (Zhang, Xu, Li, Zhang, Wang, Huang and Metaxas, 2019). We also get 4.76 \u00b1 0.43 and 46.401 on Inception Score and Fr\u00e9chet inception distance, respectively, and comparable with the current English text-to-image generation models. The mean opinion score also gives as 3.22 out of 5, which means the generated image is acceptable by humans. Link to source code: https://github.com/share424/Indonesian-Text-to-Imagesynthesis-with-Sentence-BERT-and-FastGAN"}, "cited_paper_content": {"title": "Sentence-Bert: Sentence Embeddings Using Siamese Bert-Networks", "abstract": "BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019) has set a new state-of-the-art performance on sentence-pair regression tasks like semantic textual similarity (STS). However, it requires that both sentences are fed into the network, which causes a massive computational overhead: Finding the most similar pair in a collection of 10,000 sentences requires about 50 million inference computations (~65 hours) with BERT. The construction of BERT makes it unsuitable for semantic similarity search as well as for unsupervised tasks like clustering. ::: In this publication, we present Sentence-BERT (SBERT), a modification of the pretrained BERT network that use siamese and triplet network structures to derive semantically meaningful sentence embeddings that can be compared using cosine-similarity. This reduces the effort for finding the most similar pair from 65 hours with BERT / RoBERTa to about 5 seconds with SBERT, while maintaining the accuracy from BERT. ::: We evaluate SBERT and SRoBERTa on common STS tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embeddings methods."}, "keywords": ["Sentence BERT"], "citation_intent": "background"}
{"citing_id": "2303.13352v1", "cited_id": "2003.06743", "section_title": "A. Simulation Experiments", "citation": "The second baseline, Selective Simulation #REFR (SELSIM), is a search-based algorithm that interleaves a 'planning' phase and a 'tracking' phase.", "text_before_citation": ["We assume perfect knowledge of the initial workspace state and all object parameters.", "We set a planning timeout of 120 s for 100 randomly generated MAMO problems at each level.", "Our analysis includes two versions of our algorithm -M4M refers to Algorithm 1, and M4M refers to a version which only calls CBS once (after Line 5) and does not iterate between calling CBS and finding a valid push in simulation.", "Baselines: We compare the performance of M4M against three types of baselines for solving MAMO problems with non-prehensile interactions.", "The first are standard implementations of sampling-based algorithms KPIECE #OTHEREFR and RRT #OTHEREFR from OMPL #OTHEREFR that search the entire MAMO state space X by randomly sampling robot motions."], "text_after_citation": ["The former queries the physicsbased simulator for interactions with a set of 'relevant' movable objects identified so far.", "The latter executes the solution found by the planning phase in the presence of all objects in simulation and, if any interaction constraints are violated, it adds the 'relevant' object to the set.", "It only uses the simple motion primitives described in Section III.", "Our final baseline is the work from Dogar et al.", "#OTHEREFR (DOGAR) which introduced the idea of a negative goal region (NGR) we use in M4M."], "citing_paper_content": {"title": "Planning For Complex Non-Prehensile Manipulation Among Movable Objects By Interleaving Multi-Agent Pathfinding And Physics-Based Simulation", "abstract": "Real-world manipulation problems in heavy clutter require robots to reason about potential contacts with objects in the environment. We focus on pick-and-place style tasks to retrieve a target object from a shelf where some 'movable' objects must be rearranged in order to solve the task. In particular, our motivation is to allow the robot to reason over and consider non-prehensile rearrangement actions that lead to complex robot-object and object-object interactions where multiple objects might be moved by the robot simultaneously, and objects might tilt, lean on each other, or topple. To support this, we query a physics-based simulator to forward simulate these interaction dynamics which makes action evaluation during planning computationally very expensive. To make the planner tractable, we establish a connection between the domain of Manipulation Among Movable Objects and Multi-Agent Pathfinding that lets us decompose the problem into two phases our M4M algorithm iterates over. First we solve a multiagent planning problem that reasons about the configurations of movable objects but does not forward simulate a physics model. Next, an arm motion planning problem is solved that uses a physics-based simulator but does not search over possible configurations of movable objects. We run simulated and realworld experiments with the PR2 robot and compare against relevant baseline algorithms. Our results highlight that M4M generates complex 3D interactions, and solves at least twice as many problems as the baselines with competitive performance."}, "cited_paper_content": {"title": "Planning With Selective Physics-Based Simulation For Manipulation Among Movable Objects", "abstract": "Use of physics-based simulation as a planning model enables a planner to reason and generate plans that involve non-trivial interactions with the world. For example, grasping a milk container out of a cluttered refrigerator may involve moving a robot manipulator in between other objects, pushing away the ones that are movable and avoiding interactions with certain fragile containers. A physics-based simulator allows a planner to reason about the effects of interactions with these objects and to generate a plan that grasps the milk container successfully. The use of physics-based simulation for planning however is underutilized. One of the reasons for it being that physics-based simulations are typically way too slow for being used within a planning loop that typically requires tens of thousands of actions to be evaluated within a matter of a second or two. In this work, we develop a planning algorithm that tries to address this challenge. In particular, it builds on the observation that only a small number of actions actually need to be simulated using physics, and the remaining set of actions, such as moving an arm around obstacles, can be evaluated using a much simpler internal planning model, e.g., a simple collision-checking model. Motivated by this, we develop an algorithm called Planning with Selective Physics-based Simulation that automatically discovers what should be simulated with physics and what can utilize an internal planning model for pick-and-place tasks."}, "keywords": ["'planning' phase", "Selective Simulation"], "citation_intent": "method"}
{"citing_id": "2303.17505v1", "cited_id": "1904.02639", "section_title": "Learning The Distribution Of Codes", "citation": "The reconstructed errors on local anomalies are strengthened for anomaly detection, which is similar to #REFR .", "text_before_citation": ["With the local sensitive VQVAE, the image patches are reconstructed from a learned codebook of the normal data.", "The reconstruction thus tends to be close to a normal sample."], "text_after_citation": ["Nevertheless, this separate model performs poorly at some points: (1) anomalies on global positions rather than on local textures, and (2) reconstructed anomalies due to the strong generalization of the VQVAE.", "Therefore, we employ a full-attention transformer to model the global information of images and rectify abnormal patches based on the prior distribution, which is referred to as a global sensitive transformer as shown in Fig. 2 (right) .", "Specifically, By the learned VQVAE, an image X is represented with a corresponding discrete encoding sequence", "s = {z 1 q , z 2 q , z 3 q , .", ".., z n q }."], "citing_paper_content": {"title": "Unsupervised Anomaly Detection With Local-Sensitive Vqvae And Global-Sensitive Transformers", "abstract": "Unsupervised anomaly detection (UAD) has been widely implemented in industrial and medical applications, which reduces the cost of manual annotation and improves efficiency in disease diagnosis. Recently, deep auto-encoder with its variants has demonstrated its advantages in many UAD scenarios. Training on the normal data, these models are expected to locate anomalies by producing higher reconstruction error for the abnormal areas than the normal ones. However, this assumption does not always hold because of the uncontrollable generalization capability. To solve this problem, we present LSGS, a method that builds on Vector Quantised-Variational Autoencoder (VQVAE) with a novel aggregated codebook and transformers with global attention. In this work, the VQVAE focus on feature extraction and reconstruction of images, and the transformers fit the manifold and locate anomalies in the latent space. Then, leveraging the generated encoding sequences that conform to a normal distribution, we can reconstruct a more accurate image for locating the anomalies. Experiments on various datasets demonstrate the effectiveness of the proposed method."}, "cited_paper_content": {"title": "Memorizing Normality To Detect Anomaly: Memory-Augmented Deep Autoencoder For Unsupervised Anomaly Detection", "abstract": "Deep autoencoder has been extensively used for anomaly detection. Training on the normal data, the autoencoder is expected to produce higher reconstruction error for the abnormal inputs than the normal ones, which is adopted as a criterion for identifying anomalies. However, this assumption does not always hold in practice. It has been observed that sometimes the autoencoder \"generalizes\" so well that it can also reconstruct anomalies well, leading to the miss detection of anomalies. To mitigate this drawback for autoencoder based anomaly detector, we propose to augment the autoencoder with a memory module and develop an improved autoencoder called memory-augmented autoencoder, i.e. MemAE. Given an input, MemAE firstly obtains the encoding from the encoder and then uses it as a query to retrieve the most relevant memory items for reconstruction. At the training stage, the memory contents are updated and are encouraged to represent the prototypical elements of the normal data. At the test stage, the learned memory will be fixed, and the reconstruction is obtained from a few selected memory records of the normal data. The reconstruction will thus tend to be close to a normal sample. Thus the reconstructed errors on anomalies will be strengthened for anomaly detection. MemAE is free of assumptions on the data type and thus general to be applied to different tasks. Experiments on various datasets prove the excellent generalization and high effectiveness of the proposed MemAE."}, "keywords": ["anomaly detection"], "citation_intent": "result"}
{"citing_id": "2304.08864v1", "cited_id": "1502.07571", "section_title": "Conclusion", "citation": "Our results for the weighted LIKE and weighted Balanced LIKE mechanisms match those obtained by Aleksandrov et al. #REFR perfectly.", "text_before_citation": ["These mechanisms have been developed as a first attempt to expand and adapt the online allocation with equal entitlements problem to a more general weighted condition.", "Moreover, we have proven that the proposed algorithms adhere to fundamental principles such as Strategyproofness and Envy-freeness.", "The results of this study provide a starting point for future research in the area of online fair allocation with weighted entitlements.", "In particular, our future work will focus on improving the performance of the proposed mechanisms and extending them to more complex settings with additional constraints and preferences.", "Overall, the proposed mechanisms can be valuable tools for ensuring fairness in various online allocation scenarios, such as resource allocation, job assignments, and scheduling."], "text_after_citation": ["Despite the fact that the outcomes are the same, our project's proof methodology differs from that of Aleksandrov et al.", "The authors of the Aleksandrov et al study only offer proof sketches for the theorems pertaining to strategy-proofness and envy-freeness.", "All of the proofs we did for strategy-proofness and envy-freeness would also apply in their context because our work is an extension of their study.", "We hope that our project will make it easier for readers to grasp how the theorems in #OTHEREFR are proved."], "citing_paper_content": {"title": "Online Fair Division With Arbitrary Entitlements", "abstract": "The division of goods in the online realm poses opportunities and challenges. While innovative mechanisms can be developed, uncertainty about the future may hinder effective solutions. This project aims to explore fair distribution models for goods among agents with arbitrary entitlements, specifically addressing food charity challenges in the real world. Building upon prior work in [AAGW15], which focuses on equal entitlements, our project seeks to better understand the proofs of the theorems mentioned in that paper, which currently only provide proof sketches. Our approach employs different proof techniques from those presented in [AAGW15]."}, "cited_paper_content": {"title": "Online Fair Division: Analysing A Food Bank Problem", "abstract": "We study an online model of fair division designed to capture features of a real world charity problem. We consider two simple mechanisms for this model in which agents simply declare what items they like. We analyse axiomatic properties of these mechanisms such as strategy-proofness and envy freeness. Finally, we perform a competitive analysis and compute the price of anarchy."}, "keywords": ["Balanced LIKE mechanisms"], "citation_intent": "result"}
{"citing_id": "2303.17486v1", "cited_id": "1706.02216", "section_title": "Neighbor Information Aggregation", "citation": "Therefore, in order to save computational costs while retaining important information about the domain, we use a simple and efficient aggregator, the average aggregator in GraphSAGE #REFR , for message aggregation.", "text_before_citation": ["Previous works tend to use attention mechanisms #OTHEREFR or design weighting parameters #OTHEREFR for message aggregation.", "This is the current work done by many popular GNN models who improve on the message passing and aggregation mechanisms to enhance GNN performance.", "However, assuming that we have selected the most similar neighbors for each node, the attention coefficient or weighting parameter has no obvious meaning in neighbor information aggregation.", "On the contrary, it will increase the computational cost significantly."], "text_after_citation": ["Compared with the attention mechanism, this method significantly reduces the model complexity. The formal description is as follows:", "h (l) v = \u03c3 W \u2022 AGG h (l\u22121) v \u222a h (l\u22121) v , \u2200v \u2208 N (v) (12)", "where \u03c3 is the activation function, W is the weight to be learned, and N (v) is the neighborhood of node v.", "In implementation , we use M EAN () as aggregation function AGG().", "M EAN () denotes the operation of averaging the embeddings of the nodes in the set."], "citing_paper_content": {"title": "Cost Sensitive Gnn-Based Imbalanced Learning For Mobile Social Network Fraud Detection", "abstract": "With the rapid development of mobile networks, the people's social contacts have been considerably facilitated. However, the rise of mobile social network fraud upon those networks, has caused a great deal of distress, in case of depleting personal and social wealth, then potentially doing significant economic harm. To detect fraudulent users, call detail record (CDR) data, which portrays the social behavior of users in mobile networks, has been widely utilized. But the imbalance problem in the aforementioned data, which could severely hinder the effectiveness of fraud detectors based on graph neural networks(GNN), has hardly been addressed in previous work. In this paper, we are going to present a novel Cost-Sensitive Graph Neural Network (CS-GNN) by creatively combining cost-sensitive learning and graph neural networks. We conduct extensive experiments on two open-source realworld mobile network fraud datasets. The results show that CSGNN can effectively solve the graph imbalance problem and then achieve better detection performance than the state-of-the-art algorithms. We believe that our research can be applied to solve the graph imbalance problems in other fields. The CSGNN code and datasets are publicly available at https://github.com/xxhu94/CSGNN."}, "cited_paper_content": {"title": "Inductive Representation Learning On Large Graphs", "abstract": "Low-dimensional embeddings of nodes in large graphs have proved extremely useful in a variety of prediction tasks, from content recommendation to identifying protein functions. However, most existing approaches require that all nodes in the graph are present during training of the embeddings; these previous approaches are inherently transductive and do not naturally generalize to unseen nodes. Here we present GraphSAGE, a general, inductive framework that leverages node feature information (e.g., text attributes) to efficiently generate node embeddings for previously unseen data. Instead of training individual embeddings for each node, we learn a function that generates embeddings by sampling and aggregating features from a node's local neighborhood. Our algorithm outperforms strong baselines on three inductive node-classification benchmarks: we classify the category of unseen nodes in evolving information graphs based on citation and Reddit post data, and we show that our algorithm generalizes to completely unseen graphs using a multi-graph dataset of protein-protein interactions."}, "keywords": ["message aggregation", "GraphSAGE"], "citation_intent": "method"}
{"citing_id": "2304.07883v1", "cited_id": "1703.06907", "section_title": "Synthetic Data In Deep Learning", "citation": "Domain randomization is a technique used to enhance the variability of synthetic data and has been shown to substantially increase performance in the real world #REFR .", "text_before_citation": ["A recent survey categorized hundreds of synthetic datasets and the use cases they have been devised for #OTHEREFR .", "Initially used to address low-level computer vision tasks such as optical flow #OTHEREFR , synthetic datasets are increasingly used to generate training datasets for high-level tasks such as, e.g., object recognition and detection #OTHEREFR , pose estimation #OTHEREFR , segmentation #OTHEREFR , human action recognition #OTHEREFR and pedestrian tracking and ReID #OTHEREFR .", "Works in this field typically build onto well-known repositories, including millions of virtual models with known categories or properties, which can be programmatically manipulated to automate both data generation and its labelling #OTHEREFR .", "Popular approaches for collecting synthetic data also include the use of video games #OTHEREFR , or fusing real and virtual data via compositing techniques and placing, e.g., virtual models onto real background images #OTHEREFR .", "One of the main challenges associated with synthetic data is the domain shift between real and synthetic images, which can be tackled through transfer learning or domain adaptation #OTHEREFR ."], "text_after_citation": ["With ever increasing CGI fidelity, the synthetic-to-real domain gap is progressively reducing.", "Recent exciting results showed that training DNNs on very large and diverse synthetic datasets can outperform using public real datasets on tasks such as pedestrian tracking and ReID, even without fine-tuning on real data #OTHEREFR ."], "citing_paper_content": {"title": "Bent & Broken Bicycles: Leveraging Synthetic Data For Damaged Object Re-Identification", "abstract": "Instance-level object re-identification is a fundamental computer vision task, with applications from image retrieval to intelligent monitoring and fraud detection. In this work, we propose the novel task of damaged object re-identification, which aims at distinguishing changes in visual appearance due to deformations or missing parts from subtle intra-class variations. To explore this task, we leverage the power of computer-generated imagery to create, in a semi-automatic fashion, high-quality synthetic images of the same bike before and after a damage occurs. The resulting dataset, Bent & Broken Bicycles (BB-Bicycles), contains 39,200 images and 2,800 unique bike instances spanning 20 different bike models. As a baseline for this task, we propose TransReI3D, a multi-task, transformer-based deep network unifying damage detection (framed as a multi-label classification task) with object re-identification. The BBBicycles dataset is available at https://tinyurl.com/37tepf7m keywords instance-level retrieval; re-identification; synthetic data; damage detection; transformers"}, "cited_paper_content": {"title": "Domain Randomization For Transferring Deep Neural Networks From Simulation To The Real World", "abstract": "Bridging the \u2018reality gap\u2019 that separates simulated robotics from experiments on hardware could accelerate robotic research through improved data availability. This paper explores domain randomization, a simple technique for training models on simulated images that transfer to real images by randomizing rendering in the simulator. With enough variability in the simulator, the real world may appear to the model as just another variation. We focus on the task of object localization, which is a stepping stone to general robotic manipulation skills. We find that it is possible to train a real-world object detector that is accurate to 1.5 cm and robust to distractors and partial occlusions using only data from a simulator with non-realistic random textures. To demonstrate the capabilities of our detectors, we show they can be used to perform grasping in a cluttered environment. To our knowledge, this is the first successful transfer of a deep neural network trained only on simulated RGB images (without pre-training on real images) to the real world for the purpose of robotic control."}, "keywords": ["synthetic data", "Domain randomization"], "citation_intent": "method"}
{"citing_id": "2303.00320v2", "cited_id": "1810.04805", "section_title": "Masking Strategies", "citation": "To achieve the goal of learning contextual representations of time series with a bidirectional encoding scheme, we decide to construct the corrupt input with masking strategies by following #REFR .", "text_before_citation": [], "text_after_citation": ["Suppose that S v /S m is the number of masked/visible positions.", "In this way, more comprehensive contextual representations of each position, including previous and future contexts, can be encoded, simultaneously.", "For clarity, we use Z v to denote the embeddings of visible (unmasked) positions while letting Z m be the embeddings of masked ones.", "Although some heuristic masking strategies, e.g., block masking #OTHEREFR , have been developed before, we adopt random masking strategies to form the corrupted inputs.", "It means that each sub-series unit has the same probability of being masked in constructing selfsupervised signals."], "citing_paper_content": {"title": "Timemae: Self-Supervised Representations Of Time Series With Decoupled Masked Autoencoders", "abstract": "Enhancing the expressive capacity of deep learning-based time series models with self-supervised pre-training has become ever-increasingly prevalent in time series classification. Even though numerous efforts have been devoted to developing self-supervised models for time series data, we argue that the current methods are not sufficient to learn optimal time series representations due to solely unidirectional encoding over sparse point-wise input units. In this work, we propose TimeMAE, a novel self-supervised paradigm for learning transferrable time series representations based on transformer networks. The distinct characteristics of the TimeMAE lie in processing each time series into a sequence of non-overlapping sub-series via window-slicing partitioning, followed by random masking strategies over the semantic units of localized sub-series. Such a simple yet effective setting can help us achieve the goal of killing three birds with one stone, i.e., (1) learning enriched contextual representations of time series with a bidirectional encoding scheme; (2) increasing the information density of basic semantic units; (3) efficiently encoding representations of time series using transformer networks. Nevertheless, it is a non-trivial to perform reconstructing task over such a novel formulated modeling paradigm. To solve the discrepancy issue incurred by newly injected masked embeddings, we design a decoupled autoencoder architecture, which learns the representations of visible (unmasked) positions and masked ones with two different encoder modules, respectively. Furthermore, we construct two types of informative targets to accomplish the corresponding pretext tasks. One is to create a tokenizer module that assigns a codeword to each masked region, allowing the masked codeword classification (MCC) task to be completed effectively. Another one is to adopt a siamese network structure to generate target representations for each masked input unit, aiming at performing the masked representation regression (MRR) optimization. Comprehensively pre-trained, our model can efficiently learn transferrable time series representations, thus benefiting the classification of time series. We extensively perform experiments on five benchmark datasets to verify the effectiveness of the TimeMAE. Experimental results show that the TimeMAE can significantly surpass previous competitive baselines. Furthermore, we also demonstrate the universality of the learned representations by performing transfer learning experiments. For the reproducibility of our results, we make our experiment codes publicly available to facilitate the self-supervised representations of time series in https://github.com/Mingyue-Cheng/TimeMAE."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["time series", "contextual representations"], "citation_intent": "method"}
{"citing_id": "2303.08319v1", "cited_id": "2003.12063", "section_title": "Related Works", "citation": "MEGA #REFR takes both temporal information and semantic similarity into account and aggregates local and global features jointly.", "text_before_citation": ["For example, Seq-NMS #OTHEREFR associates the bounding boxes from different frames with the IoU threshold; T-CNN #OTHEREFR links the object detection results from each frame according to optical flows.", "These methods are not trained end-to-end, and the performances are always sub-optimal.", "Feature-aggregation based models enhance the feature representations of the current frame by fusing those from the neighboring frames.", "FGFA #OTHEREFR and MANet #OTHEREFR weight average the features from the neighboring frames after being warped based on optical flows.", "SELSA #OTHEREFR , and Temporal ROI Align #OTHEREFR fuse the neighboring frames according to their semantic similarity."], "text_after_citation": ["Different from the methods mentioned above, which generate weights for aggregation using cosine similarity, TF-Blender #OTHEREFR applies a learnable network to predict the weights.", "All the methods mentioned above are designed for the classic object detectors like Faster R-CNN #OTHEREFR , or Center-Net #OTHEREFR .", "Recently, with DETR series #OTHEREFR introduced for the image domain tasks, researchers have begun to focus on how to use these Transformer-based models for video object detection.", "Among them, TransVOD series #OTHEREFR introduces two key modules to boost the performance: Temporal Query Encoder to fuse object queries and Temporal Deformable Transformer Decoder (TDTD) to obtain current frame detection results.", "Similarly, PT-SEFormer #OTHEREFR proposes the Spatial Transition Awareness Model to fuse the temporal and spatial information for a better prediction."], "citing_paper_content": {"title": "Faq: Feature Aggregated Queries For Transformer-Based Video Object Detectors", "abstract": "Video object detection needs to solve feature degradation situations that rarely happen in the image domain. One solution is to use the temporal information and fuse the features from the neighboring frames. With Transformerbased object detectors getting a better performance on the image domain tasks, recent works began to extend those methods to video object detection. However, those existing Transformer-based video object detectors still follow the same pipeline as those used for classical object detectors, like enhancing the object feature representations by aggregation. In this work, we take a different perspective on video object detection. In detail, we improve the qualities of queries for the Transformer-based models by aggregation. To achieve this goal, we first propose a vanilla query aggregation module that weighted averages the queries according to the features of the neighboring frames. Then, we extend the vanilla module to a more practical version, which generates and aggregates queries according to the features of the input frames. Extensive experimental results validate the effectiveness of our proposed methods: On the challenging ImageNet VID benchmark, when integrated with our proposed modules, the current state-of-the-art Transformerbased object detectors can be improved by more than 2.4% on mAP and 4.2% on AP 50 ."}, "cited_paper_content": {"title": "Memory Enhanced Global-Local Aggregation For Video Object Detection", "abstract": "How do humans recognize an object in a piece of video? Due to the deteriorated quality of single frame, it may be hard for people to identify an occluded object in this frame by just utilizing information within one image. We argue that there are two important cues for humans to recognize objects in videos: the global semantic information and the local localization information. Recently, plenty of methods adopt the self-attention mechanisms to enhance the features in key frame with either global semantic information or local localization information. In this paper we introduce memory enhanced global-local aggregation (MEGA) network, which is among the first trials that takes full consideration of both global and local information. Furthermore, empowered by a novel and carefully-designed Long Range Memory (LRM) module, our proposed MEGA could enable the key frame to get access to much more content than any previous methods. Enhanced by these two sources of information, our method achieves state-of-the-art performance on ImageNet VID dataset. Code is available at \\url{https://github.com/Scalsol/mega.pytorch}."}, "keywords": ["global features"], "citation_intent": "background"}
{"citing_id": "2304.10819v2", "cited_id": "1502.03167", "section_title": "38/49", "citation": "Utility also includes an MLP classifier consisting in a 2-layer network with a hidden layer using Batch-Normalization #REFR and ReLU non-linearity.", "text_before_citation": ["We also considered non-linear feature selection methods (e.g., #OTHEREFR ), but settled for this simpler one.", "\u2022 Nearest Neighbor classification. We also implemented Nearest Neighbor Classification as a classifier.", "In practice, a test feature vector is assigned a labeled corresponding to the label of the closest training sample (in L2-norm).", "For scalability, we implemented nearest neighbor retrieval using the approximate nearest neighbor library FAISS 97 .", "\u2022 MLP classification."], "text_after_citation": ["The network is optimized using the Adam optimizer #OTHEREFR with learning rate generally set to 3.0 \u00d7 10 \u22124 unless otherwise stated, and use early-stopping with patience of 3 on F1-score computed on the hold-out set.", "\u2022 MLP classification / Adversarial debiasing.", "Our MLP classifier optionally includes the option of using adversarial debiasing #OTHEREFR as a method to mitigate bias.", "This method essentially corresponds to introducing an adversary network that that is trained to infer the protected labels of the value of the protected variables of the input samples.", "In our implementation the adversary has the same architecture as the original network, but a hidden size that is 4 times smaller, and is optimized by alternating its update with the update of the original network."], "citing_paper_content": {"title": "Auditing And Generating Synthetic Data With Controllable Trust Trade-Offs", "abstract": "Data collected from the real world tends to be biased, unbalanced, and at risk of exposing sensitive and private information. This reality has given rise to the idea of creating synthetic datasets to alleviate risk, bias, harm, and privacy concerns inherent in the real data. This concept relies on Generative AI models to produce unbiased, privacy-preserving synthetic data while being true to the real data. In this new paradigm, synthetic data creators, users, and auditors must answer critical questions, such as \"How can we tell if this approach delivers on its promises?\" and \"How can we determine whether a synthetic dataset and a model built from it conform to trust requirements, such as fairness, privacy, safety, and fidelity while still being valuable for downstream tasks of interest?\". We present an auditing framework that offers a holistic assessment of synthetic datasets and AI models trained on them, centered around bias and discrimination prevention, fidelity to the real data, utility, robustness, and privacy preservation. We showcase our framework by auditing multiple generative models on diverse use cases, including education, healthcare, banking, human resources, and across different modalities, from tabular, to time-series, to natural language. Our use cases demonstrate the importance of a holistic assessment in order to ensure compliance with socio-technical safeguards that regulators and policymakers are increasingly enforcing. For this purpose, we introduce the trust index that ranks multiple synthetic datasets based on their prescribed safeguards and their desired trade-offs. Moreover, we devise a trust-index-driven model selection and cross-validation procedure via auditing in the training loop that we showcase on a class of transformer models that we dub TrustFormers, across different modalities. This trust-driven model selection allows for controllable trust trade-offs in the resulting synthetic data. We instrument our auditing framework with workflows that connect different stakeholders (e.g., data scientists, data governance experts, internal reviewers, external certifiers, and regulators) from model development to audit and certification via a synthetic data auditing report. We believe this mode of transparent reporting should become a standard to prevent bias, discrimination, and privacy breaches and ensure compliance with policy requirements, accountability, safety, and performance guarantees. Main Generative models have demonstrated impressive results in synthesizing high-quality data across multiple modalities from text 1, 2 , to images 3-6 , to chemistry 7 , to tabular and time-series data 8-10. We are entering a new era in training AI models, where synthetic data can be used to augment real data 11 , or as a complete replacement, in the most extreme case 12. One of the main motivations behind controllable synthetic data usage in training AI models is its promise to synthesize privacy-preserving data that enables safe sharing without putting the privacy of real users and individuals at risk, potentially circumventing cumbersome processes that are at the heart of many highly regulated fields such as financial services 13 and healthcare, for example 11, 14-16. Another motivation comes from controlling the generation process in order to balance the training data and reduce biases against protected groups and sensitive communities 17. Finally, synthetic data also offers new opportunities in simulating non-existing scenarios, providing grounding for causal inference via the generation of counterfactuals that would help explain some observations in the absence of real data 18. Synthetic data can be seedless 19, 20 , in the sense that it is not trained on real data but relies on a knowledge base or graph and on a rule-based generation process. While this seems ideal regarding privacy and fairness as this process does not rely on any real data, the underlying knowledge grounding can be biased and may also be at risk of linkage attacks that would leak the identities of individuals present in the real data. On the other side of the spectrum, we have synthetic data coming from inference from generative AI models trained on real data. This approach to synthetic data raises its own set of risks and pitfalls, as it often leads to memorization, privacy breaches, and data leakage 21. Another concern is the potential amplification of bias and disparate"}, "cited_paper_content": {"title": "Batch Normalization: Accelerating Deep Network Training By Reducing Internal Covariate Shift", "abstract": "Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization, and in some cases eliminates the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.82% top-5 test error, exceeding the accuracy of human raters."}, "keywords": ["Batch-Normalization"], "citation_intent": "method"}
{"citing_id": "2303.14922v1", "cited_id": "1409.1556", "section_title": "A.1. Vgg-16 Results On Cifar-10", "citation": "The white-box robustness of VGG-16 #REFR models trained using AT, ALP, TRADES, and CAT are reported in Tab. A1.", "text_before_citation": [], "text_after_citation": ["The setting for VGG-16 is the same as ResNet-18 models, i.e., \u03b1 = 1.0/20 and \u03b2 = 1.0/20.", "The improvement for CAT with VGG-16 models is as consistent with ResNet-18 models.", "CAT can boost model's robustness under AutoAttack with 2.0 points."], "citing_paper_content": {"title": "Cat: Collaborative Adversarial Training", "abstract": "Adversarial training can improve the robustness of neural networks. Previous methods focus on a single adversarial training strategy and do not consider the model property trained by different strategies. By revisiting the previous methods, we find different adversarial training methods have distinct robustness for sample instances. For example, a sample instance can be correctly classified by a model trained using standard adversarial training (AT) but not by a model trained using TRADES, and vice versa. Based on this observation, we propose a collaborative adversarial training framework to improve the robustness of neural networks. Specifically, we use different adversarial training methods to train robust models and let models interact with their knowledge during the training process. Collaborative Adversarial Training (CAT) can improve both robustness and accuracy. Extensive experiments on various networks and datasets validate the effectiveness of our method. CAT achieves state-of-the-art adversarial robustness without using any additional data on CIFAR-10 under the Auto-Attack benchmark. Code is available at https://github.com/liuxingbin/CAT."}, "cited_paper_content": {"title": "Very Deep Convolutional Networks For Large-Scale Image Recognition", "abstract": "In this work we investigate the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. Our main contribution is a thorough evaluation of networks of increasing depth using an architecture with very small (3x3) convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers. These findings were the basis of our ImageNet Challenge 2014 submission, where our team secured the first and the second places in the localisation and classification tracks respectively. We also show that our representations generalise well to other datasets, where they achieve state-of-the-art results. We have made our two best-performing ConvNet models publicly available to facilitate further research on the use of deep visual representations in computer vision."}, "keywords": ["white-box robustness", "VGG-16 models"], "citation_intent": "method"}
{"citing_id": "2304.01661v1", "cited_id": "1711.10269", "section_title": "B. State-Of-The-Art", "citation": "In #REFR , the authors jointly minimize the total transmit power and the circuit power consumption for MRT and ZF precoders.", "text_before_citation": ["The reason is that #OTHEREFR , #OTHEREFR consider the minimization of (2) from the precoding problem definition, not assuming any particular precoding scheme.", "When other BS components are included in the power consumption model, the optimal number of BS antennas to be utilized decreases.", "Similarly to #OTHEREFR , several studies have investigated the whole BS power consumption for energy-efficient mMIMO communications.", "A power consumption model for a mMIMO BS is developed in #OTHEREFR , where the total consumed power is the sum of the contributions from the PAs and from the BS circuits, plus a static consumption term.", "The circuit power consumption considers the radio frequency transceivers, the baseband processing, the backhauling, the cooling, and the power supply, and typically scales linearly with M #OTHEREFR ."], "text_after_citation": ["Differently from #OTHEREFR , they consider different traffic situations and they show that the number of activated antennas increases as a function of the traffic."], "citing_paper_content": {"title": "Energy-Saving Precoder Design For Narrowband And Wideband Massive Mimo", "abstract": "In this work, we study massive multiple-input multiple-output (MIMO) precoders optimizing power consumption while achieving the users' rate requirements. We first characterize analytically the solutions for narrowband and wideband systems minimizing the power amplifiers (PAs) consumption in low system load, where the per-antenna power constraints are not binding. After, we focus on the asymptotic wideband regime. The power consumed by the whole base station (BS) and the high-load scenario are then also investigated. We obtain simple solutions, and the optimal strategy in the asymptotic case reduces to finding the optimal number of active antennas, relying on known precoders among the active antennas. Numerical results show that large savings in power consumption are achievable in the narrowband system by employing antenna selection, while all antennas need to be activated in the wideband system when considering only the PAs consumption, and this implies lower savings. When considering the overall BS power consumption and a large number of subcarriers, we show that significant savings are achievable in the low-load regime by using a subset of the BS antennas. While optimization based on transmit power pushes to activate all antennas, optimization based on consumed power activates a number of antennas proportional to the load."}, "cited_paper_content": {"title": "Joint Transmit And Circuit Power Minimization In Massive Mimo With Downlink Sinr Constraints: When To Turn On Massive Mimo?", "abstract": "In this paper, we consider the downlink of a multi-cell multiple-input multiple-output system and find the jointly optimal number of base station (BS) antennas and transmission powers that minimize the power consumption while satisfying each user\u2019s effective signal-to-interference-and-noise-ratio constraint and the BSs\u2019 power constraints. Different from prior work, we consider a power consumption model that takes both transmitted and hardware-consumed power into account. We formulate the joint optimization problem for both single-cell and multi-cell systems. The closed-form expressions for the optimal number of BS antennas and transmission powers are derived for the single-cell case. The analysis for the multi-cell case reveals that increasing the number of BS antennas in any cell always improves the performance of the overall system in terms of both feasibility and total radiated power. A key contribution of this paper is to show that the joint optimization problem can be relaxed as a geometric programming problem that can be solved efficiently. The solution can be utilized in practice to turn on and off antennas depending on the traffic load variations. The substantial power savings are demonstrated by simulation."}, "keywords": ["total transmit power"], "citation_intent": "background"}
{"citing_id": "2303.16342v1", "cited_id": "1904.07750", "section_title": "Quantitative Results", "citation": "AudioSet has been documented by prior work #REFR to be much noisier than the videos in the other two datasets, which explains the lower performance across all approaches.", "text_before_citation": ["Similar to the reported performance on the MUSIC dataset, we also observe that VAST is comparable to the strongly supervised AVSGS approach on the SOLOS dataset.", "One possible reason behind the lower performance of AVSGS on videos with single instrument is that it is less critical to reason about context for videos with a single sounding object.", "Our better performance as compared to Co-Separation suggests that learning to perform audio separation at the coarse video region level under an MIL formulation can serve as an effective replacement for training object detectors.", "These results also indicate the great promise of latent captions in replacing ground-truth object annotations for supervision. Audio-visual source separation on AudioSet.", "Finally, we report the results of our evaluation on the AudioSet dataset in Table 3 ."], "text_after_citation": ["Unlike the SOP model, we observe that our training approach is more robust to noise in the dataset where the sounding object may not always be visible.", "L"], "citing_paper_content": {"title": "Language-Guided Audio-Visual Source Separation Via Trimodal Consistency", "abstract": "We propose a self-supervised approach for learning to perform audio source separation in videos based on natural language queries, using only unlabeled video and audio pairs as training data. A key challenge in this task is learning to associate the linguistic description of a soundemitting object to its visual features and the corresponding components of the audio waveform, all without access to annotations during training. To overcome this challenge, we adapt off-the-shelf vision-language foundation models to provide pseudo-target supervision via two novel loss functions and encourage a stronger alignment between the audio, visual and natural language modalities. During inference, our approach can separate sounds given text, video and audio input, or given text and audio input alone. We demonstrate the effectiveness of our self-supervised approach on three audiovisual separation datasets, including MUSIC, SOLOS and AudioSet, where we outperform state-of-the-art strongly supervised approaches despite not using object detectors or text labels during training. Our project page including publicly available code can be found at https://cs-people.bu.edu/rxtan/projects/VAST."}, "cited_paper_content": {"title": "Co-Separating Sounds Of Visual Objects", "abstract": "Learning how objects sound from video is challenging, since they often heavily overlap in a single audio channel. Current methods for visually-guided audio source separation sidestep the issue by training with artificially mixed video clips, but this puts unwieldy restrictions on training data collection and may even prevent learning the properties of \"true\" mixed sounds. We introduce a co-separation training paradigm that permits learning object-level sounds from unlabeled multi-source videos. Our novel training objective requires that the deep neural network's separated audio for similar-looking objects be consistently identifiable, while simultaneously reproducing accurate video-level audio tracks for each source training pair. Our approach disentangles sounds in realistic test videos, even in cases where an object was not observed individually during training. We obtain state-of-the-art results on visually-guided audio source separation and audio denoising for the MUSIC, AudioSet, and AV-Bench datasets."}, "keywords": ["AudioSet"], "citation_intent": "result"}
{"citing_id": "2303.01484v1", "cited_id": "1512.03385", "section_title": "B. Predicting Motion Plans From Images", "citation": "Function f is realized through a CNN with an ImageNet pretrained ResNet-34 backbone #REFR .", "text_before_citation": [". .] without collisions.", "We provide details about the initialization set \u0398, function f , training data, and loss function to train f .", "Initialization set \u0398 comes from the Cartesian product of a set of robot base positions in R 3 and a set of 10 arm configurations.", "We use 704 base positions (sampled in a uniform 1 m \u00d7 1.5 m 2D grid of base positions at a 10 cm resolution at 4 different heights) and 10 arm configurations, resulting in \u0398 having 7040 elements.", "The 10 arm configurations are selected in a data-driven manner, we sample 20 random configurations that satisfy the joint limits, and then select the 10 which when used with SeqIK give us the most successes across the training set."], "text_after_citation": ["We add 2 fully connected layers on the conv5 output to produce a 7040 dimensional representation.", "This is reshaped into an 80-dimensional spatial output of size 11 \u00d7 16.", "This is processed through another 3 convolutional layers to produce a (10 \u2022 4) \u00d7 11 \u00d7 16 logits tensor containing 11 \u00d7 16 spatial output for each of the 10 arm configurations at each of the 4 heights.", "Training labels are generated by decoding each candidate \u03b8 0 into motion plans using SeqIK, and testing them for endeffector pose deviation, self-collision, collision with the static environment, and collision with the articulating object in our simulator from Section III.", "Note that while testing the decoded motion plans, we interpolate between consecutive states to simulate how the plan will be executed in practice."], "citing_paper_content": {"title": "Predicting Motion Plans For Articulating Everyday Objects", "abstract": "Mobile manipulation tasks such as opening a door, pulling open a drawer, or lifting a toilet lid require constrained motion of the end-effector under environmental and task constraints. This, coupled with partial information in novel environments, makes it challenging to employ classical motion planning approaches at test time. Our key insight is to cast it as a learning problem to leverage past experience of solving similar planning problems to directly predict motion plans for mobile manipulation tasks in novel situations at test time. To enable this, we develop a simulator, ArtObjSim, that simulates articulated objects placed in real scenes. We then introduce SeqIK+\u03b80, a fast and flexible representation for motion plans. Finally, we learn models that use SeqIK+\u03b80 to quickly predict motion plans for articulating novel objects at test time. Experimental evaluation shows improved speed and accuracy at generating motion plans than pure search-based methods and pure learning methods."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["ImageNet"], "citation_intent": "method"}
{"citing_id": "2303.16585v1", "cited_id": "1803.11173", "section_title": "Properties Of Quantum Deep Hedging", "citation": "We can show however that in our setting, the well-known \"barren plateau\" problem #REFR does not arise.", "text_before_citation": ["From the definition in Section 4.4, each of the neural networks has O(T 2 ) parameters.", "As a consequence of results due to Caro and Datta #OTHEREFR , the pseudo-dimension of our parameterized model is polynomial in T .", "Therefore poly(T ) episodes suffice to ensure that empirical risk minimization over our sample converges with high probability to the true optimal expressible model over the whole distribution of futures.", "\u2022 Trainability: Finally we consider whether the task of optimizing the parameters for our model can be performed efficiently.", "The associated optimization problem is non-convex and thus training convergence cannot be guaranteed."], "text_after_citation": ["Each Quantum Compound Neural Network that we use is on O(T ) qubits and has O(T ) depth.", "Furthermore the loss function we measure can be constructed as a function of measurements in the computational basis (corresponding to a measuring the vector of observables Z on each qubit i).", "We initialize the parameters of the model as normal random variables with variance O(1/T ).", "Theorem 2 ensures that the gradients decay only polynomially with the time horizon T ."], "citing_paper_content": {"title": "Quantum Deep Hedging", "abstract": "Quantum machine learning has the potential for a transformative impact across industry sectors and in particular in finance. In our work we look at the problem of hedging where deep reinforcement learning offers a powerful framework for real markets. We develop quantum reinforcement learning methods based on policy-search and distributional actor-critic algorithms that use quantum neural network architectures with orthogonal and compound layers for the policy and value functions. We prove that the quantum neural networks we use are trainable, and we perform extensive simulations that show that quantum models can reduce the number of trainable parameters while achieving comparable performance and that the distributional approach obtains better performance than other standard approaches, both classical and quantum. We successfully implement the proposed models on a trapped-ion quantum processor, utilizing circuits with up to 16 qubits, and observe performance that agrees well with noiseless simulation. Our quantum techniques are general and can be applied to other reinforcement learning problems beyond hedging."}, "cited_paper_content": {"title": "Barren Plateaus In Quantum Neural Network Training Landscapes", "abstract": "Many experimental proposals for noisy intermediate scale quantum devices involve training a parameterized quantum circuit with a classical optimization loop. Such hybrid quantum-classical algorithms are popular for applications in quantum simulation, optimization, and machine learning. Due to its simplicity and hardware efficiency, random circuits are often proposed as initial guesses for exploring the space of quantum states. We show that the exponential dimension of Hilbert space and the gradient estimation complexity make this choice unsuitable for hybrid quantum-classical algorithms run on more than a few qubits. Specifically, we show that for a wide class of reasonable parameterized quantum circuits, the probability that the gradient along any reasonable direction is non-zero to some fixed precision is exponentially small as a function of the number of qubits. We argue that this is related to the 2-design characteristic of random circuits, and that solutions to this problem must be studied."}, "keywords": ["setting", "well-known \"barren plateau\""], "citation_intent": "background"}
{"citing_id": "2303.17977v1", "cited_id": "1808.03517", "section_title": "Introduction", "citation": "In our evaluation, we show the integration of our implemented prototype with Caterpillar #REFR , a state-ofthe-art process execution engine, to demonstrate how our approach can complement a business process management system to secure its data flow.", "text_before_citation": ["To overcome these limitations, we here propose a Multi-Authority Approach to Transaction Systems for Interoperating Applications (MARTSIA).", "In MARTSIA, encrypted data is persisted in decentralized storage, which is connected to a public permissionless blockchain system supporting process execution.", "Data owners define access policies to regulate which users are able to view specific parts of the information. No central authority can generate decryption keys alone.", "The encryption and decryption of messages are left to the individual nodes.", "To attain the desired characteristics, this approach employs hybrid encryption in combination with the Cyphertext-Policy variant of Multi-Authority Attribute-Based Encryption (henceforth, MA-ABE for the sake of conciseness), smart contracts, and InterPlanetary File System (IPFS)."], "text_after_citation": ["In the following, Sect.", "2 presents a running example, to which we will refer throughout the paper, and illustrates the problem we tackle.", "Section 3 outlines the fundamental notions that our solution is based upon. In Sect. 4, we describe our approach in detail. In Sect.", "5, we present our proof-of-concept implementation and the results of the experiments we conducted therewith. Section 6 reviews related work before Sect. 7 concludes the paper and outlines future works.", "2 Example, problem illustration, and requirements Figure 1 shows a Business Process Model and Notation (BPMN) collaboration diagram illustrating a supply chain in the automotive area: the production of a special car for a person with paraplegia."], "citing_paper_content": {"title": "Martsia: Enabling Data Confidentiality For Blockchain-Based Process Execution", "abstract": "Multi-party business processes rely on the collaboration of various players in a decentralized setting. Blockchain technology can facilitate the automation of these processes, even in cases where trust among participants is limited. Transactions are stored in a ledger, a replica of which is retained by every node of the blockchain network. The operations saved thereby are thus publicly accessible. While this enhances transparency, reliability, and persistence, it hinders the utilization of public blockchains for process automation as it violates typical confidentiality requirements in corporate settings. In this paper, we propose MARTSIA: A Multi-Authority Approach to Transaction Systems for Interoperating Applications. MARTSIA enables precise control over process data at the level of message parts. Based on Multi-Authority Attribute-Based Encryption (MA-ABE), MARTSIA realizes a number of desirable properties, including confidentiality, transparency, and auditability. We implemented our approach in proof-of-concept prototypes, with which we conduct a case study in the area of supply chain management. Also, we show the integration of MARTSIA with a state-of-the-art blockchain-based process execution engine to secure the data flow."}, "cited_paper_content": {"title": "Caterpillar: A Business Process Execution Engine On The Ethereum Blockchain", "abstract": "Blockchain platforms, such as Ethereum, allow a set of actors to maintain a ledger of transactions without relying on a central authority and to deploy scripts, called smart contracts, that are executed whenever certain transactions occur. These features can be used as basic building blocks for executing collaborative business processes between mutually untrusting parties. However, implementing business processes using the low-level primitives provided by blockchain platforms is cumbersome and error-prone. In contrast, established business process management systems, such as those based on the standard Business Process Model and Notation (BPMN), provide convenient abstractions for rapid development of process-oriented applications. This article demonstrates how to combine the advantages of a business process management system with those of a blockchain platform. The article introduces a blockchain-based BPMN execution engine, namely Caterpillar. Like any BPMN execution engine, Caterpillar supports the creation of instances of a process model and allows users to monitor the state of process instances and to execute tasks thereof. The specificity of Caterpillar is that the state of each process instance is maintained on the (Ethereum) blockchain and the workflow routing is performed by smart contracts generated by a BPMN-to-Solidity compiler. The Caterpillar compiler supports a large array of BPMN constructs, including subprocesses, multi-instances activities and event handlers. The paper describes the architecture of Caterpillar, and the interfaces it provides to support the monitoring of process instances, the allocation and execution of work items, and the execution of service tasks."}, "keywords": ["business process management"], "citation_intent": "method"}
{"citing_id": "2303.02449v1", "cited_id": "1904.05044", "section_title": "Preliminaries & Motivation", "citation": "In principle, CAM roughly indicates the contribution of local image regions to classification confidence #REFR .", "text_before_citation": ["R 3\u00d7H\u00d7W \u2192 R C\u00d7H\u00d7W .", "The classification prediction score\u0177 \u2208 R C can be calculated as\u0177 = \u03c3(GAP(f (x))), where \u03c3 denotes the sigmoid activation function.", "To train the network, the classification loss is defined as the binary cross-entropy between prediction y and ground-truth\u0177, i.e.,", "EQUATION", "After training f , the CAM of an image x related to the c-th class, denoted as M c \u2208 R C\u00d7H\u00d7W , c \u2208 {1, 2, .., C}, can be directly calculated as, Shortage of CAM."], "text_after_citation": ["The discriminating features always play a dominant role in the classification, as can be seen from Figure 2 (a), where the confidence of the bird head regions (highlighted by red boxes) is much ahead of other regions.", "The phenomenon results in CAM mainly confined to the small salient regions.", "Most methods #OTHEREFR modify the structure of the CAM generating network or introduce well-designed modules, which alter the contribution of image regions to classification confidence, leading to the expansion of the CAM.", "However, Figure 2 (a) suggests that foreground points (in red) generally lie above background points (in blue), indicating that CAM still treats non-discriminating features differently from the background. Here, we pose the question Is CAM fully exploited?", "EQUATION"], "citing_paper_content": {"title": "Exploit Cam By Itself: Complementary Learning System For Weakly Supervised Semantic Segmentation", "abstract": "Weakly Supervised Semantic Segmentation (WSSS) with image-level labels has long been suffering from fragmentary object regions led by Class Activation Map (CAM), which is incapable of generating fine-grained masks for semantic segmentation. To guide CAM to find more non-discriminating object patterns, this paper turns to an interesting working mechanism in agent learning named Complementary Learning System (CLS). CLS holds that the neocortex builds a sensation of general knowledge, while the hippocampus specially learns specific details, completing the learned patterns. Motivated by this simple but effective learning pattern, we propose a General-Specific Learning Mechanism (GSLM) to explicitly drive a coarse-grained CAM to a fine-grained pseudo mask. Specifically, GSLM develops a General Learning Module (GLM) and a Specific Learning Module (SLM). The GLM is trained with image-level supervision to extract coarse and general localization representations from CAM. Based on the general knowledge in the GLM, the SLM progressively exploits the specific spatial knowledge from the localization representations, expanding the CAM in an explicit way. To this end, we propose the Seed Reactivation to help SLM reactivate nondiscriminating regions by setting a boundary for activation values, which successively identifies more regions of CAM. Without extra refinement processes, our method is able to achieve breakthrough improvements for CAM of over 20.0% mIoU on PASCAL VOC 2012 and 10.0% mIoU on MS COCO 2014 datasets, representing a new state-of-the-art among existing WSSS methods."}, "cited_paper_content": {"title": "Weakly Supervised Learning Of Instance Segmentation With Inter-Pixel Relations", "abstract": "This paper presents a novel approach for learning instance segmentation with image-level class labels as supervision. Our approach generates pseudo instance segmentation labels of training images, which are used to train a fully supervised model. For generating the pseudo labels, we first identify confident seed areas of object classes from attention maps of an image classification model, and propagate them to discover the entire instance areas with accurate boundaries. To this end, we propose IRNet, which estimates rough areas of individual instances and detects boundaries between different object classes. It thus enables to assign instance labels to the seeds and to propagate them within the boundaries so that the entire areas of instances can be estimated accurately. Furthermore, IRNet is trained with inter-pixel relations on the attention maps, thus no extra supervision is required. Our method with IRNet achieves an outstanding performance on the PASCAL VOC 2012 dataset, surpassing not only previous state-of-the-art trained with the same level of supervision, but also some of previous models relying on stronger supervision."}, "keywords": ["local image regions"], "citation_intent": "background"}
{"citing_id": "2303.07929v1", "cited_id": "1812.04948", "section_title": "Style Transfer And Adaptive Instance Normalization", "citation": "StyleGAN #REFR draws on the idea of AdaIN style transfer, removes the traditional input, and proposes a stylebased generator, which takes a learnable constant as input.", "text_before_citation": ["This research found that replacing BN with instance normalization (IN) can significantly improve the convergence speed #OTHEREFR . Dumoulin et al.", "#OTHEREFR found that images with different styles can be generated by using different scales and displacements during IN operation, also known as conditional instance normalization (CIN). Huang et al.", "#OTHEREFR proposed that the artistic style of the image is the cross-spatial statistical information of each feature channel of the feature graph, such as mean and variance.", "Style transfer can be realized by transferring each channel's mean and standard deviation.", "This operation is named adaptive instance normalization (AdaIN), which ensures transferring any styles to the feature maps."], "text_after_citation": ["It mainly controls the visual features represented by each level by modifying the input of each level separately without affecting other levels.", "And StyleGAN2 #OTHEREFR adjusted the use of AdaIN to avoid water droplet artifacts effectively.", "With the advantage of GAN network, some scholars try to generate high-quality facial age images.", "LATS #OTHEREFR presented a method for synthesizing lifespan age transformations.", "RAGAN #OTHEREFR introduces a personalized self-guidance scheme that enables transforming the input face across various target age groups while preserving identity. Our designed network needs two modal inputs."], "citing_paper_content": {"title": "Daa: A Delta Age Adain Operation For Age Estimation Via Binary Code Transformer", "abstract": "Naked eye recognition of age is usually based on comparison with the age of others. However, this idea is ignored by computer tasks because it is difficult to obtain representative contrast images of each age. Inspired by the transfer learning, we designed the Delta Age AdaIN (DAA) operation to obtain the feature difference with each age, which obtains the style map of each age through the learned values representing the mean and standard deviation. We let the input of transfer learning as the binary code of age natural number to obtain continuous age feature information. The learned two groups of values in Binary code mapping are corresponding to the mean and standard deviation of the comparison ages. In summary, our method consists of four parts: FaceEncoder, DAA operation, Binary code mapping, and AgeDecoder modules. After getting the delta age via AgeDecoder, we take the average value of all comparison ages and delta ages as the predicted age. Compared with state-of-the-art methods, our method achieves better performance with fewer parameters on multiple facial age datasets."}, "cited_paper_content": {"title": "A Style-Based Generator Architecture For Generative Adversarial Networks", "abstract": "We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces."}, "keywords": ["AdaIN style transfer", "stylebased generator"], "citation_intent": "background"}
{"citing_id": "2303.04434v1", "cited_id": "1911.05425", "section_title": "Introduction", "citation": "Much work has been done on the optimal polynomial approximation of circular arcs (see, e.g., #REFR and the references therein).", "text_before_citation": ["One of the most important issues of computer-aided geometric design (CAGD) is the approximation of curves and surfaces by simple polynomial-based objects.", "It is well known that even fundamental geometric objects such as circular arcs or spherical patches do not possess exact polynomial representations.", "It is thus an interesting and practical issue to find their best polynomial approximants."], "text_after_citation": ["However, much less is known about the optimal approximation of spherical patches.", "The main two classes of the polynomial (or spline) surfaces used in CAGD are triangular patches and tensor product patches.", "The optimal approximation of equilateral spherical triangles by triangular B\u00e9zier patches was studied recently in #OTHEREFR .", "Approximation of rational tensor-product biquadratic B\u00e9zier surfaces (which also includes the sphere) by polynomial tensor product patches was proposed in #OTHEREFR .", "An exact sphere representation by rational S-patches was considered recently in #OTHEREFR ."], "citing_paper_content": {"title": "Optimal Approximation Of Spherical Squares By Tensor Product Quadratic B\u00e9zier Patches", "abstract": "In [1], the author considered the problem of the optimal approximation of symmetric surfaces by biquadratic B\u00e9zier patches. Unfortunately, the results therein are incorrect, which is shown in this paper by considering the optimal approximation of spherical squares. A detailed analysis and a numerical algorithm are given, providing the best approximant according to the (simplified) radial error, which differs from the one obtained in [1]. The sphere is then approximated by the continuous spline of two and six tensor product quadratic B\u00e9zier patches. It is further shown that the G 1 smooth spline of six patches approximating the sphere exists, but it is not a good approximation. The problem of an approximation of spherical rectangles is also addressed and numerical examples indicate that several optimal approximants might exist in some cases, making the problem extremely difficult to handle. Finally, numerical examples are provided that confirm theoretical results."}, "cited_paper_content": {"title": "Optimal Parametric Interpolants Of Circular Arcs.", "abstract": "The aim of this paper is a construction of quartic parametric polynomial interpolants of a circular arc, where two boundary points of a circular arc are interpolated. For every unit circular arc of inner angle not greater than $\\pi$ we find the best interpolant, where the optimality is measured by the simplified radial error."}, "keywords": ["optimal polynomial approximation"], "citation_intent": "background"}
{"citing_id": "2305.02996v1", "cited_id": "1706.03762", "section_title": "Introduction", "citation": "However, computing a single queryitem score using cross-encoders requires a forward pass of the model which can be computationally expensive as cross-encoders are typically parameterized using deep neural models such as transformers #REFR .", "text_before_citation": ["k-nearest neighbor (k-NN) search is core subroutine of a variety of tasks in NLP such as entity linking #OTHEREFR , passage retrieval for QA #OTHEREFR , and more generally, in retrieval-augmented machine learning models (Guu et al., 2020; Izacard et al., 2022) .", "Neural models are widely used for parameterizing the similarity function in such applications due to their widespread success #OTHEREFR .", "The efficiency of k-NN search depends largely on the architecture and parameterization of the similarity function.", "In this work, we focus on k-NN search for cross-encoders, a popular class of neural similarity functions.", "Cross-encoder models directly output a scalar similarity score after jointly encoding a given query-item pair, and generalize better to new domains and unseen data as compared alternative choices such as dual-encoders which independently embed query and item and compute similarity using simple functions such as dot-product of these embeddings."], "text_after_citation": ["For this reason, k-NN search with cross-encoders typically involves retrieving candidate items using additional models such as dual-encoders or BM25 #OTHEREFR , followed by re-ranking items using the cross-encoder #OTHEREFR .", "Much of recent work focuses on improving recall of the first-stage retrieval using distillation-based techniques #OTHEREFR , or on designing new architectures that trade-off effectiveness for higher efficiency while attempting to capture fine-grained interactions between the query and the item #OTHEREFR .", "Recent work #OTHEREFR explores matrix factorization based approaches for approximating cross-encoder scores. #OTHEREFR", "(2022) proposed ANNCUR, a ma-trix factorization based method, to perform k-NN search with cross-encoders without relying on additional dual-encoder style models for candidate retrieval.", "ANNCUR use an approximate-then-retrieve strategy where it approximates query-item scores using latent query and item embeddings, and performs retrieval using the approximate scores followed by optionally re-ranking them using exact CE scores."], "citing_paper_content": {"title": "Adaptive Selection Of Anchor Items For Cur-Based K-Nn Search With Cross-Encoders", "abstract": "Cross-encoder models, which jointly encode and score a query-item pair, are typically prohibitively expensive for k-nearest neighbor search. Consequently, k-NN search is performed not with a cross-encoder, but with a heuristic retrieve (e.g., using BM25 or dualencoder) and re-rank approach. Recent work proposes ANNCUR (Yadav et al., 2022) which uses CUR matrix factorization to produce an embedding space for efficient vector-based search that directly approximates the crossencoder without the need for dual-encoders. ANNCUR defines this shared query-item embedding space by scoring the test query against anchor items which are sampled uniformly at random. While this minimizes average approximation error over all items, unsuitably high approximation error on top-k items remains and leads to poor recall of top-k (and especially top-1) items. Increasing the number of anchor items is a straightforward way of improving the approximation error and hence k-NN recall of ANNCUR but at the cost of increased inference latency. In this paper, we propose a new method for adaptively choosing anchor items that minimizes the approximation error for the practically important topk neighbors for a query with minimal computational overhead. Our proposed method incrementally selects a suitable set of anchor items for a given test query over several rounds, using anchors chosen in previous rounds to inform selection of more anchor items. Empirically, our method consistently improves k-NN recall as compared to both ANNCUR and the widely-used dual-encoder-based retrieve-andrerank approach."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["single queryitem score", "deep neural models"], "citation_intent": "method"}
{"citing_id": "2303.12054v2", "cited_id": "1612.01105", "section_title": "Experimental Setting", "citation": "We follow the same standard configuration and training process of both models as in #REFR .", "text_before_citation": ["The victim class and target class of Cityscapes dataset are set to be class 13 (car) and class 0 (road), respectively.", "In this study, we use the classic Hello Kitty pattern as the backdoor trigger.", "The trigger size is set to 15 \u00d7 15 pixels for the VOC dataset and 55 \u00d7 55 for the Cityscapes dataset. Segmentation models.", "Two popular semantic segmentation architectures, namely PSPNet #OTHEREFR and DeepLabV3 #OTHEREFR , are adopted in this work.", "In both architectures, ResNet-50 #OTHEREFR pre-trained on ImageNet #OTHEREFR is used as an embedding backbone."], "text_after_citation": [], "citing_paper_content": {"title": "Influencer Backdoor Attack On Semantic Segmentation", "abstract": "When a small number of poisoned samples are injected into the training dataset of a deep neural network, the network can be induced to exhibit malicious behavior during inferences, which poses potential threats to real-world applications. While they have been intensively studied in classification, backdoor attacks on semantic segmentation have been largely overlooked. Unlike classification, semantic segmentation aims to classify every pixel within a given image. In this work, we explore backdoor attacks on segmentation models to misclassify all pixels of a victim class by injecting a specific trigger on non-victim pixels during inferences, which is dubbed Influencer Backdoor Attack (IBA). IBA is expected to maintain the classification accuracy of non-victim pixels and misleads classifications of all victim pixels in every single inference. Specifically, we consider two types of IBA scenarios, i.e., 1) Free-position IBA: the trigger can be positioned freely except for pixels of the victim class, and 2) Long-distance IBA: the trigger can only be positioned somewhere far from victim pixels, given the possible practical constraint. Based on the context aggregation ability of segmentation models, we propose techniques to improve IBA for the scenarios. Concretely, for free-position IBA, we propose a simple, yet effective Nearest Neighbor trigger injection strategy for poisoned sample creation. For long-distance IBA, we propose a novel Pixel Random Labeling strategy. Our extensive experiments reveal that current segmentation models do suffer from backdoor attacks, and verify that our proposed techniques can further increase attack performance."}, "cited_paper_content": {"title": "Pyramid Scene Parsing Network", "abstract": "Scene parsing is challenging for unrestricted open vocabulary and diverse scenes. In this paper, we exploit the capability of global context information by different-region-based context aggregation through our pyramid pooling module together with the proposed pyramid scene parsing network (PSPNet). Our global prior representation is effective to produce good quality results on the scene parsing task, while PSPNet provides a superior framework for pixel-level prediction. The proposed approach achieves state-of-the-art performance on various datasets. It came first in ImageNet scene parsing challenge 2016, PASCAL VOC 2012 benchmark and Cityscapes benchmark. A single PSPNet yields the new record of mIoU accuracy 85.4% on PASCAL VOC 2012 and accuracy 80.2% on Cityscapes."}, "keywords": ["models", "training process"], "citation_intent": "method"}
{"citing_id": "2304.14381v1", "cited_id": "1910.10683", "section_title": "Comparison With Petl Methods", "citation": "We also extend \u03c0-Tuning on a uni-modal pretrained foundation model, T5 #REFR , to solve natural language understanding tasks.", "text_before_citation": ["(c) The test error surface on RefCOCOg #OTHEREFR , as a function of model weights in a two-dimensional subspace 3 .", "It shows that checkpoints of similar tasks lie in the same basin of the surface.", "Interpolation of those checkpoints could achieve lower test error on the surface. The visualization follows #OTHEREFR", "(2018) , which derives an orthonormal basis\u00fb,v by three checkpoints.", "Pretrained foundation models."], "text_after_citation": ["We demonstrate the results in Table 3 , where we find that RTE is the task that benefits most with an improvement of 9.29% accuracy.", "Experts from PETL methods.", "In Table 5 , experiments on RefCOCO, RefCOCO+, and RefCOCOg datasets with the base-size model show that \u03c0-Tuning is agnostic to PETL methods, providing significant and consistent performance gain for LoRA and Prompt Tuning.", "\u03c0-LoRA improves a large margin of 4.20% on average accuracy across three datasets towards original LoRA and 1.94% for \u03c0-Prompt towards original Prompt Tuning."], "citing_paper_content": {"title": "\u03a0-Tuning: Transferring Multimodal Foundation Models With Optimal Multi-Task Interpolation", "abstract": "Foundation models have achieved great advances in multi-task learning with a unified interface of unimodal and multimodal tasks. However, the potential of such multi-task learners has not been exploited during transfer learning. In this work, we present a universal parameter-efficient transfer learning method, termed Predict-Interpolate Tuning (\u03c0-Tuning), for vision, language, and vision-language tasks. It aggregates the parameters of lightweight task-specific experts learned from similar tasks to aid the target downstream task. The task similarities are predicted in a unified modality-independent space, yielding a scalable graph to demonstrate task relationships. \u03c0-Tuning has several appealing benefits. First, it flexibly explores both intra-and intermodal transferability between similar tasks to improve the accuracy and robustness of transfer learning, especially in data-scarce scenarios. Second, it offers a systematical solution for transfer learning with multi-task prediction-and-theninterpolation, compatible with diverse types of parameter-efficient experts, such as prompt and adapter. Third, an extensive study of task-level mutual benefits on 14 unimodal and 6 multimodal datasets shows that \u03c0-Tuning surpasses fine-tuning and other parameter-efficient transfer learning methods both in full-shot and low-shot regimes. The task graph also enables an in-depth interpretable analysis of task transferability across modalities. The code will be available at https:"}, "cited_paper_content": {"title": "Exploring The Limits Of Transfer Learning With A Unified Text-To-Text Transformer", "abstract": "Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new \"Colossal Clean Crawled Corpus\", we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code."}, "keywords": ["uni-modal pretrained foundation", "natural language understanding"], "citation_intent": "method"}
{"citing_id": "2303.03181v1", "cited_id": "1806.07366", "section_title": "Predict", "citation": "NeuralODE #REFR The prediction dynamics corresponding to the latent NeuralODE model is given by dXt dt = F nn (X t , z \u2264r ; W 1 ) where z \u2264r = F enc (X t0 , . . .", "text_before_citation": ["\u03a6,\u03be,\u0174 (1) , . . .", ",\u0174 (M ) = arg min \u03a6,\u03be,W (1) ,...,W (M ) 1 M M i=1 R (i) (W (i) , \u03a6, \u03be) + \u03bb \u03a6 ||\u03a6|| 1 (6) + \u03bb REx Variance({R (i) (W (i) , \u03a6, \u03be)} M i=1 )", "We perform a grid search over the following hyperparameters: regularization strengths \u03bb \u03a6 \u2208 {10 \u22124 , 10 \u22123 , 5 \u00d7 10 \u22123 , 10 \u22122 }, \u03bb REx \u2208 {0, 10 \u22123 , 10 \u22122 }, and learning rates \u03b7 \u2208 {10 \u22122 , 10 \u22123 , 10 \u22124 }.", "We choose the hyperparameters that result in sparsest model (i.e., with the least ||\u03a6|| 0 ) while achieving validation loss within 5% of the best validation loss in held-out in-distribution validation data.", "B.2."], "text_after_citation": [", X tr ; W 2 ) encodes the initial observations using a recurrent neural network F enc (e.g., GRU), and F nn is a feedforward neural network.", "The model is trained with an ODE solver (dopri5) and the gradients computed using the adjoint method #OTHEREFR .", "We perform a grid search over the following hyperparameters: number of layers for F nn , L \u2208 {1, 2, 3}, size of each hidden layer of F nn , d h \u2208 {32, 64, 128}, size of the encoder representation z \u2264r , d z \u2208 {32, 64, 128}, batch sizes B \u2208 {32, 64}, and learning rates \u03b7 \u2208 {10 \u22122 , 10 \u22123 , 10 \u22124 }."], "citing_paper_content": {"title": "Metaphysica: Ood Robustness In Physics-Informed Machine Learning", "abstract": "A fundamental challenge in physics-informed machine learning (PIML) is the design of robust PIML methods for out-of-distribution (OOD) forecasting tasks. These OOD tasks require learning-to-learn from observations of the same (ODE) dynamical system with different unknown ODE parameters, and demand accurate forecasts even under out-of-support initial conditions and out-of-support ODE parameters. In this work we propose a solution for such tasks, which we define as a meta-learning procedure for causal structure discovery (including invariant risk minimization). Using three different OOD tasks, we empirically observe that the proposed approach significantly outperforms existing state-of-the-art PIML and deep learning methods."}, "cited_paper_content": {"title": "Neural Ordinary Differential Equations", "abstract": "We introduce a new family of deep neural network models. Instead of specifying a discrete sequence of hidden layers, we parameterize the derivative of the hidden state using a neural network. The output of the network is computed using a black-box differential equation solver. These continuous-depth models have constant memory cost, adapt their evaluation strategy to each input, and can explicitly trade numerical precision for speed. We demonstrate these properties in continuous-depth residual networks and continuous-time latent variable models. We also construct continuous normalizing flows, a generative model that can train by maximum likelihood, without partitioning or ordering the data dimensions. For training, we show how to scalably backpropagate through any ODE solver, without access to its internal operations. This allows end-to-end training of ODEs within larger models."}, "keywords": ["prediction dynamics", "NeuralODE model"], "citation_intent": "background"}
{"citing_id": "2303.01155v2", "cited_id": "1902.03729", "section_title": "Ii. Related Works", "citation": "Authors in #REFR proposed UcoSLAM that employs both natural and artificial landmarks in the environment to reconstruct the map.", "text_before_citation": ["VSLAM approaches have matured over the past years and many researchers contributed to this community by proposing efficient solutions to unresolved challenges in this area.", "The authors of this paper also presented a comprehensive survey on diverse VSLAM state-of-the-art works and studied their trade-offs in #OTHEREFR .", "Accordingly, many recent approaches focus on improving the front-end thread of the VSLAM systems for better pose estimation and visual odometry #OTHEREFR ."], "text_after_citation": ["It works in keypoints-only, markers-only, and mixed modes and improves the global map builder and loop closure detector modules in case ArUco markers are found in the camera's field of view.", "Another marker-based approach titled TagSLAM #OTHEREFR was proposed by Pfrommer and Daniilidis, which uses the capability of AprilTag #OTHEREFR fiducial markers for SLAM tasks.", "Although the proposed approaches perform well in many scenarios, they cannot obtain any semantic information from the environment for an improved map representation. Gomez-Ojeda et al.", "#OTHEREFR introduced PL-SLAM, which uses Oriented FAST and Rotated BRIEF (ORB) and Line Segment Detector (LSD) algorithms to extract visual features of points and lines in stereo vision cameras.", "However, this approach is computationally intensive in feature tracking and extracting data from structural lines."], "citing_paper_content": {"title": "Marker-Based Visual Slam Leveraging Hierarchical Representations", "abstract": "Fiducial markers can encode rich information about the environment and can aid Visual SLAM (VSLAM) approaches in reconstructing maps with practical semantic information. Current marker-based VSLAM approaches mainly utilize markers for improving feature detections in low-feature environments and/or for incorporating loop closure constraints, generating only low-level geometric maps of the environment prone to inaccuracies in complex environments. To bridge this gap, this paper presents a VSLAM approach utilizing a monocular camera along with fiducial markers to generate hierarchical representations of the environment while improving the camera pose estimate. The proposed approach detects semantic entities from the surroundings, including walls, corridors, and rooms encoded within markers, and appropriately adds topological constraints among them. Experimental results on a real-world dataset collected with a robot demonstrate that the proposed approach outperforms a traditional marker-based VSLAM baseline in terms of accuracy, given the addition of new constraints while creating enhanced map representations. Furthermore, it shows satisfactory results when comparing the reconstructed map quality to the one reconstructed using a LiDAR SLAM approach."}, "cited_paper_content": {"title": "Ucoslam: Simultaneous Localization And Mapping By Fusion Of Keypoints And Squared Planar Markers", "abstract": "Simultaneous Localization and Mapping is the process of simultaneously creating a map of the environment while navigating in it. Most of the SLAM approaches use natural features (e.g. keypoints) that are unstable over time, repetitive in many cases or their number insufficient for a robust tracking (e.g. in indoor buildings). Other researchers, on the other hand, have proposed the use of artificial landmarks, such as squared fiducial markers, placed in the environment to help tracking and relocalization. This paper proposes a novel SLAM approach by fusing natural and artificial landmarks in order to achieve long-term robust tracking in many scenarios. Our method has been compared to the start-of-the-art methods ORB-SLAM2 [1], LDSO [2] and SPM-SLAM [3] in the public datasets Kitti [4], Euroc-MAV [5], TUM [6] and SPM [3], obtaining better precision, robustness and speed. Our tests also show that the combination of markers and keypoints achieves better accuracy than each one of them independently."}, "keywords": ["artificial landmarks"], "citation_intent": "background"}
{"citing_id": "2304.00077v1", "cited_id": "1111.2893", "section_title": "Introduction", "citation": "This can be done by appropriately splitting up the prize (Moldovanu and Sela, 2001) , choosing a suitable reserve effort #REFR , setting an entry fee (Taylor, 1995) or by developing a revelation mechanism to select a subset of contestants from a pool of candidates (Mercier, 2018) .", "text_before_citation": ["Bug bounty programs are also used by governments and tech companies.", "3 There have been comprehensive accounts on the rules of engagement of bug bounty programs (Laszka et al., 2018) , on the effectiveness and best practices of such programs (Walshe and Simpson, 2020; Malladi and Subramanian, 2020) , and on the incentives of researchers to participate in bug bounty schemes (Maillart et al., 2017) .", "In this paper, we offer insights on some of the dimensions of bug bounty design, using a game-theoretic model of a simple contest building on the important work of Ghosh and Kleinberg (2016) and Sarne and Lepioshkin (2017) , where agents with different abilities decide on whether or not to exert costly effort for finding bugs.", "Several salient features of bug bounty differentiate our approach from that of the standard optimal contest literature.", "The design objective in traditional contests is to elicit the highest effort (or sum of efforts) from the contestants."], "text_after_citation": ["We focus on the simple problem how to maximize the likelihood to find bugs when a given amount of money is available for rewards.", "In particular, we will focus on three design variables for bug bounty systems.", "How large should the crowd of agents invited to find bugs be? Should paid experts be added to the crowd of invited bug finders? Should artificial bugs be added to the software to increase participation in bug finding and to increase the likelihood that the real bug is found?", "To answer these questions and other, general questions about the nature of equilibria in bug bounty schemes, we develop a simple model of crowd-sourced security.", "A group of individuals of arbitrary size is invited to search for a bug. Whether a bug exists is uncertain."], "citing_paper_content": {"title": "Decentralized Attack Search And The Design Of Bug Bounty Schemes *", "abstract": "Systems and blockchains often have security vulnerabilities and can be attacked by adversaries, with potentially significant negative consequences. Therefore, organizations and blockchain infrastructure providers increasingly rely on bug bounty programs, where external individuals probe the system and report any vulnerabilities (bugs) in exchange for monetary rewards (bounty). We develop a contest model for bug bounty programs with an arbitrary number of agents who decide whether to undertake a costly search for bugs or not. Search costs are private information. Besides characterizing the ensuing equilibria, we show that even inviting an unlimited crowd does not guarantee that bugs are found. Adding paid agents can increase the efficiency of the bug bounty scheme although the crowd that is attracted becomes smaller. Finally, adding (known) bugs increases the likelihood that unknown bugs are found, but to limit reward payments it may be optimal to add them only with some probability."}, "cited_paper_content": {"title": "Optimal Crowdsourcing Contests", "abstract": "We study the design and approximation of optimal crowdsourcing contests. Crowdsourcing contests can be modeled as all-pay auctions because entrants must exert effort up-front to enter. Unlike all-pay auctions where a usual design objective would be to maximize revenue, in crowdsourcing contests, the principal only benefits from the submission with the highest quality. We give a theory for optimal crowdsourcing contests that mirrors the theory of optimal auction design: the optimal crowdsourcing contest is a virtual valuation optimizer (the virtual valuation function depends on the distribution of contestant skills and the number of contestants). We also compare crowdsourcing contests with more conventional means of procurement. In this comparison, crowdsourcing contests are relatively disadvantaged because the effort of losing contestants is wasted. Nonetheless, we show that crowdsourcing contests are 2-approximations to conventional methods for a large family of \"regular\" distributions, and 4-approximations, otherwise."}, "keywords": ["prize", "contestants"], "citation_intent": "background"}
{"citing_id": "2303.13123v1", "cited_id": "1703.04977", "section_title": "How Do Aleatoric And Epistemic Uncertainty Dif-", "citation": "This is in line with the findings of #REFR , that Dropout as the epistemic component imitates the aleatoric logit distribution, unable to capture the different types of uncertainty.", "text_before_citation": ["Figure 4 shows Precision-Recall curves for the proposed model, as well as the baseline models that use the SSN and therefore model both uncertainties.", "Note, that the SSN (diag) + Dropout corresponds to the heteroscedastic Bayesian Neural Network with Dropout derived by #OTHEREFR .", "First, we find that the aleatoric component, targeting the data variation, performs better in terms of segmentation performance than the epistemic component for our model.", "Further, the epistemic component of the LSN with diagonal covariance matrix performs worse (lower AUC), underpinning that the scaling of the loss has an effect on the post-hoc Laplace approximation.", "For the Dropout-based methods, on the other hand, we see virtually no difference between the epistemic and aleatoric components."], "text_after_citation": ["What does this tell us? We see that the aleatoric and epistemic components of the LSN clearly capture different aspects of the model's uncertainty.", "To better understand where this difference comes from, Figure 5 shows epistemic and aleatoric variance maps of sigmoids for a given ISIC test set sample.", "While the epistemic component assigns low variance to this in-distribution sample, the aleatoric component expresses the variation of plausible predictions for the image.", "We find this behavior attractive: In the Bayesian framework, the aleatoric component should play the role of the predictor, as it imitates the distribution of annotations Figure 4 : Precision-Recall plots for the ISIC test dataset show that aleatoric and epistemic variance in the LSN models (first row) are not the same, opposed to the SSN + Dropout models.", "The better performance of the aleatoric component (higher AUC) indicates that the model learns to separate data variation from model uncertainty."], "citing_paper_content": {"title": "Laplacian Segmentation Networks: Improved Epistemic Uncertainty From Spatial Aleatoric Uncertainty", "abstract": "Out of distribution (OOD) medical images are frequently encountered, e.g. because of site-or scanner differences, or image corruption. OOD images come with a risk of incorrect image segmentation, potentially negatively affecting downstream diagnoses or treatment. To ensure robustness to such incorrect segmentations, we propose Laplacian Segmentation Networks (LSN) that jointly model epistemic (model) and aleatoric (data) uncertainty in image segmentation. We capture data uncertainty with a spatially correlated logit distribution. For model uncertainty, we propose the first Laplace approximation of the weight posterior that scales to large neural networks with skip connections that have high-dimensional outputs. Empirically, we demonstrate that modelling spatial pixel correlation allows the Laplacian Segmentation Network to successfully assign high epistemic uncertainty to out-of-distribution objects appearing within images."}, "cited_paper_content": {"title": "What Uncertainties Do We Need In Bayesian Deep Learning For Computer Vision?", "abstract": "There are two major types of uncertainty one can model. Aleatoric uncertainty captures noise inherent in the observations. On the other hand, epistemic uncertainty accounts for uncertainty in the model - uncertainty which can be explained away given enough data. Traditionally it has been difficult to model epistemic uncertainty in computer vision, but with new Bayesian deep learning tools this is now possible. We study the benefits of modeling epistemic vs. aleatoric uncertainty in Bayesian deep learning models for vision tasks. For this we present a Bayesian deep learning framework combining input-dependent aleatoric uncertainty together with epistemic uncertainty. We study models under the framework with per-pixel semantic segmentation and depth regression tasks. Further, our explicit uncertainty formulation leads to new loss functions for these tasks, which can be interpreted as learned attenuation. This makes the loss more robust to noisy data, also giving new state-of-the-art results on segmentation and depth regression benchmarks."}, "keywords": ["uncertainty"], "citation_intent": "result"}
{"citing_id": "2303.14368v1", "cited_id": "2003.08934", "section_title": "Neural Radiance Fields (Nerfs)", "citation": "The landscape of neural rendering changed with NeRF #REFR , which proposed a simple, yet revolutionary approach for photorealistic novel-view synthesis of static scenes.", "text_before_citation": ["NeRFs attempt to learn a scene representation for novelview synthesis by modeling the radiance field with learnable functions.", "A variety of approaches have been proposed recently for neural rendering, exploiting voxel grids #OTHEREFR , neural textures #OTHEREFR , point-clouds #OTHEREFR , and neural implicit functions #OTHEREFR ."], "text_after_citation": ["NeRF attempts to map from the 5-d light fields to 4-d space consisting of color c (RGB) and density \u03c3: likelihood that the light ray at this 5-d co-ordinate is terminated by occlusion.", "Since the introduction of original NeRF formulation, several variations and improvements #OTHEREFR have been proposed."], "citing_paper_content": {"title": "Flexnerf: Photorealistic Free-Viewpoint Rendering Of Moving Humans From Sparse Views", "abstract": "We present FlexNeRF, a method for photorealistic freeviewpoint rendering of humans in motion from monocular videos. Our approach works well with sparse views, which is a challenging scenario when the subject is exhibiting fast/complex motions. We propose a novel approach which jointly optimizes a canonical time and pose configuration, with a pose-dependent motion field and pose-independent temporal deformations complementing each other. Thanks to our novel temporal and cyclic consistency constraints along with additional losses on intermediate representation such as segmentation, our approach provides high quality outputs as the observed views become sparser. We empirically demonstrate that our method significantly outperforms the state-of-the-art on public benchmark datasets as well as a self-captured fashion dataset. The project page is available at: https://flex-nerf.github.io/."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["neural rendering"], "citation_intent": "background"}
{"citing_id": "2303.09032v1", "cited_id": "1602.04621", "section_title": "Related Work", "citation": "Other successful exploration approaches like Boot-strappedDQN #REFR are unscalable in MARL because of the exponentially large state-action space.", "text_before_citation": ["Based on the Optimism in the Face of Uncertainty (OFU) principle, the high-level idea is to capture some notion of uncertainty or novelty, and augment the extrinsic reward the environment emits with an intrinsic reward that quantifies the uncertainty.", "For instance, the count-based method #OTHEREFR measures novelty through the number of times the agent observes a state-action tuple. #OTHEREFR", "[2016] propose an intrinsic bonus based on the maximization of information gain about the agent's belief of environment dynamics.", "Despite their recent success, simply applying bonus-based methods to MARL does not guarantee effective exploration.", "As the reward signal is shared across all agents, adding one additional centralized intrinsic reward may still be inefficient to learn structured cooperation due to the multi-agent credit-assignment challenge."], "text_after_citation": ["Multi-agent exploration.", "A recent branch of research proposes to drive multi-agent exploration by promoting col-laboration among agents through the maximization of the correlation or influence of agents.", "The correlation is commonly realized by the mutual information (MI) of quantities that define or reflect agents' behaviour, such as the trajectory history of each agent.", "For instance, MAVEN #OTHEREFR ] learns a hierarchical policy to produce a latent variable that encodes the information about the joint policy, and maximizes the MI between this latent variable and the joint trajectories to encourage the correlation of agents' behaviour.", "Some other methods try to promote collaboration by maximizing pairwise MI between every two agents."], "citing_paper_content": {"title": "Conditionally Optimistic Exploration For Cooperative Deep Multi-Agent Reinforcement Learning", "abstract": "Efficient exploration is critical in cooperative deep Multi-Agent Reinforcement Learning (MARL). In this paper, we propose an exploration method that efficiently encourages cooperative exploration based on the idea of the theoretically justified tree search algorithm UCT (Upper Confidence bounds applied to Trees). The high-level intuition is that to perform optimism-based exploration, agents would achieve cooperative strategies if each agent's optimism estimate captures a structured dependency relationship with other agents. At each node (i.e., action) of the search tree, UCT performs optimismbased exploration using a bonus derived by conditioning on the visitation count of its parent node. We provide a perspective to view MARL as tree search iterations and develop a method called Conditionally Optimistic Exploration (COE). We assume agents take actions following a sequential order, and consider nodes at the same depth of the search tree as actions of one individual agent. COE computes each agent's state-action value estimate with an optimistic bonus derived from the visitation count of the state and joint actions taken by agents up to the current agent. COE is adaptable to any value decomposition method for centralized training with decentralized execution. Experiments across various cooperative MARL benchmarks show that COE outperforms current stateof-the-art exploration methods on hard-exploration tasks. Preprint. Under review."}, "cited_paper_content": {"title": "Deep Exploration Via Bootstrapped Dqn", "abstract": "Efficient exploration in complex environments remains a major challenge for reinforcement learning. We propose bootstrapped DQN, a simple algorithm that explores in a computationally and statistically efficient manner through use of randomized value functions. Unlike dithering strategies such as epsilon-greedy exploration, bootstrapped DQN carries out temporally-extended (or deep) exploration; this can lead to exponentially faster learning. We demonstrate these benefits in complex stochastic MDPs and in the large-scale Arcade Learning Environment. Bootstrapped DQN substantially improves learning times and performance across most Atari games."}, "keywords": ["successful exploration approaches"], "citation_intent": "background"}
{"citing_id": "2303.17877v1", "cited_id": "1711.03936", "section_title": "Motivating Examples", "citation": "To trigger a liquidation, the liquidator needs to send a transaction #REFR to the augmented contract along with parameters (e.g., the flash loan size and liquidation amount).", "text_before_citation": ["Such protection corresponds to an authentication mechanism, a common practice that the practitioners' community employs. We proceed with a real-world authentication example.", "Authentication Protection Example.", "In Figure 2 , we show how a liquidation contract 3 attempts to prevent a transaction imitation attack.", "The solidity code of Figure 2 's liquidation contract is not open-source, and we therefore decompile the #OTHEREFR A liquidation refers to the process of selling collateral to secure debt #OTHEREFR . #OTHEREFR Address: 0x18C0cA3947E255881f94DE50B3a906Fc2759F7FE.", "contract bytecode with a state-of-the-art EVM contract decompiler #OTHEREFR ."], "text_after_citation": ["If there is no protection mechanism, an adversary might frontrun the liquidator, by calling the same contract with identical parameters.", "However, the presented contract requires the transaction sender to match a specific address (in line 3).", "Specifically, the liquidation function printMoney() is only callable by a hard-coded address.", "If this condition is not met, the naive imitation attack from Section 2.3 fails.", "To circumvent authentication protections, one approach is to synthesize an adversarial contract that replicates the liquidator's customized contract."], "citing_paper_content": {"title": "The Blockchain Imitation Game", "abstract": "The use of blockchains for automated and adversarial trading has become commonplace. However, due to the transparent nature of blockchains, an adversary is able to observe any pending, not-yet-mined transactions, along with their execution logic. This transparency further enables a new type of adversary, which copies and front-runs profitable pending transactions in real-time, yielding significant financial gains. Shedding light on such \"copy-paste\" malpractice, this paper introduces the Blockchain Imitation Game and proposes a generalized imitation attack methodology called APE. Leveraging dynamic program analysis techniques, APE supports the automatic synthesis of adversarial smart contracts. Over a timeframe of one year (1st of August, 2021 to 31st of July, 2022), APE could have yielded 148.96M USD in profit on Ethereum, and 42.70M USD on BNB Smart Chain (BSC). Not only as a malicious attack, we further show the potential of transaction and contract imitation as a defensive strategy. Within one year, we find that APE could have successfully imitated 13 and 22 known Decentralized Finance (DeFi) attacks on Ethereum and BSC, respectively. Our findings suggest that blockchain validators can imitate attacks in real-time to prevent intrusions in DeFi."}, "cited_paper_content": {"title": "Sok: Consensus In The Age Of Blockchains", "abstract": "The core technical component of blockchains is consensus: how to reach agreement among a distributed network of nodes. A plethora of blockchain consensus protocols have been proposed---ranging from new designs, to novel modifications and extensions of consensus protocols from the classical distributed systems literature. The inherent complexity of consensus protocols and their rapid and dramatic evolution makes it hard to contextualize the design landscape. We address this challenge by conducting a systematization of knowledge of blockchain consensus protocols. After first discussing key themes in classical consensus protocols, we describe: (i) protocols based on proof-of-work; (ii) proof-of-X protocols that replace proof-of-work with more energy-efficient alternatives; and (iii) hybrid protocols that are compositions or variations of classical consensus protocols. This survey is guided by a systematization framework we develop, to highlight the various building blocks of blockchain consensus design, along with a discussion on their security and performance properties. We identify research gaps and insights for the community to consider in future research endeavours."}, "keywords": ["liquidation", "transaction"], "citation_intent": "method"}
{"citing_id": "2304.13726v1", "cited_id": "1704.04760", "section_title": "B. Efficient Ml At The Iot Edge", "citation": "Other architectures also use a MAC-array structure, such as the Google Edge TPU #REFR architecture.", "text_before_citation": ["IoT devices at the edge are resource and energy constrained.", "This advocates for efficient hardware ML accelerators, increased but limited memory resources with reduced power consumption, and optimized ML algorithms well suited for these constrained nodes.", "In #OTHEREFR an overview of the architectures for efficient deep learning on IoT devices is presented.", "Flexible approaches such as GPUs are too power-hungry to be considered for integrating ML in real IoT devices, while FPGAs are not power-efficient enough. Research focuses on specialized accelerators, especially using dataflow paradigms.", "Eyeriss #OTHEREFR is an example of such an accelerator where small processing elements using private buffers are interconnected to implement row-stationary dataflow for efficient convolution operations."], "text_after_citation": ["These accelerators focus on convolutions, however complete networks may need more flexibility.", "This is the domain of heterogeneous accelerators such as the Orlando platform from STMicroelectronics #OTHEREFR , in which a dataflow hardware engine is coupled with SIMD DSP clusters.", "To be more efficient, algorithm-architecture matching needs to be improved.", "In #OTHEREFR algorithmic and processor techniques are studied to address deep learning in the edge.", "Quantization and network pruning are popular solutions to reduce the memory footprint of a DNN network #OTHEREFR with reduced classification accuracy degradation."], "citing_paper_content": {"title": "", "abstract": "Increased capabilities such as recognition and selfadaptability are now required from IoT applications. While IoT node power consumption is a major concern for these applications, cloud-based processing is becoming unsustainable due to continuous sensor or image data transmission over the wireless network. Thus optimized ML capabilities and data transfers should be integrated in the IoT node. Moreover, IoT applications are torn between sporadic data-logging and energy-hungry data processing (e.g. image classification). Thus, the versatility of the node is key in addressing this wide diversity of energy and processing needs. This paper presents SamurAI, a versatile IoT node bridging this gap in processing and in energy by leveraging two on-chip subsystems: a low power, clock-less, event-driven Always-Responsive (AR) part and an energy-efficient On-Demand (OD) part. AR contains a 1.7MOPS event-driven, asynchronous Wake-up Controller (WuC) with a 207ns wake-up time optimized for sporadic computing, while OD combines a deep-sleep RISC-V CPU and 1.3TOPS/W Machine Learning (ML) for more complex tasks up to 36GOPS. This architecture partitioning achieves best in class versatility metrics such as peak performance to idle power ratio. On an applicative classification scenario, it demonstrates system power gains, up to 3.5x compared to cloud-based processing, and thus extended battery lifetime."}, "cited_paper_content": {"title": "In-Datacenter Performance Analysis Of A Tensor Processing Unit", "abstract": "Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware. This paper evaluates a custom ASIC---called a Tensor Processing Unit (TPU) --- deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X -- 30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X -- 80X higher. Moreover, using the CPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU."}, "keywords": ["Google Edge TPU"], "citation_intent": "method"}
{"citing_id": "2303.17504v1", "cited_id": "1707.09092", "section_title": "Line Mapping", "citation": "For Aachen #REFR , parallel lines from the line-VP association graph are colored the same.", "text_before_citation": ["From the recovered structures, we can clearly perceive the scene and easily recognize the main Manhattan directions #OTHEREFR .", "To demonstrate the scalability of the proposed system, we also run our method on two large-scale datasets: Aachen (6,697 images) #OTHEREFR and Rome city (16,179 images) #OTHEREFR . Fig.", "7 shows that our method produces reliable line maps with clear structures.", "Note that the camera poses from Bundler #OTHEREFR on Rome city are far from perfect, while our mapping still works reasonably well. The efficiency Figure 7 .", "Scalability to large-scale datasets: Aachen (6,697 images) #OTHEREFR and Rome (16,179 images) #OTHEREFR ."], "text_after_citation": ["For Rome #OTHEREFR , we visualize 10 representative components individually. Figure 8 .", "Uncertainty in line triangulation measured by the largest eigenvalue of the covariance (Sec. D in supp.).", "Left: Each segment is colored by the uncertainty in the triangulation.", "Lines that align with the epipolar lines (shown in blue) exhibit higher (red) uncertainty.", "Right: We perform a small synthetic experiment to illustrate this."], "citing_paper_content": {"title": "3D Line Mapping Revisited", "abstract": "In contrast to sparse keypoints, a handful of line segments can concisely encode the high-level scene layout, as they often delineate the main structural elements. In addition to offering strong geometric cues, they are also omnipresent in urban landscapes and indoor scenes. Despite their apparent advantages, current line-based reconstruction methods are far behind their point-based counterparts. In this paper we aim to close the gap by introducing LIMAP, a library for 3D line mapping that robustly and efficiently creates 3D line maps from multi-view imagery. This is achieved through revisiting the degeneracy problem of line triangulation, carefully crafted scoring and track building, and exploiting structural priors such as line coincidence, parallelism, and orthogonality. Our code integrates seamlessly with existing point-based Structure-from-Motion methods and can leverage their 3D points to further improve the line reconstruction. Furthermore, as a byproduct, the method is able to recover 3D association graphs between lines and points / vanishing points (VPs). In thorough experiments, we show that LIMAP significantly outperforms existing approaches for 3D line mapping. Our robust 3D line maps also open up new research directions. We show two example applications: visual localization and bundle adjustment, where integrating lines alongside points yields the best results. Code is available at https://github.com/cvg/limap."}, "cited_paper_content": {"title": "Benchmarking 6Dof Outdoor Visual Localization In Changing Conditions", "abstract": "Visual localization enables autonomous vehicles to navigate in their surroundings and augmented reality applications to link virtual to real worlds. Practical visual localization approaches need to be robust to a wide variety of viewing condition, including day-night changes, as well as weather and seasonal variations, while providing highly accurate 6 degree-of-freedom (6DOF) camera pose estimates. In this paper, we introduce the first benchmark datasets specifically designed for analyzing the impact of such factors on visual localization. Using carefully created ground truth poses for query images taken under a wide variety of conditions, we evaluate the impact of various factors on 6DOF camera pose estimation accuracy through extensive experiments with state-of-the-art localization approaches. Based on our results, we draw conclusions about the difficulty of different conditions, showing that long-term localization is far from solved, and propose promising avenues for future work, including sequence-based localization approaches and the need for better local features. Our benchmark is available at visuallocalization.net."}, "keywords": ["line-VP association graph", "Aachen"], "citation_intent": "background"}
{"citing_id": "2303.08319v1", "cited_id": "1512.03385", "section_title": "Main Results", "citation": "For a fair comparison, we use the same experimental setups and compare these models with and without our proposed modules integrated. The default backbone is ResNet-50 #REFR . We summarize the results in Table 1 .", "text_before_citation": ["In this section, we conduct experiments on the dynamic query aggregation modules with the current Transformer-based object detectors on the ImageNet VID benchmark #OTHEREFR .", "The experiments of vanilla query aggregation will be provided in the following section."], "text_after_citation": ["For most of the DETR-based object detectors, the performance can be improved by at least 2.4% on the metric of mAP compared with those not integrated with our proposed modules.", "The performance is even better than those integrated with TransVOD [100] by a large margin, which validates the effectiveness of our dynamic query aggregation modules.", "When considering the objects' sizes, we notice that the performance of large or medium objects is much better than that of small objects.", "We argue that this is because the process to generate the dynamic queries Q d i is only based on the global features, which lack enough information for the small objects."], "citing_paper_content": {"title": "Faq: Feature Aggregated Queries For Transformer-Based Video Object Detectors", "abstract": "Video object detection needs to solve feature degradation situations that rarely happen in the image domain. One solution is to use the temporal information and fuse the features from the neighboring frames. With Transformerbased object detectors getting a better performance on the image domain tasks, recent works began to extend those methods to video object detection. However, those existing Transformer-based video object detectors still follow the same pipeline as those used for classical object detectors, like enhancing the object feature representations by aggregation. In this work, we take a different perspective on video object detection. In detail, we improve the qualities of queries for the Transformer-based models by aggregation. To achieve this goal, we first propose a vanilla query aggregation module that weighted averages the queries according to the features of the neighboring frames. Then, we extend the vanilla module to a more practical version, which generates and aggregates queries according to the features of the input frames. Extensive experimental results validate the effectiveness of our proposed methods: On the challenging ImageNet VID benchmark, when integrated with our proposed modules, the current state-of-the-art Transformerbased object detectors can be improved by more than 2.4% on mAP and 4.2% on AP 50 ."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["models"], "citation_intent": "method"}
{"citing_id": "2304.14321v1", "cited_id": "1605.09653", "section_title": "Introduction", "citation": "In #REFR , an extensive evaluation of multiple combinations of feature extractors and metric learning approaches is discussed.", "text_before_citation": ["There are many difficulties for Re-ID retrieval #OTHEREFR , among them: (i) different viewpoints, (ii) possible low-image resolutions, (iii) illumination changes, (iv) occlusions, (v) difficulty of manually labeling data for training, (vi) large amount of data to be processed.", "The challenge of improving the effectiveness of these systems, especially in open-world scenarios, has attracted a lot of research efforts from the scientific community #OTHEREFR .", "Initially, person Re-ID retrieval systems were mainly based on the use of handcrafted feature representations #OTHEREFR .", "Besides that, other strategies commonly used to generic image retrieval, like bag of visual words #OTHEREFR , have also been employed.", "Aiming at further improving the quality of the results, rather than using traditional distance measures, researchers have proposed metric learning approaches for Re-ID #OTHEREFR , most of them based on supervised models."], "text_after_citation": ["Due to the significant impact of deep learning on common image retrieval and machine learning #OTHEREFR , the Convolutional Neural Networks (CNN) also have gained a lot of attention and have been widely employed to solve Re-ID tasks in the recent years #OTHEREFR .", "Among the works, existing networks have been trained for Re-ID #OTHEREFR , and some new architectures have been proposed solely focusing on Re-ID #OTHEREFR .", "Despite the success of deep learning, one of the main difficulties resides in the lack of large publicly available Re-ID datasets for training #OTHEREFR , mainly be-cause manually labeling images is a very difficulty task, especially in open-world scenarios #OTHEREFR , where the datasets may increase dynamically.", "In order to mitigate this problem, many authors have proposed multi-source training, which consists in joining multiple available datasets for training #OTHEREFR , usually increasing the network capacity for generalization.", "However, despite several advances in multi-source approaches, it is inevitable the need for intensive manual annotation for obtaining training data."], "citing_paper_content": {"title": "Person Re-Id Through Unsupervised Hypergraph Rank Selection And Fusion", "abstract": "Person Re-ID has been gaining a lot of attention and nowadays is of fundamental importance in many camera surveillance applications. The task consists of identifying individuals across multiple cameras that have no overlapping views. Most of the approaches require labeled data, which is not always available, given the huge amount of demanded data and the difficulty of manually assigning a class for each individual. Recently, studies have shown that re-ranking methods are capable of achieving significant gains, especially in the absence of labeled data. Besides that, the fusion of feature extractors and multiple-source training is another promising research direction not extensively exploited. We aim to fill this gap through a manifold rank aggregation approach capable of exploiting the complementarity of different person Re-ID rankers. In this work, we perform a completely unsupervised selection and fusion of diverse ranked lists obtained from multiple and diverse feature extractors. Among the contributions, this work proposes a query performance prediction measure that models the relationship among images considering a hypergraph structure and does not require the use of any labeled data. Expressive gains were obtained in four datasets commonly used for person Re-ID. We achieved results competitive to the state-of-the-art in most of the scenarios."}, "cited_paper_content": {"title": "A Systematic Evaluation And Benchmark For Person Re-Identification: Features, Metrics, And Datasets", "abstract": "Person re-identification (re-id) is a critical problem in video analytics applications such as security and surveillance. The public release of several datasets and code for vision algorithms has facilitated rapid progress in this area over the last few years. However, directly comparing re-id algorithms reported in the literature has become difficult since a wide variety of features, experimental protocols, and evaluation metrics are employed. In order to address this need, we present an extensive review and performance evaluation of single- and multi-shot re-id algorithms. The experimental protocol incorporates the most recent advances in both feature extraction and metric learning. To ensure a fair comparison, all of the approaches were implemented using a unified code library that includes 11 feature extraction algorithms and 22 metric learning and ranking techniques. All approaches were evaluated using a new large-scale dataset that closely mimics a real-world problem setting, in addition to 16 other publicly available datasets: VIPeR, GRID, CAVIAR, DukeMTMC4ReID, 3DPeS, PRID, V47, WARD, SAIVT-SoftBio, CUHK01, CHUK02, CUHK03, RAiD, iLIDSVID, HDA+ and Market1501. The evaluation codebase and results will be made publicly available for community use."}, "keywords": ["metric learning approaches"], "citation_intent": "background"}
