Paper_ID stringlengths 10 10 | Question stringlengths 201 1.81k | ocr_output stringlengths 252 53.7k |
|---|---|---|
2DJUXmHZ2O | Does the tree architecture described for trajectories for MDPs extend in a straightforward way to multi-agent systems with different information structures. Is it assumed that each agent sees a different tree from its own perspective? | GENERALIZING POINCARÉ POLICY REPRESENTATIONS IN MULTI-AGENT REINFORCEMENT LEARNING
Anonymous authors
Paper under double-blind review
ABSTRACT
Learning policy representations is essential for comprehending the intricacies of agent interactions and their decision-making processes. Recent studies have found that the ev... |
01ep65umEr | How does the GPT explanation help to understand the neural network’s internal decision problem? Deep learning models are known to be distributed representation, meaning that one neuron won’t determine the final decision. How could the proposed method be used to explain the cooperative behavior of the neurons in order t... | TELLME WHAT YOU SEE: USING LLMs TO EXPLAIN NEURONS IN VISION MODELS
Anonymous authors
Paper under double-blind review
ABSTRACT
As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which ar... |
RRKggDJxo2 | * What's the motivation for reservior-in-reservior? Is it similar to boosting methods, or ensembling methods in classical machine learning? Or perhaps similar to multi-head attention where each heads focus on some part of the problem, and then they are aggregated. | Real-time learning of the decay trajectory in Higgs bosons as they interact in the Higgs Field is the key to understanding and furthering of the mass providing mechanism and particle interaction mechanism beyond the [NAME] model in particle physics. We propose a novel machine learning architecture called reservoir-in-r... |
eR4W9tnJoZ | How is this study different from previous efforts studying the use of neuromarketing methods vs. plain advertisements like the ones shown to the participants (plain background with the product in the middle)? Is the novelty of the work the use of ChatGPT to generate the ad creatives? | VISUO-EMOTIONAL PERCEPTION AND HUMAN COGNITION TO ENGINEER CONTENT-GENERATION USING GENERATIVE AI
Anonymous authors
Paper under double-blind review
ABSTRACT
Media platforms compete for users’ attention. Their reach crucially depends on algorithmic real-time bidding and efficiency of hyper-personalised, rapidly gener... |
Rh1aThKliu | My primary concern is that the adversarial attack methods used in this paper appear to be very similar to those of Shin et al. (2020) and Zou et al. (2023). The optimization process, including gradient backpropagation and random position selection, seems to be identical to that of Zou et al. The only difference I disce... | LLM Lies: Hallucinations Are Not Bugs, But Features as Adversarial Examples
Anonymous authors
Paper under double-blind review
Abstract
Large Language Models (LLMs), including GPT-3.5, LLaMA, and PaLM, seem to be knowledgeable and able to adapt to many tasks. However, we still cannot completely trust their answers, s... |
Qvoe4wXWFi | * The introduced generator modules may dilute the energy efficiency brought by the low-voltage scheme. Based on Appendix E, the total computations are very large. A more ideal accuracy-saving method should introduce less overhead. | NeuralFuse: Learning to Recover the Accuracy of Access-Limited Neural Network Inference in Low-Voltage Regimes
Anonymous authors
Paper under double-blind review
Abstract
Deep neural networks (DNNs) have become ubiquitous in machine learning, but their energy consumption remains a notable issue. Lowering the supply v... |
NqQjoncEDR | What makes the difference in the sampling ratio between the selective sampling without mixup and resampling? Is this determined by the hyperparameter of mixup? When changing the lambda for the beta distribution of mixup, are similar results as Fig 15 hold? | SELECTIVE MIXUP HELPS WITH DISTRIBUTION SHIFTS, BUT NOT (ONLY) BECAUSE OF MIXUP
Anonymous authors
Paper under double-blind review
ABSTRACT
Context. [NAME] is a highly successful technique to improve generalization of neural networks by augmenting the training data with combinations of random pairs. Selective mixup i... |
zMvMwNvs4R | Even though the results have proven the effectiveness of ENT, I think the motivation of using error norm to estimate data quality can be further discussed and provide more insights about how do they correlate. | Error Norm Truncation: Robust Training in the Presence of Data Noise for Text Generation Models
[NAME], [NAME], [NAME], [NAME], [NAME]Center for Language and Speech Processing
Johns Hopkins University, Baltimore MD
[EMAIL]
Abstract
Text generation models are notoriously vulnerable to errors in the training data. Wit... |
pHaX00wxFy | What is meant by the idea, repeated many times in the text and summarized in the abstract: “the divergence between the agent’s estimation of the transition probability between the next state given current state-action pairs … in two adjacent trajectory fractions”? | REWARD-FREE EXPLORATION BY CONDITIONAL DIVERGENCE MAXIMIZATION
Anonymous authors
Paper under double-blind review
ABSTRACT
We propose maximum conditional divergence (MaxCondDiv), a new curiosity-driven exploration strategy that encourages the agent to learn in the absence of extrinsic rewards, effectively separating ... |
qWSk62REeK | > XGeoSet, encompassing over 2 million images. You should state the size of the images (on average if they are heterogeneous). 2 million 1024x1024 images is a lot different than 2 million 128x128 images. | MULTISENSORY GEOSPATIAL MODELS VIA CROSS-SENSOR PRETRAINING
Anonymous authors
Paper under double-blind review
ABSTRACT
Geospatial remote sensors, derived from optical and microwave sensors, exhibit significant diversity and provide unique capabilities due to the different observing mechanisms. By integrating multi-s... |
s5hSp7EdL3 | What is the purpose of Section 3.9? The authors say that this section is to compare with EPI-CAL, but only give a formal result for EPI-CAL ($\\Omega(\\sqrt{m})$ samples in Proposition 3.9) and didn't give any formal result for the algorithm they proposed for comparison. | THE HUMAN-AI SUBSTITUTION GAME: ACTIVE LEARNING FROM A STRATEGIC LABELER
[NAME]Machine Learning Department
Carnegie Mellon University
[EMAIL]
[NAME]Department of Computer Science
University of Arizona
[EMAIL]
ABSTRACT
The standard active learning setting assumes a willing labeler, who provides labels on i... |
mHv6wcBb0z | The choice of the optimal value for the hyperparameter α in Equation 6 is crucial. If α is set too large, the model may tend to behave like an identity mapping function, potentially reducing its effectiveness. Conversely, if α is too small, the model may not maintain a | PREVENTING MODEL COLLAPSE IN DEEP CANONICAL CORRELATION ANALYSIS BY NOISE REGULARIZATION
Anonymous authors
Paper under double-blind review
ABSTRACT
Multi-View Representation Learning (MVRL) aims to learn a unified representation of an object from multi-view data. Deep Canonical Correlation Analysis (DCCA) and its va... |
