Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
The authors developed a new efficient, single-timescale variance-reduced primal-dual method (VRPD) with an emphasis on the feasibility of solving non-convex policy evaluation problem in reinforcement learning (RL) with nonlinear function approximation, also known as the value function estimation or on-policy learning w...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors developed a new efficient, single-timescale variance-reduced primal-dual method (VRPD) with an emphasis on the feasibility of solving non-convex policy evaluation problem in reinforcement learning (RL) with nonlinear function approximation, also known as the value function estimation or on-policy le...
The paper seeks to understand architecture changes to traditional CNNs that render them as, or more, robust as Transformers when applied to out of distribution data. The main conclusions of the paper are that employing (a) an initial partition of the image into non-overlapping patches; (b) larger kernel sizes in a resn...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper seeks to understand architecture changes to traditional CNNs that render them as, or more, robust as Transformers when applied to out of distribution data. The main conclusions of the paper are that employing (a) an initial partition of the image into non-overlapping patches; (b) larger kernel sizes i...
This work studies the convergence of gradient descent in the regime of learning ReLU activations agnostically for a class of distributions that contains the standard high dimensional normal distribution. In contrast with previous results, they consider the case where bias exists and apply gradient descent on the weight...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work studies the convergence of gradient descent in the regime of learning ReLU activations agnostically for a class of distributions that contains the standard high dimensional normal distribution. In contrast with previous results, they consider the case where bias exists and apply gradient descent on th...
This paper provides a training strategy by applying sparsification to obtain a more interpretable network. The training loss combines the feature diversity loss and sparsification loss used in Glm-saga. The measure of feature diversity is proposed and the paper shows that classes with the diverse feature can improve th...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper provides a training strategy by applying sparsification to obtain a more interpretable network. The training loss combines the feature diversity loss and sparsification loss used in Glm-saga. The measure of feature diversity is proposed and the paper shows that classes with the diverse feature can im...
This work studies a notion of unlearning for graph neural networks. Basically, given a set of edges removed from the graph, it tries to address how to fast adjust the model parameters to make the model behave like the model retrained on the graph with edge removal while without retraining. The technical idea is to anal...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This work studies a notion of unlearning for graph neural networks. Basically, given a set of edges removed from the graph, it tries to address how to fast adjust the model parameters to make the model behave like the model retrained on the graph with edge removal while without retraining. The technical idea is...
This paper proposes to add set-wise connections to pair-wise connections in the Hopfield networks. The main contribution of the paper is the claim that adding such connection improves the memory storage properties of these networks. Interesting biological mechanisms of building many-body synapses are also mentioned. ...	Recommendation: 8: accept, good paper	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper proposes to add set-wise connections to pair-wise connections in the Hopfield networks. The main contribution of the paper is the claim that adding such connection improves the memory storage properties of these networks. Interesting biological mechanisms of building many-body synapses are also menti...
This paper proposes an outlier detection algorithm that is based on bidirectional mapping between the data space and the latent space. The proposed algorithm maps the normal data to a restricted region in the latent space, letting the outliers map to the outside of the region. # Strength * The paper is overall very cl...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposes an outlier detection algorithm that is based on bidirectional mapping between the data space and the latent space. The proposed algorithm maps the normal data to a restricted region in the latent space, letting the outliers map to the outside of the region. # Strength * The paper is overall...
There are now a wide crop of Spherical CNN models, which can be classified along several different axes e.g. fully Fourier/fully real/partly real and partly harmonic; or continuous models and discrete models. If we focus on the latter, existing continuous models guarantee rotational equivariance, while being computatio...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: There are now a wide crop of Spherical CNN models, which can be classified along several different axes e.g. fully Fourier/fully real/partly real and partly harmonic; or continuous models and discrete models. If we focus on the latter, existing continuous models guarantee rotational equivariance, while being co...
The paper introduces an algorithm to solve bilevel optimization problems with non-smooth inner problems. Authors propose to smooth the non-smooth term in the inner problem, and to iteratively decrease the smoothing parameter. In addition, a stochastic approach is proposed to avoid compute full hypergradient. My main ...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper introduces an algorithm to solve bilevel optimization problems with non-smooth inner problems. Authors propose to smooth the non-smooth term in the inner problem, and to iteratively decrease the smoothing parameter. In addition, a stochastic approach is proposed to avoid compute full hypergradient. ...
This paper introduces and mitigates some challenges encountered in contrastive representation distillation. In particular, the authors point out that the existing methods (CRD and wCoRD) keep track of a memory bank for the negatives that often contain inconsistent representations, which can adversely affect the student...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces and mitigates some challenges encountered in contrastive representation distillation. In particular, the authors point out that the existing methods (CRD and wCoRD) keep track of a memory bank for the negatives that often contain inconsistent representations, which can adversely affect the...
