Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

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

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
The paper describes a novel diffusion model extending the concept to modelling irregularly sampled time series by utilizing correlated noise obtained from noise processes to produce continuous samples. The paper provide an elegant framework to combine forecasting, and imputation tasks. One of the main advantage of the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The paper describes a novel diffusion model extending the concept to modelling irregularly sampled time series by utilizing correlated noise obtained from noise processes to produce continuous samples. The paper provide an elegant framework to combine forecasting, and imputation tasks. One of the main advantage...
Inspired by the recently observed neural collapse (NC) phenomena in deep learning classifiers where the feature dimension is often large than the number of classes, this paper considers a more general format of neural collapse (GNC) to include the case where feature dimension $d$ is smaller than the number of class $C$...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Inspired by the recently observed neural collapse (NC) phenomena in deep learning classifiers where the feature dimension is often large than the number of classes, this paper considers a more general format of neural collapse (GNC) to include the case where feature dimension $d$ is smaller than the number of c...
The paper argues that a good continual learner should strike a different stability - plasticity tradeoff in online and offline CL, the former being more stable and the latter more plastic. The authors propose a new method MuFAN, which comprises 3 new components : the use of multi-scale features of pretrained models, an...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper argues that a good continual learner should strike a different stability - plasticity tradeoff in online and offline CL, the former being more stable and the latter more plastic. The authors propose a new method MuFAN, which comprises 3 new components : the use of multi-scale features of pretrained mo...
This paper studies the adaptation of the well-trained transformer to downstream tasks on resource-limited devices. The authors introduced grouped connections with re-parameterization technology into the training process. The experimental results on serval vision tasks demonstrate the effectiveness of the proposed metho...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper studies the adaptation of the well-trained transformer to downstream tasks on resource-limited devices. The authors introduced grouped connections with re-parameterization technology into the training process. The experimental results on serval vision tasks demonstrate the effectiveness of the propos...
The paper investigates the pitfalls of online domain adaptation, that the model tends to collapse with small batch sizes and class imbalances at test time. The authors also propose two techniques to mitigate the problems. Strengh The paper is easy to follow. Authors also include explanations/intuitions to certain ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper investigates the pitfalls of online domain adaptation, that the model tends to collapse with small batch sizes and class imbalances at test time. The authors also propose two techniques to mitigate the problems. Strengh The paper is easy to follow. Authors also include explanations/intuitions to ...
This paper proposes a modification to IQL called MA2QL in which each agent alternately updates the parameters rather than the simultaneous update. The main contribution of this paper is the theoretical analysis of this modification. In experiments, MA2QL is demonstrated to outperform IQL in several environments. ## Str...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a modification to IQL called MA2QL in which each agent alternately updates the parameters rather than the simultaneous update. The main contribution of this paper is the theoretical analysis of this modification. In experiments, MA2QL is demonstrated to outperform IQL in several environments...
This study applies second-order optimization (K-FAC) in parameter-efficient tuning (with adapters and LoRA) of Transformer models. Measurements show that the memory cost of second-order optimization can be relatively small due to the small number of trainable parameters, suggesting a second-order method is more suitabl...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This study applies second-order optimization (K-FAC) in parameter-efficient tuning (with adapters and LoRA) of Transformer models. Measurements show that the memory cost of second-order optimization can be relatively small due to the small number of trainable parameters, suggesting a second-order method is more...
This paper presents a new diffusion-based generative model that designs protein backbone structures via a procedure that mirrors the native folding process. By considering each residue to be its own reference frame, it describes protein backbone structure as a series of consecutive angles capturing the relative orienta...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents a new diffusion-based generative model that designs protein backbone structures via a procedure that mirrors the native folding process. By considering each residue to be its own reference frame, it describes protein backbone structure as a series of consecutive angles capturing the relative...
