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 proposes a new approach to video visual relation detection that is based on extension of compositional prompt tuning with motion cues. The Relation Prompt (RePro) is designed to address the technical challenges of Open-vocabulary Video Visual Relation Detection: the prompt tokens should respect the two diffe...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new approach to video visual relation detection that is based on extension of compositional prompt tuning with motion cues. The Relation Prompt (RePro) is designed to address the technical challenges of Open-vocabulary Video Visual Relation Detection: the prompt tokens should respect the t...
The paper proposes a simple idea to test a clearly-defined hypothesis: whether representations learned by a vision model (trained on images only vs. trained on images including limited degrees of linguistic supervision) are functionally equivalent to those learned by a language model (up to a linear transformation). Th...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes a simple idea to test a clearly-defined hypothesis: whether representations learned by a vision model (trained on images only vs. trained on images including limited degrees of linguistic supervision) are functionally equivalent to those learned by a language model (up to a linear transformat...
This paper studies the de-noising recommendation problem. The authors propose a meta learning method to annotate the unlabelled data from both the loss and gradient perspective. They propose the inverse dual loss to boost the true label learning and prevent false label learning. Moreover, they also propose the inverse ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper studies the de-noising recommendation problem. The authors propose a meta learning method to annotate the unlabelled data from both the loss and gradient perspective. They propose the inverse dual loss to boost the true label learning and prevent false label learning. Moreover, they also propose the ...
This paper proposes a scene text detector, PBFormer, using the transformer with a “new” text instance representation called Polynomial Band (PB). The proposed PB is able to represent a text with a complex shape since it utilizes four polynomial curves to fit the text instance’s top, bottom, left, and right sides. The p...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a scene text detector, PBFormer, using the transformer with a “new” text instance representation called Polynomial Band (PB). The proposed PB is able to represent a text with a complex shape since it utilizes four polynomial curves to fit the text instance’s top, bottom, left, and right side...
The authors propose a graph signal sampling approach to matrix completion/recommendation systems. They propose regularization approaches for noise reduction, and also provide a Bayesian extension that takes into account model uncertainty. They show that their approaches are scalable and provide both theoretical guarant...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors propose a graph signal sampling approach to matrix completion/recommendation systems. They propose regularization approaches for noise reduction, and also provide a Bayesian extension that takes into account model uncertainty. They show that their approaches are scalable and provide both theoretical...
The main contribution of the paper is to devise an iterative differentiable and parameter-free pruning algorithm utilizing attention-based soft pruning masks. They show that the iterative learnable pruning results in improved performance on vision and NLP tasks. The differentiable pruning utilizing soft attention mas...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The main contribution of the paper is to devise an iterative differentiable and parameter-free pruning algorithm utilizing attention-based soft pruning masks. They show that the iterative learnable pruning results in improved performance on vision and NLP tasks. The differentiable pruning utilizing soft atten...
The paper presents an algorithm for producing feasible solutions for families of linearly constrained optimization problems with varying right-hand sides. The algorithm can provide feasibility guarantees under certain circumstances by 1) dilating each of the linear constraint (shrinking the feasible region), 2) trainin...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper presents an algorithm for producing feasible solutions for families of linearly constrained optimization problems with varying right-hand sides. The algorithm can provide feasibility guarantees under certain circumstances by 1) dilating each of the linear constraint (shrinking the feasible region), 2)...
Paper proposes to automatically learn which transformations to augment during training without introducing biases, for example, when the augmentations are not appropriate for the task. Authors rewrite the problem as a semi-infinite constrained problem similar to a recent adversarial robustness framework. Proposed appro...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: Paper proposes to automatically learn which transformations to augment during training without introducing biases, for example, when the augmentations are not appropriate for the task. Authors rewrite the problem as a semi-infinite constrained problem similar to a recent adversarial robustness framework. Propos...
Paper proposed a novel approach of using end-to-end Transformer network to compute 8 semantic measures based on Laban Movement Analysis (LMA) of the input video to solve the Temporal Action Proposal Generation problem. The proposed network shows comparable results against SOTA methods on 3 datasets, THUMBOS14, Activit...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Paper proposed a novel approach of using end-to-end Transformer network to compute 8 semantic measures based on Laban Movement Analysis (LMA) of the input video to solve the Temporal Action Proposal Generation problem. The proposed network shows comparable results against SOTA methods on 3 datasets, THUMBOS14,...
The paper proposes a new Decision Tree GNN (DT+GNN) architecture to help make GNN predictions more explainable. The proposed approach uses a new layer inspired by stone age model, distills all MLPs to Decision Trees, introduces a pruning mechanism for there trees, and reports experimental results to empirically evaluat...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a new Decision Tree GNN (DT+GNN) architecture to help make GNN predictions more explainable. The proposed approach uses a new layer inspired by stone age model, distills all MLPs to Decision Trees, introduces a pruning mechanism for there trees, and reports experimental results to empirically...
