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 focused on accelerating one-shot neural architecture search (NAS). In particular, the authors aims to accelerate the architecture search phase based on evolutionary approach. The authors first discuss their observation that the evolutionary approaches tend to select the same shallow blocks from the early sta...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper focused on accelerating one-shot neural architecture search (NAS). In particular, the authors aims to accelerate the architecture search phase based on evolutionary approach. The authors first discuss their observation that the evolutionary approaches tend to select the same shallow blocks from the e...
This paper provides the interpretation of IMP method in terms of the loss geometry, to better understand why and how IMP finds winning tickets. It provides various interesting empirical results. The main finding is that whether we find a matching network or not is related with whether the subnetworks at successive roun...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper provides the interpretation of IMP method in terms of the loss geometry, to better understand why and how IMP finds winning tickets. It provides various interesting empirical results. The main finding is that whether we find a matching network or not is related with whether the subnetworks at success...
This paper presents a unified perspective and analysis for almost all of the existing sample efficient RL algorithms. This notion is based on Admissible Bellman Characterisation, and provides an a bound in terms of the functional eluder dimension. Together with this analysis, the authors present an algorithm OPERA, whi...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper presents a unified perspective and analysis for almost all of the existing sample efficient RL algorithms. This notion is based on Admissible Bellman Characterisation, and provides an a bound in terms of the functional eluder dimension. Together with this analysis, the authors present an algorithm OP...
The paper considers the problem of fitting a rational function to a time series. The authors provide a coreset construction that gets a time-series and returns a small coreset that approximates its sum of (fitting) distances to any rational functions of constant degree. The size of the coreset is sub-linear in $n$ and ...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper considers the problem of fitting a rational function to a time series. The authors provide a coreset construction that gets a time-series and returns a small coreset that approximates its sum of (fitting) distances to any rational functions of constant degree. The size of the coreset is sub-linear in ...
The paper proposes BALTO, a biased-diversity-based activation learning approach for fast tensor program optimization. BALTO combines active learning and a biased-diversity-based diversity scheme. BALTO can achieve the same or higher model accuracy with only 5% of the data samples. Strength: - The application of core-se...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes BALTO, a biased-diversity-based activation learning approach for fast tensor program optimization. BALTO combines active learning and a biased-diversity-based diversity scheme. BALTO can achieve the same or higher model accuracy with only 5% of the data samples. Strength: - The application of...
In the paper, the authors use HyperTransformer (HT) in a new setting that mixes continual learning and few-shot learning. The authors propose to reuse these generated weights as input to the HT for the next task. 1. The paper is not well written. For me, it is not trivial to extract information about the setting. What ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In the paper, the authors use HyperTransformer (HT) in a new setting that mixes continual learning and few-shot learning. The authors propose to reuse these generated weights as input to the HT for the next task. 1. The paper is not well written. For me, it is not trivial to extract information about the settin...
This paper proposees a novel method of HARP, which learns both compression rate and mask each layer to achieve a a high reduction rate while maintaining both the clean and adversarial accuracy of original models. Specifically, the method is delicately designed with many skills utilized to implement the non-uniform robu...	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 proposees a novel method of HARP, which learns both compression rate and mask each layer to achieve a a high reduction rate while maintaining both the clean and adversarial accuracy of original models. Specifically, the method is delicately designed with many skills utilized to implement the non-unif...
This manuscript proposed a method for motion mimicking using differentiable physics. This manuscript used Brax as the backbone differentiable physics simulator and proposed to use demonstration to enhance the training. This manuscript showed the comparison with several baselines including DeepMimic and AMP, which are w...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This manuscript proposed a method for motion mimicking using differentiable physics. This manuscript used Brax as the backbone differentiable physics simulator and proposed to use demonstration to enhance the training. This manuscript showed the comparison with several baselines including DeepMimic and AMP, whi...
In the context of MARL, this work introduces the paradigm of centralised training with hybrid execution, a setting in which the information availability of the other agent's observation is not guaranteed during execution. A novel training approach is also introduced, MARO, in order to deal with this setting, comprising...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In the context of MARL, this work introduces the paradigm of centralised training with hybrid execution, a setting in which the information availability of the other agent's observation is not guaranteed during execution. A novel training approach is also introduced, MARO, in order to deal with this setting, co...
This paper conducts an analysis of the bottleneck size/channels of the convolutional autoencoder (CAE). The authors find that an increased bottleneck area (i.e., height × width) improves generalization in terms of reconstruction error while also speeding up training. In contrast, decreasing the feature map size will ma...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper conducts an analysis of the bottleneck size/channels of the convolutional autoencoder (CAE). The authors find that an increased bottleneck area (i.e., height × width) improves generalization in terms of reconstruction error while also speeding up training. In contrast, decreasing the feature map size...
