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 paper studies offline multi-agent reinforcement learning. The authors propose a framework for offline equilibrium finding, where the learner first trains a model $E$ by the offline dataset, and then compute an equilibrium using the model and some proper online algorithms. The authors also consider to set the behavi...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies offline multi-agent reinforcement learning. The authors propose a framework for offline equilibrium finding, where the learner first trains a model $E$ by the offline dataset, and then compute an equilibrium using the model and some proper online algorithms. The authors also consider to set th...
The paper proposes a new supervised-learning method for combinatorial optimization problems. One of the main contributions of the paper is the introduction of four types of data augmentation in training (i.e., Rotation, Symmetry, Shrink, and Noise). Another contribution is a new bidirectional loss, in which given the o...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a new supervised-learning method for combinatorial optimization problems. One of the main contributions of the paper is the introduction of four types of data augmentation in training (i.e., Rotation, Symmetry, Shrink, and Noise). Another contribution is a new bidirectional loss, in which giv...
The paper presents a method for learning a collaborative agent that is able to handle arbitrary and/or suboptimal human reward functions. The method leverages self-play, specifically fictitious co-play, where the policies the target policy plays alongside are learned according to different reward functions. A variety o...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper presents a method for learning a collaborative agent that is able to handle arbitrary and/or suboptimal human reward functions. The method leverages self-play, specifically fictitious co-play, where the policies the target policy plays alongside are learned according to different reward functions. A v...
This paper proposed an offline method to overcome the covariate shift issue in imitation learning. A context-conditioned imitation learning method was proposed, which learns a policy to map context state into ego vehicle's state without any history ego vehicle's information. To apply the method to urban driving, an eg...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed an offline method to overcome the covariate shift issue in imitation learning. A context-conditioned imitation learning method was proposed, which learns a policy to map context state into ego vehicle's state without any history ego vehicle's information. To apply the method to urban drivin...
This paper proposed ARHGA, a heterogeneous graph augmentation method for attribute reconstruction. It solved both missing attributes and defective attributes problems in heterogeneous information networks. Extensive experiments prove the effectiveness of proposed attribute reconstruction framework. Strength: -- The at...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed ARHGA, a heterogeneous graph augmentation method for attribute reconstruction. It solved both missing attributes and defective attributes problems in heterogeneous information networks. Extensive experiments prove the effectiveness of proposed attribute reconstruction framework. Strength: -...
This paper explores single-frame training for video-and language tasks. While simple, this approach can achieve state-of-the-art performance on a range of datasets. This paper also proposes an early-fusion strategy at inference which boosts performance. Strengths: 1. The single-frame training strategy is highly effici...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper explores single-frame training for video-and language tasks. While simple, this approach can achieve state-of-the-art performance on a range of datasets. This paper also proposes an early-fusion strategy at inference which boosts performance. Strengths: 1. The single-frame training strategy is highl...
This paper aims to make semi-supervised learning robust when unlabeled data contains out-of-class data. This is a practical problem. The authors improve the semi-supervised learning algorithm PAWS by calibrating the prediction of PAWS based on the data densities. Experimental results show the proposal achieves better p...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper aims to make semi-supervised learning robust when unlabeled data contains out-of-class data. This is a practical problem. The authors improve the semi-supervised learning algorithm PAWS by calibrating the prediction of PAWS based on the data densities. Experimental results show the proposal achieves ...
This paper proposed a new top-down attention framework for the task of speech separation. The knowledge is from human brain mechanism. The authors show that their method can achieve better separation results with a rather smaller model compared with recent approaches. Strengths: the authors work towards realistic app s...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed a new top-down attention framework for the task of speech separation. The knowledge is from human brain mechanism. The authors show that their method can achieve better separation results with a rather smaller model compared with recent approaches. Strengths: the authors work towards realist...
The paper introduces the problem of mapping from table data to images. The author's test strategies directly generate images from tabular data (which can include numerical values). This would give an end user the ability to create a visual summary of the data they are looking at. Generated images are evaluated using FI...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper introduces the problem of mapping from table data to images. The author's test strategies directly generate images from tabular data (which can include numerical values). This would give an end user the ability to create a visual summary of the data they are looking at. Generated images are evaluated ...
The paper considers few-shot text classification for pre-trained LMs. A supervised contrastive learning approach is taken, with consistency-regularization, and evaluation is performed across SST2, IMDB, SUBJ and PC datasets, against a collection of recent models, including BERT-SCL and DualCL. Strengths - contrasti...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper considers few-shot text classification for pre-trained LMs. A supervised contrastive learning approach is taken, with consistency-regularization, and evaluation is performed across SST2, IMDB, SUBJ and PC datasets, against a collection of recent models, including BERT-SCL and DualCL. Strengths - c...
