Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

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

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
This paper designed a simple yet effective adapter module that helps plain ViT behave well for dense prediction tasks, including object detection, instance segmentation, and semantic segmentation. Benefited by the inductive bias introduced by the spatial prior module and feature interaction module, the plain ViT archiv...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper designed a simple yet effective adapter module that helps plain ViT behave well for dense prediction tasks, including object detection, instance segmentation, and semantic segmentation. Benefited by the inductive bias introduced by the spatial prior module and feature interaction module, the plain Vi...
This paper tackles the problem of a classifier being biased to nuisance factors (shortcuts). The proposed method uses image-to-image translation to synthesize different biases in the dataset to remove the bias without collecting bias-free dataset. Strengths (+) The translation model does not require bias supervision f...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper tackles the problem of a classifier being biased to nuisance factors (shortcuts). The proposed method uses image-to-image translation to synthesize different biases in the dataset to remove the bias without collecting bias-free dataset. Strengths (+) The translation model does not require bias super...
The authors propose a technique for heteroschedastic classification, HET-XL, which does a simple trick to maintain scalability to settings with large amounts of classes. The key idea is deceptively simple: the authors elect to model the covariance over noise not over class-space in the logits, but over the empedding s...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose a technique for heteroschedastic classification, HET-XL, which does a simple trick to maintain scalability to settings with large amounts of classes. The key idea is deceptively simple: the authors elect to model the covariance over noise not over class-space in the logits, but over the emp...
The presented work establishes a theoretical link between standard diffusion models and Inverse heat dissipation models (IHDMs). They show that IHDM formulation can be equally represented as a standard diffusion model, albeit in the frequency domain. They use this theory to apparently get the best from both approaches:...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The presented work establishes a theoretical link between standard diffusion models and Inverse heat dissipation models (IHDMs). They show that IHDM formulation can be equally represented as a standard diffusion model, albeit in the frequency domain. They use this theory to apparently get the best from both app...
This paper proposes a large kernel ConvNets which scaled up the kernel size beyond 51x51 up to 61x61 upon the RepVGG-like ConvNet baseline. The authors progress the idea from ConvNeXt and show the difference between the recently published RepLKNet (CVPR 2022) with the proposed model by introducing the concept of dynami...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a large kernel ConvNets which scaled up the kernel size beyond 51x51 up to 61x61 upon the RepVGG-like ConvNet baseline. The authors progress the idea from ConvNeXt and show the difference between the recently published RepLKNet (CVPR 2022) with the proposed model by introducing the concept o...
The paper introduces the Skill Discovery Decision Transformer, an extension of Decision Transformer [1] in which that does not use the return-to-go for policy conditioning. Instead, individual states are clustered and conditioning is performed with a "skill-distribution-to-go". The general motivation is a relation to (...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces the Skill Discovery Decision Transformer, an extension of Decision Transformer [1] in which that does not use the return-to-go for policy conditioning. Instead, individual states are clustered and conditioning is performed with a "skill-distribution-to-go". The general motivation is a relat...
First, the authors make some proofs (whose correctness I haven't fully verified) based on the asymptotic equipartition property that show there exists a small set that can represent the whole data distribution. Then, the authors make some connections between RNNs and sparse coding. Finally, authors test RNNs and RNNs ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: First, the authors make some proofs (whose correctness I haven't fully verified) based on the asymptotic equipartition property that show there exists a small set that can represent the whole data distribution. Then, the authors make some connections between RNNs and sparse coding. Finally, authors test RNNs a...
The authors propose a method to sample from a distribution with a known un-normalized posterior, named Annealed Fisher Implicit Sampler. More specifically, they introduce the loss with the same gradient as the Fisher Divergence, provided a perfect estimate of the sampler's score function. They train the sampler via alt...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: The authors propose a method to sample from a distribution with a known un-normalized posterior, named Annealed Fisher Implicit Sampler. More specifically, they introduce the loss with the same gradient as the Fisher Divergence, provided a perfect estimate of the sampler's score function. They train the sampler...