6bAfAcuuZD | Secondly, the loss function requires an explicit positive/negative cue (eta). Historically this sort of global signal has been contentious in computational models. It may be worth emphasizing that this error signal is singular, and how this may be much more easily achieved than a global vector valued error/target signa... | EMERGENCE OF SURPRISE AND PREDICTIVE SIGNALS FROM LOCAL CONTRASTIVE LEARNING
Anonymous authors
Paper under double-blind review
ABSTRACT
Hierarchical predictive models are a popular view of cortical representations, and may hold promise for enhancing the robustness and generalization of machine learning architectures... |
X7gqOBG8ow | In Table 5, the authors demonstrate that the DeNS training is more efficient and results in larger performance gains than increasing max degrees of irreps. I wonder why the authors changed the model from EquiformerV2 to EquiformerV1 for this investigation. After all, the EquiformerV2 model is claimed to largely benefit... | GENERALIZING DENOISING TO NON-EQUILIBRIUM STRUCTURES IMPROVES EQUIVARIANT FORCE FIELDS
Anonymous authors
Paper under double-blind review
ABSTRACT
Understanding the interactions of atoms such as forces in 3D atomistic systems is fundamental to many applications like molecular dynamics and catalyst design. However, si... |
EMCXCTsmSx | seems like we can directly use the code from semantic tokenizer for image retrieval -- similar to how product quantization performs retrieval, what is the necessity of employing the decoder? What is the disadvantange/how much performance degrades if we just use the semantic tokenizer's code for retrieval? | IRGen: Generative Modeling for Image Retrieval
Anonymous authors
Paper under double-blind review
Abstract
While generative modeling has become prevalent across numerous research fields, its potential application to image retrieval has yet to be thoroughly justified. In this paper, we present a novel approach, refram... |
VUA9LSmC2r | - When using GPT-4 plus a simulator to collect the dataset, the location of the target object is directly obtained from the simulator? And this information be stored and used for later training? With this approach, the final complete robotic system still needs a separate vision model besides the ViT-L in the VLM. Can t... | LEARNING EMBODIED VISION-LANGUAGE PROGRAMMING FROM INSTRUCTION, EXPLORATION, AND ENVIRONMENTAL FEEDBACK
Anonymous authors
Paper under double-blind review
Figure 1: Illustration of the functionality of our vision-language programmer, Octopus, in the developed OctoGTA environment. Given a task in the form of natural la... |
ikX6D1oM1c | The experiments, to some extent, appear to be lacking in comprehensiveness. The semi-synthetic datasets utilized in this study are exclusively derived from the MIMIC-III dataset. The findings derived from this single source may not offer an adequate illustration of the model's performance. | A NEURAL FRAMEWORK FOR GENERALIZED CAUSAL SENSITIVITY ANALYSIS
Dennis Frauen\textsuperscript{1, 2, 6} Fergus Imrie\textsuperscript{3} Alicia Curth\textsuperscript{4} [NAME], 2}
[NAME], 2} [NAME], 5}
ABSTRACT
Unobserved confounding is common in many applications, making causal inference from observational data chall... |
6LyO8WTVTU | Intuitively, if you have a perfect teacher model, you can use it directly to calculate graph embeddings. Is it necessary to design such complicated contrastive learning losses to distill from the teacher model? | A Teacher-Guided Framework for Graph Representation Learning
Anonymous authors
Paper under double-blind review
Abstract
We consider the problem of unsupervised representation learning for Graph Neural Networks (GNNs). Several state-of-the-art approaches to this problem are based on Contrastive Learning (CL) principl... |
bFMpmb8p3D | Why did the author start with the inverse image restoration (motion blurring) instead of the image restoration task (deblurring)? Blurring an image seems very easy, a simple convolution with a blur kernel could achieve similar film-blurring effect. | MULTI-TASK IMAGE-TO-IMAGE DIFFUSION MODELS WITH FINE-GRAINED CONTROL
Anonymous authors
Paper under double-blind review
ABSTRACT
Diffusion models have recently been applied to various image restoration and editing tasks, showing remarkable results in commercial products, e.g., [NAME]. While recent approaches to text-... |
5xadJmgwix | The authors mentioned that vector-base approaches are inherently limited when tackling intricate and complex sketches. I would like to understand why pixel-based methods have an advantage over vector-based methods? I hope the authors can provide some intuitive explanations. | SCALE-ADAPTIVE DIFFUSION MODEL FOR COMPLEX SKETCH SYNTHESIS
Jijin Hu\textsuperscript{1} [NAME]*} [NAME]} [NAME] Song\textsuperscript{2}
\textsuperscript{1}Beijing University of Posts and Telecommunications, CN \textsuperscript{2}SketchX, CVSSP, University of Surrey, UK
[EMAIL] [EMAIL]
ABSTRACT
While diffusion models... |
I0gwsdSgsk | The equation at the end of section 3 describing CMAB-AND is rather dense/difficult to read, even given an understanding of each of the components. An explanation of the update mechanism aside from the equation might help clarify. | Memory Efficient Neural Processes via Constant Memory Attention Block
Anonymous authors
Paper under double-blind review
Abstract
Neural Processes (NPs) are popular meta-learning methods for efficiently modelling predictive uncertainty. Recent state-of-the-art methods, however, leverage expensive attention mechanisms... |
f4HohsyNEk | SDF-based reconstruction methods (e.g. NeuS) claim that the raw density field is optimized only to satisfy the rendering objective (and is also biased away from the true surface), and therefore are not ideal for surface reconstruction. In this paper, you argue that this is unnecessary and that using the density field d... | NeuManifold: Neural Watertight Manifold Reconstruction with Efficient and High-Quality Rendering Support
Anonymous authors
Paper under double-blind review
Abstract
We present a method for generating high-quality watertight manifold meshes from multi-view input images. Existing volumetric rendering methods are robust... |
OROKjdAfjs | In terms of evaluation, although in the abstract, the authors claim that the linearized LLM extends to 175B parameters, most experiments are conducted on 375M models. For the large parameter size settings, the author only reports the memory and latency cost savings. The accuracy information is missing, without which I ... | TransNormerLLM: A Faster and Better Large Language Model with Improved TransNormer
Anonymous authors
Paper under double-blind review
Abstract
We present TransNormerLLM, the first linear attention-based [NAME] (LLM) that outperforms conventional softmax attention-based models in terms of both accuracy and efficiency.... |
gyJpajLkX2 | Additionally on the upper bound, the authors state that they choose to adopt the sampled vCLUB and minimize I_{vCLUB-S}(Xo; Zi). Could they expand on why this sampling strategy was chosen and what the actual derived upper bound might be? | Enhancing Multivariate Time Series Forecasting with Mutual Information-driven Cross-variable and Temporal Modeling
Anonymous authors
Paper under double-blind review
Abstract
Recent researches have showcased the significant effectiveness of deep learning techniques for multivariate time series forecasting (MTSF). Bro... |
