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 proposes an end-to-end Transformer detector named DINO that archives impressive performance on the COCO benchmark. The proposed DINO model improves the training efficiency and the detection performance by utilizing (i) contrastive denoising training, (ii) look forward twice, and (iii) mixed query selection st...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes an end-to-end Transformer detector named DINO that archives impressive performance on the COCO benchmark. The proposed DINO model improves the training efficiency and the detection performance by utilizing (i) contrastive denoising training, (ii) look forward twice, and (iii) mixed query sele...
This paper studies asymptotic instance optimality for interactive decision making with finite decision/action set. The model includes bandits and reinforcement learning, as long as the reward is a function of the observation. The authors show that the standard, asymptotic lower bound by Lai and Robbins applies to this...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies asymptotic instance optimality for interactive decision making with finite decision/action set. The model includes bandits and reinforcement learning, as long as the reward is a function of the observation. The authors show that the standard, asymptotic lower bound by Lai and Robbins applies...
The paper presents a unified model compression to improve model robustness by learning the connection importance for each layer weights. In particular, the authors use a pre-trained model to learn the layer importance to meet a specific pruning budget, and gradually increases the pruning % unless a significant accuracy...	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 presents a unified model compression to improve model robustness by learning the connection importance for each layer weights. In particular, the authors use a pre-trained model to learn the layer importance to meet a specific pruning budget, and gradually increases the pruning % unless a significant ...
This paper considers asynchronous training in the absence of a central server. They propose a novel algorithm SWIFT to address the problem and does not rely on the common bounded delay assumption. Convergence rates and empirical validations are provided There are two aspects that concern me the most: - Algorithm 1 is...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper considers asynchronous training in the absence of a central server. They propose a novel algorithm SWIFT to address the problem and does not rely on the common bounded delay assumption. Convergence rates and empirical validations are provided There are two aspects that concern me the most: - Algori...
The paper presents CANDLE, a binary-based network trained using a teacher-student framework in which a learned weight selects the largest loss between two loss terms in the optimization problem. In this way, CANDLE ensures the minimization of the maximum loss at every iteration. The motivation of using a binary network...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper presents CANDLE, a binary-based network trained using a teacher-student framework in which a learned weight selects the largest loss between two loss terms in the optimization problem. In this way, CANDLE ensures the minimization of the maximum loss at every iteration. The motivation of using a binary...
This paper studies the problem of global and personalized federated learning, aiming to improve both at the same time. This paper proposes a method, FedHKD, that leverages hyper-knowledge distillation to improve both global and local learning. Specifically, hyper-knowledge is defined as the class-wise mean representati...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper studies the problem of global and personalized federated learning, aiming to improve both at the same time. This paper proposes a method, FedHKD, that leverages hyper-knowledge distillation to improve both global and local learning. Specifically, hyper-knowledge is defined as the class-wise mean repr...
This paper models the joint distribution of texts and images using a discrete diffusion model. The basic idea is to firstly using a VQ model to convert an image to discrete tokens, and then model tokens of both texts and images. This paper designs a specific Markov transition matrix as well as a mutual attention mechan...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper models the joint distribution of texts and images using a discrete diffusion model. The basic idea is to firstly using a VQ model to convert an image to discrete tokens, and then model tokens of both texts and images. This paper designs a specific Markov transition matrix as well as a mutual attentio...
In this paper, the authors try to answer an interesting yet challenging question： Is GNN a good representation model for MILP and whether it can predict important properties, eg., feasibility, for MILP problems? They construct a simple example showing that there exist two MILPs such that one of them is feasible while t...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In this paper, the authors try to answer an interesting yet challenging question： Is GNN a good representation model for MILP and whether it can predict important properties, eg., feasibility, for MILP problems? They construct a simple example showing that there exist two MILPs such that one of them is feasible...
This paper proposes a new initialization scheme for both vanilla convnets and convnets with residual structures. This initialization is a simple and special case to maintain dynamic isometry at initialization time. Specifically, they initialize the weight matrix to be identity. The authors also propose a spatial shift ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a new initialization scheme for both vanilla convnets and convnets with residual structures. This initialization is a simple and special case to maintain dynamic isometry at initialization time. Specifically, they initialize the weight matrix to be identity. The authors also propose a spatia...
The authors propose a novel technique to poison training data by crafting perturbations to either repel or attract training data to the corresponding class' mean feature vector. The authors demonstrate that this poisoning technique is superior to known availability attacks when defended against with adversarial trainin...	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 authors propose a novel technique to poison training data by crafting perturbations to either repel or attract training data to the corresponding class' mean feature vector. The authors demonstrate that this poisoning technique is superior to known availability attacks when defended against with adversarial...
This paper tackles the issue of computing Wasserstein gradient direction for 2-layers NNs thanks to a SDP approach with main properties that there is no need to train the underlying NN. The gradient direction is computed with the dual formulation of a least-square + a polynomial regularization term. Strenghts - Thi...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper tackles the issue of computing Wasserstein gradient direction for 2-layers NNs thanks to a SDP approach with main properties that there is no need to train the underlying NN. The gradient direction is computed with the dual formulation of a least-square + a polynomial regularization term. **Strenghts...