{"citing_id": "2303.15896v1", "cited_id": "1903.03040", "section_title": "Introduction", "citation": "Generative design and machine learning (ML) methods are increasingly being established in fluid dynamics, for example in #REFR , helping engineers to better understand their problem domains earlier on in the design process, and increase the likelihood of making the right design decisions.", "text_before_citation": ["(2021) showed that the flow around 3D cylindrical objects can be predicted based on spanwise-averaged Navier-Stokes equations.", "However, it is unclear how their method handles the wide diversity of shapes in early design stages.", "To perform the computational fluid dynamics (CFD) simulations, the lattice Boltzmann method (LBM) has made considerable progress in the prediction of turbulent flows, including direct numerical simulation (DNS), as in #OTHEREFR , under-resolved DNS, as in #OTHEREFR , and large eddy simulations (LES), as in #OTHEREFR .", "The LBM is renowned for its high performance, shown in #OTHEREFR , although 3D simulations remain a costly endeavor.", "This increases the demand for efficiency of QD methods applied in these kinds of domains even more."], "text_after_citation": ["These methods generally need large data sets with a wide variety of shapes and flow patterns.", "Although novel methods like physics-informed neural networks require fewer data points in some cases, for example in work by #OTHEREFR", "(2020) , it is unclear how they perform on the wide variety of shapes in early design stages.", "The increased integration of ML in the design process is pushing methods in fluid dynamics and optimization to their limits, due to the requirements on the amount and variation of training data points.", "Especially when data sets are unavailable for a certain problem, they need to be created first."], "citing_paper_content": {"title": "Efficient Quality Diversity Optimization Of 3D Buildings Through 2D Pre-Optimization A Preprint", "abstract": "Quality diversity algorithms can be used to efficiently create a diverse set of solutions to inform engineers' intuition. But quality diversity is not efficient in very expensive problems, needing 100.000s of evaluations. Even with the assistance of surrogate models, quality diversity needs 100s or even 1000s of evaluations, which can make it use infeasible. In this study we try to tackle this problem by using a pre-optimization strategy on a lowerdimensional optimization problem and then map the solutions to a higher-dimensional case. For a use case to design buildings that minimize wind nuisance, we show that we can predict flow features around 3D buildings from 2D flow features around building footprints. For a diverse set of building designs, by sampling the space of 2D footprints with a quality diversity algorithm, a predictive model can be trained that is more accurate than when trained on a set of footprints that were selected with a space-filling algorithm like the Sobol sequence. Simulating only 16 buildings in 3D, a set of 1024 building designs * This is the final version and has been accepted for publication in Evolutionary Computation (MIT Press)"}, "cited_paper_content": {"title": "Deep Learning Observables In Computational Fluid Dynamics", "abstract": "Many large scale problems in computational fluid dynamics such as uncertainty quantification, Bayesian inversion, data assimilation and PDE constrained optimization are considered very challenging computationally as they require a large number of expensive (forward) numerical solutions of the corresponding PDEs. We propose a machine learning algorithm, based on deep artificial neural networks, that learns the underlying input parameters to observable map from a few training samples (computed realizations of this map). By a judicious combination of theoretical arguments and empirical observations, we find suitable network architectures and training hyperparameters that result in robust and efficient neural network approximations of the parameters to observable map. Numerical experiments for realistic high dimensional test problems, demonstrate that even with approximately 100 training samples, the resulting neural networks have a prediction error of less than one to two percent, at a computational cost which is several orders of magnitude lower than the cost of the underlying PDE solver. ::: Moreover, we combine the proposed deep learning algorithm with Monte Carlo (MC) and Quasi-Monte Carlo (QMC) methods to efficiently compute uncertainty propagation for nonlinear PDEs. Under the assumption that the underlying neural networks generalize well, we prove that the deep learning MC and QMC algorithms are guaranteed to be faster than the baseline (quasi-) Monte Carlo methods. Numerical experiments demonstrating one to two orders of magnitude speed up over baseline QMC and MC algorithms, for the intricate problem of computing probability distributions of the observable, are also presented."}, "keywords": ["Generative design", "fluid dynamics"], "citation_intent": "method"}
{"citing_id": "2304.10588v1", "cited_id": "1506.01497", "section_title": "B. Annotation Pipiline", "citation": "The fixation object is automatically annotated by a pre-trained object detector #REFR with manual corrections.", "text_before_citation": ["To achieve multimodal attention detection, we developed a data annotation pipeline for efficient processing of the multimodal data.", "First, we recognized a list of significant primitives in human attention study as detailed in Table  I .", "Specifically, the eye state, gaze position, and fixation position can be extracted from Pupil Core's software system outputs but require careful calibration as errors are common in default detectors."], "text_after_citation": ["Head pose and surface tracking are derived with the aid of eight Apriltags deployed next to the assembly platform.", "Of our greatest interest, the distracted states and in-operation states are manually labelled frame-to-frame to identify subtle attention lapses.", "An example of the annotated sequence 002 is shown in Fig. 3 .", "Note ensuring ad-hoc occurrence of real-world distractors in an assembly task is critical for attention lapse study.", "We curated two pilot data sequences of a human participant assembling a dummy phone in a realistic production laboratory with manufacturing background noises, moving MiR robots, and workers around. According to K. Holmqvist and R."], "citing_paper_content": {"title": "Detecting Worker Attention Lapses In Human-Robot Interaction: An Eye Tracking And Multimodal Sensing Study", "abstract": "The advent of industrial robotics and autonomous systems endow human-robot collaboration in a massive scale. However, current industrial robots are restrained in co-working with human in close proximity due to inability of interpreting human agents' attention. Human attention study is non-trivial since it involves multiple aspects of the mind: perception, memory, problem solving, and consciousness. Human attention lapses are particularly problematic and potentially catastrophic in industrial workplace, from assembling electronics to operating machines. Attention is indeed complex and cannot be easily measured with single-modality sensors. Eye state, head pose, posture, and manifold environment stimulus could all play a part in attention lapses. To this end, we propose a pipeline to annotate multimodal dataset of human attention tracking, including eye tracking, fixation detection, third-person surveillance camera, and sound. We produce a pilot dataset containing two fully annotated phone assembly sequences in a realistic manufacturing environment. We evaluate existing fatigue and drowsiness prediction methods for attention lapse detection. Experimental results show that human attention lapses in production scenarios are more subtle and imperceptible than well-studied fatigue and drowsiness."}, "cited_paper_content": {"title": "Faster R-Cnn: Towards Real-Time Object Detection With Region Proposal Networks", "abstract": "State-of-the-art object detection networks depend on region proposal algorithms to hypothesize object locations. Advances like SPPnet [1] and Fast R-CNN [2] have reduced the running time of these detection networks, exposing region proposal computation as a bottleneck. In this work, we introduce a Region Proposal Network (RPN) that shares full-image convolutional features with the detection network, thus enabling nearly cost-free region proposals. An RPN is a fully convolutional network that simultaneously predicts object bounds and objectness scores at each position. The RPN is trained end-to-end to generate high-quality region proposals, which are used by Fast R-CNN for detection. We further merge RPN and Fast R-CNN into a single network by sharing their convolutional features\u2014using the recently popular terminology of neural networks with \u2019attention\u2019 mechanisms, the RPN component tells the unified network where to look. For the very deep VGG-16 model [3] , our detection system has a frame rate of 5 fps ( including all steps ) on a GPU, while achieving state-of-the-art object detection accuracy on PASCAL VOC 2007, 2012, and MS COCO datasets with only 300 proposals per image. In ILSVRC and COCO 2015 competitions, Faster R-CNN and RPN are the foundations of the 1st-place winning entries in several tracks. Code has been made publicly available."}, "keywords": ["fixation object", "pre-trained object detector"], "citation_intent": "method"}
{"citing_id": "2303.18044v1", "cited_id": "1801.04264", "section_title": "A. Datasets And Evaluation Metrics", "citation": "UCF-Crime #REFR contains 1,900 untrimmed videos with 13 different types of anomalies, among which 1,610 videos are for training and 290 videos are for the test.", "text_before_citation": ["Datasets.", "We evaluate the proposed method on three public datasets: ShanghaiTech #OTHEREFR , #OTHEREFR , UCF-Crime #OTHEREFR , and UBnormal #OTHEREFR .", "ShanghaiTech #OTHEREFR consists of 437 videos with 130 abnormal events in 13 scenes. In the original dataset, all training videos are normal.", "#OTHEREFR reorganized the dataset by selecting some abnormal testing videos into the training set for WS-VAD. We follow this new split way in our experiments."], "text_after_citation": ["UBnormal #OTHEREFR is a new large-scale dataset synthesized by the Cinema4D software.", "It consists of 543 videos with 22 types of anomalies on 29 virtual scenes, including 268 videos for training, 64 videos for validation, and 211 videos for test.", "Since it has a higher number of anomaly types occurring across a larger set of scenes, this dataset is more challenging for anomaly detection.", "Evaluation Metrics.", "Following the common practice #OTHEREFR , #OTHEREFR , we take the area under the curve (AUC) of the framelevel receiver operating characteristics (ROC) for evaluation. A higher AUC indicates better performance."], "citing_paper_content": {"title": "Long-Short Temporal Co-Teaching For Weakly Supervised Video Anomaly Detection", "abstract": "Weakly supervised video anomaly detection (WS-VAD) is a challenging problem that aims to learn VAD models only with video-level annotations. In this work, we propose a Long-Short Temporal Co-teaching (LSTC) method to address the WS-VAD problem. It constructs two tubelet-based spatiotemporal transformer networks to learn from short-and longterm video clips respectively. Each network is trained with respect to a multiple instance learning (MIL)-based ranking loss, together with a cross-entropy loss when clip-level pseudo labels are available. A co-teaching strategy is adopted to train the two networks. That is, clip-level pseudo labels generated from each network are used to supervise the other one at the next training round, and the two networks are learned alternatively and iteratively. Our proposed method is able to better deal with the anomalies with varying durations as well as subtle anomalies. Extensive experiments on three public datasets demonstrate that our method outperforms state-of-the-art WS-VAD methods."}, "cited_paper_content": {"title": "Real-World Anomaly Detection In Surveillance Videos", "abstract": "Surveillance videos are able to capture a variety of realistic anomalies. In this paper, we propose to learn anomalies by exploiting both normal and anomalous videos. To avoid annotating the anomalous segments or clips in training videos, which is very time consuming, we propose to learn anomaly through the deep multiple instance ranking framework by leveraging weakly labeled training videos, i.e. the training labels (anomalous or normal) are at video-level instead of clip-level. In our approach, we consider normal and anomalous videos as bags and video segments as instances in multiple instance learning (MIL), and automatically learn a deep anomaly ranking model that predicts high anomaly scores for anomalous video segments. Furthermore, we introduce sparsity and temporal smoothness constraints in the ranking loss function to better localize anomaly during training. We also introduce a new large-scale first of its kind dataset of 128 hours of videos. It consists of 1900 long and untrimmed real-world surveillance videos, with 13 realistic anomalies such as fighting, road accident, burglary, robbery, etc. as well as normal activities. This dataset can be used for two tasks. First, general anomaly detection considering all anomalies in one group and all normal activities in another group. Second, for recognizing each of 13 anomalous activities. Our experimental results show that our MIL method for anomaly detection achieves significant improvement on anomaly detection performance as compared to the state-of-the-art approaches. We provide the results of several recent deep learning baselines on anomalous activity recognition. The low recognition performance of these baselines reveals that our dataset is very challenging and opens more opportunities for future work."}, "keywords": ["1,900 untrimmed videos"], "citation_intent": "background"}
{"citing_id": "2303.02214v1", "cited_id": "1605.07277", "section_title": "D", "citation": "These adversarial attacks have also been successfully transferred between models designed to perform similar tasks, even if their architectures or training sets are different #REFR .", "text_before_citation": ["One example is the reconstruction of the faces used to train a facial recognition system #OTHEREFR .", "Model extraction exploratory attacks seek to use probing to construct a replica of the victim model.", "The replica can then be used to stage further attacks.", "In #OTHEREFR , the authors demonstrate that they can accomplish the same black-box model inversion attack performed in #OTHEREFR with fewer queries by extracting the victim model and then performing a white-box inversion attack on the replica.", "In summary, evasion, poisoning, backdoor, and exploratory attacks (illustrated in Fig. 1 ) are major challenges for ML."], "text_after_citation": ["This property dramatically complicates the challenge of defending ML models deployed in embedded systems against adversarial attacks, enabling scenarios in which exploratory attacks are used to extract the victim model and generate a surrogate model against which an evasion, poisoning, or backdoor attack can be crafted and then transferred back to the victim model."], "citing_paper_content": {"title": "Adversarial Attacks On Machine Learning In Embedded And Iot Platforms", "abstract": "Machine learning (ML) algorithms are increasingly being integrated into embedded and IoT systems that surround us, and they are vulnerable to adversarial attacks. The deployment of these ML algorithms on resource-limited embedded platforms also requires the use of model compression techniques. The impact of such model compression techniques on adversarial robustness in ML is an important and emerging area of research. This article provides an overview of the landscape of adversarial attacks and ML model compression techniques relevant to embedded systems. We then describe efforts that seek to understand the relationship between adversarial attacks and ML model compression before discussing open problems in this area."}, "cited_paper_content": {"title": "Transferability In Machine Learning: From Phenomena To Black-Box Attacks Using Adversarial Samples", "abstract": "Many machine learning models are vulnerable to adversarial examples: inputs that are specially crafted to cause a machine learning model to produce an incorrect output. Adversarial examples that affect one model often affect another model, even if the two models have different architectures or were trained on different training sets, so long as both models were trained to perform the same task. An attacker may therefore train their own substitute model, craft adversarial examples against the substitute, and transfer them to a victim model, with very little information about the victim. Recent work has further developed a technique that uses the victim model as an oracle to label a synthetic training set for the substitute, so the attacker need not even collect a training set to mount the attack. We extend these recent techniques using reservoir sampling to greatly enhance the efficiency of the training procedure for the substitute model. We introduce new transferability attacks between previously unexplored (substitute, victim) pairs of machine learning model classes, most notably SVMs and decision trees. We demonstrate our attacks on two commercial machine learning classification systems from Amazon (96.19% misclassification rate) and Google (88.94%) using only 800 queries of the victim model, thereby showing that existing machine learning approaches are in general vulnerable to systematic black-box attacks regardless of their structure."}, "keywords": ["adversarial attacks"], "citation_intent": "background"}
{"citing_id": "2303.11620v2", "cited_id": "1011.5553", "section_title": "Definition 2.", "citation": "The representation due to a perfect alignment is called a realization of the framework #REFR .", "text_before_citation": ["The consensus representation of the framework \u0398 due to an alignment S is given by \u0398(S) := H * (:, 1 :", "n) = S T BL \u2020 \u0393 (:, 1 : n) I n \u2212 n \u22121 1 n 1 T n ."], "text_after_citation": [], "citing_paper_content": {"title": "Non-Degenerate Rigid Alignment In A Patch Framework", "abstract": "Given a set of overlapping local views (patches) of a dataset, we consider the problem of finding a rigid alignment of the views that minimizes a 2-norm based alignment error. In general, the views are noisy and a perfect alignment may not exist. In this work, we characterize the non-degeneracy of an alignment in the noisy setting based on the kernel and positivity of a certain matrix. This leads to a polynomial time algorithm for testing the non-degeneracy of a given alignment. Consequently, we focus on Riemannian gradient descent for minimization of the error and obtain a sufficient condition on an alignment for the algorithm to converge (locally) linearly to it. In the case of noiseless views, a perfect alignment exists, resulting in a realization of the points that respects the geometry of the views. Under a mild condition on the views, we show that the non-degeneracy of a perfect alignment is equivalent to the local rigidity of the resulting realization. By specializing the characterization of a non-degenerate alignment to the noiseless setting, we obtain necessary and sufficient conditions on the overlapping structure of the views for a locally rigid realization. Similar results are also obtained in the context of global rigidity."}, "cited_paper_content": {"title": "On Affine Rigidity", "abstract": "We define the notion of affine rigidity of a hypergraph and prove a variety of fundamental results for this notion. First, we show that affine rigidity can be determined by the rank of a specific matrix which implies that affine rigidity is a generic property of the hypergraph.Then we prove that if a graph is is $(d+1)$-vertex-connected, then it must be \"generically neighborhood affinely rigid\" in $d$-dimensional space. This implies that if a graph is $(d+1)$-vertex-connected then any generic framework of its squared graph must be universally rigid. ::: Our results, and affine rigidity more generally, have natural applications in point registration and localization, as well as connections to manifold learning."}, "keywords": ["perfect alignment"], "citation_intent": "background"}
{"citing_id": "2303.00628v1", "cited_id": "1809.00496", "section_title": "Corpus Creation", "citation": "We acquire video tracks of the original recordings, and align processed video data with the audio one to form audio-visual data, similar to LRS3-TED #REFR .", "text_before_citation": ["To improve matching recall on the train set, we develop a fuzzy text matching strategy: we first segment TED2020 source and target sentences by punctuation if both sides of the sentence pair contain the same amount of segments.", "Then we normalize TED2020 and LRS3-TED texts by punctuation removal and lowercasing.", "Finally, we conduct exact text matching between the two corpora.", "For LRS3-TED train set examples without a match from TED2020, we acquire pseudotranslation labels from a machine translation model, M2M-100 418M #OTHEREFR with default decoding hyper-parameters.", "For non-English talks, we reuse the audio-only data, transcriptions and text translations collected by mTEDx #OTHEREFR . Our data split also follows mTEDx."], "text_after_citation": ["Although all the audio data in mTEDx is transcribed, only a subset of it is translated.", "We acquire pseudo-translation labels from M2M-100 418M #OTHEREFR for the untranslated train set examples with default decoding hyper-parameters.", "In Table 1 , we provide duration and speaker statistics for each data split in MuAViC.", "MuAViC covers 9 languages providing total 1200 hours of transcribed audio-visual data from more than 8000 speakers (number of speakers estimated based on TED/TEDx talk IDs).", "It also provides text translations into or out of English for 6 of the 9 languages."], "citing_paper_content": {"title": "Muavic: A Multilingual Audio-Visual Corpus For Robust Speech Recognition And Robust Speech-To-Text Translation", "abstract": "We introduce MuAViC, a multilingual audiovisual corpus for robust speech recognition and robust speech-to-text translation providing 1200 hours of audiovisual speech in 9 languages. It is fully transcribed and covers 6 English-to-X translation as well as 6 X-to-English translation directions. To the best of our knowledge, this is the first open benchmark for audiovisual speech-to-text translation and the largest open benchmark for multilingual audiovisual speech recognition. Our baseline results show that MuAViC is effective for building noise-robust speech recognition and translation models. We make the corpus available at https://github.com/ facebookresearch/muavic."}, "cited_paper_content": {"title": "Lrs3-Ted: A Large-Scale Dataset For Visual Speech Recognition", "abstract": "This paper introduces a new multi-modal dataset for visual and audio-visual speech recognition. It includes face tracks from over 400 hours of TED and TEDx videos, along with the corresponding subtitles and word alignment boundaries. The new dataset is substantially larger in scale compared to other public datasets that are available for general research."}, "keywords": ["audio-visual data"], "citation_intent": "method"}
{"citing_id": "2303.12359v1", "cited_id": "1703.01909", "section_title": "E. Calibration", "citation": "Nonetheless, the lack of a perfect weight calibration can be circumvented via inthe-loop training on the BrainScaleS-1 system, as shown for inference tasks in previous results #REFR .", "text_before_citation": ["12d ; this renders the search for a more precise fit function unbeneficial.", "In addition, the per wafer calibration opted over a per neuron circuit calibration introduces a dominant error due to the deviations between neuron circuits, shown in fig. 12e .", "A precise weight calibration within a reasonable runtime would be achievable via a parallel measurement of each neuron circuit.", "This would also allow to exclude neurons showing unintended behavior.", "However, this is not possible with the currently used analog readout system."], "text_after_citation": ["3) Calibration Based Exclusion of Components: The operation of the HICANNs during the calibration is similar to the operation during experiments.", "All components have to work correctly for the calibration to succeed. Failing calibrations indicate unintended behavior.", "This allows for testing the whole die, especially the analog circuits that cannot be tested directly. Additionally, thresholds can be defined to exclude outliers.", "Consequently, neurons that do not pass all calibration steps are excluded from the availability database.", "Numbers of calibration based excluded neurons on a typical wafer are given in table II."], "citing_paper_content": {"title": "From Clean Room To Machine Room: Commissioning Of The First-Generation Brainscales Wafer-Scale Neuromorphic System", "abstract": "The first-generation of BrainScaleS, also referred to as BrainScaleS-1, is a neuromorphic system for emulating large-scale networks of spiking neurons. Following a \"physical modeling\" principle, its VLSI circuits are designed to emulate the dynamics of biological examples: analog circuits implement neurons and synapses with time constants that arise from their electronic components' intrinsic properties. It operates in continuous time, with dynamics typically matching an acceleration factor of 10 000 compared to the biological regime. A fault-tolerant design allows it to achieve wafer-scale integration despite unavoidable analog variability and component failures. In this paper, we present the commissioning process of a BrainScaleS-1 wafer module, providing a short description of the system's physical components, illustrating the steps taken during its assembly and the measures taken to operate it. Furthermore, we reflect on the system's development process and the lessons learned to conclude with a demonstration of its functionality by emulating a wafer-scale synchronous firing chain, the largest spiking network emulation ran with analog components and individual synapses to date. Index Terms-Neuromorphic hardware, wafer-scale integration, spiking neural networks, emulated networks, analog neuromorphic devices, synfire chains."}, "cited_paper_content": {"title": "Neuromorphic Hardware In The Loop: Training A Deep Spiking Network On The Brainscales Wafer-Scale System", "abstract": "Emulating spiking neural networks on analog neuromorphic hardware offers several advantages over simulating them on conventional computers, particularly in terms of speed and energy consumption. However, this usually comes at the cost of reduced control over the dynamics of the emulated networks. In this paper, we demonstrate how iterative training of a hardware-emulated network can compensate for anomalies induced by the analog substrate. We first convert a deep neural network trained in software to a spiking network on the BrainScaleS wafer-scale neuromorphic system, thereby enabling an acceleration factor of 10000 compared to the biological time domain. This mapping is followed by the in-the-loop training, where in each training step, the network activity is first recorded in hardware and then used to compute the parameter updates in software via backpropagation. An essential finding is that the parameter updates do not have to be precise, but only need to approximately follow the correct gradient, which simplifies the computation of updates. Using this approach, after only several tens of iterations, the spiking network shows an accuracy close to the ideal software-emulated prototype. The presented techniques show that deep spiking networks emulated on analog neuromorphic devices can attain good computational performance despite the inherent variations of the analog substrate."}, "keywords": ["BrainScaleS-1 system", "inthe-loop training"], "citation_intent": "method"}
{"citing_id": "2303.00250v1", "cited_id": "1802.03801", "section_title": "Realization Of Slack Federated Adversarial Training", "citation": "However, some relaxation on the assumptions could be further improved by considering more practical SGD theories #REFR , which is beyond the scope of this paper and we leave in the future works.", "text_before_citation": ["In this case, the increment of the first term of B can be totally counteracted by the loss of the second term of B so that in sum B becomes smaller.", "Then, we can have a tighter upper bound for SFAT in Eq. (4) to achieve a faster convergence than FAT.", "The completed proof of Theorem 4.3 is given in Appendix B.3 and the empirical verification is presented in Appendix E.7.", "The following section will comprehensively confirm that SFAT can reach to a more robust optimum.", "Note that, the theorem we deduce here follows the same assupmitions #OTHEREFR and is mainly to compare the difference with the early convergence analysis in federated learning."], "text_after_citation": [], "citing_paper_content": {"title": "Combating Exacerbated Heterogeneity For Robust Models In Federated Learning", "abstract": "Privacy and security concerns in real-world applications have led to the development of adversarially robust federated models. However, the straightforward combination between adversarial training and federated learning in one framework can lead to the undesired robustness deterioration. We discover that the attribution behind this phenomenon is that the generated adversarial data could exacerbate the data heterogeneity among local clients, making the wrapped federated learning perform poorly. To deal with this problem, we propose a novel framework called Slack Federated Adversarial Training (SFAT), assigning the client-wise slack during aggregation to combat the intensified heterogeneity. Theoretically, we analyze the convergence of the proposed method to properly relax the objective when combining federated learning and adversarial training. Experimentally, we verify the rationality and effectiveness of SFAT on various benchmarked and real-world datasets with different adversarial training and federated optimization methods."}, "cited_paper_content": {"title": "Sgd And Hogwild! Convergence Without The Bounded Gradients Assumption", "abstract": "Stochastic gradient descent (SGD) is the optimization algorithm of choice in many machine learning applications such as regularized empirical risk minimization and training deep neural networks. The classical convergence analysis of SGD is carried out under the assumption that the norm of the stochastic gradient is uniformly bounded. While this might hold for some loss functions, it is always violated for cases where the objective function is strongly convex. In (Bottou et al.,2016), a new analysis of convergence of SGD is performed under the assumption that stochastic gradients are bounded with respect to the true gradient norm. Here we show that for stochastic problems arising in machine learning such bound always holds; and we also propose an alternative convergence analysis of SGD with diminishing learning rate regime, which results in more relaxed conditions than those in (Bottou et al.,2016). We then move on the asynchronous parallel setting, and prove convergence of Hogwild! algorithm in the same regime, obtaining the first convergence results for this method in the case of diminished learning rate."}, "keywords": ["practical SGD theories"], "citation_intent": "background"}
{"citing_id": "2305.02423v1", "cited_id": "1905.00537", "section_title": "Experimental Setup", "citation": "In fully-supervised setting, the full training set of each task in Super-GLUE #REFR is employed during the prompt tuning process.", "text_before_citation": ["NLU Dataset.", "SuperGLUE benchmark #OTHEREFR contains 8 challenging natural language understanding (NLU) tasks.", "We also include the few-shot version of SuperGLUE, FewGLUE benchmark #OTHEREFR to test the ability of our algorithm, which consists of 32 training samples in each dataset on SuperGLUE.", "Following #OTHEREFR Liu et al., 2021b) , we report results on 7 of 8 NLU tasks in few-shot settings.", "Fully & Few-shot Setting."], "text_after_citation": ["Besides, in the model selection part, we adopt the whole validation set.", "As few-shot learning ability of prompt tuning can reduce the cost of annotations in real-world applications, following #OTHEREFR Liu et al., 2021b) , we also test our algorithm under fewshot settings.", "To be specific, we use the training set provided by FewGLUE #OTHEREFR , the few-shot version of SuperGLUE, containing 32 training pairs in each task.", "Besides, we use the same version of the development set as (Liu et al., 2021b) to select models, which are created by randomly choosing 32 unused training pairs.", "Baseline Methods."], "citing_paper_content": {"title": "Ptp: Boosting Stability And Performance Of Prompt Tuning With Perturbation-Based Regularizer", "abstract": "Recent studies show that prompt tuning can better leverage the power of large language models than fine-tuning on downstream natural language understanding tasks. However, the existing prompt tuning methods have training instability issues, as the variance of scores under different random seeds is quite large. To address this critical problem, we first investigate and find that the loss landscape of vanilla prompt tuning is precipitous when it is visualized, where a slight change of input data can cause a big fluctuation in the loss landscape. This is an essential factor that leads to the instability of prompt tuning. Based on this observation, we introduce perturbation-based regularizers, which can smooth the loss landscape, into prompt tuning. We propose a new algorithm, called Prompt Tuning with Perturbationbased regularizer (PTP), which can not only alleviate training instability dramatically but also boost the performance of prompt tuning. We design two kinds of perturbation-based regularizers, including random-noise-based and adversarial-based. In particular, our proposed perturbations are flexible on both text space and embedding space. Extensive experiments show the effectiveness of our proposed methods in stabilizing the training. Our new algorithms improve the state-of-the-art prompt tuning methods by 1.94% and 2.34% on Su-perGLUE and FewGLUE benchmarks, respectively."}, "cited_paper_content": {"title": "Superglue: A Stickier Benchmark For General-Purpose Language Understanding Systems", "abstract": "In the last year, new models and methods for pretraining and transfer learning have driven striking performance improvements across a range of language understanding tasks. The GLUE benchmark, introduced a little over one year ago, offers a single-number metric that summarizes progress on a diverse set of such tasks, but performance on the benchmark has recently surpassed the level of non-expert humans, suggesting limited headroom for further research. In this paper we present SuperGLUE, a new benchmark styled after GLUE with a new set of more difficult language understanding tasks, a software toolkit, and a public leaderboard. SuperGLUE is available at https://super.gluebenchmark.com."}, "keywords": ["prompt tuning process", "fully-supervised setting"], "citation_intent": "method"}
{"citing_id": "2305.00250v1", "cited_id": "1410.1275", "section_title": "Tests With", "citation": "As shown in equation #REFR , the index function is convergent as the radius R tends to infinity, which explains the new configuration can still achieve good reconstructions. Based on the results shown in Fig.", "text_before_citation": ["Latin letters \"D\", \"S\" and \"M\" with high noise levels", "In this example, we use the network trained by the MNIST dataset with N i = 16 to reconstruct Latin letters \"D\", \"S\" and \"M\" with high noise levels 50% and 100%.", "In addition, different from the configuration for the training data, we use 60 receivers equally distributed on a circular curve with a radius 5 centered at the origin to collect the data for the scattered field."], "text_after_citation": ["12 , it can be seen that the proposed DSM-DL is very robust to high-level noise as the noise is smoothed in the DSM process.", "We can also observe that the reconstruction of the thinner part of the letters is more likely to be distorted, which is mainly because the thinner part contributes less to the scattered field."], "citing_paper_content": {"title": "A Direct Sampling-Based Deep Learning Approach For Inverse Medium Scattering Problems", "abstract": "In this work, we focus on the inverse medium scattering problem (IMSP), which aims to recover unknown scatterers based on measured scattered data. Motivated by the efficient direct sampling method (DSM) introduced in [23], we propose a novel direct sampling-based deep learning approach (DSM-DL) for reconstructing inhomogeneous scatterers. In particular, we use the U-Net neural network to learn the relation between the index functions and the true contrasts. Our proposed DSM-DL is computationally efficient, robust to noise, easy to implement, and able to naturally incorporate multiple measured data to achieve high-quality reconstructions. Some representative tests are carried out with varying numbers of incident waves and different noise levels to evaluate the performance of the proposed method. The results demonstrate the promising benefits of combining deep learning techniques with the DSM for IMSP."}, "cited_paper_content": {"title": "Direct Sampling Method For Diffusive Optical Tomography", "abstract": "In this work, we are concerned with the diffusive optical tomography (DOT) problem in the case when only one or two pairs of Cauchy data are available. We propose a simple and efficient direct sampling method to locate inhomogeneities inside a homogeneous background and solve the DOT problem in both full and limited aperture cases. This new method is easy to implement and less expensive computationally. Numerical experiments demonstrate its effectiveness and robustness against noise in the data. This provides a new promising numerical strategy for the DOT problem."}, "keywords": ["good reconstructions"], "citation_intent": "result"}
{"citing_id": "2304.00736v1", "cited_id": "1901.00596", "section_title": "C. Gnn For In-Hand Manipulation", "citation": "Recent research from molecular biology, to multi-robot path planning #REFR have demonstrated promising performance of GNNs to capture the interaction between a node in the graph for node and edge classification and message passing.", "text_before_citation": [], "text_after_citation": ["This property yields the rising research in capturing tactile sensor data using GNN #OTHEREFR , #OTHEREFR for in-hand manipulation.", "The properties of GNNs in capturing inter-node reactions inspired researchers to use GNN as a control algorithm.", "The strength of GNNs has been recently utilized in tactile sensing. Garcia-Garcia et al.", "#OTHEREFR proposed a GCN-based approach to capture taxel reading from BioTac SP taxel to binary classify grasps as stable or slippery ones. Funabashi et al.", "#OTHEREFR proposed a GCN-based pipeline to extract geodesical features from the tactile data from distributed taxels."], "citing_paper_content": {"title": "Tacgnn: Learning Tactile-Based In-Hand Manipulation With A Blind Robot Using Hierarchical Graph Neural Network", "abstract": "In this paper, we propose a novel framework for tactile-based dexterous manipulation learning with a blind anthropomorphic robotic hand, i.e. without visual sensing. First, object-related states were extracted from the raw tactile signals by a graph-based perception model-TacGNN. The resulting tactile features were then utilized in the policy learning of an in-hand manipulation task in the second stage. This method was examined by a Baoding ball task-simultaneously manipulating two spheres around each other by 180 degrees in hand. We conducted experiments on object states prediction and in-hand manipulation using a reinforcement learning algorithm (PPO). Results show that TacGNN is effective in predicting objectrelated states during manipulation by decreasing the RMSE of prediction to 0.096cm comparing to other methods, such as MLP, CNN, and GCN. Finally, the robot hand could finish an in-hand manipulation task solely relying on the robotic own perception-tactile sensing and proprioception. In addition, our methods are tested on three tasks with different difficulty levels and transferred to the real robot without further training. https://sites.google.com/view/tacgnn"}, "cited_paper_content": {"title": "A Comprehensive Survey On Graph Neural Networks", "abstract": "Deep learning has revolutionized many machine learning tasks in recent years, ranging from image classification and video processing to speech recognition and natural language understanding. The data in these tasks are typically represented in the Euclidean space. However, there is an increasing number of applications, where data are generated from non-Euclidean domains and are represented as graphs with complex relationships and interdependency between objects. The complexity of graph data has imposed significant challenges on the existing machine learning algorithms. Recently, many studies on extending deep learning approaches for graph data have emerged. In this article, we provide a comprehensive overview of graph neural networks (GNNs) in data mining and machine learning fields. We propose a new taxonomy to divide the state-of-the-art GNNs into four categories, namely, recurrent GNNs, convolutional GNNs, graph autoencoders, and spatial-temporal GNNs. We further discuss the applications of GNNs across various domains and summarize the open-source codes, benchmark data sets, and model evaluation of GNNs. Finally, we propose potential research directions in this rapidly growing field."}, "keywords": ["graph"], "citation_intent": "background"}
{"citing_id": "2303.09252v1", "cited_id": "1908.03195", "section_title": "Ablation Studies", "citation": "Comparison of different backbones and alignment methods on LVIS v1.0 #REFR with close-set settings.", "text_before_citation": ["We verify the effectiveness of grid-level and image-level alignment for undersampled categories for both close-set detection (Table 4) and open-set detection ( Table 5 ). We follow the same settings in Sec.", "4.2 except that multi-scaling training and random cropping augmentation are not used here.", "Among the experiments, \"wo align\" denotes using the original design of FCOS only with different backbones Table 4 ."], "text_after_citation": ["The top section compares supervised (ImageNet #OTHEREFR ) and unsupervised (SwAV #OTHEREFR ) pretrained visual models with CLIP as unsupervised visual-language pretrained model.", "The bottom section compares different alignment methods based on the ResNet50 version of pretrained CLIP image encoder.", "\"GridCLIP-R50 * \" uses the CLIP ResNet50 without the MHSA layer, while \"GridCLIP-R50\" uses the whole image encoder of CLIP ResNet50. Figure 3 .", "The AP on LVIS v1.0 over categories sorted by frequency in ascending order.", "(a) uses on the close-set setting, only containing 461 \"common\" categories and 405 \"frequent\" categories."], "citing_paper_content": {"title": "Gridclip: One-Stage Object Detection By Grid-Level Clip Representation Learning", "abstract": "A vision-language foundation model pretrained on very large-scale image-text paired data has the potential to provide generalizable knowledge representation for downstream visual recognition and detection tasks, especially on supplementing the undersampled categories in downstream model training. Recent studies utilizing CLIP for object detection have shown that a two-stage detector design typically outperforms a one-stage detector, while requiring more expensive training resources and longer inference time. In this work, we propose a one-stage detector GridCLIP that narrows its performance gap to those of twostage detectors, with approximately 43\u00d7 and 5\u00d7 faster than its two-stage counterpart (ViLD) in the training and test process respectively. GridCLIP learns grid-level representations to adapt to the intrinsic principle of one-stage detection learning by expanding the conventional CLIP imagetext holistic mapping to a more fine-grained, grid-text alignment. This differs from the region-text mapping in two-stage detectors that apply CLIP directly by treating regions as images. Specifically, GridCLIP performs Grid-level Alignment to adapt the CLIP image-level representations to gridlevel representations by aligning to CLIP category representations to learn the annotated (especially frequent) categories. To learn generalizable visual representations of broader categories, especially undersampled ones, we perform Image-level Alignment during training to propagate broad pre-learned categories in the CLIP image encoder from the image-level to the grid-level representations. Experiments show that the learned CLIP-based grid-level representations boost the performance of undersampled (infrequent and novel) categories, reaching comparable detection performance on the LVIS benchmark."}, "cited_paper_content": {"title": "Lvis: A Dataset For Large Vocabulary Instance Segmentation", "abstract": "Progress on object detection is enabled by datasets that focus the research community\u2019s attention on open challenges. This process led us from simple images to complex scenes and from bounding boxes to segmentation masks. In this work, we introduce LVIS (pronounced \u2018el-vis\u2019): a new dataset for Large Vocabulary Instance Segmentation. We plan to collect 2.2 million high-quality instance segmentation masks for over 1000 entry-level object categories in 164k images. Due to the Zipfian distribution of categories in natural images, LVIS naturally has a long tail of categories with few training samples. Given that state-of-the-art deep learning methods for object detection perform poorly in the low-sample regime, we believe that our dataset poses an important and exciting new scientific challenge. LVIS is available at http://www.lvisdataset.org."}, "keywords": ["alignment methods", "different backbones"], "citation_intent": "method"}
{"citing_id": "2304.04336v1", "cited_id": "1911.06971", "section_title": "Related Work", "citation": "However, such classical approaches usually start without any initialization while SMART effectively utilizes shape-dependent initialization learned by neural networks (BSP-Net) #REFR .", "text_before_citation": ["#OTHEREFR 's work is based on visibility (line of sight), which is one of the volumetric properties. Shapira et al.", "#OTHEREFR define a shape diameter function that expresses the diameter of the object's volume in the nearby point on the surface for consistent mesh partitioning. Liu et al.", "#OTHEREFR design a part-aware surface metric that considers the volumetric context when encoding part information. Kaick et al.", "#OTHEREFR merge weakly convex components taking account of the volumetric profile of the parts to obtain the final segmentation.", "In the sense of exploiting volumetric information to capture part-level segmentation, SMART is similar to these approaches."], "text_after_citation": ["Moreover, we propose an additional step to refine the errors and failures of these approaches. Efficient Search for 3D Applications.", "Efficient search is crucial for tackling high-dimensional search spaces since we cannot exhaustively search all the possible solutions.", "Continuity and extra dimensionality in 3D data make the search space extremely large, requiring dedicated algorithms for efficient searches.", "For this purpose beam search #OTHEREFR , monte carlo tree search (MCTS) #OTHEREFR 36] and learning #OTHEREFR approaches have been explored to solve various 3D computer vision tasks such as shape reconstruction, scene detection and convex decomposition. Sharma et al.", "#OTHEREFR propose a neural shape parser that learns to generate CSG grammar to challenge the huge search spaces."], "citing_paper_content": {"title": "Split, Merge, And Refine: Fitting Tight Bounding Boxes Via Learned Over-Segmentation And Iterative Search", "abstract": "Figure 1: Our framework dubbed SMART finds a set of tight bounding boxes of a 3D shape. Given a 3D shape represented as a volumetric mesh, it first takes over-segmentation from the output of BSP-Net [8] and post-processing and then performs hierarchical merging with tightness-aware criteria (second row). Then, it refines the bounding box parameters with a sequence of actions in a discrete space (third row). The results are further improved with long-sighted exploration by our MCTS accelerated with our technique (last row)."}, "cited_paper_content": {"title": "Bsp-Net: Generating Compact Meshes Via Binary Space Partitioning", "abstract": "Polygonal meshes are ubiquitous in the digital 3D domain, yet they have only played a minor role in the deep learning revolution. Leading methods for learning generative models of shapes rely on implicit functions, and generate meshes only after expensive iso-surfacing routines. To overcome these challenges, we are inspired by a classical spatial data structure from computer graphics, Binary Space Partitioning (BSP), to facilitate 3D learning. The core ingredient of BSP is an operation for recursive subdivision of space to obtain convex sets. By exploiting this property, we devise BSP-Net, a network that learns to represent a 3D shape via convex decomposition. Importantly, BSP-Net is unsupervised since no convex shape decompositions are needed for training. The network is trained to reconstruct a shape using a set of convexes obtained from a BSP-tree built on a set of planes. The convexes inferred by BSP-Net can be easily extracted to form a polygon mesh, without any need for iso-surfacing. The generated meshes are compact (i.e., low-poly) and well suited to represent sharp geometry; they are guaranteed to be watertight and can be easily parameterized. We also show that the reconstruction quality by BSP-Net is competitive with state-of-the-art methods while using much fewer primitives. Code is available at https://github.com/czq142857/BSP-NET-original."}, "keywords": ["shape-dependent initialization"], "citation_intent": "background"}
{"citing_id": "2303.08225v1", "cited_id": "2003.06988", "section_title": "Node Classification-Based Discriminator", "citation": "Moreover, we observe that HouseGAN #REFR cannot produce very discriminative rooms, leading to similar generation results for different types of rooms.", "text_before_citation": ["Next, we reshape it to a tensor of size 8\u00d732\u00d732.", "Thus, we use a shared three-layer CNN to convert it to a feature of size 16\u00d732\u00d732, followed by two rounds of Conv-MPN and downsampling.", "Lastly, we use another three-layer CNN to convert each room feature into a 128-d vector \u2212 \u2192 d r .", "To classify ground-truth samples from the generated ones, we sum-pool over all the room vectors and then apply a single linear layer to produce a scalard 1 , which can be expressed as follows,", "EQUATION"], "text_after_citation": ["To provide a more diverse generation for different rooms, we propose a novel node classification loss to learn more discriminative classspecific node representations.", "Specifically, we sum pool over all the room vectors and add another single linear layer to output a 10-d one-hot vectord 2 , classifying generated rooms to the corresponding room labels,", "EQUATION", "We use the binary cross-entropy loss between the real room label d 2 and the predicted labeld 2 ."], "citing_paper_content": {"title": "Graph Transformer Gans For Graph-Constrained House Generation", "abstract": "We present a novel graph Transformer generative adversarial network (GTGAN) to learn effective graph node relations in an end-to-end fashion for the challenging graphconstrained house generation task. The proposed graph-Transformer-based generator includes a novel graph Transformer encoder that combines graph convolutions and selfattentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node attention (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. Moreover, we propose a new node classification-based discriminator to preserve the high-level semantic and discriminative node features for different house components. Finally, we propose a novel graph-based cycle-consistency loss that aims at maintaining the relative spatial relationships between ground truth and predicted graphs. Experiments on two challenging graph-constrained house generation tasks (i.e., house layout and roof generation) with two public datasets demonstrate the effectiveness of GTGAN in terms of objective quantitative scores and subjective visual realism. New state-of-the-art results are established by large margins on both tasks."}, "cited_paper_content": {"title": "House-Gan: Relational Generative Adversarial Networks For Graph-Constrained House Layout Generation", "abstract": "This paper proposes a novel graph-constrained generative adversarial network, whose generator and discriminator are built upon relational architecture. The main idea is to encode the constraint into the graph structure of its relational networks. We have demonstrated the proposed architecture for a new house layout generation problem, whose task is to take an architectural constraint as a graph (i.e., the number and types of rooms with their spatial adjacency) and produce a set of axis-aligned bounding boxes of rooms. We measure the quality of generated house layouts with the three metrics: the realism, the diversity, and the compatibility with the input graph constraint. Our qualitative and quantitative evaluations over 117,000 real floorplan images demonstrate that the proposed approach outperforms existing methods and baselines. We will publicly share all our code and data."}, "keywords": ["discriminative rooms"], "citation_intent": "result"}
{"citing_id": "2303.04077v1", "cited_id": "1901.03035", "section_title": "B.3. Baselines", "citation": "SMNA #REFR uses visual-textual co-grounding module that encodes the past instructions and the instructions and actions to be done.", "text_before_citation": ["Seq2Seq #OTHEREFR uses sequence-to-sequence action prediciton to generate actions from the agent trajectory.", "Speaker-Follower #OTHEREFR uses the speaker model to augment natural language instructions and evaluate the candidate action sequence.", "FAST #OTHEREFR uses both local and global signals to look forward the unobserved environment during exploration and backtrack to the originally visited nodes when needed."], "text_after_citation": ["SMNA also uses a progress monitor to estimate the current progress of the agent relative to the total instructions. Regretful-Agent #OTHEREFR improves SMNA via two modules.", "The regret module decides whether to continue to explore or rollback to previous state by a learned policy, and the progress marker decides the direction the agent should head to by selecting visited nodes with progress estimates.", "RCM #OTHEREFR applies reinforcement learning to enforce the global matching between the agent trajectory and the given natural language instruction.", "Via cycle-reconstruction reward, RCM allows the agent to comprehend the natural language instruction and penalize paths that do not match with the given instructions.", "FAST-MATTN #OTHEREFR introduces a Navigator-Pointer model to both navigate to the target point and to localize the object from the navigation point according to the language guidance."], "citing_paper_content": {"title": "Meta-Explore: Exploratory Hierarchical Vision-And-Language Navigation Using Scene Object Spectrum Grounding", "abstract": "The main challenge in vision-and-language navigation (VLN) is how to understand natural-language instructions in an unseen environment. The main limitation of conventional VLN algorithms is that if an action is mistaken, the agent fails to follow the instructions or explores unnecessary regions, leading the agent to an irrecoverable path. To tackle this problem, we propose Meta-Explore, a hierarchical navigation method deploying an exploitation policy to correct misled recent actions. We show that an exploitation policy, which moves the agent toward a well-chosen local goal among unvisited but observable states, outperforms a method which moves the agent to a previously visited state. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. The key to our approach is understanding the object placements around the agent in spectral-domain. Specifically, we present a novel visual representation, called scene object spectrum (SOS), which performs category-wise 2D Fourier transform of detected objects. Combining exploitation policy and SOS features, the agent can correct its path by choosing a promising local goal. We evaluate our method in three VLN benchmarks: R2R, SOON, and REVERIE. Meta-Explore outperforms other baselines and shows significant generalization performance. In addition, local goal search using the proposed spectral-domain SOS features significantly improves the success rate by 17."}, "cited_paper_content": {"title": "Self-Monitoring Navigation Agent Via Auxiliary Progress Estimation", "abstract": "The Vision-and-Language Navigation (VLN) task entails an agent following navigational instruction in photo-realistic unknown environments. This challenging task demands that the agent be aware of which instruction was completed, which instruction is needed next, which way to go, and its navigation progress towards the goal. In this paper, we introduce a self-monitoring agent with two complementary components: (1) visual-textual co-grounding module to locate the instruction completed in the past, the instruction required for the next action, and the next moving direction from surrounding images and (2) progress monitor to ensure the grounded instruction correctly reflects the navigation progress. We test our self-monitoring agent on a standard benchmark and analyze our proposed approach through a series of ablation studies that elucidate the contributions of the primary components. Using our proposed method, we set the new state of the art by a significant margin (8% absolute increase in success rate on the unseen test set). Code is available at this https URL ."}, "keywords": ["visual-textual co-grounding module"], "citation_intent": "method"}
{"citing_id": "2303.12606v2", "cited_id": "1908.04509", "section_title": "Isolation Levels", "citation": "We present the axiomatic framework introduced by Biswas and Enea #REFR for defining isolation levels.", "text_before_citation": [], "text_after_citation": ["Isolation levels are defined as logical constraints, called axioms, over histories, which are an abstract representation of the interaction between a program and the database in an execution.", "2.2.1 Histories.", "Programs interact with a database by issuing transactions formed of begin, commit, abort, read and write instructions.", "The effect of executing one such instruction is represented using an event , type where is an identifier and type is a type.", "There are five types of events: begin, commit, abort, read( ) for reading the global variable , and write( , ) for writing value to . E denotes the set of events."], "citing_paper_content": {"title": "Dynamic Partial Order Reduction For Checking Correctness Against Transaction Isolation Levels", "abstract": "Modern applications, such as social networking systems and e-commerce platforms are centered around using large-scale databases for storing and retrieving data. Accesses to the database are typically enclosed in transactions that allow computations on shared data to be isolated from other concurrent computations and resilient to failures. Modern databases trade isolation for performance. The weaker the isolation level is, the more behaviors a database is allowed to exhibit and it is up to the developer to ensure that their application can tolerate those behaviors. In this work, we propose stateless model checking algorithms for studying correctness of such applications that rely on dynamic partial order reduction. These algorithms work for a number of widely-used weak isolation levels, including Read Committed, Causal Consistency, Snapshot Isolation, and Serializability. We show that they are complete, sound and optimal, and run with polynomial memory consumption in all cases. We report on an implementation of these algorithms in the context of Java Pathfinder applied to a number of challenging applications drawn from the literature of distributed systems and databases."}, "cited_paper_content": {"title": "On The Complexity Of Checking Transactional Consistency", "abstract": "Transactions simplify concurrent programming by enabling computations on shared data that are isolated from other concurrent computations and are resilient to failures. Modern databases provide different consistency models for transactions corresponding to different tradeoffs between consistency and availability. In this work, we investigate the problem of checking whether a given execution of a transactional database adheres to some consistency model. We show that consistency models like read committed, read atomic, and causal consistency are polynomial-time checkable while prefix consistency and snapshot isolation are NP-complete in general. These results complement a previous NP-completeness result concerning serializability. Moreover, in the context of NP-complete consistency models, we devise algorithms that are polynomial time assuming that certain parameters in the input executions, e.g., the number of sessions, are fixed. We evaluate the scalability of these algorithms in the context of several production databases."}, "keywords": ["axiomatic framework"], "citation_intent": "background"}
{"citing_id": "2304.04290v1", "cited_id": "1603.04467", "section_title": "B. Proof Of Concept: A Conditional Tabular Gan", "citation": "Distributing a tabular GAN discGAN models are written with the TensorFlow library #REFR , which provides the tf.distrbute [15] API to distribute model training across GPUs and machines.", "text_before_citation": ["These results are encouraging as the generated data both quantitatively and qualitatively resembles real data; however, it is noteworthy that the standard deviation for all unit types and ethnicities is smaller in the generated data than real data, significantly so for the Native American class, which had only 12 real examples in our training data.", "A 2D discGAN uses a default noise input of 50 followed by 2 64-node dense layers with LeakyReLU activation and a dropout layer (dropout rate of 0.1).", "Batch sizes of 16, 32, and 64 were used, and 32 produced the most accurate results.", "The following figures contain the output of the 2D GAN:", "C."], "text_after_citation": ["To demonstrate a tabular GAN in a distributed environment, we provisioned two GPU's and utilized tf.distribute's MultiWorkerMirroredStrategy which enables synchronous and distributed training using multiple GPUs on a single machine.", "MultiWorkerMirroredStrategy can distribute training across multiple workers, each with \"potentially multiple GPUs.\" [15] More specifically, the strategy replicates and updates all model parameters, weights, and computations to each device (hence it being a \"mirrored\" strategy), trains on each device, then uses distributed collective implementation -namely all-reduce.", "All-reduce is a fused algorithm that TensorFlow uses to \"aggregate tensors across all the devices by adding them up, and makes them available on each device.\" [15] TensorFlow uses the NVIDIA Collec- The strategy distributes discGAN by allowing the GPU workers to train together.", "Specifically, the devices divide the training data at each step before running forward and backward propagation separately on a mirrored copy of the the GAN models and their parameters on each device.", "The strategy then uses all-reduce to aggregate the distributed gradients and sends updated variable weights and a single aggregated gradient to all devices before starting the next training step."], "citing_paper_content": {"title": "Distributed Conditional Gan (Discgan) For Synthetic Healthcare Data Generation", "abstract": "In this paper, we propose a distributed Generative Adversarial Networks (discGANs) to generate synthetic tabular data specific to the healthcare domain. While using GANs to generate images has been well studied, little to no attention has been given to generation of tabular data. Modeling distributions of discrete and continuous tabular data is a non-trivial task with high utility. We applied discGAN to model non-Gaussian multi-modal healthcare data. We generated 249,000 synthetic records from original 2027 eICU dataset and we evaluated the performance of the model using machine learning efficacy, the Kolmogorov-Smirnov test (KS test) for continuous variables, and chi-squared test for discrete variables. Our results show that discGAN was able to generate data with distributions similar to the real data."}, "cited_paper_content": {"title": "Tensorflow: Large-Scale Machine Learning On Heterogeneous Distributed Systems", "abstract": "TensorFlow is an interface for expressing machine learning algorithms, and an implementation for executing such algorithms. A computation expressed using TensorFlow can be executed with little or no change on a wide variety of heterogeneous systems, ranging from mobile devices such as phones and tablets up to large-scale distributed systems of hundreds of machines and thousands of computational devices such as GPU cards. The system is flexible and can be used to express a wide variety of algorithms, including training and inference algorithms for deep neural network models, and it has been used for conducting research and for deploying machine learning systems into production across more than a dozen areas of computer science and other fields, including speech recognition, computer vision, robotics, information retrieval, natural language processing, geographic information extraction, and computational drug discovery. This paper describes the TensorFlow interface and an implementation of that interface that we have built at Google. The TensorFlow API and a reference implementation were released as an open-source package under the Apache 2.0 license in November, 2015 and are available at www.tensorflow.org."}, "keywords": ["tabular GAN discGAN", "TensorFlow library"], "citation_intent": "method"}
{"citing_id": "2303.06561v2", "cited_id": "1011.3027", "section_title": "Remark 7.", "citation": "We state an important theorem without proof, details of which can be found in #REFR .", "text_before_citation": ["X 1 := 1 \u03b5 2 \u03c3(\u03b5w x)\u03c3(\u03b5w y), then X 1 \u03c8 1 \u2264 C \u03c8,d , \u2022 and if X 2 := a 2 \u03c3 (1) (\u03b5w x)\u03c3 (1) (\u03b5w y) x, y , then X 2 \u03c8 1 \u2264 C \u03c8,1 . Proof.", "Let Z := w 2 2 \u223c \u03c7 2 (d).", "Since |X 1 | \u2264 |w x| |w y| = w 2 2 , then by definition, X 1 \u03c8 1 = inf s > 0 | E X 1 exp |X 1 | s \u2264 2 = inf s > 0 | E w exp |\u03c3(w x)\u03c3(w y)| s \u2264 2 \u2264 inf s > 0 | E w exp w 2 2 s \u2264 2 , hence X 1 \u03c8 1 \u2264 C \u03c8,d . (A.3)", "By similar reasoning, as", "|X 2 | \u2264 a 2 , hence X 2 \u03c8 1 \u2264 C \u03c8,1 . (A.4)"], "text_after_citation": ["Theorem 3 (Bernstein's inequality). Let {X k } m k=1 be i.i.d.", "sub-exponential random variables satisfying EX 1 = \u00b5, then for any \u03b7 \u2265 0, we have P 1 m m k=1 X k \u2212 \u00b5 \u2265 \u03b7 \u2264 2 exp \u2212C 0 m min \u03b7 2 X 1 2 \u03c8 1 , \u03b7 X 1 \u03c8 1 ,", "for some absolute constant C 0 .", "Proposition 1 (Upper and lower bounds of initial parameters). Given any \u03b4 \u2208 (0, 1), if", "m = \u2126 log 4 \u03b4 ,"], "citing_paper_content": {"title": "Phase Diagram Of Initial Condensation For Two-Layer Neural Networks", "abstract": "The phenomenon of distinct behaviors exhibited by neural networks under varying scales of initialization remains an enigma in deep learning research. In this paper, based on the earlier work by Luo et al. [16], we present a phase diagram of initial condensation for twolayer neural networks. Condensation is a phenomenon wherein the weight vectors of neural networks concentrate on isolated orientations"}, "cited_paper_content": {"title": "Introduction To The Non-Asymptotic Analysis Of Random Matrices", "abstract": "This is a tutorial on some basic non-asymptotic methods and concepts in random matrix theory. The reader will learn several tools for the analysis of the extreme singular values of random matrices with independent rows or columns. Many of these methods sprung off from the development of geometric functional analysis since the 1970's. They have applications in several fields, most notably in theoretical computer science, statistics and signal processing. A few basic applications are covered in this text, particularly for the problem of estimating covariance matrices in statistics and for validating probabilistic constructions of measurement matrices in compressed sensing. These notes are written particularly for graduate students and beginning researchers in different areas, including functional analysts, probabilists, theoretical statisticians, electrical engineers, and theoretical computer scientists."}, "keywords": ["important theorem", "details"], "citation_intent": "background"}
{"citing_id": "2305.01595v1", "cited_id": "1610.08452", "section_title": "Related Work", "citation": "They found good results balancing between overall accuracy and fairness and performed better than other fairness methods #REFR .", "text_before_citation": ["#OTHEREFR proposed a debiasing approach when creating embeddings by incorporating the loss function in the training process with a pre-trained binary (e.g., positive or negative) classifier.", "Boulitsakis-Logothetis #OTHEREFR proposed a modification to the Naive Bayes algorithm that gives guarantees on the statistical parity across protected attributes.", "This paper found benefits in terms of increases in fairness measures across protected classes but found that accuracy would often be degraded. Bechavod et al.", "#OTHEREFR proposed a method of penalizing unfairness which was inspired by regularization.", "In this way, they have introduced a data-dependent penalty to the learning process."], "text_after_citation": ["As compared to previous work, we looked at the impact of noisy training data on the level of fairness achieved by models trained with such data, compared the robustness to noise of different models, and also looked at a transfer learning scenario."], "citing_paper_content": {"title": "On The Impact Of Data Quality On Image Classification Fairness", "abstract": "With the proliferation of algorithmic decision-making, increased scrutiny has been placed on these systems. This paper explores the relationship between the quality of the training data and the overall fairness of the models trained with such data in the context of supervised classification. We measure key fairness metrics across a range of algorithms over multiple image classification datasets that have a varying level of noise in both the labels and the training data itself. We describe noise in the labels as inaccuracies in the labelling of the data in the training set and noise in the data as distortions in the data, also in the training set. By adding noise to the original datasets, we can explore the relationship between the quality of the training data and the fairness of the output of the models trained on that data."}, "cited_paper_content": {"title": "Fairness Beyond Disparate Treatment&Disparate Impact: Learning Classification Without Disparate Mistreatment", "abstract": "Automated data-driven decision making systems are increasingly being used to assist, or even replace humans in many settings. These systems function by learning from historical decisions, often taken by humans. In order to maximize the utility of these systems (or, classifiers), their training involves minimizing the errors (or, misclassifications) over the given historical data. However, it is quite possible that the optimally trained classifier makes decisions for people belonging to different social groups with different misclassification rates (e.g., misclassification rates for females are higher than for males), thereby placing these groups at an unfair disadvantage. To account for and avoid such unfairness, in this paper, we introduce a new notion of unfairness, disparate mistreatment, which is defined in terms of misclassification rates. We then propose intuitive measures of disparate mistreatment for decision boundary-based classifiers, which can be easily incorporated into their formulation as convex-concave constraints. Experiments on synthetic as well as real world datasets show that our methodology is effective at avoiding disparate mistreatment, often at a small cost in terms of accuracy."}, "keywords": ["fairness methods"], "citation_intent": "result"}
{"citing_id": "2304.08864v1", "cited_id": "1502.07571", "section_title": "Outline Of The Paper", "citation": "We also compare our findings with the previously known results of Alekasandrov et al #REFR .", "text_before_citation": ["Our report is organized as follows: In Section 1.1, we provide the motivation behind our report.", "In Section 2, we provide the necessary notations and preliminaries used in our proofs.", "Section 3 briefly mentions the results presented in #OTHEREFR , while also defining the notions of strategyproofness and envy-freeness.", "Next, in Section 4, we present our weighted LIKE and weighted Balanced LIKE mechanisms, along with a table summarizing our results."], "text_after_citation": ["We introduce the concept of the game tree, which helps in calculating expected utility in our weighted Balanced LIKE mechanism, in Section 5.", "In Section 6, we present our results related to the strategy-proofness of the weighted LIKE and weighted Balanced LIKE mechanisms.", "Finally, in Section 7, we present our results related to the envy-freeness of the weighted LIKE and weighted Balanced LIKE mechanisms."], "citing_paper_content": {"title": "Online Fair Division With Arbitrary Entitlements", "abstract": "The division of goods in the online realm poses opportunities and challenges. While innovative mechanisms can be developed, uncertainty about the future may hinder effective solutions. This project aims to explore fair distribution models for goods among agents with arbitrary entitlements, specifically addressing food charity challenges in the real world. Building upon prior work in [AAGW15], which focuses on equal entitlements, our project seeks to better understand the proofs of the theorems mentioned in that paper, which currently only provide proof sketches. Our approach employs different proof techniques from those presented in [AAGW15]."}, "cited_paper_content": {"title": "Online Fair Division: Analysing A Food Bank Problem", "abstract": "We study an online model of fair division designed to capture features of a real world charity problem. We consider two simple mechanisms for this model in which agents simply declare what items they like. We analyse axiomatic properties of these mechanisms such as strategy-proofness and envy freeness. Finally, we perform a competitive analysis and compute the price of anarchy."}, "keywords": ["previously known results", "Alekasandrov"], "citation_intent": "result"}
{"citing_id": "2303.11701v1", "cited_id": "1510.04389", "section_title": "A. Datasets And Metrics", "citation": "We evaluated our reconstructed images using PSNR and SSIM #REFR as metrics, where both metrics are of higher values indicating better quality of the recovered images.", "text_before_citation": ["We used the most widely used DIV2K #OTHEREFR for training, which has 900 images, of which 800 are used as the training set, and 100 are the validation set.", "LR images were aquired by double triple down-sampling We also tested our model on five common benchmark datasets: Set5 #OTHEREFR , Set14 #OTHEREFR , BSD100 #OTHEREFR , Urban100 #OTHEREFR , and Manga109 #OTHEREFR ."], "text_after_citation": [], "citing_paper_content": {"title": "A High-Frequency Focused Network For Lightweight Single Image Super-Resolution", "abstract": "Lightweight neural networks for single-image superresolution (SISR) tasks have made substantial breakthroughs in recent years. Compared to low-frequency information, highfrequency detail is much more difficult to reconstruct. Most SISR models allocate equal computational resources for lowfrequency and high-frequency information, which leads to redundant processing of simple low-frequency information and inadequate recovery of more challenging high-frequency information. We propose a novel High-Frequency Focused Network (HFFN) through High-Frequency Focused Blocks (HFFBs) that selectively enhance high-frequency information while minimizing redundant feature computation of low-frequency information. The HFFB effectively allocates more computational resources to the more challenging reconstruction of high-frequency information. Moreover, we propose a Local Feature Fusion Block (LFFB) effectively fuses features from multiple HFFBs in a local region, utilizing complementary information across layers to enhance feature representativeness and reduce artifacts in reconstructed images. We assess the efficacy of our proposed HFFN on five benchmark datasets and show that it significantly enhances the super-resolution performance of the network. Our experimental results demonstrate state-of-the-art performance in reconstructing high-frequency information while using a low number of parameters."}, "cited_paper_content": {"title": "Sketch-Based Manga Retrieval Using Manga109 Dataset", "abstract": "Manga (Japanese comics) are popular worldwide. However, current e-manga archives offer very limited search support, i.e., keyword-based search by title or author. To make the manga search experience more intuitive, efficient, and enjoyable, we propose a manga-specific image retrieval system. The proposed system consists of efficient margin labeling, edge orientation histogram feature description with screen tone removal, and approximate nearest-neighbor search using product quantization. For querying, the system provides a sketch-based interface. Based on the interface, two interactive reranking schemes are presented: relevance feedback and query retouch. For evaluation, we built a novel dataset of manga images, Manga109, which consists of 109 comic books of 21,142 pages drawn by professional manga artists. To the best of our knowledge, Manga109 is currently the biggest dataset of manga images available for research. Experimental results showed that the proposed framework is efficient and scalable (70 ms from 21,142 pages using a single computer with 204 MB RAM)."}, "keywords": ["reconstructed images", "better quality"], "citation_intent": "method"}
{"citing_id": "2304.04959v1", "cited_id": "1711.00108", "section_title": "Related Work", "citation": "The Soft Layer Ordering technique #REFR identifies the limitations of traditional fixed-order sharing approaches in MTL models and proposes learning task-specific scaling parameters to enable a flexible ordering of shared layers for each task.", "text_before_citation": ["Differently, our work introduces two kinds of complementary fusion modules to separately model task specific learning and shared learning.", "Also, In addition to explicitly introducing shared modules for learning commonalities across all tasks, we leverage direct task-pair fusion, based on input, to maximize the flexibility of knowledge sharing.", "Neural Architecture Search (NAS) #OTHEREFR methods have been applied to Multi-Task Learning (MTL) to automatically learn the structure of the models.", "Branched Multi-Task Networks #OTHEREFR generate a tree structure by clustering tasks based on affinity scores and assigning dissimilar tasks to the different branches.", "#OTHEREFR utilizes Gumbel-Softmax sampling for the branching operation instead of pre-calculated affinity scores, making the network endto-end trainable."], "text_after_citation": ["AdaShare #OTHEREFR learns a task-specific policy to select which layers to execute for every particular task.", "Sub-Network Routing (SNR) #OTHEREFR splits shared layers into sub-networks and learns their connections with latent variables.", "NAS methods eliminate a significant amount of manual work and improve the flexibility of the sharing patterns in the MTL model.", "However, as an exhaustive search of all possible model configurations would be combinatorially complex, these methods often rely on quite simplified assumptions such as branching #OTHEREFR , routing #OTHEREFR , layer ordering #OTHEREFR , layer selecting #OTHEREFR etc. to limit the search space. Additionally, the generated structures don't adjust to individual examples.", "In addition to works focusing on MTL architecture design, another line of works aims to improve MTL optimization."], "citing_paper_content": {"title": "Adatt: Adaptive Task-To-Task Fusion Network For Multitask Learning In Recommendations", "abstract": "Multi-task learning (MTL) aims at enhancing the performance and efficiency of machine learning models by training them on multiple tasks simultaneously. However, MTL research faces two challenges: 1) modeling the relationships between tasks to effectively share knowledge between them, and 2) jointly learning task-specific and shared knowledge. In this paper, we present a novel model Adaptive Task-to-Task Fusion Network (AdaTT) to address both challenges. AdaTT is a deep fusion network built with task specific and optional shared fusion units at multiple levels. By leveraging a residual mechanism and gating mechanism for task-to-task fusion, these units adaptively learn shared knowledge and task specific knowledge. To evaluate the performance of AdaTT, we conduct experiments on a public benchmark and an industrial recommendation dataset using various task groups. Results demonstrate AdaTT can significantly outperform existing state-of-the-art baselines. CCS CONCEPTS \u2022 Computing methodologies \u2192 Multi-task learning; \u2022 Information systems \u2192 Recommender systems."}, "cited_paper_content": {"title": "Beyond Shared Hierarchies: Deep Multitask Learning Through Soft Layer Ordering", "abstract": "Existing deep multitask learning (MTL) approaches align layers shared between tasks in a parallel ordering. Such an organization significantly constricts the types of shared structure that can be learned. The necessity of parallel ordering for deep MTL is first tested by comparing it with permuted ordering of shared layers. The results indicate that a flexible ordering can enable more effective sharing, thus motivating the development of a soft ordering approach, which learns how shared layers are applied in different ways for different tasks. Deep MTL with soft ordering outperforms parallel ordering methods across a series of domains. These results suggest that the power of deep MTL comes from learning highly general building blocks that can be assembled to meet the demands of each task."}, "keywords": ["task", "Soft Layer Ordering"], "citation_intent": "background"}
{"citing_id": "2303.05639v2", "cited_id": "1506.03365", "section_title": "Automatic Segmentation With One-Shot Learning", "citation": "For the other three classes, models trained on the LSUN dataset #REFR are used which contains more than 1 \u00d7 10 6 samples for each class.", "text_before_citation": ["For each class, we have used two different datasets for evaluation: (i) The first dataset is a large-scale unlabeled dataset on which the GAN model is pre-trained for image synthesis.", "And (ii) the second dataset is a small-size labeled dataset wherein the samples have been annotated with segmentation labels and are used to train and evaluate the one-shot learning model.", "The first section of Table 1 lists the different unlabeled datasets used for pre-training the StyleGAN models for image synthesis.", "The table also mentions the size of the StyleGAN model being pre-trained.", "Here, two main datasets are made use of for pre-training the models: For the Face class, we use a StyleGAN2-ADA model that is pre-trained on the Flickr-Faces-HQ dataset #OTHEREFR to generate human faces."], "text_after_citation": ["For the Cat class, the Animals Faces-HQ #OTHEREFR dataset is also additionally used."], "citing_paper_content": {"title": "Self-Supervised One-Shot Learning For For Automatic Segmentation Of Stylegan Images", "abstract": "We propose a framework for the automatic one-shot segmentation of synthetic images generated by a StyleGAN. Our framework is based on the observation that the multi-scale hidden features in the GAN generator hold useful semantic information that can be utilized for automatic on-the-fly segmentation of the generated images. Using these features, our framework learns to segment synthetic images using a self-supervised contrastive clustering algorithm that projects the hidden features into a compact space for per-pixel classification. This novel contrastive learner is based on using a pixel-wise swapped prediction loss for image segmentation that leads to faster learning of the feature vectors for one-shot segmentation. We have tested our implementation on a number of standard benchmarks to yield a segmentation performance that not only outperforms the semi-supervised baseline methods by an average wIoU margin of 1.02% but also improves the inference speeds by a factor of 4.5. Finally, we also show the results of using the proposed one-shot learner in implementing BagGAN, a framework for producing annotated synthetic baggage X-ray scans for threat detection. This framework was trained and tested on the PIDRay baggage benchmark to yield a performance comparable to its baseline segmenter based on manual annotations."}, "cited_paper_content": {"title": "Lsun: Construction Of A Large-Scale Image Dataset Using Deep Learning With Humans In The Loop", "abstract": "While there has been remarkable progress in the performance of visual recognition algorithms, the state-of-the-art models tend to be exceptionally data-hungry. Large labeled training datasets, expensive and tedious to produce, are required to optimize millions of parameters in deep network models. Lagging behind the growth in model capacity, the available datasets are quickly becoming outdated in terms of size and density. To circumvent this bottleneck, we propose to amplify human effort through a partially automated labeling scheme, leveraging deep learning with humans in the loop. Starting from a large set of candidate images for each category, we iteratively sample a subset, ask people to label them, classify the others with a trained model, split the set into positives, negatives, and unlabeled based on the classification confidence, and then iterate with the unlabeled set. To assess the effectiveness of this cascading procedure and enable further progress in visual recognition research, we construct a new image dataset, LSUN. It contains around one million labeled images for each of 10 scene categories and 20 object categories. We experiment with training popular convolutional networks and find that they achieve substantial performance gains when trained on this dataset."}, "keywords": ["LSUN dataset"], "citation_intent": "method"}
{"citing_id": "2303.16633v1", "cited_id": "1707.04131", "section_title": "Adversarial Noise Attack", "citation": "A very simple form of untargeted adversarial attacks are adversarial noise attacks, which were originally introduced by #REFR .", "text_before_citation": [], "text_after_citation": ["Noise attacks are applicable to both classification tasks and regression tasks.", "They perturb the input data by adding random noise, commonly Gaussian noise or uniform noise.", "In the process, the perturbation is normalized and rescaled to the desired size, e.g., with respect to the L \u221e norm.", "In addition, the perturbed samples need to be clipped afterwards so that all values are within the valid lower and upper bounds of the input data #OTHEREFR .", "Noise attacks require no prior knowledge of the model and thus represent black-box attacks."], "citing_paper_content": {"title": "Targeted Adversarial Attacks On Wind Power Forecasts", "abstract": "In recent years, researchers proposed a variety of deep learning models for wind power forecasting. These models predict the wind power generation of wind farms or entire regions more accurately than traditional machine learning algorithms or physical models. However, latest research has shown that deep learning models can often be manipulated by adversarial attacks. Since wind power forecasts are essential for the stability of modern power systems, it is important to protect them from this threat. In this work, we investigate the vulnerability of two different forecasting models to targeted, semitargeted, and untargeted adversarial attacks. We consider a Long Short-Term Memory (LSTM) network for predicting the power generation of a wind farm and a Convolutional Neural Network (CNN) for forecasting the wind power generation throughout Germany. Moreover, we propose the Total Adversarial Robustness Score (TARS), an evaluation metric for quantifying the robustness of regression models to targeted and semi-targeted adversarial attacks. It assesses the impact of attacks on the model's performance, as well as the extent to which the attacker's goal was achieved, by assigning a score between 0 (very vulnerable) and 1 (very robust). In our experiments, the LSTM forecasting model was fairly robust and achieved a TARS value of over 0.81 for all adversarial attacks investigated. The CNN forecasting model only achieved TARS values below 0.06 when trained ordinarily, and was thus very vulnerable. Yet, its robustness could be significantly improved by adversarial training, which always resulted in a TARS above 0.46."}, "cited_paper_content": {"title": "Foolbox V0.8.0: A Python Toolbox To Benchmark The Robustness Of Machine Learning Models", "abstract": "Even todays most advanced machine learning models are easily fooled by almost imperceptible perturbations of their inputs. Foolbox is a new Python package to generate such adversarial perturbations and to quantify and compare the robustness of machine learning models. It is build around the idea that the most comparable robustness measure is the minimum perturbation needed to craft an adversarial example. To this end, Foolbox provides reference implementations of most published adversarial attack methods alongside some new ones, all of which perform internal hyperparameter tuning to find the minimum adversarial perturbation. Additionally, Foolbox interfaces with most popular deep learning frameworks such as PyTorch, Keras, TensorFlow, Theano and MXNet and allows different adversarial criteria such as targeted misclassification and top-k misclassification as well as different distance measures. The code is licensed under the MIT license and is openly available at this https URL . The most up-to-date documentation can be found at this http URL ."}, "keywords": ["untargeted adversarial attacks", "adversarial noise attacks"], "citation_intent": "background"}
{"citing_id": "2303.06277v1", "cited_id": "1908.03265", "section_title": "Implementation Details", "citation": "The network is trained using the RAdam optimizer #REFR as it eliminates the need for learning rate warmup during training.", "text_before_citation": ["Our network is implemented in Pytorch.", "We use the spatial partitioning strategy from #OTHEREFR to partition the adjacency matrix representing the joint connections.", "Dilations of 1, 2, and 4 are used in the temporal convolution layers in the GCN-TCN unit in order increase their receptive field."], "text_after_citation": ["We use learning rate of 3 \u00d7 10 \u22124 and decay it by a factor of 0.99 every epoch.", "The model is regularized using a weight decay of 1e \u2212 5 and a dropout rate of 0.1 in the convolutional layers operating along the temporal dimension.", "By performing some test runs we picked the values \u03b1 = 10 and \u03b2 = 0.1.", "In the SSA unit shown in 2 we use a temporal convolution layer after the spatial attention.", "Analogous to this we use a GCN unit to extract the spatial features before extracting temporal attention features."], "citing_paper_content": {"title": "Spotr: Spatio-Temporal Pose Transformers For Human Motion Prediction", "abstract": "3D human motion prediction is a research area of high significance and a challenge in computer vision. It is useful for the design of many applications including robotics and autonomous driving. Traditionally, autogregressive models have been used to predict human motion. However, these models have high computation needs and error accumulation that make it difficult to use them for realtime applications. In this paper, we present a non-autogressive model for human motion prediction. We focus on learning spatiotemporal representations non-autoregressively for generation of plausible future motions. We propose a novel architecture that leverages the recently proposed Transformers. Human motion involves complex spatio-temporal dynamics with joints affecting the position and rotation of each other even though they are not connected directly. The proposed model extracts these dynamics using both convolutions and the self-attention mechanism. Using specialized spatial and temporal self-attention to augment the features extracted through convolution allows our model to generate spatiotemporally coherent predictions in parallel independent of the activity. Our contributions are threefold: (i) we frame human motion prediction as a sequence-to-sequence problem and propose a non-autoregressive Transformer to forecast a sequence of poses in parallel; (ii) our method is activity agnostic; (iii) we show that despite its simplicity, our approach is able to make accurate predictions, achieving better or comparable results compared to the state-of-theart on two public datasets, with far fewer parameters and much faster inference."}, "cited_paper_content": {"title": "On The Variance Of The Adaptive Learning Rate And Beyond", "abstract": "The learning rate warmup heuristic achieves remarkable success in stabilizing training, accelerating convergence and improving generalization for adaptive stochastic optimization algorithms like RMSprop and Adam. Pursuing the theory behind warmup, we identify a problem of the adaptive learning rate -- its variance is problematically large in the early stage, and presume warmup works as a variance reduction technique. We provide both empirical and theoretical evidence to verify our hypothesis. We further propose Rectified Adam (RAdam), a novel variant of Adam, by introducing a term to rectify the variance of the adaptive learning rate. Experimental results on image classification, language modeling, and neural machine translation verify our intuition and demonstrate the efficacy and robustness of RAdam."}, "keywords": ["training", "RAdam optimizer"], "citation_intent": "method"}
{"citing_id": "2303.07012v1", "cited_id": "2003.06606", "section_title": "Handwritten Text Generation", "citation": "In addition, #REFR is an adversarial augmentation method that increases handwritten text images by transforming shapes.", "text_before_citation": ["The task of handwritten text generation is similar to that of photographic ancient character generation, aiming to imitate natural handwriting in human style.", "The early two-step methods #OTHEREFR generate isolated letters, and then concatenate them to produce a whole word.", "These methods rely on handcrafted rules and only generate handwriting with limited variations. Recently, deep generative models directly generate whole-word images.", "RNN and GAN were used to generate handwriting in a variety of styles #OTHEREFR .", "#OTHEREFR and #OTHEREFR generate text images on the condition of extracting style features in a few-shot setup and textual content of a predefined fixed length."], "text_after_citation": ["The above-mentioned methods are designed to generate only Latin characters, and most of them require expensive text annotations to enhance the generation quality."], "citing_paper_content": {"title": "Agtgan: Unpaired Image Translation For Photographic Ancient Character Generation", "abstract": "The study of ancient writings has great value for archaeology and philology. Essential forms of material are photographic characters, but manual photographic character recognition is extremely timeconsuming and expertise-dependent. Automatic classification is therefore greatly desired. However, the current performance is limited due to the lack of annotated data. Data generation is an inexpensive but useful solution to data scarcity. Nevertheless, the diverse glyph shapes and complex background textures of photographic ancient characters make the generation task difficult, leading to unsatisfactory results of existing methods. To this end, we propose an unsupervised generative adversarial network called AGTGAN in this paper. By explicitly modeling global and local glyph shape style, followed by a stroke-aware texture transfer and an associate adversarial learning mechanism, our method can generate characters with diverse glyphs and realistic textures. We evaluate our method on photographic ancient character datasets, e.g., OBC306 and CSDD. Our method outperforms other state-of-the-art methods in terms of various metrics and performs much better in terms of the diversity and authenticity of generated samples. With our generated images, experiments on the largest photographic oracle bone character dataset show that our method can achieve a significant increase in classification accuracy, up to 16.34%. The source code is"}, "cited_paper_content": {"title": "Learn To Augment: Joint Data Augmentation And Network Optimization For Text Recognition", "abstract": "Handwritten text and scene text suffer from various shapes and distorted patterns. Thus training a robust recognition model requires a large amount of data to cover diversity as much as possible. In contrast to data collection and annotation, data augmentation is a low cost way. In this paper, we propose a new method for text image augmentation. Different from traditional augmentation methods such as rotation, scaling and perspective transformation, our proposed augmentation method is designed to learn proper and efficient data augmentation which is more effective and specific for training a robust recognizer. By using a set of custom fiducial points, the proposed augmentation method is flexible and controllable. Furthermore, we bridge the gap between the isolated processes of data augmentation and network optimization by joint learning. An agent network learns from the output of the recognition network and controls the fiducial points to generate more proper training samples for the recognition network. Extensive experiments on various benchmarks, including regular scene text, irregular scene text and handwritten text, show that the proposed augmentation and the joint learning methods significantly boost the performance of the recognition networks. A general toolkit for geometric augmentation is available."}, "keywords": ["handwritten text images"], "citation_intent": "method"}
{"citing_id": "2303.06266v1", "cited_id": "1107.4540", "section_title": "Draft March 14, 2023", "citation": "N-COMP based user identification N-COMP is a decoding strategy for noisy GT proposed by Chan et al. in #REFR .", "text_before_citation": ["A."], "text_after_citation": ["The basic idea is to declare an item to be defective or non-defective based on the fraction of positive tests they participate in #OTHEREFR .", "This significantly simplifies the run-time and storage requirements in comparison to the optimal exhaustive search strategy in Algorithm 1.", "As established in Section III, the active device identification framework in non-coherent ( , k)\u2212MnAC can be viewed as a GT problem.", "This motivates us to study the performance of N-COMP approach in identifying active devices in the non-coherent ( , k)\u2212MnAC.", "Note that in the existing literature, the N-COMP GT decoder was primarily analyzed in the context of symmetric noise models wherein errors in test results are equally likely #OTHEREFR , #OTHEREFR ."], "citing_paper_content": {"title": "Non-Coherent Active Device Identification For Massive Random Access", "abstract": "Massive Machine-Type Communications (mMTC) is a key service category in the current generation of wireless networks featuring an extremely high density of energy and resource-limited devices with sparse and sporadic activity patterns. In order to enable random access in such mMTC networks, base station needs to identify the active devices while operating within stringent access delay constraints. In this paper, an energy efficient active device identification protocol is proposed in which active devices transmit On-Off Keying (OOK) modulated preambles jointly and base station employs non-coherent energy detection avoiding channel estimation overheads. The minimum number of channel-uses required by the active user identification protocol is characterized in the asymptotic regime of total number of devices when the number of active devices k scales as k = \u0398(1) along with an achievability scheme relying on the equivalence of activity detection to a group testing problem. Several practical schemes based on Belief Propagation (BP) and Combinatorial Orthogonal Matching Pursuit (COMP) are also proposed. Simulation results show that BP strategies outperform COMP significantly and can operate close to the theoretical achievability bounds. In a partial-recovery setting where few misdetections are allowed, BP continues to perform well."}, "cited_paper_content": {"title": "Non-Adaptive Probabilistic Group Testing With Noisy Measurements: Near-Optimal Bounds With Efficient Algorithms", "abstract": "We consider the problem of detecting a small subset of defective items from a large set via non-adaptive \u201crandom pooling\u201d group tests. We consider both the case when the measurements are noiseless, and the case2 when the measurements are noisy (the outcome of each group test may be independently faulty with probability q). Order-optimal results for these scenarios are known in the literature. We give information-theoretic lower bounds on the query complexity of these problems, and provide corresponding computationally efficient algorithms that match the lower bounds up to a constant factor. To the best of our knowledge this work is the first to explicitly estimate such a constant that characterizes the gap between the upper and lower bounds for these problems."}, "keywords": ["identification N-COMP"], "citation_intent": "background"}
{"citing_id": "2303.06163v1", "cited_id": "2003.09754", "section_title": "Dataset", "citation": "Following #REFR 15] , we use Level-3 granularity and filter out the shapes with more than 50 pairs of joints, which leaves us with 3736 chairs, 5053 tables, 719 cabinets.", "text_before_citation": ["Following #OTHEREFR , we use the three largest furniture categories that require realworld assembly, chairs, tables, and cabinets, and adopt the PartNet official train/validation/test split.", "We use Furthest Point Sampling (FPS) to sample 1,000 points over each part mesh.", "All parts are canonicalized to be zero-centered and rotated to be local axis aligned using PCA.", "We detect joint points by computing all pair-wise part Chamfer Distance (eq.", "3) and take the closest 50 points between two connected parts with a minimum distance of less than 0.05."], "text_after_citation": [], "citing_paper_content": {"title": "Category-Level Multi-Part Multi-Joint 3D Shape Assembly", "abstract": "Shape assembly composes complex shapes geometries by arranging simple part geometries and has wide applications in autonomous robotic assembly and CAD modeling. Existing works focus on geometry reasoning and neglect the actual physical assembly process of matching and fitting joints, which are the contact surfaces connecting different parts. In this paper, we consider contacting joints for the task of multi-part assembly. A successful joint-optimized assembly needs to satisfy the bilateral objectives of shape structure and joint alignment. We propose a hierarchical graph learning approach composed of two levels of graph representation learning. The part graph takes part geometries as input to build the desired shape structure. The joint-level graph uses part joints information and focuses on matching and aligning joints. The two kinds of information are combined to achieve the bilateral objectives. Extensive experiments demonstrate that our method outperforms previous methods, achieving better shape structure and higher joint alignment accuracy."}, "cited_paper_content": {"title": "Learning 3D Part Assembly From A Single Image", "abstract": "Autonomous assembly is a crucial capability for robots in many applications. For this task, several problems such as obstacle avoidance, motion planning, and actuator control have been extensively studied in robotics. However, when it comes to task specification, the space of possibilities remains underexplored. Towards this end, we introduce a novel problem, single-image-guided 3D part assembly, along with a learningbased solution. We study this problem in the setting of furniture assembly from a given complete set of parts and a single image depicting the entire assembled object. Multiple challenges exist in this setting, including handling ambiguity among parts (e.g., slats in a chair back and leg stretchers) and 3D pose prediction for parts and part subassemblies, whether visible or occluded. We address these issues by proposing a two-module pipeline that leverages strong 2D-3D correspondences and assembly-oriented graph message-passing to infer part relationships. In experiments with a PartNet-based synthetic benchmark, we demonstrate the effectiveness of our framework as compared with three baseline approaches."}, "keywords": ["shapes", "3736 chairs"], "citation_intent": "method"}
{"citing_id": "2304.06810v1", "cited_id": "1511.01854", "section_title": "Methodology", "citation": "In the above, D(\u2022, \u2022) is a discrepancy measure between two sets of observables; for example, in the case where we are measuring the discrepancy between two density matrices, we may take D to be the Trace Distance #REFR . In Eq.", "text_before_citation": ["1 for the set of parameters \u039b and a particular realization of the vacuum noise, followed by projection of the output and noise fields onto a desired orthonormal basis; E denotes the expectation over the vacuum noise; and the operator F computes the first-order correlations which yield the desired observable.", "Given a desired observable-set, O d , describing the quantum state or any related features, our goal is to find the unknown physical parameters, \u039b, that characterize the system.", "We take a parameterized approach to solving the inverse problem, i.e.", "\u039b = \u039b(\u03b8), and solve the inverse problem by solving the optimization problem", "EQUATION"], "text_after_citation": ["3, by minimizing the discrepancy the model produces the properties of the optical system that would produce a result as close as possible to the desired quantum state. The inverse model is then given by", "EQUATION", "where I(\u2022) is our inverse solver. In order to solve the optimization problem in Eq.", "3, an approach based on gradient descent may be employed. The key is that the forward model of Eq.", "1, while quite complicated, can be expressed in such a way that it is fully differentiable."], "citing_paper_content": {"title": "Designing Nonlinear Photonic Crystals For High-Dimensional Quantum State Engineering", "abstract": "We propose a novel, physically-constrained and differentiable approach for the generation of D-dimensional qudit states via spontaneous parametric downconversion (SPDC) in quantum optics. We circumvent any limitations imposed by the inherently stochastic nature of the physical process and incorporate a set of stochastic dynamical equations governing its evolution under the SPDC Hamiltonian. We demonstrate the effectiveness of our model through the design of structured nonlinear photonic crystals (NLPCs) and shaped pump beams; and show, theoretically and experimentally, how to generate maximally entangled states in the spatial degree of freedom. The learning of NLPC structures offers a promising new avenue for shaping and controlling arbitrary quantum states and enables all-optical coherent control of the generated states. We believe that this approach can readily be extended from bulky crystals to thin Metasurfaces and potentially applied to other quantum systems sharing a similar Hamiltonian structures, such as superfluids and superconductors."}, "cited_paper_content": {"title": "Trace-Distance Measure Of Coherence", "abstract": "We show that trace distance measure of coherence is a strong monotone for all qubit and, so called, $X$ states. An expression for the trace distance coherence for all pure states and a semi definite program for arbitrary states is provided. We also explore the relation between $l_1$-norm and relative entropy based measures of coherence, and give a sharp inequality connecting the two. In addition, it is shown that both $l_p$-norm- and Schatten-$p$-norm-based measures violate the (strong) monotonicity for all $p\\in(1,\\infty)$."}, "keywords": ["two density matrices", "Trace Distance"], "citation_intent": "background"}
{"citing_id": "2304.00216v1", "cited_id": "1803.08375", "section_title": "Ablation Studies", "citation": "As the CD classification performance in Table 4 , sharing the kernel weight in the CS-MIL with ReLU #REFR achieved better performances with a higher mean value of multiple metrics.", "text_before_citation": ["Inspired by #OTHEREFR and #OTHEREFR , we explores various attention mechanism designs in MIL with different activation functions and evaluate those designs on CD dataset.", "We formed the CS-MIL into two strategies, differentiated by whether they shared the kernel weights while learning the embedding features from multiple scales."], "text_after_citation": [], "citing_paper_content": {"title": "Medical Image Analysis", "abstract": "Analyzing high resolution whole slide images (WSIs) with regard to information across multiple scales poses a significant challenge in digital pathology. Multi-instance learning (MIL) is a common solution for working with high resolution images by classifying bags of objects (i.e. sets of smaller image patches). However, such processing is typically performed at a single scale (e.g., 20\u00d7 magnification) of WSIs, disregarding the vital inter-scale information that is key to diagnoses by human pathologists. In this study, we propose a novel cross-scale MIL algorithm to explicitly aggregate inter-scale relationships into a single MIL network for pathological image diagnosis. The contribution of this paper is threefold: (1) A novel cross-scale MIL (CS-MIL) algorithm that integrates the multi-scale information and the inter-scale relationships is proposed; (2) A toy dataset with scale-specific morphological features is created and released to examine and visualize differential cross-scale attention; (3) Superior performance on both in-house and public datasets is demonstrated by our simple cross-scale MIL strategy. The official implementation is publicly available at https://github.com/hrlblab/CS-MIL."}, "cited_paper_content": {"title": "Deep Learning Using Rectified Linear Units (Relu)", "abstract": "We introduce the use of rectified linear units (ReLU) as the classification function in a deep neural network (DNN). Conventionally, ReLU is used as an activation function in DNNs, with Softmax function as their classification function. However, there have been several studies on using a classification function other than Softmax, and this study is an addition to those. We accomplish this by taking the activation of the penultimate layer $h_{n - 1}$ in a neural network, then multiply it by weight parameters $\\theta$ to get the raw scores $o_{i}$. Afterwards, we threshold the raw scores $o_{i}$ by $0$, i.e. $f(o) = \\max(0, o_{i})$, where $f(o)$ is the ReLU function. We provide class predictions $\\hat{y}$ through argmax function, i.e. argmax $f(x)$."}, "keywords": ["kernel weight", "ReLU"], "citation_intent": "result"}
{"citing_id": "2303.04503v1", "cited_id": "1906.08311", "section_title": "A. Ev Charging Modeling", "citation": "In this work, EV charging demand at the station must be calculated and included as part of the power balance equation #REFR .", "text_before_citation": ["EQUATION", "where", "P t v + , P t v \u2212", "are the charging and discharging power for vehicle v; \u03b7 v+ , \u03b7 v\u2212 are the charging and discharging efficiency, respectively.", "Charging and discharging can not happen at the same time."], "text_after_citation": ["Therefore, the set of plugged-in EVs to be charged at each time t at the station is defined as:", "EQUATION", "where t a v is the arrival time; t d v is the departure time; SoC t v is the EV battery energy level; E c v is the required charging energy of vehicle v.", "Thus, assuming charging with a constant nominal charging powerP v , the overall EV demand at the railway station level can be estimated as:", "EQUATION"], "citing_paper_content": {"title": "An Optimal Energy Management Algorithm Considering Regenerative Braking And Renewable Energy For Ev Charging In Railway Stations", "abstract": "This paper proposes a novel optimal Energy Management System (EMS) algorithm for Electric Vehicle (EV) charging in smart electric railway stations with renewable generation. As opposed to previous railway EMS methods, the proposed EMS coordinates the combined Regenerative Braking Energy (RBE), renewable generation, electric railway demand and EV charging demand at the EV parking lot of the railway station. Numerical results using a scenario-based approach on an actual railway station in Chur, Switzerland demonstrate that the proposed algorithm can effectively minimize the expected daily operating cost for the train station over an entire year. Index Terms-Electric vehicles, energy management, railway systems, regenerative braking energy, renewable energy, mixed integer linear programming."}, "cited_paper_content": {"title": "The Effect Of The Uncertainty Of Load And Renewable Generation On The Dynamic Voltage Stability Margin", "abstract": "In this paper, the impact of stochastic load and renewable generation uncertainty on the dynamic voltage stability margin is studied. Stochastic trajectories describing the uncertainty of load, wind and solar generation have been incorporated in the power system model as a set of Stochastic Differential-Algebraic Equations (SDAEs). A systematic study of Monte Carlo dynamic simulations on the IEEE 39-Bus system has been conducted to compute the stochastic load margin with all dynamic components active. Numerical results show that the uncertainty of both demand and generation may lead to a decrease on the size of the dynamic voltage stability margin, yet the variability of renewable generators may play a more significant role. Given that the integration of renewable energy will continue growing, it is of paramount importance to apply stochastic and dynamic approaches in the voltage stability study."}, "keywords": ["EV charging demand"], "citation_intent": "method"}
{"citing_id": "2303.16084v1", "cited_id": "1906.11415", "section_title": "A Common Setup For Classifier And Matching-Based Approaches", "citation": "We identify two dominant families of approaches proposed in the recent few-shot action recognition literature: the classifier-based and the matching-based methods #REFR .", "text_before_citation": [], "text_after_citation": ["Unfortunately, discrepancies in setup and architecture between existing approaches make it difficult to compare them fairly.", "We propose to follow the same representation learning strategy and start from a common frozen backbone.", "Classifier-based approaches The classifier-based approaches #OTHEREFR train the video representation using a classifier in the form of a linear layer.", "This is similar to the corresponding work on few-shot learning in the image domain #OTHEREFR .", "Classifiers are trained using the whole dataset, but episodic training cannot guarantee to see all the examples because it is intractable to sample all the possible episodes. Xian et al."], "citing_paper_content": {"title": "Rethinking Matching-Based Few-Shot Action Recognition", "abstract": "Few-shot action recognition, i.e. recognizing new action classes given only a few examples, benefits from incorporating temporal information. Prior work either encodes such information in the representation itself and learns classifiers at test time, or obtains frame-level features and performs pairwise temporal matching. We first evaluate a number of matching-based approaches using features from spatio-temporal backbones, a comparison missing from the literature, and show that the gap in performance between simple baselines and more complicated methods is significantly reduced. Inspired by this, we propose Chamfer++, a non-temporal matching function that achieves state-of-the-art results in few-shot action recognition. We show that, when starting from temporal features, our parameter-free and interpretable approach can outperform all other matching-based and classifier methods for one-shot action recognition on three common datasets without using temporal information in the matching stage. Project page: https://jbertrand89.github.io/matching-based-fsar"}, "cited_paper_content": {"title": "Few-Shot Video Classification Via Temporal Alignment", "abstract": "There is a growing interest in learning a model which could recognize novel classes with only a few labeled examples. In this paper, we propose Temporal Alignment Module (TAM), a novel few-shot learning framework that can learn to classify a previous unseen video. While most previous works neglect long-term temporal ordering information, our proposed model explicitly leverages the temporal ordering information in video data through temporal alignment. This leads to strong data-efficiency for few-shot learning. In concrete, TAM calculates the distance value of query video with respect to novel class proxies by averaging the per frame distances along its alignment path. We introduce continuous relaxation to TAM so the model can be learned in an end-to-end fashion to directly optimize the few-shot learning objective. We evaluate TAM on two challenging real-world datasets, Kinetics and Something-Something-V2, and show that our model leads to significant improvement of few-shot video classification over a wide range of competitive baselines."}, "keywords": ["recent few-shot action"], "citation_intent": "method"}