The goal of this paper is "investigating the optimization geometry of deep networks". In particular, correlations between gradients in adjacent time steps is measured in computer vision tasks (mostly convnets on CIFAR). It is observed that correlations during training have periods of positive values, implying that the ...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: The goal of this paper is "investigating the optimization geometry of deep networks". In particular, correlations between gradients in adjacent time steps is measured in computer vision tasks (mostly convnets on CIFAR). It is observed that correlations during training have periods of positive values, implying t...
This paper aims at preventing data from being used (without authorization) for training DNNs. They propose the One-Pixel Shortcut (OPS) method where all the images of a same class get the same pixel location replaced with a same color pixel. This method fools the network during training as the network will use this rep...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper aims at preventing data from being used (without authorization) for training DNNs. They propose the One-Pixel Shortcut (OPS) method where all the images of a same class get the same pixel location replaced with a same color pixel. This method fools the network during training as the network will use ...
This paper presents a derivcate construction mechanims for logical formulae. Results are very good. This paper presents a theory for obtaining derivatves of logic programs Several researches worked on this, Domingo s and colleagues had a relatively simple model for mlns, Problog used Derivatives from BDDS, Last we hav...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper presents a derivcate construction mechanims for logical formulae. Results are very good. This paper presents a theory for obtaining derivatves of logic programs Several researches worked on this, Domingo s and colleagues had a relatively simple model for mlns, Problog used Derivatives from BDDS, Las...
This paper proposed an autoencoder architecture based on trigonometric functions for unsupervised disentangled representation learning, introduced a new disentanglement metric (based on a strong assumption), and evaluated the proposed architecture on six datasets. ## Strengths - Unsupervised disentanglemen...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposed an autoencoder architecture based on trigonometric functions for unsupervised disentangled representation learning, introduced a new disentanglement metric (based on a strong assumption), and evaluated the proposed architecture on six datasets. ## Strengths - Unsupervised disent...
This paper proposes Specformer, a Transformer-based graph spectral filter that captures the magnitudes and relative dependencies of all Laplacian eigenvalues. Specformer is permutation equivariant and can perform non-local graph convolutions. Extensive experiments on the node-level and graph-level datasets demonstrate ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes Specformer, a Transformer-based graph spectral filter that captures the magnitudes and relative dependencies of all Laplacian eigenvalues. Specformer is permutation equivariant and can perform non-local graph convolutions. Extensive experiments on the node-level and graph-level datasets demo...
This paper presents ContextSpeech, a text-to-speech synthesis model that additionally takes the embedding of the previous sentence as input to improve prosody modeling for paragraph reading. The authors also utilize efficient attention to improve efficiency. Strengths - This paper tackles an important problem that has ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents ContextSpeech, a text-to-speech synthesis model that additionally takes the embedding of the previous sentence as input to improve prosody modeling for paragraph reading. The authors also utilize efficient attention to improve efficiency. Strengths - This paper tackles an important problem t...
The paper proposes a new GNN architecture inspired by a model of homophily and heterophily in social networks. Nodes with similar features are modeled with a positive edge, while nodes with different features are modeled with a negative edge. The paper notices that the weak balance theory for networks naturally applies...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a new GNN architecture inspired by a model of homophily and heterophily in social networks. Nodes with similar features are modeled with a positive edge, while nodes with different features are modeled with a negative edge. The paper notices that the weak balance theory for networks naturally...
This paper proposed an approach for distributedly multiplying two large matrices while handling active adversaries and stragglers. The design is within the framework of verifiable computing, where the adversaries are detected before recovering the final results. However, by modifying the encoding polynomials (using squ...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper proposed an approach for distributedly multiplying two large matrices while handling active adversaries and stragglers. The design is within the framework of verifiable computing, where the adversaries are detected before recovering the final results. However, by modifying the encoding polynomials (u...
This paper looks at the problem of hyperparameter tuning for the Approximate Nearest Neighbor (ANN) search. The authors propose a constrained optimization-based approach for tuning the quantization-based ANN methods. They only require a search cost or recall as input, then the Lagrange multipliers-based approach can pr...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper looks at the problem of hyperparameter tuning for the Approximate Nearest Neighbor (ANN) search. The authors propose a constrained optimization-based approach for tuning the quantization-based ANN methods. They only require a search cost or recall as input, then the Lagrange multipliers-based approac...
The paper considers a federated learning setting. The idea and novelty of the paper consist in taking fairness considerations into account. Namely, agents that contribute a lot of/little data or lower/higher quality data should not be treated the same way. Strength: - the motivation of the paper is clear, and importa...	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 a federated learning setting. The idea and novelty of the paper consist in taking fairness considerations into account. Namely, agents that contribute a lot of/little data or lower/higher quality data should not be treated the same way. Strength: - the motivation of the paper is clear, and...