The paper presents a method named TCSR which is based on EfficientZero. TCSR helps a reinforcement learning agent to learn better latent representations in a self supervised manner. Weaknesses: The paper has several weakness in writing, results, explanation of the approach. See below: Results: - The paper does not ...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper presents a method named TCSR which is based on EfficientZero. TCSR helps a reinforcement learning agent to learn better latent representations in a self supervised manner. Weaknesses: The paper has several weakness in writing, results, explanation of the approach. See below: Results: - The paper d...
The paper focuses on addressing the biases in the VCR dataset. It found that the correct answer has the most overlapping words with the question and the incorrect choices have low bert--similarity with the correct ones. The authors propose to rewrite all answer choices, generate counterfactual images, and train the mod...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper focuses on addressing the biases in the VCR dataset. It found that the correct answer has the most overlapping words with the question and the incorrect choices have low bert--similarity with the correct ones. The authors propose to rewrite all answer choices, generate counterfactual images, and train...
This paper provided a Céa-type lemma's lemma for p-laplacian. Based on this, the author provide a breaking the curse of dimensionality proof for solving the PDE. Question 1: What's the application of p-Laplacian equation in high dimension? ( I don't think clustering is a good application. Question 2: The main contrib...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper provided a Céa-type lemma's lemma for p-laplacian. Based on this, the author provide a breaking the curse of dimensionality proof for solving the PDE. Question 1: What's the application of p-Laplacian equation in high dimension? ( I don't think clustering is a good application. Question 2: The main...
The paper presents theoretical and empirical results supporting an exponential gap between a sequence-to-sequence task and its decomposition into many smaller subtasks. In this study, the authors consider a bit parity task where the objective is to determine if the number of ones in a subset of a 0/1 array of length d ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper presents theoretical and empirical results supporting an exponential gap between a sequence-to-sequence task and its decomposition into many smaller subtasks. In this study, the authors consider a bit parity task where the objective is to determine if the number of ones in a subset of a 0/1 array of l...
This paper proposes a method to achieve anytime networks by controlling the network depth during runtime. The method focuses on the typical ResNet-like architectures and proposes to divide the residual blocks in each stage into two parts that are responsible for feature learning and feature refinement, respectively. A ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a method to achieve anytime networks by controlling the network depth during runtime. The method focuses on the typical ResNet-like architectures and proposes to divide the residual blocks in each stage into two parts that are responsible for feature learning and feature refinement, respecti...
This paper provides a "fine-grained" theory for the label propagation algorithm (LPA) in the presence of prior information. Unlike the existing theory, the new error bound takes into account the local geometric properties of the graph. The new error bound enables us to discriminate two geometrically different graphs wi...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper provides a "fine-grained" theory for the label propagation algorithm (LPA) in the presence of prior information. Unlike the existing theory, the new error bound takes into account the local geometric properties of the graph. The new error bound enables us to discriminate two geometrically different g...
This paper is about understanding the functionality of layers in a deep neural network by examining the curvature of the low-dimensional manifold associated with the feature maps at different layers. To evaluate the curvature, one has to have sufficiently dense samples on the manifold and this is achieved by adding sma...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper is about understanding the functionality of layers in a deep neural network by examining the curvature of the low-dimensional manifold associated with the feature maps at different layers. To evaluate the curvature, one has to have sufficiently dense samples on the manifold and this is achieved by ad...
The paper studies the problem of noisy label learning and proposes a method that computes a score based on the agreements over an ensemble of models to identify the samples with the correct given labels. The paper conducts experiments and shows that the disagreements among the ensemble of models strongly correlate with...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper studies the problem of noisy label learning and proposes a method that computes a score based on the agreements over an ensemble of models to identify the samples with the correct given labels. The paper conducts experiments and shows that the disagreements among the ensemble of models strongly correl...
This work studies the optimization landscape of deep neural networks through tracking the correlation of adjacent gradient steps of full batch gradient descent. Surprisingly, the authors find a strong zigzag behaviour, where adjacent gradients consistently have strong negative correlation values for a large period of ...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This work studies the optimization landscape of deep neural networks through tracking the correlation of adjacent gradient steps of full batch gradient descent. Surprisingly, the authors find a strong zigzag behaviour, where adjacent gradients consistently have strong negative correlation values for a large pe...
The paper presents a new SVD-based Graph Contrastive Learning paradigm called LightGCL to learn effective representations of nodes in a user-item interaction graph from unlabeled data that are eventually used to predict users' preferences in downstream recommendation tasks. The key idea is in the augmentation step whe...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents a new SVD-based Graph Contrastive Learning paradigm called LightGCL to learn effective representations of nodes in a user-item interaction graph from unlabeled data that are eventually used to predict users' preferences in downstream recommendation tasks. The key idea is in the augmentation ...