{"citing_id": "2304.07651v1", "cited_id": "1705.05065", "section_title": "High Fidelity Simulation", "citation": "For our high fidelity simulation environment, we took inspiration from many industry standard simulations, such as Gazebo and AirSim #REFR .", "text_before_citation": [], "text_after_citation": ["Still, our high fidelity simulation does have several characteristics that cause it to vary from many common simulations.", "Two of the key differentiators is that there is no active physics in the high fidelity simulation and agent bodies consist of simple relative sized cube objects.", "In place of a physics environment, each platform type has a geometry based PID controller to actuate motion.", "Agents send actuation commands via ROS Twist messages or setpoint position commands just like on the real platforms.", "Comparable to other simulators, individual agents receive sensory information from ROS messages."], "citing_paper_content": {"title": "From Warfighting Needs To Robot Actuation: A Complete Rapid Integration Swarming Solution", "abstract": "Swarm robotics systems have the potential to transform warfighting in urban environments, but until now have not seen large-scale field testing. We present the Rapid Integration Swarming Ecosystem (RISE), a platform for future multi-agent research and deployment. RISE enables rapid integration of third-party swarm tactics and behaviors, which was demonstrated using both physical and simulated swarms. Our physical testbed is composed of more than 250 networked heterogeneous agents and has been extensively tested in mock warfare scenarios at five urban combat training ranges. RISE implements live, virtual, constructive simulation capabilities to allow the use of both virtual and physical agents simultaneously, while our \"fluid fidelity\" simulation enables adaptive scaling between low and high fidelity simulation levels based on dynamic runtime requirements. Both virtual and physical agents are controlled with a unified gesture-based interface that enables a greater than 150:1 agent-to-operator ratio. Through this interface, we enable efficient swarm-based mission execution. RISE translates mission needs to robot actuation with rapid tactic integration, a reliable testbed, and efficient operation."}, "cited_paper_content": {"title": "Airsim: High-Fidelity Visual And Physical Simulation For Autonomous Vehicles", "abstract": "Developing and testing algorithms for autonomous vehicles in real world is an expensive and time consuming process. Also, in order to utilize recent advances in machine intelligence and deep learning we need to collect a large amount of annotated training data in a variety of conditions and environments. We present a new simulator built on Unreal Engine that offers physically and visually realistic simulations for both of these goals. Our simulator includes a physics engine that can operate at a high frequency for real-time hardware-in-the-loop (HITL) simulations with support for popular protocols (e.g. MavLink). The simulator is designed from the ground up to be extensible to accommodate new types of vehicles, hardware platforms and software protocols. In addition, the modular design enables various components to be easily usable independently in other projects. We demonstrate the simulator by first implementing a quadrotor as an autonomous vehicle and then experimentally comparing the software components with real-world flights."}, "keywords": ["high fidelity simulation"], "citation_intent": "method"}
{"citing_id": "2303.06431v1", "cited_id": "1907.03313", "section_title": "B. Autoencoder", "citation": "Compared with supervised approaches for anomaly detection #REFR , employing autoencoder-based models for anomaly detection problems is efficient as only normal data are necessary for the training phase, and collecting various labeled anomalous data is costly and not always possible.", "text_before_citation": ["The goal of training an autoencoder is to capture the nonlinear correlations among input features #OTHEREFR .", "For anomaly detection tasks, autoencoders model the data distribution of normal samples by minimizing the distance between the normal sample x and the reconstructed sample x .", "As shown in #OTHEREFR , an autoencoder can solve the following optimization problem #OTHEREFR :", "EQUATION", "where \u03c6 and \u03c8 are the parameters for the encoder and decoder, respectively, and l 2 -norm is usually utilized for || \u2022 ||."], "text_after_citation": [], "citing_paper_content": {"title": "Anomaly Detection With Ensemble Of Encoder And Decoder", "abstract": "Hacking and false data injection from adversaries can threaten power grids' everyday operations and cause significant economic loss. Anomaly detection in power grids aims to detect and discriminate anomalies caused by cyber attacks against the power system, which is essential for keeping power grids working correctly and efficiently. Different methods have been applied for anomaly detection, such as statistical methods and machine learning-based methods. Usually, machine learningbased methods need to model the normal data distribution. In this work, we propose a novel anomaly detection method by modeling the data distribution of normal samples via multiple encoders and decoders. Specifically, the proposed method maps input samples into a latent space and then reconstructs output samples from latent vectors. The extra encoder finally maps reconstructed samples to latent representations. During the training phase, we optimize parameters by minimizing the reconstruction loss and encoding loss. Training samples are re-weighted to focus more on missed correlations between features of normal data. Furthermore, we employ the long short-term memory model as encoders and decoders to test its effectiveness. We also investigate a meta-learning-based framework for hyper-parameter tuning of our approach. Experiment results on network intrusion and power system datasets demonstrate the effectiveness of our proposed method, where our models consistently outperform all baselines."}, "cited_paper_content": {"title": "Smart Grid Cyber Attacks Detection Using Supervised Learning And Heuristic Feature Selection", "abstract": "False Data Injection (FDI) attacks are a common form of Cyber-attack targetting smart grids. Detection of stealthy FDI attacks is impossible by the current bad data detection systems. Machine learning is one of the alternative methods proposed to detect FDI attacks. This paper analyzes three various supervised learning techniques, each to be used with three different feature selection (FS) techniques. These methods are tested on the IEEE 14-bus, 57-bus, and 118-bus systems for evaluation of versatility. Accuracy of the classification is used as the main evaluation method for each detection technique. Simulation study clarify the supervised learning combined with heuristic FS methods result in an improved performance of the classification algorithms for FDI attack detection."}, "keywords": ["anomaly detection"], "citation_intent": "method"}
{"citing_id": "2303.09919v1", "cited_id": "1608.03983", "section_title": "Implementation Details", "citation": "As for the optimization, we use Adam optimizer with an initial learning rate of 2e-4 and a cosine annealing scheme for learning rate scheduling #REFR .", "text_before_citation": ["Implementation of EventPillars.", "In our experiments, we set the time interval of event streams to 50ms, which is same as #OTHEREFR .", "For each pillar, we set the maximum number of events per pillar M to 5 and the number of nonempty pillars K to 100k.", "In addition, the output channels of EventPillars are set to 16.", "In order to make a fair comparison for other event processing strategies, the output of all event representations are uniformly resized to 512 \u00d7 512 in both training and inference phase. Note that no extra data augmentation is applied. Implementation of DMANet. The proposed DMANet is built based on Pytorch. We use ResNet-18 #OTHEREFR as the backbone of our network."], "text_after_citation": ["The time step for training adaptive ConvLSTMs is empirically set to 10 according to GPU memory.", "All the experiments are conducted on a single NVIDIA RTX 2080Ti GPU. The code is available at https://github.com/wds320/AAAI Event based detection."], "citing_paper_content": {"title": "Dual Memory Aggregation Network For Event-Based Object Detection With Learnable Representation", "abstract": "Event-based cameras are bio-inspired sensors that capture brightness change of every pixel in an asynchronous manner. Compared with frame-based sensors, event cameras have microsecond-level latency and high dynamic range, hence showing great potential for object detection under high-speed motion and poor illumination conditions. Due to sparsity and asynchronism nature with event streams, most of existing approaches resort to hand-crafted methods to convert event data into 2D grid representation. However, they are sub-optimal in aggregating information from event stream for object detection. In this work, we propose to learn an event representation optimized for event-based object detection. Specifically, event streams are divided into grids in the x-y-t coordinates for both positive and negative polarity, producing a set of pillars as 3D tensor representation. To fully exploit information with event streams to detect objects, a dual-memory aggregation network (DMANet) is proposed to leverage both long and short memory along event streams to aggregate effective information for object detection. Long memory is encoded in the hidden state of adaptive convLSTMs while short memory is modeled by computing spatial-temporal correlation between event pillars at neighboring time intervals. Extensive experiments on the recently released event-based automotive detection dataset demonstrate the effectiveness of the proposed method."}, "cited_paper_content": {"title": "Sgdr: Stochastic Gradient Descent With Warm Restarts", "abstract": "Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets, where we demonstrate new state-of-the-art results at 3.14% and 16.21%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at https://github.com/loshchil/SGDR"}, "keywords": ["initial learning rate"], "citation_intent": "method"}
{"citing_id": "2304.05723v1", "cited_id": "math/0305433", "section_title": "C. The Dynamic Model Of A Csur", "citation": "This property is similar to the dynamics of a SIR as shown in #REFR , which provides the possibility to extend the existing results for SIRs to the virtual centers of CSURs.", "text_before_citation": ["Substituting it to (20), we have\u017c(t) = 0 which denotes that the virtual center z(t) is a static point in this case.", "Then, equation #OTHEREFR indicates that the robot is moving around z(t) along a circular orbit with a linear speed v 0 , an angular velocity \u03c9 0 , and orbit radius v 0 /|\u03c9 0 |.", "Thus, z(t) can be interpreted as the center of the circular orbit of the CSUR when it is a static point, which is why it is referred to as a virtual center.", "The relation between the CSUR position \u03b6(t) and its virtual center z(t) is shown in Fig. 2 .", "Different from the CSUR position \u03b6(t) that has to always move at a constant linear speed, the virtual center z(t) can remain static at a certain position when the CSUR is controlled with a constant input u(t) \u2261 \u03c9 0 ."], "text_after_citation": ["Therefore, in his paper, we refer to the CSUR virtual centers as CSUR agents and use them to conduct the coverage task.", "Nevertheless, it is noticed that the dynamic model of a CSUR agent in (20) is more complicated than a SIR model #OTHEREFR , which brings up challenges to this extension. Sec. III-D explains the challenges in detail.", "z(t) \u03b6(t) v 0 |w 0 |"], "citing_paper_content": {"title": "Distributed Coverage Control Of Constrained Constant-Speed Unicycle Multi-Agent Systems", "abstract": "This paper proposes a novel distributed coverage controller for a multi-agent system with constant-speed unicycle robots (CSUR). The work is motivated by the limitation of the conventional method that does not ensure the satisfaction of hard state-and input-dependent constraints and leads to feasibility issues for multi-CSUR systems. In this paper, we solve these problems by designing a novel coverage cost function and a saturated gradient-search-based control law. Invariant set theory and Lyapunov-based techniques are used to prove the state-dependent confinement and the convergence of the system state to the optimal coverage configuration, respectively. The controller is implemented in a distributed manner based on a novel communication standard among the agents. A series of simulation case studies are conducted to validate the effectiveness of the proposed coverage controller in different initial conditions and with control parameters. A comparison study in simulation reveals the advantage of the proposed method in terms of avoiding infeasibility. The experiment study verifies the applicability of the method to real robots with uncertainties. The development procedure of the method from theoretical analysis to experimental validation provides a novel framework for multiagent system coordinate control with complex agent dynamics. Note to Practitioners-This paper gives a novel solution for multiple robots to effectively cover a polygonal area. Compared to the conventional approaches, our method allows the robots to cover a target region using circular orbits, which is suitable for constant-speed unicycle robots (CSUR) like fixed-wing unmanned aerial vechicles (fUAV). The main advantage of this method is to ensure that the coverage is always successful by preventing the robots from departing the target region. Also, the method satisfies common control saturation constraints in practice and can be implemented in a reliable decentralized scheme. The method is validated to be effective for wheeled robots in experiment studies, although it can also be applied to fUAVs in theory."}, "cited_paper_content": {"title": "Nonsmooth Coordination And Geometric Optimization Via Distributed Dynamical Systems", "abstract": "This paper discusses dynamical systems for disk-covering and sphere-packing problems. We present facility location functions from geometric optimization and characterize their differentiable properties. We design and analyze a collection of distributed control laws that are related to nonsmooth gradient systems. The resulting dynamical systems promise to be of use in coordination problems for networked robots; in this setting the distributed control laws correspond to local interactions between the robots. The technical approach relies on concepts from computational geometry, nonsmooth analysis, and the dynamical system approach to algorithms."}, "keywords": ["virtual centers"], "citation_intent": "result"}
{"citing_id": "2303.15782v1", "cited_id": "1912.11913", "section_title": "Category", "citation": "Different from #REFR , we rescale and translate the closed object such that it fits in the unit cube and then backwards apply that same rescaling and translation to all objects of the same instance independent of the joint state of the object.", "text_before_citation": ["This helps with the stability of learning and simplifies hyperparameter tuning.", "This approach fails for articulated objects as their outer dimensions change depending on the joint state.", "Blindly rescaling an articulated object such that it fits inside the unit cube or unit sphere results in an inconsistent part scaling across different joint states. To mitigate this problem, #OTHEREFR proposed the NAOCS-space.", "Following their approach, first, we bring an object in its closed state (i.e.", "the lower joint limit) and put it in a canonical orientation such that for all objects, Z is pointing upwards and X back, Y is given through a right-hand coordinate frame system."], "text_after_citation": ["It is important to note that rescaling an articulated object has no impact on revolute joint states (in deg), but prismatic joint states (in m) which have to be rescaled accordingly."], "citing_paper_content": {"title": "Carto: Category And Joint Agnostic Reconstruction Of Articulated Objects", "abstract": "Figure 1. Visualization of CARTO on unseen object instances. We first use CARTO to jointly detect all objects in the scene and then articulate them while keeping the predicted shape code constant."}, "cited_paper_content": {"title": "Category-Level Articulated Object Pose Estimation", "abstract": "This paper addresses the task of category-level pose estimation for articulated objects from a single depth image. We present a novel category-level approach that correctly accommodates object instances not previously seen during training. A key aspect of the work is the new Articulation-Aware Normalized Coordinate Space Hierarchy (A-NCSH), which represents the different articulated objects for a given object category. This approach not only provides the canonical representation of each rigid part, but also normalizes the joint parameters and joint states. We developed a deep network based on PointNet++ that is capable of predicting an A-NCSH representation for unseen object instances from single depth input. The predicted A-NCSH representation is then used for global pose optimization using kinematic constraints. We demonstrate that constraints associated with joints in the kinematic chain lead to improved performance in estimating pose and relative scale for each part of the object. We also demonstrate that the approach can tolerate cases of severe occlusion in the observed data. Project webpage this https URL"}, "keywords": ["object"], "citation_intent": "method"}
{"citing_id": "2303.05227v1", "cited_id": "1902.11116", "section_title": "Impact Of References On Wikipedia:", "citation": "Our work is based on the Citation Need model #REFR , which constructs the Wikipedia verifiability taxonomy and uses machine learning to determine whether a sentence needs a citation.", "text_before_citation": ["Studies have shown that editors primarily add references to fortify the information they edit and prevent their edits from being removed #OTHEREFR .", "Numerous studies have released datasets for claim verification in specific domains, such as scientific claims #OTHEREFR .", "Studies like #OTHEREFR target cross-domain verification and show shared characteristics among these datasets.", "Our work focuses on broad topics to account for the wide variety of topics covered by Wikipedia articles.", "Compared to other work, we also consider the guidelines employed by Wikipedia's verifiability policy."], "text_after_citation": ["We expand the model into an end-to-end inference pipeline that is usable by the Wikipedia community and examine the trends seen in a decade's worth of data.", "In addition to extending previous work, we use a crowd-sourced list of domains compiled by Wikipedia editors who manually evaluated the credibility of those sources to assess the reliability #OTHEREFR ."], "citing_paper_content": {"title": "Longitudinal Assessment Of Reference Quality On Wikipedia", "abstract": "Wikipedia plays a crucial role in the integrity of the Web. This work analyzes the reliability of this global encyclopedia through the lens of its references. We operationalize the notion of reference quality by defining reference need (RN), i.e., the percentage of sentences missing a citation, and reference risk (RR), i.e., the proportion of non-authoritative references. We release Citation Detective, a tool for automatically calculating the RN score, and discover that the RN score has dropped by 20 percent point in the last decade, with more than half of verifiable statements now accompanying references. The RR score has remained below 1% over the years as a result of the efforts of the community to eliminate unreliable references. We propose pairing novice and experienced editors on the same Wikipedia article as a strategy to enhance reference quality. Our quasi-experiment indicates that such a co-editing experience can result in a lasting advantage in identifying unreliable sources in future edits. As Wikipedia is frequently used as the ground truth for numerous Web applications, our findings and suggestions on its reliability can have a far-reaching impact. We discuss the possibility of other Web services adopting Wiki-style user collaboration to eliminate unreliable content. CCS CONCEPTS \u2022 Information systems \u2192 Wikis."}, "cited_paper_content": {"title": "Citation Needed: A Taxonomy And Algorithmic Assessment Of Wikipedia'S Verifiability", "abstract": "Wikipedia is playing an increasingly central role on the web, and the policies its contributors follow when sourcing and fact-checking content affect million of readers. Among these core guiding principles, verifiability policies have a particularly important role. Verifiability requires that information included in a Wikipedia article be corroborated against reliable secondary sources. Because of the manual labor needed to curate Wikipedia at scale, however, its contents do not always evenly comply with these policies. Citations (i.e. reference to external sources) may not conform to verifiability requirements or may be missing altogether, potentially weakening the reliability of specific topic areas of the free encyclopedia. In this paper, we aim to provide an empirical characterization of the reasons why and how Wikipedia cites external sources to comply with its own verifiability guidelines. First, we construct a taxonomy of reasons why inline citations are required, by collecting labeled data from editors of multiple Wikipedia language editions. We then crowdsource a large-scale dataset of Wikipedia sentences annotated with categories derived from this taxonomy. Finally, we design algorithmic models to determine if a statement requires a citation, and to predict the citation reason . We evaluate the accuracy of such models across different classes of Wikipedia articles of varying quality, and on external datasets of claims annotated for fact-checking purposes."}, "keywords": ["Wikipedia verifiability taxonomy"], "citation_intent": "method"}
{"citing_id": "2303.05759v1", "cited_id": "1906.08237", "section_title": "Permutation Lm", "citation": "The permutation LM #REFR is a recently proposed LM that takes advantage of CLMs and masked LMs.", "text_before_citation": ["CLMs and masked LMs have their own advantages and disadvantages.", "A masked LM needs to create artificial tokens such as [mask] , which never occur in downstream tasks while CLMs only condition on preceding context."], "text_after_citation": ["Given an input sequence of linguistic units, permutation LMs randomize the order of input linguistic units and construct different permutations of the input sequence. Fig.", "2 shows an example of different permutations given an input text sequence. Let Z be the set of all possible permutations.", "Permutation LMs predict the next linguistic unit, u t , in one permutation, Z, of the sequence based on", "EQUATION", "3 Linguistic Units"], "citing_paper_content": {"title": "An Overview On Language Models: Recent Developments And Outlook", "abstract": "Language modeling studies the probability distributions over strings of texts. It is one of the most fundamental tasks in natural language processing (NLP). It has been widely used in text generation, speech recognition, machine translation, etc. Conventional language models (CLMs) aim to predict the probability of linguistic sequences in a causal manner. In contrast, pre-trained language models (PLMs) cover broader concepts and can be used in both causal sequential modeling and fine-tuning for downstream applications. PLMs have their own training paradigms (usually self-supervised) and serve as foundation models in modern NLP systems. This overview paper provides an introduction to both CLMs and PLMs from five aspects, i.e., linguistic units, structures, training methods, evaluation methods, and applications. Furthermore, we discuss the relationship between CLMs and PLMs and shed light on the future directions of language modeling in the pre-trained era."}, "cited_paper_content": {"title": "Xlnet: Generalized Autoregressive Pretraining For Language Understanding", "abstract": "With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves better performance than pretraining approaches based on autoregressive language modeling. However, relying on corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model, into pretraining. Empirically, under comparable experiment setting, XLNet outperforms BERT on 20 tasks, often by a large margin, including question answering, natural language inference, sentiment analysis, and document ranking."}, "keywords": ["recently proposed LM"], "citation_intent": "background"}
{"citing_id": "2304.03110v1", "cited_id": "1405.0312", "section_title": "A. Traditional Iod Protocol And Results", "citation": "Formally, let D = {(x, y)} be a dataset of images x with corresponding object annotations y, such as COCO 2017 #REFR , and let C = {1, . . . , C} be the set of object categories.", "text_before_citation": ["In previous work #OTHEREFR , in each phase, the incremental detector is allowed to observe all images that contain a certain type of objects.", "Because images often contain a mix of object classes, both old and new, this means that the same images can be observed in different training phases.", "This is incompatible with the standard definition of incremental learning #OTHEREFR where, with the exception of the examples deliberately stored in the exemplar memory, the images observed in different phases do not repeat.", "Thus, we provide results with our new IOD benchmark protocol in the main paper.", "For completeness and comparison, here we also evaluate performance using the traditional IOD benchmark protocol used in #OTHEREFR , and provide comparison results between our method and other top-performing IOD methods with this protocol. Traditional IOD protocol."], "text_after_citation": ["We adopt such a dataset for benchmarking IOD as follows.", "First, we partition C into M subsets C = C 1 \u222a \u2022 \u2022 \u2022 \u222a C M , one for each training phase.", "For each phase i, we modify the samples (x, y) \u2208 D, where y only contains annotations for objects of class C i and drop the others.", "In phase i of training, the model is only allowed to observe images that contain at least one annotation for objects of types C i \u2282 C. Experiment results.", "Table S1 shows that also with the traditional IOD protocol our CL-DETR consistently performs better than the state-of-the-art #OTHEREFR and other IOD methods #OTHEREFR ."], "citing_paper_content": {"title": "Continual Detection Transformer For Incremental Object Detection", "abstract": "Incremental object detection (IOD) aims to train an object detector in phases, each with annotations for new object categories. As other incremental settings, IOD is subject to catastrophic forgetting, which is often addressed by techniques such as knowledge distillation (KD) and exemplar replay (ER). However, KD and ER do not work well if applied directly to state-of-the-art transformer-based object detectors such as Deformable DETR [61] and UP-DETR [10]. In this paper, we solve these issues by proposing a ContinuaL DEtection TRansformer (CL-DETR), a new method for transformer-based IOD which enables effective usage of KD and ER in this context. First, we introduce a Detector Knowledge Distillation (DKD) loss, focusing on the most informative and reliable predictions from old versions of the model, ignoring redundant background predictions, and ensuring compatibility with the available ground-truth labels. We also improve ER by proposing a calibration strategy to preserve the label distribution of the training set, therefore better matching training and testing statistics. We conduct extensive experiments on COCO 2017 and demonstrate that CL-DETR achieves state-of-theart results in the IOD setting. 1"}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["object categories"], "citation_intent": "background"}
{"citing_id": "2304.09145v1", "cited_id": "1810.04805", "section_title": "Introduction", "citation": "Among these techniques, quantization #REFR has emerged as a general and primary paradigm for reducing both memory footprint and computational overhead.", "text_before_citation": ["Transformer language models (e.g., BERT, LLMs) have garnered significant attention due to their remarkable performance and scalable model size.", "These models have evolved from hundreds of millions of parameters #OTHEREFR to hundreds of billions of parameters #OTHEREFR .", "This necessitates the employment of compression techniques #OTHEREFR Le-Cun et al., 1989) for practical deployment."], "text_after_citation": ["However, quantization, particularly post-training quantization (PTQ) #OTHEREFR , which operates with limited data and GPU resources, has become increasingly challenging for these models (e.g., a 12% accuracy drop in BERT #OTHEREFR and catastrophic degradation in OPT-175B ).", "This is caused by the presence of detrimental outliers in activation (e.g., the width of distribution can be 80 for BERT and even 140 for OPTs), which prevents quantization from accurately expressing floating-point numbers.", "To combat the outlier bottleneck in quantization, researchers make in-depth investigations and find the concentration phenomenon that outliers typically appear in certain channels.", "Hence, some works #OTHEREFR explored various fine-grained quantization schemes to provide additional consideration for them while some #OTHEREFR attempted to diminish them through a scaling operation, which can be migrated to subsequent modules to maintain floating-point (FP) equivalence.", "However, we find that they ignore the asymmetric property of outliers that relatively small channel ranges can result in a significantly large range for the entire tensor (e.g., as shown in Fig."], "citing_paper_content": {"title": "Outlier Suppression+: Accurate Quantization Of Large Language Models By Equivalent And Optimal Shifting And Scaling", "abstract": "Quantization of transformer language models faces significant challenges due to the existence of detrimental outliers in activations. We observe that these outliers are asymmetric and concentrated in specific channels. To address this issue, we propose the Outlier Suppres-sion+ framework. First, we introduce channelwise shifting and scaling operations to eliminate asymmetric presentation and scale down problematic channels. We demonstrate that these operations can be seamlessly migrated into subsequent modules while maintaining equivalence. Second, we quantitatively analyze the optimal values for shifting and scaling, taking into account both the asymmetric property and quantization errors of weights in the next layer. Our lightweight framework can incur minimal performance degradation under static and standard post-training quantization settings. Comprehensive results across various tasks and models reveal that our approach achieves near-floating-point performance on both small models, such as BERT, and large language models (LLMs) including OPTs, BLOOM, and BLOOMZ at 8-bit and 6bit settings. Furthermore, we establish a new state of the art for 4-bit BERT."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["quantization"], "citation_intent": "background"}
{"citing_id": "2303.13477v1", "cited_id": "1812.01024", "section_title": "Deep Directional Distance Function", "citation": "Inspired by DeepVoxels #REFR , which improves the accuracy of RGB novel view synthesis with voxel features, we introduce a 3D voxel representation of the rough object shape before the decoder which ensures multi-view consistency of output depth images.", "text_before_citation": ["DeepDDF enables direct evaluation of the object shape and pose against observations in the image-space as it bypasses rendering (projection) of 3D geometry in the 2D view.", "Two fundamental challenges, however, need to be resolved to achieve this direct 2D view-dependent representation.", "The first is that the depth values for each camera ray across different viewpoints must be consistent, i.e., must represent the same rigid object shape.", "The second is the large number of samples required to span the entire space of the inputs to the decoder p ray , d ray , and z.", "We overcome these challenges by introducing a sparse 3D voxel grid aligned with the target object space."], "text_after_citation": ["This 3D voxel can be of very low-resolution as it does not encode the 3D shape itself.", "As shown in Fig.", "2 , the latent code z is decoded into latent voxels by an object instance decoder f d , and the voxels on the camera rays, defined by its 3D viewpoint p ray and 3D direction d ray , are sampled by a camera ray sampler f mlp .", "This low-resolution but three-dimensional voxel corresponding to the object instance encoded by the latent code z enforces consistency across the camera rays and makes the output depth values D multi-view consistent.", "It also enables reduction in the number of samples for p ray , d ray , and z by ensuring this multi-view consistency."], "citing_paper_content": {"title": "Transposer: Transformer As An Optimizer For Joint Object Shape And Pose Estimation", "abstract": "Transformer Decoder TransPoser momentum cross-attn Transformer Encoder Figure 1: We propose TransPoser, a novel neural optimization network for joint object pose and shape reconstruction from sequential observations (a). We also derive DeepDDF, a novel category-level neural 3D shape representation that directly outputs 2D depth images for given viewing conditions. By integrating DeepDDF and TransPoser, we realize efficient and accurate learned weighted optimization with learned momentum for joint shape and pose estimation (b)."}, "cited_paper_content": {"title": "Deepvoxels: Learning Persistent 3D Feature Embeddings", "abstract": "In this work, we address the lack of 3D understanding of generative neural networks by introducing a persistent 3D feature embedding for view synthesis. To this end, we propose DeepVoxels, a learned representation that encodes the view-dependent appearance of a 3D scene without having to explicitly model its geometry. At its core, our approach is based on a Cartesian 3D grid of persistent embedded features that learn to make use of the underlying 3D scene structure. Our approach combines insights from 3D geometric computer vision with recent advances in learning image-to-image mappings based on adversarial loss functions. DeepVoxels is supervised, without requiring a 3D reconstruction of the scene, using a 2D re-rendering loss and enforces perspective and multi-view geometry in a principled manner. We apply our persistent 3D scene representation to the problem of novel view synthesis demonstrating high-quality results for a variety of challenging scenes."}, "keywords": ["DeepVoxels"], "citation_intent": "method"}
{"citing_id": "2305.01140v1", "cited_id": "1906.04015", "section_title": "Qm9", "citation": "Following #REFR , we split the train, validation, and test partitions, with 100K, 18K, and 13K samples.", "text_before_citation": ["2014) for both unconditional and conditional molecule generation.", "QM9 is one of the most widely-used datasets for molecular machine learning research, which has also been adopted in previous 3D molecule generation studies #OTHEREFR .", "QM9 contains 3D structures together with several quantum properties for 130k small molecules, limited to 9 heavy atoms (29 atoms including hydrogens)."], "text_after_citation": ["For the molecule generation task, we also test GEOLDM on the GEOM-DRUG (Geometric Ensemble Of Molecules) dataset.", "The DRUG dataset consists of much larger organic compounds, with up to 181 atoms and 44.2 atoms on average, in 5 different atom types.", "It covers 37 million molecular conformations for around 450,000 molecules, labeled with energy and statistical weight.", "We follow the common practice #OTHEREFR to select the 30 lowest energy conformations of each molecule for training."], "citing_paper_content": {"title": "Geometric Latent Diffusion Models For 3D Molecule Generation", "abstract": "Generative models, especially diffusion models (DMs), have achieved promising results for generating feature-rich geometries and advancing foundational science problems such as molecule design. Inspired by the recent huge success of Stable (latent) Diffusion models, we propose a novel and principled method for 3D molecule generation named Geometric Latent Diffusion Models (GEOLDM). GEOLDM is the first latent DM model for the molecular geometry domain, composed of autoencoders encoding structures into continuous latent codes and DMs operating in the latent space. Our key innovation is that for modeling the 3D molecular geometries, we capture its critical roto-translational equivariance constraints by building a point-structured latent space with both invariant scalars and equivariant tensors. Extensive experiments demonstrate that GEOLDM can consistently achieve better performance on multiple molecule generation benchmarks, with up to 7% improvement for the valid percentage of large biomolecules. Results also demonstrate GEOLDM's higher capacity for controllable generation thanks to the latent modeling. Code is provided at https: //github.com/MinkaiXu/GeoLDM."}, "cited_paper_content": {"title": "Cormorant: Covariant Molecular Neural Networks", "abstract": "We propose Cormorant, a rotationally covariant neural network architecture for learning the behavior and properties of complex many-body physical systems. We apply these networks to molecular systems with two goals: learning atomic potential energy surfaces for use in Molecular Dynamics simulations, and learning ground state properties of molecules calculated by Density Functional Theory. Some of the key features of our network are that (a) each neuron explicitly corresponds to a subset of atoms; (b) the activation of each neuron is covariant to rotations, ensuring that overall the network is fully rotationally invariant. Furthermore, the non-linearity in our network is based upon tensor products and the Clebsch-Gordan decomposition, allowing the network to operate entirely in Fourier space. Cormorant significantly outperforms competing algorithms in learning molecular Potential Energy Surfaces from conformational geometries in the MD-17 dataset, and is competitive with other methods at learning geometric, energetic, electronic, and thermodynamic properties of molecules on the GDB-9 dataset."}, "keywords": ["13K samples", "train"], "citation_intent": "method"}
{"citing_id": "2303.14423v1", "cited_id": "1811.11682", "section_title": "Problem Description", "citation": "Additionally, we rely on pseudo-rehearsal through experience replay #REFR using a small memory buffer to tackle catastrophic forgetting. Figure 2 visualizes our transformer-based architecture for multimodal CL.", "text_before_citation": ["Because when the model is updated to learn the current tasks, its performance on the past learned tasks will degrade due to catastrophic forgetting #OTHEREFR .", "Figure 1 visualizes the high-level presentation of the solution that we propose to address multimodal CL.", "To use a shared model across all tasks and benefit from knowledge transfer across all tasks, we consider a base transformer model as f \u03b8 (\u2022, \u2022) and make it adaptive by adding a unique task attention layer after its final layer.", "Additionally, we modify the loss supervised learning loss by adding a knowledge distillation loss #OTHEREFR on the intermediate model layers.", "We consider the teacher model in the knowledge distillation formulation to be a copy of f i\u22121 \u03b8 (\u2022, \u2022) when training the main student model on i th task T i ."], "text_after_citation": ["The architecture consists of a shared pre-trained replaceable bimodal transformer, an independent task attention block, and MLP classification headers.", "The task attention block receives task-specific tokens that make the model adaptive.", "We provide details about these modules and the strategies that we use to train it."], "citing_paper_content": {"title": "Task-Attentive Transformer Architecture For Continual Learning Of Vision-And-Language Tasks Using Knowledge Distillation", "abstract": "The size and the computational load of fine-tuning largescale pre-trained neural network are becoming two major obstacles in adopting machine learning in many applications. Continual learning (CL) can serve as a remedy through enabling knowledge-transfer across sequentially arriving tasks which relaxes the need to fine-tune all network weights from scratch. However, existing CL algorithms primarily consider learning unimodal vision-only or language-only tasks. We develop a transformer-based CL architecture for learning bimodal vision-and-language tasks based on increasing the number of the learnable parameters dynamically and using knowledge distillation. The new additional parameters are used to specialize the network for each task. Our approach enables sharing information between the tasks while addressing the challenge of catastrophic forgetting. Our approach is scalable learning to a large number of tasks because it requires little memory and time overhead. Our model reaches state-of-the-art performance on challenging vision-and-language tasks."}, "cited_paper_content": {"title": "Experience Replay For Continual Learning", "abstract": "Interacting with a complex world involves continual learning, in which tasks and data distributions change over time. A continual learning system should demonstrate both plasticity (acquisition of new knowledge) and stability (preservation of old knowledge). Catastrophic forgetting is the failure of stability, in which new experience overwrites previous experience. In the brain, replay of past experience is widely believed to reduce forgetting, yet it has been largely overlooked as a solution to forgetting in deep reinforcement learning. Here, we introduce CLEAR, a replay-based method that greatly reduces catastrophic forgetting in multi-task reinforcement learning. CLEAR leverages off-policy learning and behavioral cloning from replay to enhance stability, as well as on-policy learning to preserve plasticity. We show that CLEAR performs better than state-of-the-art deep learning techniques for mitigating forgetting, despite being significantly less complicated and not requiring any knowledge of the individual tasks being learned."}, "keywords": ["catastrophic forgetting", "experience replay"], "citation_intent": "method"}
{"citing_id": "2303.16898v1", "cited_id": "1812.03982", "section_title": "C. Interactive Perception", "citation": "We use a SlowFast network with a ResNet-50 backbone #REFR , which takes in 32 images of size 224 \u00d7 224 sampled with a uniform interval from the video stream. Fig.", "text_before_citation": ["Thus, we tilt the gripper at an angle \u03b8 = 50 \u2022 so the grasp point is visible in the camera.", "For (2), we use a triangular trajectory, where the robot gripper first moves backward, then upward, and finally forward and downward back to the original position.", "To prevent the deformable object from translating as a whole, we use the robot's second gripper to pin the other side. See Figure 3 for a visualization. In our implementation, the trajectory takes about 5 secs.", "While the robot executes the trajectory T , the camera takes an RGB video stream of the bag.", "A video classification model M takes the video and classifies the grasp into 3 categories: 0 layer, 1 layer, and 2 layers."], "text_after_citation": ["3 illustrates the visual differences among different layers grasped on a plastic bag.", "Given the model classification, SLIP adjusts the gripper height and retries the grasp if it does not successfully grasp a single layer.", "We choose to use a fixed height adjustment each time which is similar to the strategy in #OTHEREFR , with height deltas \u2206h \u2212 and \u2206h + .", "One could also choose to let the adjustment height decay over time or use the bisection method, but we empirically find the numbers of trials these approaches take are similar while a fixed height adjustment is more robust to a long sequence of iterations than a decaying step size, and more robust to minor bag state changes between grasps and model classification errors than bisection, since once bisection enters into a wrong interval, it may not succeed. The pseudocode of SLIP is provided in Algorithm 1."], "citing_paper_content": {"title": "Bagging By Learning To Singulate Layers Using Interactive Perception", "abstract": "Many fabric handling and 2D deformable material tasks in homes and industry require singulating layers of material such as opening a bag or arranging garments for sewing. In contrast to methods requiring specialized sensing or end effectors, we use only visual observations with ordinary parallel jaw grippers. We propose SLIP: Singulating Layers using Interactive Perception, and apply SLIP to the task of autonomous bagging. We develop SLIP-Bagging, a bagging algorithm that manipulates a plastic or fabric bag from an unstructured state, and uses SLIP to grasp the top layer of the bag to open it for object insertion. In physical experiments, a YuMi robot achieves a success rate of 67% to 81% across bags of a variety of materials, shapes, and sizes, significantly improving in success rate and generality over prior work. Experiments also suggest that SLIP can be applied to tasks such as singulating layers of folded cloth and garments. Supplementary material is available at https://sites.google.com/view/slip-bagging/."}, "cited_paper_content": {"title": "Slowfast Networks For Video Recognition", "abstract": "We present SlowFast networks for video recognition. Our model involves (i) a Slow pathway, operating at low frame rate, to capture spatial semantics, and (ii) a Fast pathway, operating at high frame rate, to capture motion at fine temporal resolution. The Fast pathway can be made very lightweight by reducing its channel capacity, yet can learn useful temporal information for video recognition. Our models achieve strong performance for both action classification and detection in video, and large improvements are pin-pointed as contributions by our SlowFast concept. We report state-of-the-art accuracy on major video recognition benchmarks, Kinetics, Charades and AVA. Code has been made available at: https://github.com/facebookresearch/SlowFast."}, "keywords": ["Fig", "SlowFast network"], "citation_intent": "method"}
{"citing_id": "2305.00447v1", "cited_id": "1810.04805", "section_title": "Baseline.", "citation": "This is a variant of GRU4Rec that incorporates a pre-trained LLM model BERT #REFR to leverage the text descriptions of items.", "text_before_citation": ["We use the version of DROS implemented on GRU4Rec, provided by the authors 3 .", "The default implementation of these methods relies solely on item ID information to construct the recommender model.", "However, in our setting, we assume item text descriptions are available, and our method would utilize these text descriptions to generate recommendations.", "To ensure fair comparisons between our approach and the baseline methods, we further consider comparing the following variants of GRU4Rec and DROS:", "-GRU-BERT."], "text_after_citation": ["Specifically, we feed the item text descriptions into BERT and concatenate the resulting CLS embedding, which encodes text descriptions, with the initial item embedding of GRU4Rec.", "We take the combined embedding as the item representation used by GRU4Rec to generate recommendation predictions.", "-DROS-BERT.", "This variant of DROS is similarly modified to incorporate BERT, allowing it to utilize the text information of items."], "citing_paper_content": {"title": "Tallrec: An Effective And Efficient Tuning Framework To Align Large Language Model With Recommendation", "abstract": "Large Language Models (LLMs) have demonstrated remarkable performance across diverse domains, thereby prompting researchers to explore their potential for use in recommendation systems. Initial attempts have leveraged the exceptional capabilities of LLMs, such as rich knowledge and strong generalization through In-context Learning, which involves phrasing the recommendation task as prompts. Nevertheless, the performance of LLMs in recommendation tasks remains suboptimal due to a substantial disparity between the training tasks for LLMs and recommendation tasks, as well as inadequate recommendation data during pre-training. To bridge the gap, we consider building a Large Recommendation Language Model by tunning LLMs with recommendation data. To this end, we propose an efficient and effective Tuning framework for Aligning LLMs with Recommendation, namely TALLRec. We have demonstrated that the proposed TALLRec framework can significantly enhance the recommendation capabilities of LLMs in the movie and book domains, even with a limited dataset of fewer than 100 samples. Additionally, the proposed framework is highly efficient and can be executed on a single RTX 3090 with LLaMA-7B. Furthermore, the fine-tuned LLM exhibits robust cross-domain generalization. Our code and data are available at https://github.com/SAI990323/TALLRec."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["text descriptions", "BERT"], "citation_intent": "method"}
{"citing_id": "2304.11265v1", "cited_id": "1909.04939", "section_title": "A. Findings And Explanations", "citation": "This variability aligns with the finding of Fawaz et al., who made InceptionTime an ensemble of Inception networks due to the high variability of a single Inception network #REFR .", "text_before_citation": ["For example, the tuned Incep-tionTime tremor classifier features 7499 trainable parameters, whereas the default four-class InceptionTime includes 490 564 trainable parameters, but each produces similar scores across two different sensors (GENEActiv and Shimmer).", "This comparable performance on the test data before and after hyperparameter tuning may indicate that the hyperparameter search leads to overfitting.", "Alternatively, hyperparameters-with the exception of window length-may not substantially impact the InceptionTime performance, as seen in Section II-C.", "After all, hyperparameter optimization reliably selects long windows; the hyperparameter search might randomly set the other hyperparameters because they do not matter.", "5) InceptionTime has a high score variability: InceptionTime demonstrates greater score variability between training runs compared with other approaches, independent of whether automated hyperparameter tuning has been performed."], "text_after_citation": ["However, we show that even ensembling cannot fully compensate for the variability of Inception networks for TSC.", "InceptionTime's sensitivity to random initialization may also contribute to the futility of automated hyperparameter optimization.", "If InceptionTime reacts to several thousand trainable parameters that cannot be fully optimized by training, the impact of four additional hyperparameters may be negligible.", "This is evidenced by the large standard deviations of the model scores from the cross-validation (see Section II-C.3).", "6) ROCKET is promising, especially for dyskinesia: For dyskinesia, ROCKET has the highest mean AP and is the only classifier that substantially outperforms a RC."], "citing_paper_content": {"title": "Time Series Classification For Detecting Parkinson'S Disease From Wrist Motions", "abstract": "Parkinson's disease (PD) is a neurodegenerative disease with frequently changing motor symptoms where continuous symptom monitoring enables more targeted treatment. Classical time series classification (TSC) and deep learning techniques have limited performance for PD symptom monitoring using wearable accelerometer data because PD movement patterns are complex, but datasets are small. We investigate InceptionTime and RandOm Convolutional KErnel Transform (ROCKET) because they are state-of-the-art for TSC and promising for PD symptom monitoring: InceptionTime's high learning capacity is suited to modeling complex movement patterns while ROCKET is suited to small datasets. We used a random search to find the highest-scoring InceptionTime architecture and compared it to ROCKET with a ridge classifier and a multi-layer perceptron (MLP) on wrist motions of PD patients. We find that all approaches are suitable for estimating tremor severity and bradykinesia presence but struggle with detecting dyskinesia. ROCKET performs better for dyskinesia, whereas InceptionTime is slightly better for tremor and bradykinesia but has much higher variability in performance. Both outperform the MLP. In conclusion, both InceptionTime and ROCKET are suitable for continuous symptom monitoring, with the choice depending on the symptom of interest and desired robustness."}, "cited_paper_content": {"title": "Inceptiontime: Finding Alexnet For Time Series Classification", "abstract": "Time series classification (TSC) is the area of machine learning interested in learning how to assign labels to time series. The last few decades of work in this area have led to significant progress in the accuracy of classifiers, with the state of the art now represented by the HIVE-COTE algorithm. While extremely accurate, HIVE-COTE is infeasible to use in many applications because of its very high training time complexity in O(N^2*T^4) for a dataset with N time series of length T. For example, it takes HIVE-COTE more than 72,000s to learn from a small dataset with N=700 time series of short length T=46. Deep learning, on the other hand, has now received enormous attention because of its high scalability and state-of-the-art accuracy in computer vision and natural language processing tasks. Deep learning for TSC has only very recently started to be explored, with the first few architectures developed over the last 3 years only. The accuracy of deep learning for TSC has been raised to a competitive level, but has not quite reached the level of HIVE-COTE. This is what this paper achieves: outperforming HIVE-COTE's accuracy together with scalability. We take an important step towards finding the AlexNet network for TSC by presenting InceptionTime---an ensemble of deep Convolutional Neural Network (CNN) models, inspired by the Inception-v4 architecture. Our experiments show that InceptionTime slightly outperforms HIVE-COTE with a win/draw/loss on the UCR archive of 40/6/39. Not only is InceptionTime more accurate, but it is much faster: InceptionTime learns from that same dataset with 700 time series in 2,300s but can also learn from a dataset with 8M time series in 13 hours, a quantity of data that is fully out of reach of HIVE-COTE."}, "keywords": ["Inception networks"], "citation_intent": "result"}
{"citing_id": "2304.08027v1", "cited_id": "1801.04381", "section_title": "Human Detection", "citation": "In this research, the MobileNetV2 algorithm in the SSDLite framework #REFR was used to detect and determine the 'image's human position in real-time.", "text_before_citation": [], "text_after_citation": ["The MobileNet neural network architecture has been designed specifically for limited hardware resources such as mobiles.", "With a dramatic decrease in the number of parameters and memory required, maintaining the same accuracy has become an advanced state-of-the-art model in computer vision for mobiles #OTHEREFR . In the following section, its architecture is introduced.", "Table 2 Shows the MobileNetV2 architecture, which contains the initial fully convolution layer with 32 filters, followed by 19 residual bottleneck layers described in Table 3 #OTHEREFR .", "In this architecture, the ReLU6 activation function is used as the non-linearity because of its robustness when used with low-precision computation #OTHEREFR , #OTHEREFR .", "Table 3 MobileNetV2: Each line describes a sequence of 1 or more identical layers, repeated n times."], "citing_paper_content": {"title": "Deeplt: Personalized Lighting Facilitates By Trajectory Prediction Of Recognized Residents In The Smart Home", "abstract": "In recent years, the intelligence of various parts of the home has become one of the essential features of any modern home. One of these parts is the intelligence lighting system that personalizes the light for each person. This paper proposes an intelligent system based on machine learning that personalizes lighting in the instant future location of a recognized user, inferred by trajectory prediction. Our proposed system consists of the following modules: (I) human detection to detect and localize the person in each given video frame, (II) face recognition to identify the detected person, (III) human tracking to track the person in the sequence of video frames and (IV) trajectory prediction to forecast the future location of the user in the environment using Inverse Reinforcement Learning. The proposed method provides a unique profile for each person, including specifications, face images, and custom lighting settings. This profile is used in the lighting adjustment process. Unlike other methods that consider constant lighting for every person, our system can apply each 'person's desired lighting in terms of color and light intensity without direct user intervention. Therefore, the lighting is adjusted with higher speed and better efficiency. In addition, the predicted trajectory path makes the proposed system apply the desired lighting, creating more pleasant and comfortable conditions for the home residents. In the experimental results, the system applied the desired lighting in an average time of 1.4 seconds from the moment of entry, as well as a performance of 22.1mAp in"}, "cited_paper_content": {"title": "Mobilenetv2: Inverted Residuals And Linear Bottlenecks", "abstract": "In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes. We also describe efficient ways of applying these mobile models to object detection in a novel framework we call SSDLite. Additionally, we demonstrate how to build mobile semantic segmentation models through a reduced form of DeepLabv3 which we call Mobile DeepLabv3. The MobileNetV2 architecture is based on an inverted residual structure where the input and output of the residual block are thin bottleneck layers opposite to traditional residual models which use expanded representations in the input an MobileNetV2 uses lightweight depthwise convolutions to filter features in the intermediate expansion layer. Additionally, we find that it is important to remove non-linearities in the narrow layers in order to maintain representational power. We demonstrate that this improves performance and provide an intuition that led to this design. Finally, our approach allows decoupling of the input/output domains from the expressiveness of the transformation, which provides a convenient framework for further analysis. We measure our performance on Imagenet classification, COCO object detection, VOC image segmentation. We evaluate the trade-offs between accuracy, and number of operations measured by multiply-adds (MAdd), as well as the number of parameters"}, "keywords": ["real-time", "MobileNetV2 algorithm"], "citation_intent": "method"}
{"citing_id": "2304.14663v1", "cited_id": "1602.05629", "section_title": "Federated Learning Basics", "citation": "Federated Learning #REFR is a technique that allows for a central model to be trained over distributed data that originate and are stored locally.", "text_before_citation": [], "text_after_citation": ["This approach was originally proposed within the setting of large-scale distributed learning on mobile devices and has been widely used and researched in the medical sector given the privacy enhancing aspect as well a the regulatory restrictions on data sharing within the sector.", "Each local data source, e.g., a mobile phone or a server hosted at a hospital, is referred to as a client.", "The group of clients that contribute to training a central model, including the central server is referred to as the federation.", "The central server orchestrates the learning process over the different clients following the predefined FL algorithm.", "In that sense, the central server is holistically responsible for (i) initiating the model; (ii) providing a copy of the model to all of the clients; (iii) aggregating over the received model parameters and (iv) send back the updated model to the clients."], "citing_paper_content": {"title": "Client Recruitment For Federated Learning In Icu Length Of Stay Prediction", "abstract": "Background: Machine and deep learning methods for medical and healthcare applications have shown significant progress and performance improvement in recent years. These methods require vast amounts of training data which are available in the medical sector, albeit decentralized. Medical institutions generate vast amounts of data for which sharing and centralizing remains a challenge as the result of data and privacy regulations. The federated learning technique is well-suited to tackle these challenges. However, federated learning comes with a new set of open problems related to communication overhead, efficient parameter aggregation, client selection strategies and more. Methods: In this work, we address the step prior to the initiation of a federated network for model training, client recruitment. By intelligently recruiting clients, communication overhead and overall cost of training can be reduced without sacrificing predictive performance. Client recruitment aims at pre-excluding potential clients from partaking in the federation based on a set of criteria indicative of their eventual contributions to the federation. In this work, we propose a client recruitment approach using only the output distribution and sample size at the client site. Results: We show how a subset of clients can be recruited without sacrificing model performance whilst, at the same time, significantly improving computation time. By applying the recruitment approach to the training of federated models for accurate patient Length of Stay prediction using data from 189 Intensive Care Units, we show how the models trained in federations made up from recruited clients significantly outperform federated models trained with the standard procedure in terms of predictive power and training time."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["Federated Learning"], "citation_intent": "background"}
{"citing_id": "2303.12212v1", "cited_id": "1208.5092", "section_title": "Hierarchical Clustering", "citation": "For example, we could generate a graph using -nearest neighbours methods and define a criterion using the degrees of nodes #REFR .", "text_before_citation": ["3. average-linkage clustering", "EQUATION", "4. Ward linkage clustering", "EQUATION", "It must be noted that hierarchical clustering methods may also be based on graphs."], "text_after_citation": ["The single-linkage criterion is relatively naive and often does not produce suitable partitions. The remaining ones have higher time complexity.", "One way to bypass the complexity problem is to add additional requirements.", "For example, the Genie algorithm #OTHEREFR uses a single-linkage criterion but simultaneously monitors the partition's Gini Index value, i.e., a measure of inequality of cluster sizes.", "Only the cluster of the smallest cardinality can be merged when the algorithm reaches a pre-set threshold of the Gini Index value."], "citing_paper_content": {"title": "Community Detection In Complex Networks Via Node Similarity, Graph Representation Learning, And Hierarchical Clustering", "abstract": "Community detection is a critical challenge in the analysis of real-world graphs and complex networks, including social, transportation, citation, cybersecurity networks, and food webs. Motivated by many similarities between community detection and clustering in Euclidean spaces, we propose three algorithm frameworks to apply hierarchical clustering methods for community detection in graphs. We show that using our methods, it is possible to apply various linkagebased (single-, complete-, average-linkage, Ward, Genie) clustering algorithms to find communities based on vertex similarity matrices, eigenvector matrices thereof, and Euclidean vector representations of nodes. We convey a comprehensive analysis of choices for each framework, including state-of-the-art graph representation learning algorithms, such as Deep Neural Graph Representation, and a vertex proximity matrix known to yield high-quality results in machine learning-Positive Pointwise Mutual Information. Overall, we test over a hundred combinations of framework components and show that some-including Wasserman-Faust and PPMI proximity, DNGR representation-can compete with algorithms such as state-of-the-art Leiden and Louvain and easily outperform other known community detection algorithms. Notably, our algorithms remain hierarchical and allow the user to specify any number of clusters a priori."}, "cited_paper_content": {"title": "Graph Degree Linkage: Agglomerative Clustering On A Directed Graph", "abstract": "This paper proposes a simple but effective graph-based agglomerative algorithm, for clustering high-dimensional data. We explore the different roles of two fundamental concepts in graph theory, indegree and outdegree, in the context of clustering. The average indegree reflects the density near a sample, and the average outdegree characterizes the local geometry around a sample. Based on such insights, we define the affinity measure of clusters via the product of average indegree and average outdegree. The product-based affinity makes our algorithm robust to noise. The algorithm has three main advantages: good performance, easy implementation, and high computational efficiency. We test the algorithm on two fundamental computer vision problems: image clustering and object matching. Extensive experiments demonstrate that it outperforms the state-of-the-arts in both applications."}, "keywords": ["-nearest neighbours methods"], "citation_intent": "method"}
{"citing_id": "2305.01294v1", "cited_id": "1304.6763", "section_title": "Morphing Images", "citation": "In the next step, we extract the features using a Wavelet Scattering Network (WSN) #REFR independently on each filtered image.", "text_before_citation": ["We perform the Laplacian filtering independently on color channels that will results in", "\u2022 Novel", "L Y s , L B s , L R s and L Y t , L B t , L R t", "respectively.", "Figure 3 shows the qualitative results of the Laplacian filtering results of the example face image."], "text_after_citation": ["In this work, we construct a two-layer image scatter network with a 250 \u00d7 250 pixels with invariant scale such that two wavelets per octave in the first layer and one wavelet per octave in the second layer. Furthermore, we use six rotations of wavelets per layer.", "Therefore, the WSN used in this work has 577 paths.", "Let the WSN computed for Laplacian filtered images be denoted as:", "W Y s , W B s ,", "F D Y = W Y s \u2212 W Y t , F D Y = W B s \u2212 W B t and F D Y = W R s \u2212 W R t ."], "citing_paper_content": {"title": "Differential Newborn Face Morphing Attack Detection Using Wavelet Scatter Network", "abstract": "Face Recognition System (FRS) are shown to be vulnerable to morphed images of newborns. Detecting morphing attacks stemming from face images of newborn is important to avoid unwanted consequences, both for security and society. In this paper, we present a new reference-based/Differential Morphing Attack Detection (MAD) method to detect newborn morphing images using Wavelet Scattering Network (WSN). We propose a two-layer WSN with 250 \u00d7 250 pixels and six rotations of wavelets per layer, resulting in 577 paths. The proposed approach is validated on a dataset of 852 bona fide images and 2460 morphing images constructed using face images of 42 unique newborns. The obtained results indicate a gain of over 10% in detection accuracy over other existing D-MAD techniques."}, "cited_paper_content": {"title": "Deep Scattering Spectrum", "abstract": "A scattering transform defines a locally translation invariant representation which is stable to time-warping deformation. It extends MFCC representations by computing modulation spectrum coefficients of multiple orders, through cascades of wavelet convolutions and modulus operators. Second-order scattering coefficients characterize transient phenomena such as attacks and amplitude modulation. A frequency transposition invariant representation is obtained by applying a scattering transform along log-frequency. State-the-of-art classification results are obtained for musical genre and phone classification on GTZAN and TIMIT databases, respectively."}, "keywords": ["Wavelet Scattering Network"], "citation_intent": "method"}
{"citing_id": "2304.07939v1", "cited_id": "1911.08731", "section_title": "Domain Generalization", "citation": "For CivilComments we report the average and worst accuracy in Table 3 , where we compare with ERM and groupDRO #REFR . While performing almost on par w.r.t.", "text_before_citation": ["Our claim is that a feature space that satisfies sufficiency and minimality is more robust to spurious correlations which may affect minority groups, and should transfer better to new distributions.", "To validate this, we test on two benchmarks Waterbirds #OTHEREFR , and CivilComments #OTHEREFR ) (see Appendix C.1).", "For both, we use the train and test split of the original dataset.", "In Table 5 we report the results on the test set of Waterbirds for the different groups in the dataset (landbirds on land, landbirds on water, waterbirds on land, and waterbirds on water, respectively).", "We fit the linear head on a random subset of the training domain, balanced by class, repeat 10 times and report accuracy and standard deviation on test."], "text_after_citation": ["ERM, our method is more robust to spurious correlation in the dataset, showing the higher worst group accuracy.", "Importantly, we outperform GroupDRO, which uses information on the subdomain statistics, while we do not assume any prior knowledge about them.", "Results per group are reported in the Appendix (Table 9) .", "DomainBed.", "We evaluate the domain generalization performance on the PACS, VLCS and OfficeHome datasets from the DomainBed #OTHEREFR ) test suite (see Appendix C.1 for more details)."], "citing_paper_content": {"title": "Leveraging Sparse And Shared Feature Activations For Disentangled Representation Learning", "abstract": "Recovering the latent factors of variation of high dimensional data has so far focused on simple synthetic settings. Mostly building on unsupervised and weakly-supervised objectives, prior work missed out on the positive implications for representation learning on real world data. In this work, we propose to leverage knowledge extracted from a diversified set of supervised tasks to learn a common disentangled representation. Assuming each supervised task only depends on an unknown subset of the factors of variation, we disentangle the feature space of a supervised multi-task model, with features activating sparsely across different tasks and information being shared as appropriate. Importantly, we never directly observe the factors of variations but establish that access to multiple tasks is sufficient for identifiability under sufficiency and minimality assumptions. We validate our approach on six real world distribution shift benchmarks, and different data modalities (images, text), demonstrating how disentangled representations can be transferred to real settings."}, "cited_paper_content": {"title": "Distributionally Robust Neural Networks For Group Shifts: On The Importance Of Regularization For Worst-Case Generalization", "abstract": "Overparameterized neural networks can be highly accurate on average on an i.i.d. test set yet consistently fail on atypical groups of the data (e.g., by learning spurious correlations that hold on average but not in such groups). Distributionally robust optimization (DRO) allows us to learn models that instead minimize the worst-case training loss over a set of pre-defined groups. However, we find that naively applying group DRO to overparameterized neural networks fails: these models can perfectly fit the training data, and any model with vanishing average training loss also already has vanishing worst-case training loss. Instead, the poor worst-case performance arises from poor generalization on some groups. By coupling group DRO models with increased regularization---a stronger-than-typical L2 penalty or early stopping---we achieve substantially higher worst-group accuracies, with 10-40 percentage point improvements on a natural language inference task and two image tasks, while maintaining high average accuracies. Our results suggest that regularization is important for worst-group generalization in the overparameterized regime, even if it is not needed for average generalization. Finally, we introduce a stochastic optimization algorithm, with convergence guarantees, to efficiently train group DRO models."}, "keywords": ["worst accuracy"], "citation_intent": "result"}
{"citing_id": "2304.03159v1", "cited_id": "1910.07475", "section_title": "A. Multilingual Question Answering", "citation": "Reference #REFR input text and using extra layers to conduct classification has gradually become a popular approach to solve multilingual QA task.", "text_before_citation": ["QA is a hot topic in NLP area.", "There are massive datasets available, and these datasets pose different challenges to QA algorithms #OTHEREFR , #OTHEREFR , #OTHEREFR .", "One of these challenges is to find the answer span when the context and the problem are in different languages, and this task is called generalized crosslingual transfer(G-XLT).", "More formally, cross-lingual transfer problem requires a model to identify answer a x in context c x according to problem q x where x is the language used.", "Meanwhile, generalized cross-lingual transfer requires a model to find the answer span a z in context c z according to question q y where z and y are languages used."], "text_after_citation": ["Extensive experiments using state-of-the-art PLMs are reported in #OTHEREFR .", "Their results show that there are quite large room for improvement on both XLT and G-XLT tasks.", "One of the major challenges in G-XLT task is to enhance the connection between different languages.", "Reference #OTHEREFR uses hyperlink prediction task to firstly enhance the connection between different representations and then finetune the model on the target task data in English. Finally, they evaluate the performance on all languages.", "It is worth noting that this method creatively uses hyperlink as the bridge of different languages."], "citing_paper_content": {"title": "Bridging The Language Gap: Knowledge Injected Multilingual Question Answering", "abstract": "Question Answering (QA) is the task of automatically answering questions posed by humans in natural languages. There are different settings to answer a question, such as abstractive, extractive, boolean, and multiple-choice QA. As a popular topic in natural language processing tasks, extractive question answering task (extractive QA) has gained extensive attention in the past few years. With the continuous evolvement of the world, generalized cross-lingual transfer (G-XLT), where question and answer context are in different languages, poses some unique challenges over cross-lingual transfer (XLT), where question and answer context are in the same language. With the boost of corresponding development of related benchmarks, many works have been done to improve the performance of various language QA tasks. However, only a few works are dedicated to the G-XLT task. In this work, we propose a generalized crosslingual transfer framework to enhance the model's ability to understand different languages. Specifically, we first assemble triples from different languages to form multilingual knowledge. Since the lack of knowledge between different languages greatly limits models' reasoning ability, we further design a knowledge injection strategy via leveraging link prediction techniques to enrich the model storage of multilingual knowledge. In this way, we can profoundly exploit rich semantic knowledge. Experiment results on real-world datasets MLQA demonstrate that the proposed method can improve the performance by a large margin, outperforming the baseline method by 13.18%/12.00% F1/EM on average. Our code and models will be open-sourced."}, "cited_paper_content": {"title": "Mlqa: Evaluating Cross-Lingual Extractive Question Answering", "abstract": "Question answering (QA) models have shown rapid progress enabled by the availability of large, high-quality benchmark datasets. Such annotated datasets are difficult and costly to collect, and rarely exist in languages other than English, making training QA systems in other languages challenging. An alternative to building large monolingual training datasets is to develop cross-lingual systems which can transfer to a target language without requiring training data in that language. In order to develop such systems, it is crucial to invest in high quality multilingual evaluation benchmarks to measure progress. We present MLQA, a multi-way aligned extractive QA evaluation benchmark intended to spur research in this area. MLQA contains QA instances in 7 languages, namely English, Arabic, German, Spanish, Hindi, Vietnamese and Simplified Chinese. It consists of over 12K QA instances in English and 5K in each other language, with each QA instance being parallel between 4 languages on average. MLQA is built using a novel alignment context strategy on Wikipedia articles, and serves as a cross-lingual extension to existing extractive QA datasets. We evaluate current state-of-the-art cross-lingual representations on MLQA, and also provide machine-translation-based baselines. In all cases, transfer results are shown to be significantly behind training-language performance."}, "keywords": ["multilingual QA task"], "citation_intent": "background"}
{"citing_id": "2303.09627v1", "cited_id": "1706.03762", "section_title": "Ii. Low-Light Image Enhancement", "citation": "Most DL architectures are based on Convolutional Neural Networks (CNNs) and more recently, CNNs have been hybridized with transformer networks #REFR .", "text_before_citation": ["Another paradigm of LLIE makes use of Retinex theory, where the assumption is that a color image can be separated into reflectance and illumination #OTHEREFR .", "Non-learning Retinexbased methods such as Low-light Image Enhancement via Illumination Map Estimation (LIME) #OTHEREFR provide an effective image enhancement approach; however, post-processing denoising is typically still necessary using algorithms such as BM3D #OTHEREFR which often blurs high-frequency details.", "Alternative Retinex-based methods reformulate the traditional Retinex model to incorporate an added noise term in order to cater for noise #OTHEREFR .", "DL methods have recently achieved state-of-the-art LLIE performance.", "Some complexities of catering for a large variety of realistic low-light scenes are abstracted away by a datadriven approach."], "text_after_citation": ["DL methods either opt for incorporating denoising into a single model, or apply denoising as a post-processing step.", "S-LLNet #OTHEREFR makes use of a learned denoiser which operates sequentially after contrast enhancement.", "Retinex-net #OTHEREFR incorporates Retinex theory into a DL model with an optional illumination-varying BM3D denoiser used as post-processing.", "Raw sensor data are enhanced and denoised via CNN in #OTHEREFR .", "A variety of loss functions have also been proposed (in addition to the typical l 1 and l 2 losses) which further penalize networks based on color, smoothness, brightness and perceptual interpretation #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR - #OTHEREFR ."], "citing_paper_content": {"title": "Denoising Diffusion Post-Processing For Low-Light Image Enhancement", "abstract": "Low-light image enhancement (LLIE) techniques attempt to increase the visibility of images captured in lowlight scenarios. However, as a result of enhancement, a variety of image degradations such as noise and color bias are revealed. Furthermore, each particular LLIE approach may introduce a different form of flaw within its enhanced results. To combat these image degradations, post-processing denoisers have widely been used, which often yield oversmoothed results lacking detail. We propose using a diffusion model as a post-processing approach, and we introduce Low-light Post-processing Diffusion Model (LPDM) in order to model the conditional distribution between underexposed and normally-exposed images. We apply LPDM in a manner which avoids the computationally expensive generative reverse process of typical diffusion models, and post-process images in one pass through LPDM. Extensive experiments demonstrate that our approach outperforms competing post-processing denoisers by increasing the perceptual quality of enhanced lowlight images on a variety of challenging low-light datasets."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["Convolutional Neural Networks", "CNNs"], "citation_intent": "method"}
{"citing_id": "2303.11570v1", "cited_id": "1412.6806", "section_title": "Experimental Settings", "citation": "For CIFAR-10 dataset, we train the All-CNN model #REFR from scratch for 30 epochs using SGD with a fixed learning rate of 0.1, momentum of 0.9 and batch-size of 64.", "text_before_citation": ["Amnesiac Unlearning #OTHEREFR : amnesiac unlearning is a typical method of rapidly retrain.", "We need save the updates of parameters during several batches on the original training phase.", "Then, we can unlearn the forgetting data by subtracting the corresponding updates of parameters. Implementations.", "Our methods and other baselines are implemented in Python 3.8 and use the PyTorch library #OTHEREFR .", "All experiments are conducted on a workstation with one NVIDIA GeForce RTX 2070 GPU."], "text_after_citation": ["For Vggface2 dataset, we randomly select face images of 10 celebrities to conduct experiments.", "We obtain the original model by finetuning the pretrained ResNet50 model #OTHEREFR with the training images of the 10 celebrities.", "The original training parameters are similar to those on CIFAR-10.", "For the fine-tune process in Boundary Unlearning we use a learning rate of 10 \u22125 for 10 epochs.", "Note that we use the same forgetting class as the forgetting data for all unlearning methods and the rest of classes are the remaining data. Metrics."], "citing_paper_content": {"title": "Boundary Unlearning", "abstract": "The practical needs of the \"right to be forgotten\" and poisoned data removal call for efficient machine unlearning techniques, which enable machine learning models to unlearn, or to forget a fraction of training data and its lineage. Recent studies on machine unlearning for deep neural networks (DNNs) attempt to destroy the influence of the forgetting data by scrubbing the model parameters. However, it is prohibitively expensive due to the large dimension of the parameter space. In this paper, we refocus our attention from the parameter space to the decision space of the DNN model, and propose Boundary Unlearning, a rapid yet effective way to unlearn an entire class from a trained DNN model. The key idea is to shift the decision boundary of the original DNN model to imitate the decision behavior of the model retrained from scratch. We develop two novel boundary shift methods, namely Boundary Shrink and Boundary Expanding, both of which can rapidly achieve the utility and privacy guarantees. We extensively evaluate Boundary Unlearning on CIFAR-10 and Vggface2 datasets, and the results show that Boundary Unlearning can effectively forget the forgetting class on image classification and face recognition tasks, with an expected speed-up of 17\u00d7 and 19\u00d7, respectively, compared with retraining from the scratch."}, "cited_paper_content": {"title": "Striving For Simplicity: The All Convolutional Net", "abstract": "Most modern convolutional neural networks (CNNs) used for object recognition are built using the same principles: Alternating convolution and max-pooling layers followed by a small number of fully connected layers. We re-evaluate the state of the art for object recognition from small images with convolutional networks, questioning the necessity of different components in the pipeline. We find that max-pooling can simply be replaced by a convolutional layer with increased stride without loss in accuracy on several image recognition benchmarks. Following this finding -- and building on other recent work for finding simple network structures -- we propose a new architecture that consists solely of convolutional layers and yields competitive or state of the art performance on several object recognition datasets (CIFAR-10, CIFAR-100, ImageNet). To analyze the network we introduce a new variant of the \"deconvolution approach\" for visualizing features learned by CNNs, which can be applied to a broader range of network structures than existing approaches."}, "keywords": ["All-CNN model"], "citation_intent": "method"}
{"citing_id": "2304.12504v1", "cited_id": "0906.4725", "section_title": "The Controlled Qupit Gate", "citation": "Phase gadgets have been characterized thus far in two graphical calculi: in the qubit ZX-calculus #REFR in Ref.", "text_before_citation": ["( \u2212 \u2192 ) = \u2020 ( \u2212 \u2192 ) . (30) \u25a1", "Our approach to controlling the gate is to control each component and phase gate in Equation #OTHEREFR .", "We apply strategies from past work: generalizing controlled phase gates from qubits to qutrits #OTHEREFR and controlling the qutrit gate #OTHEREFR .", "In future work we will consider generalizing to higher dimensions, adding any number of controls to qutrit gates #OTHEREFR .", "The key structural insight leverages the generalization to arbitrary dimensions of phase gadgets, a symmetric multi-qudit interaction which applies a phase dependent on the sum of all computational basis states modulo ."], "text_after_citation": ["#OTHEREFR , and in the qutrit ZX-calculus #OTHEREFR in Ref. #OTHEREFR .", "We remark that phase gadgets are broadly applicable to all qudit dimensions as a graphical construct to relate diagonal gates in the circuit model and polynomial functions.", "Hence, this gives us an idea of what sort of circuit structure may be a good place to start.", "Keeping minimal on introducing formalism of the qudit -calculus #OTHEREFR , we introduce non-unitary circuit elements which respectively act as addition modulo and copying of basis states, labelled | \u27e9 and | \u27e9 below:", "EQUATION"], "citing_paper_content": {"title": "Scaling W State Circuits In The Qudit Clifford Hierarchy", "abstract": "We identify a novel qudit gate which we call the \u221a gate. This is an alternate generalization of the qutrit gate to any odd prime dimension , in the th level of the Clifford hierarchy. Using this gate which is efficiently realizable fault-tolerantly should a certain conjecture hold, we deterministically construct in the Clifford+ \u221a gate set,-qubit states in the qudit {|0\u27e9 , |1\u27e9} subspace. For qutrits, this gives deterministic and fault-tolerant constructions for the qubit state of sizes three with count 3, six, and powers of three. Furthermore, we adapt these constructions to recursively scale the state size to arbitrary size , in () gate count and (log) depth. This is moreover deterministic for any size qubit state, and for any prime-dimensional qudit state, size a power of. For these purposes, we devise constructions of the |0\u27e9-controlled Pauli gate and the controlled Hadamard gate in any prime qudit dimension. These decompositions, for which exact synthesis is unknown in Clifford+ for > 3, may be of independent interest. CCS CONCEPTS \u2022 Hardware \u2192 Quantum communication and cryptography; Quantum error correction and fault tolerance; \u2022 Computer systems organization \u2192 Quantum computing."}, "cited_paper_content": {"title": "Interacting Quantum Observables: Categorical Algebra And Diagrammatics", "abstract": "This paper has two tightly intertwined aims: (i) to introduce an intuitive and universal graphical calculus for multi-qubit systems, the ZX-calculus, which greatly simplifies derivations in the area of quantum computation and information. (ii) To axiomatize complementarity of quantum observables within a general framework for physical theories in terms of dagger symmetric monoidal categories. We also axiomatize phase shifts within this framework. Using the well-studied canonical correspondence between graphical calculi and dagger symmetric monoidal categories, our results provide a purely graphical formalisation of complementarity for quantum observables. Each individual observable, represented by a commutative special dagger Frobenius algebra, gives rise to an Abelian group of phase shifts, which we call the phase group. We also identify a strong form of complementarity, satisfied by the Z- and X-spin observables, which yields a scaled variant of a bialgebra."}, "keywords": ["qubit ZX"], "citation_intent": "method"}
{"citing_id": "2304.05823v1", "cited_id": "1706.09141", "section_title": "Ablation Results", "citation": "In causal discovery, larger graphs are typically more challenging to learn, as the number of possible graphs grows super exponentially with the number of nodes #REFR . The results are shown in Figure 6 .", "text_before_citation": ["Specifically, we addressed the following questions: 1) What is the impact of changing the number of replicates K n ? 2) What is the impact of changing the number of genes N ?", "To answer the first question, we varied the number of replicates per perturbation from the default setting of 500 to 100, 200, 300, and 400.", "The results, shown in Figure 5 , indicate that reducing the number of replicates from 500 to 300 only slightly decreases performance, while a more significant reduction is observed with only 100 replicates.", "These findings suggest that having more replicates generally improves performance, although the benefits are marginal beyond 300 replicates.", "To answer the second question, we varied the number of genes from 50 to 100."], "text_after_citation": ["We can see that DiscoGen's performance worsen only marginally as the number of genes increases.", "This gives a strong indication that DiscoGen is reasonably robust to different number of genes."], "citing_paper_content": {"title": "Discogen: Learning To Discover Gene Regulatory Networks", "abstract": "Accurately inferring Gene Regulatory Networks (GRNs) is a critical and challenging task in biology. GRNs model the activatory and inhibitory interactions between genes and are inherently causal in nature. To accurately identify GRNs, perturbational data is required. However, most GRN discovery methods only operate on observational data. Recent advances in neural network-based causal discovery methods have significantly improved causal discovery, including handling interventional data, improvements in performance and scalability. However, applying state-of-the-art (SOTA) causal discovery methods in biology poses challenges, such as noisy data and a large number of samples. Thus, adapting the causal discovery methods is necessary to handle these challenges. In this paper, we introduce DiscoGen, a neural network-based GRN discovery method that can denoise gene expression measurements and handle interventional data. We demonstrate that our model outperforms SOTA neural network-based causal discovery methods."}, "cited_paper_content": {"title": "Causal Structure Learning", "abstract": "Graphical models can represent a multivariate distribution in a convenient and accessible form as a graph. Causal models can be viewed as a special class of graphical models that represent not only the distribution of the observed system but also the distributions under external interventions. They hence enable predictions under hypothetical interventions, which is important for decision making. The challenging task of learning causal models from data always relies on some underlying assumptions. We discuss several recently proposed structure learning algorithms and their assumptions, and we compare their empirical performance under various scenarios."}, "keywords": ["causal discovery"], "citation_intent": "result"}
{"citing_id": "2304.14340v1", "cited_id": "1612.03144", "section_title": "Camera Detector", "citation": "B.3) go through a self-attention module, then deformable attention is conducted with the image features, where we aggregate multi-scale image features from FPN #REFR through deformable attention.", "text_before_citation": ["We extend Deformable-DETR #OTHEREFR to the 3D object detection task. The initial camera queries Q 0 C (Sec."], "text_after_citation": ["In deformable attention, each query only interacts with its corresponding single-view image features.", "The output queries are fed into a feed-forward network to get the perspective view camera candidates Q P C .", "As we do with the LiDAR detector, we add positional embeddings to all of the queries, keys, and values, which indicate their 2D locations on the image of the corresponding view. A perspective view prediction head (Sec.", "B.2) is attached to the perspective view camera candidates Q P C to get the object category as well as the 3D bounding box in camera coordinates.", "View Transformation Our view transformation module consists of two parts: feature projection and multi-view aggregation. The feature projection is already described in Eq."], "citing_paper_content": {"title": "Sparsefusion: Fusing Multi-Modal Sparse Representations For Multi-Sensor 3D Object Detection", "abstract": "By identifying four important components of existing LiDAR-camera 3D object detection methods (LiDAR and camera candidates, transformation, and fusion outputs), we observe that all existing methods either find dense candidates or yield dense representations of scenes. However, given that objects occupy only a small part of a scene, finding dense candidates and generating dense representations is noisy and inefficient. We propose SparseFusion, a novel multi-sensor 3D detection method that exclusively uses sparse candidates and sparse representations. Specifically, SparseFusion utilizes the outputs of parallel detectors in the LiDAR and camera modalities as sparse candidates for fusion. We transform the camera candidates into the Li-DAR coordinate space by disentangling the object representations. Then, we can fuse the multi-modality candidates in a unified 3D space by a lightweight self-attention module. To mitigate negative transfer between modalities, we propose novel semantic and geometric cross-modality transfer modules that are applied prior to the modality-specific detectors. SparseFusion achieves state-of-the-art performance on the nuScenes benchmark while also running at the fastest speed, even outperforming methods with stronger backbones. We perform extensive experiments to demonstrate the effectiveness and efficiency of our modules and overall method pipeline. Our code will be made publicly available at https://github.com/yichen928/SparseFusion."}, "cited_paper_content": {"title": "Feature Pyramid Networks For Object Detection", "abstract": "Feature pyramids are a basic component in recognition systems for detecting objects at different scales. But pyramid representations have been avoided in recent object detectors that are based on deep convolutional networks, partially because they are slow to compute and memory intensive. In this paper, we exploit the inherent multi-scale, pyramidal hierarchy of deep convolutional networks to construct feature pyramids with marginal extra cost. A top-down architecture with lateral connections is developed for building high-level semantic feature maps at all scales. This architecture, called a Feature Pyramid Network (FPN), shows significant improvement as a generic feature extractor in several applications. Using a basic Faster R-CNN system, our method achieves state-of-the-art single-model results on the COCO detection benchmark without bells and whistles, surpassing all existing single-model entries including those from the COCO 2016 challenge winners. In addition, our method can run at 5 FPS on a GPU and thus is a practical and accurate solution to multi-scale object detection. Code will be made publicly available."}, "keywords": ["multi-scale image features"], "citation_intent": "method"}
{"citing_id": "2305.00723v1", "cited_id": "1812.04426", "section_title": "Introduction", "citation": "Progress has been made in this direction in #REFR where a methodology for equating carefully structured convolutional layers to finite difference quotients is developed.", "text_before_citation": ["Indeed, this has led to the development of several new technologies, such as physics informed neural networks (PINNs) #OTHEREFR .", "Such models are powerful tools for approximating PDE dynamics as they typically build some level of PDE structure into the network, and may be exploited to both solve PDEs, and to identify an underlying PDE.", "Outside of PINNs, significant progress has also been made in solving the inverse problem #OTHEREFR .", "While PINNs have proven effective for solving given (families of) PDEs, they have not been shown to be efficient solution methods when using real data.", "An important step towards incorporating physical knowledge into more standard networks is to construct neural networks which capture the structure of families of PDEs but are trained on data which does not necessarily arise from a PDE."], "text_after_citation": ["Indeed, by ensuring the network contains a variety of finite difference quotients one may train a neural network which accurately simulates dynamics for remarkably large classes of PDEs based on simulated data, although the network itself is highly restricted.", "One crucial observation we make in this work (and, to a certain extent, in #OTHEREFR ) is that by imposing no restrictions on the convolutional layers one may build a convolutional neural network which contains within it a large class of finite difference discretisations of PDEs.", "Specifically, here we require particular choices of activation functions and design the network with sufficiently wide convolutional layers.", "We shall exploit this connection in our analysis in section 3.", "An additional benefit of working with unrestricted networks is they contain more than discrete solutions to PDEs, and as such are more amenable to real world datasets."], "citing_paper_content": {"title": "Predictions Based On Pixel Data: Insights From Pdes And Finite Differences", "abstract": "Neural networks are the state-of-the-art for many approximation tasks in high-dimensional spaces, as supported by an abundance of experimental evidence. However, we still need a solid theoretical understanding of what they can approximate and, more importantly, at what cost and accuracy. One network architecture of practical use, especially for approximation tasks involving images, is convolutional (residual) networks. However, due to the locality of the linear operators involved in these networks, their analysis is more complicated than for generic fully connected neural networks. This paper focuses on sequence approximation tasks, where a matrix or a higher-order tensor represents each observation. We show that when approximating sequences arising from space-time discretisations of PDEs we may use relatively small networks. We constructively derive these results by exploiting connections between discrete convolution and finite difference operators. Throughout, we design our network architecture to, while having guarantees, be similar to those typically adopted in practice for sequence approximation tasks. Our theoretical results are supported by numerical experiments which simulate linear advection, the heat equation, and the Fisher equation. The implementation used is available at the repository associated to the paper [30]."}, "cited_paper_content": {"title": "Pde-Net 2.0: Learning Pdes From Data With A Numeric-Symbolic Hybrid Deep Network", "abstract": "Partial differential equations (PDEs) are commonly derived based on empirical observations. However, recent advances of technology enable us to collect and store massive amount of data, which offers new opportunities for data-driven discovery of PDEs. In this paper, we propose a new deep neural network, called PDE-Net 2.0, to discover (time-dependent) PDEs from observed dynamic data with minor prior knowledge on the underlying mechanism that drives the dynamics. The design of PDE-Net 2.0 is based on our earlier work [1] where the original version of PDE-Net was proposed. PDE-Net 2.0 is a combination of numerical approximation of differential operators by convolutions and a symbolic multi-layer neural network for model recovery. Comparing with existing approaches, PDE-Net 2.0 has the most flexibility and expressive power by learning both differential operators and the nonlinear response function of the underlying PDE model. Numerical experiments show that the PDE-Net 2.0 has the potential to uncover the hidden PDE of the observed dynamics, and predict the dynamical behavior for a relatively long time, even in a noisy environment."}, "keywords": ["carefully structured convolutional"], "citation_intent": "method"}
{"citing_id": "2304.13871v1", "cited_id": "1602.02505", "section_title": "Sampling Atypical Solutions", "citation": "For gradinet based simulations we used the standard BinaryNet implementation (BNET) (see e.g. #REFR ).", "text_before_citation": ["We report here the details for the various algorithms used in text. Spherical Perceptron algorithms.", "For SGD we initialized the weights uniformly on the unit sphere and optimized the cross entropy loss using batchsize 200 and learning rate 1, with a maximum number of allowed epochs equal to 20000.", "In the fBP simulations (see #OTHEREFR for details) we used = 10 replicas and an increasing coupling constant with 15 exponentially spaced values in the range [0.1, 10].", "For each value of we initialized fBP with the messages obtained at the previous coupling value.", "We than made it run until the maximum absolute difference between message was less than 10 \u22125 or until it reached 500 messages updates. Binary Perceptron algorithms."], "text_after_citation": ["We then optimized the cross entropy loss using full batch gradient descent with learing rate 1 and a maximum number of 20000 epochs.", "For Stochastic Belief Propagation Inspired (SBPI) simulations (see #OTHEREFR ) we set the maximum number of allowed iterations to 1000 and used a threshold = 2 and a probability = 0.3 of updating the synapses with a stability 0 \u2264 \u0394 \u2264 .", "For fBP simulations we used the same parameters as in the continuous weights case."], "citing_paper_content": {"title": "Typical And Atypical Solutions In Non-Convex Neural Networks With Discrete And Continuous Weights", "abstract": "We study the binary and continuous negative-margin perceptrons as simple non-convex neural network models learning random rules and associations. We analyze the geometry of the landscape of solutions in both models and find important similarities and differences. Both models exhibit subdominant minimizers which are extremely flat and wide. These minimizers coexist with a background of dominant solutions which are composed by an exponential number of algorithmically inaccessible small clusters for the binary case (the frozen 1-RSB phase) or a hierarchical structure of clusters of different sizes for the spherical case (the full RSB phase). In both cases, when a certain threshold in constraint density is crossed, the local entropy of the wide flat minima becomes non-monotonic, indicating a break-up of the space of robust solutions into disconnected components. This has a strong impact on the behavior of algorithms in binary models, which cannot access the remaining isolated clusters. For the spherical case the behaviour is different, since even beyond the disappearance of the wide flat minima the remaining solutions are shown to always be surrounded by a large number of other solutions at any distance, up to capacity. Indeed, we exhibit numerical evidence that algorithms seem to find solutions up to the SAT/UNSAT transition, that we compute here using an 1RSB approximation. For both models, the generalization performance as a learning device is shown to be greatly improved by the existence of wide flat minimizers even when trained in the highly underconstrained regime of very negative margins."}, "cited_paper_content": {"title": "Binarized Neural Networks", "abstract": "We introduce a method to train Binarized Neural Networks (BNNs) - neural networks with binary weights and activations at run-time. At train-time the binary weights and activations are used for computing the parameter gradients. During the forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which is expected to substantially improve power-efficiency. To validate the effectiveness of BNNs, we conducted two sets of experiments on the Torch7 and Theano frameworks. On both, BNNs achieved nearly state-of-the-art results over the MNIST, CIFAR-10 and SVHN datasets. We also report our preliminary results on the challenging ImageNet dataset. Last but not least, we wrote a binary matrix multiplication GPU kernel with which it is possible to run our MNIST BNN 7 times faster than with an unoptimized GPU kernel, without suffering any loss in classification accuracy. The code for training and running our BNNs is available on-line."}, "keywords": ["gradinet based simulations", "standard BinaryNet implementation"], "citation_intent": "method"}
{"citing_id": "2303.13535v1", "cited_id": "1802.10591", "section_title": "Introduction", "citation": "Besides 2D images, prior work also applied NST in spherical images for 3D movies and VR #REFR .", "text_before_citation": ["Prior work explored how to reconstruct realistic scenes in VR to elicit users' different affective responses. For example, Felnhofer et al.", "#OTHEREFR created five VR parks with different environments, in which each scene elicited specific affective responses, e.g., joy, sadness, boredom, anger and anxiety.", "However, customizing all details in VR environments is challenging, especially for eliciting specific affective responses.", "Researchers have developed the neural style transfer (NST) to separate the content and style from a style image A and apply image A's style to image B's content to generate a new image with the same content as B but the same style as A #OTHEREFR .", "Here, the style image refers to an image with a specific style reference, such as some famous painters' line strokes from existing artworks."], "text_after_citation": ["Although prior work altered users' perceptions by manipulating visual elements in VR, their affective state manipulation is based on their subjective judgment.", "Therefore, our study aims to generate style-transferred 360\u00b0images that can affect users' affective states (valence and arousal levels) in VR.", "We created an affective VR generating system using the NST approach ( Fig.", "1) conducted a within-subject user study to evaluate the effectiveness of our system and approach."], "citing_paper_content": {"title": "The Exploration And Evaluation Of Generating Affective 360\u00b0Panoramic Vr Environments Through Neural Style Transfer", "abstract": "Affective virtual reality (VR) environments with varying visual style can impact users' valence and arousal responses. We applied Neural Style Transfer (NST) to generate 360\u00b0VR environments that elicited users' varied valence and arousal responses. From a user study with 30 participants, findings suggested that generative VR environments changed participants' arousal responses but not their valence levels. The generated visual features, e.g., textures and colors, also altered participants' affective perceptions. Our work contributes novel insights about how users respond to generative VR environments and provided a strategy for creating affective VR environments without altering content."}, "cited_paper_content": {"title": "Stereoscopic Neural Style Transfer", "abstract": "This paper presents the first attempt at stereoscopic neural style transfer, which responds to the emerging demand for 3D movies or AR/VR. We start with a careful examination of applying existing monocular style transfer methods to left and right views of stereoscopic images separately. This reveals that the original disparity consistency cannot be well preserved in the final stylization results, which causes 3D fatigue to the viewers. To address this issue, we incorporate a new disparity loss into the widely adopted style loss function by enforcing the bidirectional disparity constraint in non-occluded regions. For a practical real-time solution, we propose the first feed-forward network by jointly training a stylization sub-network and a disparity sub-network, and integrate them in a feature level middle domain. Our disparity sub-network is also the first end-to-end network for simultaneous bidirectional disparity and occlusion mask estimation. Finally, our network is effectively extended to stereoscopic videos, by considering both temporal coherence and disparity consistency. We will show that the proposed method clearly outperforms the baseline algorithms both quantitatively and qualitatively."}, "keywords": ["VR", "3D movies"], "citation_intent": "method"}
{"citing_id": "2303.14762v1", "cited_id": "1601.07996", "section_title": "Related Work", "citation": "Because the study #REFR shows no recommendations on how systematically identify the filter technique which fits best the concrete dataset, in the current study, we will empirically find an optimal for our dataset a filter method.", "text_before_citation": ["In the current work, based on the results of studies #OTHEREFR , to balance a dataset, we will apply the over-sampling method SMOTE.", "We also use a learning algorithm called Random Forest Classifier to build a model.", "The mentioned enhanced methods: SMOTEBoost and Weighted Random Forest, won't be used to avoid over complication of our model.", "We will use the ROC convex hull method to evaluate the built model's ability to recommend an elicitation technique.", "However, classification task metrics like precision and recall will be calculated to compare the effect of the proposed enhancements with the results achieved in the work #OTHEREFR ."], "text_after_citation": [], "citing_paper_content": {"title": "Approaches To Improving The Accuracy Of Machine Learning Models In Requirements Elicitation Techniques Selection", "abstract": "Selecting techniques is a crucial element of the business analysis approach planning in IT projects. Particular attention is paid to the choice of techniques for requirements elicitation. One of the promising methods for selecting techniques is using machine learning algorithms trained on the practitioners' experience considering different projects' contexts. The effectiveness of ML models is significantly affected by the balance of the training dataset, which is violated in the case of popular techniques. The paper aims to analyze the efficiency of the Synthetic Minority Over-sampling Technique usage in Machine Learning models for elicitation technique selection in case of the imbalanced training dataset and possible ways for positive feature importance selection. The computational experiment results confirmed the effectiveness of using the proposed approaches to improve the accuracy of machine learning models for selecting requirements elicitation techniques. Proposed approaches can be used to build Machine Learning models for business analysis activities planning in IT projects."}, "cited_paper_content": {"title": "Feature Selection: A Data Perspective", "abstract": "Feature selection, as a data preprocessing strategy, has been proven to be effective and efficient in preparing data (especially high-dimensional data) for various data-mining and machine-learning problems. The objectives of feature selection include building simpler and more comprehensible models, improving data-mining performance, and preparing clean, understandable data. The recent proliferation of big data has presented some substantial challenges and opportunities to feature selection. In this survey, we provide a comprehensive and structured overview of recent advances in feature selection research. Motivated by current challenges and opportunities in the era of big data, we revisit feature selection research from a data perspective and review representative feature selection algorithms for conventional data, structured data, heterogeneous data and streaming data. Methodologically, to emphasize the differences and similarities of most existing feature selection algorithms for conventional data, we categorize them into four main groups: similarity-based, information-theoretical-based, sparse-learning-based, and statistical-based methods. To facilitate and promote the research in this community, we also present an open source feature selection repository that consists of most of the popular feature selection algorithms (http://featureselection.asu.edu/). Also, we use it as an example to show how to evaluate feature selection algorithms. At the end of the survey, we present a discussion about some open problems and challenges that require more attention in future research."}, "keywords": ["dataset"], "citation_intent": "method"}
{"citing_id": "2303.03926v1", "cited_id": "1911.01635", "section_title": "Effect Of Language Id", "citation": "Voice Emotion Maintenance Generating the speech with a specific emotion is a difficult task for speech synthesis because conventional TTS methods require TTS data with emotion labels to train #REFR .", "text_before_citation": ["Besides, we conduct a subjection evaluation to label foreign accents from 1 to 5 on randomly selected 20 synthesized speech for both English and Chinese, where each sample is measured with a score from 1 to 5 denoting high-status foreign speakers, low-status foreign speakers, middle-status speakers, low-status native speakers, and high-status native speakers, respectively.", "As summarized in the last column of Table 6 , we observed that our VALL-E X can control the accent for the target speech by LID modules.", "For example, in English\u2192Chinese, VALL-E X Trans with right LID and without LID get the score of 4.03 and 2.35, respectively.", "This indicates that by using correct LID embedding, VALL-E X Trans is able to alleviate the foreign accent problem.", "Please also see the demo for audio examples of VALL-E X Trans with or without language ID."], "text_after_citation": ["Moreover, it is more tempting to reserve the source speaker's emotion in generated target speech for the S2ST task, which is not explored in previous S2ST work.", "In these experiments, we adopt the source prompts from the emotional voices dataset EmoV-DB #OTHEREFR as inputs of VALL-E X Trans to generate the translated target speech, whose samples are listed on our demo page.", "We found that the proposed VALL-E X can maintain emotional consistency to a certain extent between the source prompt and the synthesized speech.", "The underlying reasons are (1) our VALL-E X is trained with large-scale multi-lingual multi-speaker speech-transcription data, which contains various emotional speech records, and (2) the strong in-context learning ability of VALL-E X, like GPT-3 #OTHEREFR , promotes the generated speech to reserve the characteristic of the source prompt.", "Code-Switch Speech Synthesis It is a common phenomenon to use code-switch utterances in bilingual or multi-lingual communities #OTHEREFR ."], "citing_paper_content": {"title": "Speak Foreign Languages With Your Own Voice: Cross-Lingual Neural Codec Language Modeling", "abstract": "We propose a cross-lingual neural codec language model, VALL-E X, for crosslingual speech synthesis. Specifically, we extend VALL-E [Wang et al., 2023] and train a multilingual conditional codec language model to predict the acoustic token sequences of the target language speech by using both the source language speech and the target language text as prompts. VALL-E X inherits strong incontext learning capabilities and can be applied for zero-shot cross-lingual text-tospeech synthesis and zero-shot speech-to-speech translation tasks. Experimental results show that it can generate high-quality speech in the target language via just one speech utterance in the source language as a prompt while preserving the unseen speaker's voice, emotion, and acoustic environment. Moreover, VALL-E X effectively alleviates the foreign accent problems, which can be controlled by a language ID. Audio samples are available at https://aka.ms/vallex."}, "cited_paper_content": {"title": "Emotional Speech Synthesis With Rich And Granularized Control", "abstract": "This paper proposes an effective emotion control method for an end-to-end text-to-speech (TTS) system. To flexibly control the distinct characteristic of a target emotion category, it is essential to determine embedding vectors representing the TTS input. We introduce an inter-to-intra emotional distance ratio algorithm to the embedding vectors that can minimize the distance to the target emotion category while maximizing its distance to the other emotion categories. To further enhance the expressiveness of a target speech, we also introduce an effective interpolation technique that enables the intensity of a target emotion to be gradually changed to that of neutral speech. Subjective evaluation results in terms of emotional expressiveness and controllability show the superiority of the proposed algorithm to the conventional methods."}, "keywords": ["speech synthesis", "Voice Emotion Maintenance"], "citation_intent": "method"}
{"citing_id": "2303.13634v1", "cited_id": "1603.04467", "section_title": "Architecture", "citation": "The output of the previous action is a tensor of size B \u00d7 N \u00d7 nPDE, as shown in Fig. 2 . Using automatic differentiation #REFR , the governing PDEs (Eqs.", "text_before_citation": ["In practice, the input of PIPN is a three-dimensional tensor (i.e., numeric array) of size B \u00d7 N \u00d7 2.", "Afterward, there are two sequential shared MLPs with sizes of (ns \u00d764, ns \u00d764) and (ns \u00d764, ns \u00d7128, ns \u00d71024), as displayed in Fig. 2 .", "In the next step, the symmetric operator (s) forms the global feature with a size of ns \u00d7 1024. As exhibited in Fig.", "2 , the global feature is concatenated to the intermediate feature tensor of size B \u00d7 N \u00d7 (ns \u00d7 64), resulting in a tensor of size B \u00d7 N \u00d7 (ns \u00d7 1088).", "Next, two other sequential shared MLPs, respectively, with sizes of (ns \u00d7 512, ns \u00d7 256, ns \u00d7 128) and (ns \u00d7 128, nPDE) operate in the PIPN architecture."], "text_after_citation": ["1-2) are built and fed into the PIPN loss function.", "Note that reasonable choices for ns could be 0.125, 0.25, 0.5, 1.0, 2.0, and 4.0 such that a positive integer number represents the size of resulting shared MLPs."], "citing_paper_content": {"title": "Physics-Informed Pointnet: On How Many Irregular Geometries Can It Solve An Inverse Problem Simultaneously? Application To Linear Elasticity", "abstract": "Regular physics-informed neural networks (PINNs) predict the solution of partial differential equations using sparse labeled data but only over a single domain. On the other hand, fully supervised learning models are first trained usually over a few thousand domains with known solutions (i.e., labeled data) and then predict the solution over a few hundred unseen domains. Physics-informed PointNet (PIPN) is primarily designed to fill this gap between PINNs (as weakly supervised learning models) and fully supervised learning models. In this article, we demonstrate that PIPN predicts the solution of desired partial differential equations over a few hundred domains simultaneously, while it only uses sparse labeled data. This framework benefits fast geometric designs in the industry when only sparse labeled data are available. Particularly, we show that PIPN predicts the solution of a plane stress problem over more than 500 domains with different geometries, simultaneously. Moreover, we pioneer implementing the concept of remarkable batch size (i.e., the number of geometries fed into PIPN at each sub-epoch) into PIPN. Specifically, we try batch sizes of 7, 14, 19, 38, 76, and 133. Additionally, the effect of the PIPN size, symmetric function in the PIPN architecture, and static and dynamic weights for the component of the sparse labeled data in the loss function are investigated."}, "cited_paper_content": {"title": "Tensorflow: Large-Scale Machine Learning On Heterogeneous Distributed Systems", "abstract": "TensorFlow is an interface for expressing machine learning algorithms, and an implementation for executing such algorithms. A computation expressed using TensorFlow can be executed with little or no change on a wide variety of heterogeneous systems, ranging from mobile devices such as phones and tablets up to large-scale distributed systems of hundreds of machines and thousands of computational devices such as GPU cards. The system is flexible and can be used to express a wide variety of algorithms, including training and inference algorithms for deep neural network models, and it has been used for conducting research and for deploying machine learning systems into production across more than a dozen areas of computer science and other fields, including speech recognition, computer vision, robotics, information retrieval, natural language processing, geographic information extraction, and computational drug discovery. This paper describes the TensorFlow interface and an implementation of that interface that we have built at Google. The TensorFlow API and a reference implementation were released as an open-source package under the Apache 2.0 license in November, 2015 and are available at www.tensorflow.org."}, "keywords": ["N \u00d7 nPDE", "tensor"], "citation_intent": "method"}