This paper studies the robustness of conformal prediction to label noise, characterizing when and how it is possible to construct uncertainty sets that correctly cover the unobserved noiseless ground truth labels. Strength: This paper studies a very important problem and tries to provide some theoretical results about ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper studies the robustness of conformal prediction to label noise, characterizing when and how it is possible to construct uncertainty sets that correctly cover the unobserved noiseless ground truth labels. Strength: This paper studies a very important problem and tries to provide some theoretical result...
This paper proves a group imbalanced generalization result for a one-hidden-layer neural network under the Gaussian mixture model assuming that the labels are generated by a ground truth model of the same structure. With this result, the paper argues that the learning performance is the best when the spectral norm of t...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves a group imbalanced generalization result for a one-hidden-layer neural network under the Gaussian mixture model assuming that the labels are generated by a ground truth model of the same structure. With this result, the paper argues that the learning performance is the best when the spectral n...
This work theoretically analyzes that conventional Federated Learning (FL), heterogeneous and arbitrary client participation, is not Probably Approximately Correct (PAC) learnable. And then explore the reason and the fact that the server-aided federated learning (SA-FL) framework uses an auxiliary server dataset to red...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work theoretically analyzes that conventional Federated Learning (FL), heterogeneous and arbitrary client participation, is not Probably Approximately Correct (PAC) learnable. And then explore the reason and the fact that the server-aided federated learning (SA-FL) framework uses an auxiliary server datase...
The paper derives a GP model as a GCN with layer widths increased to infinity. To make the prediction more scalable, a low-rank approximation approach is proposed. Experiments are conducted to compare with other baseline models and show the GP model’s running time and scalability advantages. Strength: 1. The designed ...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper derives a GP model as a GCN with layer widths increased to infinity. To make the prediction more scalable, a low-rank approximation approach is proposed. Experiments are conducted to compare with other baseline models and show the GP model’s running time and scalability advantages. Strength: 1. The d...
This paper proves that, in a particular class of two-environment linear models containing both relevant and irrelevant directions, interpolation is incompatible with invariance (having low "robust error" as used in out-of-domain generalization, or also corresponding to fairness constraints). It also proposes an algori...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves that, in a particular class of two-environment linear models containing both relevant and irrelevant directions, interpolation is incompatible with invariance (having low "robust error" as used in out-of-domain generalization, or also corresponding to fairness constraints). It also proposes a...
This paper introduces ROAST (Random Operation Access Specific Tile) hashing, a model-agnostic, hardware-aware model compression framework. ROAST essentially provides a global parameter sharing method to give arbitrary control to the user over the memory footprint of model during both training and inference. Evaluation ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper introduces ROAST (Random Operation Access Specific Tile) hashing, a model-agnostic, hardware-aware model compression framework. ROAST essentially provides a global parameter sharing method to give arbitrary control to the user over the memory footprint of model during both training and inference. Eva...
The author(s) extends the Moreu-envelop based personalized FL to the composite problems. Convergence analysis under certain assumptions are developed. Experiments on synethtic data and MNIST are conducted to evaluate the proposed algorithm. Pros: - The story line of the paper is clear and related works are well-address...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The author(s) extends the Moreu-envelop based personalized FL to the composite problems. Convergence analysis under certain assumptions are developed. Experiments on synethtic data and MNIST are conducted to evaluate the proposed algorithm. Pros: - The story line of the paper is clear and related works are well...
Authors propose a method for sequential generation of Lego models. This addresses two tasks – (i) generating new class-conditional Lego models a priori; (ii) completing partial models. The proposed approach consists of a neural network to predict how good candidate locations for the 'next' brick is, and various checks ...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: Authors propose a method for sequential generation of Lego models. This addresses two tasks – (i) generating new class-conditional Lego models a priori; (ii) completing partial models. The proposed approach consists of a neural network to predict how good candidate locations for the 'next' brick is, and various...
The paper makes the observation that classic attention mechanisms do not take advantage in the hierarchical structure implicit within language and vision. The propose augmenting the attention operators within transformers with a hierarchy-aware module that progressively discovers "semantic hierarchies" between the tok...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper makes the observation that classic attention mechanisms do not take advantage in the hierarchical structure implicit within language and vision. The propose augmenting the attention operators within transformers with a hierarchy-aware module that progressively discovers "semantic hierarchies" between...
The paper proposes a novel form of group convolution called Discrete-Continuous (DISCO) group convolution, which is targeted at making group convolutions scalable and at the same time retain equivariance properties. DISCO discretizes the input signal space while the filters and output space are maintained to be continu...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes a novel form of group convolution called Discrete-Continuous (DISCO) group convolution, which is targeted at making group convolutions scalable and at the same time retain equivariance properties. DISCO discretizes the input signal space while the filters and output space are maintained to be...