This paper first proposes a novel deep graph-level anomaly detection model, which learns the graph representation with maximum mutual information between substructure features and global structure features while exploring a hypersphere anomaly decision boundary. The numerical and visualization results on a few graph da...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper first proposes a novel deep graph-level anomaly detection model, which learns the graph representation with maximum mutual information between substructure features and global structure features while exploring a hypersphere anomaly decision boundary. The numerical and visualization results on a few ...
The authors claim to propose to sample directly from the distribution over the particle positions, eliminating boundaries while adaptively focusing on the most relevant regions. It looks higher sample efficiency and improved performance of PINNs. Weaknesses: It looks that I do not understand what the authors' motivati...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors claim to propose to sample directly from the distribution over the particle positions, eliminating boundaries while adaptively focusing on the most relevant regions. It looks higher sample efficiency and improved performance of PINNs. Weaknesses: It looks that I do not understand what the authors' ...
This paper proposes novel ways to enforce an inductive bias into a temporal auto-encoder model, by: 1) enforcing that the R part of the model should be able to operate on both the original version of the time-varying embeddings and also on a transformed version. 2) separating out the per-session and per-frame parts of ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes novel ways to enforce an inductive bias into a temporal auto-encoder model, by: 1) enforcing that the R part of the model should be able to operate on both the original version of the time-varying embeddings and also on a transformed version. 2) separating out the per-session and per-frame p...
The work presents a model for motion prediction. The core component of the method is the prioritization of the agents. Agent with higher priority will have better prediction. The work comes supported with experiments and good analysis. Introduction: - Introduction is clear and work is well motivated based on the exam...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The work presents a model for motion prediction. The core component of the method is the prioritization of the agents. Agent with higher priority will have better prediction. The work comes supported with experiments and good analysis. Introduction: - Introduction is clear and work is well motivated based on ...
This paper proposes the Hidden-Utility Self-Play algorithm to explicitly model human bias via a modification to the reward function used during self-play training. This is used in the domain of the cooperative game Overcooked, where a particular reward modification is shown to be effective both in experiments where it ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes the Hidden-Utility Self-Play algorithm to explicitly model human bias via a modification to the reward function used during self-play training. This is used in the domain of the cooperative game Overcooked, where a particular reward modification is shown to be effective both in experiments w...
This paper proposes a continual learning algorithm, called Regularized Adaptive Weight Modification (RAWM) to overcome catastrophic forgetting for fake audio detection. RAWM is based on the previously published orthogonal weight modification (OWM). Because OWM does not consider the similarity of some audio, including f...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a continual learning algorithm, called Regularized Adaptive Weight Modification (RAWM) to overcome catastrophic forgetting for fake audio detection. RAWM is based on the previously published orthogonal weight modification (OWM). Because OWM does not consider the similarity of some audio, inc...
This paper proposes a domain generalization method for time series classification to improve the generalization on new unseen target domain without using domain labels. Moreover, an identification approach of the sub-distributions inside the data is proposed. Last, a min-max adversarial approach is presented to find do...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a domain generalization method for time series classification to improve the generalization on new unseen target domain without using domain labels. Moreover, an identification approach of the sub-distributions inside the data is proposed. Last, a min-max adversarial approach is presented to...
The paper provides a theoretical analysis that explains the benefit of MRP in SSL. Utilizing the multi-view data distribution and certain assumptions about the model architecture, the paper shows that the encoder trained with MRP can capture all the discriminative semantics of each semantic class in the pretraining d...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper provides a theoretical analysis that explains the benefit of MRP in SSL. Utilizing the multi-view data distribution and certain assumptions about the model architecture, the paper shows that the encoder trained with MRP can capture all the discriminative semantics of each semantic class in the pretr...
This paper provides a general template for MARL algorithmic designs. They prove that algorithms derived from the HAML template satisfy the desired properties of the monotonic improvement of the joint reward and the convergence to Nash equilibrium. +: The paper has nice result for the properties of a class of policies....	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper provides a general template for MARL algorithmic designs. They prove that algorithms derived from the HAML template satisfy the desired properties of the monotonic improvement of the joint reward and the convergence to Nash equilibrium. +: The paper has nice result for the properties of a class of p...
This paper empirically evaluates the correlation between different hyperparameters of tensor compressive layers and the classification performance. Their results show that the error is indicative of model performance, even when comparing multiple TD methods, though useful correlation only occurs at the higher compressi...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper empirically evaluates the correlation between different hyperparameters of tensor compressive layers and the classification performance. Their results show that the error is indicative of model performance, even when comparing multiple TD methods, though useful correlation only occurs at the higher c...