This paper proposes a deep learning-based Granger causality model for discovering the Granger causal graph and imputing missing values at the same time. The author claim this model can handle missing completely at random and periodic missing scenarios. The main methodology is to iteratively apply the data imputation an...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes a deep learning-based Granger causality model for discovering the Granger causal graph and imputing missing values at the same time. The author claim this model can handle missing completely at random and periodic missing scenarios. The main methodology is to iteratively apply the data imput...
This paper investigates the latent separability assumption in backdoor defenses. The paper presents adaptive backdoor poisoning attacks against this assumption and circumvents the recent defenses. Strengths: 1. The proposed adaptive attack is well-motivated by the two insights and aims to avoid the latent separability ...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper investigates the latent separability assumption in backdoor defenses. The paper presents adaptive backdoor poisoning attacks against this assumption and circumvents the recent defenses. Strengths: 1. The proposed adaptive attack is well-motivated by the two insights and aims to avoid the latent separ...
This work establishes the convergence bound of split learning (SL) for non-convex objectives with non-iid data. The convergence results provide some insight on the parameter tuning and potential benefits/limitations of split learning, compared with standard federated learning (FL, local SGD). The authors also conduct e...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This work establishes the convergence bound of split learning (SL) for non-convex objectives with non-iid data. The convergence results provide some insight on the parameter tuning and potential benefits/limitations of split learning, compared with standard federated learning (FL, local SGD). The authors also c...
The authors tackle personalized neural architecture search for federated learning, where data heterogeneity is alleviated by the architectural personalization. The authors propose a method named SPIDER that uses a supernet from which local models can be sampled while sharing the weights. The authors validate their meth...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors tackle personalized neural architecture search for federated learning, where data heterogeneity is alleviated by the architectural personalization. The authors propose a method named SPIDER that uses a supernet from which local models can be sampled while sharing the weights. The authors validate th...
This paper studies the robust overfitting (RO) problem for adversarial training (AT) which usually happens after the learning rate (LR) decay. In particular, it hypothesizes that RO is caused by a loss of balance (trainer becomes stronger than attacker) after LR decay. It verifies the hypothesis by looking into target-...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper studies the robust overfitting (RO) problem for adversarial training (AT) which usually happens after the learning rate (LR) decay. In particular, it hypothesizes that RO is caused by a loss of balance (trainer becomes stronger than attacker) after LR decay. It verifies the hypothesis by looking into...
This paper studies the how discretization over the time horizon affect the error of LQR value estimation compared with the true continuous-time LQR value. The main result is about the mean squared error (MSE) of the Monte-Carlo policy evaluation, which reveals a trade-off between discretization time interval $h$ and th...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the how discretization over the time horizon affect the error of LQR value estimation compared with the true continuous-time LQR value. The main result is about the mean squared error (MSE) of the Monte-Carlo policy evaluation, which reveals a trade-off between discretization time interval $h...
The authors investigate the robustness of the model trained to distribution shifts by varying the loss functions, datasets, and the selection of targeted shifts. They found that (1) models trained by Vision-Language (VL) and Cross-Entropy (CE) loss have drastically different robustness even when the training sets are i...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors investigate the robustness of the model trained to distribution shifts by varying the loss functions, datasets, and the selection of targeted shifts. They found that (1) models trained by Vision-Language (VL) and Cross-Entropy (CE) loss have drastically different robustness even when the training se...
This paper proposes an approach for planning-as-inference in reinforcement learning problems with sparse rewards. A method based on sequential Monte Carlo is derived, similar to Piché et al. (2019), to sample from the optimal state-action trajectories distribution by applying filtering and smoothing techniques. The alg...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes an approach for planning-as-inference in reinforcement learning problems with sparse rewards. A method based on sequential Monte Carlo is derived, similar to Piché et al. (2019), to sample from the optimal state-action trajectories distribution by applying filtering and smoothing techniques....
The paper studies model-based networks in the context of sparse coding. A model-based network is constructed by unfolding an iterative algorithm, i.e. mapping each iteration of the original algorithm to a layer of the resulting neural network. This allows to have a trainable neural network that can adapt to specific pr...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies model-based networks in the context of sparse coding. A model-based network is constructed by unfolding an iterative algorithm, i.e. mapping each iteration of the original algorithm to a layer of the resulting neural network. This allows to have a trainable neural network that can adapt to spe...
The paper aims to study the “computational complexity” of attacking versus that of defending machine learning models against adversarial attacks, which are test-time attacks that find so-called adversarial examples. An adversarial example for an input x is a bounded perturbation of x that is misclassified. The paper ...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper aims to study the “computational complexity” of attacking versus that of defending machine learning models against adversarial attacks, which are test-time attacks that find so-called adversarial examples. An adversarial example for an input x is a bounded perturbation of x that is misclassified. Th...