{"citing_id": "2304.10126v1", "cited_id": "1706.02216", "section_title": "Introduction", "citation": "However, the aggregation-like operations defined in #REFR result in the dependency of each node on its neighbors and the amount of dependent nodes for one node increases exponentially with the growth of layers, which results in the unexpected increases of batch size.", "text_before_citation": ["In recent years, neural networks #OTHEREFR , due to the impressive performance, have been extended to graph data, known as graph neural networks (GNN) #OTHEREFR .", "As GNNs significantly improve the results of graph tasks, it has been extensively investigated from different aspects, such as graph convolution network (GCN) #OTHEREFR , graph attention networks (GAT) #OTHEREFR , spatial-temporal GNN (STGNN) #OTHEREFR , graph auto-encoder [9, #OTHEREFR , graph contrastive learning #OTHEREFR , etc.", "Except for the variants that originate from different perspectives, an important topic is motivated by the well-known inefficiency of GNN.", "In classical neural networks #OTHEREFR , the optimization is usually based on stochastic algorithms with limited batch #OTHEREFR since samples are independent of each other."], "text_after_citation": ["Some works are proposed based on neighbor sampling #OTHEREFR and graph approximation #OTHEREFR to limit the batch size, while some methods #OTHEREFR attempt to directly apply high-order graph operation and sacrifice the most non-linearity.", "The training stability is a problem for neighbor sampling methods #OTHEREFR though VRGCN #OTHEREFR has attempted to control the variance via improving sampling.", "Note that the required nodes may still grow (slowly) with the increase of depth.", "ClusterGCN #OTHEREFR finds an approximate Figure 1 : Illustration of a stacked graph neural network decoupled from an L-layer GNN.", "To train each module individually, some loss (e.g., semi-supervised loss, unsupervised loss, contrastive loss) is required and it is denoted by L graph with plenty of connected components so that the batch size is strictly upper-bounded."], "citing_paper_content": {"title": "Decouple Graph Neural Networks: Train Multiple Simple Gnns Simultaneously Instead Of One", "abstract": "Graph neural networks (GNN) suffer from severe inefficiency due to the exponential growth of node dependency with the increase of layers. It extremely limits the application of stochastic optimization algorithms so that the training of GNN is usually time-consuming. To address this problem, we propose to decouple a multi-layer GNN as multiple simple modules for more efficient training, which is comprised of classical forward training (FT) and designed backward training (BT). Under the proposed framework, each module can be trained efficiently in FT by stochastic algorithms without distortion of graph information owing to its simplicity. To avoid the only unidirectional information delivery of FT and sufficiently train shallow modules with the deeper ones, we develop a backward training mechanism that makes the former modules perceive the latter modules, inspired by the classical backward propagation algorithm. The backward training introduces the reversed information delivery into the decoupled modules as well as the forward information delivery. To investigate how the decoupling and greedy training affect the representational capacity, we theoretically prove that the error produced by linear modules will not accumulate on unsupervised tasks in most cases. The theoretical and experimental results show that the proposed framework is highly efficient with reasonable performance, which may deserve more investigation. Preprint. Under review."}, "cited_paper_content": {"title": "Inductive Representation Learning On Large Graphs", "abstract": "Low-dimensional embeddings of nodes in large graphs have proved extremely useful in a variety of prediction tasks, from content recommendation to identifying protein functions. However, most existing approaches require that all nodes in the graph are present during training of the embeddings; these previous approaches are inherently transductive and do not naturally generalize to unseen nodes. Here we present GraphSAGE, a general, inductive framework that leverages node feature information (e.g., text attributes) to efficiently generate node embeddings for previously unseen data. Instead of training individual embeddings for each node, we learn a function that generates embeddings by sampling and aggregating features from a node's local neighborhood. Our algorithm outperforms strong baselines on three inductive node-classification benchmarks: we classify the category of unseen nodes in evolving information graphs based on citation and Reddit post data, and we show that our algorithm generalizes to completely unseen graphs using a multi-graph dataset of protein-protein interactions."}, "keywords": ["dependent nodes"], "citation_intent": "background"}
{"citing_id": "2304.07470v1", "cited_id": "1910.13601", "section_title": "Problem Formulation", "citation": "We will now describe the enhancements to the three stages of the anomaly detection framework compared to #REFR .", "text_before_citation": ["The objective of the anomaly detection framework is to generate anomaly scores s i for classification of data sample x i given a threshold AD th such that:", "y i = \uf8f1 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f3 0, if s i < AD th 1, if s i \u2265 AD th", "where y i denotes the predicted label.", "y i = 0 indicates that the sample x i is a normal sample, whereas y i = 1 indicates that the sample x i is an anomaly.", "The details about the anomaly score s i are discussed in Section 3.4."], "text_after_citation": [], "citing_paper_content": {"title": "Few-Shot Weakly-Supervised Cybersecurity Anomaly Detection", "abstract": "With increased reliance on Internet based technologies, cyberattacks compromising users' sensitive data are becoming more prevalent. The scale and frequency of these attacks are escalating rapidly, affecting systems and devices connected to the Internet. The traditional defense mechanisms may not be sufficiently equipped to handle the complex and ever-changing new threats. The significant breakthroughs in the machine learning methods including deep learning, had attracted interests from the cybersecurity research community for further enhancements in the existing anomaly detection methods. Unfortunately, collecting labelled anomaly data for all new evolving and sophisticated attacks is not practical. Training and tuning the machine learning model for anomaly detection using only a handful of labelled data samples is a pragmatic approach. Therefore, few-shot weakly supervised anomaly detection is an encouraging research direction. In this paper, we propose an enhancement to an existing few-shot weakly-supervised deep learning anomaly detection framework. This framework incorporates data augmentation, representation learning and ordinal regression. We then evaluated and showed the performance of our implemented framework on three benchmark datasets: NSL-KDD, CIC-IDS2018, and TON IoT."}, "cited_paper_content": {"title": "Weakly-Supervised Deep Anomaly Detection With Pairwise Relation Learning.", "abstract": "This paper studies a rarely explored but critical anomaly detection problem: weakly-supervised anomaly detection with limited labeled anomalies and a large unlabeled data set. This problem is very important because it (i) enables anomaly-informed modeling which helps identify anomalies of interests and address the notorious high false positives in unsupervised anomaly detection, and (ii) eliminates the reliance on large-scale and complete labeled anomaly data in fully-supervised settings. However, the problem is especially challenging since we have only limited labeled data for a single class, and moreover, the seen anomalies often cannot cover all types of anomalies (i.e., unseen anomalies). We address this problem by formulating the problem as a pairwise relation learning task. Particularly, our approach defines a two-stream ordinal regression network to learn the relation of randomly selected instance pairs, i.e., whether the instance pair contains labeled anomalies or just unlabeled data instances. The resulting model leverages both the labeled and unlabeled data to effectively augment the data and learn generalized representations of both normality and abnormality. Extensive empirical results show that our approach (i) significantly outperforms state-of-the-art competing methods in detecting both seen and unseen anomalies and (ii) is substantially more data-efficient."}, "keywords": ["anomaly detection framework"], "citation_intent": "method"}
{"citing_id": "2303.05629v1", "cited_id": "1603.02754", "section_title": "4) Energy Estimation Models Training:", "citation": "We use a conventional machine-learning method, namely, Extreme Gradient Boost (XGBoost), as our baseline #REFR .", "text_before_citation": [], "text_after_citation": ["The XGBoost algorithm is a boosting-based method for integrated learning.", "In our XGBoost model, we select the optimal hyperparameters using the GridSearch method in scikit-learn #OTHEREFR .", "Additionally, we use neural networks as our second model.", "Since the data distribution is not particularly complex, we construct a simple multi-layer perception with three hidden layers as our neural network model.", "As we have two input features, i.e., the distance between charging coils and the charging duration, we set the number of neurons in the input layer to two."], "citing_paper_content": {"title": "Monitoring Efficiency Of Iot Wireless Charging", "abstract": "Crowdsourcing wireless energy is a novel and convenient solution to charge nearby IoT devices. Several applications have been proposed to enable peer-to-peer wireless energy charging. However, none of them considered the energy efficiency of the wireless transfer of energy. In this paper, we propose an energy estimation framework that predicts the actual received energy. Our framework uses two machine learning algorithms, namely XGBoost and Neural Network, to estimate the received energy. The result shows that the Neural Network model is better than XGBoost at predicting the received energy. We train and evaluate our models by collecting a real wireless energy dataset."}, "cited_paper_content": {"title": "Xgboost: A Scalable Tree Boosting System", "abstract": "Tree boosting is a highly effective and widely used machine learning method. In this paper, we describe a scalable end-to-end tree boosting system called XGBoost, which is used widely by data scientists to achieve state-of-the-art results on many machine learning challenges. We propose a novel sparsity-aware algorithm for sparse data and weighted quantile sketch for approximate tree learning. More importantly, we provide insights on cache access patterns, data compression and sharding to build a scalable tree boosting system. By combining these insights, XGBoost scales beyond billions of examples using far fewer resources than existing systems."}, "keywords": ["XGBoost"], "citation_intent": "method"}
{"citing_id": "2304.07908v1", "cited_id": "1801.07587", "section_title": "Vi. The Reliability Requirement Of Xr", "citation": "These connections could be established using the same carrier frequencies (intra-frequency MC) or different carrier frequencies (inter-frequency MC) #REFR .", "text_before_citation": ["In #OTHEREFR , this requirement is referred as uniform experience.", "However, maintaining uniform connection performance for all the users during their entire sessions is challenging, especially for high mobility use cases.", "For instance, varying connection bandwidth, addressed via dynamic rate adaption, leads to varying resolution #OTHEREFR , which, coupled with latency variation, results in fluctuating XR experience or even motion sickness.", "There are two approaches to achieving an ultra-reliable connection: simultaneous multi-connectivity (MC) #OTHEREFR and massive multiple input multiple output (MIMO) #OTHEREFR .", "MC's main objective is to improve the connection's data rates and reliability, i.e., achieve uniform experience by using a diversity of connections."], "text_after_citation": ["An example of the intra-frequency MC is single-frequency networks (SFNs), where several transmitters use the same frequency to send the same signal to the same user jointly.", "In inter-frequency MC, a transmitter (or several transmitters) uses multiple frequencies simultaneously to send the same signal.", "An example of interfrequency MC is when a transmitter is connected to multiple 5G/4G base stations and Wi-Fi access points simultaneously to increase the connection reliability.", "Although MC enhances reliability significantly, it comes with the major drawbacks of more resource consumption and additional delays #OTHEREFR .", "When a transmitter utilizes multiple resources, it will negatively impact the delay experienced by the remaining transmitters."], "citing_paper_content": {"title": "Traffic Characteristics Of Extended Reality", "abstract": "This tutorial paper analyzes the traffic characteristics of immersive experiences with extended reality (XR) technologies, including Augmented reality (AR), virtual reality (VR), and mixed reality (MR). The current trend in XR applications is to offload the computation and rendering to an external server and use wireless communications between the XR head-mounted display (HMD) and the access points. This paradigm becomes essential owing to (1) its high flexibility (in terms of user mobility) compared to remote rendering through a wired connection, and (2) the high computing power available on the server compared to local rendering (on HMD). The requirements to facilitate a pleasant XR experience are analyzed in three aspects: capacity (throughput), latency, and reliability. For capacity, two VR experiences are analyzed: a human eyelike experience and an experience with the Oculus Quest 2 HMD. For latency, the key components of the motion-to-photon (MTP) delay are discussed. For reliability, the maximum packet loss rate (or the minimum packet delivery rate) is studied for different XR scenarios. Specifically, the paper reviews optimization techniques that were proposed to reduce the latency, conserve the bandwidth, extend the scalability, and/or increase the reliability to satisfy the stringent requirements of the emerging XR applications."}, "cited_paper_content": {"title": "Towards Low-Latency And Ultra-Reliable Virtual Reality", "abstract": "Virtual Reality (VR) is expected to be one of the killer-applications in 5G networks. However, many technical bottlenecks and challenges need to be overcome to facilitate its wide adoption. In particular, VR requirements in terms of high-throughput, low-latency and reliable communication call for innovative solutions and fundamental research cutting across several disciplines. In view of this, this article discusses the challenges and enablers for ultra-reliable and low-latency VR. Furthermore, in an interactive VR gaming arcade case study, we show that a smart network design that leverages the use of mmWave communication, edge computing and proactive caching can achieve the future vision of VR over wireless."}, "keywords": ["connections"], "citation_intent": "background"}
{"citing_id": "2305.00135v1", "cited_id": "1607.01668", "section_title": "A. Proof Of Theorem 1", "citation": "Problem (6) can be solved efficiently by the alternating least squares matrix factorization procedure #REFR .", "text_before_citation": [], "text_after_citation": ["Since we deal with a third order tensor, this factorization has three iterative steps.", "Then, it fixes two factor matrices and minimizes the error with respect to the considered factor matrix as:", "EQUATION", "= arg min", "EQUATION"], "citing_paper_content": {"title": "Joint Sensing, Communication, And Ai: A Trifecta For Resilient Thz User Experiences", "abstract": "In this paper a novel joint sensing, communication, and artificial intelligence (AI) framework is proposed so as to optimize extended reality (XR) experiences over terahertz (THz) wireless systems. The proposed framework consists of three main components. First, a tensor decomposition framework is proposed to extract unique sensing parameters for XR users and their environment by exploiting then THz channel sparsity. Essentially, THz band's quasi-opticality is exploited and the sensing parameters are extracted from the uplink communication signal, thereby allowing for the use of the same waveform, spectrum, and hardware for both communication and sensing functionalities. Then, the Cramer-Rao lower bound is derived to assess the accuracy of the estimated sensing parameters. Second, a non-autoregressive multi-resolution generative artificial intelligence (AI) framework integrated with an adversarial transformer is proposed to predict missing and future sensing information. The proposed framework offers robust and comprehensive historical sensing information and anticipatory forecasts of future environmental changes, which are generalizable to fluctuations in both known and unforeseen user behaviors and environmental conditions. Third, a multi-agent deep recurrent hysteretic Q-neural network is developed to control the handover policy of reconfigurable intelligent surface (RIS) subarrays, leveraging the informative nature of sensing information to minimize handover cost, maximize the individual quality of personal experiences (QoPEs), and improve the robustness and resilience of THz links. Simulation results show a high generalizability of the proposed unsupervised generative AI framework to fluctuations in user behavior and velocity, leading to a 61 % improvement in instantaneous reliability compared to schemes with known channel state information."}, "cited_paper_content": {"title": "Tensor Decomposition For Signal Processing And Machine Learning", "abstract": "Tensors or multi-way arrays are functions of three or more indices $(i,j,k,\\cdots)$ -- similar to matrices (two-way arrays), which are functions of two indices $(r,c)$ for (row,column). Tensors have a rich history, stretching over almost a century, and touching upon numerous disciplines; but they have only recently become ubiquitous in signal and data analytics at the confluence of signal processing, statistics, data mining and machine learning. This overview article aims to provide a good starting point for researchers and practitioners interested in learning about and working with tensors. As such, it focuses on fundamentals and motivation (using various application examples), aiming to strike an appropriate balance of breadth and depth that will enable someone having taken first graduate courses in matrix algebra and probability to get started doing research and/or developing tensor algorithms and software. Some background in applied optimization is useful but not strictly required. The material covered includes tensor rank and rank decomposition; basic tensor factorization models and their relationships and properties (including fairly good coverage of identifiability); broad coverage of algorithms ranging from alternating optimization to stochastic gradient; statistical performance analysis; and applications ranging from source separation to collaborative filtering, mixture and topic modeling, classification, and multilinear subspace learning."}, "keywords": ["alternating least squares", "matrix factorization procedure"], "citation_intent": "method"}
{"citing_id": "2303.14467v1", "cited_id": "1809.04802", "section_title": "Uncertain Graphs", "citation": "Miyauchi and Takeda #REFR considered the uncertainty of edge weights rather than the existence of edges.", "text_before_citation": ["Zou #OTHEREFR studied the densest subgraph problem on uncertain graphs.", "An uncertain graph is a pair of G = (V, E) and p : E \u2192 [0, 1], where e \u2208 E is present with probability p(e) whereas e \u2208 E is absent with probability 1 \u2212 p(e) #OTHEREFR .", "In the problem, given an uncertain graph G = (V, E) with p, we seek S \u2286 V that maximizes the expected density.", "Zou #OTHEREFR showed that this problem can be reduced to DSP on edge-weighted graphs, and designed a polynomial-time exact algorithm based on the reduction."], "text_after_citation": ["To model that, they assumed that there is an edge-weight space I = \u00d7 e\u2208E [l e , r e ] \u2286 \u00d7 e\u2208E [0, \u221e) that contains the unknown true edge weight w.", "To evaluate the performance of S \u2286 V without any concrete edge weight, they employed a well-known measure in the field of robust optimization, called the robust ratio.", "In their scenario, the robust ratio of S \u2286 V under I is defined as the multiplicative gap between the density of S in terms of edge weight w and the density of S * w in terms of edge weight w under the worst-case edge weight w \u2208 I, where S * w is a densest subgraph on G with w .", "Intuitively, S \u2286 V with a large robust ratio has a density close to the optimal value even on G with the edge weight selected adversarially from I.", "Using the robust ratio, they introduced the robust densest subgraph problem: Given an undirected graph G = (V, E) and an edge-weight space I = \u00d7 e\u2208E [l e , r e ] \u2286 \u00d7 e\u2208E [0, \u221e), we are asked to find S \u2286 V that maximizes the robust ratio under I."], "citing_paper_content": {"title": "A Survey On The Densest Subgraph Problem And Its Variants", "abstract": "The Densest Subgraph Problem requires to find, in a given graph, a subset of vertices whose induced subgraph maximizes a measure of density. The problem has received a great deal of attention in the algorithmic literature over the last five decades, with many variants proposed and many applications built on top of this basic definition. Recent years have witnessed a revival of research interest on this problem with several interesting contributions, including some groundbreaking results, published in 2022 and 2023. This survey provides a deep overview of the fundamental results and an exhaustive coverage of the many variants proposed in the literature, with a special attention on the most recent results. The survey also presents a comprehensive overview of applications and discusses some interesting open problems for this evergreen research topic."}, "cited_paper_content": {"title": "Robust Densest Subgraph Discovery", "abstract": "Dense subgraph discovery is an important primitive in graph mining, which has a wide variety of applications in diverse domains. In the densest subgraph problem, given an undirected graph $G=(V,E)$ with an edge-weight vector $w=(w_e)_{e\\in E}$, we aim to find $S\\subseteq V$ that maximizes the density, i.e., $w(S)/|S|$, where $w(S)$ is the sum of the weights of the edges in the subgraph induced by $S$. Although the densest subgraph problem is one of the most well-studied optimization problems for dense subgraph discovery, there is an implicit strong assumption; it is assumed that the weights of all the edges are known exactly as input. In real-world applications, there are often cases where we have only uncertain information of the edge weights. In this study, we provide a framework for dense subgraph discovery under the uncertainty of edge weights. Specifically, we address such an uncertainty issue using the theory of robust optimization. First, we formulate our fundamental problem, the robust densest subgraph problem, and present a simple algorithm. We then formulate the robust densest subgraph problem with sampling oracle that models dense subgraph discovery using an edge-weight sampling oracle, and present an algorithm with a strong theoretical performance guarantee. Computational experiments using both synthetic graphs and popular real-world graphs demonstrate the effectiveness of our proposed algorithms."}, "keywords": ["uncertainty", "edge weights"], "citation_intent": "background"}
{"citing_id": "2304.12387v1", "cited_id": "1801.08914", "section_title": "2.6.", "citation": "This approximation of the second term on the right-hand side of (20) is significantly worse that the approximation of the transformed Schur complement S by #REFR , since it involves the product of three approximations rather than one.", "text_before_citation": ["Firstly, the offdiagonal blocks B \u03b1 and B T \u03b1 do not enjoy the same favorable sparsity as D and D T , and therefore are more expensive to compute.", "Secondly, and more importantly, the Schur complement of the original system takes the form #OTHEREFR S", "= W \u03b1 + B \u03b1 M \u22121 \u03b2 B T \u03b1 = W \u03b1 + W \u03b1 DM \u22121 \u03b2 D T W \u03b1 .", "In order to construct block preconditioners, approximations of the inverse of S are required.", "As in the construction of the approximate Schur complement S in (16), it is possible to replace W \u03b1 and M \u03b2 by their diagonals, to obtain a matrix S spectrally equivalent to S ."], "text_after_citation": ["A very rough estimate would suggest that, while still asymptotically optimal, the condition number of S \u22121 S could be very roughly bounded by the cube of the condition number of S \u22121 S.", "In Figure 3 , we display computed condition numbers of the transformed preconditioned Schur complement S \u22121 S and the untransformed version S \u22121 S for increasing polynomial degree.", "We consider both the unit cube [0, 1] d , and a skewed single-element mesh (the same mesh used in the results shown in Figure 2 ).", "Although both versions of the preconditioned Schur complement have O(1) condition number with respect to polynomial degree, the empirical results show that the transformed Schur complement results in significantly smaller condition number, in accordance with arguments described above.", "These results justify the use of the transformed system (11), despite the increased cost per iteration incurred by the inverse L 2 mass matrix."], "citing_paper_content": {"title": "Matrix-Free Gpu-Accelerated Saddle-Point Solvers For High-Order Problems In H(Div)", "abstract": "This work describes the development of matrix-free GPU-accelerated solvers for high-order finite element problems in H(div). The solvers are applicable to grad-div and Darcy problems in saddle-point formulation, and have applications in radiation diffusion and porous media flow problems, among others. Using the interpolation-histopolation basis (cf. SIAM J. Sci. Comput., 45 (2023), A675-A702), efficient matrix-free preconditioners can be constructed for the (1, 1)-block and Schur complement of the block system. With these approximations, block-preconditioned MINRES converges in a number of iterations that is independent of the mesh size and polynomial degree. The approximate Schur complement takes the form of an M-matrix graph Laplacian, and therefore can be well-preconditioned by highly scalable algebraic multigrid methods. High-performance GPU-accelerated algorithms for all components of the solution algorithm are developed, discussed, and benchmarked. Numerical results are presented on a number of challenging test cases, including the \"crooked pipe\" grad-div problem, the SPE10 reservoir modeling benchmark problem, and a nonlinear radiation diffusion test case."}, "cited_paper_content": {"title": "Algebraic Hybridization And Static Condensation With Application To Scalable H(Div) Preconditioning", "abstract": "We propose a unified algebraic approach for the practical application and preconditioning of static condensation and hybridization, two popular techniques in finite element discretizations. We demo..."}, "keywords": ["transformed Schur complement"], "citation_intent": "background"}
{"citing_id": "2304.01823v1", "cited_id": "1602.04505", "section_title": "Contracting A Single Crossedge.", "citation": "The proof that T \u2032 is a tangle of order 4 is identical to the finite case in #REFR , so we only need to prove that T \u2032 remains a well-quasi-ordered set under .", "text_before_citation": ["In the context of finite graphs, #OTHEREFR proves the following lemmas that extend directly to the locally finite case:", "Lemma 3.24 (Corollary 4.24 in #OTHEREFR ).", "The graph G \u2032 resulting from the contraction of s 1 s 2 is 3-connected.", "Lemma 3.25 (Lemmas 4.26 and 4.27 in #OTHEREFR ). T \u2032 is a region tangle of order 4.", "Proof."], "text_after_citation": ["Assume for the sake of contradiction that T \u2032 contains a infinite strictly decreasing sequence of separations (Y \u2032 n , S \u2032 n , Z \u2032 n ) n\u2208N .", "By Lemma 3.24, G \u2032 is 3-connected, so these separations are tight.", "By Lemma 3.1, there are finitely many integers n for which s \u2032 \u2208 S \u2032 n .", "Up to extracting an infinite subsequence, one can assume that for all n, either s \u2032 \u2208 Y \u2032 n or s \u2032 \u2208 Z \u2032 n .", "If there exists N such that s \u2032 \u2208 Y \u2032 N then for all n N we must have s \u2032 \u2208 Y n by definition of ."], "citing_paper_content": {"title": "The Structure Of Quasi-Transitive Graphs Avoiding A Minor With Applications To The Domino Problem", "abstract": "An infinite graph is quasi-transitive if its vertex set has finitely many orbits under the action of its automorphism group. In this paper we obtain a structure theorem for locally finite quasi-transitive graphs avoiding a minor, which is reminiscent of the Robertson-Seymour Graph Minor Structure Theorem. We prove that every locally finite quasi-transitive graph G avoiding a minor has a treedecomposition whose torsos are finite or planar; moreover the tree-decomposition is canonical, i.e. invariant under the action of the automorphism group of G. As applications of this result, we prove the following. \u2022 Every locally finite quasi-transitive graph attains its Hadwiger number, that is, if such a graph contains arbitrarily large clique minors, then it contains an infinite clique minor. This extends a result of Thomassen (1992) who proved it in the 4-connected case and suggested that this assumption could be omitted. In particular, this shows that a Cayley graph excludes a finite minor if and only if it avoids the countable clique as a minor. \u2022 Locally finite quasi-transitive graphs avoiding a minor are accessible (in the sense of Thomassen and Woess), which extends known results on planar graphs to any proper minor-closed family. \u2022 Minor-excluded finitely generated groups are accessible (in the group-theoretic sense) and finitely presented, which extends classical results on planar groups. \u2022 The domino problem is decidable in a minor-excluded finitely generated group if and only if the group is virtually free, which proves the minorexcluded case of a conjecture of Ballier and Stein (2018)."}, "cited_paper_content": {"title": "Quasi-4-Connected Components", "abstract": "We introduce a new decomposition of a graphs into quasi-4-connected components, where we call a graph quasi-4-connected if it is 3-connected and it only has separations of order 3 that remove a single vertex. Moreover, we give a cubic time algorithm computing the decomposition of a given graph. Our decomposition into quasi-4-connected components refines the well-known decompositions of graphs into biconnected and triconnected components. We relate our decomposition to Robertson and Seymour's theory of tangles by establishing a correspondence between the quasi-4-connected components of a graph and its tangles of order 4."}, "keywords": ["finite case", "tangle"], "citation_intent": "background"}
{"citing_id": "2304.13615v1", "cited_id": "1511.03339", "section_title": "Semantic Image Segmentation", "citation": "However, the naive pooling is independent of the image content, which can lead to suboptimal results. Therefore, Chen et al. #REFR and Yang et al.", "text_before_citation": ["Since the breakthrough of neural network in semantic segmentation #OTHEREFR , they were improved in various aspects such as integrating context #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR , attention mechanisms #OTHEREFR , #OTHEREFR , and Transformer backbones #OTHEREFR , #OTHEREFR .", "Several architectures #OTHEREFR , #OTHEREFR , #OTHEREFR utilize intermediate features with different scales, which are generated from a single scale input, to aggregate context information.", "Furthermore, multi-scale input inference, where predictions from scaled versions of an image are combined via average or max pooling, is often used to obtain better results #OTHEREFR , #OTHEREFR ."], "text_after_citation": ["#OTHEREFR segment multi-scale inputs and learn an attention-weighted sum of the predictions. Tao et al.", "#OTHEREFR further propose a hierarchical attention that is agnostic to the number of scales during inference.", "For image classification, CNNs were shown to be sensitive to distribution shifts such as image corruptions #OTHEREFR or domain shifts #OTHEREFR .", "Recent works #OTHEREFR , #OTHEREFR show that Transformers are more robust than CNNs with respect to these properties.", "While CNNs focus on textures #OTHEREFR , Transformers put more importance on the object shape #OTHEREFR , #OTHEREFR , which is more similar to human vision #OTHEREFR ."], "citing_paper_content": {"title": "Domain Adaptive And Generalizable Network Architectures And Training Strategies For Semantic Image Segmentation", "abstract": "Unsupervised domain adaptation (UDA) and domain generalization (DG) enable machine learning models trained on a source domain to perform well on unlabeled or even unseen target domains. As previous UDA&DG semantic segmentation methods are mostly based on outdated networks, we benchmark more recent architectures, reveal the potential of Transformers, and design the DAFormer network tailored for UDA&DG. It is enabled by three training strategies to avoid overfitting to the source domain: While (1) Rare Class Sampling mitigates the bias toward common source domain classes, (2) a Thing-Class ImageNet Feature Distance and (3) a learning rate warmup promote feature transfer from ImageNet pretraining. As UDA&DG are usually GPU memory intensive, most previous methods downscale or crop images. However, low-resolution predictions often fail to preserve fine details while models trained with cropped images fall short in capturing long-range, domain-robust context information. Therefore, we propose HRDA, a multi-resolution framework for UDA&DG, that combines the strengths of small high-resolution crops to preserve fine segmentation details and large low-resolution crops to capture long-range context dependencies with a learned scale attention. DAFormer and HRDA significantly improve the state-of-the-art UDA&DG by more than 10 mIoU on 5 different benchmarks. The implementation is available at https://github.com/lhoyer/HRDA."}, "cited_paper_content": {"title": "Attention To Scale: Scale-Aware Semantic Image Segmentation", "abstract": "Incorporating multi-scale features in fully convolutional neural networks (FCNs) has been a key element to achieving state-of-the-art performance on semantic image segmentation. One common way to extract multi-scale features is to feed multiple resized input images to a shared deep network and then merge the resulting features for pixelwise classification. In this work, we propose an attention mechanism that learns to softly weight the multi-scale features at each pixel location. We adapt a state-of-the-art semantic image segmentation model, which we jointly train with multi-scale input images and the attention model. The proposed attention model not only outperforms average- and max-pooling, but allows us to diagnostically visualize the importance of features at different positions and scales. Moreover, we show that adding extra supervision to the output at each scale is essential to achieving excellent performance when merging multi-scale features. We demonstrate the effectiveness of our model with extensive experiments on three challenging datasets, including PASCAL-Person-Part, PASCAL VOC 2012 and a subset of MS-COCO 2014."}, "keywords": ["naive pooling"], "citation_intent": "background"}
{"citing_id": "2304.01172v1", "cited_id": "1903.08527", "section_title": "Experiments", "citation": "The depth accuracy (Depth) is measured as the MSE between the rendered depth and the pseudo groundtruth depth obtained from a pre-trained face reconstruction model #REFR on the face mask area.", "text_before_citation": ["We also report the training time when utilizing 8 Tesla V100 GPUs.", "Our GMNR achieves consistent improvement in performance on all metrics and different resolutions, compared to the baseline.", "Furthermore, this improvement in performance over the baseline is achieved without any significant degradation in the training time and the inference speed of the model. #OTHEREFR .", "The Frechet Inception Distance (FID) #OTHEREFR and Kernel Inception Distance (KID) [4] are computed between 50K generated images rendered at different random poses and (a) 50K real images for FFHQ; (b) 5, 065 real images with flip augmentation for AFHQv2-Cats.", "The multi-view facial identity consistency (ID) is computed by first generating 1, 024 MPI representations and then employing the mean Arcface #OTHEREFR cosine similarity score between pairs of rendered views at random poses for the same face."], "text_after_citation": ["Similarly, the 3D pose accuracy (Pose) is computed by comparing the pose input used for rendering and the yaw, pitch and roll predicted by #OTHEREFR for the rendered image."], "citing_paper_content": {"title": "Generative Multiplane Neural Radiance For 3D-Aware Image Generation", "abstract": "We present a method to efficiently generate 3D-aware high-resolution images that are view-consistent across multiple target views. The proposed multiplane neural radiance model, named GMNR, consists of a novel \u03b1-guided view-dependent representation (\u03b1-VdR) module for learning view-dependent information. The \u03b1-VdR module, faciliated by an \u03b1-guided pixel sampling technique, computes the view-dependent representation efficiently by learning viewing direction and position coefficients. Moreover, we propose a view-consistency loss to enforce photometric similarity across multiple views. The GMNR model can generate 3D-aware high-resolution images that are viewconsistent across multiple camera poses, while maintaining the computational efficiency in terms of both training and inference time. Experiments on three datasets demonstrate the effectiveness of the proposed modules, leading to favorable results in terms of both generation quality and inference time, compared to existing approaches. Our GMNR model generates 3D-aware images of 1024 \u00d7 1024 pixels with 17.6 FPS on a single V100."}, "cited_paper_content": {"title": "Accurate 3D Face Reconstruction With Weakly-Supervised Learning: From Single Image To Image Set", "abstract": "Recently, deep learning based 3D face reconstruction methods have shown promising results in both quality and efficiency.However, training deep neural networks typically requires a large volume of data, whereas face images with ground-truth 3D face shapes are scarce. In this paper, we propose a novel deep 3D face reconstruction approach that 1) leverages a robust, hybrid loss function for weakly-supervised learning which takes into account both low-level and perception-level information for supervision, and 2) performs multi-image face reconstruction by exploiting complementary information from different images for shape aggregation. Our method is fast, accurate, and robust to occlusion and large pose. We provide comprehensive experiments on three datasets, systematically comparing our method with fifteen recent methods and demonstrating its state-of-the-art performance."}, "keywords": ["pre-trained face reconstruction"], "citation_intent": "method"}
{"citing_id": "2304.09427v1", "cited_id": "1808.10654", "section_title": "Related Work", "citation": "Solving multiple problems with a single model could create efficient systems and improve recognition for general AI, such as embodied agents #REFR , Narasimhan et al. [2020] .", "text_before_citation": ["[2019] introduced a dynamic fusion model with adaptive weights for better contextual modeling. DDS Liu et al.", "[2022a] proposed a deep supervision framework that supervises all side outputs and is currently the state-of-the-art method for SBD.", "Multi-Task Learning.", "In this paper, we specify multi-task learning (MTL) as an explicit joint modeling of two or more tasks like the method introduced in #OTHEREFR , #OTHEREFR , #OTHEREFR , #OTHEREFR .", "While most of the models in computer vision are task-specific, there is great interest in joint modeling."], "text_after_citation": ["In MTL, it is common to use a multi-head architecture with a shared backbone for memory efficiency.", "The backbone is aimed to learn a shared representation between the tasks, but often times this fails due to the backbone being designed for a single-task, leading to worse results #OTHEREFR , #OTHEREFR .", "The works of explores novel mechanism for obtaining features by adding task-specific attention modules in the backbone.", "Our work, however, explores the use of two-head architecture, where the auxiliary semantic boundary detection task is complementary to the main segmentation task.", "In semantic segmentation, edges and boundaries have been used as auxiliary tasks. #OTHEREFR"], "citing_paper_content": {"title": "Boosting Semantic Segmentation By Conditioning The Backbone With Semantic Boundaries", "abstract": "In this paper, we present the Semantic Boundary Conditioned Backbone (SBCB) framework, a simple yet effective training framework that is model-agnostic and boosts segmentation performance, especially around the boundaries. Motivated by the recent development in improving semantic segmentation by incorporating boundaries as auxiliary tasks, we propose a multi-task framework that uses semantic boundary detection (SBD) as an auxiliary task. The SBCB framework utilizes the nature of the SBD task, which is complementary to semantic segmentation, to improve the backbone of the segmentation head. We apply an SBD head that exploits the multi-scale features from the backbone, where the model learns low-level features in the earlier stages, and high-level semantic understanding in the later stages. This head perfectly complements the common semantic segmentation architectures where the features from the later stages are used for classification. We can improve semantic segmentation models without additional parameters during inference by only conditioning the backbone. Through extensive evaluations, we show the effectiveness of the SBCB framework by improving various popular segmentation heads and backbones by 0.5% \u223c 3.0% IoU on the Cityscapes dataset and gains 1.6% \u223c 4.1% in boundary Fscores. We also apply this framework on customized backbones and the emerging vision transformer models and show the effectiveness of the SBCB framework."}, "cited_paper_content": {"title": "Gibson Env: Real-World Perception For Embodied Agents", "abstract": "Developing visual perception models for active agents and sensorimotor control are cumbersome to be done in the physical world, as existing algorithms are too slow to efficiently learn in real-time and robots are fragile and costly. This has given rise to learning-in-simulation which consequently casts a question on whether the results transfer to real-world. In this paper, we are concerned with the problem of developing real-world perception for active agents, propose Gibson Virtual Environment for this purpose, and showcase sample perceptual tasks learned therein. Gibson is based on virtualizing real spaces, rather than using artificially designed ones, and currently includes over 1400 floor spaces from 572 full buildings. The main characteristics of Gibson are: I. being from the real-world and reflecting its semantic complexity, II. having an internal synthesis mechanism,\"Goggles\", enabling deploying the trained models in real-world without needing further domain adaptation, III. embodiment of agents and making them subject to constraints of physics and space."}, "keywords": ["single model", "embodied agents"], "citation_intent": "background"}
{"citing_id": "2305.01400v1", "cited_id": "1810.06544", "section_title": "Related Work & Limitations", "citation": "Imitative Models #REFR learns a density model over expert data and then uses the density model as a reward in a goal conditioned open-loop planner; they demonstrate policies robust to noise in an autonomous driving simulator environment.", "text_before_citation": ["For instance, the L2 reward would be exactly 0 on the dataset of demonstrations.", "Model-based Imitation Learning (MBIL): Using world models in Imitation Learning has been explored in other contexts #OTHEREFR .", "IMPLANT #OTHEREFR uses GAIL to learn a policy and value function using online data, which then get used in a planner policy.", "While IMPLANT shows that their planner is robust in noisy deployed environments for simple locomotion environments (Hopper, HalfCheetah, and Walker2d) our method differs mainly as we use BC instead of GAIL.", "This difference allows POIR to be deployed offline which is not the case of IMPLANT."], "text_after_citation": ["A follow-up, Robust Imitative Planning (RIP) #OTHEREFR , uses an ensemble of networks to minimize uncertainty of trajectories in the planner, and then adapts the policy online by querying an expert policy #OTHEREFR .", "A similar method, RMBIL #OTHEREFR also shows improvements over BC when deployed in environments that mismatch the expert distribution.", "RMBIL is a purely offline method, that uses a neural ODE controller with a CVAE to encode states.", "In contrast to Imitative Models, RIP, RMBIL, and other similar works #OTHEREFR , our method uses a simpler policy: model-predictive control #OTHEREFR with an open-loop trajectory planning algorithm #OTHEREFR using BC priors; notably we show that POIR can run in both the offline and online settings without querying an expert.", "Generalization in IL/RL: Generalization in IL and RL seeks to improve performance on unseen interactions during deployment #OTHEREFR ."], "citing_paper_content": {"title": "Get Back Here: Robust Imitation By Return-To-Distribution Planning", "abstract": "We consider the Imitation Learning (IL) setup where expert data are not collected on the actual deployment environment but on a different version. To address the resulting distribution shift, we combine behavior cloning (BC) with a planner that is tasked to bring the agent back to states visited by the expert whenever the agent deviates from the demonstration distribution. The resulting algorithm, POIR, can be trained offline, and leverages online interactions to efficiently fine-tune its planner to improve performance over time. We test POIR on a variety of human-generated manipulation demonstrations in a realistic robotic manipulation simulator and show robustness of the learned policy to different initial state distributions and noisy dynamics."}, "cited_paper_content": {"title": "Deep Imitative Models For Flexible Inference, Planning, And Control", "abstract": "Imitation Learning (IL) is an appealing approach to learn desirable autonomous behavior. However, directing IL to achieve arbitrary goals is difficult. In contrast, planning-based algorithms use dynamics models and reward functions to achieve goals. Yet, reward functions that evoke desirable behavior are often difficult to specify. In this paper, we propose \"Imitative Models\" to combine the benefits of IL and goal-directed planning. Imitative Models are probabilistic predictive models of desirable behavior able to plan interpretable expert-like trajectories to achieve specified goals. We derive families of flexible goal objectives, including constrained goal regions, unconstrained goal sets, and energy-based goals. We show that our method can use these objectives to successfully direct behavior. Our method substantially outperforms six IL approaches and a planning-based approach in a dynamic simulated autonomous driving task, and is efficiently learned from expert demonstrations without online data collection. We also show our approach is robust to poorly-specified goals, such as goals on the wrong side of the road."}, "keywords": ["autonomous driving simulator"], "citation_intent": "method"}
{"citing_id": "2304.14630v1", "cited_id": "1907.09091", "section_title": "Introduction", "citation": "However, these methods are often limited by a predefined set of recognized entities and specific data types, such as proportion-related statistics #REFR .", "text_before_citation": ["Nonetheless, these methods share some common issues.", "First, re-arXiv:2304.14630v1 [cs.AI] 28 Apr 2023 trieving an appropriate visual design that accurately aligns with the data requires a large-scale and high-quality image corpus, which is challenging to construct.", "Studies in this direction often create new datasets, resulting in various datasets (e.g., #OTHEREFR ) that do not meet each other's requirements.", "Second, the retrieval process requires certain guidance, which can be difficult for novice users to manipulate.", "With the increasing popularity of language models, numerous works (e.g., #OTHEREFR ) have explored the automatic generation of charts from text."], "text_after_citation": ["Additionally, editing the image to bind it with data is an error-prone task due to the trade-off between fitting the chart into the image and adapting the chart to the image.", "To alleviate these limitations, we propose ChartSpark, a novel pictorial visualization authoring tool that embed semantic context into charts based on text-to-image generative model, while preserving important visual attributes such as trend.", "We firstly extract the target visualized object through text analyzer.", "The generation process will then be guided by the object and its corresponding description provided by the user.", "As an initial step towards utilizing generation techniques for creating visualization charts, we conducted a preliminary study that examines the different display formats of pictorial visualization."], "citing_paper_content": {"title": "Let The Chart Spark: Embedding Semantic Context Into Chart With Text-To-Image Generative Model", "abstract": "Pictorial visualization seamlessly integrates data and semantic context into visual representation, conveying complex information in a manner that is both engaging and informative. Extensive studies have been devoted to developing authoring tools to simplify the creation of pictorial visualizations. However, mainstream works mostly follow a retrieving-and-editing pipeline that heavily relies on retrieved visual elements from a dedicated corpus, which often compromise the data integrity. Text-guided generation methods are emerging, but may have limited applicability due to its predefined recognized entities. In this work, we propose ChartSpark, a novel system that embeds semantic context into chart based on text-to-image generative model. ChartSpark generates pictorial visualizations conditioned on both semantic context conveyed in textual inputs and data information embedded in plain charts. The method is generic for both foreground and background pictorial generation, satisfying the design practices identified from an empirical research into existing pictorial visualizations. We further develop an interactive visual interface that integrates a text analyzer, editing module, and evaluation module to enable users to generate, modify, and assess pictorial visualizations. We experimentally demonstrate the usability of our tool, and conclude with a discussion of the potential of using text-to-image generative model combined with interactive interface for visualization design."}, "cited_paper_content": {"title": "Text-To-Viz: Automatic Generation Of Infographics From Proportion-Related Natural Language Statements", "abstract": "Combining data content with visual embellishments, infographics can effectively deliver messages in an engaging and memorable manner. Various authoring tools have been proposed to facilitate the creation of infographics. However, creating a professional infographic with these authoring tools is still not an easy task, requiring much time and design expertise. Therefore, these tools are generally not attractive to casual users, who are either unwilling to take time to learn the tools or lacking in proper design expertise to create a professional infographic. In this paper, we explore an alternative approach: to automatically generate infographics from natural language statements. We first conducted a preliminary study to explore the design space of infographics. Based on the preliminary study, we built a proof-of-concept system that automatically converts statements about simple proportion-related statistics to a set of infographics with pre-designed styles. Finally, we demonstrated the usability and usefulness of the system through sample results, exhibits, and expert reviews."}, "keywords": ["recognized entities", "proportion-related statistics"], "citation_intent": "method"}
{"citing_id": "2304.03119v1", "cited_id": "1411.7766", "section_title": "Generative Model Adaptation", "citation": "Real images are from CelebA-HQ dataset #REFR and translated into two styles of images, \"Wall painting\" and \"Tolkien elf\". The yellow boxes show the key observation areas.", "text_before_citation": ["From the quality of the gener-ated images, the results of NADA have more incorrect features and noise, such as green mussy noise on hairs (Tolkien elf), ruined noses (Werewolf) and unshaped necks (Pointillism), while the results of IPL are more clear and correct.", "From the mode collapse perspective of the generated images, NADA is prone to collapse to some similar facial features for different images, such as depressed emotions (Pixar character), folded ears (Cartoon) and blue noses (Cu- Figure 5 . Real-world image translation comparison results.", "Baselines are NADA #OTHEREFR , Diff-CLIP #OTHEREFR and Diff-CLIP+ (an improved version of Diff-CLIP).", "Recon1, Recon2 and Recon3 refer to inversion results via Restyle #OTHEREFR , DDIM and diffusion autoencoders, respectively.", "GAN-IPL and Diff-IPL denote integrating IPL with NADA and Diff-CLIP+, respectively."], "text_after_citation": ["bism), while IPL presents consistently higher diversity and solve the mode collapse issue well.", "Our advantages mainly come from the fact that the latent mapper preserves sufficient image-specific and target-domain friendly features from the source-domain images.", "The produced prompt vectors provide more precise and diversified adaptation directions for the target-domain generator adaptation.", "Quantitative comparison.", "To quantify the performance improvement of IPL compared to NADA #OTHEREFR , IS, SCS, ID and SIFID are evaluated. As reported in Tab."], "citing_paper_content": {"title": "Zero-Shot Generative Model Adaptation Via Image-Specific Prompt Learning", "abstract": "Figure 1. The mode collapse issue. For NADA [21] and our method, the same generator pre-trained on the source domain of \"Photo\" is adapted to the unseen target domains of \"Disney\", \"Anime painting\", \"Wall painting\" and \"Ukiyo-e\" only with the domain labels. The images above the dotted line are some examples from the internet. The generated images of NADA exhibit some similar unseen patterns (yellow box areas) which are undesired in terms of quality and diversity. This issue is largely addressed by our method."}, "cited_paper_content": {"title": "Deep Learning Face Attributes In The Wild", "abstract": "Predicting face attributes in the wild is challenging due to complex face variations. We propose a novel deep learning framework for attribute prediction in the wild. It cascades two CNNs, LNet and ANet, which are fine-tuned jointly with attribute tags, but pre-trained differently. LNet is pre-trained by massive general object categories for face localization, while ANet is pre-trained by massive face identities for attribute prediction. This framework not only outperforms the state-of-the-art with a large margin, but also reveals valuable facts on learning face representation. (1) It shows how the performances of face localization (LNet) and attribute prediction (ANet) can be improved by different pre-training strategies. (2) It reveals that although the filters of LNet are fine-tuned only with image-level attribute tags, their response maps over entire images have strong indication of face locations. This fact enables training LNet for face localization with only image-level annotations, but without face bounding boxes or landmarks, which are required by all attribute recognition works. (3) It also demonstrates that the high-level hidden neurons of ANet automatically discover semantic concepts after pre-training with massive face identities, and such concepts are significantly enriched after fine-tuning with attribute tags. Each attribute can be well explained with a sparse linear combination of these concepts."}, "keywords": ["Real images"], "citation_intent": "method"}
{"citing_id": "2304.11870v1", "cited_id": "1611.06224", "section_title": "Data Imputation (Di)", "citation": "Meanwhile, Human-in-the-loop (HITL) methods such as #REFR also attempt to utilize constraints through human interactions.", "text_before_citation": ["First, many constraint-based ML (CML) models are mostly designed for specific models (e.g., #OTHEREFR ).", "They have shown high accuracy in the corresponding models, but it is difficult to use diverse expert inputs and they may not be robust to instances that are not known from the training data.", "Expectation regularization (ER) #OTHEREFR takes the regularization function as expert input and could be applied to most supervised models that use gradient descent updates.", "Similarly, weak supervision (WS) models like #OTHEREFR take supervision functions, which explain a law in physics and show performance gains in closeness.", "In contrast, the ER and WS models are not able to leverage how reliable the predictions of new inputs are for the learned model."], "text_after_citation": ["Nevertheless, they primarily focus on collecting high-quality training data, reusing models, and managing the existing models, which still much depend on the prediction of the learned ML models.", "For Out-of-distribution (OOD) models like DRC #OTHEREFR , they take the simple statistical base distribution as expert input and measure how training data is different from the new instances.", "However, there are no incorporating methods to correct the predictions of the existing model.", "Transfer Learning (TL) methods are also limited in modifying the predictions for exploiting the expert input.", "Many Data Imputation (DI) techniques using MICE (e.g., #OTHEREFR ) can estimate missing judgment variables from the known data, but they have not been applied to correct the predictions of the current models."], "citing_paper_content": {"title": "Incorporating Experts' Judgment Into Machine Learning Models", "abstract": "Machine learning (ML) models have been quite successful in predicting outcomes in many applications. However, in some cases, domain experts might have a judgment about the expected outcome that might conflict with the prediction of ML models. One main reason for this is that the training data might not be totally representative of the population. In this paper, we present a novel framework that aims at leveraging experts' judgment to mitigate the conflict. The underlying idea behind our framework is that we first determine, using a generative adversarial network, the degree of representation of an unlabeled data point in the training data. Then, based on such degree, we correct the machine learning model's prediction by incorporating the experts' judgment into it, where the higher that aforementioned degree of representation, the less the weight we put on the expert intuition that we add to our corrected output, and vice-versa. We perform multiple numerical experiments on synthetic data as well as two real-world case studies (one from the IT services industry and the other from the financial industry). All results show the effectiveness of our framework; it yields much higher closeness to the experts' judgment with minimal sacrifice in the prediction accuracy, when compared to multiple baseline methods. We also develop a new evaluation metric that combines prediction accuracy with the closeness to experts' judgment. Our framework yields statistically significant results when evaluated on that metric."}, "cited_paper_content": {"title": "Towards Unified Data And Lifecycle Management For Deep Learning", "abstract": "Deep learning has improved state-of-the-art results in many important fields, and has been the subject of much research in recent years, leading to the development of several systems for facilitating deep learning. Current systems, however, mainly focus on model building and training phases, while the issues of data management, model sharing, and lifecycle management are largely ignored. Deep learning modeling lifecycle generates a rich set of data artifacts, e.g., learned parameters and training logs, and it comprises of several frequently conducted tasks, e.g., to understand the model behaviors and to try out new models. Dealing with such artifacts and tasks is cumbersome and largely left to the users. This paper describes our vision and implementation of a data and lifecycle management system for deep learning. First, we generalize model exploration and model enumeration queries from commonly conducted tasks by deep learning modelers, and propose a high-level domain specific language (DSL), inspired by SQL, to raise the abstraction level and thereby accelerate the modeling process. To manage the variety of data artifacts, especially the large amount of checkpointed float parameters, we design a novel model versioning system (dlv), and a read-optimized parameter archival storage system (PAS) that minimizes storage footprint and accelerates query workloads with minimal loss of accuracy. PAS archives versioned models using deltas in a multi-resolution fashion by separately storing the less significant bits, and features a novel progressive query (inference) evaluation algorithm. Third, we develop e cient algorithms for archiving versioned models using deltas under co-retrieval constraints. We conduct extensive experiments over several real datasets from computer vision domain to show the e ciency of the proposed techniques."}, "keywords": ["constraints"], "citation_intent": "method"}
{"citing_id": "2303.13679v1", "cited_id": "1801.05507", "section_title": "Ii. Background And Motivation", "citation": "Instead, we follow the interactive and hybrid cryptographic protocol #REFR and construct our Primerbase by using GC for non-polynomial operations, FHE for polynomial operations, and SS for secure communication between multiple parties.", "text_before_citation": ["In this work, we use this interactive hybrid method to construct Primer-base, which is a starting point for our optimization techniques. Motivation.", "As Figure 2 shows, prior works like THE-X #OTHEREFR using only FHE for private inference suffer from low accuracy and enormous online latency due to polynomial approximation and expensive FHE operations.", "We use prior GC-based work #OTHEREFR to implement a GCFormer (we convert the Transformer model into a circuit based on binary gates so that GC #OTHEREFR can implement it).", "GCFormer achieves an accurate performance, i.e., 85.1% accuracy, but it takes a larger latency than THE-X.", "Thus, the FHE-based method or GCbased method cannot achieve a low-latency and accurate private Transformer inference."], "text_after_citation": ["Primer-base significantly improves the accuracy of THE-X, e.g., 7.3% accuracy increase, and reduces the latency of GCFormer. However, Primer-base still suffers from enormous online latency.", "This motivates us to propose techniques like the FHGS protocol, denoted by Primer-F, to offload the online computation to the offline phase where computations can be computed before inference.", "Considering Primer-F still has a large total latency, we have motivations to propose techniques including computation merge, i.e., combined FHGS, and tokens-first ciphertext packing techniques.", "More details about Primer and related techniques are introduced in the following section III."], "citing_paper_content": {"title": "Primer: Fast Private Transformer Inference On Encrypted Data", "abstract": "It is increasingly important to enable privacypreserving inference for cloud services based on Transformers. Post-quantum cryptographic techniques, e.g., fully homomorphic encryption (FHE), and multi-party computation (MPC), are popular methods to support private Transformer inference. However, existing works still suffer from prohibitively computational and communicational overhead. In this work, we present, Primer, to enable a fast and accurate Transformer over encrypted data for natural language processing tasks. In particular, Primer is constructed by a hybrid cryptographic protocol optimized for attention-based Transformer models, as well as techniques including computation merge and tokens-first ciphertext packing. Comprehensive experiments on encrypted language modeling show that Primer achieves state-of-the-art accuracy and reduces the inference latency by 90.6% \u223c 97.5% over previous methods."}, "cited_paper_content": {"title": "Gazelle: A Low Latency Framework For Secure Neural Network Inference", "abstract": "The growing popularity of cloud-based machine learning raises a natural question about the privacy guarantees that can be provided in such a setting. Our work tackles this problem in the context where a client wishes to classify private images using a convolutional neural network (CNN) trained by a server. Our goal is to build efficient protocols whereby the client can acquire the classification result without revealing their input to the server, while guaranteeing the privacy of the server's neural network. To this end, we design Gazelle, a scalable and low-latency system for secure neural network inference, using an intricate combination of homomorphic encryption and traditional two-party computation techniques (such as garbled circuits). Gazelle makes three contributions. First, we design the Gazelle homomorphic encryption library which provides fast algorithms for basic homomorphic operations such as SIMD (single instruction multiple data) addition, SIMD multiplication and ciphertext permutation. Second, we implement the Gazelle homomorphic linear algebra kernels which map neural network layers to optimized homomorphic matrix-vector multiplication and convolution routines. Third, we design optimized encryption switching protocols which seamlessly convert between homomorphic and garbled circuit encodings to enable implementation of complete neural network inference. We evaluate our protocols on benchmark neural networks trained on the MNIST and CIFAR-10 datasets and show that Gazelle outperforms the best existing systems such as MiniONN (ACM CCS 2017) by 20 times and Chameleon (Crypto Eprint 2017/1164) by 30 times in online runtime. Similarly when compared with fully homomorphic approaches like CryptoNets (ICML 2016) we demonstrate three orders of magnitude faster online run-time."}, "keywords": ["secure communication", "hybrid cryptographic protocol"], "citation_intent": "method"}
{"citing_id": "2304.11330v1", "cited_id": "1406.5670", "section_title": "Our Approach", "citation": "First, if the number of views is fixed as n and the view sampling strategy is constant for all objects in a dataset, we initialize a n dimension learnable parameter. n = 12 views for ModelNet40 #REFR .", "text_before_citation": ["Then we extract the feature of the source view as value and synthesize the target view.", "For the self-attention blocks following the cross-attention blocks, the value, key and query are the same (three black arrows shown in Fig. 4 upper part).", "The numbers of cross-and self-attention blocks are the same by default. Fig.", "4 (upper part) simplifies them as one cross-attention block and one self-attention block. Pose Embedding.", "As described above, both key and query input of cross-attention blocks in VSA decoder are pose embeddings. Here we provide two methods to encode poses."], "text_after_citation": ["The shape of this learnable parameter is (n, patches, dim), where patches is the number of patches for encoder or decoder input, and dim is the dimension of extracted patch features. As shown in Fig.", "4 lower part, the gray grid is the learnable parameter which has n lines (from left to right, n = 12 in this example). On the left side of Fig.", "4 lower part, when the input is a single source view, the orange vector and blue vector are corresponding to the source view and target poses. On the right side of Fig.", "4 lower part, when the inputs are two source views, the yellow and orange vectors are corresponding to two source poses.", "The key in cross-attention blocks becomes the concatenation of two source poses embeddings (red). The blue vector is corresponding to the target pose."], "citing_paper_content": {"title": "Self-Supervised Learning By View Synthesis", "abstract": "We present view-synthesis autoencoders (VSA) in this paper, which is a self-supervised learning framework designed for vision transformers. Different from traditional 2D pretraining methods, VSA can be pre-trained with multi-view data. In each iteration, the input to VSA is one view (or multiple views) of a 3D object and the output is a synthesized image in another target pose. The decoder of VSA has several cross-attention blocks, which use the source view as value, source pose as key, and target pose as query. They achieve cross-attention to synthesize the target view. This simple approach realizes large-angle view synthesis and learns spatial invariant representation, where the latter is decent initialization for transformers on downstream tasks, such as 3D classification on ModelNet40, ShapeNet Core55, and ScanObjectNN. VSA outperforms existing methods significantly for linear probing and is competitive for fine-tuning. The code will be made publicly available."}, "cited_paper_content": {"title": "3D Shapenets: A Deep Representation For Volumetric Shapes", "abstract": "3D shape is a crucial but heavily underutilized cue in today's computer vision systems, mostly due to the lack of a good generic shape representation. With the recent availability of inexpensive 2.5D depth sensors (e.g. Microsoft Kinect), it is becoming increasingly important to have a powerful 3D shape representation in the loop. Apart from category recognition, recovering full 3D shapes from view-based 2.5D depth maps is also a critical part of visual understanding. To this end, we propose to represent a geometric 3D shape as a probability distribution of binary variables on a 3D voxel grid, using a Convolutional Deep Belief Network. Our model, 3D ShapeNets, learns the distribution of complex 3D shapes across different object categories and arbitrary poses from raw CAD data, and discovers hierarchical compositional part representation automatically. It naturally supports joint object recognition and shape completion from 2.5D depth maps, and it enables active object recognition through view planning. To train our 3D deep learning model, we construct ModelNet - a large-scale 3D CAD model dataset. Extensive experiments show that our 3D deep representation enables significant performance improvement over the-state-of-the-arts in a variety of tasks."}, "keywords": ["n dimension learnable"], "citation_intent": "method"}
{"citing_id": "2303.16509v1", "cited_id": "2003.08934", "section_title": "B.2. Renderer", "citation": "In order to decode the generated voxel-grid of features into density and radiance fields, we use a NeRF-like #REFR MLP (Multi-layer perceptron).", "text_before_citation": [], "text_after_citation": ["The MLP contains 4 layers of 256 hidden units with a skip-connection on the 3rd hidden layer.", "The skip connection concatenates the input features with the intermediate hidden layer features.", "Similar to NeRF, and for the reasons described in Zhang et al. #OTHEREFR , we Figure I .", "Architecture of the RenderMLP used for decoding the features of the generated voxel grids into density and radiance fields.", "also input the view-directions at a latter layer in the MLP."], "citing_paper_content": {"title": "Holodiffusion: Training A 3D Diffusion Model Using 2D Images", "abstract": "Figure 1. We present HOLODIFFUSION as the first 3D-aware generative diffusion model that produces 3D-consistent images and is trained with only posed image supervision. Here we show a few different samples generated from models trained on different classes of the CO3D dataset [50]."}, "cited_paper_content": {"title": "Nerf: Representing Scenes As Neural Radiance Fields For View Synthesis", "abstract": "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(\\theta, \\phi)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons."}, "keywords": ["density", "radiance fields"], "citation_intent": "method"}
{"citing_id": "2303.16407v1", "cited_id": "1711.05101", "section_title": "Experimental Environment And Parameter Settings", "citation": "We trained all the models with AdamW as the optimizer, with the default parameters as described in #REFR and mini batches with size of 32. It should be noted that ConvNet Schirrmeister et al.", "text_before_citation": ["All the experiments were conducted under the Pytorch framework on a workstation with Intel(R) Xeon(R) Gold 5117 CPUs @ 2.00 GHz and Nvidia Tesla V100 GPUs."], "text_after_citation": ["(2017) has a shallow version and a deep version for EEG decoding, both EEGNet #OTHEREFR and ConvNet Schirrmeister et al.", "(2017) mentioned that shallow-ConvNet has more advantages than deep-ConvNet, thus we only selected shallow-ConvNet (Abbreviated as ConvNet below) in the experiment."], "citing_paper_content": {"title": "Lmda-Net:A Lightweight Multi-Dimensional Attention Network For General Eeg-Based Brain-Computer Interface Paradigms And Interpretability", "abstract": "Electroencephalography (EEG)-based brain-computer interfaces (BCIs) pose a challenge for decoding due to their low spatial resolution and signal-to-noise ratio. Typically, EEG-based recognition of activities and states involves the use of prior neuroscience knowledge to generate quantitative EEG features, which may limit BCI performance. Although neural network-based methods can effectively extract features, they often encounter issues such as poor generalization across datasets, high predicting volatility, and low model interpretability. To address these limitations, we propose a novel lightweight multi-dimensional attention network, called LMDA-Net. By incorporating two novel attention modules designed specifically for EEG signals, the channel attention module and the depth attention module, LMDA-Net is able to effectively integrate features from multiple dimensions, resulting in improved classification performance across various BCI tasks. LMDA-Net was evaluated on four high-impact public datasets, including motor imagery (MI) and P300-Speller paradigms, and was compared with other representative models. The experimental results demonstrate that LMDA-Net outperforms other representative methods in terms of classification accuracy and predicting volatility, achieving the highest accuracy in all datasets within 300 training epochs. Ablation experiments further confirm the effectiveness of the channel attention module and the depth attention module. To facilitate an in-depth understanding of the features extracted by LMDA-Net, we propose class-specific neural network feature interpretability algorithms that are suitable for event-related potentials (ERPs) and event-related desynchronization/synchronization (ERD/ERS). By mapping the output of the specific layer of MDA-Net to the time or spatial domain through class activation maps, the resulting feature visualizations can provide interpretable analysis and establish connections with EEG time-spatial analysis in neuroscience. In summary, LMDA-Net shows great potential as a general online decoding model for various EEG tasks."}, "cited_paper_content": {"title": "Fixing Weight Decay Regularization In Adam", "abstract": "We note that common implementations of adaptive gradient algorithms, such as Adam, limit the potential benefit of weight decay regularization, because the weights do not decay multiplicatively (as would be expected for standard weight decay) but by an additive constant factor. We propose a simple way to resolve this issue by decoupling weight decay and the optimization steps taken w.r.t. the loss function. We provide empirical evidence that our proposed modification (i) decouples the optimal choice of weight decay factor from the setting of the learning rate for both standard SGD and Adam, and (ii) substantially improves Adam's generalization performance, allowing it to compete with SGD with momentum on image classification datasets (on which it was previously typically outperformed by the latter). We also demonstrate that longer optimization runs require smaller weight decay values for optimal results and introduce a normalized variant of weight decay to reduce this dependence. Finally, we propose a version of Adam with warm restarts (AdamWR) that has strong anytime performance while achieving state-of-the-art results on CIFAR-10 and ImageNet32x32. Our source code is available at https://github.com/loshchil/AdamW-and-SGDW"}, "keywords": ["AdamW"], "citation_intent": "method"}
{"citing_id": "2304.00733v2", "cited_id": "1912.06992", "section_title": "Introduction", "citation": "Noisy scene graph annotations in Action Genome #REFR increase the uncertainty of predicted scene graphs.", "text_before_citation": ["The high recall values in prior works suggest that they may have a tendency to overfit on popular predicate classes (e.g.", "in front of / not looking at), without considering how the performances on rare classes (e.g. eating/wiping) are getting impacted #OTHEREFR .", "Predicates lying in the tails often provide more informative depictions of underlying actions and activities in the video.", "Thus, it is important to be able to measure a model's long-term performance not only on the frequently occurring relationships but also on the infrequently occurring ones.", "Data imbalance is, however, not the only challenge in dynamic SGG. As shown in Fig. 2 Figure 2 ."], "text_after_citation": ["factors, including noisy annotations, motion blur, temporal fluctuations of predictions, and a need to focus on only active objects that are involved in an action contribute to the bias in training dynamic SGG models #OTHEREFR .", "As a result, the visual relationship predictions have high uncertainty, thereby increasing the challenge of dynamic SGG manyfold.", "In this paper, we address these sources of bias in dynamic SGG and propose methods to compensate for them.", "We identify missing annotations, multi-label mapping, and triplet (<subject\u2212predicate\u2212object>) variability (Fig 2) as labeling noise, which coupled with the inherent temporal fluctuations in a video can be attributed as data noise that can be modeled as the aleatoric uncertainty #OTHEREFR .", "Another form of uncertainty called the epistemic uncertainty, relates to misleading model predictions due to a lack of sufficient observations #OTHEREFR and is more prevalent for long-tailed data #OTHEREFR ."], "citing_paper_content": {"title": "Unbiased Scene Graph Generation In Videos", "abstract": "The task of dynamic scene graph generation (SGG) from videos is complicated and challenging due to the inherent dynamics of a scene, temporal fluctuation of model predictions, and the long-tailed distribution of the visual relationships in addition to the already existing challenges in image-based SGG. Existing methods for dynamic SGG have primarily focused on capturing spatio-temporal context using complex architectures without addressing the challenges mentioned above, especially the long-tailed distribution of relationships. This often leads to the generation of biased scene graphs. To address these challenges, we introduce a new framework called TEMPURA: TEmporal consistency and Memory Prototype guided UnceRtainty Attenuation for unbiased dynamic SGG. TEMPURA employs object-level temporal consistencies via transformerbased sequence modeling, learns to synthesize unbiased relationship representations using memory-guided training, and attenuates the predictive uncertainty of visual relations using a Gaussian Mixture Model (GMM). Extensive experiments demonstrate that our method achieves significant (up to 10% in some cases) performance gain over existing methods highlighting its superiority in generating more unbiased scene graphs."}, "cited_paper_content": {"title": "Action Genome: Actions As Composition Of Spatio-Temporal Scene Graphs", "abstract": "Action recognition has typically treated actions and activities as monolithic events that occur in videos. However, there is evidence from Cognitive Science and Neuroscience that people actively encode activities into consistent hierarchical part structures. However in Computer Vision, few explorations on representations encoding event partonomies have been made. Inspired by evidence that the prototypical unit of an event is an action-object interaction, we introduce Action Genome, a representation that decomposes actions into spatio-temporal scene graphs. Action Genome captures changes between objects and their pairwise relationships while an action occurs. It contains 10K videos with 0.4M objects and 1.7M visual relationships annotated. With Action Genome, we extend an existing action recognition model by incorporating scene graphs as spatio-temporal feature banks to achieve better performance on the Charades dataset. Next, by decomposing and learning the temporal changes in visual relationships that result in an action, we demonstrate the utility of a hierarchical event decomposition by enabling few-shot action recognition, achieving 42.7% mAP using as few as 10 examples. Finally, we benchmark existing scene graph models on the new task of spatio-temporal scene graph prediction."}, "keywords": ["predicted scene graphs"], "citation_intent": "background"}
{"citing_id": "2304.07597v2", "cited_id": "1604.01685", "section_title": "I. Introduction", "citation": "Instance segmentation is the task of detecting, segmenting, and classifying each object instance in an image #REFR .", "text_before_citation": ["Many biomedical applications require the detection and segmentation of individual cells in microscopy imagery #OTHEREFR .", "For example, analyzing the cellular processes of living cells in time-lapse fluorescence microscopy (TFLM) experiments, requires accurate pixel-level segmentations of individual cells #OTHEREFR - #OTHEREFR .", "Most applications require each cell to be segmented and identified as an unique entity or instance #OTHEREFR ."], "text_after_citation": ["While powerful cell segmentation algorithms have been proposed recently (e.g.", "#OTHEREFR , #OTHEREFR ), segmenting cells in microstructured environments remains challenging, due to the perceptual similarity of microstructurs and cells (cf. Fig. 1 ) #OTHEREFR , #OTHEREFR , #OTHEREFR .", "The vast majority of current state-of-the-art segmentation algorithms utilize deep neural networks #OTHEREFR , #OTHEREFR - #OTHEREFR .", "A key factor driving the development of deep learning-based segmentation algorithms is the widespread availability of pixel-wise annotated datasets.", "Examples include Microsoft COCO #OTHEREFR , Cityscapes #OTHEREFR , ADE20K #OTHEREFR , and the 2018 Data Science Bowl dataset #OTHEREFR . While general cell segmentation datasets are available (e.g."], "citing_paper_content": {"title": "An Instance Segmentation Dataset Of Yeast Cells In Microstructures", "abstract": "Extracting single-cell information from microscopy data requires accurate instance-wise segmentations. Obtaining pixel-wise segmentations from microscopy imagery remains a challenging task, especially with the added complexity of microstructured environments. This paper presents a novel dataset for segmenting yeast cells in microstructures. We offer pixel-wise instance segmentation labels for both cells and trap microstructures. In total, we release 493 densely annotated microscopy images. To facilitate a unified comparison between novel segmentation algorithms, we propose a standardized evaluation strategy for our dataset. The aim of the dataset and evaluation strategy is to facilitate the development of new cell segmentation approaches. The dataset is publicly available at https://christophreich1996.github.io/yeast_ in_microstructures_dataset/."}, "cited_paper_content": {"title": "The Cityscapes Dataset For Semantic Urban Scene Understanding", "abstract": "Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations, 20 000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark."}, "keywords": ["Instance segmentation"], "citation_intent": "background"}
{"citing_id": "2304.05314v1", "cited_id": "1910.05092", "section_title": "I. Introduction", "citation": "Game theory combined with reinforcement learning has been used to model human decision-making and determine an optimal action sequence for intelligent agents in #REFR .", "text_before_citation": ["Emerging vehicular technologies in automation and communication have generated new opportunities to enhance traffic safety and efficiency #OTHEREFR .", "Prior research efforts have shown that in environments consisting solely of connected automated vehicles (CAVs), the traffic throughput and energy consumption can be significantly improved under different vehicle coordination strategies such as optimal control #OTHEREFR , #OTHEREFR , virtual platooning #OTHEREFR , model predictive control #OTHEREFR , and learning-based control #OTHEREFR .", "It is expected that CAVs will gradually penetrate the market and interact with human-driven vehicles (HDVs) over the next several years.", "The presence of HDVs, however, complicates vehicle coordination, given the existing limitations on predicting precise driving behavior.", "Addressing coordination of CAVs in a mixed-traffic environment, where CAVs and HDVs co-exist, has attracted considerable attention recently."], "text_after_citation": ["Mixed-integer linear programming for centralized scheduling and coordinating CAVs interacting with HDVs was proposed in #OTHEREFR .", "Other approaches include dynamic programming #OTHEREFR , statistical modeling of human uncertainties #OTHEREFR , and reachability analysis #OTHEREFR .", "In mixed autonomy, traffic bottlenecks such as merging roadways and intersections pose significant challenges to the decision-making and control of CAVs.", "A multi-agent reinforcement learning framework for CAVs was presented in #OTHEREFR for microscopic simulation where the human drivers are simulated by a car-following model.", "A hierarchical on-ramp merging control strategy for CAVs was presented in #OTHEREFR , where an upper-level planner computes an expected merging position and a low-level trajectory controller guarantees constraints under uncertainties of HDVs."], "citing_paper_content": {"title": "Coordination For Connected Automated Vehicles At Merging Roadways In Mixed Traffic Environment", "abstract": "In this paper, we present a two-level optimal control framework to address motion coordination of connected automated vehicles (CAVs) in the presence of human-driven vehicles (HDVs) in merging scenarios. Our framework combines an unconstrained trajectory solution of a low-level energyoptimal control problem with an upper-level optimization problem that yields the minimum travel time for CAVs. We predict the future trajectories of the HDVs using Newell's car-following model. To handle potential deviations of HDVs' actual behavior from the one predicted, we provide a risktriggered re-planning mechanism for the CAVs based on timeto-conflict. The effectiveness of the proposed control framework is demonstrated via simulations with heterogeneous human driving behaviors and via experiments in a scaled environment."}, "cited_paper_content": {"title": "Modeling Cyber-Physical Human Systems Via An Interplay Between Reinforcement Learning And Game Theory", "abstract": "Predicting the outcomes of cyber-physical systems with multiple human interactions is a challenging problem. This article reviews a game theoretical approach to address this issue, where reinforcement learning is employed to predict the time-extended interaction dynamics. We explain that the most attractive feature of the method is proposing a computationally feasible approach to simultaneously model multiple humans as decision makers, instead of determining the decision dynamics of the intelligent agent of interest and forcing the others to obey certain kinematic and dynamic constraints imposed by the environment. We present two recent exploitations of the method to model 1) unmanned aircraft integration into the National Airspace System and 2) highway traffic. We conclude the article by providing ongoing and future work about employing, improving and validating the method. We also provide related open problems and research opportunities."}, "keywords": ["reinforcement learning"], "citation_intent": "method"}
{"citing_id": "2304.03385v1", "cited_id": "1904.11955", "section_title": "1.2.2.", "citation": "When compared to the results presented here the clear advantage of #REFR is that it deals with deep networks, while we only consider single-layered networks.", "text_before_citation": ["Also, the kind of parameter initialization they privilege (standard Gaussian initialization with zero mean) leads to deviations as powers of n \u22121 , where n determines the width size.", "In contrast, in the present work we take a simpler approach using the Edgeworth series technique in order to explicitly compute the corrections to Gaussianity in terms of an asymptotic series in n \u2212 1 2 , with coefficients determined by the statistics of parameter initialization.", "The paper #OTHEREFR also concerns deep neural networks.", "In that setting it is shown, in particular, that at initialization the difference between the infinite limit NTK \u221e and the finite width NTK (n) can be made arbitrarily small by increasing the width of all hidden layers.", "They also show that after training the difference between the outputs of the finite and the corresponding infinite width networks can be made arbitrarily small by increasing the width."], "text_after_citation": ["Nonetheless, their bounds do not have an explicit quantitative dependence on the width of the layers and their results only apply to ReLu non-linearities; a situation that is largely fixed in our paper.", "The results in #OTHEREFR provide various estimates, in terms of the width n, controlling the behavior of the NTK, and higher order generalization of this object, and of the outputs during training.", "Once again their results apply to deep neural networks, but unfortunately the estimates they provide depend on increasing functions of training time t, a shortfall that, in particular, does not allow to control the t \u2192 \u221e limit; nonetheless the estimates provided in #OTHEREFR are sufficient to show that outputs of training data can be made arbitrarily close to their labels, by appropriately increasing the width n.", "In the, already mentioned, extensive work carried out in #OTHEREFR one can also find results concerning the behavior of outputs of wide neural networks, where it is shown that after training the difference between outputs and labels can be controlled up to an error of the order of n \u22122 , where n determines the width of the networks.", "Concerning the behavior of outputs during training the results in #OTHEREFR are, to the best of our knowledge, the ones with biggest overlap with the ones presented here."], "citing_paper_content": {"title": "Wide Neural Networks: From Non-Gaussian Random Fields At Initialization To The Ntk Geometry Of Training", "abstract": "Recent developments in applications of artificial neural networks with over n = 10 14 parameters make it extremely important to study the large n behaviour of such networks. Most works studying wide neural networks have focused on the infinite width n \u2192 +\u221e limit of such networks and have shown that, at initialization, they correspond to Gaussian processes [12, 20]. In this work we will study their behavior for large, but finite n. Our main contributions are the following: \u2022 The computation of the corrections to Gaussianity in terms of an asymptotic series in n \u2212 1 2. The coefficients in this expansion are determined by the statistics of parameter initialization and by the activation function. \u2022 Controlling the evolution of the outputs of finite width n networks, during training, by computing deviations from the limiting infinite width case (in which the network evolves through a linear flow). This improves previous estimates [1, 16, 10, 18] and along the way we also obtain sharper decay rates for the (finite width) NTK in terms of n, valid during the entire training procedure. As a corollary, we also prove that, with arbitrarily high probability, the training of sufficiently wide neural networks converges to a global minimum of the corresponding quadratic loss function. \u2022 Estimating how the deviations from Gaussianity evolve with training in terms of n. In particular, using a certain metric in the space of measures we find that, along training, the resulting measure is within n \u2212 1 2 (log n) 1+ of the time dependent Gaussian process corresponding to the infinite width network (which is explicitly given by precomposing the initial Gaussian process with the linear flow corresponding to training in the infinite width limit). Contents"}, "cited_paper_content": {"title": "On Exact Computation With An Infinitely Wide Neural Net", "abstract": "How well does a classic deep net architecture like AlexNet or VGG19 classify on a standard dataset such as CIFAR-10 when its width --- namely, number of channels in convolutional layers, and number of nodes in fully-connected internal layers --- is allowed to increase to infinity? Such questions have come to the forefront in the quest to theoretically understand deep learning and its mysteries about optimization and generalization. They also connect deep learning to notions such as Gaussian processes and kernels. A recent paper [Jacot et al., 2018] introduced the Neural Tangent Kernel (NTK) which captures the behavior of fully-connected deep nets in the infinite width limit trained by gradient descent; this object was implicit in some other recent papers. An attraction of such ideas is that a pure kernel-based method is used to capture the power of a fully-trained deep net of infinite width. ::: The current paper gives the first efficient exact algorithm for computing the extension of NTK to convolutional neural nets, which we call Convolutional NTK (CNTK), as well as an efficient GPU implementation of this algorithm. This results in a significant new benchmark for the performance of a pure kernel-based method on CIFAR-10, being $10\\%$ higher than the methods reported in [Novak et al., 2019], and only $6\\%$ lower than the performance of the corresponding finite deep net architecture (once batch normalization, etc. are turned off). Theoretically, we also give the first non-asymptotic proof showing that a fully-trained sufficiently wide net is indeed equivalent to the kernel regression predictor using NTK."}, "keywords": ["single-layered networks", "deep networks"], "citation_intent": "result"}
{"citing_id": "2304.10706v1", "cited_id": "1810.04805", "section_title": "A. Bert", "citation": "The pre-trained language model BERT #REFR has gained significant prominence in recent times for its demonstration of efficacy in comparison to traditional one-way language models and shallow splicing methods.", "text_before_citation": [], "text_after_citation": ["The masked language model (MLM) approach utilized by BERT has been shown to provide more semantic information within generated language representations.", "Our work harnesses the capabilities of BERT by incorporating sentence vectors as inputs to the model, yielding contextual features that serve as word embeddings.", "The representation of each token in sentence S is obtained through the amalgamation of its respective token embedding, segment embedding, and position embedding.", "This is achieved by utilizing the sentence encoding of the form [CLS, E S , SEP ], where CLS signifies the initiation of the sentence and SEP serves as an indicator of the sentence's conclusion."], "citing_paper_content": {"title": "Tc-Gat: Graph Attention Network For Temporal Causality Discovery", "abstract": "The present study explores the intricacies of causal relationship extraction, a vital component in the pursuit of causality knowledge. Causality is frequently intertwined with temporal elements, as the progression from cause to effect is not instantaneous but rather ensconced in a temporal dimension. Thus, the extraction of temporal causality holds paramount significance in the field. In light of this, we propose a method for extracting causality from the text that integrates both temporal and causal relations, with a particular focus on the time aspect. To this end, we first compile a dataset that encompasses temporal relationships. Subsequently, we present a novel model, TC-GAT, which employs a graph attention mechanism to assign weights to the temporal relationships and leverages a causal knowledge graph to determine the adjacency matrix. Additionally, we implement an equilibrium mechanism to regulate the interplay between temporal and causal relations. Our experiments demonstrate that our proposed method significantly surpasses baseline models in the task of causality extraction."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["pre-trained language model"], "citation_intent": "method"}
{"citing_id": "2303.03157v1", "cited_id": "2001.06116", "section_title": "Introduction", "citation": "To resolve this issue, #REFR recently proposed a method to jointly learn a stable uncontrolled dynamics model with a Lyapunov function.", "text_before_citation": ["However, polynomial approximation can be a significant restriction on the class of function approximators used.", "Related Work To this end, there has been substantial growth in literature on data-driven learning of stability certificates for dynamical systems, particularly those modeled with complex parametric representations.", "One of the earliest of such works fits a Lyapunov function for a known uncontrolled dynamical system by penalizing violations of the corresponding Lyapunov decrease condition #OTHEREFR .", "Learning certificate functions such as Lyapunov, barrier, and contraction metric functions for dynamical systems with sampled point penalties or constraints on stability violations is a ubiquitous theme in the literature #OTHEREFR .", "However, when coupled with regression on unknown dynamics, such approaches do not even guarantee the learned model is stable."], "text_after_citation": ["It guarantees stability of the learned dynamics model by restricting it to the stable halfspace described by the Lyapunov decrease condition. arXiv:2303.03157v1 [eess.SY] 6 Mar 2023", "However, its naive extension to controlled dynamics would restrict the learned model to be stabilizable by any control input, thereby hindering its ability to model general controlled nonlinear systems.", "For controlled dynamics, the paradigm of sampled pointwise constraints persists. #OTHEREFR", "(2020) jointly learn a dynamics model and certificate to regularize the dynamics model to perform well over long time horizons with sampled linear matrix inequality (LMI) constraints, yet they do not learn any specific controller. #OTHEREFR", "(2020) assume the dynamics are known, and jointly learn a controller and a contraction metric certifying the stability of the closedloop system with loss terms corresponding to sampled point violations of a stability inequality. For Lyapunov-based approaches, #OTHEREFR"], "citing_paper_content": {"title": "Data-Driven Control With Inherent Lyapunov Stability", "abstract": "Recent advances in learning-based control leverage deep function approximators, such as neural networks, to model the evolution of controlled dynamical systems over time. However, the problem of learning a dynamics model and a stabilizing controller persists, since the synthesis of a stabilizing feedback law for known nonlinear systems is a difficult task, let alone for complex parametric representations that must be fit to data. To this end, we propose a method for jointly learning parametric representations of a nonlinear dynamics model and a stabilizing controller from data. To do this, our approach simultaneously learns a parametric Lyapunov function which intrinsically constrains the dynamics model to be stabilizable by the learned controller. In addition to the stabilizability of the learned dynamics guaranteed by our novel construction, we show that the learned controller stabilizes the true dynamics under certain assumptions on the fidelity of the learned dynamics. Finally, we demonstrate the efficacy of our method on a variety of simulated nonlinear dynamical systems."}, "cited_paper_content": {"title": "Learning Stable Deep Dynamics Models", "abstract": "Deep networks are commonly used to model dynamical systems, predicting how the state of a system will evolve over time (either autonomously or in response to control inputs). Despite the predictive power of these systems, it has been difficult to make formal claims about the basic properties of the learned systems. In this paper, we propose an approach for learning dynamical systems that are guaranteed to be stable over the entire state space. The approach works by jointly learning a dynamics model and Lyapunov function that guarantees non-expansiveness of the dynamics under the learned Lyapunov function. We show that such learning systems are able to model simple dynamical systems and can be combined with additional deep generative models to learn complex dynamics, such as video textures, in a fully end-to-end fashion."}, "keywords": ["stable uncontrolled dynamics"], "citation_intent": "method"}
{"citing_id": "2304.00057v1", "cited_id": "1703.03400", "section_title": "C. Frel Learning Algorithm", "citation": "It is a simplified version of MAML #REFR that fine-tunes the last few layers which can achieve comparable performance as the original algorithm.", "text_before_citation": ["We train the embedding network and classifier jointly with a mini-dataset during the meta-learning phase, which is the same as embedding learning.", "After obtaining the embedding, we further optimize the classifier with a few samples while testing.", "In contrast to embedding learning, we believe fine-tuning can provide better flexibility and granularity, which is more suitable for a dynamic system.", "Different from meta-learning, we only retrain the simple classifier instead of the whole structure, which can reduce the computation, enabling a faster adaption to new tasks.", "This design is inspired by #OTHEREFR , which shows that the effectiveness of meta-learning is mainly due to feature reuse."], "text_after_citation": [], "citing_paper_content": {"title": "Simwisense: Simultaneous Multi-Subject Activity Classification Through Wi-Fi Signals", "abstract": "Recent advances in Wi-Fi sensing have ushered in a plethora of pervasive applications in home surveillance, remote healthcare, road safety, and home entertainment, among others. Most of the existing works are limited to the activity classification of a single human subject at a given time. Conversely, a more realistic scenario is to achieve simultaneous, multi-subject activity classification. The first key challenge in that context is that the number of classes grows exponentially with the number of subjects and activities. Moreover, it is known that Wi-Fi sensing systems struggle to adapt to new environments and subjects. To address both issues, we propose SiMWiSense, the first framework for simultaneous multi-subject activity classification based on Wi-Fi that generalizes to multiple environments and subjects. We address the scalability issue by using the Channel State Information (CSI) computed from the device positioned closest to the subject. We experimentally prove this intuition by confirming that the best accuracy is experienced when the CSI computed by the transceiver positioned closest to the subject is used for classification. To address the generalization issue, we develop a brand-new few-shot learning algorithm named Feature Reusable Embedding Learning (FREL). Through an extensive data collection campaign in 3 different environments and 3 subjects performing 20 different activities simultaneously, we demonstrate that SiMWiSense achieves classification accuracy of up to 97%, while FREL improves the accuracy by 85% in comparison to a traditional Convolutional Neural Network (CNN) and up to 20% when compared to the state-of-the-art few-shot embedding learning (FSEL), by using only 15 seconds of additional data for each class. For reproducibility purposes, we share our 1TB dataset and code repository 1 [1]."}, "cited_paper_content": {"title": "Model-Agnostic Meta-Learning For Fast Adaptation Of Deep Networks", "abstract": "We propose an algorithm for meta-learning that is model-agnostic, in the sense that it is compatible with any model trained with gradient descent and applicable to a variety of different learning problems, including classification, regression, and reinforcement learning. The goal of meta-learning is to train a model on a variety of learning tasks, such that it can solve new learning tasks using only a small number of training samples. In our approach, the parameters of the model are explicitly trained such that a small number of gradient steps with a small amount of training data from a new task will produce good generalization performance on that task. In effect, our method trains the model to be easy to fine-tune. We demonstrate that this approach leads to state-of-the-art performance on two few-shot image classification benchmarks, produces good results on few-shot regression, and accelerates fine-tuning for policy gradient reinforcement learning with neural network policies."}, "keywords": ["layers"], "citation_intent": "method"}
{"citing_id": "2305.02160v1", "cited_id": "1805.01070", "section_title": "Causal Explainability Methods", "citation": "Probing methods #REFR ) train an external model -a probe -to predict some properties of interest from the latent representations.", "text_before_citation": ["General-domain #OTHEREFR", "(2018) attempt to intervene in an unsupervised way on the hidden space by constructing several even-spaced Variational Autoencoders (VAEs) throughout a CNN, but they only train with a reconstruction loss instead of explicitly optimizing for impact."], "text_after_citation": ["To further investigate causal effects of the features learned from probing, #OTHEREFR", "(2021) assess the influence of a causal intervention by removing a feature.", "However, subsequent work #OTHEREFR shows that such methods generalize poorly to unseen samples.", "Moreover, as #OTHEREFR points out, the disconnect between the probing model and the original model may result in the properties not being utilized in the original model's prediction task.", "Causal Mediation Analysis (CMA) #OTHEREFR measures the change in an output following a counterfactual intervention in an intermediate variable, or mediator. NLP According to Feder et al."], "citing_paper_content": {"title": "Explaining Language Models' Predictions With High-Impact Concepts", "abstract": "The emergence of large-scale pretrained language models has posed unprecedented challenges in deriving explanations of why the model has made some predictions. Stemmed from the compositional nature of languages, spurious correlations have further undermined the trustworthiness of NLP systems, leading to unreliable model explanations that are merely correlated with the output predictions. To encourage fairness and transparency, there exists an urgent demand for reliable explanations that allow users to consistently understand the model's behavior. In this work, we propose a complete framework for extending conceptbased interpretability methods to NLP. Specifically, we propose a post-hoc interpretability method for extracting predictive high-level features (concepts) from the pretrained model's hidden layer activations. We optimize for features whose existence causes the output predictions to change substantially, i.e. generates a high impact. Moreover, we devise several evaluation metrics that can be universally applied. Extensive experiments on real and synthetic tasks demonstrate that our method achieves superior results on predictive impact, usability, and faithfulness compared to the baselines."}, "cited_paper_content": {"title": "What You Can Cram Into A Single Vector: Probing Sentence Embeddings For Linguistic Properties", "abstract": "Although much effort has recently been devoted to training high-quality sentence embeddings, we still have a poor understanding of what they are capturing. \"Downstream\" tasks, often based on sentence classification, are commonly used to evaluate the quality of sentence representations. The complexity of the tasks makes it however difficult to infer what kind of information is present in the representations. We introduce here 10 probing tasks designed to capture simple linguistic features of sentences, and we use them to study embeddings generated by three different encoders trained in eight distinct ways, uncovering intriguing properties of both encoders and training methods."}, "keywords": ["latent representations"], "citation_intent": "method"}
{"citing_id": "2305.02171v1", "cited_id": "1711.04574", "section_title": "Background", "citation": "More recently, the Differentiable Inductive Logic Programming ( ILP) approach #REFR was proposed, addressing cycles through negation.", "text_before_citation": ["Other exceptions would then be added as needed without changing the original rule.", "Unfortunately, this approach does not adapt well to exceptions to the exceptions such as an abnormal penguins (a hypothetical super-penguin that is capable of flying).", "At present, there is a tension between the above attempts to formalizing non-monotonicity and large-scale data-driven approaches based on neural networks and natural language that are efficient but lack any formalization.", "In this paper we seek to investigate approaches to solving PET and other simple examples that can be formalized but that work using the same tools as the large-scale network models.", "Work has been conducted to formalize NMR in neural networks starting with the Connectionist Inductive Learning and Logic Programming System (CILP) #OTHEREFR , later developed into a system for statistical relational learning."], "text_after_citation": ["Probabilistic approaches have also been developed which can implement a form of nonmonotonicity or at least avoid the problems of classical logic by assigning probabilities to beliefs expressed as Horn clauses, e.g. DeepProbLog #OTHEREFR .", "In this paper, rather than mapping symbolic representations into neural networks and vice-versa, we are interested in the interplay between learning and reasoning as part of a curriculum.", "We focus on the Logic Tensor Network (LTN) #OTHEREFR because it is a highly modular NeSy framework applicable in principle to any underlying neural network model and based on the canonical, highly expressive FOL language.", "Additionally, the LTN has shown promise for learning in continual mode #OTHEREFR .", "The LTN relies on two main ideas, the grounding of predicates and logical axioms into vectors and Real Logic which maps the satisfiability of the logical axioms to a real number in the interval {0,1} thus enabling viewing satisfiability as optimization."], "citing_paper_content": {"title": "Continual Reasoning: Non-Monotonic Reasoning In Neurosymbolic Ai Using Continual Learning", "abstract": "Despite the extensive investment and impressive recent progress at reasoning by similarity, deep learning continues to struggle with more complex forms of reasoning such as non-monotonic and commonsense reasoning. Non-monotonicity is a property of non-classical reasoning typically seen in commonsense reasoning, whereby a reasoning system is allowed (differently from classical logic) to jump to conclusions which may be retracted later, when new information becomes available. Neural-symbolic systems such as Logic Tensor Networks (LTN) have been shown to be effective at enabling deep neural networks to achieve reasoning capabilities. In this paper, we show that by combining a neural-symbolic system with methods from continual learning, LTN can obtain a higher level of accuracy when addressing nonmonotonic reasoning tasks. Continual learning is added to LTNs by adopting a curriculum of learning from knowledge and data with recall. We call this process Continual Reasoning, a new methodology for the application of neural-symbolic systems to reasoning tasks. Continual Reasoning is applied to a prototypical non-monotonic reasoning problem as well as other reasoning examples. Experimentation is conducted to compare and analyze the effects that different curriculum choices may have on overall learning and reasoning results. Results indicate significant improvement on the prototypical nonmonotonic reasoning problem and a promising outlook for the proposed approach on statistical relational learning examples."}, "cited_paper_content": {"title": "Learning Explanatory Rules From Noisy Data", "abstract": "Artificial Neural Networks are powerful function approximators capable of modelling solutions to a wide variety of problems, both supervised and unsupervised. As their size and expressivity increases, so too does the variance of the model, yielding a nearly ubiquitous overfitting problem. Although mitigated by a variety of model regularisation methods, the common cure is to seek large amounts of training data---which is not necessarily easily obtained---that sufficiently approximates the data distribution of the domain we wish to test on. In contrast, logic programming methods such as Inductive Logic Programming offer an extremely data-efficient process by which models can be trained to reason on symbolic domains. However, these methods are unable to deal with the variety of domains neural networks can be applied to: they are not robust to noise in or mislabelling of inputs, and perhaps more importantly, cannot be applied to non-symbolic domains where the data is ambiguous, such as operating on raw pixels. In this paper, we propose a Differentiable Inductive Logic framework, which can not only solve tasks which traditional ILP systems are suited for, but shows a robustness to noise and error in the training data which ILP cannot cope with. Furthermore, as it is trained by backpropagation against a likelihood objective, it can be hybridised by connecting it with neural networks over ambiguous data in order to be applied to domains which ILP cannot address, while providing data efficiency and generalisation beyond what neural networks on their own can achieve."}, "keywords": ["negation", "ILP"], "citation_intent": "method"}
{"citing_id": "2304.14092v1", "cited_id": "1803.06406", "section_title": "Ii. Related Work", "citation": "Limoyo #REFR proposed a eyeto-hand self-calibration method, which estimates robot-sensor relation by aligning a contact-based point cloud generated by manipulator and vision-based point cloud generated by the sensor.", "text_before_citation": ["Hand-eye calibration techniques for 3-D sensor specifically were also put forward, whereas calibration rig with high manufacturing accuracy and known dimensions is a must in most of relevant literatures.", "Typical instances include calibration board #OTHEREFR , standard sphere #OTHEREFR and free-form surface with design CAD model #OTHEREFR , #OTHEREFR .", "In addition to the costly high-precision machining, these methods can be unapplicable due to unavailability and accuracy degeneration of calibration rig.", "Method using common geometry like plane and its normal vector to make calibration rig dispensable was reported in #OTHEREFR .", "Nevertheless, auxiliary tool is needed to measure points coordinates on the plane beforehand, which brings extra operational complexity."], "text_after_citation": ["Although experiments were conducted successfuly with two rectangular boxes and on-board hardware only, the contact-based point cloud generation using manipulator under impedance control is rather time consuming, and this method can not be applied facing SOR and simple plane, as the registration is realized by pair-wise ICP.", "To the best of our knowledge, RegHEC is the first 3-D sensor hand-eye calibration technique that 1) uses arbitrary available object with little requirement(can be SOR or plane), 2) relies on 3-D measurements only, and 3) works in both eye-in-hand and eye-to-hand cases.", "It returns the calibrated hand-eye relation together with multi-view point clouds registration, thus is more beneficial for robotic 3-D reconstruction than for other 3-D vision-guided tasks."], "citing_paper_content": {"title": "Reghec: Hand-Eye Calibration Via Simultaneous Multi-View Point Clouds Registration Of Arbitrary Object", "abstract": "RegHEC is a registration-based hand-eye calibration technique with no need for accurate calibration rig but arbitrary available objects, applicable for both eye-in-hand and eye-to-hand cases. It tries to find the hand-eye relation which brings multi-view point clouds of arbitrary scene into simultaneous registration under a common reference frame. RegHEC first achieves initial alignment of multi-view point clouds via Bayesian optimization, where registration problem is modeled as a Gaussian process over hand-eye relation and the covariance function is modified to be compatible with distance metric in 3-D motion space SE(3), then passes the initial guess of hand-eye relation to an Anderson Accelerated ICP variant for later fine registration and accurate calibration. RegHEC has little requirement on calibration object, it is applicable with sphere, cone, cylinder and even simple plane, which can be quite challenging for correct point cloud registration and sensor motion estimation using existing methods. While suitable for most 3-D vision guided tasks, RegHEC is especially favorable for robotic 3-D reconstruction, as calibration and multi-view point clouds registration of reconstruction target are unified into a single process. Our technique is verified with extensive experiments using varieties of arbitrary objects and real hand-eye system. We release an open-source C++ implementation of RegHEC."}, "cited_paper_content": {"title": "Self-Calibration Of Mobile Manipulator Kinematic And Sensor Extrinsic Parameters Through Contact-Based Interaction", "abstract": "We present a novel approach for mobile manipulator self-calibration using contact information. Our method, based on point cloud registration, is applied to estimate the extrinsic transform between a fixed vision sensor mounted on a mobile base and an end effector. Beyond sensor calibration, we demonstrate that the method can be extended to include manipulator kinematic model parameters, which involves a non-rigid registration process. Our procedure uses on-board sensing exclusively and does not rely on any external measurement devices, fiducial markers, or calibration rigs. Further, it is fully automatic in the general case. We experimentally validate the proposed method on a custom mobile manipulator platform, and demonstrate centimetre-level post-calibration accuracy in positioning of the end effector using visual guidance only. We also discuss the stability properties of the registration algorithm, in order to determine the conditions under which calibration is possible."}, "keywords": ["eyeto-hand self-calibration method"], "citation_intent": "method"}
{"citing_id": "2305.02921v1", "cited_id": "0807.3917", "section_title": "A. Polar Codes", "citation": "A polar code of length N = 2 n is constructed by selecting a set A of indices i \u2208 [0, N \u2212 1] with high reliability #REFR .", "text_before_citation": ["T denotes the transpose of the matrix.", "The generator matrix of a polar code is formed by selecting a set of rows of G N .", "We use A to denote the set of indices of these rows and C(A) to denote the linear code generated by the set of rows of G N indexed by A.", "Note that A \u2286 [N ].", "The characterization of the information set A for polar codes relies on the channel polarization theorem #OTHEREFR and the concept of synthetic channel reliability."], "text_after_citation": ["The indices in A are dedicated to information bits, while the rest of the synthetic channels with indices in A c [0, N \u2212 1] \\ A are used to transmit a known value, '0' by default, which are called frozen bits.", "Regardless of the method we use for forming the set A for a polar code, every synthetic channel represented by index i, denoted by W i n where i \u2208 A must be more reliable than any synthetic channels in A c .", "Polar codes with the information set A can also be considered as monomial codes #OTHEREFR where the relation between the generating set of monomials I \u2282 M n discussed in Section II-B and the information set A is as follows:", "\u2200f \u2208 I, \u2203 i \u2208 A, where supp(i) = [n]\\ ind(f ). (2)", "Accordingly, we can define set A m\u2212r equivalent to I r , collecting the row indices with | supp(i)| = m \u2212 r for i \u2208 A m\u2212r where we have m = n for polar codes while m is used for decreasing monomial codes in general throughout this paper."], "citing_paper_content": {"title": "On The Closed-Form Weight Enumeration Of Polar Codes: 1.5D-Weight Codewords", "abstract": "The weight distribution of error correction codes is a critical determinant of their error-correcting performance, making enumeration of utmost importance. In the case of polar codes, the minimum weight wmin (which is equal to minimum distance d) is the only weight for which an explicit enumerator formula is currently available. Having closed-form weight enumerators for polar codewords with weights greater than the minimum weight not only simplifies the enumeration process but also provides valuable insights towards constructing better polar-like codes. In this paper, we contribute towards understanding the algebraic structure underlying higher weights by analyzing Minkowski sums of orbits. Our approach builds upon the lower triangular affine (LTA) group of decreasing monomial codes. Specifically, we propose a closed-form expression for the enumeration of codewords with weight 1.5 wmin. Our simulations demonstrate the potential for extending this method to higher weights."}, "cited_paper_content": {"title": "Channel Polarization: A Method For Constructing Capacity-Achieving Codes For Symmetric Binary-Input Memoryless Channels", "abstract": "A method is proposed, called channel polarization, to construct code sequences that achieve the symmetric capacity $I(W)$ of any given binary-input discrete memoryless channel (B-DMC) $W$. The symmetric capacity is the highest rate achievable subject to using the input letters of the channel with equal probability. Channel polarization refers to the fact that it is possible to synthesize, out of $N$ independent copies of a given B-DMC $W$, a second set of $N$ binary-input channels $\\{W_N^{(i)}:1\\le i\\le N\\}$ such that, as $N$ becomes large, the fraction of indices $i$ for which $I(W_N^{(i)})$ is near 1 approaches $I(W)$ and the fraction for which $I(W_N^{(i)})$ is near 0 approaches $1-I(W)$. The polarized channels $\\{W_N^{(i)}\\}$ are well-conditioned for channel coding: one need only send data at rate 1 through those with capacity near 1 and at rate 0 through the remaining. Codes constructed on the basis of this idea are called polar codes. The paper proves that, given any B-DMC $W$ with $I(W)>0$ and any target rate $R<I(W)$, there exists a sequence of polar codes $\\{{\\mathscr C}_n;n\\ge 1\\}$ such that ${\\mathscr C}_n$ has block-length $N=2^n$, rate $\\ge R$, and probability of block error under successive cancellation decoding bounded as $P_{e}(N,R) \\le \\bigoh(N^{-\\frac14})$ independently of the code rate. This performance is achievable by encoders and decoders with complexity $O(N\\log N)$ for each."}, "keywords": ["polar code"], "citation_intent": "background"}
