paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Overcoming barriers to the training of effective learned optimizers | 1 INTRODUCTION . Much of the success of modern deep learning has been driven by a shift from hand-designed features carefully curated by human experts , to domain-agnostic methods that can learn features from large amounts of data . By leveraging large-scale datasets with flexible models , we are now able to rapidly le... | This paper attempts to address the fundamental barriers of learned optimization. The authors identify three barriers: computational requirements, number of training tasks and lack of inductive bias. A “large-scale” evaluation and comparison of learned optimizers is then carried out using many (1024) multi-core CPUs. A ... | SP:cc0f8de167292fb0c4030746628536907cb3913a |
Overcoming barriers to the training of effective learned optimizers | 1 INTRODUCTION . Much of the success of modern deep learning has been driven by a shift from hand-designed features carefully curated by human experts , to domain-agnostic methods that can learn features from large amounts of data . By leveraging large-scale datasets with flexible models , we are now able to rapidly le... | The authors propose to use a combination of Andrychowicz et al. LSTM based approach and Metz et.al feed forward network to learn an optimizer that is useful accross any task. The authors propose to use a few thousand tasks as developed in Metz et. al. for the purpose of training the optimizer. They focus on evolutionar... | SP:cc0f8de167292fb0c4030746628536907cb3913a |
A Provably Convergent and Practical Algorithm for Min-Max Optimization with Applications to GANs | 1 INTRODUCTION . We consider the problem of min-max optimization minx∈Rd maxy∈Rd f ( x , y ) , where the loss function f may be nonconvex in x and nonconcave in y. Min-max optimization of such loss functions has many applications to machine learning , including to GANs ( Goodfellow et al. , 2014 ) and adversarial train... | The paper introduces a first-order algorithm for nonconvex-nonconcave min-max optimization problems. The proposed algorithm terminates in time polynomial in the dimension and smoothness parameters of the loss function. The points (x*,y*) returned by the algorithm satisfy the following guarantee: if the min-player propo... | SP:4985bfe9b838b639e0841b767cccf5eead3f8e47 |
A Provably Convergent and Practical Algorithm for Min-Max Optimization with Applications to GANs | 1 INTRODUCTION . We consider the problem of min-max optimization minx∈Rd maxy∈Rd f ( x , y ) , where the loss function f may be nonconvex in x and nonconcave in y. Min-max optimization of such loss functions has many applications to machine learning , including to GANs ( Goodfellow et al. , 2014 ) and adversarial train... | This paper proposes a new stochastic gradient descent-ascent-based method to approximate a stationary point (or local min-max solution) of a nonconvex-nonconcave minimax problem with application in GANs. The method is similar to the one in the GAN original paper, but the authors incorporate it with an acceptance rule a... | SP:4985bfe9b838b639e0841b767cccf5eead3f8e47 |
A Provably Convergent and Practical Algorithm for Min-Max Optimization with Applications to GANs | 1 INTRODUCTION . We consider the problem of min-max optimization minx∈Rd maxy∈Rd f ( x , y ) , where the loss function f may be nonconvex in x and nonconcave in y. Min-max optimization of such loss functions has many applications to machine learning , including to GANs ( Goodfellow et al. , 2014 ) and adversarial train... | This paper treats non-convex/non-concave min/max problems motivated by the respective problems that arise in GAN training. The main contribution is that they develop an ADAM-based algorithm that converges to \eps- local min/max points. The paper seems to be well-written and easy to follow. Moreover the proofs seem corr... | SP:4985bfe9b838b639e0841b767cccf5eead3f8e47 |
Machine Reading Comprehension with Enhanced Linguistic Verifiers | 1 INTRODUCTION . Teaching a machine to read and comprehend large-scale textual documents is a promising and longstanding goal of natural language understanding . This field , so called machine reading comprehension ( MRC ) ( Zhang et al. , 2019 ; 2020c ) , has achieved impressive milestones in recent years thanks to th... | In this paper, two linguistic verifiers are proposed to improve the model performance on machine reading comprehension datasets, such as SQuAD v2, NewsQA and TriviaQA. The first verifier rewrites the question by replacing its interrogatives with the predicted answer phrases. Then it computes a score between the rewritt... | SP:9b909cc1c71ca942c7b094c735181010d93a7d86 |
Machine Reading Comprehension with Enhanced Linguistic Verifiers | 1 INTRODUCTION . Teaching a machine to read and comprehend large-scale textual documents is a promising and longstanding goal of natural language understanding . This field , so called machine reading comprehension ( MRC ) ( Zhang et al. , 2019 ; 2020c ) , has achieved impressive milestones in recent years thanks to th... | This work addresses two main challenges of span-extraction style machine reading comprehension (MRC) tasks: how to evaluate the syntactic completeness of predicted answers and how to utilize the rich context of long documents. To handle such challenges, Question Rewritten Verifier (QRV) and Hierarchical Attention Netwo... | SP:9b909cc1c71ca942c7b094c735181010d93a7d86 |
Machine Reading Comprehension with Enhanced Linguistic Verifiers | 1 INTRODUCTION . Teaching a machine to read and comprehend large-scale textual documents is a promising and longstanding goal of natural language understanding . This field , so called machine reading comprehension ( MRC ) ( Zhang et al. , 2019 ; 2020c ) , has achieved impressive milestones in recent years thanks to th... | This paper proposes two types of linguistic verifiers for machine reading comprehension task in span extraction form. One is a rewritten question oriented verifier that checks the linguistic correctness of the extracted answers, and the other is based on a hierarchical attention network for answerability classification... | SP:9b909cc1c71ca942c7b094c735181010d93a7d86 |
Exploring the Potential of Low-Bit Training of Convolutional Neural Networks | 1 INTRODUCTION . Convolutional neural networks ( CNNs ) have achieved state-of-the-art performance in many computer vision tasks , such as image classification ( Krizhevsky et al. , 2012 ) and object detection ( Redmon et al. , 2016 ; Liu et al. , 2016 ) . However , deep CNNs are both computation and storage-intensive ... | This paper investigated the low-bit training problem and proposed a novel method, which can reduce the element-wise bit-width to simplify floating-point computations to nearly fixed-point. The major contribution can be summarized as follows: a multi-level scaling (MLS) tensor format, a dynamic quantization and the low-... | SP:35c957dccdefe660923cac3ed68afbfdea37faf7 |
Exploring the Potential of Low-Bit Training of Convolutional Neural Networks | 1 INTRODUCTION . Convolutional neural networks ( CNNs ) have achieved state-of-the-art performance in many computer vision tasks , such as image classification ( Krizhevsky et al. , 2012 ) and object detection ( Redmon et al. , 2016 ; Liu et al. , 2016 ) . However , deep CNNs are both computation and storage-intensive ... | Efficient training is becoming a crucial research topic as deep learning models get deeper and more complex to improve model accuracy. The authors propose a low-bit floating point quantization method to reduce energy and time consumption during training. To enhance training efficiency, the authors suggest multi-level s... | SP:35c957dccdefe660923cac3ed68afbfdea37faf7 |
Exploring the Potential of Low-Bit Training of Convolutional Neural Networks | 1 INTRODUCTION . Convolutional neural networks ( CNNs ) have achieved state-of-the-art performance in many computer vision tasks , such as image classification ( Krizhevsky et al. , 2012 ) and object detection ( Redmon et al. , 2016 ; Liu et al. , 2016 ) . However , deep CNNs are both computation and storage-intensive ... | This manuscript describes a new low-bit training framework as well as a new low-bit format and shows promising accuracy-precision trade-offs and better energy efficiency. The proposed method achieves no accuracy drop with 3-bit training on CIFAR dataset and 1% accuracy drop with 6-bit training on ImageNet dataset. Alth... | SP:35c957dccdefe660923cac3ed68afbfdea37faf7 |
DARTS-: Robustly Stepping out of Performance Collapse Without Indicators | 1 INTRODUCTION . Recent studies ( Zela et al. , 2020 ; Liang et al. , 2019 ; Chu et al. , 2020b ) have shown that one critical issue for differentiable architecture search ( Liu et al. , 2019b ) regarding the performance collapse due to superfluous skip connections . Accordingly , some empirical indicators for detectin... | This paper presents an interesting method to alleviate the mode collapse of DARTS (all operations degenerate to skip-connect). This is done by simply adding a skip-connect operation to complement the output of the cell function and making the coefficient of the auxiliary operation decay with time. The method is tested ... | SP:6ae92d42bc6c812252e46f6c1485b25ef6e2ef1e |