The paper claims that distance-based CO solvers are more robust than coordinates-based solvers, with concerns for two defects of DL solvers: i) coordinates can hardly depict real-world cases, and ii) coordinates can not be sufficiently 'traversed' by training data. Strength - It provides a new perspective for CO resea...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper claims that distance-based CO solvers are more robust than coordinates-based solvers, with concerns for two defects of DL solvers: i) coordinates can hardly depict real-world cases, and ii) coordinates can not be sufficiently 'traversed' by training data. Strength - It provides a new perspective for ...
This paper investigates vertical FL on graph neural networks. A communication-efficient vertically-distributed training algorithm called GLASU was proposed, where the main idea for communication saving is lazy aggregation and stale updates. It was proved that the proposed algorithm can converge given that the loss func...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper investigates vertical FL on graph neural networks. A communication-efficient vertically-distributed training algorithm called GLASU was proposed, where the main idea for communication saving is lazy aggregation and stale updates. It was proved that the proposed algorithm can converge given that the l...
This paper investigate a pevious work on stein mixtures and show that it corresponds to inference with the Renyi $\alpha$-divergence for $\alpha=0$, and using other values of $\alpha$ may stablize the inference. Strength: 1. This paper is overall well-written and the literature review is thorough. Weaknesses: 1. The...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper investigate a pevious work on stein mixtures and show that it corresponds to inference with the Renyi $\alpha$-divergence for $\alpha=0$, and using other values of $\alpha$ may stablize the inference. Strength: 1. This paper is overall well-written and the literature review is thorough. Weaknesses:...
This paper considers compositional generalization and proposes a "dynamic least-to-most prompting" method to tackle the problem. "dynamic least-to-most prompting" consists of three steps: decomposition using syntactic parsing, dynamic exemplar selection, and sequential solution. The experiment results are impressive: o...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper considers compositional generalization and proposes a "dynamic least-to-most prompting" method to tackle the problem. "dynamic least-to-most prompting" consists of three steps: decomposition using syntactic parsing, dynamic exemplar selection, and sequential solution. The experiment results are impre...
The paper performs an analysis of what the fine-tuning strategy should be when encountering various distribution shifts. They classify the different distribution shifts into three categories and show that for each category of distribution shift, different layers of the model should be surgically fine-tuned. This analys...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper performs an analysis of what the fine-tuning strategy should be when encountering various distribution shifts. They classify the different distribution shifts into three categories and show that for each category of distribution shift, different layers of the model should be surgically fine-tuned. Thi...
The paper considers the problem of continual learning in the context of fake audio detection. One of the main challenges of continual learning is catastrophic forgetting. The paper proposes a new algorithm called Regularized Adaptive Weight Modification (RAWM). The motivation is genuine audios are more similar than fak...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper considers the problem of continual learning in the context of fake audio detection. One of the main challenges of continual learning is catastrophic forgetting. The paper proposes a new algorithm called Regularized Adaptive Weight Modification (RAWM). The motivation is genuine audios are more similar ...
This paper proves that the min-max excess risk of nonparametric classification algorithms is bounded by the number of minority samples when there is a distributional shift between train and test, and therefore concludes that in the worst case, undersampling is optimal unless there is a high degree of overlap between th...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves that the min-max excess risk of nonparametric classification algorithms is bounded by the number of minority samples when there is a distributional shift between train and test, and therefore concludes that in the worst case, undersampling is optimal unless there is a high degree of overlap be...
The authors propose a method for summarising graphs in such a way that they are compressed into a smaller representation while retaining the information required for the downstream task. The authors extend an optimum transport-based framework method for graph summarization (Garg & Jaakkola (2019)) to a supervised graph...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The authors propose a method for summarising graphs in such a way that they are compressed into a smaller representation while retaining the information required for the downstream task. The authors extend an optimum transport-based framework method for graph summarization (Garg & Jaakkola (2019)) to a supervis...