{"citing_id": "2304.09670v1", "cited_id": "2003.04297", "section_title": "A. Overview", "citation": "Specifically, the global branch utilizes the standard MoCo #REFR style CL method to discriminate global representations, while the local branch aligns the local semantics by enforcing assignment consistency between position-matched pairs over a set of prototypes to learn object-separable representations and avoid semantic incompleteness and misalignment.", "text_before_citation": ["The teacher and the student share the same architecture (CNN or ViT).", "The parameters of the student are exponentially moving averaged (EMA) to the parameters of teacher.", "The student maps the masked image to latent embedding, while the teacher encodes the augmented image to preserve semantic integrity and provide contrastive supervision to guide the student.", "The MIM branch applies a MIM head to reconstruct the masked image using the latent embedding output by the student in both spatial and frequency domains.", "The other two branches project the latent embedding to different feature spaces to avoid semantics confusion and leverage CL to learn separable representations by aligning the embedding of the student and teacher at both global and local scales."], "text_after_citation": ["After pre-training using CMID with a large amount of unlabeled data, only the student is used for downstream tasks to evaluate the model performance.", "In the following subsections, we will provide a more detailed explanation of the different components of CMID."], "citing_paper_content": {"title": "Cmid: A Unified Self-Supervised Learning Framework For Remote Sensing Image Understanding", "abstract": "Self-supervised learning (SSL) has gained widespread attention in the remote sensing (RS) and earth observation (EO) communities owing to its ability to learn task-agnostic representations without human-annotated labels. Nevertheless, most existing RS SSL methods are limited to learning either global semantic separable or local spatial perceptible representations. We argue that this learning strategy is suboptimal in the realm of RS, since the required representations for different RS downstream tasks are often varied and complex. In this study, we proposed a unified SSL framework that is better suited for RS images representation learning. The proposed SSL framework, Contrastive Mask Image Distillation (CMID), is capable of learning representations with both global semantic separability and local spatial perceptibility by combining contrastive learning (CL) with masked image modeling (MIM) in a self-distillation way. Furthermore, our CMID learning framework is architecture-agnostic, which is compatible with both convolutional neural networks (CNN) and vision transformers (ViT), allowing CMID to be easily adapted to a variety of deep learning (DL) applications for RS understanding. Comprehensive experiments have been carried out on four downstream tasks (i.e. scene classification, semantic segmentation, object-detection, and change detection) and the results show that models pre-trained using CMID achieve better performance than other state-of-the-art SSL methods on multiple downstream tasks. The code and pre-trained models will be made available at https://github.com/NJU-LHRS/official-CMID to facilitate SSL research and speed up the development of RS images DL applications."}, "cited_paper_content": {"title": "Improved Baselines With Momentum Contrastive Learning", "abstract": "Contrastive unsupervised learning has recently shown encouraging progress, e.g., in Momentum Contrast (MoCo) and SimCLR. In this note, we verify the effectiveness of two of SimCLR's design improvements by implementing them in the MoCo framework. With simple modifications to MoCo---namely, using an MLP projection head and more data augmentation---we establish stronger baselines that outperform SimCLR and do not require large training batches. We hope this will make state-of-the-art unsupervised learning research more accessible. Code will be made public."}, "keywords": ["object-separable representations"], "citation_intent": "method"}
{"citing_id": "2303.03912v1", "cited_id": "1810.04805", "section_title": "Experiment Settings And Evaluation Metrics", "citation": "To implement our model, we choose uncased BERT-base #REFR as the encoder on DocRED and set the embedding dimension to 768.", "text_before_citation": [], "text_after_citation": ["For CDR dataset, we pick up BioBERT-Base v1.1 #OTHEREFR , which re-trained the BERT-base-cased model on biomedical corpora. All hyper-parameters are tuned based on the development set.", "Other parameters in the network are all obtained by random orthogonal initialization #OTHEREFR and updated during training.", "For a fair comparison, we follow the same experimental settings from previous works.", "We apply F 1 and Ign F 1 as the evaluation metrics on DocRED.", "F 1 scores can be obtained by calculation through an online interface."], "citing_paper_content": {"title": "Document-Level Relation Extraction With Cross-Sentence Reasoning Graph", "abstract": "Relation extraction (RE) has recently moved from the sentence-level to document-level, which requires aggregating document information and using entities and mentions for reasoning. Existing works put entity nodes and mention nodes with similar representations in a document-level graph, whose complex edges may incur redundant information. Furthermore, existing studies only focus on entity-level reasoning paths without considering global interactions among entities cross-sentence. To these ends, we propose a novel document-level RE model with a GRaph information Aggregation and Cross-sentence Reasoning network (GRACR). Specifically, a simplified document-level graph is constructed to model the semantic information of all mentions and sentences in a document, and an entity-level graph is designed to explore relations of long-distance crosssentence entity pairs. Experimental results show that GRACR achieves excellent performance on two public datasets of document-level RE. It is especially effective in extracting potential relations of cross-sentence entity pairs. Our code is available at https://github.com/UESTC-LHF/GRACR."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["uncased BERT-base"], "citation_intent": "method"}
{"citing_id": "2305.00632v1", "cited_id": "2003.08001", "section_title": "Embedding Model Evaluation", "citation": "Note that our knowledge graph has many entities with relatively sparse connections, making it hard to get metrics as good as those from benchmarking #REFR .", "text_before_citation": ["Still, the relative performance between the three models is similar to that observed in Table 8 .", "This suggests that our comparison results are consistent even though the knowledge graph grows over time.", "Through comparison between the embedding models, we determine that TransE performs well for the task at hand.", "The metrics for the TransE model are sufficient for good prediction, as validated by our experiments in Section 5.", "We therefore selected TransE as the embedding model for the prediction of unknown associations."], "text_after_citation": ["Fine-tuning training parameters of the embedding model may help further improve these metrics.", "Last, we investigate the influence of optimizing the knowledge graph under the TransE model, specifically removing unconnected CPE and CVE entities. The results are shown in Table 10 .", "After optimizing the knowledge graph, the MRR increases from 0.232 to 0.290.", "The Hist@N scores also increase, indicating that positive triples are more likely to be found in top-N results.", "We conclude that removing unconnected CPE and CVE entries indeed improves the quality of our knowledge graph."], "citing_paper_content": {"title": "Uncovering Cwe-Cve-Cpe Relations With Threat Knowledge Graphs", "abstract": "Security assessment relies on public information about products, vulnerabilities, and weaknesses. So far, databases in these categories have rarely been analyzed in combination. Yet, doing so could help predict unreported vulnerabilities and identify common threat patterns. In this paper, we propose a methodology for producing and optimizing a knowledge graph that aggregates knowledge from common threat databases (CVE, CWE, and CPE). We apply the threat knowledge graph to predict associations between threat databases, specifically between products, vulnerabilities, and weaknesses. We evaluate the prediction performance both in closed world with associations from the knowledge graph, and in open world with associations revealed afterward. Using rank-based metrics (i.e., Mean Rank, Mean Reciprocal Rank, and Hits@N scores), we demonstrate the ability of the threat knowledge graph to uncover many associations that are currently unknown but will be revealed in the future, which remains useful over different time periods. We propose approaches to optimize the knowledge graph, and show that they indeed help in further uncovering associations. CCS Concepts: \u2022 Security and privacy \u2192 Software security engineering; \u2022 Computing methodologies \u2192 Knowledge representation and reasoning."}, "cited_paper_content": {"title": "Realistic Re-Evaluation Of Knowledge Graph Completion Methods: An Experimental Study", "abstract": "In the active research area of employing embedding models for knowledge graph completion, particularly for the task of link prediction, most prior studies used two benchmark datasets FB15k and WN18 in evaluating such models. Most triples in these and other datasets in such studies belong to reverse and duplicate relations which exhibit high data redundancy due to semantic duplication, correlation or data incompleteness. This is a case of excessive data leakage---a model is trained using features that otherwise would not be available when the model needs to be applied for real prediction. There are also Cartesian product relations for which every triple formed by the Cartesian product of applicable subjects and objects is a true fact. Link prediction on the aforementioned relations is easy and can be achieved with even better accuracy using straightforward rules instead of sophisticated embedding models. A more fundamental defect of these models is that the link prediction scenario, given such data, is non-existent in the real-world. This paper is the first systematic study with the main objective of assessing the true effectiveness of embedding models when the unrealistic triples are removed. Our experiment results show these models are much less accurate than what we used to perceive. Their poor accuracy renders link prediction a task without truly effective automated solution. Hence, we call for re-investigation of possible effective approaches."}, "keywords": ["knowledge graph"], "citation_intent": "background"}
{"citing_id": "2304.11697v1", "cited_id": "1903.11027", "section_title": "A. Detection Model Degradation Under Noise", "citation": "The severity of the noise impact, from greatest to least, is Gaussian N oise > F rost N oise > M otion Blur #REFR In most cases, for the same class and interference level of noise, the Depth model has higher detection accuracy.", "text_before_citation": ["We list the mAP of the proposed model under different noise settings.", "Among them, the prediction accuracy of the RGB model shows a nonlinear decrease for the three types of noise, with mAP decreasing from 76.34 to 0.28(Gaussian), 6.30(Motion), and 7.23(Frost), respectively.", "Compared with the detection results of clean data, the accuracy of all 15 experiments decreases.", "Similarly, the prediction accuracy of the Depth model also decreases nonlinearly with increasing noise intensity, with mAP decreasing from 76.34 to 1.69(Gaussian), 31.62(Motion), and 24.86(Frost), respectively. The accuracy decreases in 14/15 experiments.", "In addition, the sensitivity of both modal data to these noises is consistent."], "text_after_citation": ["In all 15 sets of experiments, the Depth model outperformed the RGB model in terms of accuracy, which is the same as the comparison results on clean data.", "The average accuracy increments are, 7.74+ for Car, 20.57+ for Pedestrian, 22.26+ for Cyclist, and 16.86+ for mAP."], "citing_paper_content": {"title": "Informative Data Selection With Uncertainty For Multi-Modal Object Detection", "abstract": "Noise has always been nonnegligible trouble in object detection by creating confusion in model reasoning, thereby reducing the informativeness of the data. It can lead to inaccurate recognition due to the shift in the observed pattern, that requires a robust generalization of the models. To implement a general vision model, we need to develop deep learning models that can adaptively select valid information from multi-modal data. This is mainly based on two reasons. Multi-modal learning can break through the inherent defects of single-modal data, and adaptive information selection can reduce chaos in multi-modal data. To tackle this problem, we propose a universal uncertainty-aware multi-modal fusion model. It adopts a multi-pipeline loosely coupled architecture to combine the features and results from point clouds and images. To quantify the correlation in multimodal information, we model the uncertainty, as the inverse of data information, in different modalities and embed it in the bounding box generation. In this way, our model reduces the randomness in fusion and generates reliable output. Moreover, we conducted a completed investigation on the KITTI 2D object detection dataset and its derived dirty data. Our fusion model is proven to resist severe noise interference like Gaussian, motion blur, and frost, with only slight degradation. The experiment results demonstrate the benefits of our adaptive fusion. Our analysis on the robustness of multi-modal fusion will provide further insights for future research."}, "cited_paper_content": {"title": "Nuscenes: A Multimodal Dataset For Autonomous Driving", "abstract": "Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology. Image-based benchmark datasets have driven development in computer vision tasks such as object detection, tracking and segmentation of agents in the environment. Most autonomous vehicles, however, carry a combination of cameras and range sensors such as lidar and radar. As machine learning based methods for detection and tracking become more prevalent, there is a need to train and evaluate such methods on datasets containing range sensor data along with images. In this work we present nuTonomy scenes (nuScenes), the first published dataset to carry the full autonomous vehicle sensor suite: 6 cameras, 5 radars and 1 lidar, all with full 360 degree field of view. nuScenes comprises 1000 scenes, each 20s long and fully annotated with 3D bounding boxes for 23 classes and 8 attributes. It has 7x as many annotations and 100x as many images as the pioneering KITTI dataset. We define novel 3D detection and tracking metrics. We also provide careful dataset analysis as well as baselines for lidar and image based detection and tracking. Data, development kit and more information are available online at this http URL."}, "keywords": ["higher detection accuracy"], "citation_intent": "background"}
{"citing_id": "2303.07963v1", "cited_id": "1911.11763", "section_title": "B. Descriptor", "citation": "SuperGlue #REFR uses a module based on attention graphs that alternately stacks 'self-attention' and 'cross-attention' layers.", "text_before_citation": ["where ind is the current value index of", "g xi,", "EQUATION", "where W s E \u2208 R d\u00d7d is a learned projection matrix.", "3) Attention mechanism: A key point of recent descriptors of points cloud is the introduction of an attention mechanism that highlights some features dynamically."], "text_after_citation": ["The former links all the nodes of a point cloud to each other, while the latter links each point of set X to all points of set Y .", "Contrary to SuperGlue #OTHEREFR or MDGAT #OTHEREFR , which compute the attention weights on the encoding vectors, some methods propose adding information on local inter-points geometry at the entry of the mechanism.", "For instance, #OTHEREFR associates the 3D coordinates of each point with the descriptor, while GeoTransformer #OTHEREFR proposes to use the distances and angles between each point and its k nearest neighbours.", "Alternatively, in our approach, we propose the use of four attention heads with geometric self-attention inside each set X and Y integrating the associated normals embeddings e X and e Y , respectively, followed by cross-attention between the two sets of points and then alternate between them for L times.", "4) Self-attention: This type of layer predicts an attentionbased featuref i for each point of a point cloud (X or Y ), paying attention to all the other points of the same cloud."], "citing_paper_content": {"title": "Rocnet: 3D Robust Registration Of Point-Clouds Using Deep Learning", "abstract": "This paper introduces a new method for 3D point cloud registration based on deep learning. The architecture is composed of three distinct blocs: (i) an encoder composed of a convolutional graph-based descriptor that encodes the immediate neighbourhood of each point and an attention mechanism that encodes the variations of the surface normals. Such descriptors are refined by highlighting attention between the points of the same set and then between the points of the two sets. (ii) a matching process that estimates a matrix of correspondences using the Sinkhorn algorithm. (iii) Finally, the rigid transformation between the two point clouds is calculated by RANSAC using the K c best scores from the correspondence matrix. We conduct experiments on the ModelNet40 dataset, and our proposed architecture shows very promising results, outperforming state-of-the-art methods in most of the simulated configurations, including partial overlap and data augmentation with Gaussian noise."}, "cited_paper_content": {"title": "Superglue: Learning Feature Matching With Graph Neural Networks", "abstract": "This paper introduces SuperGlue, a neural network that matches two sets of local features by jointly finding correspondences and rejecting non-matchable points. Assignments are estimated by solving a differentiable optimal transport problem, whose costs are predicted by a graph neural network. We introduce a flexible context aggregation mechanism based on attention, enabling SuperGlue to reason about the underlying 3D scene and feature assignments jointly. Compared to traditional, hand-designed heuristics, our technique learns priors over geometric transformations and regularities of the 3D world through end-to-end training from image pairs. SuperGlue outperforms other learned approaches and achieves state-of-the-art results on the task of pose estimation in challenging real-world indoor and outdoor environments. The proposed method performs matching in real-time on a modern GPU and can be readily integrated into modern SfM or SLAM systems. The code and trained weights are publicly available at https://github.com/magicleap/SuperGluePretrainedNetwork."}, "keywords": ["attention graphs"], "citation_intent": "method"}
{"citing_id": "2304.09876v1", "cited_id": "1602.05629", "section_title": "Introduction", "citation": "Federated Learning (FL) is a well-established training algorithm that addresses this by allowing a model to be trained decentrally without physically sharing the data but instead sharing the model information only #REFR .", "text_before_citation": ["Data analytics hold the key to ensuring food security and sustainability.", "Machine learning has been widely adapted to provide technical solutions to analytical problems in agriculture and food sectors, such as crop yield prediction #OTHEREFR , consumption demand forecasting #OTHEREFR , crop and disease detection #OTHEREFR , quality control and intelligent scheduling #OTHEREFR , and several others.", "Typically, building such statistical models requires large amounts of data collected from various sources, i.e., different farms, supply chains, and other stakeholders.", "However, individually, they may not have adequate data to train competent machine learning models for the tasks.", "While combining their data into a centralized silo may improve data quality, collecting it may be challenging due to commercially sensitive information and reputational risks #OTHEREFR ."], "text_after_citation": ["Each participating device (referred to as a client) participates in training in an isolated environment and is coordinated by the central server.", "As a result, FL allows models to be collaboratively trained on large datasets while preserving data privacy, and this can be leveraged within the agri-food industry when training on local dataset alone is insufficient.", "However, with the progressive improvements in deep learning models, their number of parameters have increased exponentially #OTHEREFR .", "Since FL requires the transmission of the model between the server and the clients in each round, the communication cost often becomes a bottleneck #OTHEREFR .", "Large models also make edge device deployment challenging as they consume more memory footprint and computational power."], "citing_paper_content": {"title": "Model Pruning Enables Localized And Efficient Federated Learning For Yield Forecasting And Data Sharing", "abstract": "Federated Learning (FL) presents a decentralized approach to model training in the agrifood sector and offers the potential for improved machine learning performance, while ensuring the safety and privacy of individual farms or data silos. However, the conventional FL approach has two major limitations. First, the heterogeneous data on individual silos can cause the global model to perform well for some clients but not all, as the update direction on some clients may hinder others after they are aggregated. Second, it is lacking with respect to the efficiency perspective concerning communication costs during FL and large model sizes. This paper proposes a new technical solution that utilizes network pruning on client models and aggregates the pruned models. This method enables local models to be tailored to their respective data distribution and mitigate the data heterogeneity present in agri-food data. Moreover, it allows for more compact models that consume less data during transmission. We experiment with a soybean yield forecasting dataset and find that this approach can improve inference performance by 15.5% to 20% compared to FedAvg, while reducing local model sizes by up to 84% and the data volume communicated between the clients and the server by 57.1% to 64.7%."}, "cited_paper_content": {"title": "Communication-Efficient Learning Of Deep Networks From Decentralized Data", "abstract": "Modern mobile devices have access to a wealth of data suitable for learning models, which in turn can greatly improve the user experience on the device. For example, language models can improve speech recognition and text entry, and image models can automatically select good photos. However, this rich data is often privacy sensitive, large in quantity, or both, which may preclude logging to the data center and training there using conventional approaches. We advocate an alternative that leaves the training data distributed on the mobile devices, and learns a shared model by aggregating locally-computed updates. We term this decentralized approach Federated Learning. We present a practical method for the federated learning of deep networks based on iterative model averaging, and conduct an extensive empirical evaluation, considering five different model architectures and four datasets. These experiments demonstrate the approach is robust to the unbalanced and non-IID data distributions that are a defining characteristic of this setting. Communication costs are the principal constraint, and we show a reduction in required communication rounds by 10-100x as compared to synchronized stochastic gradient descent."}, "keywords": ["Federated Learning"], "citation_intent": "background"}
{"citing_id": "2304.02199v2", "cited_id": "1405.0312", "section_title": "Related Work", "citation": "In axis-aligned object detection, the output rectangles have four degrees of freedom, which are aligned with the image axes #REFR .", "text_before_citation": ["Rotated Detection Task.", "Rotated object detection requires the model to predict minimum area rectangles with five degrees of freedom, namely rotated bounding boxes, enclosing objects of interests #OTHEREFR ."], "text_after_citation": ["The rotated boxes occupy a much smaller area when estimating the location status of diagonally positioned elongated objects as shown in Figure 1.", "Rotated detection is strictly superior to axis-aligned detection as there is less background within the box, and the orientation can potentially convey object pose information.", "However, there are only a handful of datasets with rotated annotations #OTHEREFR comparing to large number of readily available large-scale axis-aligned datasets #OTHEREFR .", "A potential contributor to this phenomenon is the ease of annotating axis-aligned boxes by simple click-and-drag with current labelling tools.", "Rotated boxes require much more effort to tightly enclose the objects with an extra degree of freedom."], "citing_paper_content": {"title": "Knowledge Combination To Learn Rotated Detection Without Rotated Annotation", "abstract": "Rotated bounding boxes drastically reduce output ambiguity of elongated objects, making it superior to axis-aligned bounding boxes. Despite the effectiveness, rotated detectors are not widely employed. Annotating rotated bounding boxes is such a laborious process that they are not provided in many detection datasets where axis-aligned annotations are used instead. In this paper, we propose a framework that allows the model to predict precise rotated boxes only requiring cheaper axis-aligned annotation of the target dataset 1. To achieve this, we leverage the fact that neural networks are capable of learning richer representation of the target domain than what is utilized by the task. The under-utilized representation can be exploited to address a more detailed task. Our framework combines task knowledge of an out-ofdomain source dataset with stronger annotation and domain knowledge of the target dataset with weaker annotation. A novel assignment process and projection loss are used to enable the co-training on the source and target datasets. As a result, the model is able to solve the more detailed task in the target domain, without additional computation overhead during inference. We extensively evaluate the method on various target datasets including fresh-produce dataset, HRSC2016 and SSDD. Results show that the proposed method consistently performs on par with the fully supervised approach."}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["axis-aligned object detection"], "citation_intent": "background"}
{"citing_id": "2303.07130v2", "cited_id": "2001.11767", "section_title": "Introduction", "citation": "The proposed method employs various image processing algorithms and a pre-trained UNET model #REFR to extract infected regions from CT scans.", "text_before_citation": ["A criterion on the volume of infection has been used as an indicator of the severity of infection #OTHEREFR and classified the CT scans into different classes such as healthy, mild, moderate, severe, and critical.", "Radiologists commonly use the CT severity score (CTSS) to determine the extent of severity in COVID-19 patients #OTHEREFR .", "The score ranges from 0 to 25 and is calculated based on the distribution and magnitude of abnormalities (lung involvement) seen on chest CT scans.", "A higher CTSS score indicates more severe lung damage, and this information is essential in making appropriate treatment decisions #OTHEREFR .", "In line with this approach, we have developed an automatic method that categorizes CT severity into four classes: mild, moderate, severe, and critical."], "text_after_citation": ["Then, an infection rate-based feature vector is proposed for each CT scan.", "Various classical machine-learning models and ensemble-based models learn these representative feature vectors to predict the severity classes for chest CT scans.", "A fully automatic severity analysis model can significantly reduce clinicians' time to assess a patient's condition.", "The paper is organized as follows: Section 2 explains the dataset used in the study.", "Section 3 presents the proposed method for segmenting relevant clinical features and classifying COVID-19 patients into different severity classes. Section 4 discusses results and inferences. Finally, Section 5 summarizes the present study."], "citing_paper_content": {"title": "Enhancing Covid-19 Severity Analysis Through Ensemble Methods", "abstract": "Computed Tomography (CT) scans provide a detailed image of the lungs, allowing clinicians to observe the extent of damage caused by COVID-19. The CT severity score (CTSS) based scoring method is used to identify the extent of lung involvement observed on a CT scan. This paper presents a domain knowledge-based pipeline for extracting regions of infection in COVID-19 patients using a combination of imageprocessing algorithms and a pre-trained UNET model. Then, an infection rate-based feature vector is proposed for each CT scan. The severity of the infection is then classified into different categories using an ensemble of three machine-learning models: Extreme Gradient Boosting, Extremely Randomized Trees, and Support Vector Machine. The proposed system was evaluated on a validation dataset in the AI-Enabled Medical Image Analysis Workshop and COVID-19 Diagnosis Competition (AI-MIA-COV19D) and achieved a macro F1 score of 64%. These results demonstrate the potential of combining domain knowledge with machine learning techniques for accurate COVID-19 diagnosis using CT scans. The implementation of the proposed system for severity analysis is available at https://github.com/aanandt/Enhancing-COVID-19-Severity-Analysis-through-Ensemble-Methods.git"}, "cited_paper_content": {"title": "Automatic Lung Segmentation In Routine Imaging Is A Data Diversity Problem, Not A Methodology Problem", "abstract": "Automated segmentation of anatomical structures is a crucial step in many medical image analysis tasks. For lung segmentation, a variety of approaches exist, involving sophisticated pipelines trained and validated on a range of different data sets. However, during translation to clinical routine the applicability of these approaches across diseases remains limited. Here, we show that the accuracy and reliability of lung segmentation algorithms on demanding cases primarily does not depend on methodology, but on the diversity of training data. We compare 4 generic deep learning approaches and 2 published lung segmentation algorithms on routine imaging data with more than 6 different disease patterns and 3 published data sets. We show that a basic approach - U-net - performs either better, or competitively with other approaches on both routine data and published data sets, and outperforms published approaches once trained on a diverse data set covering multiple diseases. Training data composition consistently has a bigger impact than algorithm choice on accuracy across test data sets. We carefully analyse the impact of data diversity, and the specifications of annotations on both training and validation sets to provide a reference for algorithms, training data, and annotation. Results on a seemingly well understood task of lung segmentation suggest the critical importance of training data diversity compared to model choice."}, "keywords": ["CT scans"], "citation_intent": "method"}
{"citing_id": "2303.13919v1", "cited_id": "1909.11355", "section_title": "Conclusions", "citation": "We found that EigenTrust algorithm can correctly estimate low trust scores for two types of simple attacks-namely, Models A and B-which is consistent with results in past studies #REFR .", "text_before_citation": ["This paper has examined the applicability of EigenTrust algorithm to the use case of C2C services using an agent-based model.", "The algorithm was developed to calculate the trust scores of nodes in P2P communication and does not take into account the actual situation of C2C services.", "Therefore, we established transaction protocols for C2C services, in which the buyer and seller rate each other.", "When analysing the results of the simulations, we focused on changes in the counterparty's trust score over time.", "We therefore contribute to the literature (1) by having implemented a simulator that clarifies transaction procedures in purchasing goods and performs transactions between users and (2) by analysing the trust scores resulting from EigenTrust algorithm to confirm the difference in the detection of attacks between P2P and C2C."], "text_after_citation": ["Therefore, the algorithm is expected to mitigate the impact of simple attackers on C2C services.", "Furthermore, we found oscillatory patterns in trust scores for two types of advanced attacks-namely, Models E and F-whereas no previous studies have focused on such dynamics.", "That result suggests that by detecting such oscillations, EigenTrust algorithm may be able to detect advanced attacks that were considered beyond its reach in P2P communications.", "Doing so was possible because we focused on the transient dynamics of trust scores, whereas past studies have focused on trust scores at a fixed point at the end state of transactions.", "Our results provide insights for the further development of Trust Management Algorithms for the benefit of consumer services."], "citing_paper_content": {"title": "Applicability Of Trust Management Algorithm In C2C Services", "abstract": "The emergence of Consumer-to-Consumer (C2C) platforms has allowed consumers to buy and sell goods directly, but it has also created problems, such as commodity fraud and fake reviews. Trust Management Algorithms (TMAs) are expected to be a countermeasure to detect fraudulent users. However, it is unknown whether TMAs are as effective as reported as they are designed for Peer-to-Peer (P2P) communications between devices on a network. Here we examine the applicability of 'EigenTrust', a representative TMA, for the use case of C2C services using an agentbased model. First, we defined the transaction process in C2C services, assumed six types of fraudulent transactions, and then analysed the dynamics of EigenTrust in C2C systems through simulations. We found that EigenTrust could correctly estimate low trust scores for two types of simple frauds. Furthermore, we found the oscillation of trust scores for two types of advanced frauds, which previous research did not address. This suggests that by detecting such oscillations, EigenTrust may be able to detect some (but not all) advanced frauds. Our study helps increase the trustworthiness of transactions in C2C services and provides insights into further technological development for consumer services."}, "cited_paper_content": {"title": "Decentralized Trust Management: Risk Analysis And Trust Aggregation", "abstract": "Decentralized trust management is used as a referral benchmark for assisting decision making by human or intelligence machines in open collaborative systems. During any given period of time, each participant may only interact with a few other participants. Simply relying on direct trust may frequently resort to random team formation. Thus, trust aggregation becomes critical. It can leverage decentralized trust management to learn about indirect trust of every participant based on past transaction experiences. This article presents alternative designs of decentralized trust management and their efficiency and robustness from three perspectives. First, we study the risk factors and adverse effects of six common threat models. Second, we review the representative trust aggregation models and trust metrics. Third, we present an in-depth analysis and comparison of these reference trust aggregation methods with respect to effectiveness and robustness. We show our comparative study results through formal analysis and experimental evaluation. This comprehensive study advances the understanding of adverse effects of present and future threats and the robustness of different trust metrics. It may also serve as a guideline for research and development of next-generation trust aggregation algorithms and services in the anticipation of risk factors and mischievous threats."}, "keywords": ["EigenTrust algorithm", "low trust scores"], "citation_intent": "result"}
{"citing_id": "2303.13842v2", "cited_id": "1612.01105", "section_title": "Segmentation", "citation": "The model utilizes a Pyramid Pooling Module (PPM) #REFR to achieve a larger receptive field and generate effective global prior representations.", "text_before_citation": ["Considering the hierarchical structure of the features obtained using the Swin Transformer backbone, more advanced techniques for dense prediction, such as Feature Pyramid Networks (FPN) #OTHEREFR and U-Net #OTHEREFR , can be employed for linking the hierarchical information.", "In our work, we incorporate UPerNet #OTHEREFR with our proposed PCA.", "As depicted in Fig.", "3 , UPerNet features a top-down architecture with lateral connections that facilitate the fusion of high-level semantic information with lower-level details."], "text_after_citation": ["UPerNet has the capability to learn visual attributes for semantic segmentation and image completion at multiple levels.", "{F 1 , F 2 , F 3 , F 4 } denotes the set of the resulting feature maps of each stage of the backbone.", "The corresponding downsampling rate for the {F 1 , F 2 , F 3 , F 4 } are {4,8,16,32}, respectively.", "In the decoder stage, the PPM is only used at the top of F 4 , where the resulting feature map can be denoted asF 4 .", "Then,F 4 is upsampled and sum with F 3 to progressively fuse multilevel features."], "citing_paper_content": {"title": "Fishdreamer: Towards Fisheye Semantic Completion Via Unified Image Outpainting And Segmentation", "abstract": "This paper raises the new task of Fisheye Semantic Completion (FSC), where dense texture, structure, and semantics of a fisheye image are inferred even beyond the sensor field-of-view (FoV). Fisheye cameras have larger FoV than ordinary pinhole cameras, yet its unique special imaging model naturally leads to a blind area at the edge of the image plane. This is suboptimal for safety-critical applications since important perception tasks, such as semantic segmentation, become very challenging within the blind zone. Previous works considered the out-FoV outpainting and in-FoV segmentation separately. However, we observe that these two tasks are actually closely coupled. To jointly estimate the tightly intertwined complete fisheye image and scene semantics, we introduce the new FishDreamer which relies on successful ViTs enhanced with a novel Polar-aware Cross Attention module (PCA) to leverage dense context and guide semantically-consistent content generation while considering different polar distributions. In addition to the contribution of the novel task and architecture, we also derive Cityscapes-BF and KITTI360-BF datasets to facilitate training and evaluation of this new track. Our experiments demonstrate that the proposed FishDreamer outperforms methods solving each task in isolation and surpasses alternative approaches on the Fisheye Semantic Completion. Code and datasets are publicly available at FishDreamer."}, "cited_paper_content": {"title": "Pyramid Scene Parsing Network", "abstract": "Scene parsing is challenging for unrestricted open vocabulary and diverse scenes. In this paper, we exploit the capability of global context information by different-region-based context aggregation through our pyramid pooling module together with the proposed pyramid scene parsing network (PSPNet). Our global prior representation is effective to produce good quality results on the scene parsing task, while PSPNet provides a superior framework for pixel-level prediction. The proposed approach achieves state-of-the-art performance on various datasets. It came first in ImageNet scene parsing challenge 2016, PASCAL VOC 2012 benchmark and Cityscapes benchmark. A single PSPNet yields the new record of mIoU accuracy 85.4% on PASCAL VOC 2012 and accuracy 80.2% on Cityscapes."}, "keywords": ["larger receptive field", "Pyramid Pooling Module"], "citation_intent": "method"}
{"citing_id": "2304.12035v1", "cited_id": "1706.03762", "section_title": "I. Introduction", "citation": "We investigate whether combining iterative reasoning and residual learning with CNNs and Transformers #REFR , as well as decoupled image-level and patch-level discriminators, can lead to a promising design for creating a more robust and data-efficient method to tackle the fewshot image inpainting task.", "text_before_citation": ["Moreover, existing inpainting methods are not designed specifically for few-shot learning and may not work well with limited training samples.", "Domain-related prior knowledge #OTHEREFR or lightweight generative models #OTHEREFR , #OTHEREFR may be employed in image inpainting to mitigate overfitting.", "However, they do not perform well when there is a large domain gap between two tasks or the reduced network capacity affects the inpainting quality.", "To this end, we propose a novel few-shot Generative Residual Image inpaintinG framework (GRIG), which enables highquality image inpainting on few-shot datasets.", "To effectively optimize inpainting results, we use iterative reasoning to solve more accurate and generalizable algorithmic reasoning tasks #OTHEREFR along with residual learning #OTHEREFR to incrementally refine previous estimates."], "text_after_citation": ["In this framework, the inpainting process is carried out in several forward passes by feeding the generator with the output of the previous iteration and a corresponding mask.", "This enables the generator to be welltrained and focused on the residual information between the previously predicted output and the ground truth, resulting in effective image inpainting with high quality and good generality on few-shot data.", "We implemented our framework into three components: a generator, a projected discriminator, and a forgery-patch discriminator.", "The generator uses CNNs to extract shallow features of edges and textures and the Transformer blocks to capture global interactions between feature contexts at each iterative step.", "To accelerate network convergence and reduce overfitting, we decouple the image distribution learning by Fig. 1 ."], "citing_paper_content": {"title": "Grig: Few-Shot Generative Residual Image Inpainting", "abstract": "Image inpainting is the task of filling in missing or masked region of an image with semantically meaningful contents. Recent methods have shown significant improvement in dealing with large-scale missing regions. However, these methods usually require large training datasets to achieve satisfactory results and there has been limited research into training these models on a small number of samples. To address this, we present a novel few-shot generative residual image inpainting method that produces high-quality inpainting results. The core idea is to propose an iterative residual reasoning method that incorporates Convolutional Neural Networks (CNNs) for feature extraction and Transformers for global reasoning within generative adversarial networks, along with image-level and patch-level discriminators. We also propose a novel forgerypatch adversarial training strategy to create faithful textures and detailed appearances. Extensive evaluations show that our method outperforms previous methods on the few-shot image inpainting task, both quantitatively and qualitatively."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["CNNs"], "citation_intent": "method"}
{"citing_id": "2304.03199v1", "cited_id": "2003.00196", "section_title": "Comparison With State-Of-The-Art Methods", "citation": "We conduct quantitative and qualitative comparisons of the same-identity reconstruction task on the VoxCeleb dataset, where FOMM #REFR is used to generate the coarse face animation results.", "text_before_citation": ["Methods.", "We compare our FADM with five state-of-theart methods: FOMM #OTHEREFR , Face vid2vid #OTHEREFR , Bilayer #OTHEREFR , DaGAN #OTHEREFR , and PIRenderer #OTHEREFR .", "For Bilayer, FOMM, DaGAN, and PIRenderer, we use their official pre-trained models for evaluation, while for Face vid2vid, we adopt a widely recognized unofficial model due to the absence of the official code. All of these models are pre-trained on VoxCeleb.", "Same-Identity Reconstruction."], "text_after_citation": ["In or-der to accelerate the inference, we randomly select several short-time videos from the testing dataset of VoxCeleb for quantitative comparison, and the selected video list is available in the supplementary materials.", "As shown in Table 1 , compared to other SOTA methods, it is evident that FADM achieves the best performance in all metrics, especially the reconstruction faithfulness L 1 and the visual quality LPIPS, demonstrating the superiority of our FADM in generating high-fidelity face animation.", "Moreover, we show the visual results of FADM and existing SOTA methods in Fig. 4 .", "FADM exhibits overall better quality with more fine details than other methods. .", "On both testing datasets, our FADM can produce more identity-preserving and photo-realistic results compared with other methods."], "citing_paper_content": {"title": "Face Animation With An Attribute-Guided Diffusion Model", "abstract": "Figure 1. Comparison with SOTA face animation methods on reconstruction and reenactment tasks. Left: FOMM [42] often produces artifacts and over-smoothing textures, while FADM is capable of rectifying the distortions and enriching the fine-grained details. Right: Face vid2vid [49] suffers from typical unnatural facial details, while FADM prominently improves the overall visual quality."}, "cited_paper_content": {"title": "First Order Motion Model For Image Animation", "abstract": "Image animation consists of generating a video sequence so that an object in a source image is animated according to the motion of a driving video. Our framework addresses this problem without using any annotation or prior information about the specific object to animate. Once trained on a set of videos depicting objects of the same category (\\eg \\ faces, human bodies), our method can be applied to any object of this class. To achieve this, we decouple appearance and motion information using a self-supervised formulation. To support complex motions, we use a representation consisting of a set of learned keypoints along with their local affine transformations. A generator network models occlusions arising during target motions and combines the appearance extracted from the source image and the motion derived from the driving video. Our framework scores best on diverse benchmarks and on a variety of object categories. Our source code is publicly available."}, "keywords": ["coarse face animation"], "citation_intent": "method"}
{"citing_id": "2305.03045v1", "cited_id": "1904.08755", "section_title": "Network Details", "citation": "The resulting OctFormer has a similar amount of trainable parameters (39M) as MinkowskiNet (38M) #REFR .", "text_before_citation": ["The dilations of octree attention in two consecutive OctFormer blocks are set to 1 and 4, respectively.", "Downsampling.", "The downsampling module is implemented as an octree convolution with a kernel size of 2 and a stride of 2, followed by a Batch Normalization layer, which reduces the spatial resolution and increases the channel of a feature map by a factor of 2.", "Network Settings.", "By default, is set to 96, the block numbers are set to {2, 2, 18, 2}, the head numbers of octree attentions are set to 1/16 of the channel numbers, and the point numbers for window partition are set to 32."], "text_after_citation": ["The output of OctFormer is hierarchical features with four spatial resolutions, i.e., { /4, /8, /16, /32}, which can be conveniently integrated with a feature pyramid network (FPN) [Lin et al. 2017] for segmentation and detection.", "And the last feature map can be averaged as the global feature for shape classification."], "citing_paper_content": {"title": "", "abstract": ""}, "cited_paper_content": {"title": "4D Spatio-Temporal Convnets: Minkowski Convolutional Neural Networks", "abstract": "In many robotics and VR/AR applications, 3D-videos are readily-available input sources (a sequence of depth images, or LIDAR scans). However, in many cases, the 3D-videos are processed frame-by-frame either through 2D convnets or 3D perception algorithms. In this work, we propose 4-dimensional convolutional neural networks for spatio-temporal perception that can directly process such 3D-videos using high-dimensional convolutions. For this, we adopt sparse tensors and propose generalized sparse convolutions that encompass all discrete convolutions. To implement the generalized sparse convolution, we create an open-source auto-differentiation library for sparse tensors that provides extensive functions for high-dimensional convolutional neural networks. We create 4D spatio-temporal convolutional neural networks using the library and validate them on various 3D semantic segmentation benchmarks and proposed 4D datasets for 3D-video perception. To overcome challenges in 4D space, we propose the hybrid kernel, a special case of the generalized sparse convolution, and trilateral-stationary conditional random fields that enforce spatio-temporal consistency in the 7D space-time-chroma space. Experimentally, we show that a convolutional neural network with only generalized 3D sparse convolutions can outperform 2D or 2D-3D hybrid methods by a large margin. Also, we show that on 3D-videos, 4D spatio-temporal convolutional neural networks are robust to noise and outperform the 3D convolutional neural network."}, "keywords": ["resulting OctFormer", "MinkowskiNet"], "citation_intent": "method"}
{"citing_id": "2304.04972v1", "cited_id": "2002.06440", "section_title": "Related Works", "citation": "PFNM and FedMA #REFR apply the Bayesian non-parametric mechanism to study the permutation invariance of the neural network, and match the neurons of client neural networks to the global neurons.", "text_before_citation": ["Then, the server collects the updated models from each client and aggregates them to obtain a new global model by averaging the model weights.", "Unlike the privacy and communication challenges, the statistical heterogeneity challenge is a unique and popular issue in the federated learning paradigm #OTHEREFR .", "According to the two stages of federated learning mentioned above, the contributions of these studies can be roughly divided into local update improvements and aggregation improvements.", "Our work is an improvement in local update phase, so it can be combined with existing aggregation improvements without any conflict.", "As for the aggregation improvements on non-IID data, there are a series of related studies."], "text_after_citation": ["Moreover, methods such as adaptive weights #OTHEREFR , attention mechanisms #OTHEREFR , and normalization #OTHEREFR are also used to improve the aggregation effect for statistical heterogeneity.", "There are also many studies to alleviate the negative effects of statistical heterogeneity in the local update phase. #OTHEREFR", "(2020) propose the FedProx algorithm, which adds a regular term to the loss function of the client.", "This proximal term uses the 2 norm to explicitly constrain the local model to be close to the latest global model, limiting the local update of non-IID clients to be not too far apart.", "However, this explicit constraint could inhibit FedProx from quickly finding a better model in the early stage."], "citing_paper_content": {"title": "Federated Learning With Classifier Shift For Class Imbalance", "abstract": "Federated learning aims to learn a global model collaboratively while the training data belongs to different clients and is not allowed to be exchanged. However, the statistical heterogeneity challenge on non-IID data, such as class imbalance in classification, will cause client drift and significantly reduce the performance of the global model. This paper proposes a simple and effective approach named FedShift which adds the shift on the classifier output during the local training phase to alleviate the negative impact of class imbalance. We theoretically prove that the classifier shift in FedShift can make the local optimum consistent with the global optimum and ensure the convergence of the algorithm. Moreover, our experiments indicate that FedShift significantly outperforms the other state-of-the-art federated learning approaches on various datasets regarding accuracy and communication efficiency. * Use footnote for providing further information about author (webpage, alternative address)-not for acknowledging funding agencies. Preprint. Under review."}, "cited_paper_content": {"title": "Federated Learning With Matched Averaging", "abstract": "Federated learning allows edge devices to collaboratively learn a shared model while keeping the training data on device, decoupling the ability to do model training from the need to store the data in the cloud. We propose Federated matched averaging (FedMA) algorithm designed for federated learning of modern neural network architectures e.g. convolutional neural networks (CNNs) and LSTMs. FedMA constructs the shared global model in a layer-wise manner by matching and averaging hidden elements (i.e. channels for convolution layers; hidden states for LSTM; neurons for fully connected layers) with similar feature extraction signatures. Our experiments indicate that FedMA outperforms popular state-of-the-art federated learning algorithms on deep CNN and LSTM architectures trained on real world datasets, while improving the communication efficiency."}, "keywords": ["client neural networks"], "citation_intent": "method"}
{"citing_id": "2304.10739v1", "cited_id": "1805.08154", "section_title": "Related Works", "citation": "However, likelihood-based regression strategies have been proven to deal with numbers having high variance in their magnitudes #REFR .", "text_before_citation": ["Numerical Reasoning in Language Modeling While PLM has been widely used in general natural language processing tasks #OTHEREFR , handling numerals has recently started to gain focus in the NLP domain #OTHEREFR .", "Various tasks including probing numerical knowledge #OTHEREFR and probing numerical embedding #OTHEREFR showed the possibility of PLMs achieving numeracy.", "While PLMs have been known to learn the scalar magnitude of objects #OTHEREFR , some domain-specific datasets may pose challenges to its learning on regression tasks due to their distribution characteristics.", "Previous work tried a traditional linear regression architecture with simple scalar value estimation #OTHEREFR ."], "text_after_citation": ["Recently, DExp method has been introduced and reported to achieve state-of-the-art performance compared to other likelihood functions-based regression approaches such as Gaussian Mixture Model, Transformed Laplace, and LogLaplace #OTHEREFR .", "For our KitchenScale model, we adopted the DExp approach in our ingredient quantity regression task as it has shown success in previous works featuring numeral values with highly varying scales #OTHEREFR ."], "citing_paper_content": {"title": "Kitchenscale: Learning To Predict Ingredient Quantities From Recipe Contexts", "abstract": "Determining proper quantities for ingredients is an essential part of cooking practice from the perspective of enriching tastiness and promoting healthiness. We introduce KitchenScale, a fine-tuned Pre-trained Language Model (PLM) that predicts a target ingredient's quantity and measurement unit given its recipe context. To effectively train our KitchenScale model, we formulate an ingredient quantity prediction task that consists of three sub-tasks which are ingredient measurement type classification, unit classification, and quantity regression task. Furthermore, we utilized transfer learning of cooking knowledge from recipe texts to PLMs. We adopted the Discrete Latent Exponent (DExp) method to cope with high variance of numerical scales in recipe corpora. Experiments with our newly constructed dataset and recommendation examples demonstrate KitchenScale's understanding of various recipe contexts and generalizability in predicting ingredient quantities. We implemented a web application for KitchenScale to demonstrate its functionality in recommending ingredient quantities expressed in numerals (e.g., 2) with units (e.g., ounce)."}, "cited_paper_content": {"title": "Numeracy For Language Models: Evaluating And Improving Their Ability To Predict Numbers", "abstract": "Numeracy is the ability to understand and work with numbers. It is a necessary skill for composing and understanding documents in clinical, scientific, and other technical domains. In this paper, we explore different strategies for modelling numerals with language models, such as memorisation and digit-by-digit composition, and propose a novel neural architecture that uses a continuous probability density function to model numerals from an open vocabulary. Our evaluation on clinical and scientific datasets shows that using hierarchical models to distinguish numerals from words improves a perplexity metric on the subset of numerals by 2 and 4 orders of magnitude, respectively, over non-hierarchical models. A combination of strategies can further improve perplexity. Our continuous probability density function model reduces mean absolute percentage errors by 18% and 54% in comparison to the second best strategy for each dataset, respectively."}, "keywords": ["numbers"], "citation_intent": "background"}
{"citing_id": "2304.04923v1", "cited_id": "1904.04600", "section_title": "A. Related Works", "citation": "Other works use centroidal dynamics to inform a fullorder dynamics model solve with CIO #REFR .", "text_before_citation": ["A different approach is to increase the accuracy of the solution, which allows for a coarser discretization of the trajectory.", "This has been achieved through variational integrators #OTHEREFR and higher-order collocation methods #OTHEREFR , although they still remain quite slow and can struggle with convergence.", "Some researchers have blended these approaches by combining solutions of relaxed and exact problems.", "One example is to first fix the complementarity constraint relaxation to a large value and then iteratively reduce it #OTHEREFR .", "Another approach first solves the problem with first-order collocation and a large relaxation and then use this solution to warm start a problem with third-order collocation and a small relaxation #OTHEREFR ."], "text_after_citation": ["These methods ensure that the final result will meet the physical constraints of the problem and result in overall decreases in solve times, but still struggle with convergence due to the full-order complexity of the dynamics in combination with the complementarity constraints.", "More recently, MIP was applied to an iterative convex approximation of centroidal dynamics and full kinematics, achieving impressive convergence rates #OTHEREFR ."], "citing_paper_content": {"title": "Staged Contact Optimization: Combining Contact-Implicit And Multi-Phase Hybrid Trajectory Optimization", "abstract": "Trajectory optimization problems for legged robots are commonly formulated with fixed contact schedules. These multi-phase Hybrid Trajectory Optimization (HTO) methods result in locally optimal trajectories, but the result depends heavily upon the predefined contact mode sequence. Contact-Implicit Optimization (CIO) offers a potential solution to this issue by allowing the contact mode to be determined throughout the trajectory by the optimization solver. However, CIO suffers from long solve times and convergence issues. This work combines the benefits of these two methods into one algorithm: Staged Contact Optimization (SCO). SCO tightens constraints on contact in stages, eventually fixing them to allow robust and fast convergence to a feasible solution. Results on a planar biped and spatial quadruped demonstrate speed and optimality improvements over CIO and HTO. These properties make SCO well suited for offline trajectory generation or as an effective tool for exploring the dynamic capabilities of a robot."}, "cited_paper_content": {"title": "Hierarchical Planning Of Dynamic Movements Without Scheduled Contact Sequences", "abstract": "Most animal and human locomotion behaviors for solving complex tasks involve dynamic motions and rich contact interaction. In fact, complex maneuvers need to consider dynamic movement and contact events at the same time. We present a hierarchical trajectory optimization approach for planning dynamic movements with unscheduled contact sequences. We compute whole-body motions that achieve goals that cannot be reached in a kinematic fashion. First, we find a feasible CoM motion according to the centroidal dynamics of the robot. Then, we refine the solution by applying the robot's full-dynamics model, where the feasible CoM trajectory is used as a warm-start point. To accomplish the unscheduled contact behavior, we use complementarity constraints to describe the contact model, i.e. environment geometry and non-sliding active contacts. Both optimization phases are posed as Mathematical Program with Complementarity Constraints (MPCC). Experimental trials demonstrate the performance of our planning approach in a set of challenging tasks."}, "keywords": ["centroidal dynamics"], "citation_intent": "method"}
{"citing_id": "2303.01628v1", "cited_id": "1912.03572", "section_title": "I. Introduction", "citation": "A disadvantage of our approach is that it requires designing controllers per time step as in #REFR . This is the trade-off between performance and engineering effort.", "text_before_citation": ["The feedback gain is obtained by solving nonlinear optimization programs.", "In the experiment section, we compare our method with other probabilistic control methods that could track the nominal trajectory, and we show that our approach outperforms other methods.", "Apart from better performance, our approach has the following advantages.", "First, the approach applies to a general class of stochastic nonlinear systems, more precisely, systems with mixed-trigonometricpolynomial dynamics, which covers most robotic systems, and the approach applies to arbitrary probabilistic uncertainties, not necessarily Gaussian distributions.", "Second, our approach leaves room for nonlinear state feedback control, not necessarily linear state feedback control."], "text_after_citation": ["Note that we assume the open-loop nominal trajectory is given and we design feedback controllers to track the nominal trajectory.", "There are many trajectory optimization methods that could compute the nominal trajectory #OTHEREFR , and how the nominal trajectory is obtained is not the focus of this paper.", "This assumption is different from most nonlinear covariance steering literature #OTHEREFR - #OTHEREFR , especially methods similar to DDP, where they design a nominal trajectory and feedback controllers together.", "We believe that separately designing the nominal trajectory and feedback controllers can be beneficial in many cases.", "For instance, the trajectory optimization alone could incorporate more constraints, such as obstacle avoidance, chance constraints, tube safety, and so on, while it is not easy to incorporate more constraints into an algorithm that simultaneously generate nominal trajectories and feedback controllers, especially when dealing with stochastic nonlinear systems."], "citing_paper_content": {"title": "Non-Gaussian Uncertainty Minimization Based Control Of Stochastic Nonlinear Robotic Systems", "abstract": "In this paper, we consider the closed-loop control problem of nonlinear robotic systems in the presence of probabilistic uncertainties and disturbances. More precisely, we design a state feedback controller that minimizes deviations of the states of the system from the nominal state trajectories due to uncertainties and disturbances. Existing approaches to address the control problem of probabilistic systems are limited to particular classes of uncertainties and systems such as Gaussian uncertainties and processes and linearized systems. We present an approach that deals with nonlinear dynamics models and arbitrary known probabilistic uncertainties. We formulate the controller design problem as an optimization problem in terms of statistics of the probability distributions including moments and characteristic functions. In particular, in the provided optimization problem, we use moments and characteristic functions to propagate uncertainties throughout the nonlinear motion model of robotic systems. In order to reduce the tracking deviations, we minimize the uncertainty of the probabilistic states around the nominal trajectory by minimizing the trace and the determinant of the covariance matrix of the probabilistic states. To obtain the state feedback gains, we solve deterministic optimization problems in terms of moments, characteristic functions, and state feedback gains using off-the-shelf interior-point optimization solvers. To illustrate the performance of the proposed method, we compare our method with existing probabilistic control methods."}, "cited_paper_content": {"title": "Sequential Chance Optimization For Flow-Tube Based Control Of Probabilistic Nonlinear Systems", "abstract": "In this paper, we address the problem of closed-loop control of nonlinear dynamical systems subjected to probabilistic uncertainties. More precisely, we design time-varying polynomial feedback controllers to follow the given nominal trajectory and also, for safety purposes, remain in the tube around the nominal trajectory, despite all uncertainties. We formulate this problem as a chance optimization problem where we maximize the probability of achieving control objectives. To address control problems with long planning horizons, we formulate the single large chance optimization problem as a sequence of smaller chance optimization problems. To solve the obtained chance optimization problems, we leverage the theory of measures and moments and obtain convex relaxations in the form of semidefinite programs. We provide numerical examples on stabilizing controller design and motion planning of uncertain nonlinear systems to illustrate the performance of the proposed approach."}, "keywords": ["designing controllers"], "citation_intent": "method"}
{"citing_id": "2303.02244v1", "cited_id": "1905.13062", "section_title": "B. Survey Measures", "citation": "In line with this, recent research specifically on the personality of developers has not found any evidence of variation over time #REFR .", "text_before_citation": ["Given the low number of items per construct, these constitute acceptable values #OTHEREFR .", "We measured narcissism using the short version of the Narcissistic Personality Inventory (NPI-16) #OTHEREFR .", "Cronbach's \u03b1 was .70, constituting an acceptable value #OTHEREFR and even being slightly higher than that obtained in the work that developed the measure in the first place #OTHEREFR .", "As we measure personality after the analyzed commits were made, this might potentially raise concerns about an unclear direction of causality in our analysis.", "However, this is unlikely to be a problem as personality is widely considered as stable across adult life #OTHEREFR , #OTHEREFR and the sampled developers were all between 24 and 65 years of age."], "text_after_citation": ["To measure developers' age at the time of each commit, we asked developers to simply provide their age in years.", "We then subtracted the difference between 2022 and the year in which the focal commit was made from the provided age."], "citing_paper_content": {"title": "The Type To Take Out A Loan? A Study Of Developer Personality And Technical Debt", "abstract": "Background: Technical debt (TD) has been widely discussed in software engineering research, and there is an emerging literature linking it to developer characteristics. However, developer personality has not yet been studied in this context. Aims and Method: We explore the relationship between various personality traits (Five Factor Model, regulatory focus, and narcissism) of developers and the introduction and removal of TD. To this end, we complement an existing TD dataset with novel selfreport personality data gathered by surveying developers, and analyze 2,145 commits from 19 developers. Results: We find that conscientiousness, emotional stability, openness to experience, and prevention focus are negatively associated with TD. There were no significant results for extraversion, agreeableness, promotion focus, or narcissism. Conclusions: We take our results as first evidence that developer personality has a systematic influence on the introduction and removal of TD. This has implications not only for future research, which could, for example, study the effects of personality on downstream consequences of TD like defects, but also for software engineering practitioners who may, for example, consider developer personality in staffing decisions."}, "cited_paper_content": {"title": "A Large-Scale, In-Depth Analysis Of Developers' Personalities In The Apache Ecosystem", "abstract": "Context Large-scale distributed projects are typically the results of collective efforts performed by multiple developers with heterogeneous personalities. Objective We aim to find evidence that personalities can explain developers\u2019 behavior in large scale-distributed projects. For example, the propensity to trust others \u2014 a critical factor for the success of global software engineering \u2014 has been found to influence positively the result of code reviews in distributed projects. Method In this paper, we perform a quantitative analysis of ecosystem-level data from the code commits and email messages contributed by the developers working on the Apache Software Foundation (ASF) projects, as representative of large scale-distributed projects. Results We find that there are three common types of personality profiles among Apache developers, characterized in particular by their level of Agreeableness and Neuroticism. We also confirm that developers\u2019 personality is stable over time. Moreover, personality traits do not vary with their role, membership, and extent of contribution to the projects. We also find evidence that more open developers are more likely to make contributors to Apache projects. Conclusion Overall, our findings reinforce the need for future studies on human factors in software engineering to use psychometric tools to control for differences in developers\u2019 personalities."}, "keywords": ["developers"], "citation_intent": "result"}
{"citing_id": "2303.11184v1", "cited_id": "1909.01362", "section_title": "How Early Is The Warning?", "citation": "These figures generally match with those in previous research (Kementchedjhieva and S\u00f8gaard 2021), #REFR have discovered.", "text_before_citation": ["To investigate how early the warning is triggered for a conversation, we perform an analysis on the two datasets.", "Figure 5 shows the distribution of the percentage of utterances that have elapsed when the Hierarchical-Multi model makes a prediction of derailment for all the positive samples in the testing set.", "We observe that around 80% of warnings are issued when fewer than five utterances have been seen by the model and the average length of the derailing conversations in the testing set is 7.2.", "We also show the average number of utterances between the issuance of a warning and the derailment utterance for four models.", "The distance is around 3 for the WTP dataset and 4 for the CMV dataset."], "text_after_citation": [], "citing_paper_content": {"title": "Conversation Modeling To Predict Derailment", "abstract": "Conversations among online users sometimes derail, i.e., break down into personal attacks. Such derailment has a negative impact on the healthy growth of cyberspace communities. The ability to predict whether ongoing conversations are likely to derail could provide valuable real-time insight to interlocutors and moderators. Prior approaches predict conversation derailment retrospectively without the ability to forestall the derailment proactively. Some works attempt to make dynamic prediction as the conversation develops, but fail to incorporate multisource information, such as conversation structure and distance to derailment. We propose a hierarchical transformer-based framework that combines utterance-level and conversation-level information to capture fine-grained contextual semantics. We propose a domain-adaptive pretraining objective to integrate conversational structure information and a multitask learning scheme to leverage the distance from each utterance to derailment. An evaluation of our framework on two conversation derailment datasets yields improvement over F1 score for the prediction of derailment. These results demonstrate the effectiveness of incorporating multisource information."}, "cited_paper_content": {"title": "Trouble On The Horizon: Forecasting The Derailment Of Online Conversations As They Develop", "abstract": "Online discussions often derail into toxic exchanges between participants. Recent efforts mostly focused on detecting antisocial behavior after the fact, by analyzing single comments in isolation. To provide more timely notice to human moderators, a system needs to preemptively detect that a conversation is heading towards derailment before it actually turns toxic. This means modeling derailment as an emerging property of a conversation rather than as an isolated utterance-level event. ::: Forecasting emerging conversational properties, however, poses several inherent modeling challenges. First, since conversations are dynamic, a forecasting model needs to capture the flow of the discussion, rather than properties of individual comments. Second, real conversations have an unknown horizon: they can end or derail at any time; thus a practical forecasting model needs to assess the risk in an online fashion, as the conversation develops. In this work we introduce a conversational forecasting model that learns an unsupervised representation of conversational dynamics and exploits it to predict future derailment as the conversation develops. By applying this model to two new diverse datasets of online conversations with labels for antisocial events, we show that it outperforms state-of-the-art systems at forecasting derailment."}, "keywords": ["Kementchedjhieva"], "citation_intent": "result"}
{"citing_id": "2303.10916v1", "cited_id": "1709.01507", "section_title": "B. Squeeze-And-Excitation Networks Attention Mechanism", "citation": "The SE attention mechanism encodes the entire spatial feature in a channel as a global feature through automatic learning utilizing global averaging pooling by means of the squeeze operation #REFR .", "text_before_citation": [], "text_after_citation": ["EQUATION", "where z c is the output associated with the c channel and the input x is a convolutional layer with a fixed kernel size.", "An excitation operation is then performed to obtain the importance of each channel in the feature map.", "s = F ex (z, W ) = \u03c3 (g (z, W )) = \u03c3 (W 2 \u03b4 (W 1 z)) (4)", "\u03c3 is the sigmoid activation function and \u03b4 is the relu activation function."], "citing_paper_content": {"title": "Learning Behavior Recognition In Smart Classroom With Multiple Students Based On Yolov5", "abstract": "Deep learning-based computer vision technology has grown stronger in recent years, and cross-fertilization using computer vision technology has been a popular direction in recent years. The use of computer vision technology to identify students' learning behavior in the classroom can reduce the workload of traditional teachers in supervising students in the classroom, and ensure greater accuracy and comprehensiveness. However, existing student learning behavior detection systems are unable to track and detect multiple targets precisely, and the accuracy of learning behavior recognition is not high enough to meet the existing needs for the accurate recognition of student behavior in the classroom. To solve this problem, we propose a YOLOv5s network structure based on you only look once (YOLO) algorithm to recognize and analyze students' classroom behavior in this paper. Firstly, the input images taken in the smart classroom are pre-processed. Then, the pre-processed image is fed into the designed YOLOv5 networks to extract deep features through convolutional layers, and the Squeeze-and-Excitation (SE) attention detection mechanism is applied to reduce the weight of background information in the recognition process. Finally, the extracted features are classified by the Feature Pyramid Networks (FPN) and Path Aggregation Network (PAN) structures. Multiple groups of experiments were performed to compare with traditional learning behavior recognition methods to validate the effectiveness of the proposed method. When compared with YOLOv4, the proposed method is able to improve the mAP performance by 11%."}, "cited_paper_content": {"title": "Squeeze-And-Excitation Networks", "abstract": "Convolutional neural networks are built upon the convolution operation, which extracts informative features by fusing spatial and channel-wise information together within local receptive fields. In order to boost the representational power of a network, much existing work has shown the benefits of enhancing spatial encoding. In this work, we focus on channels and propose a novel architectural unit, which we term the\"Squeeze-and-Excitation\"(SE) block, that adaptively recalibrates channel-wise feature responses by explicitly modelling interdependencies between channels. We demonstrate that by stacking these blocks together, we can construct SENet architectures that generalise extremely well across challenging datasets. Crucially, we find that SE blocks produce significant performance improvements for existing state-of-the-art deep architectures at slight computational cost. SENets formed the foundation of our ILSVRC 2017 classification submission which won first place and significantly reduced the top-5 error to 2.251%, achieving a 25% relative improvement over the winning entry of 2016."}, "keywords": ["automatic learning", "entire spatial feature"], "citation_intent": "method"}
{"citing_id": "2303.14977v1", "cited_id": "1902.09630", "section_title": "Related Work 2.1 Object Detetion", "citation": "YOLOv4 achieves superior performance with these improvements and modifications.YOLOv5 adds data enhancement and FOCUs modules to the v3 backbone and replaces IOU with GIOU #REFR and NMS with DiouNMS respectively.", "text_before_citation": ["Consequently, the one-stage detectors have more efficient detection and require less computer performance.", "OverFeat #OTHEREFR was the first application of a CNN-based single-stage target detector.", "YOLO #OTHEREFR and SSD #OTHEREFR were introduced after it, and SSD achieved great performance.YOLO and YOLOv2 #OTHEREFR only use the last output feature of the backbone.", "They can both achieve a very fast detection speed, but at the expense of accuracy.", "YOLOv3 #OTHEREFR is based on YoLov1, with an FPN architecture and a backbone that was changed from DarkNet-19 to DarkNet-53."], "text_after_citation": ["Our improved network structure will be compared with YOLOv5 with backbone network of CSPDarkNet53 #OTHEREFR .", "The FPN works as the basis for a detector model that is also used in a variety of popular models, such as RetinaNet #OTHEREFR , FCOS #OTHEREFR , and their variants."], "citing_paper_content": {"title": "A Novel Multi To Single Module For Small Object Detection", "abstract": "Small object detection presents a significant challenge in computer vision and object detection. The performance of small object detectors is often compromised by a lack of pixels and less significant features. This issue stems from information misalignment caused by variations in feature scale and information loss during feature processing. In response to this challenge, this paper proposes a novel the Multi to Single Module (M2S), which enhances a specific layer through improving feature extraction and refining features. Specifically, M2S includes the proposed Cross-scale Aggregation Module (CAM) and explored Dual Relationship Module (DRM) to improve information extraction capabilities and feature refinement effects. Moreover, this paper enhances the accuracy of small object detection by utilizing M2S to generate an additional detection head. The effectiveness of the proposed method is evaluated on two datasets, VisDrone2021-DET and SeaDronesSeeV2. The experimental results demonstrate its improved performance compared with existing methods. Compared to the baseline model (YOLOv5s), M2S improves the accuracy by about 1.1% on the VisDrone2021-DET testing dataset and 15.68% on the SeaDronesSeeV2 validation set. 1 Introduction Small object detection ([1],[2],[3],[4],[5]) has long been a challenge in object detection, which aims to accurately detect small objects (those under 32 pixels by 32 pixels) with very few visual features in the image. A detection model consisting of a backbone network, an encoder, and a decoder is proposed in You Only Look One-level Feature (YOLOF) [6]. It is proposed in YOLOF that by choosing the right scale features for output to a particular level with multiple-input and single-output, comparable performance to multiple-input and multiple-output can be obtained. In the small object detection task, this paper proposes a multiple-input to the single-output module that outputs with a low-level feature with high resolution. It provides the detector with more accurate and richer information, thereby improving its performance. Feature Pyramid network (FPN) [7] success is due to the idea of divide and conquer. Divide and conquer ensures that image features are available in different layers, whereas if only a unidirectional flow of information is delivered, then FPN ends up with a limitation at each layer. The result is that each layer of character can only be thinking within its own perspective. Thus, previous work has been devoted to increasing the interaction of different levels of feature and allowing more global thinking at each level. FPN appears as the backbone network of excellent detectors SSD [8], YOLO [9], and RCNN [10]. EffientDet [11]uses a cross-edge connecting BiFPN and repeats it to obtain better accuracy and efficiency. Although the previous works have shown gratifying performance, object detection still has a problem: the poor detection of small-scale instance objects. Based on this problem, this paper proposes a module to enhance communication at different network levels and improve feature extraction efficiency. The Cross-scale Fusion Module"}, "cited_paper_content": {"title": "Generalized Intersection Over Union: A Metric And A Loss For Bounding Box Regression", "abstract": "Intersection over Union (IoU) is the most popular evaluation metric used in the object detection benchmarks. However, there is a gap between optimizing the commonly used distance losses for regressing the parameters of a bounding box and maximizing this metric value. The optimal objective for a metric is the metric itself. In the case of axis-aligned 2D bounding boxes, it can be shown that IoU can be directly used as a regression loss. However, IoU has a plateau making it infeasible to optimize in the case of non-overlapping bounding boxes. In this paper, we address the this weakness by introducing a generalized version of IoU as both a new loss and a new metric. By incorporating this generalized IoU ( GIoU) as a loss into the state-of-the art object detection frameworks, we show a consistent improvement on their performance using both the standard, IoU based, and new, GIoU based, performance measures on popular object detection benchmarks such as PASCAL VOC and MS COCO."}, "keywords": ["v3 backbone", "YOLOv4"], "citation_intent": "method"}
{"citing_id": "2303.15441v1", "cited_id": "1411.7766", "section_title": "Introduction", "citation": "In addition, we are not tied to a particular annotation from datasets (e.g., specific attributes in CelebA #REFR ).", "text_before_citation": ["Consider a user interested in evaluating which factors contribute to the lack of robustness in a cat/dog classifier (target model).", "By selecting a list of keyword attributes, the user is able to (1) see counterfactual images where semantic variations flip the target model predictions (see the classifier score in the top-right corner of the counterfactual images) and #OTHEREFR quantify the sensitivity of each attribute for the target model (see sensitivity histogram on the top).", "Instead of using a test set, we propose using a StyleGAN generator as the picture engine for sampling counterfactual images.", "CLIP transforms user's text input, and enables model diagnosis in an open-vocabulary setting.", "This is a major advantage since there is no need for collecting and annotating images and minimal user expert knowledge."], "text_after_citation": ["To summarize, our proposed work offers three major improvements over earlier efforts:", "\u2022 The user requires neither a labeled, balanced test dataset, and minimal expert knowledge in order to evaluate where a model fails (i.e., model diagnosis).", "In addition, the method provides a sensitivity histogram across the attributes of interest.", "\u2022 When a different target model or a new user-defined attribute space is introduced, it is not necessary to retrain our system, allowing for practical use.", "\u2022 The target model fine-tuned with counterfactual images not only slightly improves the classification performance, but also greatly increases the distributional robustness against counterfactual images."], "citing_paper_content": {"title": "Zero-Shot Model Diagnosis", "abstract": "When it comes to deploying deep vision models, the behavior of these systems must be explicable to ensure confidence in their reliability and fairness. A common approach to evaluate deep learning models is to build a labeled test set with attributes of interest and assess how well it performs. However, creating a balanced test set (i.e., one that is uniformly sampled over all the important traits) is often time-consuming, expensive, and prone to mistakes. The question we try to address is: can we evaluate the sensitivity of deep learning models to arbitrary visual attributes without an annotated test set? This paper argues the case that Zero-shot Model Diagnosis (ZOOM) is possible without the need for a test set nor labeling. To avoid the need for test sets, our system relies on a generative model and CLIP. The key idea is enabling the user to select a set of prompts (relevant to the problem) and our system will automatically search for semantic counterfactual images (i.e., synthesized images that flip the prediction in the case of a binary classifier) using the generative model. We evaluate several visual tasks (classification, key-point detection, and segmentation) in multiple visual domains to demonstrate the viability of our methodology. Extensive experiments demonstrate that our method is capable of producing counterfactual images and offering sensitivity analysis for model diagnosis without the need for a test set. * Equal contribution. \u2020 The code is publicly available at the project page: https://zero-shotmodel-diagnosis.github.io/. CLIP StyleGAN Generator How would a change in [attribute] affect [my model]'s prediction? User Diagnosis Request Neural Tools Diagnosis Outcomes [attribute] = \"green eye\" \"vertical pupil\" \"pointed ear\" ... [my model] = cat/dog classifier ... Zero-shot counterfactual images Zero-shot sensitivity histogram \"g re en ey e\" \"v er ti ca l p u p il \" \"p o in te d ea r\" ... ..."}, "cited_paper_content": {"title": "Deep Learning Face Attributes In The Wild", "abstract": "Predicting face attributes in the wild is challenging due to complex face variations. We propose a novel deep learning framework for attribute prediction in the wild. It cascades two CNNs, LNet and ANet, which are fine-tuned jointly with attribute tags, but pre-trained differently. LNet is pre-trained by massive general object categories for face localization, while ANet is pre-trained by massive face identities for attribute prediction. This framework not only outperforms the state-of-the-art with a large margin, but also reveals valuable facts on learning face representation. (1) It shows how the performances of face localization (LNet) and attribute prediction (ANet) can be improved by different pre-training strategies. (2) It reveals that although the filters of LNet are fine-tuned only with image-level attribute tags, their response maps over entire images have strong indication of face locations. This fact enables training LNet for face localization with only image-level annotations, but without face bounding boxes or landmarks, which are required by all attribute recognition works. (3) It also demonstrates that the high-level hidden neurons of ANet automatically discover semantic concepts after pre-training with massive face identities, and such concepts are significantly enriched after fine-tuning with attribute tags. Each attribute can be well explained with a sparse linear combination of these concepts."}, "keywords": ["datasets", "e.g., specific attributes"], "citation_intent": "background"}
{"citing_id": "2303.17504v1", "cited_id": "1712.07629", "section_title": "G.4. Extensibility To Different Line Detectors And Matchers", "citation": "While endpoint-based line matcher achieves very promising performance in Table 13 , the endpoints of the matched line still do not necessarily match each other, indicating that the performance gain may come from the effectiveness of advanced point features #REFR .", "text_before_citation": ["The endpoint-based line matcher is surprisingly effective, Table 13 .", "Extensibility of the framework to different line detectors and matchers.", "We show results of \"R1 / P1 / #supports\" for line mapping on Hypersim #OTHEREFR with only line-line proposals.", "Also, we present here two customized line matchers by matching the line endpoints with either the nearest neighbor strategy (\"Endpts SP + NN\") or an advanced point-based matcher SuperGlue #OTHEREFR (\"Endpts SP + SG\"). Table 14 .", "Comparison between endpoint and line triangulation with \"Endpts SP + SG\" matcher on Hypersim #OTHEREFR ."], "text_after_citation": ["as \"Endpoint SP + SG\" consistently achieves the best track association under all detectors.", "We further test again the comparison between endpoint triangulation and algebraic line triangulation to see whether the endpoints of the matched line from the endpoint-based line matcher correspond to each other.", "Results in Table 14 show similar trends as Table  3 in the main paper, where performing algebraic line triangulation is significantly better than directly triangulating endpoints.", "This finding indicates that the endpoint-based matcher is surprisingly effective on matching lines with, however, unmatched endpoints, which may be due to the advantages of the rich point features #OTHEREFR .", "This encourages more research towards integrating the success of existing point description and matching solutions to improve line matching."], "citing_paper_content": {"title": "3D Line Mapping Revisited", "abstract": "In contrast to sparse keypoints, a handful of line segments can concisely encode the high-level scene layout, as they often delineate the main structural elements. In addition to offering strong geometric cues, they are also omnipresent in urban landscapes and indoor scenes. Despite their apparent advantages, current line-based reconstruction methods are far behind their point-based counterparts. In this paper we aim to close the gap by introducing LIMAP, a library for 3D line mapping that robustly and efficiently creates 3D line maps from multi-view imagery. This is achieved through revisiting the degeneracy problem of line triangulation, carefully crafted scoring and track building, and exploiting structural priors such as line coincidence, parallelism, and orthogonality. Our code integrates seamlessly with existing point-based Structure-from-Motion methods and can leverage their 3D points to further improve the line reconstruction. Furthermore, as a byproduct, the method is able to recover 3D association graphs between lines and points / vanishing points (VPs). In thorough experiments, we show that LIMAP significantly outperforms existing approaches for 3D line mapping. Our robust 3D line maps also open up new research directions. We show two example applications: visual localization and bundle adjustment, where integrating lines alongside points yields the best results. Code is available at https://github.com/cvg/limap."}, "cited_paper_content": {"title": "Superpoint: Self-Supervised Interest Point Detection And Description", "abstract": "This paper presents a self-supervised framework for training interest point detectors and descriptors suitable for a large number of multiple-view geometry problems in computer vision. As opposed to patch-based neural networks, our fully-convolutional model operates on full-sized images and jointly computes pixel-level interest point locations and associated descriptors in one forward pass. We introduce Homographic Adaptation, a multi-scale, multi-homography approach for boosting interest point detection repeatability and performing cross-domain adaptation (e.g., synthetic-to-real). Our model, when trained on the MS-COCO generic image dataset using Homographic Adaptation, is able to repeatedly detect a much richer set of interest points than the initial pre-adapted deep model and any other traditional corner detector. The final system gives rise to state-of-the-art homography estimation results on HPatches when compared to LIFT, SIFT and ORB."}, "keywords": ["advanced point features"], "citation_intent": "background"}
{"citing_id": "2303.12414v2", "cited_id": "1907.02189", "section_title": "Solution.", "citation": "Due to the complex nature of the problem, we solve it via exhaustive search, performing line search over the integer values of \u03c4 in the range given in #REFR and obtain the optimum of \u03b1 \u2208 (0, 1] corresponding to each value of \u03c4 .", "text_before_citation": ["Formulation P is a non-convex mixed-integer programming problem."], "text_after_citation": ["Note that, given a value of \u03c4 , P is still non-convex with respect to \u03b1.", "Therefore, we discretize the search space of \u03b1 and perform a line search over the discretized search space of \u03b1 for each \u03c4 .", "Based on this approach, the search space of P remains to be small due to the limited ranges/choices of \u03c4 and \u03b1 (i.e., the time complexity of performing line search over \u03c4 is O(T \u2212 t k \u2212 \u2206 k\u22121 ) and the time complexity of performing line search over \u03b1 is O(1/S \u03b1 ), where S \u03b1 is the discretization step used to discretize (0, 1] interval, resulting in a time complexity of O((T \u2212 t k \u2212 \u2206 k\u22121 ) \u00d7 1/S \u03b1 )).", "Thus, we are able to solve the problem via a reasonable precision (e.g., S \u03b1 = 0.01) in a short duration of time (e.g., less than 3 seconds on a laptop with Intel(R) Xeon(R) Gold 6242 CPU @ 2.80GHz)."], "citing_paper_content": {"title": "Delay-Aware Hierarchical Federated Learning", "abstract": "Federated learning has gained popularity as a means of training models distributed across the wireless edge. The paper introduces delay-aware federated learning (DFL) to improve the efficiency of distributed machine learning (ML) model training by addressing communication delays between edge and cloud. DFL employs multiple stochastic gradient descent iterations on device datasets during each global aggregation interval and intermittently aggregates model parameters through edge servers in local subnetworks. The cloud server synchronizes the local models with the global deployed model computed via a local-global combiner at global synchronization. The convergence behavior of DFL is theoretically investigated under a generalized data heterogeneity metric. A set of conditions is obtained to achieve the sub-linear convergence rate of O(1/k). Based on these findings, an adaptive control algorithm is developed for DFL, implementing policies to mitigate energy consumption and edge-to-cloud communication latency while aiming for a sublinear convergence rate. Numerical evaluations show DFL's superior performance in terms of faster global model convergence, reduced resource consumption, and robustness against communication delays compared to existing FL algorithms. In summary, this proposed method offers improved efficiency and satisfactory results when dealing with both convex and non-convex loss functions."}, "cited_paper_content": {"title": "On The Convergence Of Fedavg On Non-Iid Data", "abstract": "Federated learning enables a large amount of edge computing devices to jointly learn a model without data sharing. As a leading algorithm in this setting, Federated Averaging (\\texttt{FedAvg}) runs Stochastic Gradient Descent (SGD) in parallel on a small subset of the total devices and averages the sequences only once in a while. Despite its simplicity, it lacks theoretical guarantees under realistic settings. In this paper, we analyze the convergence of \\texttt{FedAvg} on non-iid data and establish a convergence rate of $\\mathcal{O}(\\frac{1}{T})$ for strongly convex and smooth problems, where $T$ is the number of SGDs. Importantly, our bound demonstrates a trade-off between communication-efficiency and convergence rate. As user devices may be disconnected from the server, we relax the assumption of full device participation to partial device participation and study different averaging schemes; low device participation rate can be achieved without severely slowing down the learning. Our results indicate that heterogeneity of data slows down the convergence, which matches empirical observations. Furthermore, we provide a necessary condition for \\texttt{FedAvg} on non-iid data: the learning rate $\\eta$ must decay, even if full-gradient is used; otherwise, the solution will be $\\Omega (\\eta)$ away from the optimal."}, "keywords": ["optimum", "problem"], "citation_intent": "method"}
{"citing_id": "2304.10000v1", "cited_id": "1908.08796", "section_title": "Related Literature", "citation": "However,developing reinforcement learning solutions to increase the safety and robustness of learned strategies in healthcare remains a key challenge #REFR .", "text_before_citation": ["While this method is computationally tractable for small POMDPs, the resulting belief MDP is often too large to optimize effectively (\u00c7 elik et al., 2015) .", "Thus in many large scale setting approximate dynamic programming techniques have been proposed to approximate optimal policies #OTHEREFR .", "Our setting differs from the POMDP setting in that the transition function is also partially observable since each patient has varying metabolism rate that makes the belief state conversion challenging to implement.", "Both reinforcement learning #OTHEREFR and POMDPs have been applied in the setting of personalized healthcare #OTHEREFR .", "Several reinforcement learning methods have been proposed for personalized medicine including deep reinforcement learning #OTHEREFR , off policy learning #OTHEREFR , and some model based methods #OTHEREFR ."], "text_after_citation": ["POMDPs have also been used to study personalized dosing in applications such as heart disease #OTHEREFR , stroke #OTHEREFR ) and", "Parkinson's disease #OTHEREFR .", "POMDP models have also been used in population health and disease screening and disease management #OTHEREFR . #OTHEREFR", "(2021) incorporated a personalized readmission prediction model to manage patient flow in a hospital.", "The methods we propose in this paper expand on these methods to a healthcare setting where both states and dynamics are partially observed."], "citing_paper_content": {"title": "Model Based Reinforcement Learning For Personalized Heparin Dosing", "abstract": "One of the key challenges in sequential decision making is optimizing systems safely in the case of partial information. While much of the existing work has focused on addressing this challenge in the case of either partially known states or partially known system dynamics, it is further exacerbated in cases where both states and dynamics are partially known. For instance the setting of computing heparin doses for patients fits this paradigm since the concentration of heparin in the patient cannot be measured directly and the rates at which patients metabolize it vary greatly between individuals. While many approaches proposed to resolve the challenge in this setting are model free, they require complex models that are not transparent to decision makers, and are difficult to analyze and guarantee safety. However, if some of the structure of the dynamics is known, a model based approach can be leveraged to provide safe policies with practical empirical performance and theoretical worst case guarantees. In this paper we propose a model based framework to address the challenge of partially observed states and dynamics in the context of designing personalized doses of heparin. We use a predictive model based on pharmacokinetics (the study of how the body effects substances through absorption, distribution, and metabolism) parameterized individually by patient, and infer the current concentration of heparin and predict future therapeutic effects taking into account different patients' characteristics. We formulate the patient parameter estimation problem in to a mixed integer linear program and show that our estimates are statistically consistent. We leverage this model by developing an adaptive dosing algorithm that outputs asymptotically optimal dose sequences based on a scenario generation approach, this approach is also capable of ensuring that the required heparin doses are maintained within a safe level. We validate our models with numerical experiments by first comparing the predictive capabilities of our model against existing machine learning techniques and demonstrating how our dosing algorithm can keep patients' related medical tests within a therapeutic range in a simulated ICU environment. Our results show that our methods are capable of maintaining patients in therapeutic range for 87.7% of the treatment time as opposed to existing weight based protocols that can only do so for 55.6% of the treatment time."}, "cited_paper_content": {"title": "Reinforcement Learning In Healthcare: A Survey", "abstract": "As a subfield of machine learning, \\emph{reinforcement learning} (RL) aims at empowering one's capabilities in behavioural decision making by using interaction experience with the world and an evaluative feedback. Unlike traditional supervised learning methods that usually rely on one-shot, exhaustive and supervised reward signals, RL tackles with sequential decision making problems with sampled, evaluative and delayed feedback simultaneously. Such distinctive features make RL technique a suitable candidate for developing powerful solutions in a variety of healthcare domains, where diagnosing decisions or treatment regimes are usually characterized by a prolonged and sequential procedure. This survey will discuss the broad applications of RL techniques in healthcare domains, in order to provide the research community with systematic understanding of theoretical foundations, enabling methods and techniques, existing challenges, and new insights of this emerging paradigm. By first briefly examining theoretical foundations and key techniques in RL research from efficient and representational directions, we then provide an overview of RL applications in a variety of healthcare domains, ranging from dynamic treatment regimes in chronic diseases and critical care, automated medical diagnosis from both unstructured and structured clinical data, as well as many other control or scheduling domains that have infiltrated many aspects of a healthcare system. Finally, we summarize the challenges and open issues in current research, and point out some potential solutions and directions for future research."}, "keywords": ["reinforcement learning solutions"], "citation_intent": "background"}
{"citing_id": "2304.12158v1", "cited_id": "2001.11576", "section_title": "Introduction", "citation": "These complexity upper bounds coincide with the ones provided in #REFR in the case of safety automata.", "text_before_citation": ["In this paper we ultimately solve Problem 1.1 for non-deterministic parity tree automata, which capture the whole MSO. Theorem 1.2.", "There is an algorithm that inputs a non-deterministic parity tree automaton A and computes the coin-flipping measure of the language of trees recognised by A. Moreover, this measure is an algebraic number.", "As a consequence, given additionally a rational number q, we can decide if the resulting measure is equal, smaller, or greater than q.", "Our algorithm is able to produce a representation of the above measure in three-fold exponential time.", "The decision problems with a fixed rational number q can be solved in two-fold exponential space."], "text_after_citation": ["It is worth to note that natural stronger probabilistic variants of the satisfiability problem are known to be undecidable.", "For example, both the non-emptiness problem #OTHEREFR and the value-1 problem #OTHEREFR for probabilistic finite automata, even for finite words, are undecidable.", "Moreover, Baier, Bertrand, and Gr\u00f6sser #OTHEREFR showed that it is undecidable whether there exists an \u03c9-word that is accepted by a given probabilistic B\u00fcchi automaton with a positive probability.", "So presumably we are close to the frontiers of decidability here.", "One should note that our result does not entail the Rabin tree theorem-the satisfiability problem does not directly reduce to its probabilistic variant."], "citing_paper_content": {"title": "The Probabilistic Rabin Tree Theorem *", "abstract": "The Rabin tree theorem yields an algorithm to solve the satisfiability problem for monadic second-order logic over infinite trees. Here we solve the probabilistic variant of this problem. Namely, we show how to compute the probability that a randomly chosen tree satisfies a given formula. We additionally show that this probability is an algebraic number. This closes a line of research where similar results were shown for formalisms weaker than the full monadic second-order logic."}, "cited_paper_content": {"title": "The Uniform Measure Of Simple Regular Sets Of Infinite Trees", "abstract": "We consider the problem of computing the measure of a regular set of infinite binary trees. While the general case remains unsolved, we show that the measure of a language can be computed when the set is given in one of the following three formalisms: a first-order formula with no descendant relation; a Boolean combination of conjunctive queries (with descendant relation); or by a non-deterministic safety tree automaton. Additionally, in the first two cases the measure of the set is always rational, while in the third it is an algebraic number. Moreover, we provide an example of a first-order formula that uses descendant relation and defines a language of infinite trees having an irrational (but algebraic) measure."}, "keywords": ["safety automata"], "citation_intent": "result"}
{"citing_id": "2303.14548v2", "cited_id": "1903.11027", "section_title": "Introduction", "citation": "These algorithms demonstrate a giant leap in 3D detection performance on multi-camera datasets (E.g., Nuscenes #REFR ).", "text_before_citation": ["Our proposed VEDet encodes the 3D scene from multiview images, and decodes objects with view-conditioned queries.", "The predicted 3D bounding boxes are expressed in the underlying views of the queries, which enables us to enforce viewpoint equivariance among predictions from multiple views.", "Virtual query views are generated during training and together with the viewpoint equivariance regularization bring richer geometric learning signals to guide the model to better understand the 3D structure in the scene.", "During inference, the global predictions can be obtained by simply choosing the global frame as the query view.", "multi-view camera-based 3D algorithms #OTHEREFR proposed to jointly aggregate multi-view information at the feature level, and directly predict a single set of detections in the global frame."], "text_after_citation": ["To aggregate information from different views, one line of query-based detectors adopt transformers to query image features #OTHEREFR or bird's-eye-view (BEV) features #OTHEREFR via an attention mechanism.", "In contrast, another line of works \"lift-splat-shoot\" #OTHEREFR image features from each view into the shared BEV features to be processed by convolutional detection heads #OTHEREFR .", "To further mitigate the depth ambiguity, some concurrent works have started extending multi-view to \"multi-sweep\" across timestamps and observe a promising performance boost #OTHEREFR .", "While the works mentioned above demonstrate a strong potential for multi-view 3D detection, progress has concentrated on input aggregation and information interplay across frames and less on learning objectives.", "We argue that the learning objective can play a crucial role in ingesting the core knowledge in a multi-view setting: 3D geometry."], "citing_paper_content": {"title": "Viewpoint Equivariance For Multi-View 3D Object Detection", "abstract": "3D object detection from visual sensors is a cornerstone capability of robotic systems. State-of-the-art methods focus on reasoning and decoding object bounding boxes from multi-view camera input. In this work we gain intuition from the integral role of multi-view consistency in 3D scene understanding and geometric learning. To this end, we introduce VEDet, a novel 3D object detection framework that exploits 3D multi-view geometry to improve localization through viewpoint awareness and equivariance. VEDet leverages a query-based transformer architecture and encodes the 3D scene by augmenting image features with positional encodings from their 3D perspective geometry. We design view-conditioned queries at the output level, which enables the generation of multiple virtual frames during training to learn viewpoint equivariance by enforcing multi-view consistency. The multi-view geometry injected at the input level as positional encodings and regularized at the loss level provides rich geometric cues for 3D object detection, leading to state-of-the-art performance on the nuScenes benchmark. The code and model are made available at https://github.com/TRI-ML/VEDet."}, "cited_paper_content": {"title": "Nuscenes: A Multimodal Dataset For Autonomous Driving", "abstract": "Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology. Image-based benchmark datasets have driven development in computer vision tasks such as object detection, tracking and segmentation of agents in the environment. Most autonomous vehicles, however, carry a combination of cameras and range sensors such as lidar and radar. As machine learning based methods for detection and tracking become more prevalent, there is a need to train and evaluate such methods on datasets containing range sensor data along with images. In this work we present nuTonomy scenes (nuScenes), the first published dataset to carry the full autonomous vehicle sensor suite: 6 cameras, 5 radars and 1 lidar, all with full 360 degree field of view. nuScenes comprises 1000 scenes, each 20s long and fully annotated with 3D bounding boxes for 23 classes and 8 attributes. It has 7x as many annotations and 100x as many images as the pioneering KITTI dataset. We define novel 3D detection and tracking metrics. We also provide careful dataset analysis as well as baselines for lidar and image based detection and tracking. Data, development kit and more information are available online at this http URL."}, "keywords": ["3D detection performance", "multi-camera datasets"], "citation_intent": "background"}
{"citing_id": "2303.10653v1", "cited_id": "1901.08573", "section_title": "Motivation", "citation": "This method regards the TRADES #REFR , a state-of-the-art adversarial training method, as a special case with only zero-th term, and is able to locate flatter minima.", "text_before_citation": ["It was developed to guarantee weight smoothness by simply averaging various checkpoints throughout the training trajectory.", "SWA has been applied to semi-supervised learning, Bayesian inference, and low-precision training.", "Especially, #OTHEREFR used SWA for the first time in adversarial training in order to smooth the weights and locate flatter minima, and showed that it enhances the adversarial generalization.", "Different from SWA, we propose a novel method (with details provided in Sec.", "4) via Taylor expansion of a small Gaussian noise to smooth the update of weights."], "text_after_citation": ["To demonstrate the latter, we visualize the loss landscape with respect to both input and weight spaces. As shown in Fig.", "2 , compared to the TRADES, our method significantly flattens the rugged landscape.", "In addition, it is worth noting that our empirical results (in Sec.", "5) demonstrate that our method can further improve not only clean accuracy but also adversarial accuracy compared with other SOTA (includes SWA from #OTHEREFR )."], "citing_paper_content": {"title": "Randomized Adversarial Training Via Taylor Expansion", "abstract": "In recent years, there has been an explosion of research into developing more robust deep neural networks against adversarial examples. Adversarial training appears as one of the most successful methods. To deal with both the robustness against adversarial examples and the accuracy over clean examples, many works develop enhanced adversarial training methods to achieve various trade-offs between them [19, 38, 82]. Leveraging over the studies [8, 32] that smoothed update on weights during training may help find flat minima and improve generalization, we suggest reconciling the robustness-accuracy trade-off from another perspective, i.e., by adding random noise into deterministic weights. The randomized weights enable our design of a novel adversarial training method via Taylor expansion of a small Gaussian noise, and we show that the new adversarial training method can flatten loss landscape and find flat minima. With PGD, CW, and Auto Attacks, an extensive set of experiments demonstrate that our method enhances the state-of-the-art adversarial training methods, boosting both robustness and clean accuracy. The code is available at https://github.com/Alexkael/Randomized-Adversarial-Training."}, "cited_paper_content": {"title": "Theoretically Principled Trade-Off Between Robustness And Accuracy", "abstract": "We identify a trade-off between robustness and accuracy that serves as a guiding principle in the design of defenses against adversarial examples. Although this problem has been widely studied empirically, much remains unknown concerning the theory underlying this trade-off. In this work, we decompose the prediction error for adversarial examples (robust error) as the sum of the natural (classification) error and boundary error, and provide a differentiable upper bound using the theory of classification-calibrated loss, which is shown to be the tightest possible upper bound uniform over all probability distributions and measurable predictors. Inspired by our theoretical analysis, we also design a new defense method, TRADES, to trade adversarial robustness off against accuracy. Our proposed algorithm performs well experimentally in real-world datasets. The methodology is the foundation of our entry to the NeurIPS 2018 Adversarial Vision Challenge in which we won the 1st place out of ~2,000 submissions, surpassing the runner-up approach by $11.41\\%$ in terms of mean $\\ell_2$ perturbation distance."}, "keywords": ["flatter minima"], "citation_intent": "method"}
{"citing_id": "2304.11105v1", "cited_id": "1807.06587", "section_title": "Related Work", "citation": "To alleviate this obstacle, #REFR propose to utilize natural color distributions learned from large-scale image data when no appropriate matching region is available in the reference.", "text_before_citation": ["With the advent of deep learning, #OTHEREFR proposed a real-time user interactive colorization model which can generate visually plausible results with sparse color hints.", "However, these methods may fail to provide robust and automatic colorization when dense user annotations are not available.", "Reference-based Colorization methods colorize the target grayscale image by transferring color information from a reference image using global color #OTHEREFR or some local similarity #OTHEREFR .", "Recently, some researchers #OTHEREFR have made use of deep features extracted by a pretrained neural network for reliable spatial correspondence and generate plausible results.", "However, these methods usually require a content-related image as reference, which poses challenges to users."], "text_after_citation": ["Automatic Colorization is generally a kind of learningbased method, which uses neural networks to learn the mapping of grayscale images to color images #OTHEREFR .", "These methods can be trained on large-scale datasets and do not require user input or reference images.", "Some follow-up works promote the colorized results by integrating additional semantic guidance, like class labels #OTHEREFR , semantic segmentation maps #OTHEREFR and instance-level bounding box #OTHEREFR .", "To improve the diversity of the colors generated, #OTHEREFR proposed a method that uses a convolutional neural network (CNN) to predict a probability distribution over the possible colors for each pixel.", "Generative Adversarial Network (GAN) #OTHEREFR is adopted for realistic and vivid colorization #OTHEREFR ."], "citing_paper_content": {"title": "Improved Diffusion-Based Image Colorization Via Piggybacked Models", "abstract": "Image colorization has been attracting the research interests of the community for decades. However, existing methods still struggle to provide satisfactory colorized results given grayscale images due to a lack of human-like global understanding of colors. Recently, large-scale Textto-Image (T2I) models have been exploited to transfer the semantic information from the text prompts to the image domain, where text provides a global control for semantic objects in the image. In this work, we introduce a colorization model piggybacking on the existing powerful T2I diffusion model. Our key idea is to exploit the color prior knowledge in the pre-trained T2I diffusion model for realistic and diverse colorization. A diffusion guider is designed to incorporate the pre-trained weights of the latent diffusion model to output a latent color prior that conforms to the visual semantics of the grayscale input. A lightness-aware VQVAE will then generate the colorized result with pixelperfect alignment to the given grayscale image. Our model can also achieve conditional colorization with additional inputs (e.g. user hints and texts). Extensive experiments show that our method achieves state-of-the-art performance in terms of perceptual quality."}, "cited_paper_content": {"title": "Deep Exemplar-Based Colorization", "abstract": "We propose the first deep learning approach for exemplar-based local colorization. Given a reference color image, our convolutional neural network directly maps a grayscale image to an output colorized image. Rather than using hand-crafted rules as in traditional exemplar-based methods, our end-to-end colorization network learns how to select, propagate, and predict colors from the large-scale data. The approach performs robustly and generalizes well even when using reference images that are unrelated to the input grayscale image. More importantly, as opposed to other learning-based colorization methods, our network allows the user to achieve customizable results by simply feeding different references. In order to further reduce manual effort in selecting the references, the system automatically recommends references with our proposed image retrieval algorithm, which considers both semantic and luminance information. The colorization can be performed fully automatically by simply picking the top reference suggestion. Our approach is validated through a user study and favorable quantitative comparisons to the-state-of-the-art methods. Furthermore, our approach can be naturally extended to video colorization. Our code and models will be freely available for public use."}, "keywords": ["natural color distributions"], "citation_intent": "background"}
{"citing_id": "2303.15678v1", "cited_id": "1512.03385", "section_title": "A.1. Diswot Under Different Teacher Models.", "citation": "According to the results in Table 13, the accuracy of the student network ResNet20 #REFR is unable to make consistent gains as the size of the teacher network increases.", "text_before_citation": ["When the teacher model becomes larger, the fixed handdesigned model would have huge teacher-student gaps, limiting the performance gain.", "DisWOT aims to solve this problem by searching the suitable student architecture for different teacher models."], "text_after_citation": ["Our proposed DisWOT enables a consistent increase in student network performance as the teacher network capacity increases on search space S 0 .", "In addition, DisWOT \u2020 achieves a performance gain of about 2% when stronger distillers are adopted. Table 13 .", "Top-1 accuracy (%) of ResNet20 with KD #OTHEREFR , student (DisWOT) with KD #OTHEREFR , student (DisWOT) with DisWOT \u2020 on search space S0 under different teachers."], "citing_paper_content": {"title": "Diswot: Student Architecture Search For Distillation Without Training", "abstract": "Knowledge distillation (KD) is an effective training strategy to improve the lightweight student models under the guidance of cumbersome teachers. However, the large architecture difference across the teacher-student pairs limits the distillation gains. In contrast to previous adaptive distillation methods to reduce the teacher-student gap, we explore a novel training-free framework to search for the best student architectures for a given teacher. Our work first empirically show that the optimal model under vanilla training cannot be the winner in distillation. Secondly, we find that the similarity of feature semantics and sample relations between random-initialized teacher-student networks have good correlations with final distillation performances. Thus, we efficiently measure similarity matrixs conditioned on the semantic activation maps to select the optimal student via an evolutionary algorithm without any training. In this way, our student architecture search for Distillation WithOut Training (DisWOT) significantly improves the performance of the model in the distillation stage with at least 180\u00d7 training acceleration. Additionally, we extend similarity metrics in DisWOT as new distillers and KD-based zero-proxies. Our experiments on CIFAR, ImageNet and NAS-Bench-201 demonstrate that our technique achieves state-of-the-art results on different search spaces. Our project and code are available at https://lilujunai.github.io/DisWOT-CVPR2023/. * Corresponding author, \u2020 equal contribution, PD conducted main experiments, LL proposed ideas and led the project & writing."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["teacher network increases", "accuracy"], "citation_intent": "result"}
