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
This paper describes the typeAs of biases that can appear in automatic text summarization and presents an initial study of such biases using two corpuses: a synthetically generated one and the CNN/Daily mail corpus. The paper selects a few types of biases that can be analyzed automatically and proposes metrics (or pro...	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 describes the typeAs of biases that can appear in automatic text summarization and presents an initial study of such biases using two corpuses: a synthetically generated one and the CNN/Daily mail corpus. The paper selects a few types of biases that can be analyzed automatically and proposes metrics...
The submission "When to Trust Aggregated Gradients: Addressing Negative Client Sampling in Federated Learning" describes a re-weighting strategy that can be applied during the optimization of ML models in federated learning. Given the full gradient norm over the entire dataset at some point in training, this strategy r...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The submission "When to Trust Aggregated Gradients: Addressing Negative Client Sampling in Federated Learning" describes a re-weighting strategy that can be applied during the optimization of ML models in federated learning. Given the full gradient norm over the entire dataset at some point in training, this st...
This paper proposed a RAIN to infer interactions among agents. It learns the attentive weight and unknown system dynamics, which is also helpful to predict future trajectories. The author(s) verified that RAIN can infer the system dynamic parameters and interaction graphs, and can outperform baseline discrete models. ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposed a RAIN to infer interactions among agents. It learns the attentive weight and unknown system dynamics, which is also helpful to predict future trajectories. The author(s) verified that RAIN can infer the system dynamic parameters and interaction graphs, and can outperform baseline discrete m...
The paper is concerned with accurately estimating the logging policy in off-policy learning when it is unknown or unavailable. Specifically, the paper derives an uncertainty-aware inverse propensity score algorithm called UIPS that is more accurate for rarely-taken actions, and provides theoretical and empirical argume...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper is concerned with accurately estimating the logging policy in off-policy learning when it is unknown or unavailable. Specifically, the paper derives an uncertainty-aware inverse propensity score algorithm called UIPS that is more accurate for rarely-taken actions, and provides theoretical and empirica...
The authors propose group reweighting algorithms for achieving group fairness. +: The performance of the algorithm seems to be better than the state of the arts. The paper writing is clear. -: The technical novelty seems limited. The difference between this work and the previous work [Agarwal et al., 2018] seems v...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors propose group reweighting algorithms for achieving group fairness. +: The performance of the algorithm seems to be better than the state of the arts. The paper writing is clear. -: The technical novelty seems limited. The difference between this work and the previous work [Agarwal et al., 2018]...
This paper analyzes hindsight experience replay (HER) through divergence minimization with energy based models. With this analysis, it presents an approach that combines HER with behavioral cloning (BC) by only imitating actions in the dataset if they move the agent a certain amount towards the goal. The paper conclud...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper analyzes hindsight experience replay (HER) through divergence minimization with energy based models. With this analysis, it presents an approach that combines HER with behavioral cloning (BC) by only imitating actions in the dataset if they move the agent a certain amount towards the goal. The paper...
The paper proposes a novel model of data, and shows that under the proposed model, a model can only succeed if it performs feature learning. Specifically, the primary data model proposed in the paper involves a set of concepts, each with an equal number of vocabularies. There is an $R$ number of categories (or classes)...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper proposes a novel model of data, and shows that under the proposed model, a model can only succeed if it performs feature learning. Specifically, the primary data model proposed in the paper involves a set of concepts, each with an equal number of vocabularies. There is an $R$ number of categories (or ...
1. The motivation is to tackle the accuracy-fairness tradeoffs of fair representation learning (FRL), indicating the need of providing provable upper bounds on unfairness of downstream classifiers. 2. They propose Fairness with Restricted Encoders (FARE), the first FRL method with provable fairness guarantees. They res...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: 1. The motivation is to tackle the accuracy-fairness tradeoffs of fair representation learning (FRL), indicating the need of providing provable upper bounds on unfairness of downstream classifiers. 2. They propose Fairness with Restricted Encoders (FARE), the first FRL method with provable fairness guarantees. ...