DARTS-: Robustly Stepping out of Performance Collapse Without Indicators | 1 INTRODUCTION . Recent studies ( Zela et al. , 2020 ; Liang et al. , 2019 ; Chu et al. , 2020b ) have shown that one critical issue for differentiable architecture search ( Liu et al. , 2019b ) regarding the performance collapse due to superfluous skip connections . Accordingly , some empirical indicators for detectin... | This paper aims to improve the robustness of DARTS, and proposes to add an auxiliary skip connection branch to the “mixOp” in cells. The authors also analyze the effect of auxiliary branch on residual block from the view of gradient flow. Additionally, this paper refers to the theory of the work [1] and demonstrate tha... | SP:6ae92d42bc6c812252e46f6c1485b25ef6e2ef1e |
DARTS-: Robustly Stepping out of Performance Collapse Without Indicators | 1 INTRODUCTION . Recent studies ( Zela et al. , 2020 ; Liang et al. , 2019 ; Chu et al. , 2020b ) have shown that one critical issue for differentiable architecture search ( Liu et al. , 2019b ) regarding the performance collapse due to superfluous skip connections . Accordingly , some empirical indicators for detectin... | The paper reveals two roles of skip connection to prevent model collapse: stabilize the supernet training, and as a candidate operation to build the final network. Intuitively, the skip connection playing the first role should only be there during the training phase. Therefore, the authors propose to add an auxiliary s... | SP:6ae92d42bc6c812252e46f6c1485b25ef6e2ef1e |
Distributed Training of Graph Convolutional Networks using Subgraph Approximation | 1 INTRODUCTION . Graphs are used to model data in a diverse range of applications such as social networks ( Hamilton et al. , 2017a ) , biological networks ( Fout et al. , 2017 ) and e-commerce interactions ( Yang , 2019 ) . Recently , there has been an increased interest in applying machine learning techniques to grap... | Graph Convolutional Networks (GCNs) have inspired state-of-the-art methods for learning representations on graphs. However, training GCNs for very large graphs remains an issue because of their memory, computation demands. Towards addressing this issue, the paper proposes a distributed GCN training scheme based on subg... | SP:b7fdf5984d0bd1e4caff646d1a403b18c854e7bd |
Distributed Training of Graph Convolutional Networks using Subgraph Approximation | 1 INTRODUCTION . Graphs are used to model data in a diverse range of applications such as social networks ( Hamilton et al. , 2017a ) , biological networks ( Fout et al. , 2017 ) and e-commerce interactions ( Yang , 2019 ) . Recently , there has been an increased interest in applying machine learning techniques to grap... | This paper studies how to do distributed training for GNNs. For GNNs, nodes are connected so that it is not trivial to do distributed training, because it needs message passing across machines, incurring communication costs. If no message passing across machines, the performance will be bad. To address this problem, th... | SP:b7fdf5984d0bd1e4caff646d1a403b18c854e7bd |
Distributed Training of Graph Convolutional Networks using Subgraph Approximation | 1 INTRODUCTION . Graphs are used to model data in a diverse range of applications such as social networks ( Hamilton et al. , 2017a ) , biological networks ( Fout et al. , 2017 ) and e-commerce interactions ( Yang , 2019 ) . Recently , there has been an increased interest in applying machine learning techniques to grap... | The paper presents a subgraph approximation method to reduce communication in distributed GCN training. The authors observe that up until a certain point it is possible to delete vertices in a graph without a large accuracy loss for GCN training. Based on the observation, the authors propose the subgraph approximation... | SP:b7fdf5984d0bd1e4caff646d1a403b18c854e7bd |
Long-tail learning via logit adjustment | 1 INTRODUCTION . Real-world classification problems typically exhibit a long-tailed label distribution , wherein most labels are associated with only a few samples ( Van Horn & Perona , 2017 ; Buda et al. , 2017 ; Liu et al. , 2019 ) . Owing to this paucity of samples , generalisation on such labels is challenging ; mo... | This well-written paper re-visits the idea of logit adjustment to tackle long-tailed problems. The paper begins by setting up a statistical framework and use that to deliver two ways of realizing the logits adjustment effectively. They further prove the potential of such an approach by benchmarking it with several rela... | SP:49f0ff35a1d39031bc13b793b04f9d3c84fbfb6e |
Long-tail learning via logit adjustment | 1 INTRODUCTION . Real-world classification problems typically exhibit a long-tailed label distribution , wherein most labels are associated with only a few samples ( Van Horn & Perona , 2017 ; Buda et al. , 2017 ; Liu et al. , 2019 ) . Owing to this paucity of samples , generalisation on such labels is challenging ; mo... | This paper proposes an unifying statistical framework for imbalanced or long-tailed data, where the number of samples for some of the classes may be extremely small compared with other classes. Previous methods work empirically well, but was not consistent, meaning that even in the infinite sample limit, the minimiser ... | SP:49f0ff35a1d39031bc13b793b04f9d3c84fbfb6e |
Long-tail learning via logit adjustment | 1 INTRODUCTION . Real-world classification problems typically exhibit a long-tailed label distribution , wherein most labels are associated with only a few samples ( Van Horn & Perona , 2017 ; Buda et al. , 2017 ; Liu et al. , 2019 ) . Owing to this paucity of samples , generalisation on such labels is challenging ; mo... | This paper provides a statistical framework for long-tail learning by revisiting the idea of logit adjustment based on the label frequencies. The proposed framework then yields two variant techniques that follow the paradigm of weight normalization or loss modification. Compared with the existing methods, the proposed ... | SP:49f0ff35a1d39031bc13b793b04f9d3c84fbfb6e |
Disentangling Representations of Text by Masking Transformers | 1 INTRODUCTION AND MOTIVATION . Large-scale pretrained models such as ELMo ( Peters et al. , 2018 ) , BERT ( Devlin et al. , 2018 ) , and XLNet ( Yang et al. , 2019 ) have come to dominate in modern natural language processing ( NLP ) . Such models rely on self-supervision over large datasets to learn general-purpose r... | This paper proposes a masking strategy to identify subnetworks within language models responsible for predicting different text features. This approach requires no fine-tuning of model parameters and still achieves better results compared to previous approaches. Their experimental results on the movie domain show some ... | SP:2cfbc8524de28aae401ef94e47d242b3997c07c1 |
Disentangling Representations of Text by Masking Transformers | 1 INTRODUCTION AND MOTIVATION . Large-scale pretrained models such as ELMo ( Peters et al. , 2018 ) , BERT ( Devlin et al. , 2018 ) , and XLNet ( Yang et al. , 2019 ) have come to dominate in modern natural language processing ( NLP ) . Such models rely on self-supervision over large datasets to learn general-purpose r... | The paper presents a way to learn disentangled representations with respect to target attributes of interest by learning to mask weights or activations. A particular piece of text is encoded into distinct vectors that capture different factors of variation in the data. The method involves learning masks for each factor... | SP:2cfbc8524de28aae401ef94e47d242b3997c07c1 |
Disentangling Representations of Text by Masking Transformers | 1 INTRODUCTION AND MOTIVATION . Large-scale pretrained models such as ELMo ( Peters et al. , 2018 ) , BERT ( Devlin et al. , 2018 ) , and XLNet ( Yang et al. , 2019 ) have come to dominate in modern natural language processing ( NLP ) . Such models rely on self-supervision over large datasets to learn general-purpose r... | The paper proposes a procedure to extract disentangled representations from pretrained BERT models. In particular, the paper proposes learning binary masks over BERT weights (or, as an alternative, over BERT activations) such that the resulting representations correspond to the desired aspect representations. The model... | SP:2cfbc8524de28aae401ef94e47d242b3997c07c1 |
Contextual Transformation Networks for Online Continual Learning | Continual learning methods with fixed architectures rely on a single network to learn models that can perform well on all tasks . As a result , they often only accommodate common features of those tasks but neglect each task ’ s specific features . On the other hand , dynamic architecture methods can have a separate ne... | This paper introduces a continual learning method called Contextual Transformation Networks (CTNs). CTNs consist of a base network and a controller, which outputs task-specific feature modulators. Both these have independent memories used to reduce forgetting. Additionally, the base network has an additional regularisa... | SP:f39013d44b9908838bbd3d71564dac2ed4b55564 |
Contextual Transformation Networks for Online Continual Learning | Continual learning methods with fixed architectures rely on a single network to learn models that can perform well on all tasks . As a result , they often only accommodate common features of those tasks but neglect each task ’ s specific features . On the other hand , dynamic architecture methods can have a separate ne... | This paper tackles online continual neural network learning (following Lopez-Paz & Ranzato, 2017) with a combination of techniques: (1) a controller (or base parameter modulator, or hypernetwork) is introduced which produces task-specific scale and shift parameters, which modulate the feature maps of a base model (Pere... | SP:f39013d44b9908838bbd3d71564dac2ed4b55564 |