{"citing_id": "2303.02343v1", "cited_id": "1708.03888", "section_title": "Large-Batch Training Challenge And Improvement", "citation": "This is aligned with the observations in ERM #REFR , where the lack of stochasticity makes the optimizer difficult to escape from a sharp local minimum.", "text_before_citation": ["Recall that the full batch size (50k) was used in the existing IRM implementations (Arjovsky et al., 2019; #OTHEREFR ).", "As we can see, in the full-batch setup, IRM methods lead to widely different invariant prediction accuracies, where REX and IRMV1 significantly outperform IRMV0.", "In contrast, in the small-batch case (with size 1k), the discrepancy in accuracy across methods vanishes.", "We see that IRMv0 can be as effective as IRMv1 and other IRM variants (such as REX) only if an appropriate small batch size is used.", "Empirical evidence in Figure 1 shows that large-batch IRM training is less effective than smallbatch."], "text_after_citation": ["We also justify this issue by visualizing the loss landscapes in Figure A2 .", "Notably, the small-batch training enables IRMV1 to converge to a local optimum with a flat loss landscape, indicating better generalization #OTHEREFR .", "Small-batch training is effective versus a zoo of large-batch optimization enhancements.", "To mitigate the large-batch IRM training issue, we next investigate the effectiveness of both smallbatch training and a zoo of large-batch optimization enhancements.", "Inspired by large-batch training techniques to scale up ERM, we consider Large-batch SGD (LSGD) #OTHEREFR and Layerwise Adaptive Learning Rate (LALR) #OTHEREFR ."], "citing_paper_content": {"title": "What Is Missing In Irm Training And Evalua-Tion? Challenges And Solutions", "abstract": "Invariant risk minimization (IRM) has received increasing attention as a way to acquire environment-agnostic data representations and predictions, and as a principled solution for preventing spurious correlations from being learned and for improving models' out-of-distribution generalization. Yet, recent works have found that the optimality of the originally-proposed IRM optimization (IRMV1) may be compromised in practice or could be impossible to achieve in some scenarios. Therefore, a series of advanced IRM algorithms have been developed that show practical improvement over IRMV1. In this work, we revisit these recent IRM advancements, and identify and resolve three practical limitations in IRM training and evaluation. First, we find that the effect of batch size during training has been chronically overlooked in previous studies, leaving room for further improvement. We propose small-batch training and highlight the improvements over a set of large-batch optimization techniques. Second, we find that improper selection of evaluation environments could give a false sense of invariance for IRM. To alleviate this effect, we leverage diversified test-time environments to precisely characterize the invariance of IRM when applied in practice. Third, we revisit Ahuja et al. (2020)'s proposal to convert IRM into an ensemble game and identify a limitation when a single invariant predictor is desired instead of an ensemble of individual predictors. We propose a new IRM variant to address this limitation based on a novel viewpoint of ensemble IRM games as consensus-constrained bilevel optimization. Lastly, we conduct extensive experiments (covering 7 existing IRM variants and 7 datasets) to justify the practical significance of revisiting IRM training and evaluation in a principled manner."}, "cited_paper_content": {"title": "Scaling Sgd Batch Size To 32K For Imagenet Training.", "abstract": "The most natural way to speed-up the training of large networks is to use data-parallelism on multiple GPUs. To scale Stochastic Gradient (SG) based methods to more processors, one need to increase the batch size to make full use of the computational power of each GPU. However, keeping the accuracy of network with increase of batch size is not trivial. Currently, the state-of-the art method is to increase Learning Rate (LR) proportional to the batch size, and use special learning rate with \"warm-up\" policy to overcome initial optimization difficulty. ::: By controlling the LR during the training process, one can efficiently use large-batch in ImageNet training. For example, Batch-1024 for AlexNet and Batch-8192 for ResNet-50 are successful applications. However, for ImageNet-1k training, state-of-the-art AlexNet only scales the batch size to 1024 and ResNet50 only scales it to 8192. The reason is that we can not scale the learning rate to a large value. To enable large-batch training to general networks or datasets, we propose Layer-wise Adaptive Rate Scaling (LARS). LARS LR uses different LRs for different layers based on the norm of the weights and the norm of the gradients. By using LARS algoirithm, we can scale the batch size to 32768 for ResNet50 and 8192 for AlexNet. Large batch can make full use of the system's computational power. For example, batch-4096 can achieve 3x speedup over batch-512 for ImageNet training by AlexNet model on a DGX-1 station (8 P100 GPUs)."}, "keywords": ["optimizer"], "citation_intent": "result"}
{"citing_id": "2303.14371v1", "cited_id": "1805.07103", "section_title": "Dmri Datasets And Data Preprocessing", "citation": "Compared with raw dMRI data, the networks trained with fODF peaks data achieve better tract segmentation results #REFR .", "text_before_citation": ["We use the dataset from TractSeg #OTHEREFR , including 105 subjects from the Human Connectome Project (HCP) #OTHEREFR . Each sub- Fig. 1 .", "Overview of registration-and uncertainty-based framework for WM tract segmentation with only one labeled subject. ject has 72 tracts annotated by neuroanatomists.", "We follow the same dataset split as #OTHEREFR , using 84 subjects for training, and the rest 21 subjects for testing.", "Regarding our subjectlevel one-shot segmentation setting, we only use one annotated subject out of the 84 training subjects.", "As shown in Fig.1 , we use multi-shell multi-tissue constrained spherical deconvolution (CSD) method #OTHEREFR with all gradient directions to transform dMRI data to fiber orientation distribution function (fODF) peaks data, which has 9 channels corresponding to sagittal, axial, and coronal directions (each direction has three channels), as in #OTHEREFR ."], "text_after_citation": ["Therefore, we directly segment tracts in the field of fODF peaks, and all subjects mentioned in the following are in this field."], "citing_paper_content": {"title": "A Registration-And Uncertainty-Based Framework For White Matter Tract Segmentation With Only One Annotated Subject", "abstract": "White matter (WM) tract segmentation based on diffusion magnetic resonance imaging (dMRI) plays an important role in the analysis of human health and brain diseases. However, the annotation of WM tracts is time-consuming and needs experienced neuroanatomists. In this study, to explore tract segmentation in the challenging setting of minimal annotations, we propose a novel framework utilizing only one annotated subject (subject-level one-shot) for tract segmentation. Our method is constructed by proposed registration-based peak augmentation (RPA) and uncertaintybased refining (URe) modules. RPA module synthesizes pseudo subjects and their corresponding labels to improve the tract segmentation performance. The proposed URe module alleviates the negative influence of the low-confidence voxels on pseudo subjects. Experimental results show that our method outperforms other state-of-the-art methods by a large margin, and our proposed modules are effective. Overall, our method achieves accurate whole-brain tract segmentation with only one annotated subject. Our code is available at https://github.com/HaoXu0507/ ISBI2023-One-Shot-WM-Tract-Segmentation."}, "cited_paper_content": {"title": "Tractseg - Fast And Accurate White Matter Tract Segmentation", "abstract": "The individual course of white matter fiber tracts is an important factor for analysis of white matter characteristics in healthy and diseased brains. Diffusion-weighted MRI tractography in combination with region-based or clustering-based selection of streamlines is a unique combination of tools which enables the in-vivo delineation and analysis of anatomically well-known tracts. This, however, currently requires complex, computationally intensive processing pipelines which take a lot of time to set up. TractSeg is a novel convolutional neural network-based approach that directly segments tracts in the field of fiber orientation distribution function (fODF) peaks without using tractography, image registration or parcellation. We demonstrate that the proposed approach is much faster than existing methods while providing unprecedented accuracy, using a population of 105 subjects from the Human Connectome Project. We also show initial evidence that TractSeg is able to generalize to differently acquired data sets for most of the bundles. The code and data are openly available at https://github.com/MIC-DKFZ/TractSeg/ and https://doi.org/10.5281/zenodo.1088277 , respectively."}, "keywords": ["better tract segmentation"], "citation_intent": "result"}
{"citing_id": "2304.11744v1", "cited_id": "1704.03477", "section_title": "Methodology", "citation": "As per convention, each sh i point also contains a two-dimensional binary pen state #REFR -(1, 0): stroke is being drawn, (0, 1): the end of the stroke, (0, 0): padding points to account for the fact that all strokes have a different number of points.", "text_before_citation": ["In this section, we first introduce our classification model which is designed around strokes as sketch building blocks. We then introduce our method for model explainability.", "As a pre-processing step, we simplify all sketches by the RDP algorithm #OTHEREFR .", "For each stroke s i consisting of k points, {s i,1 , s i,2 , ..., s i,k\u22121 , s i,k }, we identify three inherent properties in s i and learn respective descriptor for each: location l i , shape sh i and stroke order o i .", "We use the coordinate value (x i , y i ) of s i,1 to represent the location of each stroke l i .", "In order to disentangle shape information sh i from its actual location, we use relative instead of absolute coordinates and move the starting point of all strokes to the canvas centre."], "text_after_citation": ["Sketch-specific encoder.", "Our proposed sketch encoder f w , which we name SketchXAINet (\"X\" for EXplainability), first learns to encode l i , sh i and o i with different learnable components before fusing them together into a Transformer for final decision.", "This tailored model design is then ready to undertake the novel explainability task defined later. A full high-level schematic is shown in Fig. 2 .", "We use a bidirectional LSTM #OTHEREFR to extract shape information of each stroke sh i , and one linear layer for location l i embedding learning.", "We pre-define the maximum number of strokes allowed and assign a learnable embedding for each order (time) embedding o i ."], "citing_paper_content": {"title": "Sketchxai: A First Look At Explainability For Human Sketches", "abstract": "Iterative Figure 1. Explainability, but for human sketches. We demonstrate a new methodology for explaining AI decisions on human sketch data. Instead of one static explanation per instance as in existing works, our proposed method supports generating infinitely many dynamic explanations with each showcasing how an AI classifier responds to a different view of the same sketch input. This is made possible by seeing stroke as the fundamental block to explain a sketch and randomising every stroke location to introduce stochasticity at every case. The infinite varieties of explanation paths allows humans to enjoy a wider coverage on how AI functions, and therefore to better scrutinise AI."}, "cited_paper_content": {"title": "A Neural Representation Of Sketch Drawings", "abstract": "We present sketch-rnn, a recurrent neural network (RNN) able to construct stroke-based drawings of common objects. The model is trained on thousands of crude human-drawn images representing hundreds of classes. We outline a framework for conditional and unconditional sketch generation, and describe new robust training methods for generating coherent sketch drawings in a vector format."}, "keywords": ["strokes", "two-dimensional binary pen"], "citation_intent": "background"}
{"citing_id": "2304.14224v1", "cited_id": "1904.05835", "section_title": "Experimental Results", "citation": "SMC-2 has better performance results on Cifar10/100 when compared to other competitors (LSR, DLB) #REFR .", "text_before_citation": ["From Table 1 , it can be seen that the improvement of SMC-2 on Cifar100 relative to vanilla its validation top-1 accuracy ranges from 0.74% to 1.87%, and on Cifar10 compared to vanilla its validation top-1 accuracy ranges from 0.26% to 0.65%, which shows that SMC-2 is effective."], "text_after_citation": ["Also compared to the advanced method SAM, our algorithm converges faster and better except for the WRN-28-10 model.", "As can be seen in Figure 3 , SMC-2 is still effective on large data sets.", "Compared with SGD, SMC-2 significantly improves the convergence speed and generalization ability.", "However, due to the limitation of our computational resources, we cannot choose the hyperparameters of SMC-2 suitable for ImageNet, so we cannot compare it with other methods.", "Figure 3 : Comparison of the validation Top-1 accuracy results of SGD and SMC-2 trained on ImageNet with AlexNet model using different seeds."], "citing_paper_content": {"title": "Self-Discipline On Multiple Channels", "abstract": "Self-distillation relies on its own information to improve the generalization ability of the model and has a bright future. Existing self-distillation methods either require additional models, model modification, or batch size expansion for training, which increases the difficulty of use, memory consumption, and computational cost. This paper developed Self-discipline on multiple channels(SMC), which combines consistency regularization with self-distillation using the concept of multiple channels. Conceptually, SMC consists of two steps: 1) each channel data is simultaneously passed through the model to obtain its corresponding soft label, and 2) the soft label saved in the previous step is read together with the soft label obtained from the current channel data through the model to calculate the loss function. SMC uses consistent regularization and self-distillation to improve the generalization ability of the model and the robustness of the model to noisy labels. We named the SMC containing only two channels as SMC-2. Comparative experimental results on both datasets show that SMC-2 outperforms Label Smoothing Regularizaion and Self-distillation From The Last Mini-batch on all models, and outperforms the state-of-the-art Sharpness-Aware Minimization method on 83% of the models.Compatibility of SMC-2 and data augmentation experimental results show that using both SMC-2 and data augmentation improves the generalization ability of the model between 0.28% and 1.80% compared to using only data augmentation. Ultimately, the results of the label noise interference experiments show that SMC-2 curbs the tendency that the model's generalization ability decreases in the late training period due to the interference of label noise."}, "cited_paper_content": {"title": "Variational Information Distillation For Knowledge Transfer", "abstract": "Transferring knowledge from a teacher neural network pretrained on the same or a similar task to a student neural network can significantly improve the performance of the student neural network. Existing knowledge transfer approaches match the activations or the corresponding hand-crafted features of the teacher and the student networks. We propose an information-theoretic framework for knowledge transfer which formulates knowledge transfer as maximizing the mutual information between the teacher and the student networks. We compare our method with existing knowledge transfer methods on both knowledge distillation and transfer learning tasks and show that our method consistently outperforms existing methods. We further demonstrate the strength of our method on knowledge transfer across heterogeneous network architectures by transferring knowledge from a convolutional neural network (CNN) to a multi-layer perceptron (MLP) on CIFAR-10. The resulting MLP significantly outperforms the-state-of-the-art methods and it achieves similar performance to the CNN with a single convolutional layer."}, "keywords": ["SMC-2", "Cifar10/100"], "citation_intent": "result"}
{"citing_id": "2303.13714v1", "cited_id": "1903.05789", "section_title": "Related Work", "citation": "Our work is perhaps most similar to #REFR , who propose a VAE on the latent space of another VAE learned directly on pixels.", "text_before_citation": ["(2022) also trains an encoder with the same loss objectives, but it outputs continuous values and trains a diffusion model over these latents.", "Our work adapts the pretrained autoencoders from #OTHEREFR since our VAEs act in continuous space.", "An alternative two-stage approach uses an upsampling stack which first creates low-resolution RGB images before upsampling them to a higher resolution #OTHEREFR .", "This method works very well for diffusion models #OTHEREFR as diffusion models have good super-resolution capabilities #OTHEREFR .", "In our hierarchical VAE experiments, we found that upsampling stack is outperformed by a latent-space approach (Section 4.3)."], "text_after_citation": ["This work is similarly motivated to ours, accounting for the fact that the ambient dimension is much greater than the manifold dimension.", "However, samples from their model are blurry since their first stage is a regular VAE trained with a pixel-space distortion term.", "In contrast, we learn the underlying manifold using a DAE augmented with perceptual and aversarial losses, and achieve much higher compression rates with little to no blurriness.", "Furthermore, our second stage is based off hierarchical VAEs instead of single-group ones to reap the benefits of high stochastic depth observed in #OTHEREFR .", "Perceptual and Adversarial Losses in VAEs."], "citing_paper_content": {"title": "High Fidelity Image Synthesis With Deep Vaes In Latent Space", "abstract": "We present fast, realistic image generation on high-resolution, multimodal datasets using hierarchical variational autoencoders (VAEs) trained on a deterministic autoencoder's latent space. In this two-stage setup, the autoencoder compresses the image into its semantic features, which are then modeled with a deep VAE. With this method, the VAE avoids modeling the fine-grained details that constitute the majority of the image's code length, allowing it to focus on learning its structural components. We demonstrate the effectiveness of our two-stage approach, achieving a FID of 9.34 on the ImageNet-256 dataset which is comparable to BigGAN. We make our implementation available online 1 ."}, "cited_paper_content": {"title": "Diagnosing And Enhancing Vae Models", "abstract": "Although variational autoencoders (VAEs) represent a widely influential deep generative model, many aspects of the underlying energy function remain poorly understood. In particular, it is commonly believed that Gaussian encoder/decoder assumptions reduce the effectiveness of VAEs in generating realistic samples. In this regard, we rigorously analyze the VAE objective, differentiating situations where this belief is and is not actually true. We then leverage the corresponding insights to develop a simple VAE enhancement that requires no additional hyperparameters or sensitive tuning. Quantitatively, this proposal produces crisp samples and stable FID scores that are actually competitive with a variety of GAN models, all while retaining desirable attributes of the original VAE architecture. A shorter version of this work will appear in the ICLR 2019 conference proceedings (Dai and Wipf, 2019). The code for our model is available at https://github.com/daib13/ TwoStageVAE."}, "keywords": ["latent space", "another VAE"], "citation_intent": "background"}
{"citing_id": "2303.14778v1", "cited_id": "1902.09630", "section_title": "C. Dynamically Guided Learning Algorithm", "citation": "L iou\u2212v and L iou\u2212t are the GIoU loss #REFR of the visible branch and thermal branch.", "text_before_citation": ["The \u03b1, \u03b2, and \u03b3 are set to 100, 1, and 100000 respectively.", "Similarly, the total loss of the visible and thermal infrared branches can be described as follows:", "EQUATION", "where L dgl f v and L dgl f t are the dynamically guided learning loss of the visible branch and thermal branch guided by the fused branch, respectively.", "Here, the settings of \u03b1, \u03b2 and \u03b3 are the same as in Equation 9."], "text_after_citation": ["Note that, we use an iterative training strategy to avoid mutual interference among branches, i.e., the gradients of other branches will be truncated when optimizing a particular branch.", "A similar training strategy is also used for the optimization of GAN network #OTHEREFR , #OTHEREFR ."], "citing_paper_content": {"title": "Rgbt Tracking Via Progressive Fusion Transformer With Dynamically Guided Learning", "abstract": "Existing Transformer-based RGBT tracking methods either use cross-attention to fuse the two modalities, or use self-attention and cross-attention to model both modality-specific and modality-sharing information. However, the significant appearance gap between modalities limits the feature representation ability of certain modalities during the fusion process. To address this problem, we propose a novel Progressive Fusion Transformer called ProFormer, which progressively integrates single-modality information into the multimodal representation for robust RGBT tracking. In particular, ProFormer first uses a self-attention module to collaboratively extract the multimodal representation, and then uses two cross-attention modules to interact it with the features of the dual modalities respectively. In this way, the modality-specific information can well be activated in the multimodal representation. Finally, a feed-forward network is used to fuse two interacted multimodal representations for the further enhancement of the final multimodal representation. In addition, existing learning methods of RGBT trackers either fuse multimodal features into one for final classification, or exploit the relationship between unimodal branches and fused branch through a competitive learning strategy. However, they either ignore the learning of single-modality branches or result in one branch failing to be well optimized. To solve these problems, we propose a dynamically guided learning algorithm that adaptively uses well-performing branches to guide the learning of other branches, for enhancing the representation ability of each branch. Extensive experiments demonstrate that our proposed ProFormer sets a new state-of-the-art performance on RGBT210, RGBT234, LasHeR, and VTUAV datasets. In particular, the improvements of ProFormer on the LasHeR dataset are up to +6.2%/ + 4.1% in PR/SR compared with existing best RGBT tracker."}, "cited_paper_content": {"title": "Generalized Intersection Over Union: A Metric And A Loss For Bounding Box Regression", "abstract": "Intersection over Union (IoU) is the most popular evaluation metric used in the object detection benchmarks. However, there is a gap between optimizing the commonly used distance losses for regressing the parameters of a bounding box and maximizing this metric value. The optimal objective for a metric is the metric itself. In the case of axis-aligned 2D bounding boxes, it can be shown that IoU can be directly used as a regression loss. However, IoU has a plateau making it infeasible to optimize in the case of non-overlapping bounding boxes. In this paper, we address the this weakness by introducing a generalized version of IoU as both a new loss and a new metric. By incorporating this generalized IoU ( GIoU) as a loss into the state-of-the art object detection frameworks, we show a consistent improvement on their performance using both the standard, IoU based, and new, GIoU based, performance measures on popular object detection benchmarks such as PASCAL VOC and MS COCO."}, "keywords": ["GIoU loss"], "citation_intent": "background"}
{"citing_id": "2303.14348v1", "cited_id": "1807.11724", "section_title": "Category-Level Zs-Sbir", "citation": "Meanwhile, we can see Ours-Ret achieves better results on TU-Berlin Ext and Sketchy Ext datasets, and works slightly worse than Ours-RN on Sketchy Ext #REFR Split and QuickDraw Ext.", "text_before_citation": ["ViT-Ret means replacing the class token in ViT with a retrieval token used for matching; while ViT-Vis uses the visual tokens for matching.", "It should be noted that all the baselines, except CC-DG #OTHEREFR , StyleGuide #OTHEREFR and ViT variants #OTHEREFR , employ external semantic information, whereas our method only relies on the learned vi-sual correspondences between sketch-photo pairs.", "We also compare two variants of our model, i.e., Ours-Ret and Ours-RN, which retrieve images based on the retrieval token and the relation network separately. Evaluation protocol.", "Mean average precision (mAP), precision on top 100 (Prec@100) and top 200 (Prec@200) are reported following the standard evaluation protocol. Results.", "From Table 1 , we can see that our proposed method achieves better or comparable results over other competitors, which is especially noteworthy since our method does not benefit from extra semantic information, i.e., text or class label."], "text_after_citation": ["Some qualitative results (including some failure cases) are illustrated in Figure 3 .", "We can observe that, compared with SAKE and DSN, most of the top images retrieved using our approach more faithfully resemble to the query sketches in terms of the overall object pose and shape characteristics.", "And the false positives given by our method are somehow reasonable, as they are superficially similar to the query. We additionally compare ours with baselines for generalized ZS-SBIR.", "Results in Table 2 show that ours clearly outperforms others, suggesting a strong generalizability of our model. Ablation study.", "An ablative study is conducted to examine the importance of each key component in our model."], "citing_paper_content": {"title": "Zero-Shot Everything Sketch-Based Image Retrieval, And In Explainable Style", "abstract": "This paper studies the problem of zero-short sketchbased image retrieval (ZS-SBIR), however with two significant differentiators to prior art (i) we tackle all variants (inter-category, intra-category, and cross datasets) of ZS-SBIR with just one network (\"everything\"), and (ii) we would really like to understand how this sketch-photo matching operates (\"explainable\"). Our key innovation lies with the realization that such a cross-modal matching problem could be reduced to comparisons of groups of key local patches-akin to the seasoned \"bag-of-words\" paradigm. Just with this change, we are able to achieve both of the aforementioned goals, with the added benefit of no longer requiring external semantic knowledge. Technically, ours is a transformer-based cross-modal network, with three novel components (i) a self-attention module with a learnable tokenizer to produce visual tokens that correspond to the most informative local regions, (ii) a cross-attention module to compute local correspondences between the visual tokens across two modalities, and finally (iii) a kernel-based relation network to assemble local putative matches and produce an overall similarity metric for a sketch-photo pair. Experiments show ours indeed delivers superior performances across all ZS-SBIR settings. The all important explainable goal is elegantly achieved by visualizing crossmodal token correspondences, and for the first time, via sketch to photo synthesis by universal replacement of all matched photo patches. Code and model are available at https://github.com/buptLinfy/ZSE-SBIR."}, "cited_paper_content": {"title": "A Zero-Shot Framework For Sketch-Based Image Retrieval", "abstract": "Sketch-based image retrieval (SBIR) is the task of retrieving images from a natural image database that correspond to a given hand-drawn sketch. Ideally, an SBIR model should learn to associate components in the sketch (say, feet, tail, etc.) with the corresponding components in the image having similar shape characteristics. However, current evaluation methods simply focus only on coarse-grained evaluation where the focus is on retrieving images which belong to the same class as the sketch but not necessarily having the same shape characteristics as in the sketch. As a result, existing methods simply learn to associate sketches with classes seen during training and hence fail to generalize to unseen classes. In this paper, we propose a new benchmark for zero-shot SBIR where the model is evaluated in novel classes that are not seen during training. We show through extensive experiments that existing models for SBIR that are trained in a discriminative setting learn only class specific mappings and fail to generalize to the proposed zero-shot setting. To circumvent this, we propose a generative approach for the SBIR task by proposing deep conditional generative models that take the sketch as an input and fill the missing information stochastically. Experiments on this new benchmark created from the \"Sketchy\" dataset, which is a large-scale database of sketch-photo pairs demonstrate that the performance of these generative models is significantly better than several state-of-the-art approaches in the proposed zero-shot framework of the coarse-grained SBIR task."}, "keywords": ["Ret", "QuickDraw Ext"], "citation_intent": "result"}
{"citing_id": "2303.10121v1", "cited_id": "1907.11692", "section_title": "Data Annotation", "citation": "Specifically, we first use the RoBERTa transformer #REFR to extract the vector embeddings of both claims and tweets, and then we compute the cosine similarity between each claim and tweet in our data, retaining the top-100 most similar tweets for each claim.", "text_before_citation": ["Given the collected fact-checked information, our goal is to find tweets reporting such verified claims.", "However, as manually annotating the whole corpus of tweets in our dataset is not a suitable solution, we deploy a machine learning-based annotation strategy to maximize the likelihood of finding tweets related to one of the claims under analysis."], "text_after_citation": ["Consequently, we end up with 8,300 unique tweet-claim pairs that were inspected via a manual labeling process.", "It is worth noting that the choice of the RoBERTa transformer depends on the higher similarity scores provided by this model with respect to other transformers (ms-marco-MiniLM-L-4-v2 and quora-roberta-base), which allows us to maximize the chance of finding a matching pair.", "Next, we annotate the above-mentioned tweet-claim pairs considering a strict definition of relevance, i.e., the tweet discusses the claim if it explicitly mentions the same entities and events reported in the fact-checked information.", "Table 1 shows some examples of false claims and matched tweets, which highlight how the matched tweet can discuss a claim without expressing any stance (examples no.", "1 and 2) or can support or refute the claim (examples no. 3 and 4, respectively)."], "citing_paper_content": {"title": "Retrieving False Claims On Twitter During The Russia-Ukraine Conflict", "abstract": "Nowadays, false and unverified information on social media sway individuals' perceptions during major geo-political events and threaten the quality of the whole digital information ecosystem. Since the Russian invasion of Ukraine, several fact-checking organizations have been actively involved in verifying stories related to the conflict that circulated online. In this paper, we leverage a public repository of fact-checked claims to build a methodological framework for automatically identifying false and unsubstantiated claims spreading on Twitter in February 2022. Our framework consists of two sequential models: First, the claim detection model identifies whether tweets incorporate a (false) claim among those considered in our collection. Then, the claim retrieval model matches the tweets with fact-checked information by ranking verified claims according to their relevance with the input tweet. Both models are based on pre-trained language models and fine-tuned to perform a text classification task and an information retrieval task, respectively. In particular, to validate the effectiveness of our methodology, we consider 83 verified false claims that spread on Twitter during the first week of the invasion, and manually annotate 5,872 tweets according to the claim(s) they report. Our experiments show that our proposed methodology outperforms standard baselines for both claim detection and claim retrieval. Overall, our results highlight how social media providers could effectively leverage semi-automated * V. La Gatta and C. Wei contributed equally to this work."}, "cited_paper_content": {"title": "Roberta: A Robustly Optimized Bert Pretraining Approach", "abstract": "Language model pretraining has led to significant performance gains but careful comparison between different approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes, and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results highlight the importance of previously overlooked design choices, and raise questions about the source of recently reported improvements. We release our models and code."}, "keywords": ["tweets", "vector embeddings"], "citation_intent": "method"}
{"citing_id": "2303.18011v1", "cited_id": "1604.02201", "section_title": "I. Introduction", "citation": "Another advantage is transfer learning i.e., training of low-resource languages in combination with high resource languages improves the translation quality of low-resource languages #REFR .", "text_before_citation": ["Neural machine translation (NMT) has established a number of cutting-edge benchmarks in machine translation tasks #OTHEREFR , #OTHEREFR .", "It is an encoder-decoder framework based on a sequence-to-sequence prediction model where the encoder generates the context vector by taking a source-side word as input, and the decoder decodes this generated context vector into target sequences.", "Recent works ( [3] - #OTHEREFR ) have extended the NMT approach to support multilingual translation, i.e., training a single model that can translate between multiple languages.", "There are several reasons for shifting researchers' interest from bilingual to multilingual machine translation.", "First, single model training for large number of languages makes multilingual model more cost-effective than multiple bilingual models."], "text_after_citation": ["One of the fine examples of transfer learning-based NMT under multilingual scenario is Zero-Shot Translation (ZST) #OTHEREFR .", "ZST is an unsupervised approach in which pretrained translation models are tested on related unseen language pairs.", "There has been a significant amount of works done on multilingual machine translation, with the majority of the works focusing on translation between language pairs with English as a prominent language #OTHEREFR - #OTHEREFR .", "However, very few works on non-English language pairs exist #OTHEREFR , #OTHEREFR .", "One of the reasons for less multilingual systems on non-English language pairs is insufficient training data for such language pairs."], "citing_paper_content": {"title": "Exploiting Multilingualism In Low-Resource Neural Machine Translation Via Adversarial Learning", "abstract": "Generative Adversarial Networks (GAN) offer a promising approach for Neural Machine Translation (NMT). However, feeding multiple morphologically languages into a single model during training reduces the NMT's performance. In GAN, similar to bilingual models, multilingual NMT only considers one reference translation for each sentence during model training. This single reference translation limits the GAN model from learning sufficient information about the source sentence representation. Thus, in this article, we propose Denoising Adversarial Auto-encoderbased Sentence Interpolation (DAASI) approach to perform sentence interpolation by learning the intermediate latent representation of the source and target sentences of multilingual language pairs. Apart from latent representation, we also use the Wasserstein-GAN approach for the multilingual NMT model by incorporating the model generated sentences of multiple languages for reward computation. This computed reward optimizes the performance of the GAN-based multilingual model in an effective manner. We demonstrate the experiments on low-resource language pairs and find that our approach outperforms the existing state-of-the-art approaches for multilingual NMT with a performance gain of up to 4 BLEU points. Moreover, we use our trained model on zero-shot language pairs under an unsupervised scenario and show the robustness of the proposed approach."}, "cited_paper_content": {"title": "Transfer Learning For Low-Resource Neural Machine Translation", "abstract": "The encoder-decoder framework for neural machine translation (NMT) has been shown effective in large data scenarios, but is much less effective for low-resource languages. We present a transfer learning method that significantly improves Bleu scores across a range of low-resource languages. Our key idea is to first train a high-resource language pair (the parent model), then transfer some of the learned parameters to the low-resource pair (the child model) to initialize and constrain training. Using our transfer learning method we improve baseline NMT models by an average of 5.6 Bleu on four low-resource language pairs. Ensembling and unknown word replacement add another 2 Bleu which brings the NMT performance on low-resource machine translation close to a strong syntax based machine translation (SBMT) system, exceeding its performance on one language pair. Additionally, using the transfer learning model for re-scoring, we can improve the SBMT system by an average of 1.3 Bleu, improving the state-of-the-art on low-resource machine translation."}, "keywords": ["translation quality"], "citation_intent": "background"}
{"citing_id": "2303.02251v1", "cited_id": "2002.11569", "section_title": "Combatting Robust Overfitting", "citation": "Table 1 reports the test error at the 200 epoch mark, replicating the results reported in Table 2 in #REFR .", "text_before_citation": ["We replicate here the experiment of Rice et al. (2020) exhibiting robust overfitting for AT training. As in Rice et al.", "(2020) we use preactivation ResNet18 on CIFAR-10 with PGD attacks of 10 steps, and with SGD learning rate decay at epochs 100 and 150. Further experimental setup is discussed in Appendix E.1. We train AT with PGD of #OTHEREFR , TRADES of #OTHEREFR", "(2019) , and HR with r \u2265 0, > 0 and \u03b1 = 0.", "For all algorithms, we use the same PGD attack on training with = 8/255 and 10 attack steps.", "Figure 3 shows the test loss and error up to 200 training epochs when using pre-activation ResNet18 under \u221e attacks."], "text_after_citation": ["AT clearly exhibits robust overfitting as the test loss drastically increases with more training epochs.", "The test error also increases but perhaps not as drastically.", "This indicates that although the number of errors made eventually settles, worryingly, the errors are made with increasing confidence as quantified by the crossentropy loss. Similar to Rice et al.", "(2020), we find that TRADES #OTHEREFR without early stopping also experiences robust overfitting although to a lesser extent than regular PGD training.", "HR training with modest parameter values r effectively mitigates this phenomenon and results in a monotonically decreasing test loss as a function of training epochs eliminating the need for early stopping."], "citing_paper_content": {"title": "Certified Robust Neural Networks: Generalization And Corruption Resistance", "abstract": "Adversarial training aims to reduce the problematic susceptibility of modern neural networks to small data perturbations. Surprisingly, overfitting is a major concern in adversarial training of neural networks despite being mostly absent in standard training. We provide here theoretical evidence for this peculiar \"robust overfitting\" phenomenon. Subsequently, we advance a novel loss function which we show both theoretically as well as empirically to enjoy a certified level of robustness against data evasion and poisoning attacks while ensuring guaranteed generalization. We indicate through careful numerical experiments that our resulting holistic robust (HR) training procedure yields state of the art performance in terms of adversarial error loss. Finally, we indicate that HR training can be interpreted as a direct extension of adversarial training and comes with a negligible additional computational burden. A readyto-use python library implementing our algorithm is available at https://github.com/ RyanLucas3/HR_Neural_Networks."}, "cited_paper_content": {"title": "Overfitting In Adversarially Robust Deep Learning", "abstract": "It is common practice in deep learning to use overparameterized networks and train for as long as possible; there are numerous studies that show, both theoretically and empirically, that such practices surprisingly do not unduly harm the generalization performance of the classifier. In this paper, we empirically study this phenomenon in the setting of adversarially trained deep networks, which are trained to minimize the loss under worst-case adversarial perturbations. We find that overfitting to the training set does in fact harm robust performance to a very large degree in adversarially robust training across multiple datasets (SVHN, CIFAR-10, CIFAR-100, and ImageNet) and perturbation models ($\\ell_\\infty$ and $\\ell_2$). Based upon this observed effect, we show that the performance gains of virtually all recent algorithmic improvements upon adversarial training can be matched by simply using early stopping. We also show that effects such as the double descent curve do still occur in adversarially trained models, yet fail to explain the observed overfitting. Finally, we study several classical and modern deep learning remedies for overfitting, including regularization and data augmentation, and find that no approach in isolation improves significantly upon the gains achieved by early stopping. All code for reproducing the experiments as well as pretrained model weights and training logs can be found at this https URL."}, "keywords": ["200 epoch mark"], "citation_intent": "result"}
{"citing_id": "2303.06808v1", "cited_id": "1005.2243", "section_title": "Study Design", "citation": "Since robustness #REFR is another important metric that evaluates the generalization ability to handle unseen data of the trained model, we obtain multiple trained code models and study their adversarial robustness in this RQ.", "text_before_citation": ["In order to assess the effectiveness of the data augmentation methods in Section 3 in source code learning, we design three research questions, as follows:", "\u2022 RQ1: Can existing data augmentation methods produce accurate code models? Accuracy is the basic metric to evaluate the performance of a trained model, and therefore, we first assess whether data augmentation methods from NLP and graph learning (as introduced in Section 3) are able to improve the accuracy of the code models, based on the comparisons with two baseline approaches, namely, the training approach that does not use any data augmentation method, and the training approach that only uses the simple code refactoring method (as introduced in Section 2.2).", "In this RQ, first, we prepare the original training data and randomly initialize code models, and then train these models using different data augmentation methods as listed in Fig. 2 .", "Specifically, we compare these methods in terms of two metrics, namely, 1) the convergence speed of the model and 2) the final accuracy of the model with fixed training epochs.", "\u2022 RQ2: Can existing data augmentation methods produce robust code models?"], "text_after_citation": ["According to the literature #OTHEREFR , data augmentation is helpful for improving the adversarial robustness of DNN models in other fields (e.g., computer vision).", "Thus, we also explore whether this conclusion also holds for code models.", "Concretely, we employ two state-of-the-art adversarial attacks, namely, MHM #OTHEREFR and ALERT #OTHEREFR , to evaluate the robustness improvement of the code models that are trained using data augmentation.", "\u2022 RQ3: How does data volume affect the effectiveness of data augmentation methods? Finally, since the initial purpose of data augmentation is to solve the problem of the lack of labeled data, it is necessary to further investigate the effectiveness of data augmentation methods in a more practical scenario, namely, the case when there is no sufficient training data.", "To do so, we reduce the size of the training set of each dataset and repeat the experiments in RQ1 and RQ2, and we then check whether data augmentation methods are still useful."], "citing_paper_content": {"title": "Boosting Source Code Learning With Data Augmentation: An Empirical Study", "abstract": "The next era of program understanding is being propelled by the use of machine learning to solve software problems. Recent studies have shown surprising results of source code learning, which applies deep neural networks (DNNs) to various critical software tasks, e.g., bug detection and clone detection. This success can be greatly attributed to the utilization of massive high-quality training data, and in practice, data augmentation, which is a technique used to produce additional training data, has been widely adopted in various domains, such as computer vision. However, in source code learning, data augmentation has not been extensively studied, and existing practice is limited to simple syntax-preserved methods, such as code refactoring. Essentially, source code is often represented in two ways, namely, sequentially as text data and structurally as graph data, when it is used as training data in source code learning. Inspired by these analogy relations, we take an early step to investigate whether data augmentation methods that are originally used for text and graphs are effective in improving the training quality of source code learning. To that end, we first collect and categorize data augmentation methods in the literature. Second, we conduct a comprehensive empirical study on four critical tasks and 11 DNN architectures to explore the effectiveness of 12 data augmentation methods (including code refactoring and 11 other methods for text and graph data). Our results identify the data augmentation methods that can produce more accurate and robust models for source code learning, including those based on mixup (e.g., SenMixup for texts and Manifold-Mixup for graphs), and those that slightly break the syntax of source code (e.g., random swap and random deletion for texts)."}, "cited_paper_content": {"title": "Robustness And Generalization", "abstract": "We derive generalization bounds for learning algorithms based on their robustness: the property that if a testing sample is\"similar\"to a training sample, then the testing error is close to the training error. This provides a novel approach, different from the complexity or stability arguments, to study generalization of learning algorithms. We further show that a weak notion of robustness is both sufficient and necessary for generalizability, which implies that robustness is a fundamental property for learning algorithms to work."}, "keywords": ["multiple trained code", "adversarial robustness"], "citation_intent": "method"}
{"citing_id": "2304.03797v1", "cited_id": "1804.05088", "section_title": "Conclusion", "citation": "For example, while Spanish hashtags are more entertainment-oriented, Catalan hashtags tend to be more political; this is consistent with prior work that shows Catalan is used online to mark political identity #REFR .", "text_before_citation": ["Effects are greatest when C y is 2-3 hours ahead (i.e. much further east) of C x .", "Figure S2 : Fraction of unique hashtags associated with sample of languages for each of four macro-topics, sorted left to right from largest to smallest proportion of political hashtags.", "Finnish and German hashtags are largely political, while Russian, Spanish, and Hungarian hashtags tend to be more entertainment-oriented.", "Portuguese hashtags are much more focused on entertainment and sports compared to politics.", "This figure also illustrates how language choice can be a meaningful sociolinguistic variable."], "text_after_citation": ["Table S5 : Summary of locus-specific causal estimates for each topic macro-category.", "Out of the 464 possible loci from 232 MCPs, ATT estimates were only calculated for a subset of loci with sufficient data.", "The remaining columns show the percent of estimates where multilinguals have a significantly positive, negative, and no significant effect (based on heteroskedasticityconsistent standard error estimation with p < 0.05.", "Figure S3 : Distributions of tweets and hashtags assigned to all 50 topics with multilingual contextualized topic modeling.", "The top plot shows the fraction of all tweets assigned to each topic."], "citing_paper_content": {"title": "Bridging Nations: Quantifying The Role Of Multilinguals In Communication On Social Media", "abstract": "Social media enables the rapid spread of many kinds of information, from memes to social movements. However, little is known about how information crosses linguistic boundaries. We apply causal inference techniques on the European Twitter network to quantify multilingual users' structural role and communication influence in cross-lingual information exchange. Overall, multilinguals play an essential role; posting in multiple languages increases betweenness centrality by 13%, and having a multilingual network neighbor increases monolinguals' odds of sharing domains and hashtags from another language 16-fold and 4-fold, respectively. We further show that multilinguals have a greater impact on diffusing information less accessible to their monolingual compatriots, such as information from faraway countries and content about regional politics, nascent social movements, and job opportunities. By highlighting information exchange across borders, this work sheds light on a crucial component of how information and ideas spread around the world."}, "cited_paper_content": {"title": "S\\'I O No, Qu\\`E Penses? Catalonian Independence And Linguistic Identity On Social Media", "abstract": "Political identity is often manifested in language variation, but the relationship between the two is still relatively unexplored from a quantitative perspective. This study examines the use of Catalan, a language local to the semi-autonomous region of Catalonia in Spain, on Twitter in discourse related to the 2017 independence referendum. We corroborate prior findings that pro-independence tweets are more likely to include the local language than anti-independence tweets. We also find that Catalan is used more often in referendum-related discourse than in other contexts, contrary to prior findings on language variation. This suggests a strong role for the Catalan language in the expression of Catalonian political identity."}, "keywords": ["Spanish hashtags", "entertainment-oriented, Catalan hashtags"], "citation_intent": "result"}
{"citing_id": "2303.11906v2", "cited_id": "1409.0575", "section_title": "Implementation Details", "citation": "We validate the performance of our algorithm on the Im-ageNet dataset #REFR , which consists of 1.2M training images and 50, 000 validation images.", "text_before_citation": [], "text_after_citation": ["We take 16 batches of data to perform the reconstruction of PTQ.", "We fix the batch size to 256 for ResNet-18 and MobileNetV2, and 128 for ResNet-50.", "In addition, our experimental analysis of the expanded batch size is shown in Sec. 4.3. Our data preprocessing follows #OTHEREFR as everyone does.", "We use the Pytorch library #OTHEREFR to complete our algorithm.", "We perform our experiments on an Nvidia Tesla A100 as well as an Intel(R) Xeon(R) Platinum 8336C CPU."], "citing_paper_content": {"title": "Solving Oscillation Problem In Post-Training Quantization Through A Theoretical Perspective", "abstract": "Post-training quantization (PTQ) is widely regarded as one of the most efficient compression methods practically, benefitting from its data privacy and low computation costs. We argue that an overlooked problem of oscillation is in the PTQ methods. In this paper, we take the initiative to explore and present a theoretical proof to explain why such a problem is essential in PTQ. And then, we try to solve this problem by introducing a principled and generalized framework theoretically. In particular, we first formulate the oscillation in PTQ and prove the problem is caused by the difference in module capacity. To this end, we define the module capacity (ModCap) under data-dependent and data-free scenarios, where the differentials between adjacent modules are used to measure the degree of oscillation. The problem is then solved by selecting top-k differentials, in which the corresponding modules are jointly optimized and quantized. Extensive experiments demonstrate that our method successfully reduces the performance drop and is generalized to different neural networks and PTQ methods. For example, with 2/4 bit ResNet-50 quantization, our method surpasses the previous state-of-the-art method by 1.9%. It becomes more significant on small model quantization, e.g. surpasses BRECQ method by 6.61% on MobileNetV2 \u00d70.5."}, "cited_paper_content": {"title": "Imagenet Large Scale Visual Recognition Challenge", "abstract": "The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements."}, "keywords": ["1.2M training images", "Im-ageNet dataset"], "citation_intent": "method"}
{"citing_id": "2303.01812v1", "cited_id": "1904.03814", "section_title": "Main Results", "citation": "To illustrate the difficulty of training a UniKW-AT model, we also ran baseline experiments using a common TC-ResNet8 #REFR model.", "text_before_citation": ["The core results of our work on the GSCV1 and Audioset can be seen in Table 2 ."], "text_after_citation": ["The results show that even though TC-ResNet8 can achieve excellent performance on GSCV1 (96.72 Acc), it fails to provide meaningful performance on AT (8.67 mAP).", "Note that TC-ResNet8's performance on GSCV1 improves against the publicly reported result (96.10 \u2192 96.72) due to UniKW-AT training, where 60h of \"noise\" samples from Audioset enhance the model's robustness to false alarms.", "As we can see, our proposed UiT-XS achieves competitive results compared to the previous Mo-bileNetV2 (MBv2) based method for UniKW-AT, as well as other works in the literature regarding KWS and AT performance."], "citing_paper_content": {"title": "Unified Keyword Spotting And Audio Tagging On Mobile Devices With Transformers", "abstract": "Keyword spotting (KWS) is a core human-machine-interaction front-end task for most modern intelligent assistants. Recently, a unified (UniKW-AT) framework has been proposed that adds additional capabilities in the form of audio tagging (AT) to a KWS model. However, previous work did not consider the real-world deployment of a UniKW-AT model, where factors such as model size and inference speed are more important than performance alone. This work introduces three mobile-device deployable models named Unified Transformers (UiT). Our best model achieves an mAP of 34.09 on Audioset, and an accuracy of 97.76 on the public Google Speech Commands V1 dataset. Further, we benchmark our proposed approaches on four mobile platforms, revealing that the proposed UiT models can achieve a speedup of 2-6 times against a competitive MobileNetV2."}, "cited_paper_content": {"title": "Temporal Convolution For Real-Time Keyword Spotting On Mobile Devices", "abstract": "Keyword spotting (KWS) plays a critical role in enabling speech-based user interactions on smart devices. Recent developments in the field of deep learning have led to wide adoption of convolutional neural networks (CNNs) in KWS systems due to their exceptional accuracy and robustness. The main challenge faced by KWS systems is the trade-off between high accuracy and low latency. Unfortunately, there has been little quantitative analysis of the actual latency of KWS models on mobile devices. This is especially concerning since conventional convolution-based KWS approaches are known to require a large number of operations to attain an adequate level of performance. In this paper, we propose a temporal convolution for real-time KWS on mobile devices. Unlike most of the 2D convolution-based KWS approaches that require a deep architecture to fully capture both low- and high-frequency domains, we exploit temporal convolutions with a compact ResNet architecture. In Google Speech Command Dataset, we achieve more than \\textbf{385x} speedup on Google Pixel 1 and surpass the accuracy compared to the state-of-the-art model. In addition, we release the implementation of the proposed and the baseline models including an end-to-end pipeline for training models and evaluating them on mobile devices."}, "keywords": ["UniKW-AT model", "common TC-ResNet8 model"], "citation_intent": "method"}
{"citing_id": "2303.01080v1", "cited_id": "1609.02907", "section_title": "Related Works", "citation": "Scene Graph Generation: There are mainly two mainstream methods for scene graph generation: Based on context modeling or graph convolutional network (GCN) #REFR .", "text_before_citation": [], "text_after_citation": ["The first approach is focused on modeling global information and widely adapted message passing between all entities features #OTHEREFR .", "A number of them #OTHEREFR harnessed sequential models e.g., LSTM #OTHEREFR , GRU #OTHEREFR and Transformer #OTHEREFR . Chen et al.", "#OTHEREFR use two stacks of Transformer to encode global information. Zellers et al.", "#OTHEREFR uses LSTM to encode global context that informs relation prediction.", "However, merely modeling global context is not sufficient for Scene graph tasks."], "citing_paper_content": {"title": "Landmark: Language-Guided Representation Enhancement Framework For Scene Graph Generation", "abstract": "Scene graph generation (SGG) is a sophisticated task that suffers from both complex visual features and dataset long-tail problem. Recently, various unbiased strategies have been proposed by designing novel loss functions and data balancing strategies. Unfortunately, these unbiased methods fail to emphasize language priors in feature refinement perspective. Inspired by the fact that predicates are highly correlated with semantics hidden in subject-object pair and global context, we propose LANDMARK (LANguage-guiDed representation enhanceMent frAmewoRK) that learns predicate-relevant representations from language-vision interactive patterns, global language context and pair-predicate correlation. Specifically, we first project object labels to three distinctive semantic embeddings for different representation learning. Then, Language Attention Module (LAM) and Experience Estimation Module (EEM) process subject-object word embeddings to attention vector and predicate distribution, respectively. Language Context Module (LCM) encodes global context from each word embedding, which avoids isolated learning from local information. Finally, modules outputs are used to update visual representations and SGG model's prediction. All language representations are purely generated 1"}, "cited_paper_content": {"title": "Semi-Supervised Classification With Graph Convolutional Networks", "abstract": "We present a scalable approach for semi-supervised learning on graph-structured data that is based on an efficient variant of convolutional neural networks which operate directly on graphs. We motivate the choice of our convolutional architecture via a localized first-order approximation of spectral graph convolutions. Our model scales linearly in the number of graph edges and learns hidden layer representations that encode both local graph structure and features of nodes. In a number of experiments on citation networks and on a knowledge graph dataset we demonstrate that our approach outperforms related methods by a significant margin."}, "keywords": ["scene graph generation", "graph convolutional network"], "citation_intent": "method"}
{"citing_id": "2303.14005v1", "cited_id": "1405.0312", "section_title": "Datasets", "citation": "V-COCO is a subset of MS-COCO #REFR with HOI annotations, including 10,346 images (2,533 for training, 2,867 for validation and 4,946 for testing).", "text_before_citation": ["HICO-DET [3] and V-COCO #OTHEREFR are two widely-used HOI benchmarks.", "HICO-DET contains 47,776 images, with 38,118 for training and 9,658 for testing.", "There are 600 HOI categories in HICO-DET, consisting of 117 interaction classes and 80 object classes.", "Each HOI category is composed of an interaction and an object."], "text_after_citation": ["It has 80 object categories same with HICO-DET and 29 interaction categories. Evaluation metrics.", "For HICO-DET, we adopt the commonly used mAP metric [3] . Each prediction is a \u2329human, interaction, object\u232a triplet.", "A prediction is a true positive only when the human and object bounding boxes both have IoU >0.5 w.r.t. ground truth and the interaction classification result is correct.", "We evaluate the performance in two different settings following [3] .", "In the known object setting, for each HOI category, we evaluate the prediction only on the images containing the target object category."], "citing_paper_content": {"title": "Category Query Learning For Human-Object Interaction Classification", "abstract": "Unlike most previous HOI methods that focus on learning better human-object features, we propose a novel and complementary approach called category query learning. Such queries are explicitly associated to interaction categories, converted to image specific category representation via a transformer decoder, and learnt via an auxiliary image-level classification task. This idea is motivated by an earlier multi-label image classification method, but is for the first time applied for the challenging humanobject interaction classification task. Our method is simple, general and effective. It is validated on three representative HOI baselines and achieves new state-of-theart results on two benchmarks. Code will be available at https://github.com/charles-xie/CQL."}, "cited_paper_content": {"title": "Microsoft Coco: Common Objects In Context", "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model."}, "keywords": ["HOI annotations", "MS-COCO"], "citation_intent": "method"}
{"citing_id": "2304.10326v1", "cited_id": "1512.03385", "section_title": "Baseline", "citation": "For feature extraction, after examine the performance of several backbones, we select the ResNet152 #REFR as the backbone for Mask R-CNN and ResNet101 for DeepLabv3.", "text_before_citation": ["In the training session of our baseline method for panoptic segmentation, we trained a single model for instance segmentation using Mask R-CNN #OTHEREFR , and a single model for semantic segmentation using DeepLabv3 #OTHEREFR architecture."], "text_after_citation": ["No ensemble, test time tricks or post processing tricks at any level used in this baseline.", "Note that for all experiments including the baseline, the optimizer is consistent. We use Stochastic Gradient Descent(SGD) with momentum.", "For instance segmentation, a multi-step learning scheduler (LR) with initial warm-up is adopted, whereas for semantic segmentation, an additional LR is used to generate a decaying LR curve using a multiplicative factor.", "Our panoptic segmentation builds upon instance and semantic segmentation therefore the performance of these two has crucial influence on the final panoptic segmentation results.", "In the following section, we will report in detail, how we improve this baseline by studying different strategies we used for instance and semantic segmentation respectively."], "citing_paper_content": {"title": "Joint Coco And Mapillary Workshop At Iccv 2019: Coco Panoptic Segmentation Challenge Track Technical Report: Ensembling Instance And Semantic For Panoptic Segmentation", "abstract": "In this technical report, we demonstrate our solution for the 2019 COCO panoptic segmentation task. Our method first performs instance segmentation and semantic segmentation separately, then combines the two to generate panoptic segmentation results. To enhance the performance, we add several expert models of Mask R-CNN in instance segmentation to tackle the data imbalance problem in the training data; also HTC model is adopted yielding our best instance segmentation results. In semantic segmentation, we trained several models with various backbones and use an ensemble strategy which further boosts the segmentation results. In the end, we analyze various combinations of instance and semantic segmentation, and report on their performance for the final panoptic segmentation results. Our best model achieves P Q 47.1 on 2019 COCO panoptic testdev data."}, "cited_paper_content": {"title": "Deep Residual Learning For Image Recognition", "abstract": "Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers\u20148\u00d7 deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation."}, "keywords": ["Mask R-CNN"], "citation_intent": "method"}
{"citing_id": "2304.03945v1", "cited_id": "1609.02907", "section_title": "Related Works", "citation": "The GCN model can update the presentation of the nodes in the graph by itself and their neighbors #REFR .", "text_before_citation": ["The RKT model combines past interactions with the attention mechanism to make predictions #OTHEREFR without considering the relationship between students, exercises, and KCs.", "Graph neural networks.", "A non-linear conducted graph is more complicated than a tree structure.", "Information with a graph structure can describe connections and entities in the real world in an intuitive manner #OTHEREFR .", "A unique type of neural network called a GCN operates directly on data with a graph structure."], "text_after_citation": ["Applying numerous graph convolutional layers, the GCN model can make sure that the updated nodes properly represent both the features of higher-order neighbors and the features of neighboring nodes.", "There exist some knowledge tracing models focusing on investigating the potential usage of knowledge graph structures to capture more knowledge.", "According to the works of #OTHEREFR , each node represents a student; the nodes are fixed, but the edges are dynamic. This creates a dynamic graph that changes over time.", "However, this method ignores the heterogeneous interactions between the students-exercises-KCs.", "The GIKT model is proposed to treat the exercise-skill relationship graph as a bipartite graph and then applies the embedding propagation to the GCN to integrate the exercise-skill correlation #OTHEREFR without considering student behavior."], "citing_paper_content": {"title": "Knowledge Relation Rank Enhanced Heterogeneous Learning Interaction Modeling For Neural Graph Forgetting Knowledge Tracing", "abstract": "Recently, knowledge tracing models have been applied in educational data mining such as the Self-attention knowledge tracing model(SAKT), which models the relationship between exercises and Knowledge concepts(Kcs). However, relation modeling in traditional Knowledge tracing models only considers the static question-knowledge relationship and knowledge-knowledge relationship and treats these relationships with equal importance. This kind of relation modeling is difficult to avoid the influence of subjective labeling and considers the relationship between exercises and KCs, or KCs and KCs separately. In this work, a novel knowledge tracing model, named Knowledge Relation Rank Enhanced Heterogeneous Learning Interaction Modeling for Neural Graph Forgetting Knowledge Tracing(NGFKT), is proposed to reduce the impact of the subjective labeling by calibrating the skill relation matrix and the Q-matrix and apply the Graph Convolutional Network(GCN) to model the heterogeneous interactions between students, exercises, and skills. Specifically, the skill relation matrix and Q-matrix are generated by the Knowledge Relation Importance Rank Calibration method(KRIRC). Then the calibrated skill relation matrix, Q-matrix, and the heterogeneous interactions are treated as the input of the GCN to generate the exercise embedding and skill embedding. Next, the exercise embedding, skill embedding, item difficulty, and contingency table are incorporated to generate an exercise relation matrix as the inputs of the Position-Relation-Forgetting attention mechanism. Finally, the Position-Relation-Forgetting attention mechanism is applied to make the predictions. Experiments are conducted on the two public educational datasets and results indicate that the NGFKT model outperforms all baseline models in terms of AUC, ACC, and Performance Stability(PS)."}, "cited_paper_content": {"title": "Semi-Supervised Classification With Graph Convolutional Networks", "abstract": "We present a scalable approach for semi-supervised learning on graph-structured data that is based on an efficient variant of convolutional neural networks which operate directly on graphs. We motivate the choice of our convolutional architecture via a localized first-order approximation of spectral graph convolutions. Our model scales linearly in the number of graph edges and learns hidden layer representations that encode both local graph structure and features of nodes. In a number of experiments on citation networks and on a knowledge graph dataset we demonstrate that our approach outperforms related methods by a significant margin."}, "keywords": ["graph"], "citation_intent": "background"}
{"citing_id": "2304.06788v1", "cited_id": "1009.3604", "section_title": "Introduction", "citation": "Support vector machines are utilized as decision tree generator to incorporate the spatial structure of featured class distribution #REFR .", "text_before_citation": ["Some other methods include decision trees based on fuzzy inference systems #OTHEREFR , feature space #OTHEREFR ensemble and support vector ma-chine based decision trees #OTHEREFR .", "Weights of random feature for ensemble decision tree assign weight to each individual attribute for better model variation #OTHEREFR .", "The interpretability of featured decision trees has been studied recently inorder to better understand how they make decisions #OTHEREFR .", "More literature information about decision trees can be studied from #OTHEREFR .", "The problem with all of the impurity measures, as mentioned in #OTHEREFR , is that they are different class distribution functions of the hyperplane on both of its sides and disregard the spatial structure of the regions of its class #OTHEREFR and if the labels of the data are changed without any modification in the significant features of hyperplane edges, it has no effect on the impurity metric."], "text_after_citation": ["The proximal hyperplanes generated by multisurface proximal support vector machines (MPSVM) are in such a way that each proximal hyperplane is closer to the class samples and far away from another class samples #OTHEREFR .", "Inorder to make the node pure, authors #OTHEREFR produced planes of two clustering at non-leaf node level by selecting an angle bisector. MPSVM algorithm is considered in the binary class.", "Therefore, the multiclass problem has been split into binary class considering superior class samples in one group and remaining into another class sample.", "As the tree grows bigger, the nodes reach towards its purest form and the subsequent nodes receives less samples.", "To avoid this issue, NULL space algorithm #OTHEREFR is deployed #OTHEREFR ."], "citing_paper_content": {"title": "Heterogeneous Oblique Double Random Forest", "abstract": "The decision tree ensembles use a single data feature at each node for splitting the data. However, splitting in this manner may fail to capture the geometric properties of the data. Thus, oblique decision trees generate the oblique hyperplane for splitting the data at each non-leaf node. Oblique decision trees capture the geometric properties of the data and hence, show better generalization. The performance of the oblique decision trees depends on the way oblique hyperplanes are generate and the data used for the generation of those hyperplanes. Recently, multiple classifiers have been used in a heterogeneous random forest (RaF) classifier, however, it fails"}, "cited_paper_content": {"title": "Geometric Decision Tree", "abstract": "In this paper, we present a new algorithm for learning oblique decision trees. Most of the current decision tree algorithms rely on impurity measures to assess the goodness of hyperplanes at each node while learning a decision tree in top-down fashion. These impurity measures do not properly capture the geometric structures in the data. Motivated by this, our algorithm uses a strategy for assessing the hyperplanes in such a way that the geometric structure in the data is taken into account. At each node of the decision tree, we find the clustering hyperplanes for both the classes and use their angle bisectors as the split rule at that node. We show through empirical studies that this idea leads to small decision trees and better performance. We also present some analysis to show that the angle bisectors of clustering hyperplanes that we use as the split rules at each node are solutions of an interesting optimization problem and hence argue that this is a principled method of learning a decision tree."}, "keywords": ["decision tree generator"], "citation_intent": "method"}
{"citing_id": "2303.11728v2", "cited_id": "1804.00582", "section_title": "Related Work", "citation": "We use #REFR to extract the pseudoalbedo ground truth of the inputs before the start of the training phase. Recently, Ye et al.", "text_before_citation": ["Some works #OTHEREFR have built datasets with variation in both attributes, but still have limitations in that their application is limited to outdoor scene collections #OTHEREFR or object-centric scenes #OTHEREFR (see Tab. 1).", "Intrinsic image decomposition. As Weiss et al.", "#OTHEREFR have proved, multiple images with varying illumination can provide rich guidance to intrinsic decomposition. Li et al.", "#OTHEREFR has succeeded in disentangling intrinsic components without supervision, by enforcing multiple images with varying illumination to have identical albedo. Das et al.", "#OTHEREFR has shown state-of-the-art performance by leveraging crosscolor-ratio, the illumination-invariant image gradients in an intrinsic decomposition."], "text_after_citation": ["#OTHEREFR have suggested a NeRF framework that enables intrinsic decomposition for a purpose of scene editing.", "However, did not address the problem of few-shot view synthesis under varying illumination."], "citing_paper_content": {"title": "Extremenerf: Few-Shot Neural Radiance Fields Under Unconstrained Illumination", "abstract": "In this paper, we propose a new challenge that synthesizes a novel view in a more practical environment, where the number of input multi-view images is limited and illumination variations are significant. Despite recent success, neural radiance fields (NeRF) require a massive amount of input multi-view images taken under constrained illuminations. To address the problem, we suggest ExtremeNeRF, which utilizes occlusion-aware multiview albedo consistency, supported by geometric alignment and depth consistency. We extract intrinsic image components that should be illumination-invariant across different views, enabling direct appearance comparison between the input and novel view under unconstrained illumination. We provide extensive experimental results for an evaluation of the task, using the newly built NeRF Extreme benchmark, which is the first in-the-wild novel view synthesis benchmark taken under multiple viewing directions and varying illuminations. The project page is at https://seokyeong94.github.io/ExtremeNeRF/"}, "cited_paper_content": {"title": "Learning Intrinsic Image Decomposition From Watching The World", "abstract": "Single-view intrinsic image decomposition is a highly ill-posed problem, and so a promising approach is to learn from large amounts of data. However, it is difficult to collect ground truth training data at scale for intrinsic images. In this paper, we explore a different approach to learning intrinsic images: observing image sequences over time depicting the same scene under changing illumination, and learning single-view decompositions that are consistent with these changes. This approach allows us to learn without ground truth decompositions, and to instead exploit information available from multiple images when training. Our trained model can then be applied at test time to single views. We describe a new learning framework based on this idea, including new loss functions that can be efficiently evaluated over entire sequences. While prior learning-based methods achieve good performance on specific benchmarks, we show that our approach generalizes well to several diverse datasets, including MIT intrinsic images, Intrinsic Images in the Wild and Shading Annotations in the Wild.1"}, "keywords": ["pseudoalbedo ground truth"], "citation_intent": "method"}
{"citing_id": "2305.02820v1", "cited_id": "1607.06520", "section_title": "Results", "citation": "The evaluation on attribute accuracies is similar to the automatic results, except that the accuracy of gender slightly decreases as human evaluators don't take neutral content that fits common gender prototypes as correctly reflecting the gender style, like soldier for male and baby-carer for female #REFR .", "text_before_citation": ["The CTRL model leveraged the attributes in finetuning, and achieved better control accuracy and BertScore, but it doesn't produce more diverse responses overall.", "Both Director and DASC exhibit the best controllability, and DASC produces more diverse and reasonable responses according to Distinct-2 and BertScore.", "We then show human judgement results in Table 2.", "The inter-annotator agreement for Acc G , Acc E and Acc Q are 0.65, 0.55 and 0.64 in Cohen's \u03ba, which indicates moderate to substantial agreement.", "The agreement of Interestingness and Sensibleness is 0.48 and 0.44 in Pearson's r."], "text_after_citation": ["The annotators also check questions without a question mark, which explains the slight difference in Acc Q .", "Overall, the rankings of controllability still hold according to human evaluation, with DASC performs the best.", "Baseline, Rerank and CTRL have slightly better Sensibleness than weighted decoding methods, which agrees with the commonly observed controllability-quality trade-off in previous literature #OTHEREFR .", "All controllable generation methods achieved higher Interestingness score than baseline, which supports the benefits of controllable generation.", "DASC achieved the best Interestingness given similar attributes accuracy as Director, indicating the effectiveness of attribute semantic space, which can establish better representations of attribute semantics and a more reasonable approach to compose the control attributes in weighted decoding."], "citing_paper_content": {"title": "Semantic Space Grounded Weighted Decoding For Multi-Attribute Controllable Dialogue Generation", "abstract": "Controlling chatbot utterance generation with multiple attributes such as personalities, emotions and dialogue acts is a practically useful but under-studied problem. We propose a novel controllable generation framework called DASC that possesses strong controllability with weighted decoding paradigm, while improving generation quality with the grounding in an attribute semantics space. Generation with multiple attributes is then intuitively implemented with an interpolation of multiple attribute embeddings. Experiments show that DASC can achieve state-of-the-art control accuracy in 3-aspect controllable generation tasks while also producing interesting and reasonably sensible responses, even if in an outof-distribution robustness test. Visualization of the meaningful representations learned in the attribute semantic space also supports its effectiveness. 1"}, "cited_paper_content": {"title": "Man Is To Computer Programmer As Woman Is To Homemaker? Debiasing Word Embeddings", "abstract": "The blind application of machine learning runs the risk of amplifying biases present in data. Such a danger is facing us with word embedding, a popular framework to represent text data as vectors which has been used in many machine learning and natural language processing tasks. We show that even word embeddings trained on Google News articles exhibit female/male gender stereotypes to a disturbing extent. This raises concerns because their widespread use, as we describe, often tends to amplify these biases. Geometrically, gender bias is first shown to be captured by a direction in the word embedding. Second, gender neutral words are shown to be linearly separable from gender definition words in the word embedding. Using these properties, we provide a methodology for modifying an embedding to remove gender stereotypes, such as the association between the words receptionist and female, while maintaining desired associations such as between the words queen and female. Using crowd-worker evaluation as well as standard benchmarks, we empirically demonstrate that our algorithms significantly reduce gender bias in embeddings while preserving the its useful properties such as the ability to cluster related concepts and to solve analogy tasks. The resulting embeddings can be used in applications without amplifying gender bias."}, "keywords": ["attribute accuracies", "common gender prototypes"], "citation_intent": "result"}
{"citing_id": "2304.01502v1", "cited_id": "1810.06060", "section_title": "Model-Split-Only", "citation": "A typical paradigm in this group is Vanilla Split Learning #REFR , shown in Figure 2 , which partitions the original model into two parts, namely the server model and the client model.", "text_before_citation": [], "text_after_citation": ["In this approach, the client trains the model locally using its own data and then transmits the forward-propagated data, also known as the \"Smashed Data,\" to the server.", "The server, in turn, uses this data to update the server model and sends the backward-propagated gradients back to the client for updating the client model.", "It is important to note that in this method, the clients are trained in a synchronous way, i.e., each client must wait for the previous client to complete a training round before interacting with the server.", "This paradigm can be further divided into two setups: (1) centralized setup: the client uploads the model-weight-containing files to the server or a trusted third party, and the next client downloads the model weights from that server and trains the model, as shown in Figure 2 (a).", "(2) the peer-to-peer (P2P) setup: the server sends the address of the previous trained client to the current client for transferring the model weights, as shown in Figure 2 (b)."], "citing_paper_content": {"title": "Slperf: A Unified Framework For Benchmarking Split Learning", "abstract": "Data privacy concerns has made centralized training of data, which is scattered across silos, infeasible, leading to the need for collaborative learning frameworks. To address that, two prominent frameworks emerged, i.e., federated learning (FL) and split learning (SL). While FL has established various benchmark frameworks and research libraries, SL currently lacks a unified library despite its diversity in terms of label sharing, model aggregation, and cut layer choice. This lack of standardization makes comparing SL paradigms difficult. To address this, we propose SLPerf, a unified research framework and open research library for SL, and conduct extensive experiments on four widely-used datasets under both IID and Non-IID data settings. Our contributions include a comprehensive survey of recently proposed SL paradigms, a detailed benchmark comparison of different SL paradigms in different situations, and rich engineering takeaway messages and research insights for improving SL paradigms. SLPerf can facilitate SL algorithm development and fair performance comparisons."}, "cited_paper_content": {"title": "Distributed Learning Of Deep Neural Network Over Multiple Agents", "abstract": "In domains such as health care and finance, shortage of labeled data and computational resources is a critical issue while developing machine learning algorithms. To address the issue of labeled data scarcity in training and deployment of neural network-based systems, we propose a new technique to train deep neural networks over several data sources. Our method allows for deep neural networks to be trained using data from multiple entities in a distributed fashion. We evaluate our algorithm on existing datasets and show that it obtains performance which is similar to a regular neural network trained on a single machine. We further extend it to incorporate semi-supervised learning when training with few labeled samples, and analyze any security concerns that may arise. Our algorithm paves the way for distributed training of deep neural networks in data sensitive applications when raw data may not be shared directly."}, "keywords": ["Vanilla Split Learning"], "citation_intent": "background"}
{"citing_id": "2304.12690v1", "cited_id": "1606.09065", "section_title": "An Algorithm That Computes Canonical Forms Of Psd Factorizations", "citation": "For this, we now formulate the task of computing canonical forms of PSD factorizations as an optimization problem, where the variables are two sets of matrices {C x } and {D y } belonging to the positive semidefinite cone: This optimization problem is non-convex and NP-hard #REFR .", "text_before_citation": ["According to Theorem 2.2, to determine whether or not |\u03c8 can produce P by local operations only, one needs to characterize all possible canonical forms of PSD factorizations for P , and find out whether one of them is consistent with |\u03c8 .", "In the current section, we propose an algorithm to compute canonical forms of PSD factorizations for an arbitrary correlation P ."], "text_after_citation": ["Since the constraints and objective function are polynomial functions of the variables, some established methods such as quantifier elimination #OTHEREFR and Lasserre's hierarchy #OTHEREFR could be employed to solve the optimization problem, albeit with exponential running time.", "To obtain a heuristic approach for this problem, we fix one of the two sets of matrices to simplify it to a convex optimization problem, similar to the algorithm in #OTHEREFR for computing PSD factorization.", "We accordingly develop an algorithm that computes canonical forms of PSD factorizations by alternately optimizing over {C x } and {D y }. The pseudo-code is represented in Algorithm 1. where {D y } are supposed to be fixed.", "We use CVXPY to solve these subproblems #OTHEREFR , and the pseudocode is given in Algorithm 2. We now see a nontrivial application of this algorithm.", "Suppose we need to determine whether a pure state |\u03c8 can generate a target classical correlation P ."], "citing_paper_content": {"title": "The Generations Of Classical Correlations Via Quantum Schemes", "abstract": "Suppose two separated parties, Alice and Bob, share a bipartite quantum state or a classical correlation called a seed, and they try to generate a target classical correlation by performing local quantum or classical operations on the seed, i.e., any communications are not allowed. We consider the following fundamental problem about this setting: whether Alice and Bob can use a given seed to generate a target classical correlation. We show that this problem has rich mathematical structures. Firstly, we prove that even if the seed is a pure bipartite state, the above decision problem is already NP-hard and a similar conclusion can also be drawn when the seed is also a classical correlation, implying that this problem is hard to solve generally. Furthermore, we prove that when the seed is a pure quantum state, solving the problem is equivalent to finding out whether the target classical correlation has some canonical form of positive semi-definite factorizations that matches the seed pure state, revealing an interesting connection between the current problem and optimization theory. Based on this observation and other insights, we give several necessary conditions where the seed pure state has to satisfy to generate the target classical correlation, and it turns out that these conditions can also be generalized to the case that the seed is a mixed quantum state. Lastly, since canonical forms of positive semi-definite factorizations play a crucial role in solving the problem, we develop an algorithm that can compute them for an arbitrary classical correlation, which has decent performance on the cases we test."}, "cited_paper_content": {"title": "The Complexity Of Positive Semidefinite Matrix Factorization", "abstract": "Let $A$ be an $m\\times n$ matrix with nonnegative real entries. The PSD rank of $A$ is the smallest integer $k$ for which there exist $k\\times k$ real PSD matrices $B_1,\\ldots,B_m$, $C_1,\\ldots,C_n$ satisfying $A(i|j)=\\operatorname{tr}(B_iC_j)$ for all $i,j$. This paper determines the computational complexity status of the PSD rank. Namely, we show that the problem of computing this function is polynomial-time equivalent to the existential theory of the reals."}, "keywords": ["PSD factorizations"], "citation_intent": "background"}
{"citing_id": "2304.10038v1", "cited_id": "1801.01423", "section_title": "Accuracy And Forgetting Rate Results And Analysis", "citation": "The performance of the multi-head method HAT #REFR using task-id prediction is only 62.68, which is lower than many other baselines.", "text_before_citation": ["Its average accuracy over Accuracy performances of the baselines and our method MORE with smaller memory sizes.", "We reduce the size of the memory buffer by half.", "The new sizes are 100, 1000, 1000 for CIFAR10, CIFAR100, and Tiny-ImageNet.", "Numbers in bold are the best results in each column.", "all the datasets is 68.25, which is still poorer than our method's performance of 71.59."], "text_after_citation": ["Its performance is particularly low in experiments where the number of classes per task is small.", "For instance, its accuracy on C100-20T is 56.72, much lower than our method of 70.53 trained based on OOD detection. Accuracy with Smaller Memory Sizes.", "For all the datasets, we run additional experiments with half of the original memory size and show that our method is even stronger with a smaller memory.", "The new memory sizes are 100, 1000, and 1000 for CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively.", "Table 6 shows that MORE has experienced almost no performance drop with the reduced memory size while the memory-based baselines suffer from major performance reduction."], "citing_paper_content": {"title": "Open-World Continual Learning: Unifying Novelty Detection And Continual Learning", "abstract": "As AI agents are increasingly used in the real open world with unknowns or novelties, they need the ability to (1) recognize objects that (i) they have learned and (ii) detect items that they have not seen or learned before, and (2) learn the new items incrementally to become more and more knowledgeable and powerful. (1) is called novelty detection or out-of-distribution (OOD) detection and (2) is called class incremental learning (CIL), which is a setting of continual learning (CL). In existing research, OOD detection and CIL are regarded as two completely different problems. This paper theoretically proves that OOD detection actually is necessary for CIL. We first show that CIL can be decomposed into two sub-problems: within-task prediction (WP) and task-id prediction (TP). We then prove that TP is correlated with OOD detection. The key theoretical result is that regardless of whether WP and OOD detection (or TP) are defined explicitly or implicitly by a CIL algorithm, good WP and good OOD detection are necessary and sufficient conditions for good CIL, which unifies novelty or OOD detection and continual learning (CIL, in particular). A good CIL algorithm based on our theory"}, "cited_paper_content": {"title": "Overcoming Catastrophic Forgetting With Hard Attention To The Task", "abstract": "Catastrophic forgetting occurs when a neural network loses the information learned in a previous task after training on subsequent tasks. This problem remains a hurdle for artificial intelligence systems with sequential learning capabilities. In this paper, we propose a task-based hard attention mechanism that preserves previous tasks' information without affecting the current task's learning. A hard attention mask is learned concurrently to every task, through stochastic gradient descent, and previous masks are exploited to condition such learning. We show that the proposed mechanism is effective for reducing catastrophic forgetting, cutting current rates by 45 to 80%. We also show that it is robust to different hyperparameter choices, and that it offers a number of monitoring capabilities. The approach features the possibility to control both the stability and compactness of the learned knowledge, which we believe makes it also attractive for online learning or network compression applications."}, "keywords": ["task-id prediction"], "citation_intent": "result"}
{"citing_id": "2304.10893v1", "cited_id": "1511.04136", "section_title": "Corpus Collection", "citation": "Four members of our team write down the corresponding sentences and paragraphs with their own writing habits and imagination by observing the images in UA-DETRAC #REFR .", "text_before_citation": ["All volunteers are welleducated and have adequate English linguistic knowledge.", "Thirdly, we insert the target vehicles with their attributes into these patterns by our sentence auto-generation framework.", "As the sample in Fig.", "1 presents, the corpus with complex context indicates narrative long sentences or paragraphs with persons' subjective emotions and imagination.", "Instead of generating a corpus with simple context by combining labor and computers, a corpus with complex context is made by human beings only."], "text_after_citation": [], "citing_paper_content": {"title": "Findvehicle And Vehiclefinder: A Ner Dataset For Natural Language-Based Vehicle Retrieval And A Keyword-Based Cross-Modal Vehicle Retrieval System", "abstract": "Natural language (NL) based vehicle retrieval is a task aiming to retrieve a vehicle that is most consistent with a given NL query from among all candidate vehicles. Because NL query can be easily obtained, such a task has a promising prospect in building an interactive intelligent traffic system (ITS). Current solutions mainly focus on extracting both text and image features and mapping them to the same latent space to compare the similarity. However, existing methods usually use dependency analysis or semantic role-labelling techniques to find keywords related to vehicle attributes. These techniques may require a lot of pre-processing and post-processing work, and also suffer from extracting the wrong keyword when the NL query is complex. To tackle these problems and simplify, we borrow the idea from named entity recognition (NER) and construct FindVehicle, a NER dataset in the traffic domain. It has 42.3k labelled NL descriptions of vehicle tracks, containing information such as the location, orientation, type and colour of the vehicle. FindVehicle also adopts both overlapping entities and fine-grained entities to meet further requirements. To verify its effectiveness, we propose a baseline NL-based vehicle retrieval model called VehicleFinder. Our experiment shows that by using text encoders pre-trained by FindVehicle, VehicleFinder achieves 87.7% precision and 89.4% recall when retrieving a target vehicle by text command on our homemade dataset based on UA-DETRAC [1]. From loading the command into VehicleFinder to identifying whether the target vehicle is consistent with the command, the time cost is 279.35 ms on one ARM v8.2 CPU and 93.72 ms on one RTX A4000 GPU, which is much faster than the Transformer-based system. The dataset is open-source via the link https://github.com/GuanRunwei/FindVehicle, and the implementation can be found via the link https://github.com/GuanRunwei/VehicleFinder-CTIM."}, "cited_paper_content": {"title": "Ua-Detrac: A New Benchmark And Protocol For Multi-Object Detection And Tracking", "abstract": "Effective multi-object tracking (MOT) methods have been developed in recent years for a wide range of applications including visual surveillance and behavior understanding. Existing performance evaluations of MOT methods usually separate the tracking step from the detection step by using one single predefined setting of object detection for comparisons. In this work, we propose a new University at Albany DEtection and TRACking (UA-DETRAC) dataset for comprehensive performance evaluation of MOT systems especially on detectors. The UA-DETRAC benchmark dataset consists of 100 challenging videos captured from real-world traffic scenes (over 140,000 frames with rich annotations, including illumination, vehicle type, occlusion, truncation ratio, and vehicle bounding boxes) for multi-object detection and tracking. We evaluate complete MOT systems constructed from combinations of state-of-the-art object detection and tracking methods. Our analysis shows the complex effects of detection accuracy on MOT system performance. Based on these observations, we propose effective and informative evaluation metrics for MOT systems that consider the effect of object detection for comprehensive performance analysis."}, "keywords": ["corresponding sentences", "UA-DETRAC"], "citation_intent": "background"}
{"citing_id": "2305.02645v1", "cited_id": "1907.01341", "section_title": "Edge Enhancement", "citation": "After the first prediction by a monocular depth estimation #REFR , which contains sharp edges on object borders and small details, depth maps get more blurred as compared to the previous iterations.", "text_before_citation": ["In this section, we address the issue of depth blurriness, which occurs when edges and high-frequency information are lost during the TTT process. As an example of this problem, in Fig.", "2 , the depth map changes with each iteration after 0, 1, and 2 epochs."], "text_after_citation": ["During the process, we aim to preserve the high-frequency information that is present on the initial depth maps (Fig. 2,  2nd row) .", "We achieve this by using gradient-based loss during training between the original depth map and the current depth estimation (at each iteration).", "The proposed method can be applied to both monocular and stereo-based depth setups. In the following, we describe two gradient-based losses. Multi-scale Gradient Loss.", "We propose to adopt a multiscale gradient loss #OTHEREFR as the scale-invariant gradient loss. This discrete scale-invariant gradient g is defined as:", "EQUATION"], "citing_paper_content": {"title": "Edge-Aware Consistent Stereo Video Depth Estimation", "abstract": "Video depth estimation is crucial in various applications, such as scene reconstruction and augmented reality. In contrast to the naive method of estimating depths from images, a more sophisticated approach uses temporal information, thereby eliminating flickering and geometrical inconsistencies. We propose a consistent method for dense video depth estimation; however, unlike the existing monocular methods, ours relates to stereo videos. This technique overcomes the limitations arising from the monocular input. As a benefit of using stereo inputs, a left-right consistency loss is introduced to improve the performance. Besides, we use SLAM-based camera pose estimation in the process. To address the problem of depth blurriness during test-time training (TTT), we present an edge-preserving loss function that improves the visibility of fine details while preserving geometrical consistency. We show that our edge-aware stereo video model can accurately estimate the dense depth maps."}, "cited_paper_content": {"title": "Towards Robust Monocular Depth Estimation: Mixing Datasets For Zero-Shot Cross-Dataset Transfer", "abstract": "The success of monocular depth estimation relies on large and diverse training sets. Due to the challenges associated with acquiring dense ground-truth depth across different environments at scale, a number of datasets with distinct characteristics and biases have emerged. We develop tools that enable mixing multiple datasets during training, even if their annotations are incompatible. In particular, we propose a training objective that is invariant to changes in depth range and scale. Armed with this objective, we explore an abundant source of training data: 3D films. We demonstrate that despite pervasive inaccuracies, 3D films constitute a useful source of data that is complementary to existing training sets. We evaluate the presented approach on diverse datasets, focusing on zero-shot cross-dataset transfer: testing the generality of the learned model by evaluating it on datasets that were not seen during training. The experiments confirm that mixing data from complementary sources yields improved depth estimates, particularly on previously unseen datasets. Some results are shown in the supplementary video: this https URL"}, "keywords": ["monocular depth estimation"], "citation_intent": "method"}
{"citing_id": "2303.06879v1", "cited_id": "1802.04431", "section_title": "Related Work", "citation": "Based on LSTMs, NASA Jet Propulsion Laboratory designs an unsupervised nonparametric anomaly thresholding ap-proach for spacecraft anomaly detection #REFR ; Park et al.", "text_before_citation": ["Recently, deep neural network architectures have achieved leading performance on various time series anomaly detection tasks.", "Especially, LSTMs and RNNs can effectively process time-series data, capturing valuable historical information for future prediction."], "text_after_citation": ["#OTHEREFR propose variational auto-encoders that fuse signals and reconstruct expected distribution.", "Most telemetry data are multi-dimensional due to the interrelation of components in the satellite's internal structure. As illustrated in Fig.", "1 , each curve corresponds to a variable (or channel) in spacecraft multivariate time series, and an anomaly in one channel of the spacecraft also causes abnormality in other channels.", "For online anomaly de- #OTHEREFR utilize stochastic variable connection and planar normalizing flow for multivariate time series anomaly detection, which performs well in datasets of NASA. Zhao et al.", "#OTHEREFR utilizes the graph attention method to capture temporal and variable dependency for anomaly detection. Li et al."], "citing_paper_content": {"title": "Spacecraft Anomaly Detection With Attention Temporal Convolution Networks", "abstract": "Spacecraft faces various situations when carrying out exploration missions in complex space, thus monitoring the anomaly status of spacecraft is crucial to the development of the aerospace industry. The time series telemetry data generated by on-orbit spacecraft contains important information about the status of spacecraft. However, traditional domain knowledge-based spacecraft anomaly detection methods are not effective due to high dimensionality and complex correlation among variables. In this work, we propose an anomaly detection framework for spacecraft multivariate time-series data based on temporal convolution networks (TCNs). First, we employ dynamic graph attention to model the complex correlation among variables and time series. Second, temporal convolution networks with parallel processing ability are used to extract multidimensional features for the downstream prediction task. Finally, many potential anomalies are detected by the best threshold. Experiments on real NASA SMAP/MSL spacecraft datasets show the superiority of our proposed model with respect to state-of-the-art methods."}, "cited_paper_content": {"title": "Detecting Spacecraft Anomalies Using Lstms And Nonparametric Dynamic Thresholding", "abstract": "As spacecraft send back increasing amounts of telemetry data, improved anomaly detection systems are needed to lessen the monitoring burden placed on operations engineers and reduce operational risk. Current spacecraft monitoring systems only target a subset of anomaly types and often require costly expert knowledge to develop and maintain due to challenges involving scale and complexity. We demonstrate the effectiveness of Long Short-Term Memory (LSTMs) networks, a type of Recurrent Neural Network (RNN), in overcoming these issues using expert-labeled telemetry anomaly data from the Soil Moisture Active Passive (SMAP) satellite and the Mars Science Laboratory (MSL) rover, Curiosity. We also propose a complementary unsupervised and nonparametric anomaly thresholding approach developed during a pilot implementation of an anomaly detection system for SMAP, and offer false positive mitigation strategies along with other key improvements and lessons learned during development."}, "keywords": ["spacecraft anomaly detection"], "citation_intent": "method"}
{"citing_id": "2303.10800v1", "cited_id": "2003.11539", "section_title": "I. Introduction Most Modern Synthetic Aperture Radar (Sar) Automatic", "citation": "SSL has shown great promise for transfer and few-shot learning in the natural imagery domain #REFR , but has not been studied widely in the context of SAR ATR.", "text_before_citation": ["In Stage 2, for an arbitrary downstream (M-way, N-shot) few-shot learning (FSL) task we train a lightweight classifier using the global model as a fixed feature extractor.", "A key part of our methodology is how we train the global model.", "Importantly, we do not assume to have labeled data in this step, which eliminates the option for a meta-learning approach #OTHEREFR , #OTHEREFR .", "To learn quality representations we instead aggregate publicly available data from recent papers and leverage a modern unsupervised learning paradigm called Self-Supervised Learning (SSL).", "Specifically, we use the SimCLR #OTHEREFR method to represent the SSL class of algorithms in this work."], "text_after_citation": ["Our hypothesis is that these modern algorithms can learn a highly flexible / expressive / useful / transferable feature extractor from unlabeled SAR data sourced from different sensors, imaging modes, polarizations, resolutions, and target types.", "Further, the unlabeled data does not necessarily have to represent the anticipated downstream conditions.", "This is contrary to several lines of concurrent research whose goal is to generate highfidelity synthetic imagery as close to the downstream task conditions as possible for supervised pre-training #OTHEREFR .", "The second key part of our methodology is how we improve robustness to OOD inputs.", "This is accomplished in the classifier training stage and given our stated operating environment it is nearly a \"free-lunch.\" Practically, we can use the SSL model's pretraining dataset as an OOD dataset for outlier arXiv:2303.10800v1 [cs.CV] 20 Mar 2023 exposure (OE) training #OTHEREFR , #OTHEREFR ."], "citing_paper_content": {"title": "A Global Model Approach To Robust Few-Shot Sar Automatic Target Recognition", "abstract": "In real-world scenarios, it may not always be possible to collect hundreds of labeled samples per class for training deep learning-based SAR Automatic Target Recognition (ATR) models. This work specifically tackles the few-shot SAR ATR problem, where only a handful of labeled samples may be available to support the task of interest. Our approach is composed of two stages. In the first, a global representation model is trained via self-supervised learning on a large pool of diverse and unlabeled SAR data. In the second stage, the global model is used as a fixed feature extractor and a classifier is trained to partition the feature space given the few-shot support samples, while simultaneously being calibrated to detect anomalous inputs. Unlike competing approaches which require a pristine labeled dataset for pretraining via meta-learning, our approach learns highly transferable features from unlabeled data that have littleto-no relation to the downstream task. We evaluate our method in standard and extended MSTAR operating conditions and find it to achieve high accuracy and robust out-of-distribution detection in many different few-shot settings. Our results are particularly significant because they show the merit of a global model approach to SAR ATR, which makes minimal assumptions, and provides many axes for extendability."}, "cited_paper_content": {"title": "Rethinking Few-Shot Image Classification: A Good Embedding Is All You Need?", "abstract": "The focus of recent meta-learning research has been on the development of learning algorithms that can quickly adapt to test time tasks with limited data and low computational cost. Few-shot learning is widely used as one of the standard benchmarks in meta-learning. In this work, we show that a simple baseline: learning a supervised or self-supervised representation on the meta-training set, followed by training a linear classifier on top of this representation, outperforms state-of-the-art few-shot learning methods. An additional boost can be achieved through the use of self-distillation. This demonstrates that using a good learned embedding model can be more effective than sophisticated meta-learning algorithms. We believe that our findings motivate a rethinking of few-shot image classification benchmarks and the associated role of meta-learning algorithms. Code is available at: http://github.com/WangYueFt/rfs/."}, "keywords": ["few-shot learning"], "citation_intent": "background"}
{"citing_id": "2304.02317v1", "cited_id": "1903.02551", "section_title": "A. The Structure Of Deep Jscc Framework", "citation": "Next, we consider the transmission of the encoded signal over the wireless channels after the encoding and normalization operations, i.e., y = hx + n, where the communication channels h are integrated as non-trainable layers into the whole deep JSCC framework to enable end-to-end training #REFR .", "text_before_citation": ["x \u2208 R 2b of e(\u2022, \u03b8) are denoted asx = e(s, \u03b8).", "Then we combine (reshape)x into b complexvalued symbols to form the encoded signal x. After encoding the task-relevant semantic signal", "x from the transmitted signal s, we normalize it a\u015d", "EQUATION", "such that the transmit signalx satisfies the average transmit power constraint in (1)."], "text_after_citation": ["Finally, we discuss the processing of the received signal y at the receiver to perform image recovery and classification tasks simultaneously, which is different from semantic communication frameworks dedicatedly designed for only one task in previous works #OTHEREFR - #OTHEREFR , #OTHEREFR - #OTHEREFR .", "First, the receiver applies channel equalization according to #OTHEREFR Extract semantic features via e(S, \u03b8)."], "citing_paper_content": {"title": "Semantic Communications For Image Recovery And Classification Via Deep Joint Source And Channel Coding", "abstract": "With the recent advancements in edge artificial intelligence (AI), future sixth-generation (6G) networks need to support new AI tasks such as classification and clustering apart from data recovery. Motivated by the success of deep learning, the semantic-aware and task-oriented communications with deep joint source and channel coding (JSCC) have emerged as new paradigm shifts in 6G from the conventional data-oriented communications with separate source and channel coding (SSCC). However, most existing works focused on the deep JSCC designs for one task of data recovery or AI task execution independently, which cannot be transferred to other unintended tasks. Differently, this paper investigates the JSCC semantic communications to support multi-task services, by performing the image data recovery and classification task execution simultaneously. First, we propose a new end-to-end deep JSCC framework by unifying the coding rate reduction maximization and the mean square error (MSE) minimization in the loss function. Here, the coding rate reduction maximization facilitates the learning of Part of this paper has been submitted to IEEE Global Communications Conference (GLOBECOM) 2023 [1]."}, "cited_paper_content": {"title": "Deep Learning Based End-To-End Wireless Communication Systems With Conditional Gan As Unknown Channel", "abstract": "In this article, we develop an end-to-end wireless communication system using deep neural networks (DNNs), in which DNNs are employed to perform several key functions, including encoding, decoding, modulation, and demodulation. However, an accurate estimation of instantaneous channel transfer function, i.e., channel state information (CSI), is needed in order for the transmitter DNN to learn to optimize the receiver gain in decoding. This is very much a challenge since CSI varies with time and location in wireless communications and is hard to obtain when designing transceivers. We propose to use a conditional generative adversarial net (GAN) to represent channel effects and to bridge the transmitter DNN and the receiver DNN so that the gradient of the transmitter DNN can be back-propagated from the receiver DNN. In particular, a conditional GAN is employed to model the channel effects in a data-driven way, where the received signal corresponding to the pilot symbols is added as a part of the conditioning information of the GAN. To address the curse of dimensionality when the transmit symbol sequence is long, convolutional layers are utilized. From the simulation results, the proposed method is effective on additive white Gaussian noise (AWGN) channels, Rayleigh fading channels, and frequency-selective channels, which opens a new door for building data-driven DNNs for end-to-end communication systems."}, "keywords": ["encoded signal", "wireless channels"], "citation_intent": "method"}
{"citing_id": "2304.14019v1", "cited_id": "1904.08990", "section_title": "Dataset And Models", "citation": "This convolutional network architecture introduced by #REFR processes audio sample as one-dimensional raw waveforms in time domain.", "text_before_citation": ["\u2022 Gain between \u221212 and \u22121 .", "\u2022 Noise with signal-to-noise-ratio between 10 \u22124 and 10 \u22121 .", "\u2022 Delay between 1 and 300 .", "\u2022 Bandpass filter with cutoff frequencies in the intervals , = [1400 ; 4000 ] and , = [500 ; 1200 ] for a low pass (LP) and high pass (HP) filter, respectively.", "1DCNN."], "text_after_citation": ["It includes four convolutional layers and contains 2.5\u00d710 6 parameters in total.", "We train 1DCNN on the UrbanSound8k audio event classification task and achieve a training set accuracy of 75.22 \u00b1 0.44%, and test set accuracy of 54.94 \u00b1 5.19% over a 10-fold cross-validation.", "YAMNet (Yet Another Merging NETwork).", "This is a convolutional neural network, that processes audio samples as two-dimensional logmel-spectrograms in the time-frequency domain.", "It is built on the architecture of Mobilenet V1 #OTHEREFR ."], "citing_paper_content": {"title": "Xai-Based Comparison Of Input Representations For Audio Event Classification", "abstract": "Deep neural networks are a promising tool for Audio Event Classification. In contrast to other data like natural images, there are many sensible and non-obvious representations for audio data, which could serve as input to these models. Due to their black-box nature, the effect of different input representations has so far mostly been investigated by measuring classification performance. In this work, we leverage eXplainable AI (XAI), to understand the underlying classification strategies of models trained on different input representations. Specifically, we compare two model architectures with regard to relevant input features used for Audio Event Detection: one directly processes the signal as the raw waveform, and the other takes in its time-frequency spectrogram representation. We show how relevance heatmaps obtained via \"Siren\"Layer-wise Relevance Propagation uncover representation-dependent decision strategies. With these insights, we can make a well-informed decision about the best input representation in terms of robustness and representativity and confirm that the model's classification strategies align with human requirements."}, "cited_paper_content": {"title": "End-To-End Environmental Sound Classification Using A 1D Convolutional Neural Network", "abstract": "In this paper, we present an end-to-end approach for environmental sound classification based on a 1D Convolution Neural Network (CNN) that learns a representation directly from the audio signal. Several convolutional layers are used to capture the signal's fine time structure and learn diverse filters that are relevant to the classification task. The proposed approach can deal with audio signals of any length as it splits the signal into overlapped frames using a sliding window. Different architectures considering several input sizes are evaluated, including the initialization of the first convolutional layer with a Gammatone filterbank that models the human auditory filter response in the cochlea. The performance of the proposed end-to-end approach in classifying environmental sounds was assessed on the UrbanSound8k dataset and the experimental results have shown that it achieves 89% of mean accuracy. Therefore, the propose approach outperforms most of the state-of-the-art approaches that use handcrafted features or 2D representations as input. Furthermore, the proposed approach has a small number of parameters compared to other architectures found in the literature, which reduces the amount of data required for training."}, "keywords": ["convolutional network architecture"], "citation_intent": "background"}
{"citing_id": "2304.02736v1", "cited_id": "1801.09847", "section_title": "Rgb-D Mesh Reconstruction", "citation": "We use the Open3D #REFR library and the location of the points to create a point cloud.", "text_before_citation": ["We follow the four steps below to do the same. Fig.", "3 illustrates how a mesh is generated from RGB-D information from HoloLens.", "Step 1: Creating the point cloud.", "We create a set of point clouds from each depth image in the dataset.", "For this, we first undistort the images and for each pixel in the undistorted image, we use the intrinsic parameters such as the focal length and the principal point and the depth value to calculate the 3D position of the point corresponding to that pixel in the camera's coordinate system."], "text_after_citation": ["Since the point cloud was noisy, we tried to remove them using a radius outlier filter and a statistical outlier filter from Open3D.", "Using these two filters, we were effectively able to reduce the noise.", "Depth cameras generally do not work well with highly reflective, transparent, or black surfaces and hence some of the points were very far away from or very close to the camera.", "This is an inherent problem of the depth sensor as it uses time-of-flight to measure the distance.", "To overcome this, we used the fact that the data was recorded inside a room and away from various artefacts."], "citing_paper_content": {"title": "Image Stabilization For Hololens Camera In Remote Collaboration", "abstract": "With the advent of new technologies, Augmented Reality (AR) has become an effective tool in remote collaboration. Narrow field-of-view (FoV) and motion blur can offer an unpleasant experience with limited cognition for remote viewers of AR headsets. In this article, we propose a two-stage pipeline to tackle this issue and ensure a stable viewing experience with a larger FoV. The solution involves an offline 3D reconstruction of the indoor environment, followed by enhanced rendering using only the live poses of AR device. We experiment with and evaluate the two different 3D reconstruction methods, RGB-D geometric approach and Neural Radiance Fields (NeRF), based on their data requirements, reconstruction quality, rendering, and training times. The generated sequences from these methods had smoother transitions and provided a better perspective of the environment. The geometry-based enhanced FoV method had better renderings as it lacked blurry outputs making it better than the other attempted approaches. Structural Similarity Index (SSIM) and Peak Signal to Noise Ratio (PSNR) metrics were used to quantitatively show that the rendering quality using the geometry-based en"}, "cited_paper_content": {"title": "Open3D: A Modern Library For 3D Data Processing", "abstract": "Open3D is an open-source library that supports rapid development of software that deals with 3D data. The Open3D frontend exposes a set of carefully selected data structures and algorithms in both C++ and Python. The backend is highly optimized and is set up for parallelization. Open3D was developed from a clean slate with a small and carefully considered set of dependencies. It can be set up on different platforms and compiled from source with minimal effort. The code is clean, consistently styled, and maintained via a clear code review mechanism. Open3D has been used in a number of published research projects and is actively deployed in the cloud. We welcome contributions from the open-source community."}, "keywords": ["point cloud", "Open3D library"], "citation_intent": "method"}
