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 work present NORM, a two-stage feature distillation method. It relies on a linear feature transform module that is inserted after the last convolutional layer of the student network to enable a many-to-one representation matching mechanism conditioned on a single teacher-student layer pair via feature expansion, ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work present NORM, a two-stage feature distillation method. It relies on a linear feature transform module that is inserted after the last convolutional layer of the student network to enable a many-to-one representation matching mechanism conditioned on a single teacher-student layer pair via feature exp...
The paper studies the problem of representation learning of the analog circuit. It propose a two-level GNN model, in which the outer graph is composed with multiple non overlapping subgraphs. The representation can be used to perform different tasks. Then it also present a dataset called Open Circuit Benchmark which co...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper studies the problem of representation learning of the analog circuit. It propose a two-level GNN model, in which the outer graph is composed with multiple non overlapping subgraphs. The representation can be used to perform different tasks. Then it also present a dataset called Open Circuit Benchmark ...
This work proposes a new exploration objective in meta-reinforcement learning to discover clusters of related tasks. The claim is that tasks may be heterogeneous, where knowledge may only be transferable to other tasks within the same cluster, so exploring to identify these clusters could find the useful transferable i...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work proposes a new exploration objective in meta-reinforcement learning to discover clusters of related tasks. The claim is that tasks may be heterogeneous, where knowledge may only be transferable to other tasks within the same cluster, so exploring to identify these clusters could find the useful transf...
Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry fo...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. ...
This paper propose a new model architecture and training method to classify digital pathology whole-slide images (WSI). These images are in the giga-pixel range and cannot be classified directly by a model. The paper proposes a multi-instance learning approach with contrastive self-supervised learning for patch level f...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper propose a new model architecture and training method to classify digital pathology whole-slide images (WSI). These images are in the giga-pixel range and cannot be classified directly by a model. The paper proposes a multi-instance learning approach with contrastive self-supervised learning for patch...
While previous work has established that adversarial robust models lead to perceptually-aligned gradients, this paper tries the answer if the reverse condition also holds that if a model has perceptually-aligned gradients, is it robust to adversarial examples? The answer given by this paper is yes. The crux of the appr...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: While previous work has established that adversarial robust models lead to perceptually-aligned gradients, this paper tries the answer if the reverse condition also holds that if a model has perceptually-aligned gradients, is it robust to adversarial examples? The answer given by this paper is yes. The crux of ...
The paper presents an isometry aware training approach that improves robustness. The main intuition is that if the latent representations (pre final layer in this case), are (locally) Lipschitz, the network will be robust to adversarial perturbations. The algorithm enforces this Lipschitz constraint using a regularizer...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper presents an isometry aware training approach that improves robustness. The main intuition is that if the latent representations (pre final layer in this case), are (locally) Lipschitz, the network will be robust to adversarial perturbations. The algorithm enforces this Lipschitz constraint using a reg...
The paper explores intrinsic dimension estimation of the data using VAE. They found that this estimation can be made after a few steps of training and proposed a method FONDUE to provide a more principled method for selecting latent dimensions for the VAE. Strength: 1. A thorough experiments across a lot of settings fo...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: The paper explores intrinsic dimension estimation of the data using VAE. They found that this estimation can be made after a few steps of training and proposed a method FONDUE to provide a more principled method for selecting latent dimensions for the VAE. Strength: 1. A thorough experiments across a lot of set...
The paper presents an initialization strategy for Koopman operators and a loss regularization in terms of spectral information for Koopman autoencoders. For initialisations, the paper claims that approximated Koopman operators should be initialized such that their spectral norm should be less than 1. For regularization...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper presents an initialization strategy for Koopman operators and a loss regularization in terms of spectral information for Koopman autoencoders. For initialisations, the paper claims that approximated Koopman operators should be initialized such that their spectral norm should be less than 1. For regula...
In this paper, the authors propose to extend the idea of interval bounds from the provably robust training literature to few-shot classification and introduce task-level interval bounds based regularization for the training of few-shot classification models (MAML + protonets). To further improve the model robustness u...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors propose to extend the idea of interval bounds from the provably robust training literature to few-shot classification and introduce task-level interval bounds based regularization for the training of few-shot classification models (MAML + protonets). To further improve the model robu...
## The problem The paper introduces an asynchronous aggregation scheme to address the straggler problem and also taking in the effect of non-iid. The paper focuses on "heavyweight" learning task such as video action recognition that are out of reach for mobile devices processing power. Overall, the paper aims to addre...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: ## The problem The paper introduces an asynchronous aggregation scheme to address the straggler problem and also taking in the effect of non-iid. The paper focuses on "heavyweight" learning task such as video action recognition that are out of reach for mobile devices processing power. Overall, the paper aims ...
