Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

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

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
The authors study mini-batch stochastic iterative hard thresholding (IHT). IHT is a well known algorithm for sparse optimization with extensive literature. IHT is essentially $\ell_2$-projected gradient descent on the set of s-sparse vectors. Given the ML-wide focus on stochastic optimization algorithms, it is natural ...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors study mini-batch stochastic iterative hard thresholding (IHT). IHT is a well known algorithm for sparse optimization with extensive literature. IHT is essentially $\ell_2$-projected gradient descent on the set of s-sparse vectors. Given the ML-wide focus on stochastic optimization algorithms, it is ...
The authors study from a theoretical point of view the extrapolation capabilities of linear RNNs in the teacher-student framework. They show that under certain conditions, students with 0 error may not extrapolate to longer time horizons than the ones they were trained on. They also show, that under the right parameter...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors study from a theoretical point of view the extrapolation capabilities of linear RNNs in the teacher-student framework. They show that under certain conditions, students with 0 error may not extrapolate to longer time horizons than the ones they were trained on. They also show, that under the right p...
This paper prooses the novel concept of populated region set, and make connection between the proposed PRS ratio and the adversarial robustness of the model. The paper further propose PRS regualrization that can improve adversarial robustness without adversarial training. ## Strength 1. This paper provides novel concep...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper prooses the novel concept of populated region set, and make connection between the proposed PRS ratio and the adversarial robustness of the model. The paper further propose PRS regualrization that can improve adversarial robustness without adversarial training. ## Strength 1. This paper provides nove...
This work studies a novel and interesting problem: muscular control with overactuated action spaces. It first demonstrates that the main issue with overactuated action spaces is how to effectively and efficiently perform exploration using a clear torquearm example with 2 and 600 DoF theoretically and empirically. It fu...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work studies a novel and interesting problem: muscular control with overactuated action spaces. It first demonstrates that the main issue with overactuated action spaces is how to effectively and efficiently perform exploration using a clear torquearm example with 2 and 600 DoF theoretically and empiricall...
This paper proposes an approach for Offline RL by explicitly modeling risk through CVaR. The paper targets stochastic environments and makes the case that modeling risk is essential for performing well in these settings by demonstrating improved performance in standard control tasks with noise injection in the rewards...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes an approach for Offline RL by explicitly modeling risk through CVaR. The paper targets stochastic environments and makes the case that modeling risk is essential for performing well in these settings by demonstrating improved performance in standard control tasks with noise injection in the...
Motivated by the ’flash-to-bang' phenomenon, in this paper, the authors propose a new audio-visual learning model for sound source depth estimation. In particular, the authors formulate the sound source depth estimation as an audio-visual collision event localization task. To solve the task and increase depth estimatio...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Motivated by the ’flash-to-bang' phenomenon, in this paper, the authors propose a new audio-visual learning model for sound source depth estimation. In particular, the authors formulate the sound source depth estimation as an audio-visual collision event localization task. To solve the task and increase depth e...
This paper focuses on sparse training, which aims to reduce the computational overhead of deep neural networks. Specifically, this work proposes a non-linear gradient-based method, namely, Gradient Annealing (GA), to address the trade-off between model sparsity and accuracy. Meanwhile, this paper combines one latest sp...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper focuses on sparse training, which aims to reduce the computational overhead of deep neural networks. Specifically, this work proposes a non-linear gradient-based method, namely, Gradient Annealing (GA), to address the trade-off between model sparsity and accuracy. Meanwhile, this paper combines one l...
This paper proposes to use superpixel-based tokens instead of fixed-shape patch tokens for image segmentation, and proposes a graph pooling between transformer blocks to create an increasing size of segmentation from the oversegmented superpixel map. Strength: 1. The motivations to use adaptive segment tokens are natur...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to use superpixel-based tokens instead of fixed-shape patch tokens for image segmentation, and proposes a graph pooling between transformer blocks to create an increasing size of segmentation from the oversegmented superpixel map. Strength: 1. The motivations to use adaptive segment tokens a...