{"citing_id": "2303.11143v1", "cited_id": "1911.02142", "section_title": "Limitations And Future Works", "citation": "Alongside looking at pre-processing techniques #REFR , we find that an interesting direction for defensive research could be to increase the resilience of code models by designing more robust feature mapping processes.", "text_before_citation": ["Our experiments suggest that, depending on the characteristics of a given model and pair of functions, the success of an attack may be affected by factors like the initial difference in code size and CFG topology, among others.", "In this respect, it could be interesting to explore how to alternate our dead-branch addition perturbation, for example, with the insertion of dead fragments within existing blocks.", "We believe both limitations could be addressed in future work with implementation effort, whereas the main goal of this paper was to show that adversarial attacks against binary similarity systems are a concrete possibility.", "For defensive attempts, we could rely on off-the-shelf obfuscation methods (Section 3.3).", "To enhance our attacks, we could explore more complex patching implementation strategies based on binary rewriting or a modified compiler back-end."], "text_after_citation": [], "citing_paper_content": {"title": "Adversarial Attacks Against Binary Similarity Systems", "abstract": "In recent years, binary analysis gained traction as a fundamental approach to inspect software and guarantee its security. Due to the exponential increase of devices running software, much research is now moving towards new autonomous solutions based on deep learning models, as they have been showing state-of-the-art performances in solving binary analysis problems. One of the hot topics in this context is binary similarity, which consists in determining if two functions in assembly code are compiled from the same source code. However, it is unclear how deep learning models for binary similarity behave in an adversarial context. In this paper, we study the resilience of binary similarity models against adversarial examples, showing that they are susceptible to both targeted and untargeted attacks (w.r.t. similarity goals) performed by black-box and white-box attackers. In more detail, we extensively test three current state-of-the-art solutions for binary similarity against two black-box greedy attacks, including a new technique that we call Spatial Greedy, and one white-box attack in which we repurpose a gradient-guided strategy used in attacks to image classifiers."}, "cited_paper_content": {"title": "Intriguing Properties Of Adversarial Ml Attacks In The Problem Space", "abstract": "Recent research efforts on adversarial ML have investigated problem-space attacks, focusing on the generation of real evasive objects in domains where, unlike images, there is no clear inverse mapping to the feature space (e.g., software). However, the design, comparison, and real-world implications of problem-space attacks remain underexplored. This paper makes two major contributions. First, we propose a general formalization for adversarial ML evasion attacks in the problem-space, which includes the definition of a comprehensive set of constraints on available transformations, preserved semantics, absent artifacts, and plausibility. We shed light on the relationship between feature space and problem space, and we introduce the concept of side-effect features as the by-product of the inverse feature-mapping problem. This enables us to define and prove necessary and sufficient conditions for the existence of problem-space attacks. We further demonstrate the expressive power of our formalization by using it to describe several attacks from related literature across different domains. Second, building on our general formalization, we propose a novel problem-space attack on Android malware that overcomes past limitations in terms of semantics and artifacts. Experiments on a dataset with 170K Android apps from 2017 and 2018 show the practical feasibility of evading a state-of-the-art malware classifier, DREBIN, along with its hardened version, Sec-SVM. Our results demonstrate that \"adversarial-malware as a service\" is a realistic threat, as we automatically generate thousands of realistic and inconspicuous adversarial applications at scale, where on average it takes only a few minutes to generate an adversarial app. Our formalization of problem-space attacks paves the way to more principled research in this domain."}, "keywords": ["code models", "defensive research"], "citation_intent": "background"}
{"citing_id": "2304.06495v1", "cited_id": "1708.08012", "section_title": "Discussion", "citation": "We notice that our implementation of FBCSP obtains a lower score than what was reported by the authors of #REFR on the same dataset.", "text_before_citation": ["We decided to use the High-Gamma Dataset in our study even though it was not specifically created to learn embeddings or benchmark embedding approaches.", "While its high number of trials per subject (around 1000) allows for the benchmarking of various scenarios, results need to be judged in the light of the mediocre number of 14 subjects in this dataset.", "Clear advantages of this dataset are that its authors #OTHEREFR", "[2017] were successful in training deep models on it, and that it has been used in follow-up publications.", "Performance of the FBCSP baseline method."], "text_after_citation": ["We qualify this point by reminding that we use different dataset preprocessings.", "We low-pass filter to frequencies up to 40 Hz before downsampling to 128 Hz, while they allowed up to 128 Hz frequencies before downsampling to 256 Hz.", "In addition, Schirrmeister and colleagues used a larger trial time window (from 0.5 s to 4 s relative to trial start, while we have removed the first 0.5 s of every trial.", "In addition, we can not exclude that our implementation of FBCSP may differ from theirs on some details.", "Balancing the influence of subject and task classes The results of the embedding models on the cross-subject training (cf. Result 1) revealed a slightly unexpected finding."], "citing_paper_content": {"title": "An Embedding For Eeg Signals Learned Using A Triplet Loss", "abstract": "Neurophysiological time series recordings like the electroencephalogram (EEG) or local field potentials are obtained from multiple sensors. They can be decoded by machine learning models in order to estimate the ongoing brain state of a patient or healthy user. In a brain-computer interface (BCI), this decoded brain state information can be used with minimal time delay to either control an application, e.g., for communication or for rehabilitation after stroke, or to passively monitor the ongoing brain state of the subject, e.g., in a demanding work environment. A specific challenge in such decoding tasks is posed by the small dataset sizes in BCI compared to other domains of machine learning like computer vision or natural language processing. A possibility to tackle classification or regression problems in BCI despite small training data sets is through transfer learning, which utilizes data from other sessions, subjects or even datasets to train a model. In this exploratory study, we propose novel domain-specific embeddings for neurophysiological data. Our approach is based on metric learning and builds upon the recently proposed ladder loss. Using embeddings allowed us to benefit, both from the good generalisation abilities and robustness of deep learning and from the fast training of classical machine learning models for subject-specific calibration. In offline analyses using EEG data of 14 subjects, we tested the embeddings' feasibility and compared their efficiency with state-of-the-art deep learning models and conventional machine learning pipelines. For within-subject classification of motor tasks, our data suggests that embeddings significantly improve the decoding accuracy compared to an established baseline model. In between-subjects transfer scenarios, pre-trained embeddings allow for an extremely fast fine-tuning of a classifier for a novel subject, even if only a few trials per class are available, which currently can not be achieved with regular deep learning models. And in the scenarios where one subject is completely set aside for testing without specific calibration, the embeddings were not competitive. In summary, we propose the use of metric learning to obtain pre-trained embeddings of EEG-BCI data as a means to incorporate domain knowledge and to reach competitive performance on novel subjects with minimal calibration requirements."}, "cited_paper_content": {"title": "Deep Learning With Convolutional Neural Networks For Decoding And Visualization Of Eeg Pathology", "abstract": "We apply convolutional neural networks (ConvNets) to the task of distinguishing pathological from normal EEG recordings in the Temple University Hospital EEG Abnormal Corpus. We use two basic, shallow and deep ConvNet architectures recently shown to decode task-related information from EEG at least as well as established algorithms designed for this purpose. In decoding EEG pathology, both ConvNets reached substantially better accuracies (about 6% better, ~85% vs. ~79%) than the only published result for this dataset, and were still better when using only 1 minute of each recording for training and only six seconds of each recording for testing. We used automated methods to optimize architectural hyperparameters and found intriguingly different ConvNet architectures, e.g., with max pooling as the only nonlinearity. Visualizations of the ConvNet decoding behavior showed that they used spectral power changes in the delta (0-4 Hz) and theta (4-8 Hz) frequency range, possibly alongside other features, consistent with expectations derived from spectral analysis of the EEG data and from the textual medical reports. Analysis of the textual medical reports also highlighted the potential for accuracy increases by integrating contextual information, such as the age of subjects. In summary, the ConvNets and visualization techniques used in this study constitute a next step towards clinically useful automated EEG diagnosis and establish a new baseline for future work on this topic."}, "keywords": ["dataset"], "citation_intent": "result"}
{"citing_id": "2304.06233v1", "cited_id": "1912.01703", "section_title": "Model Setting", "citation": "The case studies were conducted with Adam optimizer in a Pytorch environment #REFR using an Ampere A-100 GPU.", "text_before_citation": [], "text_after_citation": ["We tuned the hyperparameters, including batch size, sequence length and learning rate, of SA-MGCRN model for each instance (i.e., each date and each data delivery data scenario).", "We adopted an early-stopping technique to prevent the model from overfitting and improve the training efficiency.", "If the trained model does not have an accuracy improvement during the last 200 training epochs, we consider the model has already converged.", "In our case study, we update (i.e., retune) SA-MGCRN 14 times (from October 23, 2019 to November 6, 2019) in total.", "The testing set is set as the travel demand on the following day after each model updates."], "citing_paper_content": {"title": "Situational-Aware Multi-Graph Convolutional Recurrent Network (Sa-Mgcrn) For Travel Demand Forecasting During Wildfires", "abstract": "Real-time forecasting of travel demand during wildfire evacuations is crucial for emergency managers and transportation planners to make timely and better-informed decisions. However, few studies focus on accurate travel demand forecasting in large-scale emergency evacuations. Therefore, this study develops and tests a new methodological framework for modeling trip generation in wildfire evacuations by using (a) large-scale GPS data generated by mobile devices and (b) state-of-the-art AI technologies. The proposed methodology aims at forecasting evacuation trips and other types of trips. Based on the travel demand inferred from the GPS data, we develop a new deep learning model, i.e., Situational-Aware Multi-Graph Convolutional Recurrent Network (SA-MGCRN), along with a model updating scheme to achieve real-time forecasting of travel demand during wildfire evacuations. The proposed methodological framework is tested in this study for a real-world case study: the 2019 Kincade Fire in Sonoma County, CA. The results show that SA-MGCRN significantly outperforms all the selected state-of-the-art benchmarks in terms of prediction performance. Our finding suggests that the most important model components of SA-MGCRN are evacuation order/warning information, proximity to fire, and population change, which are consistent with behavioral theories and empirical findings."}, "cited_paper_content": {"title": "Pytorch: An Imperative Style, High-Performance Deep Learning Library", "abstract": "Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. ::: In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. ::: We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks."}, "keywords": ["Adam optimizer", "Pytorch environment"], "citation_intent": "method"}
{"citing_id": "2304.08796v1", "cited_id": "2003.12039", "section_title": "I. Introduction", "citation": "By inferring a pixelwise displacement field #REFR from the distorted image to the distortion-free one with a deep network, a distorted document image can be unwarped with the predicted flow field at a high speed.", "text_before_citation": ["Other methods #OTHEREFR - #OTHEREFR capture multiview images of a document as an alternative.", "However, the extra hardwares and the demand for multiview images inevitably limit their usage in real applications.", "To overcome these limitations, some other methods assume a parametric model on the deformed surface of documents and optimize the model by extracting specific attributes, such as shading #OTHEREFR , #OTHEREFR , boundaries #OTHEREFR , textlines #OTHEREFR - #OTHEREFR , or texture flow #OTHEREFR .", "Nevertheless, the oversimplified parametric models usually lead to limited performance, and the optimization process introduces non-negligible computational cost.", "Recently, deep learning based solutions #OTHEREFR - #OTHEREFR have been shown as an alternative to the traditional methods, with promising performance and efficiency."], "text_after_citation": ["However, most existing algorithms #OTHEREFR - #OTHEREFR are dedicated to processing restricted document images, i.e., the input images must contain complete document boundaries (see Fig. 1(a) ).", "Once the distorted images do not include complete document boundaries (see Fig. 1(b, c) ), the rectification quality is degraded.", "To overcome this issue, DocProj #OTHEREFR divides the input images into multiple patches and infers per-patch displacement field #OTHEREFR .", "However, the stitching of the per-patch displacement field involves heavy computational cost #OTHEREFR and this method is limited to rectifying the document images without any background regions (see Fig. 1(c) ).", "For document images with some background regions, we need to manually cut out the document region first before the rectification."], "citing_paper_content": {"title": "Deep Unrestricted Document Image Rectification", "abstract": "In recent years, tremendous efforts have been made on document image rectification, but existing advanced algorithms are limited to processing restricted document images, i.e., the input images must incorporate a complete document. Once the captured image merely involves a local text region, its rectification quality is degraded and unsatisfactory. Our previously proposed DocTr, a transformer-assisted network for document image rectification, also suffers from this limitation. In this work, we present DocTr++, a novel unified framework for document image rectification, without any restrictions on the input distorted images. Our major technical improvements can be concluded in three aspects. Firstly, we upgrade the original architecture by adopting a hierarchical encoder-decoder structure for multi-scale representation extraction and parsing. Secondly, we reformulate the pixel-wise mapping relationship between the unrestricted distorted document images and the distortion-free counterparts. The obtained data is used to train our DocTr++ for unrestricted document image rectification. Thirdly, we contribute a real-world test set and metrics applicable for evaluating the rectification quality. To our best knowledge, this is the first learning-based method for the rectification of unrestricted document images. Extensive experiments are conducted, and the results demonstrate the effectiveness and superiority of our method. We hope our DocTr++ will serve as a strong baseline for generic document image rectification, prompting the further advancement and application of learningbased algorithms. The source code and the proposed dataset are publicly available at https://github.com/fh2019ustc/DocTr-Plus."}, "cited_paper_content": {"title": "Raft: Recurrent All-Pairs Field Transforms For Optical Flow", "abstract": "We introduce Recurrent All-Pairs Field Transforms (RAFT), a new deep network architecture for optical flow. RAFT extracts per-pixel features, builds multi-scale 4D correlation volumes for all pairs of pixels, and iteratively updates a flow field through a recurrent unit that performs lookups on the correlation volumes. RAFT achieves state-of-the-art performance, with strong cross-dataset generalization and high efficiency in inference time, training speed, and parameter count. Code is available \\url{https://github.com/princeton-vl/RAFT}."}, "keywords": ["distorted document image", "deep network"], "citation_intent": "method"}
{"citing_id": "2303.02615v1", "cited_id": "1711.10684", "section_title": "Ablation Study", "citation": "The evaluation was performed by applying the Residual-Unet backbone #REFR and a proposed Transformer-Encoder-based cross-attention method to image pairs from the StreeLearn dataset with large overlaps.", "text_before_citation": ["The expressive power of the Transformer-Encoder depends on the number of heads and layers. The more are used, the better the expressive power.", "However, using an excessive number might lead to overfit, and the optimal constellation, in terms of accuracy in Table 4 , is given by h = 4, l = 2.", "In particular, this constellation is a sweetspot such that increasing the number of heads or layers results in reduced accuracy.", "3D rotation encoding and training losses.", "The ablations of different rotation encodings and their corresponding training losses were evaluated and are presented in Table 6 ."], "text_after_citation": ["For the discrete formulation, in line with Cai et al.", "#OTHEREFR , the pitch and yaw angles were discretized into 360 bins \u2208 [\u2212180 \u2022 , 180 \u2022 ], and a cross-entropy loss was used to train the network.", "These results were compared to those obtained using the L 2 regression loss, as described in Eq. 2 in our scheme.", "The results in Table 6 demonstrate that our L 2 regression approach outperforms the discrete Euler angle approach proposed by Cai et al. #OTHEREFR ."], "citing_paper_content": {"title": "Estimating Extreme 3D Image Rotation With Transformer Cross-Attention", "abstract": "The estimation of large and extreme image rotation plays a key role in multiple computer vision domains, where the rotated images are related by a limited or a nonoverlapping field of view. Contemporary approaches apply convolutional neural networks to compute a 4D correlation volume to estimate the relative rotation between image pairs. In this work, we propose a cross-attentionbased approach that utilizes CNN feature maps and a Transformer-Encoder, to compute the cross-attention between the activation maps of the image pairs, which is shown to be an improved equivalent of the 4D correlation volume, used in previous works. In the suggested approach, higher attention scores are associated with image regions that encode visual cues of rotation. Our approach is end-to-end trainable and optimizes a simple regression loss. It is experimentally shown to outperform contemporary state-of-the-art schemes when applied to commonly used image rotation datasets and benchmarks, and establishes a new state-of-the-art accuracy on these datasets. We make our code publicly available 1 ."}, "cited_paper_content": {"title": "Road Extraction By Deep Residual U-Net", "abstract": "Road extraction from aerial images has been a hot research topic in the field of remote sensing image analysis. In this letter, a semantic segmentation neural network, which combines the strengths of residual learning and U-Net, is proposed for road area extraction. The network is built with residual units and has similar architecture to that of U-Net. The benefits of this model are twofold: first, residual units ease training of deep networks. Second, the rich skip connections within the network could facilitate information propagation, allowing us to design networks with fewer parameters, however, better performance. We test our network on a public road data set and compare it with U-Net and other two state-of-the-art deep-learning-based road extraction methods. The proposed approach outperforms all the comparing methods, which demonstrates its superiority over recently developed state of the arts."}, "keywords": ["proposed Transformer-Encoder-based cross-attention", "StreeLearn dataset"], "citation_intent": "method"}
{"citing_id": "2303.13243v1", "cited_id": "1412.5567", "section_title": "C. After Multi-Branch", "citation": "To reduce the number of parameters, we set a decay factor R, which means to reduce the number of convolution kernels in the first FC layer to #REFR R .", "text_before_citation": ["The last part is mainly composed of SENet #OTHEREFR , Feed Forward module.", "Unlike self-attention, SENet weights the channel dimension of speech features to obtain the interdependence between its modeling channels. Fig.", "6 shows its structure; it mainly includes Squeeze and Excitation.", "The Squeeze operation is a global average pooling, after the compression operation, the feature map is compressed into 1 \u00d7 C, where C is the number of channels, and it has a global receptive field.", "Excitation is similar to the gate mechanism in RNN, utilizing the weight matrix W \u2208 R 1\u00d7C to regenerate the weights for each feature channel."], "text_after_citation": ["Let the input of SENet be x, its formula is: 6 . SENet.", "The Global Pooling is applied to condense the input into a 1D vector, which is then processed by a bottleneck structure formed by two fully connected (FC) layers with swish activation functions.", "The output goes through a sigmoid function to be mapped to (0, 1), and then applied to the input vector using pointwise multiplications.", "EQUATION", "EQUATION"], "citing_paper_content": {"title": "Pyramid Multi-Branch Fusion Dcnn With Multi-Head Self-Attention For Mandarin Speech Recognition", "abstract": "As one of the major branches of automatic speech recognition, attention-based models greatly improves the feature representation ability of the model. In particular, the multi-head mechanism is employed in the attention, hoping to learn speech features of more aspects in different attention subspaces. For speech recognition of complex languages, on the one hand, a small head size will lead to an obvious shortage of learnable aspects. On the other hand, we need to reduce the dimension of each subspace to keep the size of the overall feature space unchanged when we increase the number of heads, which will significantly weaken the ability to represent the feature of each subspace. Therefore, this paper explores how to use a small attention subspace to represent complete speech features while ensuring many heads. In this work we propose a novel neural network architecture, namely, pyramid multi-branch fusion DCNN with multi-head self-attention. The proposed architecture is inspired by Dilated Convolution Neural Networks (DCNN), it uses multiple branches with DCNN to extract the feature of the input speech under different receptive fields. To reduce the number of parameters, every two branches are merged until all the branches are merged into one. Thus, its shape is like a pyramid rotated 90 degrees. We demonstrate that on Aishell-1, a widely used Mandarin speech dataset, our model achieves a character error rate (CER) of 6.45% on the test sets."}, "cited_paper_content": {"title": "Deep Speech: Scaling Up End-To-End Speech Recognition", "abstract": "We present a state-of-the-art speech recognition system developed using end-to-end deep learning. Our architecture is significantly simpler than traditional speech systems, which rely on laboriously engineered processing pipelines; these traditional systems also tend to perform poorly when used in noisy environments. In contrast, our system does not need hand-designed components to model background noise, reverberation, or speaker variation, but instead directly learns a function that is robust to such effects. We do not need a phoneme dictionary, nor even the concept of a \"phoneme.\" Key to our approach is a well-optimized RNN training system that uses multiple GPUs, as well as a set of novel data synthesis techniques that allow us to efficiently obtain a large amount of varied data for training. Our system, called Deep Speech, outperforms previously published results on the widely studied Switchboard Hub5'00, achieving 16.0% error on the full test set. Deep Speech also handles challenging noisy environments better than widely used, state-of-the-art commercial speech systems."}, "keywords": ["convolution kernels"], "citation_intent": "method"}
{"citing_id": "2304.14630v1", "cited_id": "1301.3781", "section_title": "Feature Extraction", "citation": "This objective is accomplished by estimating word similarity through Word2Vec #REFR to convert words into vectors.", "text_before_citation": ["This data preview also functions as an indicator to detect any deviations in the subsequent generation process.", "Furthermore, we extract the data annotation encompassing the x and y axes and the title into an SVG format file, making it available for subsequent user editing. Semantic context.", "For the semantic context extraction, we employ a two-step approach, namely keyword extraction and relevant word retrieval.", "Initially, we extract the keyword using MPNet #OTHEREFR , a pretrained model featuring a sentence transformer architecture.", "However, since the keyword may not be sufficiently explicit to inspire a concrete visual element, we also provide relevant words to stimulate users' creativity, particularly for individuals with limited design expertise, in accordance with Fig. 2 ."], "text_after_citation": ["To accomplish this, we employ the English Wikipedia Dump from November 2021, comprising a corpus of 199,430 words, each represented by a 300-dimensional vector.", "In our experiments, we discovered that some top similar words are more closely associated with proper nouns, which could decrease the recognition capacity in visualization.", "With the assumption that the occurrence frequency of the word in the corpus is related to its usage in everyday life, we integrate frequency and similarity as two criteria to rank the relevant semantic context, yielding the top 7 words as a result ( Fig. 5 (A 3 ) )."], "citing_paper_content": {"title": "Let The Chart Spark: Embedding Semantic Context Into Chart With Text-To-Image Generative Model", "abstract": "Pictorial visualization seamlessly integrates data and semantic context into visual representation, conveying complex information in a manner that is both engaging and informative. Extensive studies have been devoted to developing authoring tools to simplify the creation of pictorial visualizations. However, mainstream works mostly follow a retrieving-and-editing pipeline that heavily relies on retrieved visual elements from a dedicated corpus, which often compromise the data integrity. Text-guided generation methods are emerging, but may have limited applicability due to its predefined recognized entities. In this work, we propose ChartSpark, a novel system that embeds semantic context into chart based on text-to-image generative model. ChartSpark generates pictorial visualizations conditioned on both semantic context conveyed in textual inputs and data information embedded in plain charts. The method is generic for both foreground and background pictorial generation, satisfying the design practices identified from an empirical research into existing pictorial visualizations. We further develop an interactive visual interface that integrates a text analyzer, editing module, and evaluation module to enable users to generate, modify, and assess pictorial visualizations. We experimentally demonstrate the usability of our tool, and conclude with a discussion of the potential of using text-to-image generative model combined with interactive interface for visualization design."}, "cited_paper_content": {"title": "Efficient Estimation Of Word Representations In Vector Space", "abstract": "We propose two novel model architectures for computing continuous vector representations of words from very large data sets. The quality of these representations is measured in a word similarity task, and the results are compared to the previously best performing techniques based on different types of neural networks. We observe large improvements in accuracy at much lower computational cost, i.e. it takes less than a day to learn high quality word vectors from a 1.6 billion words data set. Furthermore, we show that these vectors provide state-of-the-art performance on our test set for measuring syntactic and semantic word similarities."}, "keywords": ["Word2Vec"], "citation_intent": "method"}
{"citing_id": "2303.10594v1", "cited_id": "1902.11237", "section_title": "Backdoor Attack", "citation": "Another pattern of trigger #REFR uses a designed perturbation that can be directly added to the training data.", "text_before_citation": ["In the domain adaptation task, model providers randomly select a portion of the source data to poison and specify a targeted category label y b .", "Blended #OTHEREFR uses a fixed image as a trigger whose size is the same as the training data x s i .", "The backdoor source data x b,s i is calculated through a mixup manner as follows,", "EQUATION", "where I tri denotes the designed trigger and \u03b1 controls the mix ratio."], "text_after_citation": ["This perturbation needs to have a bounded amplitude and the backdoor source data must be clamped to the meaningful interval as,", "EQUATION", "where \u2206 tri represents the trigger perturbation and clamp denotes the clamp operation to make sure all pixel values are in the range between 0 and 255.", "The providers replace the selected set with its poison version", "{(x b,s i , y b )} N b i=1"], "citing_paper_content": {"title": "Adaptguard: Defending Against Universal Attacks For Model Adaptation", "abstract": "Model adaptation aims at solving the domain transfer problem under the constraint of only accessing the pretrained source models. With the increasing considerations of data privacy and transmission efficiency, this paradigm has been gaining recent popularity. This paper studies the vulnerability to universal attacks transferred from the source domain during model adaptation algorithms due to the existence of the malicious providers. We explore both universal adversarial perturbations and backdoor attacks as loopholes on the source side and discover that they still survive in the target models after adaptation. To address this issue, we propose a model preprocessing framework, named AdaptGuard, to improve the security of model adaptation algorithms. AdaptGuard avoids direct use of the risky source parameters through knowledge distillation and utilizes the pseudo adversarial samples under adjusted radius to enhance the robustness. AdaptGuard is a plug-and-play module that requires neither robust pretrained models nor any changes for the following model adaptation algorithms. Extensive results on three commonly used datasets and two popular adaptation methods validate that AdaptGuard can effectively defend against universal attacks and maintain clean accuracy in the target domain simultaneously. We hope this research will shed light on the safety and robustness of transfer learning."}, "cited_paper_content": {"title": "A New Backdoor Attack In Cnns By Training Set Corruption Without Label Poisoning", "abstract": "Backdoor attacks against CNNs represent a new threat against deep learning systems, due to the possibility of corrupting the training set so to induce an incorrect behaviour at test time. To avoid that the trainer recognises the presence of the corrupted samples, the corruption of the training set must be as stealthy as possible. Previous works have focused on the stealthiness of the perturbation injected into the training samples, however they all assume that the labels of the corrupted samples are also poisoned. This greatly reduces the stealthiness of the attack, since samples whose content does not agree with the label can be identified by visual inspection of the training set or by running a pre-classification step. In this paper we present a new backdoor attack without label poisoning Since the attack works by corrupting only samples of the target class, it has the additional advantage that it does not need to identify beforehand the class of the samples to be attacked at test time. Results obtained on the MNIST digits recognition task and the traffic signs classification task show that backdoor attacks without label poisoning are indeed possible, thus raising a new alarm regarding the use of deep learning in security-critical applications."}, "keywords": ["training data"], "citation_intent": "background"}
{"citing_id": "2303.01295v1", "cited_id": "1905.04223", "section_title": "I. Introduction", "citation": "In these DNN life cycle models, development and operational stages are linked in a loop #REFR , aiming to assess and improve the accuracy the DNN according to the operating conditions.", "text_before_citation": ["Metamorphic testing #OTHEREFR and mutation testing #OTHEREFR , #OTHEREFR represent the most common strategies to evaluate the robustness of the DNN and to forecast the reliability of these systems in the operational environment.", "However, the accuracy estimated before release can substantially diverge from the one obtained during operation (operational accuracy). Retch et al.", "demonstrated how the accuracy scores of classifiers can significantly drop when completely new data are submitted #OTHEREFR .", "This problem grows up when unexpected phenomena occur in operation, such as distribution shift or label shift #OTHEREFR .", "Iterative life cycles specific for DNN -such as MLOps #OTHEREFR , #OTHEREFR -have been envisaged by companies like Google."], "text_after_citation": ["In particular, they may exploit operational data for remodeling and/or retraining the DNN before the new deployment (experimental stage), and for both the automatic evaluation of the accuracy and the automatic re-training of the models on the field (deployment stage).", "The true label of operational data collected by monitoring the DNN is generally unknown.", "This is a general issue in software testing, known as the oracle problem #OTHEREFR .", "For ML systems, according to Murphy et al., \"there is no reliable test oracle to indicate what the correct output should be for arbitrary input\" #OTHEREFR .", "The problem clearly affects also DNN accuracy estimation #OTHEREFR ."], "citing_paper_content": {"title": "Iterative Assessment And Improvement Of Dnn Operational Accuracy", "abstract": "Deep Neural Networks (DNN) are nowadays largely adopted in many application domains thanks to their human-like, or even superhuman, performance in specific tasks. However, due to unpredictable/unconsidered operating conditions, unexpected failures show up on field, making the performance of a DNN in operation very different from the one estimated prior to release. In the life cycle of DNN systems, the assessment of accuracy is typically addressed in two ways: offline, via sampling of operational inputs, or online, via pseudo-oracles. The former is considered more expensive due to the need for manual labeling of the sampled inputs. The latter is automatic but less accurate. We believe that emerging iterative industrial-strength life cycle models for Machine Learning systems, like MLOps, offer the possibility to leverage inputs observed in operation not only to provide faithful estimates of a DNN accuracy, but also to improve it through remodeling/retraining actions. We propose DAIC (DNN Assessment and Improvement Cycle), an approach which combines \"low-cost\" online pseudo-oracles and \"high-cost\" offline sampling techniques to estimate and improve the operational accuracy of a DNN in the iterations of its life cycle. Preliminary results show the benefits of combining the two approaches and integrating them in the DNN life cycle."}, "cited_paper_content": {"title": "Assuring The Machine Learning Lifecycle: Desiderata, Methods, And Challenges", "abstract": "Machine learning has evolved into an enabling technology for a wide range of highly successful applications. The potential for this success to continue and accelerate has placed machine learning (ML) at the top of research, economic and political agendas. Such unprecedented interest is fuelled by a vision of ML applicability extending to healthcare, transportation, defence and other domains of great societal importance. Achieving this vision requires the use of ML in safety-critical applications that demand levels of assurance beyond those needed for current ML applications. Our paper provides a comprehensive survey of the state-of-the-art in the assurance of ML, i.e. in the generation of evidence that ML is sufficiently safe for its intended use. The survey covers the methods capable of providing such evidence at different stages of the machine learning lifecycle, i.e. of the complex, iterative process that starts with the collection of the data used to train an ML component for a system, and ends with the deployment of that component within the system. The paper begins with a systematic presentation of the ML lifecycle and its stages. We then define assurance desiderata for each stage, review existing methods that contribute to achieving these desiderata, and identify open challenges that require further research."}, "keywords": ["DNN life cycle", "accuracy"], "citation_intent": "method"}
{"citing_id": "2303.10467v2", "cited_id": "1706.00112", "section_title": "Introduction", "citation": "Indeed, the best-known sub-packetization of MSR array code construction is exp(O(n)) while the sub-packetization of (high-rate) scalar MSR codes needs to be exp(\u0398(n log(n)), according to a lower bound in #REFR .", "text_before_citation": ["In particular, Guruswami and Wootters viewed each codeword coordinate as a vector over a subfield of RS codes' symbol field and proposed repair schemes with smaller bandwidth than the naive repair scheme. The degree of field extension is called sub-packetization.", "Building on the framework in #OTHEREFR , Tamo et al.", "#OTHEREFR proposed an RS code construction with optimal repair bandwidth, which is a scalar MSR code.", "In contrast, for an (n, k, \u2113) MSR array code, each codeword coordinate is a vector in F \u2113 q for some field F q , where the parameter \u2113 is sub-packetization.", "Compared to scalar MSR codes, their array code counterparts allow more flexibility in the code construction, which results in a smaller sub-packetization."], "text_after_citation": ["For this reason, we only consider MSR array codes in this paper."], "citing_paper_content": {"title": "Msr Codes With Linear Field Size And Smallest Sub-Packetization For Any Number Of Helper Nodes", "abstract": "An (n, k, \u2113) array code has k information coordinates and r = n \u2212 k parity coordinates, where each coordinate is a vector in F \u2113 q for some field F q. An (n, k, \u2113) MDS array code has the additional property that any k out of n coordinates suffice to recover the whole codeword. Dimakis et al. considered the problem of repairing the erasure of a single coordinate and proved a lower bound on the amount of data transmission that is needed for the repair. A minimum storage regenerating (MSR) array code with repair degree d is an MDS array code that achieves this lower bound for the repair of any single erased coordinate from any d out of n \u2212 1 remaining coordinates. An MSR code has the optimal access property if the amount of accessed data is the same as the amount of transmitted data in the repair procedure. The sub-packetization \u2113 and the field size q are of paramount importance in the MSR array code constructions. For optimal-access MSR codes, Balaji et al. proved that \u2113 \u2265 s \u2308n/s\u2309 , where s = d \u2212 k + 1. Rawat et al. showed that this lower bound is attainable for all admissible values of d when the field size is exponential in n. After that, tremendous efforts have been devoted to reducing the field size. However, till now, reduction to linear field size is only available for d \u2208 {k + 1, k + 2, k + 3} and d = n \u2212 1. In this paper, we construct the first class of explicit optimal-access MSR codes with the smallest sub-packetization \u2113 = s \u2308n/s\u2309 for all d between k + 1 and n\u22121, resolving an open problem in the survey (Ramkumar et al., Foundations and Trends in Communications and Information Theory: Vol. 19: No. 4). We further propose another class of explicit MSR code constructions (not optimal-access) with even smaller sub-packetization s \u2308n/(s+1)\u2309 for all admissible values of d, making significant progress on another open problem in the survey. Previously, MSR codes with \u2113 = s \u2308n/(s+1)\u2309 and q = O(n) were only known for d = k + 1 and d = n \u2212 1. The key insight that enables a linear field size in our construction is to reduce n r global constraints of non-vanishing determinants to O s (n) local ones, which is achieved by carefully designing the parity check matrices."}, "cited_paper_content": {"title": "Optimal Repair Of Reed-Solomon Codes: Achieving The Cut-Set Bound", "abstract": "Coding for distributed storage gives rise to a new set of problems in coding theory related to the need of reducing inter-node communication in the system. A large number of recent papers addressed the problem of optimizing the total amount of information downloaded for repair of a single failed node (the repair bandwidth) by accessing information on $d$ {\\em helper nodes}, where $k\\le d\\le n-1.$ By the so-called cut-set bound (Dimakis et al., 2010), the repair bandwidth of an $(n,k=n-r)$ MDS code using $d$ helper nodes is at least $dl/(d+1-k),$ where $l$ is the size of the node. Also, a number of known constructions of MDS array codes meet this bound with equality. In a related but separate line of work, Guruswami and Wootters (2016) studied repair of Reed-Solomon (RS) codes, showing that these codes can be repaired using a smaller bandwidth than under the trivial approach. At the same time, their work as well as follow-up papers stopped short of constructing RS codes (or any scalar MDS codes) that meet the cut-set bound with equality, which has been an open problem in coding theory. In this work we present a solution to this problem, constructing RS codes of length $n$ over the field $q^l, l=\\exp((1+o(1))n\\log n)$ that meet the cut-set bound. We also prove an almost matching lower bound on $l$, showing that the super-exponential scaling is both necessary and sufficient for achieving the cut-set bound using linear repair schemes. More precisely, we prove that for scalar MDS codes (including the RS codes) to meet this bound, the sub-packetization $l$ must satisfy $l \\ge \\exp((1+o(1)) k\\log k).$"}, "keywords": ["MSR array code", "packetization"], "citation_intent": "background"}
{"citing_id": "2305.02735v1", "cited_id": "1407.6329", "section_title": "I. Introduction", "citation": "Motivated by a possibility to cover by algebraic constructions more parameters than the parameters of linear codes, the concept of additive codes was suggested in #REFR .", "text_before_citation": ["However, the sphere-packing necessary condition on the existence of 1-perfect codes requires only the divisibility of the number 16 m \u2022 4 n of the vertices of the graph by the cardinality 1 + 6m + 3n of a radius-1 ball.", "The family of parameters satisfying this condition is much wider than the ones meeting #OTHEREFR .", "For example, all 10 Doob graphs of diameter 21, D #OTHEREFR 19) , D(2, 17), . . .", ", D(10, 1), have radius-1 balls of size 64, dividing 4 21 , while only one of them, D #OTHEREFR , falls into restrictions #OTHEREFR .", "As was shown in #OTHEREFR , the sphere-packing divisibility condition is sufficient for the existence of unrestricted 1-perfect codes in a Doob graph, where \"unrestricted\" means that the code is not implied to have any algebraic structure."], "text_after_citation": ["A Doob graph D(m, n) inherits the structure of the module", "Z 2m+n \u2032\u2032 4 \u00d7 Z 2n \u2032 2 ,", "where n = n \u2032 + n \u2032\u2032 , and a code is additive if it is additively closed, or, equivalently, forms a submodule of this module. Necessary conditions", "2m + n \u2032 + n \u2032\u2032 = 2 \u0393+2\u2206 \u2212 1 3 , 3n \u2032 + n \u2032\u2032 = 2 \u0393+\u2206 \u2212 1, 0 \u2264 n \u2032\u2032 \u2264 2 \u2206 \u2212 1, n \u2032\u2032 = 1, \u0393 \u2265 0, \u2206 \u2265 2 (2)", "were given and first examples constructed in #OTHEREFR , while constructions developed in #OTHEREFR cover the whole three-parameter family of Doob graphs D(m, n \u2032 + n \u2032\u2032 ) satisfying #OTHEREFR ."], "citing_paper_content": {"title": "Quasi-Cyclic Perfect Codes In Doob Graphs And Special Partitions Of Galois Rings", "abstract": "The Galois ring GR(4 \u2206) is the residue ring Z4[x]/(h(x)), where h(x) is a basic primitive polynomial of degree \u2206 over Z4. For any odd \u2206 larger than 1, we construct a partition of GR(4 \u2206)\\{0} into 6-subsets of type {a, b, \u2212a\u2212b, \u2212a, \u2212b, a+b} and 3-subsets of type {c, \u2212c, 2c} such that the partition is invariant under the multiplication by a nonzero element of the Teichmuller set in GR(4 \u2206) and, if \u2206 is not a multiple of 3, under the action of the automorphism group of GR(4 \u2206). As a corollary, this implies the existence of quasi-cyclic additive 1-perfect codes of index (2 \u2206 \u2212 1) in D((2 \u2206 \u2212 1)(2 \u2206 \u2212 2)/6, 2 \u2206 \u2212 1) where D(m, n) is the Doob metric scheme on Z 2m+n ."}, "cited_paper_content": {"title": "Perfect Codes In Doob Graphs", "abstract": "We study $$1$$1-perfect codes in Doob graphs $$D(m,n)$$D(m,n). We show that such codes that are linear over the Galois ring $$\\mathrm {GR}(4^2)$$GR(42) exist if and only if there exist integers $$\\gamma \\ge 0$$\u017a\u017a0 and $$\\delta >0$$\u017a>0 such that $$n=(4^{\\gamma +\\delta }-1)/3$$n=(4\u017a+\u017a-1)/3 and $$m=(4^{\\gamma +2\\delta }-4^{\\gamma +\\delta })/6$$m=(4\u017a+2\u017a-4\u017a+\u017a)/6. We also prove necessary conditions on $$(m,n)$$(m,n) for $$1$$1-perfect codes that are linear over $$Z_4$$Z4 (we call such codes additive) to exist in $$D(m,n)$$D(m,n) graphs; for some of these parameters, we show the existence of codes. For every $$m$$m and $$n$$n satisfying $$2m+n = (4^\\mu - 1)/3$$2m+n=(4\u03bc-1)/3 and $$m\\le (4^\\mu -5\\cdot 2^{\\mu -1}+1)/9$$m\u2264(4\u03bc-5\u00b72\u03bc-1+1)/9, we construct $$1$$1-perfect codes in $$D(m,n)$$D(m,n), which do not necessarily have a group structure."}, "keywords": ["additive codes"], "citation_intent": "background"}
{"citing_id": "2305.02629v1", "cited_id": "1609.05807", "section_title": "A. Fairness", "citation": "Though AC is not broadly tied to understanding people within organizations, examining fairness through this lens is highly illustrative of some of the difficulties and inherent trade-offs (cf. #REFR ) in fairness considerations within AC.", "text_before_citation": ["Fairness has no universal definition as it is a social, not psychometric, concept rooted in value judgments #OTHEREFR .", "Fairness is a subjective evaluation (e.g., justice, morality), varying across cultures and societies, and in the context of organizations such as schools, hospitals, or corporations, organizational justice, has been the predominant theoretical concept used to recognize perceptions of unfairness #OTHEREFR ."], "text_after_citation": ["Organizational justice involves three key dimensions: distributive fairness, interactional fairness, and procedural fairness #OTHEREFR .", "Distributive fairness regards the perceived fairness of outcomes and allocations of important resources (e.g., jobs).", "Interactional fairness regards how people perceive the explanations, rationales, and justifications for organizational decisions and how they perceive the interpersonal treatment they receive along the way.", "Procedural fairness regards the perceived fairness of the elements of the decision-making process.", "Procedural fairness is emphasized in the Standards because it is crucial that an assessment (e.g., ML predictions) does not generate different scores between subgroups if they have equivalent true scores."], "citing_paper_content": {"title": "Integrating Psychometrics And Computing Perspectives On Bias And Fairness In Affective Computing: A Case Study Of Automated Video Interviews", "abstract": "We provide a psychometric-grounded exposition of bias and fairness as applied to a typical machine learning pipeline for affective computing. We expand on an interpersonal communication framework to elucidate how to identify sources of bias that may arise in the process of inferring human emotions and other psychological constructs from observed behavior. Various methods and metrics for measuring fairness and bias are discussed along with pertinent implications within the United States legal context. We illustrate how to measure some types of bias and fairness in a case study involving automatic personality and hireability inference from multimodal data collected in video interviews for mock job applications. We encourage affective computing researchers and practitioners to encapsulate bias and fairness in their research processes and products and to consider their role, agency, and responsibility in promoting equitable and just systems."}, "cited_paper_content": {"title": "Inherent Trade-Offs In The Fair Determination Of Risk Scores", "abstract": "Recent discussion in the public sphere about algorithmic classification has involved tension between competing notions of what it means for a probabilistic classification to be fair to different groups. We formalize three fairness conditions that lie at the heart of these debates, and we prove that except in highly constrained special cases, there is no method that can satisfy these three conditions simultaneously. Moreover, even satisfying all three conditions approximately requires that the data lie in an approximate version of one of the constrained special cases identified by our theorem. These results suggest some of the ways in which key notions of fairness are incompatible with each other, and hence provide a framework for thinking about the trade-offs between them."}, "keywords": ["fairness considerations"], "citation_intent": "background"}
{"citing_id": "2303.11076v1", "cited_id": "1705.03550", "section_title": "Introduction", "citation": "Other datasets, such as Continuous Object Recognition (CORe50) #REFR , specifically designed for LL, still make the same assumptions to generate tasks.", "text_before_citation": ["The result is that the benchmark panorama of CL in computer vision is quite fragmented, and therefore it has become tougher to measure catastrophic forgetting and domain adaptation in a fair and homogeneous way for the many CL models we have in the literature these days.", "Furthermore, all previous evaluation protocols are designed to highlight some specific model characteristics and, as such, are generally over-simplified w.r.t. real-world data #OTHEREFR .", "For example, one of the most popular evaluation protocol for CL models in computer vision is to design different tasks to classify different (subsets of the) classes of a single dataset #OTHEREFR .", "The most prominent example is splitMNIST in which the 10 digits from MNIST #OTHEREFR are (usually) divided into 5 tasks consisting of 2 digits each.", "Similar approaches are proposed for CIFAR10 [20] , and TinyImagenet #OTHEREFR ."], "text_after_citation": ["Clearly, these protocols are not suited to detect distribution shifts, due to the high inter-task similarity.", "Consequently, catastrophic forgetting is much easier to prevent in these cases.", "Therefore the reported metrics for models evaluated in this way can be overly optimistic. Table 1 Benchmarks comparison considering only multi-dataset benchmarks.", "Columns refer to: i) supporting multiple heterogeneous tasks; ii) varying task complexity and quality; iii) evaluating curriculum strategies; iv) rigorous way to measure generalization and forgetting, and v) exactly reproducible out-of-the-box.", "In addition, we consider the coverage of class (CI) and task-incremental (TI) learning settings."], "citing_paper_content": {"title": "From Mnist To Imagenet And Back: Benchmarking Continual Curriculum Learning", "abstract": "Continual learning (CL) is one of the most promising trends in recent machine learning research. Its goal is to go beyond classical assumptions in machine learning and develop models and learning strategies that present high robustness in dynamic environments. This goal is realized by designing strategies that simultaneously foster the incorporation of new knowledge while avoiding forgetting past knowledge. The landscape of CL research is fragmented into several learning evaluation protocols, comprising different learning tasks, datasets, and evaluation metrics. Additionally, the benchmarks adopted so far are still distant from the complexity of real-world scenarios, and are usually tailored to highlight capabilities specific to certain strategies. In such a landscape, it is hard to clearly and objectively assess models and strategies. In this work, we fill this gap for CL on image data by introducing two novel CL benchmarks that involve multiple heterogeneous tasks from six image datasets, with varying levels of complexity and quality. Our aim is to fairly evaluate current state-of-the-art CL strategies on a common ground that is closer to complex real-world scenarios. We additionally structure our benchmarks so that tasks are presented in increasing and decreasing order of complexityaccording to a curriculum-in order to evaluate if current CL models are able to exploit structure across tasks. We devote particular emphasis to providing the CL community with a rigorous and reproducible evaluation protocol for measuring the ability of a model to generalize and not to forget while learning. Furthermore, we provide an extensive experimental evaluation showing that popular CL strategies,"}, "cited_paper_content": {"title": "Core50: A New Dataset And Benchmark For Continuous Object Recognition", "abstract": "Continuous/Lifelong learning of high-dimensional data streams is a challenging research problem. In fact, fully retraining models each time new data become available is infeasible, due to computational and storage issues, while naive incremental strategies have been shown to suffer from catastrophic forgetting. In the context of real-world object recognition applications (e.g., robotic vision), where continuous learning is crucial, very few datasets and benchmarks are available to evaluate and compare emerging techniques. In this work we propose a new dataset and benchmark CORe50, specifically designed for continuous object recognition, and introduce baseline approaches for different continuous learning scenarios."}, "keywords": ["Continuous Object Recognition"], "citation_intent": "background"}
{"citing_id": "2304.10317v1", "cited_id": "1506.03365", "section_title": "D. Results For Cifar10, Lsun And Ffhq", "citation": "In Figure 5 and 9 (more samples only for our method), we compare the generated images for LSUN 2 #REFR bedroom dataset for 32 \u00d7 32 images; we find that the generated images are close to state-of-the-art.", "text_before_citation": ["Since for CIFAR10, a standard metric to compare is inception score, in subfigure 6c in Figure 6 , we plot the inception scores for these methods, we find that our method achieves high inception score much earlier than both ACGD and ConOpt for the same GAN architecture as mentioned above.", "For reference, we have also plotted discriminator and generator losses in Figure 6a and 6b respectively.", "However, we must mention here that the first order methods such as ADAM never achieved comparable inception scores as those of second order methods; this observation was also found in previous works such as #OTHEREFR , #OTHEREFR , where other first order methods were also compared.", "Hence, existing literature and our comparison suggest that qualitative improvements are seen with second order methods.", "Also, as shown before, our method is fastest to train and requires less memory among second order methods."], "text_after_citation": ["Lastly, in Figure 7c and 8 (more samples only for our method), we show images generated from FFHQ dataset #OTHEREFR ; this dataset offers a significant variety in terms of ethnicity, age, viewpoint, image background, and lighting", "t \u2190 t + 1 3: D x \u2190 \u2207 x f (x t , y t ) 4: D 2 xx \u2206x \u2190 \u2207 x (D x )(x t \u2212 x t\u22121 )", "{Update second order derivative} #OTHEREFR :", "D x t \u2190 D 2", "xx \u2206x + D x {Update first order derivative} #OTHEREFR :"], "citing_paper_content": {"title": "Adaptive Consensus Optimization Method For Gans", "abstract": "We propose a second order gradient based method with ADAM and RMSprop for the training of generative adversarial networks. The proposed method is fastest to obtain similar accuracy when compared to prominent second order methods. Unlike state-of-the-art recent methods, it does not require solving a linear system, or it does not require additional mixed second derivative terms. We derive the fixed point iteration corresponding to proposed method, and show that the proposed method is convergent. The proposed method produces better or comparable inception scores, and comparable quality of images compared to other recently proposed state-of-the-art second order methods. Compared to first order methods such as ADAM, it produces significantly better inception scores. The proposed method is compared and validated on popular datasets such as FFHQ, LSUN, CIFAR10, MNIST, and Fashion MNIST for image generation tasks 1 ."}, "cited_paper_content": {"title": "Lsun: Construction Of A Large-Scale Image Dataset Using Deep Learning With Humans In The Loop", "abstract": "While there has been remarkable progress in the performance of visual recognition algorithms, the state-of-the-art models tend to be exceptionally data-hungry. Large labeled training datasets, expensive and tedious to produce, are required to optimize millions of parameters in deep network models. Lagging behind the growth in model capacity, the available datasets are quickly becoming outdated in terms of size and density. To circumvent this bottleneck, we propose to amplify human effort through a partially automated labeling scheme, leveraging deep learning with humans in the loop. Starting from a large set of candidate images for each category, we iteratively sample a subset, ask people to label them, classify the others with a trained model, split the set into positives, negatives, and unlabeled based on the classification confidence, and then iterate with the unlabeled set. To assess the effectiveness of this cascading procedure and enable further progress in visual recognition research, we construct a new image dataset, LSUN. It contains around one million labeled images for each of 10 scene categories and 20 object categories. We experiment with training popular convolutional networks and find that they achieve substantial performance gains when trained on this dataset."}, "keywords": ["generated images", "dataset"], "citation_intent": "result"}
{"citing_id": "2303.09988v1", "cited_id": "1703.10593", "section_title": "Quantitative Evaluations", "citation": "This is demonstrated by the fact that we achieve a PSNR gain of 7.98dB and a SSIM improvement of 0.07 compared to Cycle-GAN #REFR .", "text_before_citation": ["Star-Net achieves a PSNR gain of 4.12dB and an SSIM gain of 0.05 compared to Tran-sWeather #OTHEREFR .", "In the SRRS dataset, Restormer #OTHEREFR shows a PSNR of 32.24dB and SSIM of 0.96.", "Star-Net shows a PSNR of 32.41dB and SSIM of 0.97.", "As shown in Table 6 , Star-Net performs best on the quantitative evaluation of the Snow100K dataset.", "Specially, we score a PSNR of 34.79dB and a SSIM of 0.96."], "text_after_citation": ["Compared to All in One #OTHEREFR , Star-Net achieves a PSNR improvement of 8.72dB and SSIM gain of 0.08.", "Furthermore, Star-Net achieves a PSNR gain of 0.28dB and an SSIM improvement of 0.01 compared to MPRNet #OTHEREFR .", "Figure 15 shows the results of the qualitative comparison of Star-Net with several algorithms on CSD dataset.", "Although these methods achieve considerable snow removal on the CSD dataset, they still have some shortcomings.", "DesnowNet #OTHEREFR and JSTASR #OTHEREFR suffer from snow particle residual issue, and HDCW-Net #OTHEREFR suffers from image distortion."], "citing_paper_content": {"title": "Star-Net: Improving Single Image Desnowing Model With More Efficient Connection And Diverse Feature Interaction", "abstract": "Compared to other severe weather image restoration tasks, single image desnowing is a more challenging task. This is mainly due to the diversity and irregularity of snow shape, which makes it extremely difficult to restore images in snowy scenes. Moreover, snow particles also have a veiling effect similar to haze or mist. Although current works can effectively remove snow particles with various shapes, they also bring distortion to the restored image. To address these issues, we propose a novel single image desnowing network called Star-Net. First, we design a Star type Skip Connection (SSC) to establish information channels for all different scale features, which can deal with the complex shape of snow particles. Second, we present a Multi-Stage Interactive Transformer (MIT) as the base module of Star-Net, which is designed to better understand snow particle shapes and to address image distortion by explicitly modeling a variety of important image recovery features. Finally, we propose a Degenerate Filter Module (DFM) to filter the snow particle and snow fog residual in the SSC on the spatial and channel domains. Extensive experiments show that our Star-Net achieves state-of-the-art snow removal performances on three standard snow removal datasets and retains the original sharpness of the images."}, "cited_paper_content": {"title": "Unpaired Image-To-Image Translation Using Cycle-Consistent Adversarial Networks", "abstract": "Image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between an input image and an output image using a training set of aligned image pairs. However, for many tasks, paired training data will not be available. We present an approach for learning to translate an image from a source domain $X$ to a target domain $Y$ in the absence of paired examples. Our goal is to learn a mapping $G: X \\rightarrow Y$ such that the distribution of images from $G(X)$ is indistinguishable from the distribution $Y$ using an adversarial loss. Because this mapping is highly under-constrained, we couple it with an inverse mapping $F: Y \\rightarrow X$ and introduce a cycle consistency loss to push $F(G(X)) \\approx X$ (and vice versa). Qualitative results are presented on several tasks where paired training data does not exist, including collection style transfer, object transfiguration, season transfer, photo enhancement, etc. Quantitative comparisons against several prior methods demonstrate the superiority of our approach."}, "keywords": ["Cycle-GAN"], "citation_intent": "result"}
{"citing_id": "2303.00979v1", "cited_id": "1506.01186", "section_title": "Vi. Experiments", "citation": "The network is trained with an initial learning rate of 2 \u00d7 10 \u22122 and cyclic learning rate scheduling #REFR .", "text_before_citation": ["A.", "Experimental setting 1) Training environment: We used PyTorch implementation of ESPNetv2 #OTHEREFR with modifications of adding an auxiliary classification branch, and normalizing the features and classification weights.", "The models were trained and evaluated on one NVIDIA Quadro RTX 8000 with 48GB RAM."], "text_after_citation": ["2) Datasets: As source datasets, we use CamVid #OTHEREFR , Cityscapes #OTHEREFR , and Freiburg Forest #OTHEREFR .", "As target datasets, we use Greenhouse A #OTHEREFR , TUT Park, a dataset of images around a in-campus park, and Toyohashi Trail, images in unstructured mountain paths in Toyohashi Nature Trail.", "The datasets used in the experiments are summarized in Table I.", "3) Baselines: In our comparative studies, we employ baseline methods as follows.", "Training without DA As baselines without DA, we use two methods, namely supervised and ensemble."], "citing_paper_content": {"title": "Multi-Source Soft Pseudo-Label Learning With Domain Similarity-Based Weighting For Semantic Segmentation", "abstract": "This paper describes a method of domain adaptive training for semantic segmentation using multiple source datasets that are not necessarily relevant to the target dataset. We propose a soft pseudo-label generation method by integrating predicted object probabilities from multiple source models. The prediction of each source model is weighted based on the estimated domain similarity between the source and the target datasets to emphasize contribution of a model trained on a source that is more similar to the target and generate reasonable pseudo-labels. We also propose a training method using the soft pseudo-labels considering their entropy to fully exploit information from the source datasets while suppressing the influence of possibly misclassified pixels. The experiments show comparative or better performance than our previous work and another existing multi-source domain adaptation method, and applicability to a variety of target environments."}, "cited_paper_content": {"title": "Cyclical Learning Rates For Training Neural Networks", "abstract": "It is known that the learning rate is the most important hyper-parameter to tune for training deep neural networks. This paper describes a new method for setting the learning rate, named cyclical learning rates, which practically eliminates the need to experimentally find the best values and schedule for the global learning rates. Instead of monotonically decreasing the learning rate, this method lets the learning rate cyclically vary between reasonable boundary values. Training with cyclical learning rates instead of fixed values achieves improved classification accuracy without a need to tune and often in fewer iterations. This paper also describes a simple way to estimate \"reasonable bounds\" -- linearly increasing the learning rate of the network for a few epochs. In addition, cyclical learning rates are demonstrated on the CIFAR-10 and CIFAR-100 datasets with ResNets, Stochastic Depth networks, and DenseNets, and the ImageNet dataset with the AlexNet and GoogLeNet architectures. These are practical tools for everyone who trains neural networks."}, "keywords": ["network", "cyclic learning rate"], "citation_intent": "method"}
{"citing_id": "2303.10138v1", "cited_id": "1907.06488", "section_title": "Model Agnostic Tool-Use Detector", "citation": "Our use of such simple detection methods contrasts with well-studied error detection methods in NLP #REFR Fomicheva et al., 2020) as a sign that tools are reasonable model-agnostic extensions even with simple detection heuristics.", "text_before_citation": ["Sequence log-probability (SeqLogProb) is the length-normalised sequence log-probability from a trained model F (", "EQUATION", "We expect low-confidence answers are likely to be incorrect.", "Input length, as we have seen in Section 3, poses a challenge to all models irrespective of size and training objective.", "We leverage the size of the table measured by the number of rows as a simple feature to decide when to apply a tool."], "text_after_citation": [], "citing_paper_content": {"title": "Generate, Transform, Answer: Question Specific Tool Synthesis For Tabular Data", "abstract": "Tabular question answering (TQA) presents a challenging setting for neural systems by requiring joint reasoning of natural language with large amounts of semi-structured data. Unlike humans who use programmatic tools like filters to transform data before processing, language models in TQA process tables directly, resulting in information loss as table size increases. In this paper we propose Tool-Writer to generate query specific programs and detect when to apply them to transform tables and align them with the TQA model's capabilities. Focusing ToolWriter to generate rowfiltering tools improves the state-of-the-art for WikiTableQuestions and WikiSQL with the most performance gained on long tables. By investigating headroom, our work highlights the broader potential for programmatic tools combined with neural components to manipulate large amounts of structured data. Question: How many monarchs died before the age of 35? Deceased James I aged 68 Alfonso I aged 27 Sancho aged 28 James II aged 69 295 rows x 5 columns TQA Model James III aged 34 Alfonso I aged 27 Sancho aged 28 >>> table.apply(lambda row: float(row['Deceased'] .split('aged')[1]) < 35"}, "cited_paper_content": {"title": "Naver Labs Europe'S Systems For The Wmt19 Machine Translation Robustness Task", "abstract": "This paper describes the systems that we submitted to the WMT19 Machine Translation robustness task. This task aims to improve MT's robustness to noise found on social media, like informal language, spelling mistakes and other orthographic variations. The organizers provide parallel data extracted from a social media website in two language pairs: French-English and Japanese-English (in both translation directions). The goal is to obtain the best scores on unseen test sets from the same source, according to automatic metrics (BLEU) and human evaluation. We proposed one single and one ensemble system for each translation direction. Our ensemble models ranked first in all language pairs, according to BLEU evaluation. We discuss the pre-processing choices that we made, and present our solutions for robustness to noise and domain adaptation."}, "keywords": ["NLP Fomicheva", "well-studied error detection"], "citation_intent": "method"}
{"citing_id": "2304.13229v1", "cited_id": "1802.00420", "section_title": "D.7 Comparison With Standard Attacks", "citation": "While it is acknowledged that the targeted attack is a more common protocol in attacking the ImageNet dataset #REFR , however, we believe that our significant improvement on the untargeted attack is still worth noting.", "text_before_citation": ["We report experimental results with different settings in Table 19 , where RE/LE/-TAR/UNTAR represents Robust Ensemble/Less-Robust Ensemble/Targeted Attack/Untargeted Attack, respectively.", "It can be seen that, in the robust ensemble setting (RE-TAR and RE-UNTAR), our MOO achieves a similar performance compared to the baseline, while TA-MOO has a further improvement over MOO.", "The gap of SAR-All between TA-MOO and the uniform weighting strategy is 0.1% in the targeted attack setting (RE-TAR), while that in the untargeted attack setting is 1.2%.", "In the less-robust ensemble setting (LE-TAR and LE-UNTAR), the improvement of our methods over the baseline is higher than in the robust ensemble setting.", "With the gap of SAR-All between TA-MOO and the uniform strategy is 0.38% with the targeted attack setting (LE-TAR), while the gap in the untargeted setting (LE-UNTAR) is 15.22% a significantly higher."], "text_after_citation": ["We conduct an additional experiment on the EoT setting with the ImageNet dataset and report result in Table 20 .", "In this experiment, we use the robust pretrained ResNet18 model (model ID: Salman2020Do_R18) as the victim model.", "We use the standard attack setting, i.e., targeted attack with = 4/255, \u03b7 = 1/255 with 20 steps.", "It can be seen that both MOO and TA-MOO could obtain a better attack performance than the uniform strategy.", "It is a worth noting that, in the experiment on the CIFAR10/CIFAR100 datasets (i.e., Table 6 in the main paper) the dominating issue of the vertical filliping exists and prevents MOO to obtain a better performance."], "citing_paper_content": {"title": "Generating Adversarial Examples With Task Oriented Multi-Objective Optimization", "abstract": "Deep learning models, even the-state-of-the-art ones, are highly vulnerable to adversarial examples. Adversarial training is one of the most efficient methods to improve the model's robustness. The key factor for the success of adversarial training is the capability to generate qualified and divergent adversarial examples which satisfy some objectives/goals (e.g., finding adversarial examples that maximize the model losses for simultaneously attacking multiple models). Therefore, multi-objective optimization (MOO) is a natural tool for adversarial example generation to achieve multiple objectives/goals simultaneously. However, we observe that a naive application of MOO tends to maximize all objectives/goals equally, without caring if an objective/goal has been achieved yet. This leads to useless effort to further improve the goal-achieved tasks, while putting less focus on the goal-unachieved tasks. In this paper, we propose Task Oriented MOO to address this issue, in the context where we can explicitly define the goal achievement for a task. Our principle is to only maintain the goal-achieved tasks, while letting the optimizer spend more effort on improving the goal-unachieved tasks. We conduct comprehensive experiments for our Task Oriented MOO on various adversarial example generation schemes. The experimental results firmly demonstrate the merit of our proposed approach. Our code is available at https://github.com/tuananhbui89/TAMOO."}, "cited_paper_content": {"title": "Obfuscated Gradients Give A False Sense Of Security: Circumventing Defenses To Adversarial Examples", "abstract": "We identify obfuscated gradients as a phenomenon that leads to a false sense of security in defenses against adversarial examples. While defenses that cause obfuscated gradients appear to defeat optimization-based attacks, we find defenses relying on this effect can be circumvented. ::: For each of the three types of obfuscated gradients we discover, we describe indicators of defenses exhibiting this effect and develop attack techniques to overcome it. In a case study, examining all defenses accepted to ICLR 2018, we find obfuscated gradients are a common occurrence, with 7 of 8 defenses relying on obfuscated gradients. Using our new attack techniques, we successfully circumvent all 7 of them."}, "keywords": ["ImageNet", "untargeted attack"], "citation_intent": "background"}
{"citing_id": "2303.12558v2", "cited_id": "1611.00712", "section_title": "B.8 Hyperparameters", "citation": "The importance of the temperature parameters for the continuous relaxation of discrete distributions is consistent with the results of #REFR , revealing that the success of the relaxation depends on the choice of the temperature for the different latent space sizes.", "text_before_citation": ["Concerning the continuous relaxation of discrete distributions, we used a different temperature for each distribution, as #OTHEREFR", "(2017) pointed out that doing so is valuable to improve the results. We further followed the guidelines of #OTHEREFR", "(2017) to choose the interval of temperatures and did not schedule any annealing scheme (in contrast to VAE-MDPs).", "Essentially, the search reveals that the regularizer scale factors \u03b2\u00a8(defining the optimization direction) as well as the encoder and latent transition temperatures are important to improve the performance of distilled policies.", "For the encoder temperature, we found a nice spot in \u03bb \u03c6\u03b9 \" 2 {3, which provides the best performance in general, whereas the choice of \u03bb P \u03b8 and \u03b2\u00a8are (latent-) environment dependent."], "text_after_citation": ["Labeling functions.", "We used the same labeling functions as those described by #OTHEREFR .", "For completeness, we recall the labeling function used for each environment in Table 4 .", "Let \u03b8 1 , \u03b8 2 P r0, 2\u03c0s be the angles of the two rotational joints,", "\u03bb P \u03b8 t 0.1, 1 {3, 1 {2, 2 {3, 3 {5, 0.99 u \u03bb \u03c6\u03b9 t 0.1, 1 {3, 1 {2, 2 {3, 3 {5, 0.99 u \u03bb \u03c0 \u03b8 t 1 {|A|\u00b41, 1 {p|A|\u00b41q\u00a81.5 u \u03bb \u03c6 A \u03b9 t 1 {|A|\u00b41, 1 {p|A|\u00b41q\u00a81.5 u"], "citing_paper_content": {"title": "Wasserstein Auto-Encoded Mdps Formal Verification Of Efficiently Distilled Rl Policies With Many-Sided Guarantees", "abstract": "Although deep reinforcement learning (DRL) has many success stories, the largescale deployment of policies learned through these advanced techniques in safetycritical scenarios is hindered by their lack of formal guarantees. Variational Markov Decision Processes (VAE-MDPs) are discrete latent space models that provide a reliable framework for distilling formally verifiable controllers from any RL policy. While the related guarantees address relevant practical aspects such as the satisfaction of performance and safety properties, the VAE approach suffers from several learning flaws (posterior collapse, slow learning speed, poor dynamics estimates), primarily due to the absence of abstraction and representation guarantees to support latent optimization. We introduce the Wasserstein auto-encoded MDP (WAE-MDP), a latent space model that fixes those issues by minimizing a penalized form of the optimal transport between the behaviors of the agent executing the original policy and the distilled policy, for which the formal guarantees apply. Our approach yields bisimulation guarantees while learning the distilled policy, allowing concrete optimization of the abstraction and representation model quality. Our experiments show that, besides distilling policies up to 10 times faster, the latent model quality is indeed better in general. Moreover, we present experiments from a simple time-to-failure verification algorithm on the latent space. The fact that our approach enables such simple verification techniques highlights its applicability."}, "cited_paper_content": {"title": "The Concrete Distribution: A Continuous Relaxation Of Discrete Random Variables", "abstract": "The reparameterization trick enables optimizing large scale stochastic computation graphs via gradient descent. The essence of the trick is to refactor each stochastic node into a differentiable function of its parameters and a random variable with fixed distribution. After refactoring, the gradients of the loss propagated by the chain rule through the graph are low variance unbiased estimators of the gradients of the expected loss. While many continuous random variables have such reparameterizations, discrete random variables lack useful reparameterizations due to the discontinuous nature of discrete states. In this work we introduce Concrete random variables -- continuous relaxations of discrete random variables. The Concrete distribution is a new family of distributions with closed form densities and a simple reparameterization. Whenever a discrete stochastic node of a computation graph can be refactored into a one-hot bit representation that is treated continuously, Concrete stochastic nodes can be used with automatic differentiation to produce low-variance biased gradients of objectives (including objectives that depend on the log-probability of latent stochastic nodes) on the corresponding discrete graph. We demonstrate the effectiveness of Concrete relaxations on density estimation and structured prediction tasks using neural networks."}, "keywords": ["continuous relaxation", "discrete distributions"], "citation_intent": "result"}
{"citing_id": "2305.00181v1", "cited_id": "1812.07035", "section_title": "Training Objective", "citation": "Following ProHMR, we use 6D representation #REFR to model rotations and the L orth loss to force the 6D representations of the samples drawn from the distribution to be close to the orthonormal 6D representation.", "text_before_citation": ["where", "EQUATION", "For each image I, the mode \u03b8 * I of the output distribution corresponds to the transformation of z = 0.", "We do this by explicitly supervising \u03b8 * I with all the available annotations as in a standard regression framework and minimize", "EQUATION"], "text_after_citation": ["The final training objective becomes:", "EQUATION", "The loss weights are defined as: \u03bb nll = 0.001, \u03bb exp,2D = 0.001, \u03bb exp,adv = \u03bb mode,adv = 0.01, \u03bb mode,2D = 0.01, \u03bb mode,3D = 0.05, \u03bb mode,\u03b8 = 0.001, \u03bb mode,\u03b2 = 0.0005 and \u03bb orth = 0.1."], "citing_paper_content": {"title": "Tape: Temporal Attention-Based Probabilistic Human Pose And Shape Estimation", "abstract": "Reconstructing 3D human pose and shape from monocular videos is a well-studied but challenging problem. Common challenges include occlusions, the inherent ambiguities in the 2D to 3D mapping and the computational complexity of video processing. Existing methods ignore the ambiguities of the reconstruction and provide a single deterministic estimate for the 3D pose. In order to address these issues, we present a Temporal Attention based Probabilistic human pose and shape Estimation method (TAPE) that operates on an RGB video. More specifically, we propose to use a neural network to encode video frames to temporal features using an attention-based neural network. Given these features, we output a per-frame but temporally-informed probability distribution for the human pose using Normalizing Flows. We show that TAPE outperforms state-of-the-art methods in standard benchmarks and serves as an effective video-based prior for optimizationbased human pose and shape estimation."}, "cited_paper_content": {"title": "On The Continuity Of Rotation Representations In Neural Networks", "abstract": "In neural networks, it is often desirable to work with various representations of the same space. For example, 3D rotations can be represented with quaternions or Euler angles. In this paper, we advance a definition of a continuous representation, which can be helpful for training deep neural networks. We relate this to topological concepts such as homeomorphism and embedding. We then investigate what are continuous and discontinuous representations for 2D, 3D, and n-dimensional rotations. We demonstrate that for 3D rotations, all representations are discontinuous in the real Euclidean spaces of four or fewer dimensions. Thus, widely used representations such as quaternions and Euler angles are discontinuous and difficult for neural networks to learn. We show that the 3D rotations have continuous representations in 5D and 6D, which are more suitable for learning. We also present continuous representations for the general case of the n-dimensional rotation group SO(n). While our main focus is on rotations, we also show that our constructions apply to other groups such as the orthogonal group and similarity transforms. We finally present empirical results, which show that our continuous rotation representations outperform discontinuous ones for several practical problems in graphics and vision, including a simple autoencoder sanity test, a rotation estimator for 3D point clouds, and an inverse kinematics solver for 3D human poses."}, "keywords": ["model rotations"], "citation_intent": "method"}
{"citing_id": "2303.16280v1", "cited_id": "1801.03924", "section_title": "D. Metrics Of Faithfulness To The Source", "citation": "Such observations mirror the conclusion of #REFR , about the effectiveness of deep features as perceptual metrics.", "text_before_citation": ["On the contrary, the Type-2 examples suggest that I-L 2 , the L 2 on Inception latent features, might be a better measurement of semantic consistency.", "While EGSDE fails to maintain features such as background, facial expression (neural v.s.", "smile), eye movement, and prominent bone structure and produces over-generalized translations, UVC-GANv2 translations with significantly lower I-L 2 manage to preserve those features and appear more individualized.", "These examples illustrate that the pixel-wise L 2 faithfulness metric may be in poor agreement with a human judgment of image faithfulness.", "They also point to a possibility that the I-L 2 distances, based on deep features of Inception-v3, may better capture the perceived image faithfulness."], "text_after_citation": ["However, we stress again, that the main purpose of this paper is to improve the performance of the classic Cycle-GAN architecture, not the development of better faithfulness metrics.", "While this section points to a possibility of I-L 2 being a better faithfulness metric, a full-scale investigation needs to be conducted to conclusively establish this. We leave such a study for future work."], "citing_paper_content": {"title": "Rethinking Cyclegan: Improving Quality Of Gans For Unpaired Image-To-Image Translation", "abstract": "An unpaired image-to-image (I2I) translation technique seeks to find a mapping between two domains of data in a fully unsupervised manner. While the initial solutions to the I2I problem were provided by the generative adversarial neural networks (GANs), currently, diffusion models (DM) hold the state-of-the-art status on the I2I translation benchmarks in terms of FID. Yet, they suffer from some limitations, such as not using data from the source domain during the training, or maintaining consistency of the source and translated images only via simple pixel-wise errors. This work revisits the classic CycleGAN model and equips it with recent advancements in model architectures and model training procedures. The revised model is shown to significantly outperform other advanced GAN-and DM-based competitors on a variety of benchmarks. In the case of Male-to-Female translation of CelebA, the model achieves over 40% improvement in FID score compared to the stateof-the-art results. This work also demonstrates the ineffectiveness of the pixel-wise I2I translation faithfulness metrics and suggests their revision. The code and trained models are available at https://github.com/LS4GAN/ uvcgan2."}, "cited_paper_content": {"title": "The Unreasonable Effectiveness Of Deep Features As A Perceptual Metric", "abstract": "While it is nearly effortless for humans to quickly assess the perceptual similarity between two images, the underlying processes are thought to be quite complex. Despite this, the most widely used perceptual metrics today, such as PSNR and SSIM, are simple, shallow functions, and fail to account for many nuances of human perception. Recently, the deep learning community has found that features of the VGG network trained on the ImageNet classification task has been remarkably useful as a training loss for image synthesis. But how perceptual are these so-called\"perceptual losses\"? What elements are critical for their success? To answer these questions, we introduce a new Full Reference Image Quality Assessment (FR-IQA) dataset of perceptual human judgments, orders of magnitude larger than previous datasets. We systematically evaluate deep features across different architectures and tasks and compare them with classic metrics. We find that deep features outperform all previous metrics by huge margins. More surprisingly, this result is not restricted to ImageNet-trained VGG features, but holds across different deep architectures and levels of supervision (supervised, self-supervised, or even unsupervised). Our results suggest that perceptual similarity is an emergent property shared across deep visual representations."}, "keywords": ["deep features"], "citation_intent": "result"}
{"citing_id": "2305.02810v1", "cited_id": "1409.0473", "section_title": "5.1", "citation": "Two example alignments between source and target tokens induced by attention values in the work of #REFR .", "text_before_citation": [], "text_after_citation": ["The brightness of each cell is proportional to the value of s(h i , c j ), where the columns i correspond to source tokens and the rows j correspond to target tokens. . . . . . . . ."], "citing_paper_content": {"title": "", "abstract": "Recent progress in NLP can largely be attributed to progress in Deep Representation Learning techniques, which are based on opaque statistical learning methods. The success of these methods is due to the availability of large volumes of data which they leverage in pre-training schemes to improve generalization when fine-tuned on specific NLP tasks. Producing such high performance representations with an understandable meaning most often calls for the availability of annotated data. However, annotated data comes at great costs, and incurs recurring annotation effort which is impractical when deploying models at scale, i.e. for different languages, different tasks, and on different domains."}, "cited_paper_content": {"title": "Neural Machine Translation By Jointly Learning To Align And Translate", "abstract": "Neural machine translation is a recently proposed approach to machine translation. Unlike the traditional statistical machine translation, the neural machine translation aims at building a single neural network that can be jointly tuned to maximize the translation performance. The models proposed recently for neural machine translation often belong to a family of encoder-decoders and consists of an encoder that encodes a source sentence into a fixed-length vector from which a decoder generates a translation. In this paper, we conjecture that the use of a fixed-length vector is a bottleneck in improving the performance of this basic encoder-decoder architecture, and propose to extend this by allowing a model to automatically (soft-)search for parts of a source sentence that are relevant to predicting a target word, without having to form these parts as a hard segment explicitly. With this new approach, we achieve a translation performance comparable to the existing state-of-the-art phrase-based system on the task of English-to-French translation. Furthermore, qualitative analysis reveals that the (soft-)alignments found by the model agree well with our intuition."}, "keywords": ["attention values", "Two example alignments"], "citation_intent": "background"}
{"citing_id": "2303.01768v1", "cited_id": "1911.00025", "section_title": "Exploration In Rl", "citation": "CMAE #REFR creates an exploration compartment for each agent that is not shared with other agents, drastically reducing the searching space via cooperative behavior.", "text_before_citation": ["Although these methods are effective in the single-agent domain, na\u00efvely applying them separately and independently to each agent in the MARL setting is bound to result in suboptimal performance.", "This is because the agents mutually influence one another, creating additional uncertainty that has to be taken into account.", "One way of accounting for such uncertainty is to require cooperative behavior between agents.", "Recently, there have been some works on ensuring cooperation between the agents.", "MAVEN #OTHEREFR uses a hierarchical architecture to generate a shared latent vector for each agent to explore the space cooperatively."], "text_after_citation": ["However, none of these algorithms account for inherent uncertainty and are applied to other MARL algorithms."], "citing_paper_content": {"title": "Toward Risk-Based Optimistic Exploration For Cooperative Multi-Agent Reinforcement Learning", "abstract": "The multi-agent setting is intricate and unpredictable since the behaviors of multiple agents influence one another. To address this environmental uncertainty, distributional reinforcement learning algorithms that incorporate uncertainty via distributional output have been integrated with multi-agent reinforcement learning (MARL) methods, achieving state-of-the-art performance. However, distributional MARL algorithms still rely on the traditional-greedy, which does not take cooperative strategy into account. In this paper, we present a risk-based exploration that leads to collaboratively optimistic behavior by shifting the sampling region of distribution. Initially, we take expectations from the upper quantiles of state-action values for exploration, which are optimistic actions, and gradually shift the sampling region of quantiles to the full distribution for exploitation. By ensuring that each agent is exposed to the same level of risk, we can force them to take cooperatively optimistic actions. Our method shows remarkable performance in multi-agent settings requiring cooperative exploration based on quantile regression appropriately controlling the level of risk."}, "cited_paper_content": {"title": "Pic: Permutation Invariant Critic For Multi-Agent Deep Reinforcement Learning", "abstract": "Sample efficiency and scalability to a large number of agents are two important goals for multi-agent reinforcement learning systems. Recent works got us closer to those goals, addressing non-stationarity of the environment from a single agent's perspective by utilizing a deep net critic which depends on all observations and actions. The critic input concatenates agent observations and actions in a user-specified order. However, since deep nets aren't permutation invariant, a permuted input changes the critic output despite the environment remaining identical. To avoid this inefficiency, we propose a 'permutation invariant critic' (PIC), which yields identical output irrespective of the agent permutation. This consistent representation enables our model to scale to 30 times more agents and to achieve improvements of test episode reward between 15% to 50% on the challenging multi-agent particle environment (MPE)."}, "keywords": ["cooperative behavior", "agent"], "citation_intent": "background"}
{"citing_id": "2303.15749v1", "cited_id": "1503.02531", "section_title": "Introduction", "citation": "In this regard, we further propose a teacher-student #REFR approach to generate pseudo labels effectively and fine-tune g(x) simultaneously.", "text_before_citation": ["However, the domain shift between WSI and natural images may lead to sub-optimal representations, so recently there have been methods proposed to fine-tune g(x) using self-supervised techniques #OTHEREFR or weakly-supervised techniques #OTHEREFR (as shown in Fig.2(c) ).", "Nevertheless, since these two processes are still trained separately with different supervision signals, they lack joint optimization and may still leads to inconsistencies within the entire MIL pipeline.", "To address the challenges mentioned above, we propose a novel MIL framework called ICMIL, which can iteratively couple the patch feature embedding process with the bag-level classification process to enhance the effectiveness of MIL training (as illustrated in Fig. 2(d) ).", "Unlike previous works that mainly focused on designing sophisticated instance aggregators a(x) #OTHEREFR and bag classifiers f (x) #OTHEREFR , we aim to bridge the loss back-propagation process from f (x) to g(x) to improve g(x)'s ability to perceive slide-level labels.", "Specifically, we propose to use the bag-level classifier f (x) to initialize an instance-level classifier f (x), enabling f (x) to use the category knowledge learned from bag-level features to determine each instance's category."], "text_after_citation": ["After fine-tuning, the domain shift problem is alleviated in g(x), leading to better patch representations.", "The new representations can be used to train a better bag-level classifier in return for the next round of iteration.", "In summary, our contributions are: #OTHEREFR We propose ICMIL which bridges the loss propagation from the bag classifier to the patch embedder by iteratively coupling them during training.", "This framework fine-tunes the patch embedder based on the bag-level classifier, and the refined embeddings, in turn, help train a more accurate bag-level classifier.", "2We propose a teacher-student approach to achieve effective and robust knowledge transfer from the bag-level classifier f (x) to the instance-level representation embedder g(x)."], "citing_paper_content": {"title": "Iteratively Coupled Multiple Instance Learning From Instance To Bag Classifier For Whole Slide Image Classification", "abstract": "Whole Slide Image (WSI) classification remains a challenge due to their extremely high resolution and the absence of fine-grained labels. Presently, WSIs are usually classified as a Multiple Instance Learning (MIL) problem when only slide-level labels are available. MIL methods involve a patch embedding process and a bag-level classification process, but they are prohibitively expensive to be trained end-to-end. Therefore, existing methods usually train them separately, or directly skip the training of the embedder. Such schemes hinder the patch embedder's access to slide-level labels, resulting in inconsistencies within the entire MIL pipeline. To overcome this issue, we propose a novel framework called Iteratively Coupled MIL (ICMIL), which bridges the loss back-propagation process from the bag-level classifier to the patch embedder. In ICMIL, we use category information in the bag-level classifier to guide the patch-level fine-tuning of the patch feature extractor. The refined embedder then generates better instance representations for achieving a more accurate bag-level classifier. By coupling the patch embedder and bag classifier at a low cost, our proposed framework enables information exchange between the two processes, benefiting the entire MIL classification model. We tested our framework on two datasets using three different backbones, and our experimental results demonstrate consistent performance improvements over state-of-the-art MIL methods. Code will be made available upon acceptance."}, "cited_paper_content": {"title": "Distilling The Knowledge In A Neural Network", "abstract": "A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel."}, "keywords": ["pseudo labels"], "citation_intent": "method"}
{"citing_id": "2304.14343v2", "cited_id": "1912.01703", "section_title": "I. Introduction", "citation": "The library is built entirely on Py-Torch #REFR and comprises five modules: Configuration, Data, Model, Evaluation, and Execution.", "text_before_citation": ["Based on this, we introduce LibCity #OTHEREFR , an open-source library that supports standardized measurement of models.", "Using LibCity, we conducted a series of experiments to validate the effectiveness of different models and components.", "By providing a standard library for urban spatial-temporal prediction, we aim to enhance reproducibility, comparability, and the advancement of this field.", "The main features of LibCity can be summarized in four aspects:", "\u2022 Unified and Modular Framework Design: LibCity adopts a comprehensive and standardized approach to implementing, deploying, and evaluating spatial-temporal prediction models."], "text_after_citation": ["Each module has a well-defined scope and collaborates seamlessly with others to deliver the complete functionality of the library.", "We design basic spatial-temporal data storage, unified model instantiation interfaces, and standardized evaluation procedure within these modules.", "Users can effortlessly train and evaluate existing models with simple configurations.", "On the other hand, developers can concentrate solely on the interfaces that are relevant to their models without worrying about the implementation details of other modules.", "\u2022 General and Extensible Data Storage Format: Opensource spatial-temporal datasets are available in various storage formats."], "citing_paper_content": {"title": "Towards Efficient And Comprehensive Urban Spatial-Temporal Prediction: A Unified Library And Performance Benchmark", "abstract": "As deep learning technology advances and more urban spatial-temporal data accumulates, an increasing number of deep learning models are being proposed to solve urban spatial-temporal prediction problems. However, there are limitations in the existing field, including open-source data being in various formats and difficult to use, few papers making their code and data openly available, and open-source models often using different frameworks and platforms, making comparisons challenging. A standardized framework is urgently needed to implement and evaluate these methods. To address these issues, we provide a comprehensive review of urban spatial-temporal prediction and propose a unified storage format for spatialtemporal data called atomic files. We also propose LibCity, an open-source library that offers researchers a credible experimental tool and a convenient development framework. In this library, we have reproduced 65 spatial-temporal prediction models and collected 55 spatial-temporal datasets, allowing researchers to conduct comprehensive experiments conveniently. Using LibCity, we conducted a series of experiments to validate the effectiveness of different models and components, and we summarized promising future technology developments and research directions for spatial-temporal prediction. By enabling fair model comparisons, designing a unified data storage format, and simplifying the process of developing new models, LibCity is poised to make significant contributions to the spatial-temporal prediction field."}, "cited_paper_content": {"title": "Pytorch: An Imperative Style, High-Performance Deep Learning Library", "abstract": "Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. ::: In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. ::: We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks."}, "keywords": ["Data", "Py-Torch"], "citation_intent": "method"}
{"citing_id": "2303.00844v1", "cited_id": "1810.11115", "section_title": "A.3 Proof Of Theorem 4", "citation": "Specifically, we derive greedy selection rules for \u2113 0 w -based variants of the SR-LASSO (Appendix B.1) and LAD-LASSO (Appendix B.2), extending the \u2113 0 w -based LASSO setting considered in #REFR .", "text_before_citation": ["h(t) =", "l(t) = m i=1 |r i \u2212 tA ij | + \u03bbw j |x j + t| = m+1 i=1 r j i \u2212 t A ij", "and, thanks to Lemma 1, l(t) is minimized at t = r \u0135 i(j) / A\u00ee (j),j , where\u00ee(j) = \u03c4 (k(j)) andk(j) is defined as in #OTHEREFR B Greedy selection rules for \u2113 0 w -based loss functions", ".", "In this appendix we show how to derive greedy selection rules for \u2113 0 w -regularized loss functions."], "text_after_citation": ["The corresponding weighted OMP algorithms are numerically tested in \u00a74, Experiments I and II."], "citing_paper_content": {"title": "The Greedy Side Of The Lasso: New Algorithms For Weighted Sparse Recovery Via Loss Function-Based Orthogonal Matching Pursuit", "abstract": "We propose a class of greedy algorithms for weighted sparse recovery by considering new loss function-based generalizations of Orthogonal Matching Pursuit (OMP). Given a (regularized) loss function, the proposed algorithms alternate the iterative construction of the signal support via greedy index selection and a signal update based on solving a local data-fitting problem restricted to the current support. We show that greedy selection rules associated with popular weighted sparsity-promoting loss functions admit explicitly computable and simple formulas. Specifically, we consider \u2113 0-and \u2113 1-based versions of the weighted LASSO (Least Absolute Shrinkage and Selection Operator), the Square-Root LASSO (SR-LASSO) and the Least Absolute Deviations LASSO (LAD-LASSO). Through numerical experiments on Gaussian compressive sensing and high-dimensional function approximation, we demonstrate the effectiveness of the proposed algorithms and empirically show that they inherit desirable characteristics from the corresponding loss functions, such as SR-LASSO's noise-blind optimal parameter tuning and LAD-LASSO's fault tolerance. In doing so, our study sheds new light on the connection between greedy sparse recovery and convex relaxation."}, "cited_paper_content": {"title": "Sparse Approximation Of Multivariate Functions From Small Datasets Via Weighted Orthogonal Matching Pursuit", "abstract": "We show the potential of greedy recovery strategies for the sparse approximation of multivariate functions from a small dataset of pointwise evaluations by considering an extension of the orthogonal matching pursuit to the setting of weighted sparsity. The proposed recovery strategy is based on a formal derivation of the greedy index selection rule. Numerical experiments show that the proposed weighted orthogonal matching pursuit algorithm is able to reach accuracy levels similar to those of weighted l1 minimization programs while considerably improving the computational efficiency."}, "keywords": ["-based LASSO setting"], "citation_intent": "method"}
{"citing_id": "2304.03973v2", "cited_id": "1608.04644", "section_title": "B. Evaluations Under The Carlini Wagner (Cw) Attack", "citation": "In fact, it reaches a 100% fooling rate for all three datasets. Similar findings were also made in #REFR .", "text_before_citation": ["For our estimations, we set a maximum of 10 iterations for the MNIST and 5 iterations for the CIFAR10 dataset.", "In addition, for the attacks on CIFAR10, the algorithm has been forced to set the confidence of the targeted class to 0.5 higher than the confidence of the true label.", "Table II reports the fooling rate and the mean distortion for both the datasets.", "CapsNets vs.", "CNNs: The CW attack is very effective for traditional CNNs."], "text_after_citation": ["On the other hand, both CapsNets show greater robustness (i.e., lower fooling rate) than CNNs, for the MNIST dataset (and also for the GTSRB, even if the fooling rate of the DeepCaps is just a little bit lower than 100%).", "The CapsNets also require a higher mean distortion than the CNNs, which makes the resulting adversarial example more perceptible.", "For the CIFAR10 dataset, the CW attack shows its effectiveness because, for all the networks, the fooling rate is 100%.", "However, we can notice that CapsNets are more robust due to a higher mean distortion. DeepCaps vs.", "ShallowCaps: The DeepCaps appears to be more robust than the ShallowCaps, because of a lower fooling rate, despite having slightly lower mean distortion."], "citing_paper_content": {"title": "Robcaps: Evaluating The Robustness Of Capsule Networks Against Affine Transformations And Adversarial Attacks", "abstract": "Capsule Networks (CapsNets) are able to hierarchically preserve the pose relationships between multiple objects for image classification tasks. Other than achieving high accuracy, another relevant factor in deploying CapsNets in safety-critical applications is the robustness against input transformations and malicious adversarial attacks. In this paper, we systematically analyze and evaluate different factors affecting the robustness of CapsNets, compared to traditional Convolutional Neural Networks (CNNs). Towards a comprehensive comparison, we test two CapsNet models and two CNN models on the MNIST, GTSRB, and CIFAR10 datasets, as well as on the affine-transformed versions of such datasets. With a thorough analysis, we show which properties of these architectures better contribute to increasing the robustness and their limitations. Overall, CapsNets achieve better robustness against adversarial examples and affine transformations, compared to a traditional CNN with a similar number of parameters. Similar conclusions have been derived for deeper versions of CapsNets and CNNs. Moreover, our results unleash a key finding that the dynamic routing does not contribute much to improving the CapsNets' robustness. Indeed, the main generalization contribution is due to the hierarchical feature learning through capsules."}, "cited_paper_content": {"title": "Towards Evaluating The Robustness Of Neural Networks", "abstract": "Neural networks provide state-of-the-art results for most machine learning tasks. Unfortunately, neural networks are vulnerable to adversarial examples: given an input $x$ and any target classification $t$, it is possible to find a new input $x'$ that is similar to $x$ but classified as $t$. This makes it difficult to apply neural networks in security-critical areas. Defensive distillation is a recently proposed approach that can take an arbitrary neural network, and increase its robustness, reducing the success rate of current attacks' ability to find adversarial examples from $95\\%$ to $0.5\\%$. ::: In this paper, we demonstrate that defensive distillation does not significantly increase the robustness of neural networks by introducing three new attack algorithms that are successful on both distilled and undistilled neural networks with $100\\%$ probability. Our attacks are tailored to three distance metrics used previously in the literature, and when compared to previous adversarial example generation algorithms, our attacks are often much more effective (and never worse). Furthermore, we propose using high-confidence adversarial examples in a simple transferability test we show can also be used to break defensive distillation. We hope our attacks will be used as a benchmark in future defense attempts to create neural networks that resist adversarial examples."}, "keywords": ["100% fooling rate"], "citation_intent": "result"}
{"citing_id": "2304.02419v1", "cited_id": "2002.03761", "section_title": "Evaluation On Music Conditioned Dance Generation", "citation": "And for DanceNet #REFR , it generates dance in a repeat motion pattern, nearly zero freezing but with low diversity.", "text_before_citation": ["The existing evaluation metrics are not sufficient to reflect the quality of generated dance in practice, which motivates us to propose a PFF and AUC f to evaluate the percentage of freezing frames.", "It is worth noting that our architecture outperforms FACT #OTHEREFR and Bailando #OTHEREFR in terms of PFF and AUC f , which shows that our method rarely generates frozen frames. Meanwhile, Li et al.", "#OTHEREFR and DanceNet #OTHEREFR also gain nearly zero freezing in PFF and AUC f .", "The reason is that the generated dances of Li et al.", "#OTHEREFR are highly jittery making its velocity variation extremely high, which is also reported in #OTHEREFR , and leads to the non-freezing issue."], "text_after_citation": ["Furthermore, the results of user study show that the generated 3D dance is more visually realistic compared to other methods.", "Even in comparison to the ground truth, 41.9% of our generated dance is voted as Table 3 : The result of text2motion, only mean reported.", "the better in average.", "We also report the result trained with dance-only data, which shows comparable performance.", "In addition to the evaluation on the test set of AIST++, we also show the experimental results on our in-the-wild dataset in Table 1 ."], "citing_paper_content": {"title": "Fig-1: Demo", "abstract": "Audio input Effect range: 6s ~ 8s A person does a cartwheel. Effect range : 16s ~ 18s A person does a spin."}, "cited_paper_content": {"title": "Music2Dance: Dancenet For Music-Driven Dance Generation", "abstract": "Synthesize human motions from music, i.e., music to dance, is appealing and attracts lots of research interests in recent years. It is challenging due to not only the requirement of realistic and complex human motions for dance, but more importantly, the synthesized motions should be consistent with the style, rhythm and melody of the music. In this paper, we propose a novel autoregressive generative model, DanceNet, to take the style, rhythm and melody of music as the control signals to generate 3D dance motions with high realism and diversity. To boost the performance of our proposed model, we capture several synchronized music-dance pairs by professional dancers, and build a high-quality music-dance pair dataset. Experiments have demonstrated that the proposed method can achieve the state-of-the-art results."}, "keywords": ["DanceNet"], "citation_intent": "background"}
{"citing_id": "2305.00678v1", "cited_id": "1901.04056", "section_title": "Experimental Results", "citation": "On LiTS17 #REFR , as shown in Table 2, our model achieves 91.50% in Dice and 84.59% in IoU, outperforming SOTA methods by 0.26% and 0.45%, respectively.", "text_before_citation": ["We compare our CTO with the state-of-the-art (SOTA) methods including U-Net #OTHEREFR , ResUNet #OTHEREFR with ResNet-50 #OTHEREFR as the backbone, VNet #OTHEREFR , ViT #OTHEREFR , TransUNet #OTHEREFR , and Swin-Unet #OTHEREFR .", "On ISIC 2016 #OTHEREFR in IoU, which outperforms the SOTA methods by 0.05% and 0.88%, respectively.", "On ISIC 2018 #OTHEREFR , our CTO achieves 91.2% in Dice and 84.5% in IoU by 5-fold crossvalidation, which outperforms the SOTA methods by 1.8% and 2.3%, respectively.", "On CoNIC #OTHEREFR , results are shown in Table 2 , we can observe that our CTO achieves 79.77%, 66.42%, and 65.58% in Dice, IoU, and PQ, respectively, consistently outperforming other methods. Qualitative result comparisons are illustrated in Figure 2 .", "We can observe that our CTO delineates more accurate object contours than other methods regarding diverse shapes and sizes of nuclei, especially on some blurred nuclei objects. We also conduct experiments on 3D MISeg tasks."], "text_after_citation": ["On BTCV #OTHEREFR , as shown in Table 3 , our CTO achieves 81.10% in Dice and 18.75% in HD, which outperforms the SOTA methods.", "In particular, as for Dice, our CTO outperforms the second, third, and fourth best methods by 1.98%, 3.33%, and 3.62, respectively. Besides, the distinct improvements Table 2 .", "Comparisons with other methods on CoNIC #OTHEREFR and LiTS17 #OTHEREFR .", "can be markedly observed for organs with blurry boundaries, e.g., the \"pancreas\" and the \"stomach\", where our model achieves significant gains over the SOTA methods, i.e., 4.70% and 3.60% in Dice, respectively.", "As for the model efficiency, we can observe that CTO achieves competitive performance improvements with comparable FLOPs and parameters. Ablation Study."], "citing_paper_content": {"title": "Rethinking Boundary Detection In Deep Learning Models For Medical Image Segmentation", "abstract": "Medical image segmentation is a fundamental task in the community of medical image analysis. In this paper, a novel network architecture, referred to as Convolution, Transformer, and Operator (CTO), is proposed. CTO employs a combination of Convolutional Neural Networks (CNNs), Vision Transformer (ViT), and an explicit boundary detection operator to achieve high recognition accuracy while maintaining an optimal balance between accuracy and efficiency. The proposed CTO follows the standard encoder-decoder segmentation paradigm, where the encoder network incorporates a popular CNN backbone for capturing local semantic information, and a lightweight ViT assistant for integrating long-range dependencies. To enhance the learning capacity on boundary, a boundary-guided decoder network is proposed that uses a boundary mask obtained from a dedicated boundary detection operator as explicit supervision to guide the decoding learning process. The performance of the proposed method is evaluated on six challenging medical image segmentation datasets, demonstrating that CTO achieves state-of-the-art accuracy with a competitive model complexity."}, "cited_paper_content": {"title": "The Liver Tumor Segmentation Benchmark (Lits)", "abstract": "In this work, we report the set-up and results of the Liver Tumor Segmentation Benchmark (LITS) organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI) 2016 and International Conference On Medical Image Computing Computer Assisted Intervention (MICCAI) 2017. Twenty four valid state-of-the-art liver and liver tumor segmentation algorithms were applied to a set of 131 computed tomography (CT) volumes with different types of tumor contrast levels (hyper-/hypo-intense), abnormalities in tissues (metastasectomie) size and varying amount of lesions. The submitted algorithms have been tested on 70 undisclosed volumes. The dataset is created in collaboration with seven hospitals and research institutions and manually reviewed by independent three radiologists. We found that not a single algorithm performed best for liver and tumors. The best liver segmentation algorithm achieved a Dice score of 0.96(MICCAI) whereas for tumor segmentation the best algorithm evaluated at 0.67(ISBI) and 0.70(MICCAI). The LITS image data and manual annotations continue to be publicly available through an online evaluation system as an ongoing benchmarking resource."}, "keywords": ["Dice"], "citation_intent": "result"}
{"citing_id": "2304.01102v1", "cited_id": "1712.03854", "section_title": "Introduction", "citation": "Section 5 presents characteristics of the final dataset including statistical information on programming languages, fix size, bug labels, exceptions and errors as well as data splits. We also present initial result from two APR baselines. One is Cardumen, a Generate-and-Validate approach #REFR .", "text_before_citation": ["We also ensure that the bugs are of adequate type (semantic, rather than syntactic bugs) and have an adequate distribution in terms of difficulty. We also address issues such as duplication and flakyness.", "Moreover, we have labeled most of the bugs into fine-grained categories, allowing to diagnose \"weak spots\" in an APR system.", "In section 4 we explain how we built RunBugRun.", "The bugs were extracted from Project CodeNet #OTHEREFR , a large dataset of code submissions to programming contest websites.", "In this section we describe in detail the extensive filtering, pre-processing, execution and labeling steps we undertook."], "text_after_citation": ["The second is an NPR baseline, based on a fine-tuned pretrained language model, CodeT5 #OTHEREFR .", "Beyond the raw performance comparison, we also discuss additional findings of our baseline evaluation.", "We show that there is a strong correlation between repair performance and the number of fixing changes, we highlight strong and weak spot for both baselines, and provide initial evidence that NPR models are able to learn and apply fixes across language boundaries (knowledge transfer).", "From these results, we conclude that there is ample room for improvement for both G&V and NPR approaches.", "We discuss some of the way forwards and the challenges ahead in Section 6."], "citing_paper_content": {"title": "", "abstract": "Recently, we can notice a transition to data-driven techniques in Automated Program Repair (APR), in particular towards deep neural networks. This entails training on hundreds of thousands or even millions of non-executable code fragments. We would like to bring more attention to an aspect of code often neglected in Neural Program Repair (NPR), namely its execution. Code execution has several significant advantages. It allows for test-based evaluation of candidate fixes and can provide valuable information to aid repair. In this work we present a fully executable dataset of 450,000 small buggy/fixed program pairs originally submitted to programming competition websites written in eight different programming languages. Along with the dataset we provide infrastructure to compile, safely execute and test programs as well as fine-grained bug-type labels. To give a point of reference, we provide basic evaluation results for two baselines, one based on a generate-andvalidate approach and one on deep learning. With this dataset we follow several goals: we want to lift Neural Program Repair beyond fully static code representations, foster the use of executionbased features and, by including several different languages, counterbalance the predominance of Java in the current landscape of APR datasets and benchmarks. Keywords automated program repair, data-driven software engineering, fault localization Recently, more and more APR research builds on deep learning. So much so that this sub-field has been given its own name: Neural Program Repair (NPR). NPR systems are trained on up to millions of buggy/fixed code fragment pairs. So far, there is a strong focus on static code features, in particular textual features [13, 47, 32], less commonly, tree or graph representations [43, 17]. Because NPR systems are data hungry and manually collecting and isolating bugs is infeasible on a large scale, copious amounts of bug data are mined from open source code repositories (e.g., GitHub) [71]."}, "cited_paper_content": {"title": "Ultra-Large Repair Search Space With Automatically Mined Templates: The Cardumen Mode Of Astor", "abstract": "Astor is a program repair library which has different modes. In this paper, we present the Cardumen mode of Astor, a repair approach based mined templates that has an ultra-large search space. We evaluate the capacity of Cardumen to discover test-suite adequate patches (aka plausible patches) over the 356 real bugs from Defects4J [11]. Cardumen finds 8935 patches over 77 bugs of Defects4J. This is the largest number of automatically synthesized patches ever reported, all patches being available in an open-science repository. Moreover, Cardumen identifies 8 unique patches, that are patches for Defects4J bugs that were never repaired in the whole history of program repair."}, "keywords": ["bug labels"], "citation_intent": "method"}