FItPCl4uEc | However, in the general feature-based knowledge distillation framework [1], both teacher and student features can be transformed before minimizing their distances. This makes the proposed method a simple variant in the feature-based knowledge distillation framework and thus lack novelty. | [NAME] from Arbitrary Pre-Trained Models
Anonymous authors
Paper under double-blind review
Abstract
Transfer learning typically involves loading pre-trained weights as an initialization, followed by fine-tuning on a downstream task. As pre-trained models become ever larger, this procedure is becoming prohibitively e... |
MQ4JJIYKkh | Scaled-up, however, both the reward model and the approximate planning model would need to be far more complex, to the point where we would not expect (an approximation of) Bayesian IRL to be any more sample efficient that behavioral cloning with a similarly complex policy model | CONCEPT ALIGNMENT AS A PREREQUISITE FOR VALUE ALIGNMENT
Anonymous authors
Paper under double-blind review
ABSTRACT
Value alignment is essential for building AI systems that can safely and reliably interact with people. However, what a person values—and is even capable of valuing—depends on the concepts that they are... |
ZULjcYLWKe | In the experiments, results show DMBP significantly improve the robustness performance of existing offline RL methods. The review noticed that most offline RL methods used are value-based. How will DMBP perform when used with weighted imitation learning methods, e.g., IQL? | DMBP: Diffusion Model-Based Predictor for Robust Offline Reinforcement Learning Against State Observation Perturbations
Zhihe Yang\textsuperscript{1,2} [NAME]} *
\textsuperscript{1}The Chinese University of Hong Kong, Hong Kong SAR, China
\textsuperscript{2}The Chinese University of Hong Kong, Shenzhen Research Instit... |
EYtga9mSdT | - Self-supervised pretraining has proven to be extremely helpful for few-shot learning [2, 3]. As most of the other methods compared in this paper do not use self-supervised pretraining, the comparisons are not fair. | BALANCED LEARNING WITH TOKEN SELECTION FOR FEW-SHOT CLASSIFICATION
Anonymous authors
Paper under double-blind review
ABSTRACT
In recent years, patch-based approaches have shown promise in few-shot learning, with further improvements observed through the use of self-supervised learning. However, we observe that the m... |
AN5uo4ByWH | You claim Euclidean Transformers are inadequate for modeling hierarchical and cyclic graphs. However, recent works like Graphormer show strong performance on tasks like molecular property prediction that involve such structures. Can you provide more concrete evidence on the limitations of existing methods? Comparisons ... | Curve Your Attention:
Mixed-Curvature Transformers
for Graph Representation Learning
Anonymous authors
Paper under double-blind review
Abstract
Real-world graphs naturally exhibit hierarchical trees and cyclic structures that are unfit for the typical Euclidean space. While there exist graph neural networks that uti... |
RsztjXcvUf | It seems to me that equation (14) is one of the crucial steps: it avoids Lemma 1 in Yang et al, where the gap is bounded as a function of the distance to the optimal VI solution. However, I struggle to see where/how exactly the improvements were achieved. | A Primal-Dual Approach to Solving Variational Inequalities with General Constraints
Tatjana Chavdarova∗
University of California, Berkeley
[EMAIL]
Matteo Pagliardini
University of California, Berkeley & EPFL
[EMAIL]
[NAME]Carnegie Mellon University
[EMAIL]
[NAME]
University of California, Berkeley
[EMAIL]
Abstract... |
SerYSFntLh | In the sensitivity analysis, the model performance has ups and downs instead of rising and then stopping or no obvious change as described in the paper. How to explain this phenomenon? In addition, when taking two extreme values (0.2 and 1.0), the performance difference on some datasets(such as lambda in book-domian) i... | Variational Disentangled Cross-domain Knowledge Alignment for Multimodal Recommendation
Anonymous authors
Paper under double-blind review
Abstract
Multimodal recommendation systems have been widely used in e-commerce and short video platforms. Due to the large differences in data volume and data distribution in diff... |
yLClGs770I | I'm curious about whether the generated annotation converted from TheoremQA can be successfully executed, because the problems there would require many advanced calculation like integral and derivative computation. | MAmmoTH: Building Math [NAME] through Hybrid Instruction Tuning
*[NAME], †Xingwei [NAME], [NAME], *[NAME], §[NAME],
*[NAME], *[NAME], †Wenhu [NAME]*
†University of [NAME], *The Ohio State University, †HKUST, °University of Edinburgh, §01.AI
[EMAIL], [EMAIL]
Abstract
We introduce MAmmoTH, a series of open-source larg... |
djcciHhCrt | Additionally, the authors claim that the attack is stealthy because it does not alter the semantic meanings of the output answers. They support this claim with the evidence that the answers under attack are 10% less natural compared to the original ones. My question is, why is a 10% difference considered a small one? | MISUSING TOOLS IN LARGE LANGUAGE MODELS WITH VISUAL ADVERSARIAL EXAMPLES
Anonymous authors
Paper under double-blind review
ABSTRACT
Large Language Models (LLMs) are being enhanced with the ability to use tools and to process multiple modalities. These new capabilities bring new benefits and also new security risks. ... |
ekdurSMmbH | * Why do the authors believe clustering improves performance? If the initial state is unique to participants wouldn't it be possible to learn a single policy that performs well across all states? Why do they think this doesn't happen? Is the policy class being trained on offline data not rich enough? | Universal Off-Policy Selection for Human-Centric Systems via Participant Sub-grouping
Anonymous authors
Paper under double-blind review
Abstract
Human-centric tasks like healthcare and education are characterized by heterogeneity among patients and students, resulting in different disease trajectories and learning s... |
oDYXpvnv5f | In networks with normalization layers, e.g. BatchNorm, the network would become scale-invariance, then increasing the weight in intermediate layers would not change the output. The proposed method could therefore be less effective. | DEEP ANTI-REGULARIZED ENSEMBLES
Anonymous authors
Paper under double-blind review
ABSTRACT
We consider the problem of uncertainty quantification in high dimensional regression and classification, for which deep ensemble have proven to be promising methods. Recent observations have shown that deep ensemble return ove... |
jTsnuWsda2 | What can I2T-Bridge specifically correct when it is used to correct textual instructions? If the images generated by the text-to-image model are not good enough, can the images be further corrected based on the corrected text? | MULTIMODAL PROCEDURAL PLANNING VIA DUAL TEXT-IMAGE PROMPTING
Anonymous authors
Paper under double-blind review
Figure 1: Our dual Text-Image Prompting (TIP) model generates coherent and authentic multimodal procedural plans towards a high-level goal, providing useful guidelines in task completion.