This paper proposed denoising diffusion probabilistic model for image restoration. They specially applied null-space projection at every iteration to target variable x and used A^T y as range spaced values. By considering only null-space contents in reverse process, they achieved realistic images with data consistency....	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed denoising diffusion probabilistic model for image restoration. They specially applied null-space projection at every iteration to target variable x and used A^T y as range spaced values. By considering only null-space contents in reverse process, they achieved realistic images with data cons...
This article aims to experimentally and theoretically define the best SG across different stress tests to reduce the future need for grid search and proposes techniques to improve the performance of SNN training from multiple perspectives, such as the dampening of the SG and the normalization method. Strengths: 1. The ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This article aims to experimentally and theoretically define the best SG across different stress tests to reduce the future need for grid search and proposes techniques to improve the performance of SNN training from multiple perspectives, such as the dampening of the SG and the normalization method. Strengths:...
The paper proposes a new dataset of garment motions under different human movements as well as wind forces. It also introduces a new network that can predict garment motions of different layers at the same time. Strengths: 1. Multi-layered cloth motion dataset is indeed in great need. This dataset can boost the advance...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a new dataset of garment motions under different human movements as well as wind forces. It also introduces a new network that can predict garment motions of different layers at the same time. Strengths: 1. Multi-layered cloth motion dataset is indeed in great need. This dataset can boost the...
This paper points out the spatial redundancy on background points and useless group operation in MSG for the inappropriate receptive field. The authors present an algorithm named dynamic ball query, which dynamically generates gate masks for each group of MSG to process useful points and block redundant background poin...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper points out the spatial redundancy on background points and useless group operation in MSG for the inappropriate receptive field. The authors present an algorithm named dynamic ball query, which dynamically generates gate masks for each group of MSG to process useful points and block redundant backgro...
The paper proposes a curriculum learning paradigm to improve the strength of human-like chess-playing models. The paper studies how the occurrence of problems at a level and the labelling strategy affect the learning process. Through various experiments and analysis of the strength of the teacher and the selection of p...	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 curriculum learning paradigm to improve the strength of human-like chess-playing models. The paper studies how the occurrence of problems at a level and the labelling strategy affect the learning process. Through various experiments and analysis of the strength of the teacher and the select...
This paper provides the first algorithm based on sign SGD, which achieves Byzantine resilience and differential privacy. 1)Author motivates the subject from a theoretical and practical point of view. 2)In the introduction, the author reviewed the previous works carefully. 3)the author successfully addressed the issue...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper provides the first algorithm based on sign SGD, which achieves Byzantine resilience and differential privacy. 1)Author motivates the subject from a theoretical and practical point of view. 2)In the introduction, the author reviewed the previous works carefully. 3)the author successfully addressed t...
This paper proposes the first preventive learning based approach to guarantee DNN solution feasible for optimization problem with linear constrains. The key idea is to obtain an conservative satisfication of constriants and adjusting the size of DNN accordlying to guarantee that the redundant feasible region is large e...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes the first preventive learning based approach to guarantee DNN solution feasible for optimization problem with linear constrains. The key idea is to obtain an conservative satisfication of constriants and adjusting the size of DNN accordlying to guarantee that the redundant feasible region is...
This paper proposes a simple method for unlearning training samples, to comply with GDPR (and other privacy acts') right-to-be-forgotten statement, which gives each person the right to delete their data at any time they want. The proposed unlearning method for language models involves doing a gradient ascent (instead o...	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 a simple method for unlearning training samples, to comply with GDPR (and other privacy acts') right-to-be-forgotten statement, which gives each person the right to delete their data at any time they want. The proposed unlearning method for language models involves doing a gradient ascent (i...
The author proposed a new structural equation model, named Rhino, that can be used to model causal relations between variables in temporal data. Compared to ANM, the proposed model is more flexible, but there is no sacrificing in identifiable properties. The proposed Rhino model is actually identifiable under some comm...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The author proposed a new structural equation model, named Rhino, that can be used to model causal relations between variables in temporal data. Compared to ANM, the proposed model is more flexible, but there is no sacrificing in identifiable properties. The proposed Rhino model is actually identifiable under s...
The goal of this study is to measure the “oppeness” of CLIP-based models. The authors are adding class names that are not part of the target dataset domain and evaluating the change in accuracy from that. They then show how poorly models do and propose a new approach to this via two metrics: inter-modal alignment and i...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The goal of this study is to measure the “oppeness” of CLIP-based models. The authors are adding class names that are not part of the target dataset domain and evaluating the change in accuracy from that. They then show how poorly models do and propose a new approach to this via two metrics: inter-modal alignme...