{"citing_id": "2303.00320v2", "cited_id": "1810.04805", "section_title": "Introduction", "citation": "What is worse, these methods essentially adopt the scheme of the unidirectional encoder to learning representations of time series, which largely restrict the extraction of contextual representations #REFR .", "text_before_citation": ["Numerous efforts have been proposed by successfully applying self-supervision techniques to the time series areas.", "Among these current works, contrastive learning paradigm #OTHEREFR has nearly become the most prevalent solutions.", "The common schemes underlying these contrastive methods are to learn embeddings that are invariant to distortions of various scale inputs with the cooperation of data augmentation and negative sampling strategies #OTHEREFR .", "Though their effectiveness and prevalence, the invariance assumptions may not always hold in real-world scenarios.", "In addition, it would bring too much inductive bias in developing data augmentation strategies #OTHEREFR and raise extra biases in negative sampling #OTHEREFR ."], "text_after_citation": ["Masked autoencoders #OTHEREFR , #OTHEREFR , which can well overcome these aforementioned limitations, have been proposed before, whose main philosophy is to encode the maskingstyle corrupt input into latent space followed by a recovery of the raw inputs via the encoder and decoder.", "One successful instantiation is BERT #OTHEREFR , which is built based on a transformer network and has achieved a great milestone in representing language.", "Inspired by that, we notice that a pioneering work #OTHEREFR has been presented based on transformer networks, which directly treats the raw time series as inputs, followed by recovering these full inputs with point-wise regression optimization.", "However, this methods often require very expensive computation cost due to ignoring the quadratic computation complexity incurred by self-attention mechanism #OTHEREFR , #OTHEREFR .", "As reported by recently proposed methods #OTHEREFR , #OTHEREFR , #OTHEREFR , it seems that this method only produces limited representation results due to its weak generalization performance."], "citing_paper_content": {"title": "Timemae: Self-Supervised Representations Of Time Series With Decoupled Masked Autoencoders", "abstract": "Enhancing the expressive capacity of deep learning-based time series models with self-supervised pre-training has become ever-increasingly prevalent in time series classification. Even though numerous efforts have been devoted to developing self-supervised models for time series data, we argue that the current methods are not sufficient to learn optimal time series representations due to solely unidirectional encoding over sparse point-wise input units. In this work, we propose TimeMAE, a novel self-supervised paradigm for learning transferrable time series representations based on transformer networks. The distinct characteristics of the TimeMAE lie in processing each time series into a sequence of non-overlapping sub-series via window-slicing partitioning, followed by random masking strategies over the semantic units of localized sub-series. Such a simple yet effective setting can help us achieve the goal of killing three birds with one stone, i.e., (1) learning enriched contextual representations of time series with a bidirectional encoding scheme; (2) increasing the information density of basic semantic units; (3) efficiently encoding representations of time series using transformer networks. Nevertheless, it is a non-trivial to perform reconstructing task over such a novel formulated modeling paradigm. To solve the discrepancy issue incurred by newly injected masked embeddings, we design a decoupled autoencoder architecture, which learns the representations of visible (unmasked) positions and masked ones with two different encoder modules, respectively. Furthermore, we construct two types of informative targets to accomplish the corresponding pretext tasks. One is to create a tokenizer module that assigns a codeword to each masked region, allowing the masked codeword classification (MCC) task to be completed effectively. Another one is to adopt a siamese network structure to generate target representations for each masked input unit, aiming at performing the masked representation regression (MRR) optimization. Comprehensively pre-trained, our model can efficiently learn transferrable time series representations, thus benefiting the classification of time series. We extensively perform experiments on five benchmark datasets to verify the effectiveness of the TimeMAE. Experimental results show that the TimeMAE can significantly surpass previous competitive baselines. Furthermore, we also demonstrate the universality of the learned representations by performing transfer learning experiments. For the reproducibility of our results, we make our experiment codes publicly available to facilitate the self-supervised representations of time series in https://github.com/Mingyue-Cheng/TimeMAE."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["time series", "contextual representations"], "citation_intent": "method"}
{"citing_id": "2304.08439v1", "cited_id": "1809.03443", "section_title": "A. Self-Supervised Learning", "citation": "L f ld as defined in #REFR , penalizes the anatomically infeasible deformations where the retinal tissue folds onto itself.", "text_before_citation": ["The network weights of \u03c8 are not updated through backpropagation as it may learn to collapse the features of I t+k and I t , even though they appear different.", "Inspired by BYOL #OTHEREFR , we update \u03c8 with an exponential moving average of the Encoder weights obtained while it is being trained.", "Additional regularization loss terms are also incorporated to obtain an anatomically feasible T D,A .", "D is encouraged to be diffeomorphic by penalizing it to be smooth with L smt and prevent folding with L f ld .", "The L smth = p\u2208\u2126 ||\u2207D(p)|| 2 2 was defined as in #OTHEREFR , where the spatial gradient \u2207D(p) is computed at all voxel positions through discrete numerical approximation."], "text_after_citation": ["Finally, the sparsity of A is ensured with an L1regularization L add = p\u2208\u2126 |A(p)|. Thus the total loss is", "L = L mse + \u03bb 3 .L prc + \u03bb 4 .L smt + \u03bb 5 .L f ld + \u03bb 6 L add , (3)", "where \u03bb 3 = 10 1 , \u03bb 4 = 10 \u22121 , \u03bb 5 = 10 6 and \u03bb 6 = 10 \u22125 are empirically fixed, based on their relative importance and also to scale the different loss terms to a similar range.", "The range of the L f ld is orders of magnitude lower than the other terms, thus requiring a significantly larger scaling weight.", "Network Architecture: The separable 3D Convolution Block (S3DConv) depicted in Fig. 2a ."], "citing_paper_content": {"title": "Morph-Ssl: Self-Supervision With Longitudinal Morphing To Predict Amd Progression From Oct", "abstract": "The lack of reliable biomarkers makes predicting the conversion from intermediate to neovascular age-related macular degeneration (iAMD, nAMD) a challenging task. We develop a Deep Learning (DL) model to predict the future risk of conversion of an eye from iAMD to nAMD from its current OCT scan. Although eye clinics generate vast amounts of longitudinal OCT scans to monitor AMD progression, only a small subset can be manually labeled for supervised DL. To address this issue, we propose Morph-SSL, a novel Self-supervised Learning (SSL) method for longitudinal data. It uses pairs of unlabelled OCT scans from different visits and involves morphing the scan from the previous visit to the next. The Decoder predicts the transformation for morphing and ensures a smooth feature manifold that can generate intermediate scans between visits through linear interpolation. Next, the Morph-SSL trained features are input to a Classifier which is trained in a supervised manner to model the cumulative probability distribution of the time to conversion with a sigmoidal function. Morph-SSL was trained on unlabelled scans of 399 eyes (3570 visits). The Classifier was evaluated with a five-fold cross-validation on 2418 scans from 343 eyes with clinical labels of the conversion date. The Morph-SSL features achieved an AUC of 0.766 in predicting the conversion to nAMD within the next 6 months, outperforming the same network when trained end-to-end from scratch or pretrained with popular SSL methods. Automated prediction of the future risk of nAMD onset can enable timely treatment and individualized AMD management."}, "cited_paper_content": {"title": "Inverse-Consistent Deep Networks For Unsupervised Deformable Image Registration", "abstract": "Deformable image registration is a fundamental task in medical image analysis, aiming to establish a dense and non-linear correspondence between a pair of images. Previous deep-learning studies usually employ supervised neural networks to directly learn the spatial transformation from one image to another, requiring task-specific ground-truth registration for model training. Due to the difficulty in collecting precise ground-truth registration, implementation of these supervised methods is practically challenging. Although several unsupervised networks have been recently developed, these methods usually ignore the inherent inverse-consistent property (essential for diffeomorphic mapping) of transformations between a pair of images. Also, existing approaches usually encourage the to-be-estimated transformation to be locally smooth via a smoothness constraint only, which could not completely avoid folding in the resulting transformation. To this end, we propose an Inverse-Consistent deep Network (ICNet) for unsupervised deformable image registration. Specifically, we develop an inverse-consistent constraint to encourage that a pair of images are symmetrically deformed toward one another, until both warped images are matched. Besides using the conventional smoothness constraint, we also propose an anti-folding constraint to further avoid folding in the transformation. The proposed method does not require any supervision information, while encouraging the diffeomoprhic property of the transformation via the proposed inverse-consistent and anti-folding constraints. We evaluate our method on T1-weighted brain magnetic resonance imaging (MRI) scans for tissue segmentation and anatomical landmark detection, with results demonstrating the superior performance of our ICNet over several state-of-the-art approaches for deformable image registration. Our code will be made publicly available."}, "keywords": ["retinal tissue", "anatomically infeasible deformations"], "citation_intent": "background"}
{"citing_id": "2305.00656v1", "cited_id": "1908.06148", "section_title": "C. Experimental Setup", "citation": "We did not perform tuning of convolution kernel size; we used the kernel sizes from previous work #REFR that proved to be effective.", "text_before_citation": ["All of our experiments were conducted on a dual Intel Xeon CPU E5-2620 CPUs @ 2.40 GHz (12 physical cores, 24 logical cores), 192 GB RAM, and a single Nvidia Titan X GPU, running on Ubuntu 20.04 Operating System. Pytorch 1.5.0 was used for neural network design.", "We used automated hyper-parameter tuning for learning rate, optimizer choice, and activation function using TPE #OTHEREFR implemented through Optuna #OTHEREFR .", "We found out that Adam optimizer #OTHEREFR and Hardswish #OTHEREFR are the best optimizer and activation function, respectively."], "text_after_citation": ["The kernel sizes for different branches of the inception block were taken as 11,19, and 27.", "The accuracy for different networks was calculated as follows:", "Accuracy = T P + T N T P + T N + F P + F N (7)", "where TP is True Positive, TN is True Negative, FP is False Positive, and FN is False Negative.", "To achieve higher accuracy from smaller models, we pretrained our network on corresponding fragment size dataset."], "citing_paper_content": {"title": "File Fragment Classification Using Light-Weight Convolutional Neural Networks", "abstract": "In digital forensics, file fragment classification is an important step toward completing file carving process. There exist several techniques to identify the type of file fragments without relying on meta-data, such as using features like header/footer and N-gram to identify the fragment type. Recently, convolutional neural network (CNN) models have been used to build classification models to achieve this task. However, the number of parameters in CNNs tends to grow exponentially as the number of layers increases. This results in a dramatic increase in training and inference time. In this paper, we propose lightweight file fragment classification models based on depthwise separable CNNs. The evaluation results show that our proposed models provide faster inference time with comparable accuracy as compared to the state-of-art CNN based models. In particular, our models were able to achieve an accuracy of 79% on the FFT-75 dataset with nearly 100K parameters and 164M FLOPs, which is 4x smaller and 6x faster than the state-of-the-art classifier in the literature."}, "cited_paper_content": {"title": "Fifty: Large-Scale File Fragment Type Identification Using Neural Networks", "abstract": "We present FiFTy, a modern file type identification tool for memory forensics and data carving. In contrast to previous approaches based on hand-crafted features, we design a compact neural network architecture, which uses a trainable embedding space, akin to successful natural language processing models. Our approach dispenses with explicit feature extraction which is a bottleneck in legacy systems. We evaluate the proposed method on a novel dataset with 75 file types - the most diverse and balanced dataset reported to date. FiFTy consistently outperforms all baselines in terms of speed, accuracy and individual misclassification rates. We achieved an average accuracy of 77.5% with processing speed of approx 38 sec/GB, which is better and more than an order of magnitude faster than the previous state-of-the-art tool - Sceadan (69% at 9 min/GB). Our tool and the corresponding dataset are available publicly online."}, "keywords": ["convolution kernel size", "kernel sizes"], "citation_intent": "method"}
{"citing_id": "2303.03889v1", "cited_id": "1311.5202", "section_title": "Accuracy And Computational Complexity", "citation": "Compared with the error of the dFMM in #REFR from which our algorithm is transformed, the error becomes greater.", "text_before_citation": ["In the highest frequency case with \u03baD = 201.1, the diameter is 32 wavelengths, as illustrated in Figure   3 , and over 1 million points are sampled on the boundary.", "The computational results are summarized in Table 1 , where T c , T m , T p , T t are the computational time cost by curvelet construction, sparse system matrix computation, matrix-vector multiplication, and the overall running time, M Q , M m , M t are the memory cost by the Q matrices, sparse system matrix, and the overall memory consumption, respectively.", "Generally the overall error \u03b5 a is of the same magnitude with the predetermined parameter \u03b5.", "The error \u03b5 a seems to grow with log N .", "This is reasonable since in each level, some error of order O(\u03b5) is introduced, and there are O(log N ) levels in the octree."], "text_after_citation": ["This is because in our CBM, besides the low rank approximation of the kernel, extra error is introduced in the construction of curvelets with quasi-vanishing moments and the a-posteriori compression. Nevertheless, the overall error is still controllable.", "\u03b5 a T c T m T p T t M Q M m M", "In for each cone and its theoretical upper bound\u0393 l are listed in Table 2 .", "It is shown that in real-world numerical cases, the number of nonempty directional cones may be much less than the upper bound\u0393 l \u223c O(\u03ba 2 w 2 ), and it tends to increase with frequency.", "This may be because the dimensionless curvature \u03bb/R of the surface on which the points locate reduces with the frequency."], "citing_paper_content": {"title": "A Nearly Optimal Explicitly-Sparse Representation For Oscillatory Kernels With Curvelet-Like Functions", "abstract": "A nearly optimal explicitly-sparse representation for oscillatory kernels is presented in this work by developing a curvelet based method. Multilevel curvelet-like functions are constructed as the transform of the original nodal basis. Then the system matrix in a new non-standard form is derived with respect to the curvelet basis, which would be nearly optimally sparse due to the directional low rank property of the oscillatory kernel. Its sparsity is further enhanced via a-posteriori compression. Finally its nearly optimial log-linear computational complexity with controllable accuracy is demonstrated with numerical results. This explicitly-sparse representation is expected to lay ground to future work related to fast direct solvers and effective preconditioners for high frequency problems. It may also be viewed as the generalization of wavelet based methods to high frequency cases, and used as a new wideband fast algorithm for wave problems."}, "cited_paper_content": {"title": "A Fast Directional Bem For Large-Scale Acoustic Problems Based On The Burton-Miller Formulation", "abstract": "In this paper, a highly efficient fast boundary element method (BEM) for solving large-scale engineering acoustic problems in a broad frequency range is developed and implemented. The acoustic problems are modeled by the Burton\u2013Miller boundary integral equation (BIE), thus the fictitious frequency issue is completely avoided. The BIE is discretized by using the Nystrom method based on the curved quadratic elements, leading to simple numerical implementation (no edge or corner problems) and high accuracy in the BEM analysis. The linear systems are solved iteratively and accelerated by using a newly developed kernel-independent wideband fast directional algorithm (FDA) for fast summation of oscillatory kernels. In addition, the computational efficiency of the FDA is further promoted by exploiting the low-rank features of the translation matrices, resulting in two- to three-fold reduction in the computational time of the multipole-to-local translations. The high accuracy and nearly linear computational complexity of the present method are clearly demonstrated by typical examples. An acoustic scattering problem with dimensionless wave number kD (where k is the wave number and D is the typical length of the obstacle) up to 1000 and the degrees of freedom up to 4 million is successfully solved within 10 h on a computer with one core and the memory usage is 24 GB."}, "keywords": ["algorithm"], "citation_intent": "result"}
{"citing_id": "2303.06164v1", "cited_id": "1801.01290", "section_title": "Integrating Advances In Rl Within Qd-Rl", "citation": "Most interestingly, we observe that PGA-ME (SAC) is able to achieve in the Humanoid task a maximum fitness that corresponds to solving the task according to previous literature #REFR .", "text_before_citation": ["We see that SAC performs as well as TD3 on HalfCheetah and Ant.", "In the complex Humanoid task, TD3 completely fails while SAC successfully achieves high-performance.", "This result shows that both action space exploration coupled with the maximum-entropy objective incentive is crucial to enable Humanoid task to be achievable.", "6.2.1 Benefits of PGA-ME (SAC) -maximum entropy RL.", "We first evaluate the effectiveness of maximum entropy objectives in the QD-RL setting through PGA-ME (SAC) (Section 5.2.2). Figure 5 shows the maximum fitness and QD-score metrics."], "text_after_citation": ["Such scores were never reached by any QD-RL algorithm (including PGA-ME (TD3)) before this.", "While Deep RL literature (confirmed by experiments in Figure 4 ) has already shown that SAC is capable of solving the Humanoid task while TD3 is unable to do so, realizing the same results in the QD-RL setting just by changing the RL algorithm in the framework is a promising indication of the possibilities.", "As seen in the Deep RL results, this also demonstrates that action space exploration and maximum-entropy objective of SAC enables the Humanoid task to be solved by PGA-ME (SAC).", "We would like to stress here that this results is only permitted by sufficient training of the actor-critic models, as will be shown in Section 6.3.", "Results for HalfCheetah similarly follow the Deep RL trend, where PGA-ME (SAC) and PGA-ME (TD3) perform equivalently."], "citing_paper_content": {"title": "Understanding The Synergies Between Quality-Diversity And Deep Reinforcement Learning", "abstract": "Figure 1: Illustrative figure displaying the settings compared in this work: QD-RL, Actor-Critic QD-RL and Actor-Critic Deep RL. QD-RL refers generally to applying QD to RL domains to find collection of diverse policies, while Actor-Critic QD-RL approaches integrate elements from Actor-Critic Deep RL methods within QD to take advantage of the properties of the Deep RL domains. This representation highlights key differences between Actor-Critic QD-RL and Actor-Critic Deep RL settings: (1) update frequency of the actor-critic models and (2) policies used to collect data."}, "cited_paper_content": {"title": "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning With A Stochastic Actor", "abstract": "Model-free deep reinforcement learning (RL) algorithms have been demonstrated on a range of challenging decision making and control tasks. However, these methods typically suffer from two major challenges: very high sample complexity and brittle convergence properties, which necessitate meticulous hyperparameter tuning. Both of these challenges severely limit the applicability of such methods to complex, real-world domains. In this paper, we propose soft actor-critic, an off-policy actor-critic deep RL algorithm based on the maximum entropy reinforcement learning framework. In this framework, the actor aims to maximize expected reward while also maximizing entropy - that is, succeed at the task while acting as randomly as possible. Prior deep RL methods based on this framework have been formulated as Q-learning methods. By combining off-policy updates with a stable stochastic actor-critic formulation, our method achieves state-of-the-art performance on a range of continuous control benchmark tasks, outperforming prior on-policy and off-policy methods. Furthermore, we demonstrate that, in contrast to other off-policy algorithms, our approach is very stable, achieving very similar performance across different random seeds."}, "keywords": ["Humanoid task"], "citation_intent": "result"}
{"citing_id": "2303.06800v1", "cited_id": "1803.10683", "section_title": "B.2. Ochuman Dataset", "citation": "Compared to Pose2Seg #REFR , our method achieves better performance in both detection and segmentation tasks (Table 7).", "text_before_citation": ["In Fig.", "6 in the main paper, we present several qualitative results on the OCHuman dataset #OTHEREFR .", "In this section, we present quantitative results on the dataset."], "text_after_citation": ["Compared to a state-of-the-art segmentation method (Mask2Former #OTHEREFR ), our method shows a comparable accuracy in segmentation.", "Compared to our baseline multitasking model (BaseNet-DPS), our approach still induces significant improvements in pose estimation and segmentation on this different dataset.", "We show additional visual results of our methods on the COCO and OCHuman datasets."], "citing_paper_content": {"title": "Object-Centric Multi-Task Learning For Human Instances", "abstract": "Human is one of the most essential classes in visual recognition tasks such as detection, segmentation, and pose estimation. Although much effort has been put into individual tasks, multi-task learning for these three tasks has been rarely studied. In this paper, we explore a compact multi-task network architecture that maximally shares the parameters of the multiple tasks via object-centric learning. To this end, we propose a novel query design to encode the human instance information effectively, called humancentric query (HCQ). HCQ enables for the query to learn explicit and structural information of human as well such as keypoints. Besides, we utilize HCQ in prediction heads of the target tasks directly and also interweave HCQ with the deformable attention in Transformer decoders to exploit a well-learned object-centric representation. Experimental results show that the proposed multi-task network achieves comparable accuracy to state-of-the-art task-specific models in human detection, segmentation, and pose estimation task, while it consumes less computational costs."}, "cited_paper_content": {"title": "Pose2Seg: Detection Free Human Instance Segmentation", "abstract": "The standard approach to image instance segmentation is to perform the object detection first, and then segment the object from the detection bounding-box. More recently, deep learning methods like Mask R-CNN perform them jointly. However, little research takes into account the uniqueness of the \"human\" category, which can be well defined by the pose skeleton. Moreover, the human pose skeleton can be used to better distinguish instances with heavy occlusion than using bounding-boxes. In this paper, we present a brand new pose-based instance segmentation framework for humans which separates instances based on human pose, rather than proposal region detection. We demonstrate that our pose-based framework can achieve better accuracy than the state-of-art detection-based approach on the human instance segmentation problem, and can moreover better handle occlusion. Furthermore, there are few public datasets containing many heavily occluded humans along with comprehensive annotations, which makes this a challenging problem seldom noticed by researchers. Therefore, in this paper we introduce a new benchmark \"Occluded Human (OCHuman)\", which focuses on occluded humans with comprehensive annotations including bounding-box, human pose and instance masks. This dataset contains 8110 detailed annotated human instances within 4731 images. With an average 0.67 MaxIoU for each person, OCHuman is the most complex and challenging dataset related to human instance segmentation. Through this dataset, we want to emphasize occlusion as a challenging problem for researchers to study."}, "keywords": ["segmentation tasks", "Pose2Seg"], "citation_intent": "result"}
{"citing_id": "2304.05212v1", "cited_id": "1912.04958", "section_title": "Related Work 2.1. Ai-Synthesized Faces And Their Detection", "citation": "Inspired by the superior performance of the StyleGAN series #REFR in synthesizing highresolution and high-quality images, StyleGAN architectures have been adopted for image editing, achieving highquality edited images.", "text_before_citation": ["Artificial Intelligence (AI)-synthesized faces can be either fully synthetic when the faces are generated from scratch using generative models, or locally manipulated, e.g.", "when a single facial attribute or multiple attributes are modified by the model while the other attributes remain unchanged.", "A wide variety of generative models, notably GANs #OTHEREFR and diffusion models #OTHEREFR , are nowadays able to generate high-resolution images from scratch with an unprecedented level of realism."], "text_after_citation": ["Among them, we mention InterFace-GAN #OTHEREFR and StyleCLIP [27] .", "InterFaceGAN proposes a framework to interpret the disentangled face representation learned by the StyleGAN model and studies the properties of the facial semantics encoded in the latent space, showing that it is possible to edit the semantic attribute through linear subspace projection. StyleCLIP is a text-based interface for StyleGAN-based image manipulation.", "StyleCLIP mainly uses the Contrastive Language-Image Pre-training (CLIP) model to edit the latent code through the user input language description, so as to achieve the purpose of editing the image.", "With regard to the defences, several methods have been developed in the last years to discriminate between fake and real #OTHEREFR .", "Attempts have also been made to attribute the GAN model and/or the type of architecture generating the image."], "citing_paper_content": {"title": "Classification Of Gan-Based Image Manipulations Via A Vit-Based Hybrid Architecture", "abstract": "Classification of AI-manipulated content is receiving great attention, for distinguishing different types of manipulations. Most of the methods developed so far fail in the open-set scenario, that is when the algorithm used for the manipulation is not represented by the training set. In this paper, we focus on the classification of synthetic face generation and manipulation in open-set scenarios, and propose a method for classification with a rejection option. The proposed method combines the use of Vision Transformers (ViT) with a hybrid approach for simultaneous classification and localization. Feature map correlation is exploited by the ViT module, while a localization branch is employed as an attention mechanism to force the model to learn per-class discriminative features associated with the forgery when the manipulation is performed locally in the image. Rejection is performed by considering several strategies and analyzing the model output layers. The effectiveness of the proposed method is assessed for the task of classification of facial attribute editing and GAN attribution."}, "cited_paper_content": {"title": "Analyzing And Improving The Image Quality Of Stylegan", "abstract": "The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality."}, "keywords": ["highquality edited images", "StyleGAN architectures"], "citation_intent": "method"}
{"citing_id": "2304.11193v1", "cited_id": "1710.11252", "section_title": "Ii. Related Works", "citation": "To reduce uncertainty, which manifests as image blur #REFR , and produce sharper prediction image quality, recent models have estimated and sampled from these latent variables. Babaeizadeh et al.", "text_before_citation": ["To perform predictions over longer time horizons, Srivastava et al.", "#OTHEREFR introduced autoencoders and LSTM units to model the temporal coherence.", "Action-conditioned video prediction models provide further information to the prediction model [9; 23] for use in reinforcement learning and enabling model predictive control with video prediction systems #OTHEREFR . Villegas et al.", "#OTHEREFR split frames into content and motion streams and devolved the problem into predicting the pose and dynamics of landmarks #OTHEREFR .", "More recent methods have applied a stochastic assumption to the video prediction problem, stating that there are multiple possible outputs for a single given input, due to a set of latent variables."], "text_after_citation": ["#OTHEREFR applied this method to the optical flow method proposed in #OTHEREFR . Denton et al.", "#OTHEREFR proposed a similar approach but with a simpler model using only basic layers, which was later built upon by #OTHEREFR for larger high-fidelity video prediction. Lee et al.", "#OTHEREFR also built upon the method in #OTHEREFR by appending adversarial training techniques.", "The existing video prediction models do not use the tactile sensation and assume the changes caused by PRI are fully observed in visual information and/or the models provide a prediction with high uncertainty.", "b) Physical Robot Interaction: Opera et al."], "citing_paper_content": {"title": "Combining Vision And Tactile Sensation For Video Prediction", "abstract": "In this paper, we explore the impact of adding tactile sensation to video prediction models for physical robot interactions. Predicting the impact of robotic actions on the environment is a fundamental challenge in robotics. Current methods leverage visual and robot action data to generate video predictions over a given time period, which can then be used to adjust robot actions. However, humans rely on both visual and tactile feedback to develop and maintain a mental model of their physical surroundings. In this paper, we investigate the impact of integrating tactile feedback into video prediction models for physical robot interactions. We propose three multimodal integration approaches and compare the performance of these tactile-enhanced video prediction models. Additionally, we introduce two new datasets of robot pushing that use a magneticbased tactile sensor for unsupervised learning. The first dataset contains visually identical objects with different physical properties, while the second dataset mimics existing robot-pushing datasets of household object clusters. Our results demonstrate that incorporating tactile feedback into video prediction models improves scene prediction accuracy and enhances the agent's perception of physical interactions and understanding of causeeffect relationships during physical robot interactions. Index Terms-Deep learning in robotics and automation, perception for grasping and manipulation, physical robot interaction, video prediction, force and tactile sensing."}, "cited_paper_content": {"title": "Stochastic Variational Video Prediction", "abstract": "Predicting the future in real-world settings, particularly from raw sensory observations such as images, is exceptionally challenging. Real-world events can be stochastic and unpredictable, and the high dimensionality and complexity of natural images requires the predictive model to build an intricate understanding of the natural world. Many existing methods tackle this problem by making simplifying assumptions about the environment. One common assumption is that the outcome is deterministic and there is only one plausible future. This can lead to low-quality predictions in real-world settings with stochastic dynamics. In this paper, we develop a stochastic variational video prediction (SV2P) method that predicts a different possible future for each sample of its latent variables. To the best of our knowledge, our model is the first to provide effective stochastic multi-frame prediction for real-world video. We demonstrate the capability of the proposed method in predicting detailed future frames of videos on multiple real-world datasets, both action-free and action-conditioned. We find that our proposed method produces substantially improved video predictions when compared to the same model without stochasticity, and to other stochastic video prediction methods. Our SV2P implementation will be open sourced upon publication."}, "keywords": ["sharper prediction image"], "citation_intent": "background"}
{"citing_id": "2303.04226v1", "cited_id": "1906.08237", "section_title": "Foundation Model", "citation": "XL-Net #REFR , which is also based on BERT, incorporates permutation operations to change the prediction order for each training iteration, allowing the model to learn more information across tokens.", "text_before_citation": ["Since the introduction of the Transformer architecture, it has become the dominant choice in natural language processing due to its parallelism and learning capabilities.", "Generally, these transformer based pre-trained language models can be commonly classified into two types based on their training tasks: autoregressive language modeling and masked language modeling #OTHEREFR .", "Given a sentence, which is composed of several tokens, the objective of masked language modeling, e.g., BERT #OTHEREFR and RoBERTa #OTHEREFR , refers to predicting the probability of a masked token given context information.", "The most notable example of masked language modeling is BERT #OTHEREFR , which includes masked language modeling and next sentence prediction tasks.", "RoBERTa #OTHEREFR , which uses the same architecture as BERT, improves its performance by increasing the amount of pre-training data and incorporating more challenging pre-training objectives."], "text_after_citation": ["While autoregressive language modeling, e.g., GPT-3 #OTHEREFR and OPT [45] , is to model the probability of the next token given previous tokens, hence, left-to-right language modeling.", "Different from masked language models, autoregressive models are more suitable for generative tasks.", "We will introduce more about autoregressive models in Section 4.1.1.", "Decoder (GPT) ! \" # Trm Trm Trm Trm Trm Trm ! \" # \u2026 \u2026 \u2026 \u2026 ! \" # Trm Trm Trm Trm Trm Trm ! \" # \u2026 \u2026 \u2026 \u2026 Encoder (BERT) ! \" # \u2026 Trm Trm Trm \u2026 Trm Trm Trm \u2026 Trm Trm Trm \u2026 ! \" # \u2026 Encoder-Decoder (T5/BART)"], "citing_paper_content": {"title": "A Comprehensive Survey Of Ai-Generated Content (Aigc): A History Of Generative Ai From Gan To Chatgpt", "abstract": "Recently, ChatGPT, along with DALL-E-2 [1] and Codex [2],has been gaining significant attention from society. As a result, many individuals have become interested in related resources and are seeking to uncover the background and secrets behind its impressive performance. In fact, ChatGPT and other Generative AI (GAI) techniques belong to the category of Artificial Intelligence Generated Content (AIGC), which involves the creation of digital content, such as images, music, and natural language, through AI models. The goal of AIGC is to make the content creation process more efficient and accessible, allowing for the production of high-quality content at a faster pace. AIGC is achieved by extracting and understanding intent information from instructions provided by human, and generating the content according to its knowledge and the intent information. In recent years, large-scale models have become increasingly important in AIGC as they provide better intent extraction and thus, improved generation results. With the growth of data and the size of the models, the distribution that the model can learn becomes more comprehensive and closer to reality, leading to more realistic and high-quality content generation. This survey provides a comprehensive review on the history of generative models, and basic components, recent advances in AIGC from unimodal interaction and multimodal interaction. From the perspective of unimodality, we introduce the generation tasks and relative models of text and image. From the perspective of multimodality, we introduce the cross-application between the modalities mentioned above. Finally, we discuss the existing open problems and future challenges in AIGC."}, "cited_paper_content": {"title": "Xlnet: Generalized Autoregressive Pretraining For Language Understanding", "abstract": "With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves better performance than pretraining approaches based on autoregressive language modeling. However, relying on corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model, into pretraining. Empirically, under comparable experiment setting, XLNet outperforms BERT on 20 tasks, often by a large margin, including question answering, natural language inference, sentiment analysis, and document ranking."}, "keywords": ["model", "training iteration"], "citation_intent": "method"}
{"citing_id": "2304.08880v1", "cited_id": "1605.08695", "section_title": "Experiment Setup", "citation": "We use CentOS 7 for the operating system, Docker 19 for deployment, and TensorFlow 2.2 #REFR for training and embedding generation.", "text_before_citation": ["For training and testing, we partitioned the graph samples collected from 200M instructions with 70% for training and 30% for testing.", "The entire training takes less than 20 hours on Nvidia Tesla V100.", "After the model is trained, we deploy it to our prototype system to generate NPS embeddings for hundreds of billions of instructions. System Prototype.", "Our prototype system runs on 10 Intel Xeon Platinum 8163 servers.", "Each server has 48 CPU cores running on 2.5GHz with 800GB of memory."], "text_after_citation": [], "citing_paper_content": {"title": "Nps: A Framework For Accurate Program Sampling Using Graph Neural Network", "abstract": "With the end of Moore's Law, there is a growing demand for rapid architectural innovations in modern processors, such as RISC-V custom extensions, to continue performance scaling. Program sampling is a crucial step in microprocessor design, as it selects representative simulation points for workload simulation. While SimPoint has been the de-facto approach for decades, its limited expressiveness with Basic Block Vector (BBV) requires time-consuming human tuning, often taking months, which impedes fast innovation and agile hardware development. This paper introduces Neural Program Sampling (NPS), a novel framework that learns execution embeddings using dynamic snapshots of a Graph Neural Network. NPS deploys AssemblyNet for embedding generation, leveraging an application's code structures and runtime states. AssemblyNet serves as NPS's graph model and neural architecture, capturing a program's behavior in aspects such as data computation, code path, and data flow. AssemblyNet is trained with a data prefetch task that predicts consecutive memory addresses. In the experiments, NPS outperforms SimPoint by up to 63%, reducing the average sampling error by 38%. Additionally, NPS demonstrates strong robustness with increased accuracy, reducing the expensive accuracy tuning overhead. Furthermore, NPS shows higher accuracy and generality than the state-of-the-art GNN approach in code behavior learning, enabling the generation of high-quality execution embeddings."}, "cited_paper_content": {"title": "Tensorflow: A System For Large-Scale Machine Learning", "abstract": "TensorFlow is a machine learning system that operates at large scale and in heterogeneous environments. Tensor-Flow uses dataflow graphs to represent computation, shared state, and the operations that mutate that state. It maps the nodes of a dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, general-purpose GPUs, and custom-designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility to the application developer: whereas in previous \"parameter server\" designs the management of shared state is built into the system, TensorFlow enables developers to experiment with novel optimizations and training algorithms. TensorFlow supports a variety of applications, with a focus on training and inference on deep neural networks. Several Google services use TensorFlow in production, we have released it as an open-source project, and it has become widely used for machine learning research. In this paper, we describe the TensorFlow dataflow model and demonstrate the compelling performance that TensorFlow achieves for several real-world applications."}, "keywords": ["TensorFlow"], "citation_intent": "method"}
{"citing_id": "2304.10168v1", "cited_id": "2003.12039", "section_title": "Context Adaptation And Propagation", "citation": "The frame flow is computed on x s t\u22121 and x s t using image-based flow estimator #REFR .", "text_before_citation": ["We introduce the Context Adaptation and Propagation (CAP) module to make the model produce smooth videos. The illustration of CAP is shown in Fig. 5 .", "First, the raw feature f r t\u22121 of the previous source frame is sequentially warped with the frame flow w t\u22121\u2192t and the current local deformation w s\u2192d local (Two-Step Warping)."], "text_after_citation": ["Second, the warped previous raw featur\u1ebd f r t\u22121 , hidden feature h t\u22121 , adapted feature f c t\u22121 and the current aligned feature f a t are concatenated.", "We further feed the concatenated feature to the Context Adaptation module, which is composed of several convolution blocks, to get the feature in the same spatial and channel size with f a t .", "Then, the feature is further refined with the Feature Refinement module.", "After that, we get the adapted feature f c t for the current frame, and the hidden feature is updated with the adapted feature f c t using a ConvGRU block #OTHEREFR . The proposed CAP module can be formed as,", "EQUATION"], "citing_paper_content": {"title": "High-Fidelity And Freely Controllable Talking Head Video Generation", "abstract": "PECHead (a). Frontalization Input Yaw Pitch Roll Input Neutral Smile Laugh Input Smile Expressionless Roll (b). Free pose and expression editing Figure 1. Presented herein are representative results showcasing the effectiveness of our proposed method in the tasks of frontalization, as well as free pose and expression editing. (a) FOMM [46] often produces face distortion issues, while our model PECHead generates high-fidelity results. (b) Our proposed framework empowers the generation of talking head videos that offer free control over the head pose and expression. More results can be found in the supplementary materials."}, "cited_paper_content": {"title": "Raft: Recurrent All-Pairs Field Transforms For Optical Flow", "abstract": "We introduce Recurrent All-Pairs Field Transforms (RAFT), a new deep network architecture for optical flow. RAFT extracts per-pixel features, builds multi-scale 4D correlation volumes for all pairs of pixels, and iteratively updates a flow field through a recurrent unit that performs lookups on the correlation volumes. RAFT achieves state-of-the-art performance, with strong cross-dataset generalization and high efficiency in inference time, training speed, and parameter count. Code is available \\url{https://github.com/princeton-vl/RAFT}."}, "keywords": ["frame flow", "image-based flow estimator"], "citation_intent": "method"}
{"citing_id": "2304.04668v1", "cited_id": "1703.03400", "section_title": "Evaluating Meta-Learning For The Principal", "citation": "We now compare the learning behavior of a principal that is meta-trained (MAML) #REFR versus one trained with standard policy gradients (RL), in a simple Stackelberg game between a principal and an agent.", "text_before_citation": [], "text_after_citation": ["The Stackelberg game features a principal (leader) who acts first (intervene or not), and the agent (follower) acts second according to a best-response policy (Von Stackelberg, 2010).", "Our goal is to learn an intervention policy that can adapt to different agent types and find the Stackelberg equilibrium.", "The agent's actions are \"cooperate\" and \"defect\", while the principal can choose whether or not to intervene.", "Assuming the row player is the agent and the column player is the principal, we define a 2 \u00d7 2 payoff matrix", "No intervention (NI) Intervene (IN) Cooperate (C) u, 1 u + 1, 1 \u2212 c Defect (D) 1 \u2212 u, 0 \u2212u, \u2212c ,"], "citing_paper_content": {"title": "Mermaide: Learning To Align Learners Using Model-Based Meta-Learning", "abstract": "We study how a principal can efficiently and effectively intervene on the rewards of a previously unseen learning agent in order to induce desirable outcomes. This is relevant to many realworld settings like auctions or taxation, where the principal may not know the learning behavior nor the rewards of real people. Moreover, the principal should be few-shot adaptable and minimize the number of interventions, because interventions are often costly. We introduce MER-MAIDE, a model-based meta-learning framework to train a principal that can quickly adapt to out-ofdistribution agents with different learning strategies and reward functions. We validate this approach step-by-step. First, in a Stackelberg setting with a best-response agent, we show that meta-learning enables quick convergence to the theoretically known Stackelberg equilibrium at test time, although noisy observations severely increase the sample complexity. We then show that our model-based meta-learning approach is cost-effective in intervening on bandit agents with unseen explore-exploit strategies. Finally, we outperform baselines that use either meta-learning or agent behavior modeling, in both 0-shot and K = 1-shot settings with partial agent information."}, "cited_paper_content": {"title": "Model-Agnostic Meta-Learning For Fast Adaptation Of Deep Networks", "abstract": "We propose an algorithm for meta-learning that is model-agnostic, in the sense that it is compatible with any model trained with gradient descent and applicable to a variety of different learning problems, including classification, regression, and reinforcement learning. The goal of meta-learning is to train a model on a variety of learning tasks, such that it can solve new learning tasks using only a small number of training samples. In our approach, the parameters of the model are explicitly trained such that a small number of gradient steps with a small amount of training data from a new task will produce good generalization performance on that task. In effect, our method trains the model to be easy to fine-tune. We demonstrate that this approach leads to state-of-the-art performance on two few-shot image classification benchmarks, produces good results on few-shot regression, and accelerates fine-tuning for policy gradient reinforcement learning with neural network policies."}, "keywords": ["meta-trained (MAML"], "citation_intent": "method"}
{"citing_id": "2303.16886v1", "cited_id": "1810.04805", "section_title": "Iv. Main Experiments And Results", "citation": "We show the scores obtained through both the 3-class and 2-class modeling in Table III #REFR . We note that both approaches achieved similar scores.", "text_before_citation": ["The natural way to model the original 3-class setup is to use all three tags (from Table II ), which inherently gives per-class F1-scores.", "But to compute the ANY-COMB scores, we just collapse the POS and COMB labels to the same ANY-COMB label and compute the ANY-COMB F1-score.", "A slightly different way is to use a 2-class set up during the training and testing through a different labeling scheme upfront.", "For the POS classifier, we simply use two tags: @POS@ and @NON-POS@, where the latter is the tag given to instances in the union of COMB and NOCOMB.", "For the ANY-COMB classifier, we use a new tag @ANY-COMB@ for instances in the union of POS and COMB classes and the @NOCOMB@ is the tag for the other class."], "text_after_citation": ["The the 3-way setup results in a slightly better POS score and the 2-way labeling improves by 1.6% in ANY-COMB F1-score compared with the 3-class approach.", "Precision and recall are also close to each other but overall recall is slightly less than precision.", "Next, we compare our results with prior state-of-the-art in Table IV , where the last row contains the F1-scores of our 3-way model (from Table III ).", "We see that for both POS and ANY-COMB settings, our Seq2Rel model has better performance.", "Compared with the penultimate row, which shows the prior SoTA score on this dataset, we have an almost 5% improvement in POS score even though our model is end-to-end while all other rows in the table assume drugs names are spotted."], "citing_paper_content": {"title": "End-To-End N-Ary Relation Extraction For Combination Drug Therapies", "abstract": "Combination drug therapies are treatment regimens that involve two or more drugs, administered more commonly for patients with cancer, HIV, malaria, or tuberculosis. Currently there are over 350K articles in PubMed that use the combination drug therapy MeSH heading with at least 10K articles published per year over the past two decades. Extracting combination therapies from scientific literature inherently constitutes an n-ary relation extraction problem. Unlike in the general n-ary setting where n is fixed (e.g., drug-gene-mutation relations where n = 3), extracting combination therapies is a special setting where n \u2265 2 is dynamic, depending on each instance. Recently, Tiktinsky et al. (NAACL 2022) introduced a first of its kind dataset, CombDrugExt, for extracting such therapies from literature. Here, we use a sequence-to-sequence style end-to-end extraction method to achieve an F1-Score of 66.7% on the CombDrugExt test set for positive (or effective) combinations. This is an absolute \u2248 5% F1score improvement even over the prior best relation classification score with spotted drug entities (hence, not end-to-end). Thus our effort introduces a state-of-the-art first model for end-to-end extraction that is already superior to the best prior non end-toend model for this task. Our model seamlessly extracts all drug entities and relations in a single pass and is highly suitable for dynamic n-ary extraction scenarios."}, "cited_paper_content": {"title": "Bert: Pre-Training Of Deep Bidirectional Transformers For Language Understanding", "abstract": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. ::: BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."}, "keywords": ["2-class modeling"], "citation_intent": "result"}
{"citing_id": "2304.07832v1", "cited_id": "1403.0361", "section_title": "Kernel Operators And Delay Coordinates", "citation": "In this setting, kernel operators have demonstrated an impressive ability in capturing complex nonlinear phenomena #REFR .", "text_before_citation": ["Kernel methods have a long-standing history in machine learning #OTHEREFR , with established applications in detecting nonlinear data manifolds and order reduction [Coifman and Lafon, 2006a] .", "Recently, researchers have studied dynamics-dependent kernels in the context of dynamical systems #OTHEREFR , with a particular focus on delay coordinate space."], "text_after_citation": ["Given a a collection of N samples", "F (x 1 ), ..., F (x N )", ", organized in a time-ordered manner, where each F (x i ) \u2208 R d .", "The value of x n is determined by the function \u03a6 n\u2206t (x 0 ), where \u2206t is the interval at which the data is sampled.", "A delay coordinate map is constructed from F by embedding X in a manifold in R Qd as"], "citing_paper_content": {"title": "Characterizing The Load Profile In Power Grids By Koopman Mode Decomposition Of Interconnected Dynamics", "abstract": "Electricity load forecasting is crucial for effectively managing and optimizing power grids. Over the past few decades, various statistical and deep learning approaches have been used to develop load forecasting models. This paper presents an interpretable machine learning approach that identifies load dynamics using data-driven methods within an operator-theoretic framework. We represent the load data using the Koopman operator, which is inherent to the underlying dynamics. By computing the corresponding eigenfunctions, we decompose the load dynamics into coherent spatiotemporal patterns that are the most robust features of the dynamics. Each pattern evolves independently according to its single frequency, making its predictability based on linear dynamics. We emphasize that the load dynamics are constructed based on coherent spatiotemporal patterns that are intrinsic to the dynamics and are capable of encoding rich dynamical features at multiple time scales. These features are related to complex interactions over interconnected power grids and different exogenous effects. To implement the Koopman operator approach more efficiently, we cluster the load data using a modern kernel-based clustering approach and identify power stations with similar load patterns, particularly those with synchronized dynamics. We evaluate our approach using a large-scale dataset from a renewable electric power system within the continental European electricity system and show that the Koopman-based approach outperforms a deep learning (LSTM) architecture in terms of accuracy and computational efficiency."}, "cited_paper_content": {"title": "Dynamics-Adapted Cone Kernels", "abstract": "We present a family of kernels for analysis of data generated by dynamical systems. These so-called cone kernels feature a dependence on the dynamical vector field operating in the phase space manifold, estimated empirically through finite differences of time-ordered data samples. In particular, cone kernels assign strong affinity to pairs of samples whose relative displacement vector lies within a narrow cone aligned with the dynamical vector field. The outcome of this explicit dependence on the dynamics is that, in a suitable asymptotic limit, Laplace--Beltrami operators for data analysis constructed from cone kernels generate diffusions along the integral curves of the dynamical vector field. This property is independent of the observation modality, and endows these operators with invariance under a weakly restrictive class of transformations of the data (including conformal transformations), while it also enhances their capability to extract intrinsic dynamical timescales via eigenfunctions. Here, we ..."}, "keywords": ["kernel operators"], "citation_intent": "background"}
{"citing_id": "2304.07696v1", "cited_id": "1507.04452", "section_title": "B. Uncoded System: Symbol Error Rate", "citation": "In our simulations, the one-bit channel estimation method developed in #REFR is adopted to provide the estimated CSI.", "text_before_citation": ["To evaluate the data detection performance of the proposed methods in the multiuser massive MIMO system, we compare the following detection methods: 1) Naive learning-based one-bit ML 2) ADL-based one-bit ML (proposed) 3) ADL-based one-bit ML with estimated SNR (proposed) 4) Minimum-Center-Distance (MCD) #OTHEREFR 5) One-bit ZF with perfect CSI #OTHEREFR 6) One-bit ML with perfect CSI (optimal one-bit detection) 7) One-bit ML with estimated CSI 8) Infinite-bit ML with perfect CSI (optimal detection) We note that the learning-based methods: 1) Naive onebit ML, 2) ADL one-bit ML, 3) ADL one-bit ML with estimated SNR, and 4) MCD, do not require explicit channel estimation; however, the other methods either assume perfect CSI or estimated CSI at the AP.", "The learning-based methods transmit N tr pilot signals per each training symbol vector, which requires KN tr pilot signals in total.", "Accordingly, we consider that the conventional one-bit ML detection with an estimated channel also uses KN tr pilot signals to estimate the channel."], "text_after_citation": ["For readability of the curves, we compare MCD for the 16-QAM case shown in Fig. 8 . Fig.", "6 presents the SER results for N r = 32 antennas, N u = 4 users, N tr = 45 pilot signals, and 4-QAM. As expected from Fig.", "5 , the naive-learning approach shows the catastrophic result from the medium to high SNR due to the large number of zero-valued likelihood functions.", "The onebit ZF detection that applies the pseudo-inverse matrix of the perfectly-known channel matrix onto the one-bit observations shows the large performance degradation with the error floor at the medium and high SNR regime.", "The one-bit ML detection with the one-bit estimated channels shows a larger deviation from the optimal one-bit ML detection with perfect CSI as the SNR increases due to the channel estimation error."], "citing_paper_content": {"title": "Adaptive Learning-Based Maximum Likelihood And Channel-Coded Detection For Massive Mimo Systems With One-Bit Adcs", "abstract": "In this paper, we propose a learning-based detection framework for uplink massive multiple-input and multiple-output (MIMO) systems with one-bit analog-to-digital converters. The learning-based detection only requires counting the occurrences of the quantized outputs of-1 and +1 for estimating a likelihood probability at each antenna. Accordingly, the key advantage of this approach is to perform maximum likelihood detection without explicit channel estimation which has been one of the primary challenges of one-bit quantized systems. The learning in the high signal-to-noise ratio (SNR) regime, however, needs excessive training to estimate the extremely small likelihood probabilities. To address this drawback, we propose a dither-andlearning technique to estimate likelihood functions from dithered signals. First, we add a dithering signal to artificially decrease the SNR and then infer the likelihood function from the quantized dithered signals by using an SNR estimate derived from a deep neural network-based offline estimator. We extend our technique by developing an adaptive dither-and-learning method that updates the dithering power according the patterns observed in the quantized dithered signals. The proposed framework is also applied to state-of-the-art channel-coded MIMO systems by computing a bit-wise and user-wise log-likelihood ratio from the refined likelihood probabilities. Simulation results validate the detection performance of the proposed methods in both uncoded and coded systems. Index Terms-Massive MIMO, one-bit ADC, dithering, maximum likelihood detection, deep neural network. I. INTRODUCTION Massive MIMO systems for sub-6 GHz wireless communications [2], [3] and millimeter wave (mmWave) communications [4]-[6] have been considered as one of the emerging technologies for future communications because of the outstanding gain in spectral efficiency and capacity [7]. As wireless communication systems continue to grow in popularity and importance, there is a need to investigate communication systems that are not only reliable and high-performing, but also energy-efficient for various future wireless applications such as internet-of-things (IoT), extended reality, and smart grid [8], [9]. Because of the small wavelength of mmWave"}, "cited_paper_content": {"title": "Near Maximum-Likelihood Detector And Channel Estimator For Uplink Multiuser Massive Mimo Systems With One-Bit Adcs", "abstract": "In massive multiple-input multiple-output (MIMO) systems, it may not be power efficient to have a pair of high-resolution analog-to-digital converters (ADCs) for each antenna element. In this paper, a near maximum likelihood (nML) detector for uplink multiuser massive MIMO systems is proposed where each antenna is connected to a pair of one-bit ADCs, i.e., one for each real and imaginary component of the baseband signal. The exhaustive search over all the possible transmitted vectors required in the original maximum likelihood (ML) detection problem is relaxed to formulate an ML estimation problem. Then, the ML estimation problem is converted into a convex optimization problem which can be efficiently solved. Using the solution, the base station can perform simple symbol-by-symbol detection for the transmitted signals from multiple users. To further improve detection performance, we also develop a two-stage nML detector that exploits the structures of both the original ML and the proposed (one-stage) nML detectors. Numerical results show that the proposed nML detectors are efficient enough to simultaneously support multiple uplink users adopting higher-order constellations, e.g., 16 quadrature amplitude modulation. Since our detectors exploit the channel state information as part of the detection, an ML channel estimation technique with one-bit ADCs that shares the same structure with our proposed nML detector is also developed. The proposed detectors and channel estimator provide a complete low power solution for the uplink of a massive MIMO system."}, "keywords": ["one-bit channel estimation"], "citation_intent": "method"}
{"citing_id": "2303.01928v1", "cited_id": "1411.7766", "section_title": "Fairshap When A = Y And Fair T", "citation": "The training set D is derived from 2 publicly available, largescale (and biased) image datasets, CelebA and LFWA #REFR .", "text_before_citation": ["In this scenario, the model's task is to predict the sensitive attribute, i.e.", "A = Y , and the reference dataset T is fair.", "Specifically, we perform a gender classification task from facial images by means of a deep convolutional network using FairShap for data re-weighting.", "We use Inception Resnet V1 #OTHEREFR as the gender classification model."], "text_after_citation": ["FairFace #OTHEREFR is the reference (and fair) dataset T .", "Gender is both the protected attribute (A) and the target variable (Y ). The pipeline (see Fig.", "2 for an illustration) to obtain the instance-level pre-processing weights in this scenario is as follows:", "(1) we pre-train an Inception-v1 model with the CelebA dataset;", "(2) we use this model to obtain the embeddings for the LFWA and FairFace datasets;"], "citing_paper_content": {"title": "Fairshap: A Data Re-Weighting Approach For Algorithmic Fairness Based On Shapley Values", "abstract": "In this paper, we propose FairShap, a novel and interpretable pre-processing (re-weighting) method for fair algorithmic decision-making through data valuation. FairShap is based on the Shapley Value, a well-known mathematical framework from game theory to achieve a fair allocation of resources. Our approach is easily interpretable, as it measures the contribution of each training data point to a predefined fairness metric. We empirically validate FairShap on several state-of-the-art datasets of different nature, with different training scenarios and models. The proposed approach outperforms other methods, yielding significantly fairer models with similar levels of accuracy. In addition, we illustrate FairShap's interpretability by means of histograms and latent space visualizations. We believe this work represents a promising direction in interpretable, model-agnostic approaches to algorithmic fairness."}, "cited_paper_content": {"title": "Deep Learning Face Attributes In The Wild", "abstract": "Predicting face attributes in the wild is challenging due to complex face variations. We propose a novel deep learning framework for attribute prediction in the wild. It cascades two CNNs, LNet and ANet, which are fine-tuned jointly with attribute tags, but pre-trained differently. LNet is pre-trained by massive general object categories for face localization, while ANet is pre-trained by massive face identities for attribute prediction. This framework not only outperforms the state-of-the-art with a large margin, but also reveals valuable facts on learning face representation. (1) It shows how the performances of face localization (LNet) and attribute prediction (ANet) can be improved by different pre-training strategies. (2) It reveals that although the filters of LNet are fine-tuned only with image-level attribute tags, their response maps over entire images have strong indication of face locations. This fact enables training LNet for face localization with only image-level annotations, but without face bounding boxes or landmarks, which are required by all attribute recognition works. (3) It also demonstrates that the high-level hidden neurons of ANet automatically discover semantic concepts after pre-training with massive face identities, and such concepts are significantly enriched after fine-tuning with attribute tags. Each attribute can be well explained with a sparse linear combination of these concepts."}, "keywords": ["biased) image datasets"], "citation_intent": "method"}
{"citing_id": "2304.11341v1", "cited_id": "2003.03964", "section_title": "", "citation": "More specifically, in #REFR , several relaying schemes were developed to address the blockage issue in THz communications, in which the performance gain from using relaying was rigorously evaluated.", "text_before_citation": ["In #OTHEREFR , to capture the obstruction of walls and human bodies, etc., the coverage probability was studied by using the stochastic geometry.", "The analytical results revealed the detrimental impact of the node density, transmission dis-tance, and atmosphere upon the coverage probability.", "Recently, some prior works have tried to solve this dilemma.", "In #OTHEREFR , massive multiple-input multiple-output (MIMO) aided THz communications were proposed for the blockage mitigation and a network sum-rate based optimization scheme was devised by considering only statistical channel state information (CSI).", "Moreover, in order to substantially overcome blockage impairment, one basic idea is altering the communication links so as to circumvent harsh propagation environment, which can be realized by combining relaying, reconfigurable intelligent surfaces (RIS), unmanned aerial vehicles (UAVs), etc. #OTHEREFR ."], "text_after_citation": ["In #OTHEREFR , RIS was employed to aid THz communications and it is uncovered that the THz communications performance can be significantly enhanced by increasing the number of RIS's elements.", "Besides, a block coordinate searching algorithm was developed to jointly optimize the coordinates and phase shifts of RIS for THz communications in #OTHEREFR .", "Some other latest works #OTHEREFR , #OTHEREFR also verified that integrating RIS into THz systems not only avoids the effect of blockage but also expands the coverage area.", "Moreover, in #OTHEREFR , the UAV was deployed to assist THz communications and the total transmission delay between the UAV and the ground user was minimized through optimizing the location, bandwidth and power, which can be solved with alternating algorithm.", "Unfortunately, most of prior optimization methodologies for designing THz communication systems have a prohibitively high computational complexity because of the complex fading characteristics of THz communications."], "citing_paper_content": {"title": "Performance Analysis And Optimal Design Of Harq-Ir-Aided Terahertz Communications", "abstract": "Terahertz (THz) communications are envisioned to be a promising technology for 6G thanks to its broad bandwidth. However, the large path loss, antenna misalignment, and atmospheric influence of THz communications severely deteriorate its reliability. To address this, hybrid automatic repeat request (HARQ) is recognized as an effective technique to ensure reliable THz communications. This paper delves into the performance analysis of HARQ with incremental redundancy (HARQ-IR)-aided THz communications in the presence/absence of blockage. More specifically, the analytical expression of the outage probability of HARQ-IR-aided THz communications is derived, with which the asymptotic outage analysis is enabled to gain meaningful insights, including diversity order, power allocation gain, modulation and coding gain, etc. Then the long term average throughput (LTAT) is expressed in terms of the outage probability based on renewal theory. Moreover, to combat the blockage effects, a multi-hop HARQ-IR-aided THz communication scheme is proposed and its performance is examined. To demonstrate the superiority of the proposed scheme, the other two HARQ-aided schemes, i.e., Type-I HARQ and HARQ with chase combining (HARQ-CC), are used for benchmarking in the simulations. In addition, a deep neural network (DNN) based outage evaluation framework with low computational complexity is devised to reap the benefits of using both asymptotic and simulation results in low and high outage regimes, respectively. This novel outage evaluation framework is finally employed for the optimal rate selection, which outperforms the asymptotic based optimization. Index Terms-Blockage, hybrid automatic repeat request (HARQ), outage probability, terahertz (THz) communications. I. INTRODUCTION T O meet the unprecedented increase of wireless devices and data traffic, there is an urgent need to boost the transmission rate. To accommodate the ever-growing demand, terahertz (THz) communications have emerged as a potential technique to support the peak data rate of more than 1 Tbps [1]-[3]. Different from the conventional low-frequency channel propagation characteristics, THz communications are susceptible to atmospheric turbulence, pointing errors, rain attenuation, etc. [4]-[12]. As uncovered by [13], [14], these"}, "cited_paper_content": {"title": "Relay-Based Blockage And Antenna Misalignment Mitigation In Thz Wireless Communications", "abstract": "The proliferation of wireless devices in recent years has caused a spectrum shortage, which led the scientific community to explore the potential of using terahertz (THz) communications. However, THz systems suffer from severe path attenuation, blockage, and antenna misalignment. In this paper, we present a relay-based blockage and antenna misalignment mitigation approach. In more detail, two relay selection policies are employed, namely best and random relay selection. The system performance under both policies is evaluated and compared in terms of average throughput and the probability that throughput of a link is below the quality of service (QoS) threshold, using Monte Carlo simulations. It was observed that the effect of both blockage and misalignment can be mitigated using relays. Moreover, the gain of using relaying to mitigate blockage is much more significant."}, "keywords": ["THz communications"], "citation_intent": "method"}
{"citing_id": "2303.00152v1", "cited_id": "0902.2137", "section_title": "Introduction", "citation": "One solution is to design a provably correct compiler, but this is a complex and long-term endeavour #REFR .", "text_before_citation": ["However, the original specification in the Yellow Paper #OTHEREFR has some known shortcomings: (i) it is hard to read and does not provide a formal semantics of the EVM and the bytecode; (ii) the lack of a formal semantics makes it hard for Ethereum client developers to guarantee that they interpret the Yellow Paper in a consistent way; (iii) designing compilers from high-level languages (e.g., Solidity 1 ) to EVM bytecode without a formal semantics is error-prone and, without a precise semantics of the EVM, it is hard to design certified compilers (preserving of semantics from a high to a low-level language.).", "One can argue that existing implementations of the EVM (e.g., in Go, Java) provide a de facto semantics for it.", "Whilst this is true to some extent, such implementations do not enable formal reasoning about bytecode.", "Furthermore, whilst smart contracts can be written in highlevel languages like Solidity, they must be compiled into EVM bytecode before being executed on the EVM.", "Tools for checking safety properties (e.g., absence of overflow, division by zero, etc) at the Solidity level are problematic if they cannot guarantee such properties hold at the bytecode level."], "text_after_citation": ["Alternatively we can provide techniques, supported by tools, to reason about properties of the bytecode.", "This is what we propose to do in this work.", "Our Contribution.", "We present a complete and formal specification of the EVM in Dafny, available at https://github.com/ConsenSys/evm-dafny.", "We provide a formal semantics where the meaning of an instruction is given as a partial function that maps states to states."], "citing_paper_content": {"title": "Formal And Executable Semantics Of The Ethereum Virtual Machine In Dafny", "abstract": "The Ethereum protocol implements a replicated state machine. The network participants keep track of the system state by: 1) agreeing on the sequence of transactions to be processed and 2) computing the state transitions that correspond to the sequence of transactions. Ethereum transactions are programs, called smart contracts, and computing a state transition requires executing some code. The Ethereum Virtual Machine (EVM) provides this capability and can execute programs written in EVM bytecode. We present a formal and executable semantics of the EVM written in the verification-friendly language Dafny: it provides (i) a readable, formal and verified specification of the semantics of the EVM; (ii) a framework to formally reason about bytecode."}, "cited_paper_content": {"title": "A Formally Verified Compiler Back-End", "abstract": "This article describes the development and formal verification (proof of semantic preservation) of a compiler back-end from Cminor (a simple imperative intermediate language) to PowerPC assembly code, using the Coq proof assistant both for programming the compiler and for proving its soundness. Such a verified compiler is useful in the context of formal methods applied to the certification of critical software: the verification of the compiler guarantees that the safety properties proved on the source code hold for the executable compiled code as well."}, "keywords": ["provably correct compiler"], "citation_intent": "background"}
{"citing_id": "2303.18008v1", "cited_id": "1907.08063", "section_title": "Theorem 10. The Capacity Of The Trapdoor Channel With Delayed Feedback Of Two Time Instances Is", "citation": "As explained in Remark 2, the scheme is based on the posterior matching principle, and the exact details regarding the constriction of the coding scheme are given in #REFR .", "text_before_citation": ["In other words, the delayed feedback capacity in Theorem 10 does not seem to be the feedforward capacity, which remains an open problem.", "Remark 6.", "The achievability proof of Theorem 10 is based on the Q-graph lower bound, which was presented in Section IV-B.", "That is, the lower bound was established by showing that a particular graphbased encoder, given by the Q-graph in #OTHEREFR and the input distribution in #OTHEREFR , provides an achievable rate of log 2 (3/2).", "This graph-based encoder implies a simple coding scheme that, in our case, achieves the capacity."], "text_after_citation": ["Remark 7.", "Following the formulation in Section IV-A, we present in Fig.", "5 the trapdoor channel with delayed feedback of two time instances as a new unifilar FSC with instantaneous feedback.", "For the new FSC, it is interesting to note that the capacity of each individual channel (per state) is zero.", "Nevertheless, as we already demonstrated, the capacity of the overall FSC is not zero."], "citing_paper_content": {"title": "Capacity Of Finite-State Channels With Delayed Feedback", "abstract": "In this paper, we investigate the capacity of finite-state channels (FSCs) in presence of delayed feedback. We show that the capacity of a FSC with delayed feedback can be computed as that of a new FSC with instantaneous feedback and an extended state. Consequently, graph-based methods to obtain computable upper and lower bounds on the delayed feedback capacity of unifilar FSCs are proposed. Based on these methods, we establish that the capacity of the trapdoor channel with delayed feedback of two time instances is given by log 2 3 2. In addition, we derive an analytical upper bound on the delayed feedback capacity of the binary symmetric channel with a no consecutive ones input constraint. This bound also serves as a novel upper bound on its non-feedback capacity, which outperforms all previously known bounds. Lastly, we demonstrate that feedback does improve the capacity of the dicode erasure channel. This paper was presented in part at the 2022 IEEE International Symposium on Information Theory [1]. 1 The Q-graph is an auxiliary directed graph that is used to map channel output sequences onto one of the graph nodes (e.g. see Fig. 4)."}, "cited_paper_content": {"title": "Graph-Based Encoders And Their Performance For Finite-State Channels With Feedback", "abstract": "The capacity of unifilar finite-state channels in the presence of feedback is investigated. We derive a new evaluation method to extract graph-based encoders with their achievable rates, and to compute upper bounds to examine their performance. The evaluation method is built upon a recent methodology to derive simple bounds on the capacity using auxiliary directed graphs. While it is not clear whether the upper bound is convex, we manage to formulate it as a convex optimization problem using transformation of the argument with proper constraints. The lower bound is formulated as a non-convex optimization problem, yet, any feasible point to the optimization problem induces a graph-based encoder. In all examples, the numerical results show near-tight upper and lower bounds that can be easily converted to analytic results. For the non-symmetric Trapdoor channel and binary fading channels (BFCs), new capacity results are established by computing the corresponding bounds. For all other instances, including the Ising channel, the near-tightness of the achievable rates is shown via a comparison with corresponding upper bounds. Finally, we show that any graph-based encoder implies a simple coding scheme that is based on the posterior matching principle and achieves the lower bound."}, "keywords": ["coding scheme"], "citation_intent": "method"}
{"citing_id": "2304.10036v1", "cited_id": "1911.13299", "section_title": "Ignoring Specific Attributes", "citation": "We observe that all feature extractors considered can somewhat succeed at the task, including the ResNet-50 with random weights, which we expect to still produce valuable features #REFR .", "text_before_citation": ["CelebA sensitivities Tab. 2 presents results as averaged over all attributes (i.e.", "with a = hat and b being all other attributes, then a = glasses and b all others, etc.).", "The results clearly show that neuron granularity is crucial for success as fd, which operates layer-wise, falls short against dna-emd and dna-fd.", "Averaged over all attributes, our approach can discard 95.5% of the distance over the attributes on which we remove sensitivity, while only causing a 9.6% of deviation in distances over other attributes.", "dna-emd performs slightly better than dna-fd, but both do very well."], "text_after_citation": ["However, we obtain the best results using self-supervised models which are likely to produce more informative features.", "Finding similar images Qualitatively, we expect neurons to react to general and consistent features, which should also apply to comparing image pairs and different datasets.", "To verify this, we compare image pairs from a different dataset, FFHQ #OTHEREFR , with and without specific attributes (eyeglasses and wearing hat) and select the neuron(s) with the highest dna-emd sensitivity on CelebA. Using the selected neurons, we compare the DNA hist. of a selected reference image to DNAs hist. of 2000 random FFHQ samples. We present our results in Fig. 6 .", "We can verify that very few neurons are required to focus on high-level semantic attributes, even when selected on a different dataset.", "We still observe some errors, possibly due to neurons reacting to several attributes simultaneously."], "citing_paper_content": {"title": "Visual Dna: Representing And Comparing Images Using Distributions Of Neuron Activations", "abstract": "Selecting appropriate datasets is critical in modern computer vision. However, no general-purpose tools exist to evaluate the extent to which two datasets differ. For this, we propose representing images-and by extension datasetsusing Distributions of Neuron Activations (DNAs). DNAs fit distributions, such as histograms or Gaussians, to activations of neurons in a pre-trained feature extractor through which we pass the image(s) to represent. This extractor is frozen for all datasets, and we rely on its generally expressive power in feature space. By comparing two DNAs, we can evaluate the extent to which two datasets differ with granular control over the comparison attributes of interest, providing the ability to customise the way distances are measured to suit the requirements of the task at hand. Furthermore, DNAs are compact, representing datasets of any size with less than 15 megabytes. We demonstrate the value of DNAs by evaluating their applicability on several tasks, including conditional dataset comparison, synthetic image evaluation, and transfer learning, and across diverse datasets, ranging from synthetic cat images to celebrity faces and urban driving scenes. Real horses Fake horse Person 1 Person 2 (b) The DNA can also be used to compare individual images to datasets-for example, to measure the realism and semantic consistency of a synthetic image-or pairs of images-for example, to verify the presence of similar attributes such as smiling or wearing a hat."}, "cited_paper_content": {"title": "What'S Hidden In A Randomly Weighted Neural Network?", "abstract": "Training a neural network is synonymous with learning the values of the weights. By contrast, we demonstrate that randomly weighted neural networks contain subnetworks which achieve impressive performance without ever training the weight values. Hidden in a randomly weighted Wide ResNet-50 we show that there is a subnetwork (with random weights) that is smaller than, but matches the performance of a ResNet-34 trained on ImageNet. Not only do these\"untrained subnetworks\"exist, but we provide an algorithm to effectively find them. We empirically show that as randomly weighted neural networks with fixed weights grow wider and deeper, an\"untrained subnetwork\"approaches a network with learned weights in accuracy. Our code and pretrained models are available at https://github.com/allenai/hidden-networks."}, "keywords": ["feature extractors", "random weights"], "citation_intent": "background"}
{"citing_id": "2304.02022v1", "cited_id": "1706.03880", "section_title": "Ec.6. Additional Experiments On Parameter Estimation", "citation": "In our problem, the inventory is limited, the numbers of arriving customers (M t ) t\u2208[T ] are general, and the assortment set can dynamically change within each epoch whenever stochastic stockout events occur. To apply the estimator in #REFR to our setup, we consider three natural extensions.", "text_before_citation": ["In this section, we compare our attraction parameter estimatorv = (v 1, , . . .", ",v N, ) proposed in Sec- the same assortment until a no-purchase option occurs, and then record the number of purchases of each assorted product before the first no-purchase during this epoch.", "Then the average of each product's summary statistics recorded in all previous epochs where that product is assorted gives an estimator for the corresponding attraction parameter. #OTHEREFR", "(2019) shows that this estimator is consistent in their setup where there is no inventory limit and the assortment set is fixed in each epoch."], "text_after_citation": ["These extensions are all based on a similar epoch structure: in each epoch , we offer the same inventory level at the beginning of every cycle until a stopping criterion is met and then we record a summary statistic v i, for each assorted product i \u2208 [N ] .", "With slight abuse of notations, we still denote the subset of epochs up to in which product i is assorted by T i, and its cardinality by T i, .", "The estimators in the three extensions all have the same formv i, := 1/T i,", "\u2022 \u03c4 \u2208T i, v i,\u03c4 ,", "just with slightly different definitions of epoch stopping criteria and summary statistics v i,\u03c4 's."], "citing_paper_content": {"title": "Online Joint Assortment-Inventory Optimization Under Mnl Choices", "abstract": "We study an online joint assortment-inventory optimization problem, in which we assume that the choice behavior of each customer follows the Multinomial Logit (MNL) choice model, and the attraction parameters are unknown a priori. The retailer makes periodic assortment and inventory decisions to dynamically learn from the realized demands about the attraction parameters while maximizing the expected total profit over time. In this paper, we propose a novel algorithm that can effectively balance the exploration and exploitation in the online decision-making of assortment and inventory. Our algorithm builds on a new estimator for the MNL attraction parameters, a novel approach to incentivize exploration by adaptively tuning certain known and unknown parameters, and an optimization oracle to static single-cycle assortment-inventory planning problems with given parameters. We establish a regret upper bound for our algorithm and a lower bound for the online joint assortment-inventory optimization problem, suggesting that our algorithm achieves nearly optimal regret rate, provided that the static optimization oracle is exact. Then we incorporate more practical approximate static optimization oracles into our algorithm, and bound from above the impact of static optimization errors on the regret of our algorithm. At last, we perform numerical studies to demonstrate the effectiveness of our proposed algorithm."}, "cited_paper_content": {"title": "Mnl-Bandit: A Dynamic Learning Approach To Assortment Selection", "abstract": "We consider a dynamic assortment selection problem where in every round the retailer offers a subset (assortment) of N substitutable products to a consumer, who selects one of these products accord..."}, "keywords": ["assortment"], "citation_intent": "background"}
{"citing_id": "2304.12923v1", "cited_id": "1910.06151", "section_title": "Discussion", "citation": "Although we demonstrate the feasibility of using QBO to optimize hyperparameters of a machine learning model, the potential benefits of employing quantum kernels over classical machine learning methods in tasks using classical data remain uncertain #REFR .", "text_before_citation": ["In this study, we apply QGP models to one and multi-dimensional regression problems and show that they can be used as a surrogate model for BO to create a QBO.", "We demonstrate that QBO can be used to solve real-world hyperparameter optimization problems.", "Our encoding strategy allows for effectively using the variational parameters of the data embedding circuit as hyperparameters for the quantum kernels.", "In our simulations, we observe that the posterior variance of the QGP remains intact under the influence of samplingnoise and even for the calculation NISQ devices, although the influence of the various error sources in the latter affect the result.", "Nevertheless, since the results from the sampling-based simulations can be seen as an upper-bound for future hardware capabilities, the outlook is optimistic."], "text_after_citation": ["However, it is reasonable to expect that QBO may provide advantages in problems where quantum data can be leveraged to achieve a quantum advantage #OTHEREFR .", "Notably, QBO is potentially well-suited for active learning tasks in expensive molecular simulations, where the evaluation of the potential energy surface is based on quantum mechanics and is computationally expensive #OTHEREFR .", "The performance of the QGP model remains unexplored in several avenues within this work.", "For example, the choice of feature map is a crucial aspect and it has been shown that choosing problem-specific feature map with an inductive bias that is tailored to the dataset has various advantages such as improved performance and trainability #OTHEREFR .", "It is also known that using parametrized feature maps require special care when scaling the number of qubits which can lead to exponential concentration #OTHEREFR ."], "citing_paper_content": {"title": "Quantum Gaussian Process Regression For Bayesian Optimization", "abstract": "Gaussian process regression is a well-established Bayesian machine learning method. We propose a new approach to Gaussian process regression using quantum kernels based on parameterized quantum circuits. By employing a hardware-efficient feature map and careful regularization of the Gram matrix, we demonstrate that the variance information of the resulting quantum Gaussian process can be preserved. We also show that quantum Gaussian processes can be used as a surrogate model for Bayesian optimization, a task that critically relies on the variance of the surrogate model. To demonstrate the performance of this quantum Bayesian optimization algorithm, we apply it to the hyperparameter optimization of a machine learning model which performs regression on a real-world dataset. We benchmark the quantum Bayesian optimization against its classical counterpart and show that quantum version can match its performance."}, "cited_paper_content": {"title": "Sampling-Based Sublinear Low-Rank Matrix Arithmetic Framework For Dequantizing Quantum Machine Learning", "abstract": "We present an algorithmic framework generalizing quantum-inspired polylogarithmic-time algorithms on low-rank matrices. Our work follows the line of research started by Tang's breakthrough classical algorithm for recommendation systems [STOC'19]. The main result of this work is an algorithm for singular value transformation on low-rank inputs in the quantum-inspired regime, where singular value transformation is a framework proposed by Gilyen et al. [STOC'19] to study various quantum speedups. Since singular value transformation encompasses a vast range of matrix arithmetic, this result, combined with simple sampling lemmas from previous work, suffices to generalize all results dequantizing quantum machine learning algorithms to the authors' knowledge. Via simple black-box applications of our singular value transformation framework, we recover the dequantization results on recommendation systems, principal component analysis, supervised clustering, low-rank matrix inversion, low-rank semidefinite programming, and support vector machines. We also give additional dequantizations results on low-rank Hamiltonian simulation and discriminant analysis."}, "keywords": ["quantum kernels"], "citation_intent": "method"}
{"citing_id": "2303.09172v1", "cited_id": "1608.01946", "section_title": "Pomcp Integration Results.", "citation": "In Figure 2a , we show the discounted return achieved by POMCP (with and without rule bias) when the number of particles is varied between 2 8 and 2 #REFR .", "text_before_citation": [], "text_after_citation": ["Introducing learned ASP axioms does not significantly affect the performance of the solver.", "This is probably due to the limited number of actions involved in the task (only 3, i.e., advance, check and recharge), thus few branches must be explored at each simulation step.", "Figure 2b shows the discounted return achieved by POMCP when the length of the path increases, while the number of particles is constantly 2 #OTHEREFR as in the training setup.", "In this case, the solver performs significantly better when learned rules are employed (the discounted return increases up to 20% when the path length is 75, i.e., almost double than the training scenario).", "In fact, as the path length increases, the risk of battery discharge is higher, since failing to recharge is more probable."], "citing_paper_content": {"title": "Learning Logic Specifications For Soft Policy Guidance In Pomcp", "abstract": "Partially Observable Monte Carlo Planning (POMCP) is an efficient solver for Partially Observable Markov Decision Processes (POMDPs). It allows scaling to large state spaces by computing an approximation of the optimal policy locally and online, using a Monte Carlo Tree Search based strategy. However, POMCP suffers from sparse reward function, namely, rewards achieved only when the final goal is reached, particularly in environments with large state spaces and long horizons. Recently, logic specifications have been integrated into POMCP to guide exploration and to satisfy safety requirements. However, such policy-related rules require manual definition by domain experts, especially in real-world scenarios. In this paper, we use inductive logic programming to learn logic specifications from traces of POMCP executions, i.e., sets of belief-action pairs generated by the planner. Specifically, we learn rules expressed in the paradigm of answer set programming. We then integrate them inside POMCP to provide soft policy bias toward promising actions. In the context of two benchmark scenarios, rocksample and battery, we show that the integration of learned rules from small task instances can improve performance with fewer Monte Carlo simulations and in larger task instances."}, "cited_paper_content": {"title": "Iterative Learning Of Answer Set Programs From Context Dependent Examples", "abstract": "In recent years, several frameworks and systems have been proposed that extend Inductive Logic Programming (ILP) to the Answer Set Programming (ASP) paradigm. In ILP, examples must all be explained by a hypothesis together with a given background knowledge. In existing systems, the background knowledge is the same for all examples; however, examples may be context-dependent. This means that some examples should be explained in the context of some information, whereas others should be explained in different contexts. In this paper, we capture this notion and present a context-dependent extension of the Learning from Ordered Answer Sets framework. In this extension, contexts can be used to further structure the background knowledge. We then propose a new iterative algorithm, ILASP2i, which exploits this feature to scale up the existing ILASP2 system to learning tasks with large numbers of examples. We demonstrate the gain in scalability by applying both algorithms to various learning tasks. Our results show that, compared to ILASP2, the newly proposed ILASP2i system can be two orders of magnitude faster and use two orders of magnitude less memory, whilst preserving the same average accuracy. This paper is under consideration for acceptance in TPLP."}, "keywords": ["POMCP", "rule bias"], "citation_intent": "method"}
{"citing_id": "2303.17598v1", "cited_id": "1906.08240", "section_title": "Consistent Long-Term View Synthesis", "citation": "To reduce the variance of the backward process across different views, we use the same initialization noise x T and diffusion noises {\u03f5 t } #REFR t=T to generate all images in the same video.", "text_before_citation": ["We use the following two solutions to address the issue. Stochastic conditioning.", "We find that using stochastic conditioning #OTHEREFR slightly improves the temporal flickering.", "Specifically, at each step in the backward process described in (3), instead of using the previous frame x i\u22121 , we randomly sample the source view image x j from the set of prior frames x j \u223c Uniform({x k , \u2022 \u2022 \u2022 , x i\u22121 }).", "Such a strategy encourages the diffusion model to be guided by all the previous frames, thus improving the temporal consistency. Fixing noises in the backward process.", "The noises introduced during the backward process illustrated in (3) also contribute to the temporal inconsistency."], "text_after_citation": ["Neverthe-less, we observe noticeable artifacts if we fix all diffusion noises {\u03f5 t } 1 t=T during the backward process, as demonstrated in Figure 3 .", "In practice, fixing the diffusion noises {\u03f5 t } t \u2032 t=T to a certain backward step t \u2032 alleviates the issue and improves the temporal consistency. #OTHEREFR"], "citing_paper_content": {"title": "Consistent View Synthesis With Pose-Guided Diffusion Models", "abstract": "https://poseguided-diffusion.github.io/ Outputs Input Figure 1. Consistent view synthesis via pose-guided diffusion model. (top) Given an input image and a sequence of camera poses, we present a pose-guided diffusion model to generate a sequence of frames that are photorealistic and 3D consistent. (bottom) Our proposed method can synthesize diverse sequences from the same set of inputs."}, "cited_paper_content": {"title": "Neural Point-Based Graphics", "abstract": "We present a new point-based approach for modeling the appearance of real scenes. The approach uses a raw point cloud as the geometric representation of a scene, and augments each point with a learnable neural descriptor that encodes local geometry and appearance. A deep rendering network is learned in parallel with the descriptors, so that new views of the scene can be obtained by passing the rasterizations of a point cloud from new viewpoints through this network. The input rasterizations use the learned descriptors as point pseudo-colors. We show that the proposed approach can be used for modeling complex scenes and obtaining their photorealistic views, while avoiding explicit surface estimation and meshing. In particular, compelling results are obtained for scene scanned using hand-held commodity RGB-D sensors as well as standard RGB cameras even in the presence of objects that are challenging for standard mesh-based modeling."}, "keywords": ["diffusion noises", "images"], "citation_intent": "method"}
{"citing_id": "2304.11547v1", "cited_id": "1811.00002", "section_title": "B. Model Configuration", "citation": "We use similar parameters setup to those in #REFR . Each audio clip contains 16,000 samples.", "text_before_citation": ["The first FC layer contains 128 hidden units, whereas the output FC layer consists of 80 units.", "As for training, the batch size used is 64, and Adam Optimizer is applied with 1e-4 learning rate.", "For checkpoint selection, an early stop is applied regarding the validation loss.", "The \u03b1 is sampled from a uniform distribution of interval [0, 0.2].", "For downstream tasks in our experiments, Waveglow is selected as a neural vocoder."], "text_after_citation": ["The batch size used is 8 with Adam Optimizer of learning rate 1e-4.", "For ground-truth acoustic feature extraction, Fast Fourier Transform (FFT) window size is selected to be 1024 with a hop size equal to 16 milliseconds (256 for 16000 sampling rate, and 128 for 8000 sampling rate).", "As for the acoustic model, Tacotron2 #OTHEREFR is utilized to predict acoustic features.", "The Prenet dimension is set to 256, and the number of embedding dimensions is selected to be 512 with a dimension of attention recurrent neural network (RNN) equal to 1024.", "The dimension of decoder RNN is set to 1024 with the dimension of Postnet embedding equal to 512."], "citing_paper_content": {"title": "Sar: Self-Supervised Anti-Distortion Representation For End-To-End Speech Model", "abstract": "In recent Text-to-Speech (TTS) systems, a neural vocoder often generates speech samples by solely conditioning on acoustic features predicted from an acoustic model. However, there are always distortions existing in the predicted acoustic features, compared to those of the groundtruth, especially in the common case of poor acoustic modeling due to lowquality training data. To overcome such limits, we propose a Self-supervised learning framework to learn an Anti-distortion acoustic Representation (SAR) to replace human-crafted acoustic features by introducing distortion prior to an auto-encoder pretraining process. The learned acoustic representation from the proposed framework is proved anti-distortion compared to the most commonly used mel-spectrogram through both objective and subjective evaluation."}, "cited_paper_content": {"title": "Waveglow: A Flow-Based Generative Network For Speech Synthesis", "abstract": "In this paper we propose WaveGlow: a flow-based network capable of generating high quality speech from mel-spectrograms. WaveGlow combines insights from Glow [1] and WaveNet [2] in order to provide fast, efficient and high-quality audio synthesis, without the need for auto-regression. WaveGlow is implemented using only a single network, trained using only a single cost function: maximizing the likelihood of the training data, which makes the training procedure simple and stable. Our PyTorch implementation produces audio samples at a rate of more than 500 kHz on an NVIDIA V100 GPU. Mean Opinion Scores show that it delivers audio quality as good as the best publicly available WaveNet implementation. All code will be made publicly available online [3]."}, "keywords": ["16,000 samples"], "citation_intent": "method"}
{"citing_id": "2303.12341v1", "cited_id": "1706.03762", "section_title": "Endogenous Dynamics Encoding", "citation": "Note that e (t) u,i essentially models the associations between clusters since it implicitly computes the interaction between cluster embeddings, analogous to the positional encoding #REFR .", "text_before_citation": ["Let v u denote the query node and set N u denotes its neighborhood, then we encode such dynamics into representation s", "EQUATION", "where", "v (t) i;l is the i th column of V (t) l , e (t)", "u,i is the attention coefficient which can be implemented in various ways (e.g., SA #OTHEREFR , GAT #OTHEREFR and GATv2 #OTHEREFR )."], "text_after_citation": [], "citing_paper_content": {"title": "Easydgl: Encode, Train And Interpret For Continuous-Time Dynamic Graph Learning", "abstract": "Dynamic graphs arise in various real-world applications, and it is often welcomed to model the dynamics directly in continuous time domain for its flexibility. This paper aims to design an easy-to-use pipeline (termed as EasyDGL which is also due to its implementation by DGL toolkit) composed of three key modules with both strong fitting ability and interpretability. Specifically the proposed pipeline which involves encoding, training and interpreting: i) a temporal point process (TPP) modulated attention architecture to endow the continuous-time resolution with the coupled spatiotemporal dynamics of the observed graph with edge-addition events; ii) a principled loss composed of task-agnostic TPP posterior maximization based on observed events on the graph, and a task-aware loss with a masking strategy over dynamic graph, where the covered tasks include dynamic link prediction, dynamic node classification and node traffic forecasting; iii) interpretation of the model outputs (e.g., representations and predictions) with scalable perturbation-based quantitative analysis in the graph Fourier domain, which could more comprehensively reflect the behavior of the learned model. Extensive experimental results on public benchmarks show the superior performance of our EasyDGL for time-conditioned predictive tasks, and in particular demonstrate that EasyDGL can effectively quantify the predictive power of frequency content that a model learn from the evolving graph data."}, "cited_paper_content": {"title": "Attention Is All You Need", "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."}, "keywords": ["cluster embeddings", "positional encoding"], "citation_intent": "background"}
{"citing_id": "2303.16435v1", "cited_id": "1606.00915", "section_title": "Datasets", "citation": "The initial learning rate is set as 2.5 \u00d7 10 \u22124 and is decreased using the polynomial decay with power of 0.9 as mentioned in #REFR , and the maximum iteration number is Table 2 : Performance gap between the adapted model and the fully-supervised(oracle)model.", "text_before_citation": ["To train the proposed adaptation model, in each train batch, we first forward the source image X s to optimize the segmentation network for L s seg in (9) and generate the output P s .", "For the target image X t , we obtain the segmentation output P t and pass it along with P s for optimizing L OT in (7).", "For the multi-level training objective in (10), we simply repeat the same procedure for each adaptation module.", "Our method is implemented with the PyTorch library on Nvidia GPU GTX 2080Ti with 12GB memory.", "To train the segmentation network, we use the Stochastic Gradient Descent(SGD) optimizer with Nesterov acceleration where the momentum is 0.9 and the weight decay is 5 \u00d7 10 \u22124 ."], "text_after_citation": ["We first compare results with state-of-the-art methods using the VGG based model, and then show our result using the ResNet one. Our baseline model is AdaptSegNet #OTHEREFR .", "Other previous state-of-the-art methods include CDA #OTHEREFR , CyCADA #OTHEREFR ."], "citing_paper_content": {"title": "Domain Adaptive Semantic Segmentation By Optimal Transport", "abstract": "Scene segmentation is widely used in the field of autonomous driving for environment perception, and semantic scene segmentation (3S) has received a great deal of attention due to the richness of the semantic information it contains. It aims to assign labels to pixels in an image, thus enabling automatic image labeling. Current approaches are mainly based on convolutional neural networks (CNN), but they rely on a large number of labels. Therefore, how to use a small size of labeled data to achieve semantic segmentation becomes more and more important. In this paper, we propose a domain adaptation (DA) framework based on optimal transport (OT) and attention mechanism to address this issue. Concretely, first we generate the output space via CNN due to its superiority of feature representation. Second, we utilize OT to achieve a more robust alignment of source and target domains in output space, where the OT plan defines a well attention mechanism to improve the adaptation of the model. In particular, with OT, the number of network parameters has been reduced and the network has been better interpretable. Third, to better describe the multi-scale property of features, we construct a multi-scale segmentation network to perform domain adaptation. Finally, in order to verify the performance of our proposed method, we conduct experimental comparison with three benchmark and four SOTA methods on three scene datasets, and the mean intersection-over-union (mIOU) has been significant improved, and visualization results under multiple domain adaptation scenarios also show that our proposed method has better performance than compared semantic segmentation methods."}, "cited_paper_content": {"title": "Deeplab: Semantic Image Segmentation With Deep Convolutional Nets, Atrous Convolution, And Fully Connected Crfs", "abstract": "In this work we address the task of semantic image segmentation with Deep Learning and make three main contributions that are experimentally shown to have substantial practical merit. First , we highlight convolution with upsampled filters, or \u2018atrous convolution\u2019, as a powerful tool in dense prediction tasks. Atrous convolution allows us to explicitly control the resolution at which feature responses are computed within Deep Convolutional Neural Networks. It also allows us to effectively enlarge the field of view of filters to incorporate larger context without increasing the number of parameters or the amount of computation. Second , we propose atrous spatial pyramid pooling (ASPP) to robustly segment objects at multiple scales. ASPP probes an incoming convolutional feature layer with filters at multiple sampling rates and effective fields-of-views, thus capturing objects as well as image context at multiple scales. Third , we improve the localization of object boundaries by combining methods from DCNNs and probabilistic graphical models. The commonly deployed combination of max-pooling and downsampling in DCNNs achieves invariance but has a toll on localization accuracy. We overcome this by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF), which is shown both qualitatively and quantitatively to improve localization performance. Our proposed \u201cDeepLab\u201d system sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 79.7 percent mIOU in the test set, and advances the results on three other datasets: PASCAL-Context, PASCAL-Person-Part, and Cityscapes. All of our code is made publicly available online."}, "keywords": ["fully-supervised(oracle)model"], "citation_intent": "method"}
{"citing_id": "2303.00633v1", "cited_id": "1611.01353", "section_title": "Background & Preliminaries", "citation": "Finally, #REFR found that minimizing a stochastic network with a regularizer is equivalent to minimizing the cross-entropy over deterministic DNNs with multiplicative noise.", "text_before_citation": ["As a result, the mutual information between the input and the DNN representation in such objectives would be infinite, resulting in ill-posed optimization problems.", "To avoid this problem, stochastic DNNs with variational bounds could be used, where the output of the deterministic network is used as the parameters of the conditional distribution #OTHEREFR ; Shwartz-Ziv and Alemi, 2020). Dubois et al.", "(2021) assumed that the randomness of data augmentation among the two views is the source of stochasticity in the network.", "Other work assumed a random input, but without making any assumptions about the properties of the distribution of the network's output, to analyze the objective and relied on general lower bounds (Wang and Isola, 2020; Zimmermann et al., 2021) . For supervised learning, #OTHEREFR", "(2018) introduced an auxiliary (noisy) DNN by injecting additive noise into the model and demonstrated that the resulting model is a good proxy for the original (deterministic) DNN in terms of both performance and representation."], "text_after_citation": ["All of these methods assume that the source of randomness comes from the DNN, contradicting common practice."], "citing_paper_content": {"title": "An Information-Theoretic Perspective On Variance-Invariance-Covariance Regularization", "abstract": "In this paper, we provide an information-theoretic perspective on Variance-Invariance-Covariance Regularization (VICReg) for self-supervised learning. To do so, we first demonstrate how information-theoretic quantities can be obtained for deterministic networks as an alternative to the commonly used unrealistic stochastic networks assumption. Next, we relate the VICReg objective to mutual information maximization and use it to highlight the underlying assumptions of the objective. Based on this relationship, we derive a generalization bound for VICReg, providing generalization guarantees for downstream supervised learning tasks and present new self-supervised learning methods, derived from a mutual information maximization objective, that outperform existing methods in terms of performance. This work provides a new information-theoretic perspective on self-supervised learning and Variance-Invariance-Covariance Regularization in particular and guides the way for improved transfer learning via information-theoretic self-supervised learning objectives."}, "cited_paper_content": {"title": "Information Dropout: Learning Optimal Representations Through Noisy Computation", "abstract": "The cross-entropy loss commonly used in deep learning is closely related to the defining properties of optimal representations, but does not enforce some of the key properties. We show that this can be solved by adding a regularization term, which is in turn related to injecting multiplicative noise in the activations of a Deep Neural Network, a special case of which is the common practice of dropout. We show that our regularized loss function can be efficiently minimized using Information Dropout, a generalization of dropout rooted in information theoretic principles that automatically adapts to the data and can better exploit architectures of limited capacity. When the task is the reconstruction of the input, we show that our loss function yields a Variational Autoencoder as a special case, thus providing a link between representation learning, information theory and variational inference. Finally, we prove that we can promote the creation of optimal disentangled representations simply by enforcing a factorized prior, a fact that has been observed empirically in recent work. Our experiments validate the theoretical intuitions behind our method, and we find that Information Dropout achieves a comparable or better generalization performance than binary dropout, especially on smaller models, since it can automatically adapt the noise to the structure of the network, as well as to the test sample."}, "keywords": ["deterministic DNNs"], "citation_intent": "background"}
{"citing_id": "2303.15796v1", "cited_id": "1602.07332", "section_title": "Retrieved Facts", "citation": "To obtain knowledge for view images, facts (i.e., knowledge described by language descriptions) are retrieved from the knowledge base constructed on the Visual Genome dataset #REFR .", "text_before_citation": ["Second, the knowledge improves the generalization ability of the agent.", "As the alignment between the instruction and the navigable candidate is learned in limited-seen environments, leveraging knowledge benefits the alignment in the unseen environment, as there is no specific regularity for target object arrangement. Third, knowledge increases the capability of VLN models.", "As rich conceptual information is injected into VLN models, the correlations among numerous concepts are learned.", "The learned correlations are able to benefit visual and language alignment, especially for tasks with high-level instructions.", "In this work, we incorporate knowledge into the VLN task."], "text_after_citation": ["The retrieved facts by CLIP #OTHEREFR provide rich and complementary information for visual view images.", "And then, a knowledge enhanced reasoning model (KERM) which leverages knowledge for sufficient interaction and better alignment between vision and language information is proposed.", "Especially, the proposed KERM consists of a purification module, a fact-aware interaction module, and an instruction-guided aggregation module.", "The purification model aims to extract key information in the fact representations, the visual region representations, and the historical representations respectively guided by the instruction.", "The fact-aware interaction module allows visual and historical representations to obtain the interaction of the facts with cross-attention encoders."], "citing_paper_content": {"title": "Kerm: Knowledge Enhanced Reasoning For Vision-And-Language Navigation", "abstract": "Vision-and-language navigation (VLN) is the task to enable an embodied agent to navigate to a remote location following the natural language instruction in real scenes. Most of the previous approaches utilize the entire features or object-centric features to represent navigable candidates. However, these representations are not efficient enough for an agent to perform actions to arrive the target location. As knowledge provides crucial information which is complementary to visible content, in this paper, we propose a Knowledge Enhanced Reasoning Model (KERM) to leverage knowledge to improve agent navigation ability. Specifically, we first retrieve facts (i.e., knowledge described by language descriptions) for the navigation views based on local regions from the constructed knowledge base. The retrieved facts range from properties of a single object (e.g., color, shape) to relationships between objects (e.g., action, spatial position), providing crucial information for VLN. We further present the KERM which contains the purification, fact-aware interaction, and instruction-guided aggregation modules to integrate visual, history, instruction, and fact features. The proposed KERM can automatically select and gather crucial and relevant cues, obtaining more accurate action prediction. Experimental results on the REVERIE, R2R, and SOON datasets demonstrate the effectiveness of the proposed method. The source code is available at https://github.com/XiangyangLi20/KERM ."}, "cited_paper_content": {"title": "Visual Genome: Connecting Language And Vision Using Crowdsourced Dense Image Annotations", "abstract": "Despite progress in perceptual tasks such as image classification, computers still perform poorly on cognitive tasks such as image description and question answering. Cognition is core to tasks that involve not just recognizing, but reasoning about our visual world. However, models used to tackle the rich content in images for cognitive tasks are still being trained using the same datasets designed for perceptual tasks. To achieve success at cognitive tasks, models need to understand the interactions and relationships between objects in an image. When asked \"What vehicle is the person riding?\", computers will need to identify the objects in an image as well as the relationships riding(man, carriage) and pulling(horse, carriage) to answer correctly that \"the person is riding a horse-drawn carriage.\" In this paper, we present the Visual Genome dataset to enable the modeling of such relationships. We collect dense annotations of objects, attributes, and relationships within each image to learn these models. Specifically, our dataset contains over 108K images where each image has an average of $$35$$35 objects, $$26$$26 attributes, and $$21$$21 pairwise relationships between objects. We canonicalize the objects, attributes, relationships, and noun phrases in region descriptions and questions answer pairs to WordNet synsets. Together, these annotations represent the densest and largest dataset of image descriptions, objects, attributes, relationships, and question answer pairs."}, "keywords": ["knowledge base", "language descriptions"], "citation_intent": "method"}
{"citing_id": "2305.02244v1", "cited_id": "1903.05714", "section_title": "I. Introduction", "citation": "It provides high persistence guarantees, higher read and write bandwidth and lower latencies than most persistent block devices #REFR .", "text_before_citation": ["The emergence of modern NVMM is a great opportunity to implement known designs and adapt them, or invent new ones.", "We tried these two approaches with caching mechanisms for a file system stored in secondary storage.", "Indeed, caching data for slower tier storage devices (SSD or HDD) is a great use case for NVMM."], "text_after_citation": ["In this study, we target applications that require a high level of data consistency, which would highly solicit a regular disk with frequent calls to fsync.", "For such applications, we propose a persistent cache able to give fast persistence guarantees without having to wait for a slow secondary storage."], "citing_paper_content": {"title": "Nvmm Cache Design: Logging Vs. Paging", "abstract": "Modern NVMM is closing the gap between DRAM and persistent storage, both in terms of performance and features. Having both byte addressability and persistence on the same device gives NVMM an unprecedented set of features, leading to the following question: How should we design an NVMMbased caching system to fully exploit its potential? We build two caching mechanisms, NVPages and NVLog, based on two radically different design approaches. NVPages stores memory pages in NVMM, similar to the Linux page cache (LPC). NVLog uses NVMM to store a log of pending write operations to be submitted to the LPC, while it ensures reads with a small DRAM cache. Our study shows and quantifies advantages and flaws for both designs."}, "cited_paper_content": {"title": "Basic Performance Measurements Of The Intel Optane Dc Persistent Memory Module", "abstract": "Scalable nonvolatile memory DIMMs will finally be commercially available with the release of the Intel Optane DC Persistent Memory Module (or just \"Optane DC PMM\"). This new nonvolatile DIMM supports byte-granularity accesses with access times on the order of DRAM, while also providing data storage that survives power outages. This work comprises the first in-depth, scholarly, performance review of Intel's Optane DC PMM, exploring its capabilities as a main memory device, and as persistent, byte-addressable memory exposed to user-space applications. This report details the technologies performance under a number of modes and scenarios, and across a wide variety of macro-scale benchmarks. Optane DC PMMs can be used as large memory devices with a DRAM cache to hide their lower bandwidth and higher latency. When used in this Memory (or cached) mode, Optane DC memory has little impact on applications with small memory footprints. Applications with larger memory footprints may experience some slow-down relative to DRAM, but are now able to keep much more data in memory. When used under a file system, Optane DC PMMs can result in significant performance gains, especially when the file system is optimized to use the load/store interface of the Optane DC PMM and the application uses many small, persistent writes. For instance, using the NOVA-relaxed NVMM file system, we can improve the performance of Kyoto Cabinet by almost 2x. Optane DC PMMs can also enable user-space persistence where the application explicitly controls its writes into persistent Optane DC media. In our experiments, modified applications that used user-space Optane DC persistence generally outperformed their file system counterparts. For instance, the persistent version of RocksDB performed almost 2x faster than the equivalent program utilizing an NVMM-aware file system."}, "keywords": ["high persistence guarantees"], "citation_intent": "background"}