This paper describes a tranformer-based architecture for policies for multi-agent (video) games, that tokenises image-based inputs plus entity-component-system-based inputs (representing data of entities in video games), allowing the architecture to process inputs from various different games in a single and consistent...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper describes a tranformer-based architecture for policies for multi-agent (video) games, that tokenises image-based inputs plus entity-component-system-based inputs (representing data of entities in video games), allowing the architecture to process inputs from various different games in a single and co...
This paper investigates learning a task representation space for use in unsupervised environment design (UED), with an algorithm called PAIRED. PAIRED trains an RL agent, the adversary, to configure the environment parameters throughout training, in tandem with two RL agents that learn to solve these configurations (ea...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper investigates learning a task representation space for use in unsupervised environment design (UED), with an algorithm called PAIRED. PAIRED trains an RL agent, the adversary, to configure the environment parameters throughout training, in tandem with two RL agents that learn to solve these configurat...
This proposes to solve online combinatorial optimization problems using reinforcement learning composed of latent MDPs and curriculum learning. The paper presents formal formulations and theoretical results regarding the performance bound based on the relative condition number. In addition, the paper also presents theo...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This proposes to solve online combinatorial optimization problems using reinforcement learning composed of latent MDPs and curriculum learning. The paper presents formal formulations and theoretical results regarding the performance bound based on the relative condition number. In addition, the paper also prese...
This paper explores how CLIP can be used effectively for multi-source domain adaptation. In this problem setting there is a single labeled source domain dataset and multiple target domain datasets that are not labeled. Several methods have been proposed in past to address this problem, but they all have limited accurac...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper explores how CLIP can be used effectively for multi-source domain adaptation. In this problem setting there is a single labeled source domain dataset and multiple target domain datasets that are not labeled. Several methods have been proposed in past to address this problem, but they all have limited...
The submission explores information-theoretically motivated objectives for self-supervised learning (SSL). The submission derives a bound on an objective to maximize the mutual information between neural network inputs and outputs using assumptions about Gaussianity of the input distribution and a spline framing of ne...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The submission explores information-theoretically motivated objectives for self-supervised learning (SSL). The submission derives a bound on an objective to maximize the mutual information between neural network inputs and outputs using assumptions about Gaussianity of the input distribution and a spline frami...
This paper studies coresets for robust clustering. A coreset is a small proxy of the dataset, which is yet sufficient to perform the task at hand while not significantly compromising the quality of the solution. The goal of the robust clustering task is to cluster a set of data points that contains a number of adversar...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies coresets for robust clustering. A coreset is a small proxy of the dataset, which is yet sufficient to perform the task at hand while not significantly compromising the quality of the solution. The goal of the robust clustering task is to cluster a set of data points that contains a number of ...
This paper identifies an issue they call "over-conservatism" with adversarial training methods, which causes more conservative behavior than is desirable. This is adressed by instead optimizing for a mixture of average case and worst case performance. Strengths: * The conservatism (If I've understood it) that they id...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper identifies an issue they call "over-conservatism" with adversarial training methods, which causes more conservative behavior than is desirable. This is adressed by instead optimizing for a mixture of average case and worst case performance. Strengths: * The conservatism (If I've understood it) that...
The paper considers training dynamics of simple neural networks on contrastive learning tasks, and studies the effect of non-linearities such as the ReLU on the resulting optima. In the setting considered, the authors find that linear activations lead to simple solutions involving leading eigenvectors of some fixed cov...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper considers training dynamics of simple neural networks on contrastive learning tasks, and studies the effect of non-linearities such as the ReLU on the resulting optima. In the setting considered, the authors find that linear activations lead to simple solutions involving leading eigenvectors of some f...
This paper considers statistical inference in linear Fisher markets. In their framework, a market formed by a finite number of items sampled from an underlying distribution is observed and their goal is to infer several important equilibrium quantities (individual utilities, pacing multipliers, and social welfare) of t...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers statistical inference in linear Fisher markets. In their framework, a market formed by a finite number of items sampled from an underlying distribution is observed and their goal is to infer several important equilibrium quantities (individual utilities, pacing multipliers, and social welfa...
This paper presents a recurrent convolutional architecture (LocRNN) to improve extrapolation in visual tasks. The novelty of the architecture is presented as being long-range lateral connections between neurons from the same layer. The architecture is tested on two tasks: Mazes (route segmentation) and PathFinder (curv...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper presents a recurrent convolutional architecture (LocRNN) to improve extrapolation in visual tasks. The novelty of the architecture is presented as being long-range lateral connections between neurons from the same layer. The architecture is tested on two tasks: Mazes (route segmentation) and PathFind...