Contextual Transformation Networks for Online Continual Learning | Continual learning methods with fixed architectures rely on a single network to learn models that can perform well on all tasks . As a result , they often only accommodate common features of those tasks but neglect each task ’ s specific features . On the other hand , dynamic architecture methods can have a separate ne... | This paper proposes CTN (Contextual Trasnformer Networks) for online continual learning. In particular, the authors introduce a dual memory framework that contains an episodic memory for base networks and semantic memory for task controllers. The overall framework is optimized with bi-level optimization. In addition, t... | SP:f39013d44b9908838bbd3d71564dac2ed4b55564 |
The Logical Options Framework | 1 INTRODUCTION . To operate in the real world , intelligent agents must be able to make long-term plans by reasoning over symbolic abstractions while also maintaining the ability to react to low-level stimuli in their environment ( Zhang & Sridharan , 2020 ) . Many environments obey rules that can be represented as log... | This paper is on a new RL framework that leverages logical reasoning to improve the learning performance of RL agents. In particular, the knowledge is encoded using LTL, and includes both safety knowledge (used for reward function definition) and liveness knowledge (used for constructing FSA). THe developed framework h... | SP:0731025a467f9c3e648608248eada5a23aff5b49 |
The Logical Options Framework | 1 INTRODUCTION . To operate in the real world , intelligent agents must be able to make long-term plans by reasoning over symbolic abstractions while also maintaining the ability to react to low-level stimuli in their environment ( Zhang & Sridharan , 2020 ) . Many environments obey rules that can be represented as log... | The authors propose the Logical Options Framework (LOF) --- a framework for reasoning over high-level plans and learning low-level control policies. This framework uses Linear Temporal Logic (LTL) to specify properties (high-level tasks) in terms of propositions. The authors propose a framework in which a separate sub-... | SP:0731025a467f9c3e648608248eada5a23aff5b49 |
The Logical Options Framework | 1 INTRODUCTION . To operate in the real world , intelligent agents must be able to make long-term plans by reasoning over symbolic abstractions while also maintaining the ability to react to low-level stimuli in their environment ( Zhang & Sridharan , 2020 ) . Many environments obey rules that can be represented as log... | In this paper, the author proposes logic option framework for learning policies that satisfy the logic constraints. Temporal logic rules are converted into finite state machine and each of them is then converted to a learnable option, where the logic proposition is used as the reward function. Given the options, it s... | SP:0731025a467f9c3e648608248eada5a23aff5b49 |
Filter pre-pruning for improved fine-tuning of quantized deep neural networks | Deep Neural Networks ( DNNs ) have many parameters and activation data , and these both are expensive to implement . One method to reduce the size of the DNN is to quantize the pre-trained model by using a low-bit expression for weights and activations , using fine-tuning to recover the drop in accuracy . However , it ... | This work present a Pruning mechanism for Quantization scenario. Duo to the low-bits effects, the quantized network is hard to train properly. Therefore, authors provide a new method call Pruning for Quantization (PfQ) and a workflow to solve the model compression problem practically. Comparing to some current quantiza... | SP:43fc643b6ac6560e48804865dd7a57398743b664 |
Filter pre-pruning for improved fine-tuning of quantized deep neural networks | Deep Neural Networks ( DNNs ) have many parameters and activation data , and these both are expensive to implement . One method to reduce the size of the DNN is to quantize the pre-trained model by using a low-bit expression for weights and activations , using fine-tuning to recover the drop in accuracy . However , it ... | This paper studies the effect of quantization during training together with batch normalization in quantized deep neural networks. The compound effect of convolution and batch normalization on the dynamic range of activations has implications on the progress of training. The authors propose a protocol for training a qu... | SP:43fc643b6ac6560e48804865dd7a57398743b664 |
Filter pre-pruning for improved fine-tuning of quantized deep neural networks | Deep Neural Networks ( DNNs ) have many parameters and activation data , and these both are expensive to implement . One method to reduce the size of the DNN is to quantize the pre-trained model by using a low-bit expression for weights and activations , using fine-tuning to recover the drop in accuracy . However , it ... | This paper proposed to prune certain channels to improve the accuracy of quantized DNN model. The motivation comes from the observation that the channels which have small variance are actually harmful to the quantization-aware training. The authors show that these channels with small variance can be pruned without sign... | SP:43fc643b6ac6560e48804865dd7a57398743b664 |
It Is Likely That Your Loss Should be a Likelihood | 1 INTRODUCTION . Choosing the right loss matters . Many common losses arise from likelihoods , such as the squared error loss from the normal distribution , absolute error from the Laplace distribution , and the cross entropy loss from the softmax distribution . The same is true of regularizers , where L2 arises from a... | The paper proposes the use of complete parametrized likelihoods for providing supervision in place of the commonly used loss functions. The normal distribution, the categorical distribution defined by softmax and the likelihood of the robust rho-estimator are considered. The main idea is that by including the parameter... | SP:c3b3b5c69080aba6363c9120e42b3e68ea633c68 |
It Is Likely That Your Loss Should be a Likelihood | 1 INTRODUCTION . Choosing the right loss matters . Many common losses arise from likelihoods , such as the squared error loss from the normal distribution , absolute error from the Laplace distribution , and the cross entropy loss from the softmax distribution . The same is true of regularizers , where L2 arises from a... | This paper studied loss functions by interpreting them from a likelihood viewpoint and by proposing to optimize "full" likelihoods for robust modeling, outlier-detection, and re-calibration purposes. Many loss functions stem from maximum likelihood estimation (MLE). For instance, the quadratic loss stems from MLE under... | SP:c3b3b5c69080aba6363c9120e42b3e68ea633c68 |
It Is Likely That Your Loss Should be a Likelihood | 1 INTRODUCTION . Choosing the right loss matters . Many common losses arise from likelihoods , such as the squared error loss from the normal distribution , absolute error from the Laplace distribution , and the cross entropy loss from the softmax distribution . The same is true of regularizers , where L2 arises from a... | This paper's starting point is a probabilistic interpretation of three losses: MSE, cross-entropy, and the loss introduced in reference (Barron 2019). It proposes the minimize the average loss jointly over the predictor's parameters $\theta$, but also the loss' own parameters $\phi$, vary, yielding an adaptive loss. It... | SP:c3b3b5c69080aba6363c9120e42b3e68ea633c68 |
Rethinking Embedding Coupling in Pre-trained Language Models | 1 INTRODUCTION . The performance of models in natural language processing ( NLP ) has dramatically improved in recent years , mainly driven by advances in transfer learning from large amounts of unlabeled data ( Howard & Ruder , 2018 ; Devlin et al. , 2019 ) . The most successful paradigm consists of pre-training a lar... | This work studied the impact of embedding coupling in pre-trained language models, by taking a multilingual model as backbone. The major finding is that decoupling the input and output embedding shapes can bring benefits, and the output embedding plays an important role in the transferability of pre-trained representat... | SP:456ad1760ac9923417ff5158e569452b75036646 |
Rethinking Embedding Coupling in Pre-trained Language Models | 1 INTRODUCTION . The performance of models in natural language processing ( NLP ) has dramatically improved in recent years , mainly driven by advances in transfer learning from large amounts of unlabeled data ( Howard & Ruder , 2018 ; Devlin et al. , 2019 ) . The most successful paradigm consists of pre-training a lar... | This work investigated the strategy of reallocating parameters of multilingual language models for improving their cross lingual transferability. Authors first decoupled the input and output embeddings and showed that the capacity of output embedding is more important than input embedding. Then, they proposed a Rebala... | SP:456ad1760ac9923417ff5158e569452b75036646 |
Rethinking Embedding Coupling in Pre-trained Language Models | 1 INTRODUCTION . The performance of models in natural language processing ( NLP ) has dramatically improved in recent years , mainly driven by advances in transfer learning from large amounts of unlabeled data ( Howard & Ruder , 2018 ; Devlin et al. , 2019 ) . The most successful paradigm consists of pre-training a lar... | This paper systematically studies the impact of embedding coupling with multilingual language models. The authors observe that while na¨ıvely decoupling the input and output embedding parameters does not consistently improve downstream evaluation metrics, decoupling their shapes comes with a host of benefits. Moreover... | SP:456ad1760ac9923417ff5158e569452b75036646 |