The presented work proposes a method for estimating model performance and mixed precision configuration without network retraining. The introduced method is based on the Fisher information, what allows for faster computations compared to other existing methods based on e.g., Hessians. The work is clear and well struct...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The presented work proposes a method for estimating model performance and mixed precision configuration without network retraining. The introduced method is based on the Fisher information, what allows for faster computations compared to other existing methods based on e.g., Hessians. The work is clear and wel...
This work aims to improve the cross-lingual retrieval capability of dual encoder models. Given access to multi-lingual document corpus and the intuition that sentence ordering across languages are often similar, this work proposes to use sequential sentence relation to facilitate cross-lingual representation learning. ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work aims to improve the cross-lingual retrieval capability of dual encoder models. Given access to multi-lingual document corpus and the intuition that sentence ordering across languages are often similar, this work proposes to use sequential sentence relation to facilitate cross-lingual representation le...
This main contribution of this paper is that it provides an idea and corresponding theoretical proof about degenerating subgraph matching problem to subtree matching with the help of Graph Neural Network. Strength: Theoretical proof and support for each claims mentioned. Weakness: 1. The datasets used in the experime...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This main contribution of this paper is that it provides an idea and corresponding theoretical proof about degenerating subgraph matching problem to subtree matching with the help of Graph Neural Network. Strength: Theoretical proof and support for each claims mentioned. Weakness: 1. The datasets used in the ...
The authors tackle the problem of updating existing models in production-scale systems while performing Positive-Congruent Training (minimize negative flip rate and error rate simultaneously). To do so, they make two key observations: - Previous studies have demonstrated that ensembles can reduce NFR without negativ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors tackle the problem of updating existing models in production-scale systems while performing Positive-Congruent Training (minimize negative flip rate and error rate simultaneously). To do so, they make two key observations: - Previous studies have demonstrated that ensembles can reduce NFR without ...
The paper proposed a slimmable contrastive self-supervised learning framework for building small models with SimCLR. It applied the well-known slimmable network and solved the gradient imbalance problem in the training by slow start training, online distillation, and loss reweighting, etc. 1. The proposed approach ach...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper proposed a slimmable contrastive self-supervised learning framework for building small models with SimCLR. It applied the well-known slimmable network and solved the gradient imbalance problem in the training by slow start training, online distillation, and loss reweighting, etc. 1. The proposed appr...
This paper provides bounds on the multi-task sample complexity of reinforcement learning in bilinear classes (recently introduced by Du et al., generalizing several existing notions) when the Bellman errors of the tasks can be parameterized in terms of a small set of common features. The main strength is that the work ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper provides bounds on the multi-task sample complexity of reinforcement learning in bilinear classes (recently introduced by Du et al., generalizing several existing notions) when the Bellman errors of the tasks can be parameterized in terms of a small set of common features. The main strength is that t...
This paper proposes PMixUp that combines two augmentation methods (replacement-based augmentation and feature space augmentation) for the topic classification and sentiment analysis. They emphasis the importance of POS tags for those NLP tasks based on their study, and apply the feature space interpolation to swap out ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes PMixUp that combines two augmentation methods (replacement-based augmentation and feature space augmentation) for the topic classification and sentiment analysis. They emphasis the importance of POS tags for those NLP tasks based on their study, and apply the feature space interpolation to s...
This paper introduces a new variant of CTC that can encode conditions to achieve different preferable properties. The authors gave two examples, one is to force the model to emit the last non-blank token as early as possible; the other is to avoid the drift latency of non-blank tokens to provide better trade-off betwee...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a new variant of CTC that can encode conditions to achieve different preferable properties. The authors gave two examples, one is to force the model to emit the last non-blank token as early as possible; the other is to avoid the drift latency of non-blank tokens to provide better trade-of...
The paper compares two forms of meta-learning, with inner-loop updates based either on modulated Hebbian plasticity, or on gradient descent over a network-generated synthetic loss function. Various experiments show that both forms of episodic plasticity (Hebbian and gradient-based) improve performance over using fixe...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The paper compares two forms of meta-learning, with inner-loop updates based either on modulated Hebbian plasticity, or on gradient descent over a network-generated synthetic loss function. Various experiments show that both forms of episodic plasticity (Hebbian and gradient-based) improve performance over us...
In this paper, the authors focus on open set recognition (detection) and propose a method when the known classes are imbalanced. The proposed method is built based on evidential learning and distributionally robust optimization. To this end, the authors first integrate the evidential learning loss function to distribut...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors focus on open set recognition (detection) and propose a method when the known classes are imbalanced. The proposed method is built based on evidential learning and distributionally robust optimization. To this end, the authors first integrate the evidential learning loss function to d...