- a benchmark for retrosynthesis is proposed - a model for retrosynthesis is presented - the authors claim state of the art there are only limited novel contributions in this paper ### strengths - clear description - interesting memory model ### weaknesses: - the paper has several incorrect definitions - the paper ye...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: - a benchmark for retrosynthesis is proposed - a model for retrosynthesis is presented - the authors claim state of the art there are only limited novel contributions in this paper ### strengths - clear description - interesting memory model ### weaknesses: - the paper has several incorrect definitions - the ...
This paper describes a learning and inference algorithm for deep Gaussian processes based on operator variational inference (OVI). Particularly, OVI (Ranganath et al., 2016, Hu et al. 2018, Grathwohl et al. 2020) generalizes Kernelized Stein Discrepancy (KSD) to situations where the RKHS space the smooth function (of...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper describes a learning and inference algorithm for deep Gaussian processes based on operator variational inference (OVI). Particularly, OVI (Ranganath et al., 2016, Hu et al. 2018, Grathwohl et al. 2020) generalizes Kernelized Stein Discrepancy (KSD) to situations where the RKHS space the smooth func...
This paper tackles the vertical Federated Learning (FL) scenario for graph-structured data. In particular, the authors first propose to update the node features calculated from every layer of Graph Neural Network (GNNs), distributed to multiple clients, simultaneously (i.e., at l-th layer, all clients share their node ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper tackles the vertical Federated Learning (FL) scenario for graph-structured data. In particular, the authors first propose to update the node features calculated from every layer of Graph Neural Network (GNNs), distributed to multiple clients, simultaneously (i.e., at l-th layer, all clients share the...
This is primarily a theoretical paper addressing the problem of solving differential equations with neural networks. Existing approaches do not provide any guarantees about their error relative to the true solution. Specifically, optimising the loss function doesn’t bound the error of the solution. Section 2 of the pa...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This is primarily a theoretical paper addressing the problem of solving differential equations with neural networks. Existing approaches do not provide any guarantees about their error relative to the true solution. Specifically, optimising the loss function doesn’t bound the error of the solution. Section 2 o...
The paper studies offline RL problems with perturbed rewards. In particular, the Q function will be parametrized as the minimum of $M$ neural networks, trained on $M$ perturbed datasets. The benefit of the proposed algorithm, compared to the UCB-based method is reducing the time complexity of action selection. On a tec...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies offline RL problems with perturbed rewards. In particular, the Q function will be parametrized as the minimum of $M$ neural networks, trained on $M$ perturbed datasets. The benefit of the proposed algorithm, compared to the UCB-based method is reducing the time complexity of action selection. ...
To address the problem of the anisotropy issue (i.e., hidden representations of transformer models are squeezed into a tiny cone), this paper proposes to use contrastive learning for language generation (henceforth, ContraGen). The model was evaluated on both text generation and code generation. It was also assessed on...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: To address the problem of the anisotropy issue (i.e., hidden representations of transformer models are squeezed into a tiny cone), this paper proposes to use contrastive learning for language generation (henceforth, ContraGen). The model was evaluated on both text generation and code generation. It was also ass...
This paper refines the study of linear regions in neural networks by analyzing how each linear region can be decomposed into simplices. The authors argue through theoretical results and experiments that most linear regions can be decomposed into very few simplices; hence implying in yet another form that neural network...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper refines the study of linear regions in neural networks by analyzing how each linear region can be decomposed into simplices. The authors argue through theoretical results and experiments that most linear regions can be decomposed into very few simplices; hence implying in yet another form that neural...
This work considers the crowdsourcing problems where there exist ambiguous tasks. To address this problem, a new framework that used a similar technique in AUM method to identify ambiguous tasks has been proposed. To show the effectiveness of the proposed method, evaluations on both the synthetic datasets and the CIFAR...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This work considers the crowdsourcing problems where there exist ambiguous tasks. To address this problem, a new framework that used a similar technique in AUM method to identify ambiguous tasks has been proposed. To show the effectiveness of the proposed method, evaluations on both the synthetic datasets and t...
In this paper, the authors investigate whether or not pretrained language models like BERT have the ability to maintain previously learned knowledge in the long term. To do this, they track the encoding ability of BERT for specific tasks before, during, and after learning new tasks. They find that BERT can actually ref...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors investigate whether or not pretrained language models like BERT have the ability to maintain previously learned knowledge in the long term. To do this, they track the encoding ability of BERT for specific tasks before, during, and after learning new tasks. They find that BERT can actu...