ABSTRACT
Embodied... |
rUx0zQFwD1 | Thm. 2.3: it would be better to explain the difference between the consistent version and the usual version of amplitude estimation before stating the theorem. $f$ is not defined. Is it some fixed function? So, when $p$ and $s$ are fixed, with high probability, the algorithm's output will be the same? | Quantum Speedups in Linear Programming via Sublinear Multi-Gibbs Sampling
Anonymous authors
Paper under double-blind review
Abstract
As a basic optimization technique, linear programming has found wide applications in many areas. In this paper, we propose an improved quantum algorithm for solving a linear programmin... |
uJPWeZffgl | As a follow-up question: why should we be interested in developing this equivalent formulation as a bilevel problem? What could one possibly gain that improves upon the exact inference proposed in semantic loss, deepproblog, NeSy entropy and semantic probabilistic layers? | CONVEX AND BILEVEL OPTIMIZATION FOR NEURO-SYMBOLIC INFERENCE AND LEARNING
Anonymous authors
Paper under double-blind review
ABSTRACT
We address a key challenge for neuro-symbolic (NeSy) systems by leveraging convex and bilevel optimization techniques to develop a general gradient-based framework for end-to-end neura... |
xUe1YqEgd6 | However, if I understand it right, the temporal consistency in equation (7) remains constrained to neighboring consecutive frames, similar to ST-MS. The primary distinction lies in using B-spline to model long-term motion changes and a different network architecture. | UNSUPERVISED MOTION SEGMENTATION IN ONE GO:
SMOOTH LONG-TERM MODEL OVER A VIDEO
Anonymous authors
Paper under double-blind review
ABSTRACT
Human beings have the ability to continuously analyze a video and immediately extract the main motion components. Motion segmentation methods often proceed frame by frame. We wan... |
tzlGWqRA0T | For example, while “contextual signals” is mentioned and later modeled, the reader can’t know what that means with respect to this experiment without an initial explanation of the experimental setup and task(s) | SNN-LPCG: Spiking Neural Networks with Local Plasticity Context Gating for Lifelong Learning
Anonymous authors
Paper under double-blind review
Abstract
Humans learn multiple tasks in succession with minimal mutual interference, through the context gating mechanism in the prefrontal cortex (PFC). The brain-inspired m... |
PbpJnyewVM | It seems that the total number of trajectories is 4, and the authors use a K-means clustering to separate them further. I am wondering about the necessity of doing this. Are you labeling the preference among clusters or within clusters? | ZERO-SHOT CROSS-TASK PREFERENCE ALIGNMENT FOR OFFLINE RL VIA OPTIMAL TRANSPORT
Anonymous authors
Paper under double-blind review
ABSTRACT
In preference-based Reinforcement Learning (PbRL), aligning rewards with human intentions often necessitates a substantial volume of human-provided labels. Furthermore, the expens... |
TfbpnxTJt3 | If the server does not aggregate the DP private label feedback by users, but only uses the DP private label of any individual user to perform the steps in line 9 of Algorithm 1, can the server calculate the real local label distribution of the user? Does this imply a leak of privacy? | Federated Learning with Local OpenSet Noisy Labels
Anonymous authors
Paper under double-blind review
Abstract
Federated learning is a learning paradigm that allows the central server to learn from different data sources while keeping the data private locally. Without controlling and monitoring the local data collect... |
KncRpAnprQ | A simple test to verify the obfuscating gradient behaviour is to measure whether the perturbation, found based on the attack methods in the paper, indeed reaches the specified radius of the $\ell_p$ ball. | A NOVEL APPROACH FOR ADVERSARIAL ROBUSTNESS
Anonymous authors
Paper under double-blind review
ABSTRACT
Deep learning has made tremendous progress in the last decades; however, it is not robust to adversarial attacks. To deal with this issue, perhaps the most effective approach is adversarial training at a high compu... |
Q00CO1Tm6M | As in previous work, the OSI in the hindsight is common some applications, such as data center scheduling, where the state is revealed after the action is take (so the effect of the action can be observed), However, the query model in this paper allows the agent to query part of the state before the action is taken. Ev... | THEORETICAL HARDNESS AND TRACTABILITY OF POMDPs IN RL WITH PARTIAL ONLINE STATE INFORMATION
Anonymous authors
Paper under double-blind review
ABSTRACT
Partially observable [NAME] decision processes (POMDPs) have been widely applied to capture many real-world applications. However, existing theoretical results have s... |
MCUvAc1GTg | While you state DeepWalk to be one of your baseline models, it is absent in Table 3 and only used for subgraph matching in Figure 3. Could you please explain why you choose not to use it for graph matching? | Network Alignment with Transferable Graph Autoencoders
Anonymous authors
Paper under double-blind review
Abstract
Network alignment is the task of establishing one-to-one correspondences between the nodes of different graphs and finds a plethora of applications in high-impact domains. However, this task is known to ... |
NgaLU2fP5D | The assessment of the model's interpretability is not entirely convincing. The limited dimensionality of hidden learner representations in deep learning methods (e.g., DKT, AKT) at just 16 may constrain the neural networks' capabilities. Furthermore, there is no supporting evidence indicating that the learner represent... | PREDICTIVE, SCALABLE AND INTERPRETABLE KNOWLEDGE TRACING ON STRUCTURED DOMAINS
[NAME]}, [NAME], [NAME]}, & [NAME]}
\textsuperscript{1}University of Tübingen, \textsuperscript{2}Cluster of Excellence Machine Learning, \textsuperscript{3}Tübingen AI Center, \textsuperscript{4}IMPRS-IS
\[EMAIL]
ABSTRACT
Intelligent tut... |
uqxBTcWRnj | How can $theta$ be optimized in Eq. 1 if it does not appear in the two terms? The definition of the decoder g(·) is not consistent. Does it take $theta$ as a condition or not? Seemingly Eq. 2 is the appropriate form. | Bridging Neural and Symbolic Representations with Transitional Dictionary Learning
[NAME]
Thayer School of Engineering
Dartmouth College
Hanover, NH 03755, USA
[EMAIL]
[NAME] of Engineering
Dartmouth College
Hanover, NH 03755, USA
[EMAIL]
Abstract
This paper introduces a novel Transitional Dictionary ... |
vfzRRjumpX | About not using random token replacement, this makes a lot of sense to me. But we would probably have the same problem for natural language, if we consider multiple languages, right? Is the same practice common in training multi-lingual text embeddings? | CODE REPRESENTATION LEARNING AT SCALE
[NAME]* & Wasi [NAME]
[EMAIL]
Ming Tan & Hantian Ding
[EMAIL]
Ramesh Nallapati & Dan Roth & [NAME]
[EMAIL]
AWS [NAME]
ABSTRACT
Recent studies have shown that code language models at scale demonstrate significant performance gains on downstream tasks, i.e., code generation. Ho... |