On the geometry of generalization and memorization in deep neural networks | 1 INTRODUCTION . Deep neural networks have many more learnable parameters than training examples , and could simply memorize the data instead of converging to a generalizable solution ( Novak et al. , 2018 ) . Moreover , standard regularization methods are insufficient to eliminate memorization of random labels , and n... | This paper analyses memorization in DNNs, from the lens of memorization = fitting random labels, and finds that it seems to happen in later layers. These results are obtained using the MFTMA framework, a manifold analysis tool, testing geometric properties of individual layers. The analysis also attempts to explain why... | SP:75052bcb31029d6f128915eedcb3c9ee5143a9a4 |
On the geometry of generalization and memorization in deep neural networks | 1 INTRODUCTION . Deep neural networks have many more learnable parameters than training examples , and could simply memorize the data instead of converging to a generalizable solution ( Novak et al. , 2018 ) . Moreover , standard regularization methods are insufficient to eliminate memorization of random labels , and n... | This paper investigates memorization in deep neural networks (DNNs). Authors leverage mean field theoretic geometric analysis method (MFTMA) to analyze when and where memorization occurs in a DNN. Through empirical analysis, they show that i) generalizing feature are learned initially and that memorization happen late... | SP:75052bcb31029d6f128915eedcb3c9ee5143a9a4 |
On the geometry of generalization and memorization in deep neural networks | 1 INTRODUCTION . Deep neural networks have many more learnable parameters than training examples , and could simply memorize the data instead of converging to a generalizable solution ( Novak et al. , 2018 ) . Moreover , standard regularization methods are insufficient to eliminate memorization of random labels , and n... | The authors apply MFTMA to DNNs trained on CIFAR with label noise to analyze their behaviors between generalization and memorization. Based on experimental results, they claim that what is involved in memorization are not lower layers but higher layers. This claim is convincing. Another claim that this is not caused by... | SP:75052bcb31029d6f128915eedcb3c9ee5143a9a4 |
Grounding Physical Concepts of Objects and Events Through Dynamic Visual Reasoning | 1 INTRODUCTION . Visual reasoning in dynamic scenes involves both the understanding of compositional properties , relationships , and events of objects , and the inference and prediction of their temporal and causal structures . As depicted in Fig . 1 , to answer the question “ What will happen next ? ” based on the ob... | The paper presents a neural-symbolic approach for video question answering. In particular, the authors propose the Dynamic Concept Learner (DCL), a network architecture to localize object trajectories over time and further build a graph network to model the physical interaction between objects and events. Finally, a st... | SP:a403544a76c5222b75621e25723ec8d7c9d98e5f |
Grounding Physical Concepts of Objects and Events Through Dynamic Visual Reasoning | 1 INTRODUCTION . Visual reasoning in dynamic scenes involves both the understanding of compositional properties , relationships , and events of objects , and the inference and prediction of their temporal and causal structures . As depicted in Fig . 1 , to answer the question “ What will happen next ? ” based on the ob... | The paper studies the problem of dynamic visual reasoning on the recently proposed synthetic video QA dataset CLEVRER for understanding visual properties, physical events, the causal relationship between events, and making future and counterfactual predictions. The paper proposes a joint framework called Dynamic Concep... | SP:a403544a76c5222b75621e25723ec8d7c9d98e5f |
Grounding Physical Concepts of Objects and Events Through Dynamic Visual Reasoning | 1 INTRODUCTION . Visual reasoning in dynamic scenes involves both the understanding of compositional properties , relationships , and events of objects , and the inference and prediction of their temporal and causal structures . As depicted in Fig . 1 , to answer the question “ What will happen next ? ” based on the ob... | The paper proposes a new framework, Dynamic Concept Learner (DCL), which learns by watching videos and reading questions/answers. It is inspired by prior work which combines symbolic representations with video dynamics modeling, but unlike prior work, here the authors do not use any additional supervision except for q... | SP:a403544a76c5222b75621e25723ec8d7c9d98e5f |
GeDi: Generative Discriminator Guided Sequence Generation | 1 INTRODUCTION . Natural language generation has seen great progress with the advent of Transformers ( Vaswani et al. , 2017 ) and large scale training ( Radford et al. , 2017 ; 2018 ; 2019 ; Brown et al. , 2020 ) . Large language models ( LMs ) like GPT-2 ( Radford et al. , 2019 ) and GPT-3 ( Brown et al. , 2020 ) are... | In this paper, the authors propose an efficient method for controllable language generation of large pre-trained LMs (e.g., GPT2). The main idea is to use a smaller, compared to the LM to control, language model trained with control code (Keskar et al., 2019) to generate a per-token score $P(c|x_{1:t}$) to steer the or... | SP:bd6cf6c4ef7d78b8f5d239f539d94f337c8ec4b1 |
GeDi: Generative Discriminator Guided Sequence Generation | 1 INTRODUCTION . Natural language generation has seen great progress with the advent of Transformers ( Vaswani et al. , 2017 ) and large scale training ( Radford et al. , 2017 ; 2018 ; 2019 ; Brown et al. , 2020 ) . Large language models ( LMs ) like GPT-2 ( Radford et al. , 2019 ) and GPT-3 ( Brown et al. , 2020 ) are... | The paper proposed a method —- GeDi — to generate guided and controlled text from a large language model (LM). The method utilizes smaller LMs as generative discriminators to guide generation from large LMs to make them safer and more controllable. By safer and controllable they emphasis on the toxicity, hate, bias, an... | SP:bd6cf6c4ef7d78b8f5d239f539d94f337c8ec4b1 |
GeDi: Generative Discriminator Guided Sequence Generation | 1 INTRODUCTION . Natural language generation has seen great progress with the advent of Transformers ( Vaswani et al. , 2017 ) and large scale training ( Radford et al. , 2017 ; 2018 ; 2019 ; Brown et al. , 2020 ) . Large language models ( LMs ) like GPT-2 ( Radford et al. , 2019 ) and GPT-3 ( Brown et al. , 2020 ) are... | The paper considers the problem of attribute-based sequence generation, particularly in language models. Authors propose a framework “GeDi” which learns a generative classifier for controlling generation from a large language model. With experiments on publicly available datasets and models, and including human-evaluat... | SP:bd6cf6c4ef7d78b8f5d239f539d94f337c8ec4b1 |
Efficient Reinforcement Learning in Resource Allocation Problems Through Permutation Invariant Multi-task Learning | 1 INTRODUCTION . Sample efficiency in reinforcement learning ( RL ) is an elusive goal . Recent attempts at increasing the sample efficiency of RL implementations have focused to a large extent on incorporating models into the training process : Xu et al . ( 2019 ) ; Clavera et al . ( 2018 ) ; Zhang et al . ( 2018 ) ; ... | This paper proposes an approach to reducing the sample complexity in multi-task reinforcement learning using permutation invariant policies. The main premise of the paper is that certain families of tasks exhibit approximate forms of symmetry, i.e., applying a permutation to the state/action variables would make all ta... | SP:2ab4f27ae293c71faa682bb0500dd42213f2694b |
Efficient Reinforcement Learning in Resource Allocation Problems Through Permutation Invariant Multi-task Learning | 1 INTRODUCTION . Sample efficiency in reinforcement learning ( RL ) is an elusive goal . Recent attempts at increasing the sample efficiency of RL implementations have focused to a large extent on incorporating models into the training process : Xu et al . ( 2019 ) ; Clavera et al . ( 2018 ) ; Zhang et al . ( 2018 ) ; ... | This paper addresses sequential resource allocation problem using reinforcement learning, where sample efficiency is the focus of the paper. The authors identify a key property in the targeted resource allocation problems -- the permutation invariance -- which intrinsically implies the independency of samples at differ... | SP:2ab4f27ae293c71faa682bb0500dd42213f2694b |
Efficient Reinforcement Learning in Resource Allocation Problems Through Permutation Invariant Multi-task Learning | 1 INTRODUCTION . Sample efficiency in reinforcement learning ( RL ) is an elusive goal . Recent attempts at increasing the sample efficiency of RL implementations have focused to a large extent on incorporating models into the training process : Xu et al . ( 2019 ) ; Clavera et al . ( 2018 ) ; Zhang et al . ( 2018 ) ; ... | The paper proposes an RL algorithm for multi-task learning. Under certain assumptions, the paper proves a sample complexity result for this setting. The paper presents a new algorithm based on this approach. The empirical results on the task of sequential portfolio optimization shows that this approach performs better ... | SP:2ab4f27ae293c71faa682bb0500dd42213f2694b |