The authors introduces a new framework for OOD detection by defining more broadly the in-distribution (ID) samples using the notion of texture and semantic. Based on this definition, they create new benchmark scenarios. Finally, they propose a new method which perform well on both old and new benchmarks. Strength: - a ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors introduces a new framework for OOD detection by defining more broadly the in-distribution (ID) samples using the notion of texture and semantic. Based on this definition, they create new benchmark scenarios. Finally, they propose a new method which perform well on both old and new benchmarks. Streng...
This paper presents a GNN design that directly aggregates the multi-hop neighbors by assigning the node features based on the different hop numbers. The explicit node coloring augments the message passing process and thus enables the model to differentiate some non-isomorphic subgraphs that a regular 1-WL will fail to ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a GNN design that directly aggregates the multi-hop neighbors by assigning the node features based on the different hop numbers. The explicit node coloring augments the message passing process and thus enables the model to differentiate some non-isomorphic subgraphs that a regular 1-WL will ...
Cascading ensembling of certifiably adversarially robust classifiers, which has been used by previous works, is shown to be unsound, which is demonstrated by theoretical proof as as well as experimental results on real test set data. Strengths: - The theoretical introduction and methodology are clearly presented and a...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Cascading ensembling of certifiably adversarially robust classifiers, which has been used by previous works, is shown to be unsound, which is demonstrated by theoretical proof as as well as experimental results on real test set data. Strengths: - The theoretical introduction and methodology are clearly present...
This paper presents an image generation model called fast weight painters (FPA), applying the technique of fast weight programmer. The proposed method treats the generated image as the generated weight matrix. 1. This work seems to be one of the first attempts to apply fast weight programmer method to image generation...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents an image generation model called fast weight painters (FPA), applying the technique of fast weight programmer. The proposed method treats the generated image as the generated weight matrix. 1. This work seems to be one of the first attempts to apply fast weight programmer method to image ge...
Summary: Authors propose an active learning pipeline that incorporates domain knowledge into the pipeline for more efficient active learning & yielding models more consistent with known domain rules. Specifically, the model employs first-order-logic (FOL) based encoding to employ domain rules to inform examples to be s...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Summary: Authors propose an active learning pipeline that incorporates domain knowledge into the pipeline for more efficient active learning & yielding models more consistent with known domain rules. Specifically, the model employs first-order-logic (FOL) based encoding to employ domain rules to inform examples...
This paper seeks to investigate the worst-case performance of neural network models in the low-data regime with respect to data sample index choice. Studying this problem is motivated by wanting to investigate neural network’s tendency to rely on spurious correlations for classification. The authors propose a method ca...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper seeks to investigate the worst-case performance of neural network models in the low-data regime with respect to data sample index choice. Studying this problem is motivated by wanting to investigate neural network’s tendency to rely on spurious correlations for classification. The authors propose a m...
This paper proposes an algorithm to optimally compress a model while preserving performance and improving inference speed. The algorithm achieves compression by modelling the weights as quantized latent representations sampled from a Gaussian prior which are further optimized using an entropy penalty. In order to ensur...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an algorithm to optimally compress a model while preserving performance and improving inference speed. The algorithm achieves compression by modelling the weights as quantized latent representations sampled from a Gaussian prior which are further optimized using an entropy penalty. In order ...
This paper investigates a multi-agent backdoor attack scenario, where multiple attackers attempt to backdoor a victim model simultaneously. Backdoor attacks, where an attacker poisons a model during training to successfully achieve targeted misclassification, are a major concern to train-time robustness. A consistent b...	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 investigates a multi-agent backdoor attack scenario, where multiple attackers attempt to backdoor a victim model simultaneously. Backdoor attacks, where an attacker poisons a model during training to successfully achieve targeted misclassification, are a major concern to train-time robustness. A cons...