Bilevel optimization problems, in which one optimization problem is nested within another, occur in multiple machine learning applications such as hyperparameter optimization. The paper considers a generic distributed bilevel optimization problem and proposes a novel asynchronous distributed algorithm. The proposed alg...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: Bilevel optimization problems, in which one optimization problem is nested within another, occur in multiple machine learning applications such as hyperparameter optimization. The paper considers a generic distributed bilevel optimization problem and proposes a novel asynchronous distributed algorithm. The prop...
This paper proposes a probabilistic graph generator that models the distribution over good architectures using GNNs. This converts the neural architectures to a learnable computational graph. Extensive experiments indicate the effectiveness and efficiency of the proposed method. This paper is generally well-written, an...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a probabilistic graph generator that models the distribution over good architectures using GNNs. This converts the neural architectures to a learnable computational graph. Extensive experiments indicate the effectiveness and efficiency of the proposed method. This paper is generally well-wri...
Building on top of BDD100k, Cityscapes and Sim10k, the proposed DetectBench in this work is an object detection benchmark for evaluating the OOD generalization performance of object detectors. The benchmark consists of 4 types of domain shifts, i.e., sim2real, weather, scene and time. The authors compared a number of o...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Building on top of BDD100k, Cityscapes and Sim10k, the proposed DetectBench in this work is an object detection benchmark for evaluating the OOD generalization performance of object detectors. The benchmark consists of 4 types of domain shifts, i.e., sim2real, weather, scene and time. The authors compared a num...
While reconstructing data in FL from gradients leads to a continuous optimization problem, tabular data with categorical variables yields a mixed integer programming optimization problem. The authors propose a softmax based continuous relaxation and an ensemble strategy that makes the underlying solution algorithm more...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: While reconstructing data in FL from gradients leads to a continuous optimization problem, tabular data with categorical variables yields a mixed integer programming optimization problem. The authors propose a softmax based continuous relaxation and an ensemble strategy that makes the underlying solution algori...
The paper proposes a generalizable optimization strategy that codifies a way to trade-off between group attribution and component-wise attributions calculated from explanation methods. The end product is a group attribution version of a given explainer (they propose G-SHAP as an example for groupwise SHAP) that balance...	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 generalizable optimization strategy that codifies a way to trade-off between group attribution and component-wise attributions calculated from explanation methods. The end product is a group attribution version of a given explainer (they propose G-SHAP as an example for groupwise SHAP) that...
This paper proposes “Diff Comb Explainer”: a neuro-symbolic architecture for selecting an answer (from a given set of candidates) as well as explanations (from a corpus of facts) for a multiple-choice QA task. Each candidate answer is first converted into a hypothesis by concatenating the question to it, and then scor...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes “Diff Comb Explainer”: a neuro-symbolic architecture for selecting an answer (from a given set of candidates) as well as explanations (from a corpus of facts) for a multiple-choice QA task. Each candidate answer is first converted into a hypothesis by concatenating the question to it, and t...
The authors introduce a mechanism for semi-parametric prediction using neural attention mechanisms and a variant of inducing points. In this work, the authors combine ideas such as cross-attention between attributes and between datapoints, yielding a framework -denoted SPIN- which given a new dataset produces a set of...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors introduce a mechanism for semi-parametric prediction using neural attention mechanisms and a variant of inducing points. In this work, the authors combine ideas such as cross-attention between attributes and between datapoints, yielding a framework -denoted SPIN- which given a new dataset produces ...
The paper investigates the reproduction of training examples from pre-trained language models. The amount of memorization is measured across 1) different model sizes, 2) the number of times an example is duplicated in the training data, 3) the length of prompt given as input. It is concluded that larger models, more du...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper investigates the reproduction of training examples from pre-trained language models. The amount of memorization is measured across 1) different model sizes, 2) the number of times an example is duplicated in the training data, 3) the length of prompt given as input. It is concluded that larger models,...