This paper examines the capacity of renormalizing nondeterministic stack RNN (RNS-RNN), a stack RNN variant proposed by previous work, and proves that it can recognize all context-free languages. It empirically shows that, surprisingly, it performs well in recognizing non-context-free languages, even when the language’...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper examines the capacity of renormalizing nondeterministic stack RNN (RNS-RNN), a stack RNN variant proposed by previous work, and proves that it can recognize all context-free languages. It empirically shows that, surprisingly, it performs well in recognizing non-context-free languages, even when the l...
This paper proposes to solve the potential classification problem of the adversarial samples with the help of diffusion models. Specifically, the authors firstly give some analysis of the regions of the conditional generation. They claim that the conditionally generated samples will concentrate on the regions around th...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes to solve the potential classification problem of the adversarial samples with the help of diffusion models. Specifically, the authors firstly give some analysis of the regions of the conditional generation. They claim that the conditionally generated samples will concentrate on the regions a...
(1) This paper aims to solve the long-tailed recognition task. (I) This paper proposes that the class-conditional distribution shift can not be ignored in practice with empirical analysis. (II) Based on this observation, this paper proposes to utilize DRO to minimize the worst risk and It can be understood as ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: (1) This paper aims to solve the long-tailed recognition task. (I) This paper proposes that the class-conditional distribution shift can not be ignored in practice with empirical analysis. (II) Based on this observation, this paper proposes to utilize DRO to minimize the worst risk and It can be unders...
The goal of this work is to automatically learn the activation functions via the combinations of different candidate activation functions for PINN. Specifically, it adopted gate function with a learnable parameter to identify the coefficients for different candidate activation functions. Extensive experiments are carri...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The goal of this work is to automatically learn the activation functions via the combinations of different candidate activation functions for PINN. Specifically, it adopted gate function with a learnable parameter to identify the coefficients for different candidate activation functions. Extensive experiments a...
This paper presents a general masking strategy for visual representation learning with siamese neural network architectures. The approach is applicable to CNN-based or ViT-based neural networks. It consists in pre-processing the input with a high-pass filter and then doing different augmentations that mask out parts of...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper presents a general masking strategy for visual representation learning with siamese neural network architectures. The approach is applicable to CNN-based or ViT-based neural networks. It consists in pre-processing the input with a high-pass filter and then doing different augmentations that mask out ...
This paper proposes a homotopy learning method for solving constrained optimization problems. The idea of homotopy is pretty intuitive: we have a parameter lambda_h \in [0,1], indicating the difficulty of the current problem. lambda_h=0 means a trivial initial problem H0 and lambda_h=1 means the target difficult proble...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a homotopy learning method for solving constrained optimization problems. The idea of homotopy is pretty intuitive: we have a parameter lambda_h \in [0,1], indicating the difficulty of the current problem. lambda_h=0 means a trivial initial problem H0 and lambda_h=1 means the target difficul...
This paper attempts to prove algorithmic stability-based generalization error bounds for neural nets with ReLU activation functions. To do this, the modified definition of "Almost (Sure) Support Stability" is proposed, which weakens the standard uniform stability definition and requires it to hold with a certain probab...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper attempts to prove algorithmic stability-based generalization error bounds for neural nets with ReLU activation functions. To do this, the modified definition of "Almost (Sure) Support Stability" is proposed, which weakens the standard uniform stability definition and requires it to hold with a certai...
By considering it as an empirical Replay Memory MDP (RM-MDP), the authors in this study take advantage of the information contained in the experience replay memory. The authors discovered a conservative value estimate that solely considers transitions seen within the replay memory by solving it using dynamic programmin...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: By considering it as an empirical Replay Memory MDP (RM-MDP), the authors in this study take advantage of the information contained in the experience replay memory. The authors discovered a conservative value estimate that solely considers transitions seen within the replay memory by solving it using dynamic pr...