This paper describes an adversarial imitation learning algorithm that learns a discriminator with patch pixel rewards rather than directly from images. Recent work in the vision deep learning community have shown that using patches is effective in embedding pixels. The authors argue that patch pixels rewards are more i...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper describes an adversarial imitation learning algorithm that learns a discriminator with patch pixel rewards rather than directly from images. Recent work in the vision deep learning community have shown that using patches is effective in embedding pixels. The authors argue that patch pixels rewards ar...
-This paper studies the problem of federated learning with semi-supervised client (local) data. They propose a method that first pre-trains the client models with their labeled training data, then performs cross-client label propagation based on the embeddings of the labeled and unlabeled training data obtained from th...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: -This paper studies the problem of federated learning with semi-supervised client (local) data. They propose a method that first pre-trains the client models with their labeled training data, then performs cross-client label propagation based on the embeddings of the labeled and unlabeled training data obtained...
This paper proposes an object-centric learning approach that scales to real-world data. The method builds on Slot Attention, but instead of reconstructing pixels, it reconstructs image features obtained by an encoder (e.g. ViT) pre-trained with self-supervision (DINO). Compared to standard Slot Attention, the paper sh...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes an object-centric learning approach that scales to real-world data. The method builds on Slot Attention, but instead of reconstructing pixels, it reconstructs image features obtained by an encoder (e.g. ViT) pre-trained with self-supervision (DINO). Compared to standard Slot Attention, the ...
This paper proposes to combine a measure of the difficulty of an example in a supervised classification task (AUM) with the performance of annotators, calculated from their confusion matrix on a crowd-sourcing task. This combined estimator, called WAUM (weighted AUM), allows to take into account both the intrinsect dif...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes to combine a measure of the difficulty of an example in a supervised classification task (AUM) with the performance of annotators, calculated from their confusion matrix on a crowd-sourcing task. This combined estimator, called WAUM (weighted AUM), allows to take into account both the intrin...
The current study, as a strategy to effectively solve large-scale CVRP, presents a strategy for generating TSP sub-problems by clustering nodes and simultaneously solving the generated TSP problems. In particular, the sequential node selection approach has been used to decompose nodes. The decomposition strategy is a m...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The current study, as a strategy to effectively solve large-scale CVRP, presents a strategy for generating TSP sub-problems by clustering nodes and simultaneously solving the generated TSP problems. In particular, the sequential node selection approach has been used to decompose nodes. The decomposition strateg...
This paper proposes MULTIVIZ which is a method to analyze the workings of multi-modal models. The authors propose to break down the visualization / analysis problem into four primary pillars, namely: unimodal importance: how each modality contributes to the downstream tasks cross-modal interactions: how the multiple m...	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 proposes MULTIVIZ which is a method to analyze the workings of multi-modal models. The authors propose to break down the visualization / analysis problem into four primary pillars, namely: unimodal importance: how each modality contributes to the downstream tasks cross-modal interactions: how the mu...
The paper investigates the latent separability of the existing backdoor attacks, which makes them easily caught by existing defenses that look at the latent space. The paper then crafts an attack using data poisoning without latent separability. The proposed adaptive backdoor attack method uses regularized samples (sam...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper investigates the latent separability of the existing backdoor attacks, which makes them easily caught by existing defenses that look at the latent space. The paper then crafts an attack using data poisoning without latent separability. The proposed adaptive backdoor attack method uses regularized samp...
This paper provides a new method, WinIT, for interpreting predictions of multivariate time series. It provides an importance score for each feature in a multivariate time series according to explainability, considering the dependencies of features between time steps and variation in feature importance over time. The me...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper provides a new method, WinIT, for interpreting predictions of multivariate time series. It provides an importance score for each feature in a multivariate time series according to explainability, considering the dependencies of features between time steps and variation in feature importance over time...