The paper proposes a new approach for adversarial attacks on NMT models. The authors design a way to propagate the gradient for embeddings and propose a specific loss that targets reasonable criteria. The propagation is based on the common idea of relaxation with a new additional FC part to create embeddings for compar...	Recommendation: 6: marginally above 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 approach for adversarial attacks on NMT models. The authors design a way to propagate the gradient for embeddings and propose a specific loss that targets reasonable criteria. The propagation is based on the common idea of relaxation with a new additional FC part to create embeddings fo...
This paper investigates why iterative magnitude pruning (IMP) is successful. The authors hypothesize that the masks discovered during iterations of IMP encode subspaces that intersect the linearly connected loss sublevel sets associated with the previous iterate. They present multiple experiments to support this hypo...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper investigates why iterative magnitude pruning (IMP) is successful. The authors hypothesize that the masks discovered during iterations of IMP encode subspaces that intersect the linearly connected loss sublevel sets associated with the previous iterate. They present multiple experiments to support t...
This paper considers designing dynamic network topologies to improve the training time and performance of the cross silo FL training. The authors propose an algorithm and verify its effectiveness through empirical evaluation. - Strength: This is an interesting problem, especially the dynamic topology part which is unde...	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 considers designing dynamic network topologies to improve the training time and performance of the cross silo FL training. The authors propose an algorithm and verify its effectiveness through empirical evaluation. - Strength: This is an interesting problem, especially the dynamic topology part which...
This paper proposes a method to learn image-based representations from video. The paper makes two contributions: (1) on the network side, the authors introduce simple modifications to the way augmentations are applied (e.g. change the scale and uniform temporal sampling ) and a simple attention head for attention pooli...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a method to learn image-based representations from video. The paper makes two contributions: (1) on the network side, the authors introduce simple modifications to the way augmentations are applied (e.g. change the scale and uniform temporal sampling ) and a simple attention head for attenti...
The focus of this paper is on Explanatory Interactive Learning (XIL), which consists of a loop in which the user queries the system, the system explains its predictions and the loop restarts. An assumption here is that this interactive loop is interesting to both the system (improving its prediction accuracy) and the...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The focus of this paper is on Explanatory Interactive Learning (XIL), which consists of a loop in which the user queries the system, the system explains its predictions and the loop restarts. An assumption here is that this interactive loop is interesting to both the system (improving its prediction accuracy)...
The paper proposes a new method for deterministic generation of images called TACoS. It allows to invert arbitrary image transformations, with a few investigated experimentally: deterministic gaussian noise, blurs, animorphosis, masking, pixellating and snowing. The paper first evaluates it on the task of reconstructio...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The paper proposes a new method for deterministic generation of images called TACoS. It allows to invert arbitrary image transformations, with a few investigated experimentally: deterministic gaussian noise, blurs, animorphosis, masking, pixellating and snowing. The paper first evaluates it on the task of recon...
The paper studies fairness in federated learning where clients have heterogeneous local data. In particular, the paper introduces a new fairness definition/measure to leverage the excess risk at each local client. Subsequently, the paper proposes an algorithm that provably outperforms FedAvg in terms of this new fairne...	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 studies fairness in federated learning where clients have heterogeneous local data. In particular, the paper introduces a new fairness definition/measure to leverage the excess risk at each local client. Subsequently, the paper proposes an algorithm that provably outperforms FedAvg in terms of this ne...
The work proposes to improve the pipeline of amputated limbs scanning by using the deep learning-based image segmentation before rendering them with photogrammetry. In this way, the density of obtained samples can be increased to improve coverage and accuracy of 3D models with more samples. The proposed method is verif...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The work proposes to improve the pipeline of amputated limbs scanning by using the deep learning-based image segmentation before rendering them with photogrammetry. In this way, the density of obtained samples can be increased to improve coverage and accuracy of 3D models with more samples. The proposed method ...