The paper proposes a new framework for obtaining bounds on quantiles of the loss distribution of a given predictor. The method uses a lower bound on the CDF to obtain high-confidence bounds on the quantile loss profiles of the predictor. Strength The framework produced by the paper generalizes exisitng known framework...	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 a new framework for obtaining bounds on quantiles of the loss distribution of a given predictor. The method uses a lower bound on the CDF to obtain high-confidence bounds on the quantile loss profiles of the predictor. Strength The framework produced by the paper generalizes exisitng known f...
This paper proposes an extension of GP that is applicable to situations where a single task is assigned several different output labels. The proposed model is evaluated using real-world datasets. S1. The authors are trying to solve a spoken language assessment task using a Gaussian process. The problem addressed herein...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes an extension of GP that is applicable to situations where a single task is assigned several different output labels. The proposed model is evaluated using real-world datasets. S1. The authors are trying to solve a spoken language assessment task using a Gaussian process. The problem addresse...
This paper proposes a simple method of randomly masking past tokens during causal language modeling that boosts zero-shot capabilities and fine-tuning results by a non-trivial margin. The problem motivation is capturing the best of both decoder and encoder LMs. While previous methods achieve this as well, the authors ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a simple method of randomly masking past tokens during causal language modeling that boosts zero-shot capabilities and fine-tuning results by a non-trivial margin. The problem motivation is capturing the best of both decoder and encoder LMs. While previous methods achieve this as well, the ...
The paper presents a novel method for targeted adversarial attacks on Deep Reinforcement Learning algorithms by modifying agent observations. It learns an adversarial reward function via the preferences of a human-in-the-loop, and uses it to guide an intention policy that itself generates behaviors the adversarial poli...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper presents a novel method for targeted adversarial attacks on Deep Reinforcement Learning algorithms by modifying agent observations. It learns an adversarial reward function via the preferences of a human-in-the-loop, and uses it to guide an intention policy that itself generates behaviors the adversar...
Prior work by Saunshi et al. 2022 had highlighted the insufficiency of previous theoretical analyses that are function-class agnostic to completely explain the success of self-supervised learning. This paper provides theoretical guarantees for self-supervised learning that do incorporate the function class building on ...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: Prior work by Saunshi et al. 2022 had highlighted the insufficiency of previous theoretical analyses that are function-class agnostic to completely explain the success of self-supervised learning. This paper provides theoretical guarantees for self-supervised learning that do incorporate the function class buil...
L2B is a learnable loss objective that enables a joint reweighting of instances and labels at once. L2B dynamically adjusts the per-sample importance weight between the given labels and pseudo-labels in a meta way. Strengths: 1) L2B dynamically adjusts the per-sample importance weight between the given labels and pse...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: L2B is a learnable loss objective that enables a joint reweighting of instances and labels at once. L2B dynamically adjusts the per-sample importance weight between the given labels and pseudo-labels in a meta way. Strengths: 1) L2B dynamically adjusts the per-sample importance weight between the given labels...
In this paper, a novel second-order technique is proposed for deep neural networks training. strength + it is a good try to further reduce the computational and storage burden for second-order method for training DNNs. + the method is not complicated and has shown great advantage in numerical experiments weaknesses ...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: In this paper, a novel second-order technique is proposed for deep neural networks training. strength + it is a good try to further reduce the computational and storage burden for second-order method for training DNNs. + the method is not complicated and has shown great advantage in numerical experiments wea...
This paper proposes Transformer-M, a Transformer-based model that can take both 2D and 3D molecular formats as input. It adopts various positional encoding techniques to unify 2D and 3D information into transformer as attention bias terms. This model is claimed to be a general-purpose model for molecular tasks. Experim...	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 Transformer-M, a Transformer-based model that can take both 2D and 3D molecular formats as input. It adopts various positional encoding techniques to unify 2D and 3D information into transformer as attention bias terms. This model is claimed to be a general-purpose model for molecular tasks....