Hellinger Distance Constrained Regression | 1 INTRODUCTION . Policy gradient algorithms are methods of model-free reinforcement learning that optimize policy through differentiating expected discounted return . Despite the simplicity , to converge , these methods should stay on-policy because of the first-order approximation of state visitation frequencies . Thi... | This paper proposes a supervised learning for off-policy reinforcement learning. It exploits the Hellinger distance instead of KL divergence. Thus it achieves tighter lower bound of the expected culmulative return than that using the KL divergence. Moreover, the new lower bound is policy independent. The experimental r... | SP:ef7e83384f580d3bc6ad5104d87f7f42fb8bd5c6 |
Hellinger Distance Constrained Regression | 1 INTRODUCTION . Policy gradient algorithms are methods of model-free reinforcement learning that optimize policy through differentiating expected discounted return . Despite the simplicity , to converge , these methods should stay on-policy because of the first-order approximation of state visitation frequencies . Thi... | The authors propose the use of the Hellinger distance instead of KL divergence to constrain new policies to remain close to the behavior policy. The technical aspects are straightforward, noting that Hellinger provides tighter bounds on total variation than KL, and can straightforwardly be plugged into the CPI/TRPO bou... | SP:ef7e83384f580d3bc6ad5104d87f7f42fb8bd5c6 |
Hellinger Distance Constrained Regression | 1 INTRODUCTION . Policy gradient algorithms are methods of model-free reinforcement learning that optimize policy through differentiating expected discounted return . Despite the simplicity , to converge , these methods should stay on-policy because of the first-order approximation of state visitation frequencies . Thi... | The paper provides a new metric - Hellinger distance to be combined with trust region ideas in policy optimization. The major difference from prior work is the change of this distance metric. The paper shows that with this distance metric, along with Lagrangian relaxation, one could show analytic results of improved po... | SP:ef7e83384f580d3bc6ad5104d87f7f42fb8bd5c6 |
Differentiable Segmentation of Sequences | 1 INTRODUCTION . Non-stationarity is a classical challenge in the analysis of sequential data . A common source of non-stationarity is the presence of change points , where the data-generating process switches its dynamics from one regime to another regime . In some applications , the detection of change points is of p... | The paper describes the use of two-sided power functions for differentiable approximation of segmentation in discrete sequences with monotonic segmentation (each event in the sequence is defined by one continuous interval). The authors show that their particular parametrization allows to control exactly the length, sta... | SP:99a4cd5e5df612e8319b5abd1162696a2f7453fd |
Differentiable Segmentation of Sequences | 1 INTRODUCTION . Non-stationarity is a classical challenge in the analysis of sequential data . A common source of non-stationarity is the presence of change points , where the data-generating process switches its dynamics from one regime to another regime . In some applications , the detection of change points is of p... | The paper proposes a relaxed way to solve the segmentation of sequence that can directly leverage the deep learning architectures. The relaxed model allows each segmentation parameter to be a linear interpolations between two consecutive parameters depends on a continuous warping function. The paper then proposes to us... | SP:99a4cd5e5df612e8319b5abd1162696a2f7453fd |
Differentiable Segmentation of Sequences | 1 INTRODUCTION . Non-stationarity is a classical challenge in the analysis of sequential data . A common source of non-stationarity is the presence of change points , where the data-generating process switches its dynamics from one regime to another regime . In some applications , the detection of change points is of p... | The proposed paper introduces a novel approach for the segmentation of sequences. The proposed method is based on two-sided power distributions (TSP) that are mathematically well-define and enable differentiability. The main goal of the method is to jointly optimize model parameters, including the segmentation function... | SP:99a4cd5e5df612e8319b5abd1162696a2f7453fd |
Bounded Myopic Adversaries for Deep Reinforcement Learning Agents | Adversarial attacks against deep neural networks have been widely studied . Adversarial examples for deep reinforcement learning ( DeepRL ) have significant security implications , due to the deployment of these algorithms in many application domains . In this work we formalize an optimal myopic adversary for deep rein... | This paper proposes an optimal myopic adversary for deep reinforcement learning agent, in which the adversary finds a bounded perturbation of the state that minimizes the value of the action taken by the agent. The authors introduce a differentiable approximation for the optimal myopic adversarial formulation that lead... | SP:f9aec02ea7ba21b7840c78f46af8cb6aa504e4ab |
Bounded Myopic Adversaries for Deep Reinforcement Learning Agents | Adversarial attacks against deep neural networks have been widely studied . Adversarial examples for deep reinforcement learning ( DeepRL ) have significant security implications , due to the deployment of these algorithms in many application domains . In this work we formalize an optimal myopic adversary for deep rein... | The work introduces a new method for observation-perturbation adversarial attacks against deep neural network policies. The idealized version of the method is an optimal attack assuming (a) the agent's Q-values are calibrated; (b) the attacker can only act once, at the current time step. While the attack is in fact con... | SP:f9aec02ea7ba21b7840c78f46af8cb6aa504e4ab |
Bounded Myopic Adversaries for Deep Reinforcement Learning Agents | Adversarial attacks against deep neural networks have been widely studied . Adversarial examples for deep reinforcement learning ( DeepRL ) have significant security implications , due to the deployment of these algorithms in many application domains . In this work we formalize an optimal myopic adversary for deep rein... | This paper proposes using a better quantitative metric to conduct an attack on a DRL learner. The attack is limited to an attack on a state (not over multiple states) and aims to lower the Q value from this state by making the worst action be the action chosen to be played. The experiment on Atari games show promising ... | SP:f9aec02ea7ba21b7840c78f46af8cb6aa504e4ab |
Meta-learning Symmetries by Reparameterization | 1 INTRODUCTION . In deep learning , the convolutional neural network ( CNN ) ( LeCun et al. , 1998 ) is a prime example of exploiting equivariance to a symmetry transformation to conserve parameters and improve generalization . In image classification ( Russakovsky et al. , 2015 ; Krizhevsky et al. , 2012 ) and audio p... | In this paper, the authors propose MSR, a parametrization of convolutional kernels that allows for meta-learning symmetries shared between several tasks. Each kernel is represented as a product of a structure matrix and a vector of the kernel weights. The kernel weights are updated during the inner loop. The structure... | SP:4c76919a2aa2b645c86a072f7bafbb2c384dc58a |
Meta-learning Symmetries by Reparameterization | 1 INTRODUCTION . In deep learning , the convolutional neural network ( CNN ) ( LeCun et al. , 1998 ) is a prime example of exploiting equivariance to a symmetry transformation to conserve parameters and improve generalization . In image classification ( Russakovsky et al. , 2015 ; Krizhevsky et al. , 2012 ) and audio p... | The authors present a meta-learning approach for network equivariance where the key idea is that equivariance to a finite group of transformations can be achieved by identifying the sharing pattern of weights. Their proposition claims that a fully connected layer $\phi: \mathbb{R}^n \rightarrow \mathbb{R}^m$ with weigh... | SP:4c76919a2aa2b645c86a072f7bafbb2c384dc58a |
Meta-learning Symmetries by Reparameterization | 1 INTRODUCTION . In deep learning , the convolutional neural network ( CNN ) ( LeCun et al. , 1998 ) is a prime example of exploiting equivariance to a symmetry transformation to conserve parameters and improve generalization . In image classification ( Russakovsky et al. , 2015 ; Krizhevsky et al. , 2012 ) and audio p... | The paper presents a meta-learning algorithm to learn/encode equivariance into deep nets. The main idea is to decompose the model parameters into two parts, a spatial sharing pattern, and the trainable weights. When transferring to a new task, the sharing pattern is fixed and only the remaining trainable weights are tu... | SP:4c76919a2aa2b645c86a072f7bafbb2c384dc58a |
Extrapolatable Relational Reasoning With Comparators in Low-Dimensional Manifolds | 1 INTRODUCTION . The goal of Artificial Intelligence research , first proposed in the 1950s and reiterated many times , is to create machine intelligence comparable to that of a human being . While today ’ s deep learning based systems achieve human-comparable performances in specific tasks such as object classificatio... | The paper proposes a simple low-dimensional projection and comparison module for extrapolatable relational reasoning. The proposed module first projects object embeddings into lower-dimensional manifolds, compare and concatenate them to form a vector, before finally making a prediction. The paper shows that on the maxi... | SP:66cc7af855b0af31f435da307b8c6edf48e80c42 |