The authors present a novel method that decouples back propagation gradient computation in artificial neural networks. The authors identify issues that result from long gradient flows in back propagation, and propose a method that tackles some issues, which include vanishing/exploding gradients, and update locking. In ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors present a novel method that decouples back propagation gradient computation in artificial neural networks. The authors identify issues that result from long gradient flows in back propagation, and propose a method that tackles some issues, which include vanishing/exploding gradients, and update lock...
The paper provides a stochastic algorithm for a class of problem characterized by weak minty variational inequality. The algorithm modifies stochastic extra-gradient by adding a bias-correction term in the exploration step. Strength: (1) The design of the algorithm is interesting, as it only uses diminishing stepsize...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper provides a stochastic algorithm for a class of problem characterized by weak minty variational inequality. The algorithm modifies stochastic extra-gradient by adding a bias-correction term in the exploration step. Strength: (1) The design of the algorithm is interesting, as it only uses diminishing ...
This paper aims to address the problem of robust overfitting in the adversarial training, where the robust model overfits to the adversarial examples. Prior works address this by assigning different weight to different instances according to their importance. However, this has not been well extended to multi-class clas...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper aims to address the problem of robust overfitting in the adversarial training, where the robust model overfits to the adversarial examples. Prior works address this by assigning different weight to different instances according to their importance. However, this has not been well extended to multi-cl...
The paper proposes a new zero-cost proxy (called ZiCo) for neural architecture search motivated by theoretical insights about the relationship of statistics of gradient across different input samples (absolute mean and standard deviation) and a network's converge speed. The proposed proxy is evaluated by investigating ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a new zero-cost proxy (called ZiCo) for neural architecture search motivated by theoretical insights about the relationship of statistics of gradient across different input samples (absolute mean and standard deviation) and a network's converge speed. The proposed proxy is evaluated by invest...
The paper proposes deep nearest centroids, a new non-parametric classification layer. For each class, a number of sub-centroids are learned during the course of training. At test-time, the prediction is the class corresponding to the closest sub-centroid. The sub-centroids are learned in an iterative fashion: 1) Networ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes deep nearest centroids, a new non-parametric classification layer. For each class, a number of sub-centroids are learned during the course of training. At test-time, the prediction is the class corresponding to the closest sub-centroid. The sub-centroids are learned in an iterative fashion: 1...
The paper presents a framework that allows using any reinforcement learning (RL) algorithm within a population of agents. The contribution is to use quality diversity methods to evolve populations and maintain their diversity. The most commonly used method is MAP-ELITES, but MAP-ELITES does not work well on high-dimens...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper presents a framework that allows using any reinforcement learning (RL) algorithm within a population of agents. The contribution is to use quality diversity methods to evolve populations and maintain their diversity. The most commonly used method is MAP-ELITES, but MAP-ELITES does not work well on hig...
The paper proposes an feature compensation method by two MLPs that computes the distance of data points to data centers. .The method is evaluated on a set of benchmark datasets for transfer learning from ImageNet, showing somewhat on par result using this technique. Strengths: the paper is well written, methodology is ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes an feature compensation method by two MLPs that computes the distance of data points to data centers. .The method is evaluated on a set of benchmark datasets for transfer learning from ImageNet, showing somewhat on par result using this technique. Strengths: the paper is well written, methodo...
This paper proposes a semi-supervised learning (SSL) method that pre-trains on both SEEG and EEG data. It learns the correlation graph between channels from three SSL tasks: instantaneous time shift task, delayed time shift task, and replace discriminative task. Experiments on seizure detection is described with promis...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a semi-supervised learning (SSL) method that pre-trains on both SEEG and EEG data. It learns the correlation graph between channels from three SSL tasks: instantaneous time shift task, delayed time shift task, and replace discriminative task. Experiments on seizure detection is described wit...
This submission provides a theoretical result of a lower bound of adversarial risk when data contains noise. Experiments on MNIST and CIFAR10 also demonstrate the correctness of the fact that uniform label noise is more harmful than typical real-world label noise. The authors also show the connection between inductive ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This submission provides a theoretical result of a lower bound of adversarial risk when data contains noise. Experiments on MNIST and CIFAR10 also demonstrate the correctness of the fact that uniform label noise is more harmful than typical real-world label noise. The authors also show the connection between in...
Deep learning has become an important method for tabular data. Even though pretraining based DL has been demonstrated successful in computer vision and NLP problems, it is challenging to make reliable conclusion about pretraining efficacy in tabular DL. This work provides a fully labeled tabular datasets to understand ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Deep learning has become an important method for tabular data. Even though pretraining based DL has been demonstrated successful in computer vision and NLP problems, it is challenging to make reliable conclusion about pretraining efficacy in tabular DL. This work provides a fully labeled tabular datasets to und...