This paper focuses on text-supervised semantic segmentation by studying a multi-view consistency learning framework. Specifically, a text-to-views consistency and a cross-view segmentation consistency training strategy are proposed and added upon existing work Group VIT. The proposed method tries to construct more trai...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper focuses on text-supervised semantic segmentation by studying a multi-view consistency learning framework. Specifically, a text-to-views consistency and a cross-view segmentation consistency training strategy are proposed and added upon existing work Group VIT. The proposed method tries to construct m...
The manuscript proposes a novel multi-channel self-supervised framework for SEEG and EEG data. The framework comprises delayed-time-shift prediction, instantaneous time shift, replacement discriminative tasks, and an additional graph module. The results show improvements over previous related work and supervised models...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The manuscript proposes a novel multi-channel self-supervised framework for SEEG and EEG data. The framework comprises delayed-time-shift prediction, instantaneous time shift, replacement discriminative tasks, and an additional graph module. The results show improvements over previous related work and supervise...
This paper proposes a novel curriculum-based co-design method for evolving soft voxel-based robots. The main contributions are: - A novel curriculum mechanism that effectively evolves robots from simple to large design spaces. - A transformer-based control policy representation, enabling adaptation to arbitrary dimensi...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a novel curriculum-based co-design method for evolving soft voxel-based robots. The main contributions are: - A novel curriculum mechanism that effectively evolves robots from simple to large design spaces. - A transformer-based control policy representation, enabling adaptation to arbitrary...
The paper presents a discussion on the definition of the in-distribution (ID) vs. out-of-distribution (OOD) for OOD detection problems, arguing that existing methods fail to semantically extrapolate the in-distribution (e.g. the ID of flying and sitting birds should include and generalize to various backgrounds, e.g. s...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper presents a discussion on the definition of the in-distribution (ID) vs. out-of-distribution (OOD) for OOD detection problems, arguing that existing methods fail to semantically extrapolate the in-distribution (e.g. the ID of flying and sitting birds should include and generalize to various backgrounds...
The authors propose low-precision model memory (LPMM), which quantizes a model's parameters, gradients, and optimizer state to low precision. Prior works have focused only on params/gradients or optimizer state separately. The main challenge is that params, gradients, and momentum may have very different dynamic range,...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose low-precision model memory (LPMM), which quantizes a model's parameters, gradients, and optimizer state to low precision. Prior works have focused only on params/gradients or optimizer state separately. The main challenge is that params, gradients, and momentum may have very different dynami...
This paper proposes to leverage visual attention as a diagnosis tool for reflecting the reason for error during models' task execution. The authors leverage a 2D language navigation task as the test bed and tested mainly using goal-conditioned imitation learning methods. They visualize attention on the 2D map to show w...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes to leverage visual attention as a diagnosis tool for reflecting the reason for error during models' task execution. The authors leverage a 2D language navigation task as the test bed and tested mainly using goal-conditioned imitation learning methods. They visualize attention on the 2D map t...
This paper proposes adding topic model upon pretrained language models (PLMs), in order to increase its interpretability and downstream performance. More specifically, the authors design CLDA, a new LDA model compatible with continuous word counts, which are computed via aggregated attention scores of the last layer in...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes adding topic model upon pretrained language models (PLMs), in order to increase its interpretability and downstream performance. More specifically, the authors design CLDA, a new LDA model compatible with continuous word counts, which are computed via aggregated attention scores of the last ...
The paper proposes to use multi-objective optimization to generate adversarial examples for model ensembles, adversarial examples against several transformations, and universal adversarial perturbation. Specifically, by setting the different models in the ensemble as different objectives, it turns the problem into mult...	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 paper proposes to use multi-objective optimization to generate adversarial examples for model ensembles, adversarial examples against several transformations, and universal adversarial perturbation. Specifically, by setting the different models in the ensemble as different objectives, it turns the problem i...