This paper proposes a self-supervised representation learning method by applying high-pass or low-pass filters on the image and restoring the missing frequency components. The experiments also include results for self-supervised learning via other common image restoration methods such as denoising, super-resolution and...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a self-supervised representation learning method by applying high-pass or low-pass filters on the image and restoring the missing frequency components. The experiments also include results for self-supervised learning via other common image restoration methods such as denoising, super-resolu...
This work propose a federated learning algorithm, named FedMIM, which adopts the multi-step inertial momentum on the edge devices and enhances the local consistency for free during the training to improve the robustness of the heterogeneity. Specifically, the authors incorporate the weighted global gradient estimations...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This work propose a federated learning algorithm, named FedMIM, which adopts the multi-step inertial momentum on the edge devices and enhances the local consistency for free during the training to improve the robustness of the heterogeneity. Specifically, the authors incorporate the weighted global gradient est...
The submitted article proposes a practical solution to the rotated object detection problem. To locate arbitrarily oriented objects in images, detectors often rely on the calculation of the SkewIoU. However, its closed-form calculation cannot always be provided and its implementation in popular deep learning framework ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The submitted article proposes a practical solution to the rotated object detection problem. To locate arbitrarily oriented objects in images, detectors often rely on the calculation of the SkewIoU. However, its closed-form calculation cannot always be provided and its implementation in popular deep learning fr...
The authors analyze the new challenges in offline communication learning, and introduce a benchmark of offline communication learning which contains diverse tasks. The author propose an effective algorithm, Multi-head Communication Imitation (MHCI), which aims to address the problem of learning from single-source or mu...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors analyze the new challenges in offline communication learning, and introduce a benchmark of offline communication learning which contains diverse tasks. The author propose an effective algorithm, Multi-head Communication Imitation (MHCI), which aims to address the problem of learning from single-sour...
This paper proposes a new approach Decision Transformers (DT) have demonstrated strong performances in offline reinforcement learning settings, but quickly adapting to unseen novel tasks remains challenging. To address this challenge, we propose a new framework, called Hyper-Decision Transformer (HDT), that can genera...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a new approach Decision Transformers (DT) have demonstrated strong performances in offline reinforcement learning settings, but quickly adapting to unseen novel tasks remains challenging. To address this challenge, we propose a new framework, called Hyper-Decision Transformer (HDT), that ca...
This paper proposes a contextualized generative retrieval model. The generative retrieval model usually performs worse on unseen data and the traditional KNN retrieval model has the advantage of good generalization but also fails on documents with long sequences. The authors propose a new architecture to overcome thes...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a contextualized generative retrieval model. The generative retrieval model usually performs worse on unseen data and the traditional KNN retrieval model has the advantage of good generalization but also fails on documents with long sequences. The authors propose a new architecture to overc...
This paper identifies the gradient misalignment as the main reason why applying privacy-enhancing methods may lead to more unfairness. The authors use an existing method, DPSGD-Global, and a variant of it, DPSGD-Global-Adapt, to prevent misalignment and reduce this unfairness. Major strengths: 1. Proposition 3 is close...	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 identifies the gradient misalignment as the main reason why applying privacy-enhancing methods may lead to more unfairness. The authors use an existing method, DPSGD-Global, and a variant of it, DPSGD-Global-Adapt, to prevent misalignment and reduce this unfairness. Major strengths: 1. Proposition 3 ...
This paper proposed a novel gradient free quantum learning framework called QUARK, which is designed for classification tasks. Different numbers of qubits are assigned to Data, Weight and Output session, and the data are encoded with basis encoding method. Weights are applied to data through controlled CNOT gates and t...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposed a novel gradient free quantum learning framework called QUARK, which is designed for classification tasks. Different numbers of qubits are assigned to Data, Weight and Output session, and the data are encoded with basis encoding method. Weights are applied to data through controlled CNOT gat...