The paper proposes a learnable data augmentation procedure for speeding up reinforcement learning from pixels. In particular in applies a transformation to the input images, which is parameterized by a Gaussain distribution. The proposed method combines three main components: 1. Standard Q-learning approach with data ...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a learnable data augmentation procedure for speeding up reinforcement learning from pixels. In particular in applies a transformation to the input images, which is parameterized by a Gaussain distribution. The proposed method combines three main components: 1. Standard Q-learning approach wi...
Long-range spatial information can be passed with a low computational cost through pooling and unpooling operations. On an irregular mesh, there is a challenge of how to select pivot nodes and assign non-pivot nodes to a pivot node for pooling so that all assignments are at most 2 hops away. Previous methods have used ...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: Long-range spatial information can be passed with a low computational cost through pooling and unpooling operations. On an irregular mesh, there is a challenge of how to select pivot nodes and assign non-pivot nodes to a pivot node for pooling so that all assignments are at most 2 hops away. Previous methods ha...
This paper presents a unified framework for simultaneously detoxifying and debiasing language models. They consider protected groups as well as toxicity as an attribute of the generated text which can be controlled during the decoding time during an attribute classifier. Authors propose to use an embedding representati...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper presents a unified framework for simultaneously detoxifying and debiasing language models. They consider protected groups as well as toxicity as an attribute of the generated text which can be controlled during the decoding time during an attribute classifier. Authors propose to use an embedding repr...
This paper describes a method for characterizing the amount of added toxicity in a multilingual machine translation system’s output. The toxicity detection method is a list-based approach that is unchanged from an earlier publication (NLLB Team et al 2022, cited prominently and repeatedly). The main contributions here ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper describes a method for characterizing the amount of added toxicity in a multilingual machine translation system’s output. The toxicity detection method is a list-based approach that is unchanged from an earlier publication (NLLB Team et al 2022, cited prominently and repeatedly). The main contributio...
This paper follows-up on recent work around "chain-of-thought" (CoT) prompting, where the expected output of the model is not only the desired output, but also rationales that would reflect reasoning steps that a human would use to solve that same problem. The additional insight of this paper is that this process can ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper follows-up on recent work around "chain-of-thought" (CoT) prompting, where the expected output of the model is not only the desired output, but also rationales that would reflect reasoning steps that a human would use to solve that same problem. The additional insight of this paper is that this proc...
This paper proposes federated composite optimization, i.e., the nonconvex minimization of sum_i f_i + h in a federated setting. Here, f_i is smooth, h is not smooth but proximable. Note that the f_i are not assumed to be equal to each other (heterogeneous data). The Forward Backward Envelope (FBE) of f_i + h is a func...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes federated composite optimization, i.e., the nonconvex minimization of sum_i f_i + h in a federated setting. Here, f_i is smooth, h is not smooth but proximable. Note that the f_i are not assumed to be equal to each other (heterogeneous data). The Forward Backward Envelope (FBE) of f_i + h i...
This paper conducts a rigorous analysis on the fundamental properties of the policy induced value functions in multi-objective reinforcement learning (MORL). Based on the analysis, the authors further distinguish and consolidate three existing definitions of Pareto optimal policies and identify issues of training polic...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper conducts a rigorous analysis on the fundamental properties of the policy induced value functions in multi-objective reinforcement learning (MORL). Based on the analysis, the authors further distinguish and consolidate three existing definitions of Pareto optimal policies and identify issues of traini...
This paper considers the risk measure instead of the classically expected return in the offline RL setting. This paper shows that under some assumptions it is equivalent to minimize a risk measure in the latent space and in the natural space. A practical approach is proposed to use CVaR as the objective function. Stren...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper considers the risk measure instead of the classically expected return in the offline RL setting. This paper shows that under some assumptions it is equivalent to minimize a risk measure in the latent space and in the natural space. A practical approach is proposed to use CVaR as the objective functio...