LH2JNpfwdH | Also, in the METHODOLOGY section, the authors are repeatedly mentioning that many components of the framework are motived by previous methods. How could authors persuade reviewers the proposed the framework is innocative instead of an incremental work with the combination of previous works? | Towards 4D Human Video Stylization
Anonymous authors
Paper under double-blind review
Abstract
We present a first step towards 4D (3D and time) human video stylization, which addresses style transfer, novel view synthesis and human animation within a unified framework. While numerous video stylization methods have be... |
gIiz7tBtYZ | Another point is that the cost function proposed in Prop.1 contains a quadruple sum over samples: a discussion about its variance would be welcome. Also, the proposition mentions that it is an estimator: in which sense? | Neural Optimal Transport with General Cost Functionals
Arip Asadulaev∗1,3 Alexander Korotin∗2,1 Vage Egiazarian4,5 Petr Mokrov2 [NAME]
1 Artificial Intelligence Research Institute 2 Skolkovo Institute of Science and Technology
3 Moscow Institute of Physics and Technology 4 HSE University 5 Yandex
[EMAIL],[EMAIL]
Abst... |
duyA42HlCK | The paper's objective is to address incoherent parts and unnatural poses. However, it falls short in terms of providing quantitative metrics to evaluate the effectiveness of the proposed method in addressing these issues. | **HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion**
Xian Liu\(^1,2\)*, [NAME] Ren\(^1\)† Aliaksandr Siarohin\(^1\) [NAME]) Yanyu Li\(^1\) [NAME]) [NAME]) [NAME]) Sergey Tulyakov\(^1\)
\(^1\)Snap Inc. \(^2\)CUHK \(^3\)HKU \(^4\)NTU
Project Page: https://snap-research.github.io/HyperHuman... |
YNmnGzttMJ | Q3: Tables 1 + 2: There are several instances where the minimal data ratio for higher IPC is lower. Why does this happen, and why for these specific examples (e.g. CAFÉ for CIFAR10 + MNIST, or MTT for CIFAR 10)? | DISTILL GOLD FROM MASSIVE ORES: EFFICIENT DATASET DISTILLATION VIA CRITICAL SAMPLES SELECTION
Anonymous authors
Paper under double-blind review
ABSTRACT
Data-efficient learning has drawn significant attention, especially given the current trend of large multi-modal models, where dataset distillation can be an effect... |
eOCvA8iwXH | With respect to the work of Miyato et al., (2022) that considers the properties of the learned model to be intra-orbital homogeneity or full equivariance, which of these properties that the proposed model owns in each setting (U-NFT, G-NFT, g-NFT) ? | Neural Fourier Transform: A General Approach to Equivariant Representation Learning
Masanori Koyama\textsuperscript{1} Kenji Fukumizu\textsuperscript{2,1} Kohei Hayashi\textsuperscript{1} Takeru Miyato\textsuperscript{3,1}
\textsuperscript{1}Preferred Networks, Inc. \textsuperscript{2}The Institute of Statistical Math... |
dUDwK38MVC | - When comparing FVDs on UCF-101, the paper included many previous works — but only if their FVD score is higher: https://paperswithcode.com/sota/video-generation-on-ucf-101. In this way, it omitted Make-A-Video, VDIM, LVDM, etc. despite comparing it in other regards. I do not see how it could be justified. | VIDEOFACTORY: SWAP ATTENTION IN SPATIOTEMPORAL DIFFUSIONS FOR TEXT-TO-VIDEO GENERATION
Anonymous authors
Paper under double-blind review
ABSTRACT
We present VideoFactory, an innovative framework for generating high-quality open-domain videos. VideoFactory excels in producing high-definition (1376×768), widescreen (1... |
AN5uo4ByWH | There are no significant theoretical analysis that would provide insights on specific features of FPS-T: e.g when graphs whose sectional curvature distribution has a large variance what would be a good constant curvature space (or product of spaces) to discriminate them ? | Curve Your Attention:
Mixed-Curvature Transformers
for Graph Representation Learning
Anonymous authors
Paper under double-blind review
Abstract
Real-world graphs naturally exhibit hierarchical trees and cyclic structures that are unfit for the typical Euclidean space. While there exist graph neural networks that uti... |
HSESApr9r7 | The so called ``period drift'' comes from the stochastic sampling of clients. If we see sampling clients as sampling data in SGD, such a period drift also happens during SGD -- each batch of data has distinct data distribution from other batches. Authors should provide a more rigorous definition of period drift and sho... | FedEve: On Bridging the Client Drift and Period Drift for Cross-device Federated Learning
Anonymous authors
Paper under double-blind review
Abstract
Federated learning (FL) is a machine learning paradigm that allows multiple clients to collaboratively train a shared model without exposing their private data. Data he... |
X4ATu1huMJ | This paper only investigates the offline model selection, which may have limited insights to address the challenge of online model selection for TTA methods, particularly in light of the batch dependency analyzed in [3]. | Realistic Evaluation of Test-Time Adaptation: Surrogate-Based Model Selection Strategies
Anonymous authors
Paper under double-blind review
Abstract
Test-Time Adaptation (TTA) has recently emerged as a promising strategy for tackling the problem of machine learning model robustness under distribution shifts. This set... |
TjGJFkU3xL | I understand that $q_0$ represents the maximum value for the equation in Lemma 5.1. However, I am unsure about the connection between Lemma 5.1 and Equation (8). In other words, why do we need to minimize the empirical quantity for the equation in Lemma 5.1? | Doubly Robust Proximal Causal Learning for Continuous Treatments
Yong Wu\textsuperscript{1,3,4,5} Yanwei Fu\textsuperscript{2} Shouyan Wang\textsuperscript{1,3,4,5,6,7} [NAME]*}
\textsuperscript{1}Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
\textsuperscript{2}School of Data S... |
cSSHiLnjsJ | * Figure 1 and Figure 4 suggest that work particles travel along the path determined by residual updates, but such a description is very general. Are there more specific properties within the residual updates? | Traveling Words: A Geometric Interpretation of Transformers
Anonymous authors
Paper under double-blind review
Abstract
Transformers have significantly advanced the field of natural language processing, but comprehending their internal mechanisms remains a challenge. In this paper, we introduce a novel geometric pers... |
UVb0g26xyH | CMOW is a simple sentence embedding method that represents every word as a matrix and composes words in a sentence via matrix multiplication. Hence, CMOW builds a linear deep neural network dynamically. Could you comment on in how far this relates to your idea in the conclusion to construct novel NLP models by embeddin... | VOCABULARY FOR UNIVERSAL APPROXIMATION:
A LINGUISTIC PERSPECTIVE OF MAPPING COMPOSITIONS
Anonymous authors
Paper under double-blind review
ABSTRACT
In recent years, deep learning-based sequence modelings, such as language models, have received much attention and success, which pushes researchers to explore the possi... |
gYcft1HIaU | Would it be considered a correct hit if the model predicts 'GI tract' instead of 'digestive system' in the examples from Table 3? What kind of standardization was performed in evaluating LLMs response with the experts output? | DO CURRENT LARGE LANGUAGE MODELS MASTER ADEQUATE CLINICAL KNOWLEDGE?