The presented work focuses on improving the efficiency of the training process by reducing memory cached during the backward propagation. Activation maps are split into low and high frequency components based on the observation that they don't affect the accuracy of the model equally. High precision copy of low frequen...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The presented work focuses on improving the efficiency of the training process by reducing memory cached during the backward propagation. Activation maps are split into low and high frequency components based on the observation that they don't affect the accuracy of the model equally. High precision copy of low...
The paper proposed the use of Equal size hard EM algorithm to address the problem of diverse response generation in open domain dialog modeling. The paper modified the EM algorithm to overcome mode collapse issues with hard EM and synchronous training collapse issues with soft EM. Experimental results show improvement ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposed the use of Equal size hard EM algorithm to address the problem of diverse response generation in open domain dialog modeling. The paper modified the EM algorithm to overcome mode collapse issues with hard EM and synchronous training collapse issues with soft EM. Experimental results show impr...
This paper proposes SIMPLEKT, a simple but tough-to-beat KT baseline that is simple to implement, computationally friendly and robust to a wide range of KT datasets across different domains. Strength 1. The paper is well written and well presented. 2. The source code is publicly avaiblale. 3. The experimental resul...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes SIMPLEKT, a simple but tough-to-beat KT baseline that is simple to implement, computationally friendly and robust to a wide range of KT datasets across different domains. Strength 1. The paper is well written and well presented. 2. The source code is publicly avaiblale. 3. The experiment...
The paper experimentally investigates the effect of differentially private stochastic gradient training on robustness against three types of distribution shifts. The experiments involve data with synthetically generated distribution shifts as well as natural ones. The main strength of the paper is that it investigates ...	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 experimentally investigates the effect of differentially private stochastic gradient training on robustness against three types of distribution shifts. The experiments involve data with synthetically generated distribution shifts as well as natural ones. The main strength of the paper is that it inves...
This paper presents a method that reduces the memory usage of training/fine-tuning for convolutional neural networks and vision transformers. The idea is to prune away insignificant intermediate tensors (i.e., those that are low in value) required for the back-propagation. The method has been tested on DeiT and R-CNN m...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a method that reduces the memory usage of training/fine-tuning for convolutional neural networks and vision transformers. The idea is to prune away insignificant intermediate tensors (i.e., those that are low in value) required for the back-propagation. The method has been tested on DeiT and...
This paper proposes Masking Imputing Aggregation (MIA), a framework based on masking and aggregation, to achieve certified robustness for time series forecasting models against temporally localized $\ell_0$-norm-bounded perturbations. For the masked time steps, an imputation model is trained to fill in and then new tim...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes Masking Imputing Aggregation (MIA), a framework based on masking and aggregation, to achieve certified robustness for time series forecasting models against temporally localized $\ell_0$-norm-bounded perturbations. For the masked time steps, an imputation model is trained to fill in and then...
The authors propose the generation algorithm of vectorized HD map from onboard sensors such as cameras and LiDARs. They represent map elements as a set of polylines and the positions are learned from extracted BEV features. After element keypoints are detected, the trained decoder outputs polylines of every elements....	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose the generation algorithm of vectorized HD map from onboard sensors such as cameras and LiDARs. They represent map elements as a set of polylines and the positions are learned from extracted BEV features. After element keypoints are detected, the trained decoder outputs polylines of every e...
Based on observation in Liang el al., 2022 on the modality gap, the paper further proves that the modality gap in multi-modal settings has not influence on the prediction of classifiers, and thus enables cross-modal transferability. The papers further shows that texts can be used as a proxy for image inputs, and provid...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Based on observation in Liang el al., 2022 on the modality gap, the paper further proves that the modality gap in multi-modal settings has not influence on the prediction of classifiers, and thus enables cross-modal transferability. The papers further shows that texts can be used as a proxy for image inputs, an...