Extrapolatable Relational Reasoning With Comparators in Low-Dimensional Manifolds | 1 INTRODUCTION . The goal of Artificial Intelligence research , first proposed in the 1950s and reiterated many times , is to create machine intelligence comparable to that of a human being . While today ’ s deep learning based systems achieve human-comparable performances in specific tasks such as object classificatio... | The paper presents two neural network design ideas: low-dimensional projection and arithmetic comparator. By integrating these two ideas into CNNs, the model can solve a set of tasks that require recognizing 1-dimensional properties of objects (such as the size, the color, etc), and making a comparison of these propert... | SP:66cc7af855b0af31f435da307b8c6edf48e80c42 |
Extrapolatable Relational Reasoning With Comparators in Low-Dimensional Manifolds | 1 INTRODUCTION . The goal of Artificial Intelligence research , first proposed in the 1950s and reiterated many times , is to create machine intelligence comparable to that of a human being . While today ’ s deep learning based systems achieve human-comparable performances in specific tasks such as object classificatio... | This paper addresses an inductive bias for relational reasoning tasks to improve generalization performance on out-of-distribution scenarios (so called extrapolation), which the value ranges (continuous variables) or value itself (discrete variables) of the training/test dataset do not overlap. The proposed idea is to ... | SP:66cc7af855b0af31f435da307b8c6edf48e80c42 |
Provably robust classification of adversarial examples with detection | 1 INTRODUCTION . Despite popularity and success of deep neural networks in many applications , their performance declines sharply in adversarial settings . Small adversarial perturbations are shown to greatly deteriorate the performance of neural network classifiers , which creates a growing concern for utilizing them ... | This paper aims to train networks that can map a possibly $\ell_\infty$-perturbed input to its class provably or map this input to the “abstain class” provably. This is achieved by training on the IBP output boxes together with a new loss function. The method diverts from the classical setting in which the classifier n... | SP:a6006c32d8bdbbd5f9d30303e688f9a0266ad29e |
Provably robust classification of adversarial examples with detection | 1 INTRODUCTION . Despite popularity and success of deep neural networks in many applications , their performance declines sharply in adversarial settings . Small adversarial perturbations are shown to greatly deteriorate the performance of neural network classifiers , which creates a growing concern for utilizing them ... | This paper deals with the problem of bounding the amount of errors that a model can make when attacked by an adversary limited to small perturbation of an input image. Similarly to previous paper, it proposes to extend the classifier with a detector to identify adversarial examples. As opposed to previous papers, the a... | SP:a6006c32d8bdbbd5f9d30303e688f9a0266ad29e |
Provably robust classification of adversarial examples with detection | 1 INTRODUCTION . Despite popularity and success of deep neural networks in many applications , their performance declines sharply in adversarial settings . Small adversarial perturbations are shown to greatly deteriorate the performance of neural network classifiers , which creates a growing concern for utilizing them ... | In this paper, the authors propose an additional "abstain/detection" loss term into training, so that the classifier can either robustly classify or detect an adversarial attack. They extend the interval bound propagation method for certified robustness under L_infity perturbations (a simple bounding in changes of the ... | SP:a6006c32d8bdbbd5f9d30303e688f9a0266ad29e |
Go with the flow: Adaptive control for Neural ODEs | 1 INTRODUCTION . The interpretation of artificial neural networks as continuous-time dynamical systems has led to both theoretical and practical advances in representation learning . According to this interpretation , the separate layers of a deep neural network are understood to be a discretization of a continuous-tim... | The paper proposes a new class of neural ODE based models, called neurally-controlled ODEs (N-CODE). Instead of directly learning the weights of the neural network parameterizing the vector field f of the ODE, the paper instead proposes to learn a controller that takes as input the initial state and outputs the initial... | SP:f13fb67c966196fbc7fe3a264a1f2a224e651321 |
Go with the flow: Adaptive control for Neural ODEs | 1 INTRODUCTION . The interpretation of artificial neural networks as continuous-time dynamical systems has led to both theoretical and practical advances in representation learning . According to this interpretation , the separate layers of a deep neural network are understood to be a discretization of a continuous-tim... | Comment: Summary: This paper presents a technique for more expressive neural ordinary differential equations (NODE) flows. Instead of learning a fixed set of parameters $\theta$ that governs the ODE dynamic, the proposed approach learns dynamic parameters that evolve over time. The authors propose two variants of the m... | SP:f13fb67c966196fbc7fe3a264a1f2a224e651321 |
Go with the flow: Adaptive control for Neural ODEs | 1 INTRODUCTION . The interpretation of artificial neural networks as continuous-time dynamical systems has led to both theoretical and practical advances in representation learning . According to this interpretation , the separate layers of a deep neural network are understood to be a discretization of a continuous-tim... | The paper introduces a novel approach N-CODE, based on Neural Ordinary Differential Equations (NODE), that increases the expressivity of continuous-time neural nets by using approaches from Control theory. N-CODE, in contrast to NODE, can learn a family of vector fields and is therefore able to flexibly adjust the flo... | SP:f13fb67c966196fbc7fe3a264a1f2a224e651321 |
Automatic Data Augmentation for Generalization in Reinforcement Learning | 1 Introduction . Generalization to new environments remains a major challenge in deep reinforcement learning ( RL ) . Current methods fail to generalize to unseen environments even when trained on similar settings [ 19 , 51 , 71 , 11 , 21 , 12 , 60 ] . This indicates that standard RL agents memorize specific trajectori... | This paper presents a method that utilizes data augmentation for image-based reinforcement learning. The data augmentation is used to regularize the policy and function approximation in the proposed method. In addition, a method for automatically identifying effective ways of data augmentation is proposed. The experime... | SP:f2325f97eb57b82ce95c86638af9d733d325b45d |
Automatic Data Augmentation for Generalization in Reinforcement Learning | 1 Introduction . Generalization to new environments remains a major challenge in deep reinforcement learning ( RL ) . Current methods fail to generalize to unseen environments even when trained on similar settings [ 19 , 51 , 71 , 11 , 21 , 12 , 60 ] . This indicates that standard RL agents memorize specific trajectori... | This paper tackles the problem of generalization in deep RL via data augmentation. It provides a framework for automatic data augmentation based on UCB, RL^2, or MAML. When UCB is combined with regularization of the policy and value function so that their outputs are invariant to transformations (such as rotation, crop... | SP:f2325f97eb57b82ce95c86638af9d733d325b45d |
Automatic Data Augmentation for Generalization in Reinforcement Learning | 1 Introduction . Generalization to new environments remains a major challenge in deep reinforcement learning ( RL ) . Current methods fail to generalize to unseen environments even when trained on similar settings [ 19 , 51 , 71 , 11 , 21 , 12 , 60 ] . This indicates that standard RL agents memorize specific trajectori... | This paper proposes an automatic data augmentation approach for RL tasks. Specifically, it takes UCB for data selection and introduces two regularization terms for actor-critic algorithms' policy and value function. Then this paper evaluated the approach based on the Procgen benchmark and demonstrated that it outperfor... | SP:f2325f97eb57b82ce95c86638af9d733d325b45d |
Noise against noise: stochastic label noise helps combat inherent label noise | The noise in stochastic gradient descent ( SGD ) provides a crucial implicit regularization effect , previously studied in optimization by analyzing the dynamics of parameter updates . In this paper , we are interested in learning with noisy labels , where we have a collection of samples with potential mislabeling . We... | The paper tackles the problem of training under noisy labels. It proposes adding random zero-mean Gaussian noise to the labels during training. It is shown that such a noise induces a variable gradient noise which adaptively increases 1) when the learnt network function has higher curvature around training points and ... | SP:3ef2a669af95e6ba4c15e71b13e080a21885b03b |