The paper focuses on the ad-hoc teaming aspect of coordination. Unlike most previous works it focuses on coordination of different groups of agents rather than a single-agent with different teams. Moreover they show adaption ability in the context of within an episode as well in various multi-agent cooperative scenario...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper focuses on the ad-hoc teaming aspect of coordination. Unlike most previous works it focuses on coordination of different groups of agents rather than a single-agent with different teams. Moreover they show adaption ability in the context of within an episode as well in various multi-agent cooperative ...
The paper outlines an approach to use successor representations to drive a discrete state space abstraction. The abstract states are clusters in successor space, so "nearby" states have similar successors. The paper also contributes a way to utilize the abstraction learned to solve tasks efficiently by interpreting tra...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper outlines an approach to use successor representations to drive a discrete state space abstraction. The abstract states are clusters in successor space, so "nearby" states have similar successors. The paper also contributes a way to utilize the abstraction learned to solve tasks efficiently by interpre...
Contrastive Learning (CL) has become popular for both unsupervised and fully supervised settings. Recent empirical studies have indicated that CL can also be beneficial under weakly (ex noisy label) or semi-supervised settings. This paper studies CL under weakly supervised setting and proposed an approach Mix where t...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: Contrastive Learning (CL) has become popular for both unsupervised and fully supervised settings. Recent empirical studies have indicated that CL can also be beneficial under weakly (ex noisy label) or semi-supervised settings. This paper studies CL under weakly supervised setting and proposed an approach Mix...
This paper proposes a model for detecting true hate (vs. sarcasm) speech in texts, i.e., hate speech detection taking into account sarcasm. The proposed model uses game theory (Prisoners’ Dilemma) and Nash equilibrium. The experiments show that the proposed model outperforms baselines and state-of-the-art models. str...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a model for detecting true hate (vs. sarcasm) speech in texts, i.e., hate speech detection taking into account sarcasm. The proposed model uses game theory (Prisoners’ Dilemma) and Nash equilibrium. The experiments show that the proposed model outperforms baselines and state-of-the-art mode...
In this paper, the authors analyze the characteristics of spatial information in document understanding and propose a framework to leverage spatial information for document OOD detection. They start from an analysis on the effects across different OOD types and emphasize the importance of spatial information for docume...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this paper, the authors analyze the characteristics of spatial information in document understanding and propose a framework to leverage spatial information for document OOD detection. They start from an analysis on the effects across different OOD types and emphasize the importance of spatial information fo...
This paper introduces a contrastive learning-based method for learning video representation. The core idea is to use the temporal order of video segments and sentences to align the temporal-sensitive representation. Meanwhile, a negative sampling strategy based on temporal granularity and shuffling the units in the pos...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces a contrastive learning-based method for learning video representation. The core idea is to use the temporal order of video segments and sentences to align the temporal-sensitive representation. Meanwhile, a negative sampling strategy based on temporal granularity and shuffling the units in...
This paper proposes a FL framework named F2L. It combines hierarchical network design and knowledge distillation to solve the problem of non-IID of data. Its contributions are threefold:(1) it shows the traditional FL algorithms are unstable for non-IID data. (2) it proposes a new knowledge distillation algorithm Label...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a FL framework named F2L. It combines hierarchical network design and knowledge distillation to solve the problem of non-IID of data. Its contributions are threefold:(1) it shows the traditional FL algorithms are unstable for non-IID data. (2) it proposes a new knowledge distillation algorit...
This paper presents the Hardware-friendly regrouping towards block-based pruning (HRBP) method, and the HRBP++ method. It tests the method on CIFAR-10, CIFAR-100 and ImageNet dataset. Strength: 1. The network pruning is an important topic, and the author has given sufficient background introduction. 2. The paper illu...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents the Hardware-friendly regrouping towards block-based pruning (HRBP) method, and the HRBP++ method. It tests the method on CIFAR-10, CIFAR-100 and ImageNet dataset. Strength: 1. The network pruning is an important topic, and the author has given sufficient background introduction. 2. The pa...
This work considers the problem of finding optimal activation functions for a random features regression setting under a performance vs. sensitivity trade-off. Under the same setup as in [Mei, Montanari '22] (spherical input data, linear target function + noise, Gaussian random features projection), the starting point ...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work considers the problem of finding optimal activation functions for a random features regression setting under a performance vs. sensitivity trade-off. Under the same setup as in [Mei, Montanari '22] (spherical input data, linear target function + noise, Gaussian random features projection), the startin...
This paper proposes a method for training normalizing flows using a mass-covering objective which does not require samples from the target distribution. Existing methods often optimize a mode-seeking objective when target samples are not available, although the experiments section highlights applications (such a lear...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a method for training normalizing flows using a mass-covering objective which does not require samples from the target distribution. Existing methods often optimize a mode-seeking objective when target samples are not available, although the experiments section highlights applications (suc...