Anonymous authors
Paper under double-blind review
ABSTRACT
Large Language Models (LLMs) show promising potential in solving clinical problems. Current LLMs, including so-called medical LLMs, are reported to achieve excellent performance on certain ... |
9g8h5HwZMy | Incorporating a well-designed prior distribution could also reduce the challenging learning task, e.g., [EigenFold](https://arxiv.org/abs/2304.02198). The model should be able to figure out the correlation between data dimensions during training. Would this be more effective and easier to learn than introducing masks f... | [NAME] for 3D Molecular Representation Learning: Combining Continuous and Discrete
Anonymous authors
Paper under double-blind review
Abstract
Molecular representation learning has shown great success in AI-based drug discovery. The 3D geometric structure contains crucial information about the underlying energy funct... |
Mhb5fpA1T0 | I am concerned by the reported results for the BC baseline. Due to action data being available, as well as the robot data being in-domain for the task a simple BC or kNN baseline should work very well. There are many cases where the results are < 5% success. This should be addressed. | Learning to Act from Actionless Videos through Dense Correspondences
Po-Chen Ko†
National Taiwan University
[NAME]
[NAME]
Shao-Hua Sun
National Taiwan University
[NAME]
MIT BCS, CBMM, CSAIL
Abstract
In this work, we present an approach to construct a video-based robot policy capable of reliably executing diverse... |
xbXASfz8MD | If I understood correctly, the proposed method works for compact groups only. The experiments demonstrate that the method can learn trajectories that are isomorphic to circles. How will the method behave on the data which has translation symmetry only? Will it fail? If so, the set of admissible symmetries seems more li... | [NAME] authors
Paper under double-blind review
ABSTRACT
Equivariant neural networks require explicit knowledge of the symmetry group. Automatic symmetry discovery methods aim to relax this constraint and learn invariance and equivariance from data. However, existing symmetry discovery methods are limited to linear sy... |
AxYTFpdlvj | The GRPDG method requires knowing in advance the number of positive/negative eigenvalues, so it's very important to know how the authors addressed this. Is it always the same? Is it chosen for each dataset? How? | [NAME] via Generalized Random Dot Product Graph
Anonymous authors
Paper under double-blind review
Abstract
Graph Neural Networks (GNNs) have established themselves as the state-of-the-art methodology for a multitude of graph-related tasks, including but not limited to link prediction, node clustering, and classifica... |
W3T9rql5eo | It is interesting to model the preference-to-objective mapping to characterize the PF in a more direct way, but I wonder how can we generate certain Pareto solution given a specific preference from the learned Pareto front. It seems that the proposed model does not explicitly involve the solutions in the decision space... | Uniform as [NAME]: Gliding over the Pareto Front with Neural Adaptive Preferences
Anonymous authors
Paper under double-blind review
Abstract
Multiobjective optimization (MOO) is prevalent in numerous real-world applications, in which a Pareto front (PF) is constructed to display optima under various preferences. Pre... |
bA5o5eZplk | At the initial stages, the egonet dissimilarity of anomalies significantly outweighs that of normal instances as illustrated in Fig.2. Therefore, it's worth considering if exploiting this difference directly could lead to more efficient detection. | NEW RECIPES FOR GRAPH ANOMALY DETECTION: FORWARD DIFFUSION DYNAMICS AND GRAPH GENERATION
Anonymous authors
Paper under double-blind review
ABSTRACT
Distinguishing atypical nodes in a graph, which is known as graph anomaly detection, is more crucial than the generic node classification in real applications, such as f... |
lCLdLlXAvt | The average sensitivity of SHC is $O(\lambda D n)$, which can be made small by setting $\lambda \ll 1$. In the experiments, this is not always the case, for instance in Section 8. This signifies that SHC in practice doesn't really have a low average sensitivity. | Average Sensitivity of Hierarchical Clustering
Anonymous authors
Paper under double-blind review
Abstract
Hierarchical clustering is one of the most popular methods used to extract cluster structures in a dataset. However, if the hierarchical clustering algorithm is sensitive to a small perturbation to the dataset, ... |
libLqoInAd | I'm a bit confused as to why the n-dim DS classifier handles epistemic uncertainty better than the softmax classifier, especially as demonstrated in Figure 3. As noted in the text, the n-dim classifier loses the ability to distinguish between aleatoric and epistemic uncertainty (since uncertainty is only able to be mea... | Reliable Classifications with Guaranteed Confidence using the Dempster-Shafer Theory of Evidence
Anonymous authors
Paper under double-blind review
Abstract
Reliably capturing predictive uncertainty is indispensable for the deployment of machine learning (ML) models in safety-critical domains. The most commonly used ... |
xUe1YqEgd6 | What exactly is the motion model. It seems to be some kind of linear decomposition of the optical flow. is that correct? If so please clarify and denote all notation in Eq(2), for instance what is $n$? | UNSUPERVISED MOTION SEGMENTATION IN ONE GO:
SMOOTH LONG-TERM MODEL OVER A VIDEO
Anonymous authors
Paper under double-blind review
ABSTRACT
Human beings have the ability to continuously analyze a video and immediately extract the main motion components. Motion segmentation methods often proceed frame by frame. We wan... |
9XdLlbxZCC | The proposed method performs on par or slightly worse against other flow estimation methods on Sintel benchmark and Kitti, which can not demonstrate the benefit of the proposed ''Joint-Embedding Predictive Architecture''. The authors don't give convincing analysis for this issue. | MC-JEPA: A JOINT-EMBEDDING PREDICTIVE ARCHITECTURE FOR SELF-SUPERVISED LEARNING OF MOTION AND CONTENT FEATURES
Anonymous authors
Paper under double-blind review
ABSTRACT
Self-supervised learning of visual representations has been focusing on learning content features, which do not capture object motion or location, ... |
K7l94Z81bH | Can you clarify the data sparsity issue, why 3000 drivers with 13000 orders a day is considered sparse. And what are the effect on the learning algorithm? The performance of other DRL algorithms are not too bad, which is not what I would expect for a data sparse environment. | Sparsity-Aware Grouped Reinforcement Learning for Designated Driver Dispatch
Anonymous authors
Paper under double-blind review
Abstract
Designated driving service is a fast-growing market that provides drivers to transport customers in their own cars. The main technical challenge in this business is the design of dr... |
LegZeFYugN | What are the author’s intuitions for why Time2Image underperforms the regular ResNet? Is there something about the Gaussian representation that is particularly suited to Transformers over Convulational approaches? | TIME2IMAGE: A UNIFIED ADAPTIVE IMAGE REPRESENTATION FRAMEWORK FOR TIME SERIES CLASSIFICATION
Anonymous authors
Paper under double-blind review
ABSTRACT
Time Series Classification (TSC) is a crucial and challenging task that holds significant importance across various domains, of which one of the kernel ingredients i... |
4aJg9e4nvF | In table2, multiple ViT models and CNN models are compared to show that the ViTs are better at using background information to predict correct classes. The issue here is that the used ViTs are more powerful than the CNNs with more parameters and more computations. ViTs have consistently better classification accuracies... | What do Vision Transformers Learn? A Visual Exploration
Anonymous authors
Paper under double-blind review
Abstract