Noise against noise: stochastic label noise helps combat inherent label noise | The noise in stochastic gradient descent ( SGD ) provides a crucial implicit regularization effect , previously studied in optimization by analyzing the dynamics of parameter updates . In this paper , we are interested in learning with noisy labels , where we have a collection of samples with potential mislabeling . We... | This paper studies learning robust models with noisy labels. The authors argue that a specific SGD noise induced by stochastic label noise (SLN) can mitigate the effect of label noise. But the common SGD noise cannot achieve this. Then they apply the proposed SLN induced SGD noise to the existing label-correction metho... | SP:3ef2a669af95e6ba4c15e71b13e080a21885b03b |
Noise against noise: stochastic label noise helps combat inherent label noise | The noise in stochastic gradient descent ( SGD ) provides a crucial implicit regularization effect , previously studied in optimization by analyzing the dynamics of parameter updates . In this paper , we are interested in learning with noisy labels , where we have a collection of samples with potential mislabeling . We... | This paper studies the noisy stochastic gradient descent algorithm in noisy label learning. Concretely, the authors added Gaussian noise on the labels rather than the gradient itself. By comparing different noisy SGD algorithms, the authors demonstrate that the proposed SLN algorithm not only help the model to escape f... | SP:3ef2a669af95e6ba4c15e71b13e080a21885b03b |
Poisoned classifiers are not only backdoored, they are fundamentally broken | 1 INTRODUCTION . Backdoor attacks ( Gu et al. , 2017 ; Chen et al. , 2017 ; Turner et al. , 2019 ; Saha et al. , 2020 ) have emerged as a prominent strategy for poisoning classification models . An adversary controlling ( even a relatively small amount of ) the training data can inject a “ trigger ” into the training d... | This submission just describes some phenomenons in a strange setting but not proposes any valuable questions. The authors claim that "anyone with access to the classifier, even without access to any original training data or trigger, can construct several alternative triggers that are as effective or more so at elicit... | SP:2bd729b7aa045bf74e31229c9e76e57af36e804b |
Poisoned classifiers are not only backdoored, they are fundamentally broken | 1 INTRODUCTION . Backdoor attacks ( Gu et al. , 2017 ; Chen et al. , 2017 ; Turner et al. , 2019 ; Saha et al. , 2020 ) have emerged as a prominent strategy for poisoning classification models . An adversary controlling ( even a relatively small amount of ) the training data can inject a “ trigger ” into the training d... | This paper demonstrates that backdoor-poisoned machine learning models can also be vulnerable to alternative triggers. Specifically, adversarial samples that are generated against models robustified with Denoised Smoothing often show backdoor patterns. Therefore, these adversarial samples can be used to create new trig... | SP:2bd729b7aa045bf74e31229c9e76e57af36e804b |
Poisoned classifiers are not only backdoored, they are fundamentally broken | 1 INTRODUCTION . Backdoor attacks ( Gu et al. , 2017 ; Chen et al. , 2017 ; Turner et al. , 2019 ; Saha et al. , 2020 ) have emerged as a prominent strategy for poisoning classification models . An adversary controlling ( even a relatively small amount of ) the training data can inject a “ trigger ” into the training d... | This is an interesting study on the analysis of poisoned classifiers and backdoor attacks. The authors showed that with some post-processing analysis on a poisoned classifier, it is possible to construct effective alternative triggers against a backdoor classifier. In particular, after creating several poisoned classif... | SP:2bd729b7aa045bf74e31229c9e76e57af36e804b |
Towards Defending Multiple Adversarial Perturbations via Gated Batch Normalization | 1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved remarkable performance across a wide areas of applications ( Krizhevsky et al. , 2012 ; Bahdanau et al. , 2014 ; Hinton et al. , 2012 ) , but they are susceptible to adversarial examples ( Szegedy et al. , 2013 ) . These elaborately designed perturbations are... | This paper introduces an algorithm for defending against multiple adversarial attacks (L1, L2, L-inf) by learning separate batch-norm statistics for each attack type. At inference time, the batch-normalized outputs corresponding to the different attack types are averaged according to the probability of attack types as ... | SP:ae5069e4b8fe352a287a7487aff8cf012cbd035d |
Towards Defending Multiple Adversarial Perturbations via Gated Batch Normalization | 1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved remarkable performance across a wide areas of applications ( Krizhevsky et al. , 2012 ; Bahdanau et al. , 2014 ; Hinton et al. , 2012 ) , but they are susceptible to adversarial examples ( Szegedy et al. , 2013 ) . These elaborately designed perturbations are... | The authors propose to improve the robustness against adversarial attacks in deep networks via the use of a customized normalization strategy. Their central methodological suggestion is called gated batch normalization (GBN), and involves a (soft, but some other variants are briefly explored as well) gating mechanism t... | SP:ae5069e4b8fe352a287a7487aff8cf012cbd035d |
Towards Defending Multiple Adversarial Perturbations via Gated Batch Normalization | 1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved remarkable performance across a wide areas of applications ( Krizhevsky et al. , 2012 ; Bahdanau et al. , 2014 ; Hinton et al. , 2012 ) , but they are susceptible to adversarial examples ( Szegedy et al. , 2013 ) . These elaborately designed perturbations are... | Existing defenses are usually robust against certain types of attacks. This paper proposes to build a defense against different types of norm-bounded attacks via gated batch normalization. This work treats different types of adversarial attacks as data from different domains and proposes to use gated batch normalizatio... | SP:ae5069e4b8fe352a287a7487aff8cf012cbd035d |
RetCL: A Selection-based Approach for Retrosynthesis via Contrastive Learning | Retrosynthesis , of which the goal is to find a set of reactants for synthesizing a target product , is an emerging research area of deep learning . While the existing approaches have shown promising results , they currently lack the ability to consider availability ( e.g. , stability or purchasability ) of the reactan... | This submission describes an approach to single-step retrosynthesis based on contrastive learning that selects reactants that can be used to synthesize a target product in a single step. The stated contributions are (1) an approach to retrosynthesis that is constrained to only select “available” starting materials; and... | SP:88387e597974d941b7f4a9c71c7bc42fc433fd4a |
RetCL: A Selection-based Approach for Retrosynthesis via Contrastive Learning | Retrosynthesis , of which the goal is to find a set of reactants for synthesizing a target product , is an emerging research area of deep learning . While the existing approaches have shown promising results , they currently lack the ability to consider availability ( e.g. , stability or purchasability ) of the reactan... | RETCL enumerates all of the candidate molecules (all US patent dataset, all 671k) based on selection scores computed by graph neural networks. The cosine similarity between products and reactants are used to design scores, which is later used for training. The way of ‘cosine similarity to bridge reactants to products a... | SP:88387e597974d941b7f4a9c71c7bc42fc433fd4a |
RetCL: A Selection-based Approach for Retrosynthesis via Contrastive Learning | Retrosynthesis , of which the goal is to find a set of reactants for synthesizing a target product , is an emerging research area of deep learning . While the existing approaches have shown promising results , they currently lack the ability to consider availability ( e.g. , stability or purchasability ) of the reactan... | This paper poses an approach to retrosynthesis that addresses the challenges of (i) availability of reactants and (ii) generalization to unseen templates. To achieve this the authors reformulate retrosynthesis as the selection of reactants from a fixed set, in the case of the USPTO database this is the set of 671,578 c... | SP:88387e597974d941b7f4a9c71c7bc42fc433fd4a |
Ranking Cost: One-Stage Circuit Routing by Directly Optimizing Global Objective Function | 1 INTRODUCTION . As described in Moore ’ s Law ( Schaller , 1997 ) , the number of transistors in a dense integrated circuit ( IC ) increases exponentially over time and the complexity of chips and printed circuit boards ( PCBs ) becomes higher and higher . Such high complexity makes the IC design a time-consuming and ... | In this paper the authors attempt to solve the circuit routing problem through a novel approach of combining search-based routing techniques (e.g. A* search) and evolutionary strategies (e.g. OpenAI-ES). The authors define new parameters e.g. cost maps and a ranking parameter which improve the efficiency and effectiven... | SP:d72d696772354772231126f097a0cb17218625c2 |
Ranking Cost: One-Stage Circuit Routing by Directly Optimizing Global Objective Function | 1 INTRODUCTION . As described in Moore ’ s Law ( Schaller , 1997 ) , the number of transistors in a dense integrated circuit ( IC ) increases exponentially over time and the complexity of chips and printed circuit boards ( PCBs ) becomes higher and higher . Such high complexity makes the IC design a time-consuming and ... | Author considers the circuit routing problem, where each net contains two nodes. The exact circuit routing problem is NP-hard in general, and author proposed to an approximate method. It uses several parameters to control the solution returned by a greedy search, i.e. the order of the nets to be routed and the greedy r... | SP:d72d696772354772231126f097a0cb17218625c2 |