The paper proposes a new loss function for time-series called TILDE-Q, which has 3 components. Each loss component is invariant to different distortions, including amplitude shifting, phase shifting, and uniform amplification. The paper starts with formal definitions of common time-series distortions and then designs 3...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes a new loss function for time-series called TILDE-Q, which has 3 components. Each loss component is invariant to different distortions, including amplitude shifting, phase shifting, and uniform amplification. The paper starts with formal definitions of common time-series distortions and then d...
This paper presents a method for learning solutions to PDE that are guaranteed to satisfy the constraints. This is done in eq. (5) by learning a basis for the solutions and finding the best linear combination of them. This results in the overall learning objective in eq. (8) that learns the basis that best-minimizes th...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents a method for learning solutions to PDE that are guaranteed to satisfy the constraints. This is done in eq. (5) by learning a basis for the solutions and finding the best linear combination of them. This results in the overall learning objective in eq. (8) that learns the basis that best-mini...
- The paper addresses the problem of model extraction attacks (i.e., learning a surrogate model to mimic a black-box classifier under a budgeted query constraint) - The main contributions of the paper is two-fold: 1. formally posing model extraction as a bi-level optimization problem, thereby unifying some work ar...	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 addresses the problem of model extraction attacks (i.e., learning a surrogate model to mimic a black-box classifier under a budgeted query constraint) - The main contributions of the paper is two-fold: 1. formally posing model extraction as a bi-level optimization problem, thereby unifying some...
This paper discusses the challenge of "relay generalization", where a policy trained with RL is evaluated when starting from initial states observed under an independently trained policy. Relay generalization is intended to measure the out-of-distribution generalization error, by testing how well the policy performs s...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper discusses the challenge of "relay generalization", where a policy trained with RL is evaluated when starting from initial states observed under an independently trained policy. Relay generalization is intended to measure the out-of-distribution generalization error, by testing how well the policy pe...
The paper proposes to divide computational graph of the neural network into the fixed path and float path and then shows the improvement of robust accuracy is mainly gained from correcting the vulnerability floating neurons. Specifically, the neural network graph is first divided into several activation regions and thu...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes to divide computational graph of the neural network into the fixed path and float path and then shows the improvement of robust accuracy is mainly gained from correcting the vulnerability floating neurons. Specifically, the neural network graph is first divided into several activation regions...
The paper proposes an interesting supervised learning setting for causal discovery. Firstly, it creates synthetic training data ( observational and international data ) with various simulators and settings. And then it designs a neural network that outputs the causal graphs. Based on such a supervised learning manner, ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes an interesting supervised learning setting for causal discovery. Firstly, it creates synthetic training data ( observational and international data ) with various simulators and settings. And then it designs a neural network that outputs the causal graphs. Based on such a supervised learning ...
In this paper, the authors proposed a method for subset selection, which is useful for applications such as active learning. Specifically, in the proposed method, when constructing the subset, data is selected by importance sampling, where the sampling probability relates to the model's entropy. In the subset, the sele...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: In this paper, the authors proposed a method for subset selection, which is useful for applications such as active learning. Specifically, in the proposed method, when constructing the subset, data is selected by importance sampling, where the sampling probability relates to the model's entropy. In the subset, ...
This work studies a relatively new problem in self-supervised learning, which is called sustainable self-supervised learning. Based on a pretrained SSL model (or partially trained with certain epochs), the authors want to further improve it with the proposed target-enhanced conditional mask-reconstruction. The authors ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This work studies a relatively new problem in self-supervised learning, which is called sustainable self-supervised learning. Based on a pretrained SSL model (or partially trained with certain epochs), the authors want to further improve it with the proposed target-enhanced conditional mask-reconstruction. The ...