Vision transformers (ViTs) are quickly becoming the de-facto architecture for computer vision, yet we understand very little about why they work and what they learn. While existing studies visually anal... |
YKIGyf215Q | - How do you assess the cost/benefit trade-off of (a) requiring parseable code and (b) increasing the length of the program's representation in terms of tokens given that the most common use-case of LLMs for code is code generation, often in incomplete contexts. | Structured Fine-Tuning Enables Data-Efficient Adaptation of Code Language Models
Anonymous authors
Paper under double-blind review
Abstract
Current models tailored for code tasks often adopt the successful pre-training-then-fine-tuning paradigm from natural language processing, treating source code in plain text as ... |
I5MquO1g7R | In the second sentence, it is also mention that $Q$ can be modeled using non parametric techniques, but the following sentence (and the presented algorithm) only deal with a parametric distribution (with parameter $\Phi$). It might be good to mention how the algorithm is changed when a non-parametric density estimation... | CHANGE POINT DETECTION VIA VARIATIONAL TIME-VARYING HIDDEN MARKOV MODEL
Anonymous authors
Paper under double-blind review
ABSTRACT
The task of modeling time series data that exhibit sudden regime shifts has been an enduring focus of research due to its inherent complexity. Among the various strategies to tackle this... |
X7nz6ljg9Y | End of 3.1: “allowing us to reject the hypothesis that the labeling functions are drawn uniformly at random with extremely high confidence.” Who would make the claim that labeling functions for real-world datasets are drawn uniformly at random? | The No Free Lunch Theorem, Kolmogorov Complexity, and the Role of Inductive Biases in Machine Learning
Anonymous authors
Paper under double-blind review
Abstract
No free lunch theorems for supervised learning state that no learner can solve all problems or that all learners achieve exactly the same accuracy on avera... |
JGTC6WgO4T | Furthermore, when the cardinality of 'Ic' is 'M', there is another possible scenario where all source models have identical predictions. The paper does not explicitly mention this, which is a lack of rigor. | [NAME] PSEUDO LABEL REFINEMENT FOR MULTI-SOURCE BLACK-BOX DOMAIN ADAPTATION
Anonymous authors
Paper under double-blind review
ABSTRACT
Unsupervised Domain Adaptation (UDA) aims to train a model for an unlabeled target domain by transferring knowledge from a source domain. However, standard UDA requires access to sou... |
vLJg4wgBPu | - I was able to verify that GPT-4 correctly outputs a trace for the example shown in Prompt 1. However, the same prompt on a slightly larger input produces a correct answer but starts to include comments like one might see in code. Is this expected? | GPT IS BECOMING A TURING MACHINE:
HERE ARE SOME WAYS TO PROGRAM IT
Anonymous authors
Paper under double-blind review
ABSTRACT
We demonstrate that, through appropriate prompting, GPT-3 can be triggered to perform iterative behaviours necessary to execute (rather than just write or recall) programs that involve loops,... |
hujS6bmduD | The architecture of the tagging adapter is not well motivated. The TextEnc and ImageEnc in the the tagging adapter as described in Eq. 6 should be explained in more detail, including details of it model architecture. How does the TextEnc and ImageEnc relate to the cross attention modules in Figure 2. | HARNESSING TEXT-TO-IMAGE DIFFUSION FOR DENSE PREDICTION TASKS
Anonymous authors
Paper under double-blind review
ABSTRACT
Equipped with large-scale training data, text-to-image diffusion models have demonstrated the capacity to generate high-quality images that semantically correspond to the given textual description... |
QXRScRrwNr | I wonder how to fairly compare ATLAS and LLM since they may not use the same datasets. (1) Saying ATLAS uses data A to perform pretraining and data B as a corpus, and LLM uses data C for pretraining, does A+B=C? If not, how to fairly compare them? | In-Context Learning with Retrieval Augmented Encoder-Decoder Language Models
Anonymous authors
Paper under double-blind review
Abstract
In this paper, we investigate the in-context learning ability of retrieval-augmented encoder-decoder language models. We first conduct a comprehensive analysis of the state-of-the-a... |
HEcbGXzIHK | * Could you please discuss in detail the difficulties of applying the proposed theory to nonlinear RNNs? In Appendix A.4, you have shown that a nonlinear RNN has a similar form of the linear system as linear RNNs instead of a different $W_{hh}$. | Episodic Memory Theory for the Mechanistic Interpretation of Recurrent Neural Networks
Anonymous authors
Paper under double-blind review
Abstract
Understanding the intricate operations of Recurrent Neural Networks (RNNs) mechanistically is pivotal for advancing their capabilities and applications. In this pursuit, w... |
QKqWnNkwPL | I am not sure that this statement in introduction (second paragraph) about the current distillation techniques is true — “The student is trained to mimic the teacher, but in fewer iterations, so that, eventually the teacher and the student diverge during training.” In my experience with distillation techniques, I haven... | SELF-DISTILLATION FOR DIFFUSION
Anonymous authors
Paper under double-blind review
ABSTRACT
In recent years, diffusion models have demonstrated powerful generative capabilities. As they continue to grow in both ability and complexity, performance optimization becomes more relevant. Knowledge Distillation (KD), where ... |
Bo6GpQ3B9a | When n=0, no out-of-domain samples are utilized and the problem reduces to simple ERM. But in Theorem 4.2, when n=0, the dependence of the error on dimension is d^{3/8}, meaning that this reduction in the exponent of the dimension is not related to the utilization of out-of-domain samples. | OUT-OF-DOMAIN UNLABELED DATA IMPROVES GENERALIZATION
[NAME] †∗ [NAME] † [NAME] †
[NAME] † [NAME] Motahari † [NAME]
∗ Department of Electrical Engineering,
† Department of Computer Engineering,
‡ Sharif Center for Information Systems and Data Science.
§ Sharif Institute for Convergence Science & Technology,
Sharif Univ... |
RpKA1wqgk0 | The paper does not demonstrate any efficiency gains from the proposed attention design. This is unconvincing, as one of the primary motivations described in the introduction is to reduce the computational cost of attention in ViTs when adapting for Meta-learning. | MetaFormer with Holistic Attention Modelling Improves Few-Shot Classification
Anonymous authors
Paper under double-blind review
Abstract
Pre-trained vision transformers have revolutionized few-shot image classification, and it has been recently demonstrated that the previous common practice of meta-learning in syner... |
QVVSb0GMXK | For time series from different domains, it is hard to determine a universal window length which works for all time series and contains data of single scale. The authors may wanna clarify that how they resolve this problem. | NewTime: Numerically Multi-Scaled Embedding for Large-Scale Time Series Pretraining
Anonymous authors
Paper under double-blind review
Abstract
Recent research on time-series self-supervised models shows great promise in learning semantic representations. However, it has been limited to small-scale datasets, e.g., th... |
77N93tc3o5 | Could you please describe the dataset used in the experiments, and relate them to the DeepIVA model? What aspect of the neuroimaging data used in the experiments is supposed to be modelled with a statistical dependence across datasets/subjects (i.e., among components of $\mathbf{s}_i$)? | DEEP INDEPENDENT VECTOR ANALYSIS
Anonymous authors
Paper under double-blind review
ABSTRACT
We introduce a deep multivariate latent variable model, Deep Independent Vector Analysis (DeepIVA), for learning linked and identifiable latent sources across multiple data modalities by unifying multidataset independent subs... |