Ranking Cost: One-Stage Circuit Routing by Directly Optimizing Global Objective Function | 1 INTRODUCTION . As described in Moore ’ s Law ( Schaller , 1997 ) , the number of transistors in a dense integrated circuit ( IC ) increases exponentially over time and the complexity of chips and printed circuit boards ( PCBs ) becomes higher and higher . Such high complexity makes the IC design a time-consuming and ... | The paper considers the problem of determining efficient routes for connecting pairs of source, destination points over a circuit board. Due to physical constraints no two routes may intersect, and the routes may not pass through obstacles if there are any. For efficient performance it is desirable that the paths are a... | SP:d72d696772354772231126f097a0cb17218625c2 |
Teleport Graph Convolutional Networks | 1 INTRODUCTION . Graph neural networks ( GNNs ) have shown great capability in solving challenging tasks on graph data such as node classification ( Grover & Leskovec , 2016 ; Kipf & Welling , 2017 ; Veličković et al. , 2017 ; Gao et al. , 2018 ) , graph classification ( Xu et al. , 2018 ; Gao & Ji , 2019 ; You et al... | A new architecture for graph neural networks, which the authors name as Teleport Graph Convolutional Networks (TGL), is proposed in this paper. Teleport graph convolution layer is proposed to address the limitations in message-passing operations of graph neural networks: 1. over-smoothing and 2. over-fitting. The ar... | SP:fc0f6a067f55acbf6b9ad8166ce128831a1829da |
Teleport Graph Convolutional Networks | 1 INTRODUCTION . Graph neural networks ( GNNs ) have shown great capability in solving challenging tasks on graph data such as node classification ( Grover & Leskovec , 2016 ; Kipf & Welling , 2017 ; Veličković et al. , 2017 ; Gao et al. , 2018 ) , graph classification ( Xu et al. , 2018 ; Gao & Ji , 2019 ; You et al... | The paper proposes a method to increase the receptive field of GNNs, while avoiding oversmoothing. The idea is to create extra connections by linking distant nodes based on two criteria: node feature similarity and node structure similarity. Pairs of nodes that are more similar than a threshold are connected. For struc... | SP:fc0f6a067f55acbf6b9ad8166ce128831a1829da |
Teleport Graph Convolutional Networks | 1 INTRODUCTION . Graph neural networks ( GNNs ) have shown great capability in solving challenging tasks on graph data such as node classification ( Grover & Leskovec , 2016 ; Kipf & Welling , 2017 ; Veličković et al. , 2017 ; Gao et al. , 2018 ) , graph classification ( Xu et al. , 2018 ; Gao & Ji , 2019 ; You et al... | This paper analyzed the key issues of the existing message-passing graph convolutional networks. That is, the multiply stacked layer might be over-fitting and over-smoothing. Thus it proposed to choose the neighbors from the entire graph based on the structure-aware and feature-aware relatedness rather than simply choo... | SP:fc0f6a067f55acbf6b9ad8166ce128831a1829da |
Learning to Noise: Application-Agnostic Data Sharing with Local Differential Privacy | 1 INTRODUCTION . The collection of personal data is ubiquitous , and unavoidable for many in everyday life . While this has undeniably improved the quality and user experience of many products and services , evidence of data misuse and data breaches ( Sweeney , 1997 ; Jolly , 2020 ) have brought the concept of data pri... | This paper presents a new privatization mechanism for Local Differential Privacy based on representation learning. The proposed VAE-based method is used for the low-dimensional latent representation of the data and uses the Laplace mechanism to satisfy Local DP. The paper shows this mechanism can be used across various... | SP:bf68937b0e23da990383e48608af226ad5f250cf |
Learning to Noise: Application-Agnostic Data Sharing with Local Differential Privacy | 1 INTRODUCTION . The collection of personal data is ubiquitous , and unavoidable for many in everyday life . While this has undeniably improved the quality and user experience of many products and services , evidence of data misuse and data breaches ( Sweeney , 1997 ; Jolly , 2020 ) have brought the concept of data pri... | For LDP, when applying noise directly to high-dimensional data, the required noise entirely destroys data utility. In this paper, authors introduce a novel, application-agnostic privatization mechanism that leverages representation learning to overcome the prohibitive noise requirements of direct methods. They further ... | SP:bf68937b0e23da990383e48608af226ad5f250cf |
Learning to Noise: Application-Agnostic Data Sharing with Local Differential Privacy | 1 INTRODUCTION . The collection of personal data is ubiquitous , and unavoidable for many in everyday life . While this has undeniably improved the quality and user experience of many products and services , evidence of data misuse and data breaches ( Sweeney , 1997 ; Jolly , 2020 ) have brought the concept of data pri... | In this paper, the authors present a generative-model-based Laplace mechanism. By training the VAE on some dataset, the trained encoder can be used to privatize raw data towards epsilon, delta-LDP. Though the method is novel, the privacy guarantee of the proposed method is not clearly stated and proved. Related experim... | SP:bf68937b0e23da990383e48608af226ad5f250cf |
Efficiently Troubleshooting Image Segmentation Models with Human-In-The-Loop | 1 INTRODUCTION . Image segmentation ( i.e. , pixel-level image labelling ) has recently risen to explosive popularity , due in part to its profound impact on many high-stakes vision applications , such as autonomous driving and medical image analysis . While the performance of segmentation models , as measured by exces... | This work used a variety of existing segmentation algorithms to discover most "controversial" samples from massive online unlabeled images. Those representative controversial samples were believed to have the best chance to confuse the algorithm being trained and to expose its weakness. They are rated by annotators on ... | SP:5fce6baf0f2dcde5b8519a5850484aa627f64f68 |
Efficiently Troubleshooting Image Segmentation Models with Human-In-The-Loop | 1 INTRODUCTION . Image segmentation ( i.e. , pixel-level image labelling ) has recently risen to explosive popularity , due in part to its profound impact on many high-stakes vision applications , such as autonomous driving and medical image analysis . While the performance of segmentation models , as measured by exces... | Annotating images for training of segmentation models is time consuming and it can be difficult to annotate enough examples to ensure good performance on the rare difficult examples that often occur when methods are applied to real world data. These cases are referred to as corner-cases. The paper therefore proposes a ... | SP:5fce6baf0f2dcde5b8519a5850484aa627f64f68 |
Efficiently Troubleshooting Image Segmentation Models with Human-In-The-Loop | 1 INTRODUCTION . Image segmentation ( i.e. , pixel-level image labelling ) has recently risen to explosive popularity , due in part to its profound impact on many high-stakes vision applications , such as autonomous driving and medical image analysis . While the performance of segmentation models , as measured by exces... | In this work, the authors seek to leverage external sources of data to improve the generalization of segmentation models. In particular, they seek to identify images which generate discordance among models, hypothesizing that they would be well-suited to improve model performance. Once selected, they leverage human ann... | SP:5fce6baf0f2dcde5b8519a5850484aa627f64f68 |
Disentangling 3D Prototypical Networks for Few-Shot Concept Learning | 1 INTRODUCTION . Humans can learn new concepts from just one or a few samples . Consider the example in Figure 1 . Assuming there is a person who has no prior knowledge about blue and carrot , by showing this person an image of a blue carrot and telling him “ this is an carrot with blue color ” , the person can easily ... | This paper describes an approach that learns a disentangled shape and style representation of objects in a self-supervised way from RGB-D images. The approach is based on various components, like a 3D feature volume, a bounding box detector, and a disentanglement network. Neural rendering (e.g. recomposing the various... | SP:8f46cbd5fff3557fe870cebf1ba231309fceab14 |
Disentangling 3D Prototypical Networks for Few-Shot Concept Learning | 1 INTRODUCTION . Humans can learn new concepts from just one or a few samples . Consider the example in Figure 1 . Assuming there is a person who has no prior knowledge about blue and carrot , by showing this person an image of a blue carrot and telling him “ this is an carrot with blue color ” , the person can easily ... | The paper presents a framework that combines 1) multi-view prediction for 3D reconstruction from a single image and 2) content-style disentangled representation learning using instance-norm-based auto-encoders. It aims at learning disentangled 3D representation of input images. The authors show that these methods can b... | SP:8f46cbd5fff3557fe870cebf1ba231309fceab14 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.