This paper introduces a general framework of discretization invariant network, DI-Net. The authors give a theoretical upper bound w.r.t. variation and discrepancy in the finite samples case. They also reveal that DI-Net is a generalization of existing neural fields networks such as CNN. They demonstrate the performance...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper introduces a general framework of discretization invariant network, DI-Net. The authors give a theoretical upper bound w.r.t. variation and discrepancy in the finite samples case. They also reveal that DI-Net is a generalization of existing neural fields networks such as CNN. They demonstrate the per...
This paper studies diffusion model of Oja's algorithm for leading eigenvector problem. When the noise is Gaussian and isotropic, the matrix is diagonal, then author establish diffusion convergence to equilibrium state. They show that the algorithm exhibits a escaping property in they laboratory setting which is similar...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper studies diffusion model of Oja's algorithm for leading eigenvector problem. When the noise is Gaussian and isotropic, the matrix is diagonal, then author establish diffusion convergence to equilibrium state. They show that the algorithm exhibits a escaping property in they laboratory setting which is...
This paper proposed an adaptive structured dropout method, ProbDropBlock, which drops contiguous blocks from feature maps with a probability given by the normalized feature salience values. The authors evaluate ProbDropBlock on both vision and language tasks. Strengths: 1. The paper proposed an adaptive structured dro...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposed an adaptive structured dropout method, ProbDropBlock, which drops contiguous blocks from feature maps with a probability given by the normalized feature salience values. The authors evaluate ProbDropBlock on both vision and language tasks. Strengths: 1. The paper proposed an adaptive struct...
This paper proposes a novel token pruning method based on attention back tracking for efficient transformer inference. The goal (and the difference from the previous token pruning methods) is to mitigate the mistakes of removing important tokens at lower layers. An attention approximation network (ApproxNet) is trained...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a novel token pruning method based on attention back tracking for efficient transformer inference. The goal (and the difference from the previous token pruning methods) is to mitigate the mistakes of removing important tokens at lower layers. An attention approximation network (ApproxNet) is...
This paper proposed a uniform-in-time analysis of the propagation of chaos for the mean-field Langevin dynamics to conduct neural network optimization. The authors avoid the double-loop structure and enable to deal with convergence guarantee based on finite-width neural network with vanilla noisy gradient descent algor...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proposed a uniform-in-time analysis of the propagation of chaos for the mean-field Langevin dynamics to conduct neural network optimization. The authors avoid the double-loop structure and enable to deal with convergence guarantee based on finite-width neural network with vanilla noisy gradient desce...
This paper proposes a simple multimodal vision and language mask model for multimodal training, called MaskedVLM. The main idea is to leverage both masked image patches with full text description and masked language tokens with full image patches for semantic align. Additionally, the method employs two pretraining los...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a simple multimodal vision and language mask model for multimodal training, called MaskedVLM. The main idea is to leverage both masked image patches with full text description and masked language tokens with full image patches for semantic align. Additionally, the method employs two pretrai...
This paper presents a model for graph representation learning called AgentNet. The main idea of the paper is to combine graph walks with neural networks: the model consists of agents that, in parallel, explore the graph by performing the following steps: 1. Updating the state of the current node 2. Aggregate inform...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper presents a model for graph representation learning called AgentNet. The main idea of the paper is to combine graph walks with neural networks: the model consists of agents that, in parallel, explore the graph by performing the following steps: 1. Updating the state of the current node 2. Aggregat...
This paper proposes a simple estimator for Bayes error in binary classification. The authors are able to show several properties of the proposed estimator including unbiasedness and some convergence guarantee. Experiments on synthetic datasets, CIFAR-10, and Fashion-MNIST show the effectiveness of the proposed estimato...	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 a simple estimator for Bayes error in binary classification. The authors are able to show several properties of the proposed estimator including unbiasedness and some convergence guarantee. Experiments on synthetic datasets, CIFAR-10, and Fashion-MNIST show the effectiveness of the proposed ...