FwdnG0xR02 | In this work, author only adopted COCO Captions dataset for tasking and gender as the debias attribute. It would be more convincing if author provide discussion and empirical result of how does insight draw from this work also applicable to other dataset and attribute. | Balancing the Picture: Debiasing Vision-Language Datasets with Synthetic Contrast Sets
Anonymous authors
Paper under double-blind review
Abstract
Vision-language models are growing in popularity and public visibility to generate, edit, and caption images at scale; but their outputs can perpetuate and amplify societa... |
BBD4cFDKxQ | The justification of this step where rather cryptic and I could not distinguish the argument of AdaProj over AdaCos from arguments for the whole design of centroid-based embeddings for outlier detection. | ADAProj: Adaptively Scaled Angular Margin Subspace Projections for Anomaly Detection with Auxiliary Classification Tasks
Anonymous authors
Paper under double-blind review
Abstract
One of the state-of-the-art approaches for semi-supervised anomaly detection is to first learn an embedding space and then estimate the d... |
h922Qhkmx1 | Q1: The separation result of ISDM and MSDM in Table 2 suggests a trade-off in terms of generation quality and separation performance. Is it confirmed? If so, why using a unified approach for source separation and generation? I understand on one hand, ISDM helps to show the diffusion method is comparable to the SOTA met... | MULTI-SOURCE DIFFUSION MODELS FOR SIMULTANEOUS MUSIC GENERATION AND SEPARATION
[NAME] Mariani∗
Sapienza University of Rome
[EMAIL]
Emilian Postolache∗
Sapienza University of Rome
[EMAIL]
Luca Cosmo†
Ca’ Foscari University of Venice
[EMAIL]
Irene Tallini∗
Sapienza University of Rome
[EMAIL]
[NAME] University of Rom... |
xmQMz9OPF5 | Following my question above, another perspective to understand this phenomenon is due to the limited scalability [1] of asymmetric masking architecture proposed in MAE, which is also utilized as the main architecture in this work. While in Swinv2 and EVA, the MIM pre-training has been quite important for ViT training o... | EXPLORING TARGET REPRESENTATIONS FOR MASKED AUTOENCODERS
[NAME] Liu\textsuperscript{1,2}\textsuperscript{*} [NAME] Zhou\textsuperscript{2}\textsuperscript{*} [NAME] Kong\textsuperscript{2}\textsuperscript{*} Xianming Lin\textsuperscript{1}\textsuperscript{†} Rongrong Ji\textsuperscript{1}
\textsuperscript{1}Xiamen Uni... |
EDXkkUAIFW | In the empirical setting part, I'm confused about the setting of *50 distinct architectures*. Since it's noted in the footnote that all the experiments were conducted with the same GPU and CPU, I'm wondering whether it's still necessary to employ such method? | ONE-SHOT ACTIVE LEARNING BASED ON LEWIS WEIGHT SAMPLING FOR MULTIPLE DEEP MODELS
[NAME]}, [NAME]}, [NAME]} [NAME],*}
College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics\textsuperscript{1}
School of Physical and Mathematical Sciences, Nanyang Technological University\textsuper... |
m52uU0dVbH | How does the proposed E-TS algorithm work compared to sequential arm elimination proposed in S. Shahrampour, M. Noshad and V. Tarokh. On Sequential Elimination Algorithms for Best-Arm Identification in Multi-Armed Bandits. IEEE Transactions on Signal Processing, vol. 65, no. 16, pp. 4281-4292, 15 Aug.15, 2017 | CONSTRUCTING ADVERSARIAL EXAMPLES FOR VERTICAL FEDERATED LEARNING: OPTIMAL CLIENT CORRUPTION THROUGH MULTI-ARMED BANDIT
Duanyi, [NAME]
Southeast University
[EMAIL]
[NAME], [NAME] of Big Data, CUHK(SZ)
[EMAIL]HKUST(GZ)
[EMAIL]
ABSTRACT
Vertical federated learning (VFL), where each participating client holds ... |
pPJTQYOpNI | If the goal is to first learn to follow earlier parts of trajectories first, and then move forward once policy learns, why not simply put a scheduler on truncating the expert trajectories, instead of on the discount factor? | IMITATION LEARNING FROM OBSERVATION WITH AUTOMATIC DISCOUNT SCHEDULING
Yuyang Liu\textsuperscript{1,2,*}, Weijun Dong\textsuperscript{1,2,*}, [NAME]}, [NAME] Wen\textsuperscript{1,2}, [NAME]}, Chongjie Zhang\textsuperscript{4}, [NAME]}
\textsuperscript{1}Institute for Interdisciplinary Information Sciences, Tsinghua U... |
O1lR4vSw5x | I don’t see how eqs 14 and 16 are any different from standard neural ODE loss of eq 3. To me this is a reordering of the original simple loss into a more complicated version, which still looks like the same thing. The eqs 17-19 seem to be just gradient updates with Hessians. Ok, but can’t we directly use a second-order... | Recursive Neural Ordinary Differential Equations for Partially Observed Systems
Anonymous authors
Paper under double-blind review
Abstract
Identifying spatiotemporal dynamics is a difficult task, especially in scenarios where latent states are partially observed and/or represent physical quantities. In this context,... |
MOtZlKkvdz | I’m slightly confused about the insights we can draw from the experiments. Specifically, the authors propose 3 different explanation strategies that seem to perform reasonably similarly. I understand the overall message that LLMs have the potential to be used as explainers. However, the confusing part to me is there ar... | Are Large Language Models Post Hoc Explainers?
Anonymous authors
Paper under double-blind review
Abstract
Large Language Models (LLMs) are increasingly used as powerful tools for a plethora of natural language processing (NLP) applications. A recent innovation, in-context learning (ICL), enables LLMs to learn new ta... |
wriKDQqiOQ | As far as I know, the speed gain obtained by choosing a larger batch size is naturally explained by the fact that we can benefit from the parallelization of computations (in the system side). However, I would find it interesting to understand whether the methods presented are faster to reach a given accuracy with a lar... | ON THE EFFECT OF BATCH SIZE IN BYZANTINE-ROBUST DISTRIBUTED LEARNING
Yi-[NAME] [NAME] [NAME]
National Key Laboratory for Novel Software Technology,
Department of Computer Science and Technology,
Nanjing University, Nanjing, China
[EMAIL], [EMAIL]
ABSTRACT
Byzantine-robust distributed learning (BRDL), in ... |
AgM3MzT99c | This is quite important in a practical open-ended environment. Without considering the survival component, the aspects that need to be considered during choosing the task become less adequate and complex, which goes against the motivation of using LLMs here. | OMNI: Open-endedness via Models of Human Notions of Interestingness
Jenny Zhang\textsuperscript{1,2} Joel Lehman\textsuperscript{3} Kenneth Stanley\textsuperscript{4} Jeff Clune\textsuperscript{1,2,5}
\textsuperscript{1}Department of Computer Science, University of British Columbia \textsuperscript{2}Vector Institute
... |
mrRbIcyouU | * Concatenating features does not ensure true adaptability to new classes, as it might lead to a model that is more of an ensemble of old and new knowledge rather than a seamlessly adapted model, raising the following concerns: (3) Lack of Interaction Between Old and New Knowledge: The concatenation approach lacks a me... | REVISITING CLASS-INCREMENTAL LEARNING WITH PRE-TRAINED MODELS: GENERALIZABILITY AND ADAPTIVITY ARE ALL YOU NEED
Anonymous authors
Paper under double-blind review
ABSTRACT
Class-incremental learning (CIL) aims to adapt to emerging new classes without forgetting old ones. Traditional CIL models are trained from scratc... |
488A64eOf6 | The method requires sampling in order to approximate the new parametric distribution (via the WIS algorithm). The runtime of this algorithm, and hence the additional computational complexity incurred by this decoding method, is not discussed. Further, there is an additional sampling + renormalization step that must hap... | LANGUAGE MODEL DECODING AS DIRECT METRICS OPTIMIZATION
[NAME] Ke∗ Hongning Wang [NAME]
The CoAI Group, DCST, BNRist, Tsinghua University, Beijing 100084, China
[EMAIL] [EMAIL]
ABSTRACT
Despite the remarkable advances in language modeling, current mainstream decoding methods still struggle to generate texts ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.