The authors propose an Attention Retractable Transformer (ART), which incorporates sparse and dense attention to widen the receptive field size. The ART method achieves pretty good results in several image restoration tasks, like super-resolution, denoising, compression artifact reduction. The idea about using sparse a...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose an Attention Retractable Transformer (ART), which incorporates sparse and dense attention to widen the receptive field size. The ART method achieves pretty good results in several image restoration tasks, like super-resolution, denoising, compression artifact reduction. The idea about using ...
This paper shows that separate batch statistics for co-training on clean and adversarial inputs are not necessary. An extremely lightweight adapter using the class token is enough to achieve comparable performance compared with the dual norm setting. It also enables model soup instead of model ensembling for faster mod...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper shows that separate batch statistics for co-training on clean and adversarial inputs are not necessary. An extremely lightweight adapter using the class token is enough to achieve comparable performance compared with the dual norm setting. It also enables model soup instead of model ensembling for fa...
The paper proposes an approach to estimate a manifold using an implicit model. The authors use a constrained energy-based model to learn the data distribution within the manifold. A Langevin dynamics is used to sample from the manifold while training and at an inference time. Strength: - a manifold learning and density...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: The paper proposes an approach to estimate a manifold using an implicit model. The authors use a constrained energy-based model to learn the data distribution within the manifold. A Langevin dynamics is used to sample from the manifold while training and at an inference time. Strength: - a manifold learning and...
It's always hard to summarize papers that report on such a big project, but fundamentally this is a tech report of training an English+Chinese LLM with techniques that improve on the state of the art for pretraining methodology and training stability. GLM-130 is concurrent with OPT and BLOOM, two other GPT-3-class LLM...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: It's always hard to summarize papers that report on such a big project, but fundamentally this is a tech report of training an English+Chinese LLM with techniques that improve on the state of the art for pretraining methodology and training stability. GLM-130 is concurrent with OPT and BLOOM, two other GPT-3-c...
The paper introduces CAMA, a safety approach for MARL agents based on constraint augmentation of the search space, and especially the reward function under the notion of a safety budget and hazard values. CAMA is integrated into two MARL paradigms: CTDE and IL. Experiments are performed to compare CAMA to baselines met...	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 introduces CAMA, a safety approach for MARL agents based on constraint augmentation of the search space, and especially the reward function under the notion of a safety budget and hazard values. CAMA is integrated into two MARL paradigms: CTDE and IL. Experiments are performed to compare CAMA to basel...
In this paper, the authors propose the method Edgeformers, which represents the text-attributed edges in a graph. They utilize the pre-trained language model (BERT) and the concept of virtual node to represent each text in the edge with regards to its adjacent neighboring edges. Specifically, in Edgeformer-E, they r...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors propose the method Edgeformers, which represents the text-attributed edges in a graph. They utilize the pre-trained language model (BERT) and the concept of virtual node to represent each text in the edge with regards to its adjacent neighboring edges. Specifically, in Edgeformer-E...
The paper proposes a variation-based framework that is able to capture the compositionality in emergent languages. Contrary to most previous work, the authors show that compositionality is in fact correlated with generalization. The difference being that the learned compositionality is obscure and posses a high degree ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a variation-based framework that is able to capture the compositionality in emergent languages. Contrary to most previous work, the authors show that compositionality is in fact correlated with generalization. The difference being that the learned compositionality is obscure and posses a high...
The authors propose to analyze the learned representation in a robotic setting by utilizing graph neural networks (GNNs). Using GNNs and Layer-wise Relevance Propagation (LRP), they represent the observations as an entity-relationship to allow us to interpret the learned policy. Finally, they evaluate their approach in...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose to analyze the learned representation in a robotic setting by utilizing graph neural networks (GNNs). Using GNNs and Layer-wise Relevance Propagation (LRP), they represent the observations as an entity-relationship to allow us to interpret the learned policy. Finally, they evaluate their app...
This paper studies the infinite width limit of Multi-layer perceptrons, in particular the existence of feature learning. In contrast to previous work, it studies adaptive optimizers instead of Gradient Descent, but results are roughly the same: NTK limit still results in lazy training (and no feature learning), while ...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the infinite width limit of Multi-layer perceptrons, in particular the existence of feature learning. In contrast to previous work, it studies adaptive optimizers instead of Gradient Descent, but results are roughly the same: NTK limit still results in lazy training (and no feature learning)...
This paper proposes neural attention memory (NAM), an alternative to the standard attention mechanisms, which suffer from well known limitations: their quadratic dependency with respect to the sequence length, and thus their inefficiency when dealing with long sequences, as well as their difficulty to solve tasks that ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes neural attention memory (NAM), an alternative to the standard attention mechanisms, which suffer from well known limitations: their quadratic dependency with respect to the sequence length, and thus their inefficiency when dealing with long sequences, as well as their difficulty to solve tas...