This paper investigates the example selection problem which is selecting reasoning examples that make the most effective prompts when prompting large language models with a chain of thoughts (CoT) prompt. The authors apply their approach at both prompting and decoding time. Evaluating on multi-step reasoning task, they...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper investigates the example selection problem which is selecting reasoning examples that make the most effective prompts when prompting large language models with a chain of thoughts (CoT) prompt. The authors apply their approach at both prompting and decoding time. Evaluating on multi-step reasoning ta...
The paper presents Autoregressive Conditional Neural Processes, a way to use conditional neural processes in an autoregressive way, by feeding newly sampled points back into the CNP and re-evaluating. The paper vouches that using CNPs in this autoregressive way gives better predictive performance, since there are no in...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper presents Autoregressive Conditional Neural Processes, a way to use conditional neural processes in an autoregressive way, by feeding newly sampled points back into the CNP and re-evaluating. The paper vouches that using CNPs in this autoregressive way gives better predictive performance, since there a...
The paper studies offline learning of the Nash equilibrium of the congestion game. In a congestion game, agents choose which facility to use. The reward the agent obtains from using the facility depends only on the number of total agents using that facility. In this paper, the authors consider three types of feedback: ...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies offline learning of the Nash equilibrium of the congestion game. In a congestion game, agents choose which facility to use. The reward the agent obtains from using the facility depends only on the number of total agents using that facility. In this paper, the authors consider three types of fe...
The paper develops a multi-agent algorithm suitable for handling a large number of agents by means of the mean-field approximation. Strength: - an interesting and important problem - natural approach Weaknesses - poor empirical evolution The paper develops a multi-agent algorithm suitable for handling a large number o...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper develops a multi-agent algorithm suitable for handling a large number of agents by means of the mean-field approximation. Strength: - an interesting and important problem - natural approach Weaknesses - poor empirical evolution The paper develops a multi-agent algorithm suitable for handling a large ...
This paper revisited federated and decentralized learning and showed the linear convergence rate for both of them when the objective loss satisfied the PL condition without generic assumptions such as bounded variance. Specifically, the authors leveraged the sample-wise smoothness of the local loss functions to capture...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper revisited federated and decentralized learning and showed the linear convergence rate for both of them when the objective loss satisfied the PL condition without generic assumptions such as bounded variance. Specifically, the authors leveraged the sample-wise smoothness of the local loss functions to...
# Intro / Section 2 - This paper introduces the property of diffused redundancy: any randomly chosen subset of a layer can perform similarly to the full layer on downstream tasks. - The degree of diffuse redundancy is investigated for the final layers of image classification models trained on imagenet. - Two downstream...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: # Intro / Section 2 - This paper introduces the property of diffused redundancy: any randomly chosen subset of a layer can perform similarly to the full layer on downstream tasks. - The degree of diffuse redundancy is investigated for the final layers of image classification models trained on imagenet. - Two do...
The paper proposes LINEX, an explainable AI(XAI) method. In a typical categorization of XAI, LINEX deals with a method to explain a given black-box predictor locally with features, i.e. providing an explanation of a given data with important features. LINEX employs IRM and game theory in its formulation and the optimiz...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes LINEX, an explainable AI(XAI) method. In a typical categorization of XAI, LINEX deals with a method to explain a given black-box predictor locally with features, i.e. providing an explanation of a given data with important features. LINEX employs IRM and game theory in its formulation and the...
This paper uses protein structure embedding as as features to predict protein-protein interaction. Strengths: + Using protein structural embedding for PPI prediction is an important direction. Weaknesses: + I find this paper to be not very clear. For example, what does the AUC actually measure in the protein-protein i...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper uses protein structure embedding as as features to predict protein-protein interaction. Strengths: + Using protein structural embedding for PPI prediction is an important direction. Weaknesses: + I find this paper to be not very clear. For example, what does the AUC actually measure in the protein-p...
This paper discusses the graph biconnectivity and the potential of existing GNNs on recognizing it. It first proves most of the existing GNNs and 1-WL cannot distinguish biconnectivity. The author proves ESAN are powerful enough for this problem due to the subgraph aggregation policies. Furthermore, author discusses th...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper discusses the graph biconnectivity and the potential of existing GNNs on recognizing it. It first proves most of the existing GNNs and 1-WL cannot distinguish biconnectivity. The author proves ESAN are powerful enough for this problem due to the subgraph aggregation policies. Furthermore, author disc...
The paper experimentally discusses the emergence of sparse activations in the feed-forward layer of Transformer neural networks. The authors highlight that sparse activations emerge for a variety of tasks and architectures. They then discuss how sparsity can be taken advantage of to reduce FLOP count and the implicatio...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper experimentally discusses the emergence of sparse activations in the feed-forward layer of Transformer neural networks. The authors highlight that sparse activations emerge for a variety of tasks and architectures. They then discuss how sparsity can be taken advantage of to reduce FLOP count and the im...