This paper introduces an approach for learning to perform high-quality long-term video prediction via an object-centric latent bottleneck. The approach hinges on two key ideas. The first is to use a pre-trained temporal object discovery method to extract a sequence of aligned slots. The second is to use an autoregressi...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This paper introduces an approach for learning to perform high-quality long-term video prediction via an object-centric latent bottleneck. The approach hinges on two key ideas. The first is to use a pre-trained temporal object discovery method to extract a sequence of aligned slots. The second is to use an auto...
This paper proposed a new ViT architecture called LipsFormer, which replaces non-Lipschitz or unstable modules with Lipschitz continuous counterparts. Experiments show that LipsFormer obtains better Top-1 accuracy on Imagenet-1k dataset compared with previous works. Strength: This paper analyzes the Lipschitz bound o...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposed a new ViT architecture called LipsFormer, which replaces non-Lipschitz or unstable modules with Lipschitz continuous counterparts. Experiments show that LipsFormer obtains better Top-1 accuracy on Imagenet-1k dataset compared with previous works. Strength: This paper analyzes the Lipschitz...
This paper studies the peculiar observation that, with very sample sizes, robust training can be detrimental for robust error. The authors show that in a classification task given by a linear model, the robust risk of the max-margin classier can increase as a function of the perturbation size used during robust trainin...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the peculiar observation that, with very sample sizes, robust training can be detrimental for robust error. The authors show that in a classification task given by a linear model, the robust risk of the max-margin classier can increase as a function of the perturbation size used during robust...
The paper proposes a new neural network interpretation and attribution framework which looks at heat flow to achieve multi-scale interpretation of a scalar-valued function Pros: - It is an interesting and novel idea to use the diffusion equation for the attribution problem - The paper is generally well written Cons: ...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes a new neural network interpretation and attribution framework which looks at heat flow to achieve multi-scale interpretation of a scalar-valued function Pros: - It is an interesting and novel idea to use the diffusion equation for the attribution problem - The paper is generally well written...
This paper proposes a hyper-network to solve partial differential equations. The strengths of this paper are: $\mathbf{1}.$ The novelty of separating time and space by the hyper-network architecture. $\mathbf{2}.$ This paper is clearly written and well organized, with detailed explanation of theories and experiment...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a hyper-network to solve partial differential equations. The strengths of this paper are: $\mathbf{1}.$ The novelty of separating time and space by the hyper-network architecture. $\mathbf{2}.$ This paper is clearly written and well organized, with detailed explanation of theories and ex...
This paper investigates how the order of batch normalization (BN) placed in the network affects the performance, when using the unbounded activation functions (e.g., Tanh). It shows that the Swap model (BN after the nonlinearity) using unbounded activation functions has significantly better performance than the convent...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper investigates how the order of batch normalization (BN) placed in the network affects the performance, when using the unbounded activation functions (e.g., Tanh). It shows that the Swap model (BN after the nonlinearity) using unbounded activation functions has significantly better performance than the...
I would like to begin by saying that this paper makes repeated and relentless reference to a given set of papers that induced this reviewer to suspect that it was self reference, and it took barely a minute of google search to verify this fact. I find this very disturbing, since the other papers I have reviewed have by...	Recommendation: 8: accept, good paper	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: I would like to begin by saying that this paper makes repeated and relentless reference to a given set of papers that induced this reviewer to suspect that it was self reference, and it took barely a minute of google search to verify this fact. I find this very disturbing, since the other papers I have reviewed...
Model ensembling is a common technology to further boost performance. While most paper study how to ensemble on supervised models, this paper focus on self-supervised models (SSL). Then, this paper discusses where to ensemble and how to ensemble the models training from SSL, proposing a downstream-efficient ensemble me...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: Model ensembling is a common technology to further boost performance. While most paper study how to ensemble on supervised models, this paper focus on self-supervised models (SSL). Then, this paper discusses where to ensemble and how to ensemble the models training from SSL, proposing a downstream-efficient ens...