Disco-Dance is a method for improving skill learning by expanding the diversity of the states that skills can reach. The method first selects guide skills by finding skills most likely to lead to new states, then trains new skills (or old skills that have low discriminability) to reach novel states using mutual-inform...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: Disco-Dance is a method for improving skill learning by expanding the diversity of the states that skills can reach. The method first selects guide skills by finding skills most likely to lead to new states, then trains new skills (or old skills that have low discriminability) to reach novel states using mutua...
The authors of this paper present a semisupervised adaptation of the linear least squares-SVM. This formulation trivially contains the LS-SVM and linear spectral clustering as special cases. The bulk of the paper and perhaps its most interesting contribution is its theoretical analysis of the method which utilizes rand...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors of this paper present a semisupervised adaptation of the linear least squares-SVM. This formulation trivially contains the LS-SVM and linear spectral clustering as special cases. The bulk of the paper and perhaps its most interesting contribution is its theoretical analysis of the method which utili...
This paper extends the well-known compositional generalization benchmark dataset SCAN to cSCAN, driven by the idea of evaluating machine learning models' ability in semantic parsing with context and rule-like constraints. ### Strengths: - The authors provided extensive details about how this dataset is generated, hence...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper extends the well-known compositional generalization benchmark dataset SCAN to cSCAN, driven by the idea of evaluating machine learning models' ability in semantic parsing with context and rule-like constraints. ### Strengths: - The authors provided extensive details about how this dataset is generate...
This paper targets transferring multi-source knowledge to the target domain. While the previous studies tried to utilize the generalized knowledge of the multiple source domains, the proposed algorithm focus on domain-specific information. To accomplish the extraction of the domain-specific information, the paper prese...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper targets transferring multi-source knowledge to the target domain. While the previous studies tried to utilize the generalized knowledge of the multiple source domains, the proposed algorithm focus on domain-specific information. To accomplish the extraction of the domain-specific information, the pap...
This paper targets solving knowledge transfer problems. Specifically, the authors distill large models to compact models for efficient machine learning. Teacher-guided training (TGT) framework is proposed for training high-quality compact models which learn knowledge from pretrained generative models. Experiments show ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper targets solving knowledge transfer problems. Specifically, the authors distill large models to compact models for efficient machine learning. Teacher-guided training (TGT) framework is proposed for training high-quality compact models which learn knowledge from pretrained generative models. Experimen...
Score-based generative models are learned via denoising score matching at many time steps during a diffusion process. This paper identifies an important problem with such loss function---estimating this loss via simple monte carlo could have high variance in the middle stage of the diffusion. The authors then propose t...	Recommendation: 8: accept, good paper	Area: Generative models	Review: Score-based generative models are learned via denoising score matching at many time steps during a diffusion process. This paper identifies an important problem with such loss function---estimating this loss via simple monte carlo could have high variance in the middle stage of the diffusion. The authors then p...
This paper proposed a learning-based approach for multi-view image coding. Specifically, the proposed method decouples the inter-view operations during the encoding and thus is suitable for the distributed system. Thorough evaluations demonstrate the effectiveness of the proposed method in terms of PSNR and SSIM. - Str...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposed a learning-based approach for multi-view image coding. Specifically, the proposed method decouples the inter-view operations during the encoding and thus is suitable for the distributed system. Thorough evaluations demonstrate the effectiveness of the proposed method in terms of PSNR and SSI...
This paper considers an important problem in denoising diffusion models (DPM), that is, how to accurately conduct posterior sampling given a pre-learned score function and an inverse problem. Previous works address this problem by circumventing the calculation of $p(y\|x^t)$ and resorting to projections onto the measure...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper considers an important problem in denoising diffusion models (DPM), that is, how to accurately conduct posterior sampling given a pre-learned score function and an inverse problem. Previous works address this problem by circumventing the calculation of $p(y\|x^t)$ and resorting to projections onto the...
The paper proposes to use C-scores as a score to select representative samples to insert into a memory buffer in continual learning. More concretely the authors: - describe current limitations of memory selection methods - propose to use C-score, an expensive but accurate score capable to identify learnable samples, to...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper proposes to use C-scores as a score to select representative samples to insert into a memory buffer in continual learning. More concretely the authors: - describe current limitations of memory selection methods - propose to use C-score, an expensive but accurate score capable to identify learnable sam...
The paper proposes a new knowledge distillation method called NORM, based on the idea of many-to-one representation matching. The method follows a two-stage distillation process using a linear Feature Transform (FT) module. The FT module has two layers: the first expands the number of channels to N times those from the...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a new knowledge distillation method called NORM, based on the idea of many-to-one representation matching. The method follows a two-stage distillation process using a linear Feature Transform (FT) module. The FT module has two layers: the first expands the number of channels to N times those ...