The authors propose a method based on the Past Motion Dropout hyperparameter from ChaufferNet, the behavioral cloning approach to autonomous vehicle motion planning. With the same motivation as ChaufferNet authors, they suggest to set the dropout parameter to 100%, such that neural network policy does not have access t...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose a method based on the Past Motion Dropout hyperparameter from ChaufferNet, the behavioral cloning approach to autonomous vehicle motion planning. With the same motivation as ChaufferNet authors, they suggest to set the dropout parameter to 100%, such that neural network policy does not have ...
This paper proposes a universal vision-language dense retrieval with two techniques, using modality-balanced hard negatives for optimization and bridging the modality gap by the image verbalization method. Experiments are conducted on the built open-domain dataset from WebQA and compared to existing models for single-...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a universal vision-language dense retrieval with two techniques, using modality-balanced hard negatives for optimization and bridging the modality gap by the image verbalization method. Experiments are conducted on the built open-domain dataset from WebQA and compared to existing models for...
This paper proposed `CrossFormer`, a Transformer based neural network for multivariate time series (MTS) forecasting. The author(s) claimed that cross-dimension dependency was not well utilized and developed a method to explicitly explore and utilize cross-dimension dependency for MTS forecasting. In order to achieve t...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed `CrossFormer`, a Transformer based neural network for multivariate time series (MTS) forecasting. The author(s) claimed that cross-dimension dependency was not well utilized and developed a method to explicitly explore and utilize cross-dimension dependency for MTS forecasting. In order to a...
This manuscript proposes a binary classifier that determines if an environment is used in the training of an agent, by inspecting $n$ trajectories that are generated from that environment. The manuscript claims this to be an attack toward a value function of an RL agent. The manuscript subsequently proposes to perturb ...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This manuscript proposes a binary classifier that determines if an environment is used in the training of an agent, by inspecting $n$ trajectories that are generated from that environment. The manuscript claims this to be an attack toward a value function of an RL agent. The manuscript subsequently proposes to ...
This paper offers a straightforward extension of the exploration options framework (Bagot et al., 2020)—which was initially introduced as a prototype for tabular environments—to a deep-learning, function approximation setting. The method is empirically evaluated in the Atari Learning Environment, following a standardiz...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper offers a straightforward extension of the exploration options framework (Bagot et al., 2020)—which was initially introduced as a prototype for tabular environments—to a deep-learning, function approximation setting. The method is empirically evaluated in the Atari Learning Environment, following a st...
This paper proposes Multi-Agent Joint-Predictive Representations (MAJOR), which applys the self-supervised learning (SSL) technique to MARL, trying to improve the learning efficiency. Specifically, MAJOR builds a transformer-based transition model, which takes all agents observations and actions as inputs and predicts ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes Multi-Agent Joint-Predictive Representations (MAJOR), which applys the self-supervised learning (SSL) technique to MARL, trying to improve the learning efficiency. Specifically, MAJOR builds a transformer-based transition model, which takes all agents observations and actions as inputs and p...
The paper studies communication in multiagent reinforcement learning that preserves privacy. It adopts the concept of differential privacy to approach the problem and proposes an algorithm that is able to protect individual agents' information privacy. The authors also use a stochastic message sender for each agent to ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies communication in multiagent reinforcement learning that preserves privacy. It adopts the concept of differential privacy to approach the problem and proposes an algorithm that is able to protect individual agents' information privacy. The authors also use a stochastic message sender for each a...
A major bottleneck in large-scale distributed deep learning is the communication bottleneck. The computational cost of the majority of existing compression algorithms to reduce the communication cost is too high. The authors proposed a new distributed optimization algorithm, Binary SGD-Momentum which 1- compresses the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: A major bottleneck in large-scale distributed deep learning is the communication bottleneck. The computational cost of the majority of existing compression algorithms to reduce the communication cost is too high. The authors proposed a new distributed optimization algorithm, Binary SGD-Momentum which 1- compres...
This work tackles the challenges of lifelong reinforcement learning. In particular, authors propose an algorithm (UCBvld) for solving sequential contextual Markov decision processes with linear representation which (1) guarantees sublinear regret with (2) a sublinear number of planning calls, even when the sequence of ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work tackles the challenges of lifelong reinforcement learning. In particular, authors propose an algorithm (UCBvld) for solving sequential contextual Markov decision processes with linear representation which (1) guarantees sublinear regret with (2) a sublinear number of planning calls, even when the sequ...
This work cautions the application of CKA similarity measure in comparing neural representations in practice and points out a few issues including 1) its sensitivity to the family of the subset translations where certain transformations do not change the functional behaviors of the model; 2) its value can be directly m...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work cautions the application of CKA similarity measure in comparing neural representations in practice and points out a few issues including 1) its sensitivity to the family of the subset translations where certain transformations do not change the functional behaviors of the model; 2) its value can be di...