The paper considers the task of unsupervised pretraining of the reinforcement learning (RL) agents. It analyses and compares different design choices for pretraining and fine-tuning the pretrained components; it shows the improvement of using proposed model-based reinforcement learning over existing model-free techniqu...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper considers the task of unsupervised pretraining of the reinforcement learning (RL) agents. It analyses and compares different design choices for pretraining and fine-tuning the pretrained components; it shows the improvement of using proposed model-based reinforcement learning over existing model-free ...
The paper presents a soft (continuous) prompt tuning method called MPT. In traditional soft prompt tuning, prompts are often sensitive to initialization when trained from scratch and performance may still lag behind full model fine tuning. In this work, the manuscript presents a method for multitask prompt tuning where...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents a soft (continuous) prompt tuning method called MPT. In traditional soft prompt tuning, prompts are often sensitive to initialization when trained from scratch and performance may still lag behind full model fine tuning. In this work, the manuscript presents a method for multitask prompt tuni...
The paper concerns itself with continual learning and observes that current methods are making assumptions on the schedule that can be unrealistic. In order to ameliorate this, the notion of \emph{schedule-robustness} is introduced. Intuitively schedule-robustness for CL means that the performance is "independent" of t...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper concerns itself with continual learning and observes that current methods are making assumptions on the schedule that can be unrealistic. In order to ameliorate this, the notion of \emph{schedule-robustness} is introduced. Intuitively schedule-robustness for CL means that the performance is "independe...
This paper studies the model-free linear constrained Markov Decision Process with infinite horizon average rewards. The goal is to learn a policy for selecting actions that minimize the regret (the difference between the maximum achievable average reward and the policy's total average reward) while keeping the total co...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the model-free linear constrained Markov Decision Process with infinite horizon average rewards. The goal is to learn a policy for selecting actions that minimize the regret (the difference between the maximum achievable average reward and the policy's total average reward) while keeping the ...
This paper proposes a new architecture that is useful for generalization to new tasks, that may be previous combination of prior tasks. It adapts from existing literature on successor features and GPI, and shows that a modified architecture that predicts useful representations while also learning representations can be...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a new architecture that is useful for generalization to new tasks, that may be previous combination of prior tasks. It adapts from existing literature on successor features and GPI, and shows that a modified architecture that predicts useful representations while also learning representation...
This paper proposed a method to train a network that can run at different depth at test time. The author claimed that a residual block could be considered to perform two functions (1) learning new features (2) refine features. The refinement stage/layers do not change feature semantics thus can be skipped. The proposed...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposed a method to train a network that can run at different depth at test time. The author claimed that a residual block could be considered to perform two functions (1) learning new features (2) refine features. The refinement stage/layers do not change feature semantics thus can be skipped. The ...
The authors propose a LDA-based counterfactual contrastive learning model for robust text classification and counterfactual data augmentation methods. To help readers better understand the motivations of the proposed model, the authors should not only state them by sentences but also with figures. E.g., the authors sta...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors propose a LDA-based counterfactual contrastive learning model for robust text classification and counterfactual data augmentation methods. To help readers better understand the motivations of the proposed model, the authors should not only state them by sentences but also with figures. E.g., the aut...
Summary: This paper presents a new Test-time adaptation (TTA) method called DELTA for debiased fully TTA. To be specific, the authors 1) introduce batch renormalization to alleviate the bias in normalization statistics, 2) propose dynamic online re-weighting (DOT) to address the class bias within optimization. **P...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Summary: This paper presents a new Test-time adaptation (TTA) method called DELTA for debiased fully TTA. To be specific, the authors 1) introduce batch renormalization to alleviate the bias in normalization statistics, 2) propose dynamic online re-weighting (DOT) to address the class bias within optimizat...