The core questions in the paper are 1) What is the geometry of representations learned by contrastive learning with data augmentation? 2) What are the determinants of good generalization in downstream tasks? Strengths: A set of intuitive descriptors "manifold graph metrics" is proposed and evidence is presented tha...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The core questions in the paper are 1) What is the geometry of representations learned by contrastive learning with data augmentation? 2) What are the determinants of good generalization in downstream tasks? Strengths: A set of intuitive descriptors "manifold graph metrics" is proposed and evidence is prese...
In this paper, the authors describe a hierarchical strategy for storing large batches of GNN training data in multiple levels of memory in a single system. By partitioning the storage of GNN training data between the extremely large out-of-core disk, smaller but faster CPU DRAM main memory, and much more limited but fa...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors describe a hierarchical strategy for storing large batches of GNN training data in multiple levels of memory in a single system. By partitioning the storage of GNN training data between the extremely large out-of-core disk, smaller but faster CPU DRAM main memory, and much more limite...
Summary: This paper provides some explanations for regularization in self-supervised learning algorithms and, to some extent, other phenomena beyond self-supervised learning. Its analysis is reasonable and comprehensive. However, this analysis does not yield very useful insights into the improvement of self-supervised...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Summary: This paper provides some explanations for regularization in self-supervised learning algorithms and, to some extent, other phenomena beyond self-supervised learning. Its analysis is reasonable and comprehensive. However, this analysis does not yield very useful insights into the improvement of self-su...
This paper points out an issue in a recent theoretical result that motivates the adversarial training algorithm. The authors construct counter examples and provide explanations about the issue they identify, and then further propose a Danskin's Descent Direction for training robust neural networks. Strengths: The cla...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper points out an issue in a recent theoretical result that motivates the adversarial training algorithm. The authors construct counter examples and provide explanations about the issue they identify, and then further propose a Danskin's Descent Direction for training robust neural networks. Strengths: ...
This paper explores the active learning problem under OOD data scenarios and incorporates AL and OOD objectives within a multi-objective optimization framework to balance their conflict. Specifically, they propose a Monte-Carlo Pareto optimization mechanism to enable efficient Pareto optimization, which selects optimal...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper explores the active learning problem under OOD data scenarios and incorporates AL and OOD objectives within a multi-objective optimization framework to balance their conflict. Specifically, they propose a Monte-Carlo Pareto optimization mechanism to enable efficient Pareto optimization, which selects...
This paper proposes an unsupervised keypoint discovery method, especially for videos. To this end, authors propose to build an entropy layer to measure the information, and a deep learning model to predict keypoint. The model is learned in an unsupervised manner, by maximizing the information in a single frame and acro...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes an unsupervised keypoint discovery method, especially for videos. To this end, authors propose to build an entropy layer to measure the information, and a deep learning model to predict keypoint. The model is learned in an unsupervised manner, by maximizing the information in a single frame ...
This paper proposes a method for 3D shape part segmentation. It learns an image-conditioned NeRF together with a semantic field. The method enables novel view (2D image) part segmentation and 3D point cloud part segmentation. The authors evaluate the proposed method on the PartNet dataset and show some real-world demos...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a method for 3D shape part segmentation. It learns an image-conditioned NeRF together with a semantic field. The method enables novel view (2D image) part segmentation and 3D point cloud part segmentation. The authors evaluate the proposed method on the PartNet dataset and show some real-wor...
Summary: - The paper proposed out-of-distribution (OOD) evaluation sets for the COIN and Breakfast dataset, together referred to as the new GAIN benchmark. To formulate the novel OOD evaluation sets, the authors ensured that the training set and the OOD evaluation set share the same set of steps, but different sets of ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Summary: - The paper proposed out-of-distribution (OOD) evaluation sets for the COIN and Breakfast dataset, together referred to as the new GAIN benchmark. To formulate the novel OOD evaluation sets, the authors ensured that the training set and the OOD evaluation set share the same set of steps, but different ...