This work claims that contrastive learning can find a minimax-optimal representation for linear predictors when the prediction function is approximately view-invariant. More precisely, the authors demonstrate that learning a representation via contrastive losses such as NT-XEnt, NT-Logistic, and Spectral can be seen ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This work claims that contrastive learning can find a minimax-optimal representation for linear predictors when the prediction function is approximately view-invariant. More precisely, the authors demonstrate that learning a representation via contrastive losses such as NT-XEnt, NT-Logistic, and Spectral can ...
This paper proposes an explanation method to provide consistent and truthful explanations of black-box models. Experimental results verify the effectiveness of the proposed method. [Strengths] 1. This paper focused on an important topic, i.e., learning consistent and truthful explanations. 2. This paper is well written...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proposes an explanation method to provide consistent and truthful explanations of black-box models. Experimental results verify the effectiveness of the proposed method. [Strengths] 1. This paper focused on an important topic, i.e., learning consistent and truthful explanations. 2. This paper is well...
This paper proposed an interesting idea of turning two unimodal pre-trained models into a multi-modal model without training. Specifically, each sample is encoded into a relative representation with respect to paired multi-modal data in its own modality. The relative representation serves as the bridge between two moda...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed an interesting idea of turning two unimodal pre-trained models into a multi-modal model without training. Specifically, each sample is encoded into a relative representation with respect to paired multi-modal data in its own modality. The relative representation serves as the bridge between ...
This work proposes an approach for solving the inverse problem via diffusion models. Particularly, the work considers structured noise that contaminates the observations. To solve the problem, another line of the diffusion process that describes noise is used except for the image generative process. The score models ...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This work proposes an approach for solving the inverse problem via diffusion models. Particularly, the work considers structured noise that contaminates the observations. To solve the problem, another line of the diffusion process that describes noise is used except for the image generative process. The score...
The authors consider domain generalization as a non-random sample selection problem. It is the first paper of kind to utilize this method for DG problem. Via assessment on both simulation and benchmarking data sets, the authors demonstrate the efficacy of our method both theoretically and empirically on simulated data ...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The authors consider domain generalization as a non-random sample selection problem. It is the first paper of kind to utilize this method for DG problem. Via assessment on both simulation and benchmarking data sets, the authors demonstrate the efficacy of our method both theoretically and empirically on simulat...
This paper proposes a pre-training framework, called VALM, to jointly train on image-text data. The novelty of this work, compared to previous works in similar field, is how the image-text pairs are created. While previous works use pre-curated image-text aligned pairs, this work instead uses images retrieved using tex...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a pre-training framework, called VALM, to jointly train on image-text data. The novelty of this work, compared to previous works in similar field, is how the image-text pairs are created. While previous works use pre-curated image-text aligned pairs, this work instead uses images retrieved u...
The paper proposes gamma frailty for modulating proportional hazards (NFM-PF) or nonparametric hazards (NFM-FN) functions. The assumed hazard functions are modeled with neural networks where model parameters are learned via maximum likelihood. Additionally, the paper provides theoretical convergence analysis for th...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper proposes gamma frailty for modulating proportional hazards (NFM-PF) or nonparametric hazards (NFM-FN) functions. The assumed hazard functions are modeled with neural networks where model parameters are learned via maximum likelihood. Additionally, the paper provides theoretical convergence analysi...
The authors revisit the paradigm of supervised learning with recent successful breakthroughs in self-supervised learning, including multi-crop data augmentation [Caron et al., 2020], expendable projector head [Chen et al., 2020a], and a variant of nearest class means classifier [Mensink et al., 2012]. They found that t...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors revisit the paradigm of supervised learning with recent successful breakthroughs in self-supervised learning, including multi-crop data augmentation [Caron et al., 2020], expendable projector head [Chen et al., 2020a], and a variant of nearest class means classifier [Mensink et al., 2012]. They foun...
This paper considers the problem of imputing the missing attributes of nodes in a network.This is done by identifying the most informative nodes for renewing the training set after each training epoch. Some uncertainty and representativeness metrics are proposed to achieve the task of identification. Their method achie...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper considers the problem of imputing the missing attributes of nodes in a network.This is done by identifying the most informative nodes for renewing the training set after each training epoch. Some uncertainty and representativeness metrics are proposed to achieve the task of identification. Their meth...
The authors proposed a method that enables us to learn a broader family of stochastic dynamics by directly learning an underlying mechanism to move samples in time without modeling the distributions at each time-step. The contributions are as follows: 1. Action Matching relies only on samples and does not require any ...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: The authors proposed a method that enables us to learn a broader family of stochastic dynamics by directly learning an underlying mechanism to move samples in time without modeling the distributions at each time-step. The contributions are as follows: 1. Action Matching relies only on samples and does not requ...
This paper proposes a method for constraint generation within the branch-and-cut framework. Differing from prior work, the paper casts the constraint generation problem as an MDP and proposes a hierarchical policy to generate a tuple of cuts of variable length depending on problem state. The constraint generation polic...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a method for constraint generation within the branch-and-cut framework. Differing from prior work, the paper casts the constraint generation problem as an MDP and proposes a hierarchical policy to generate a tuple of cuts of variable length depending on problem state. The constraint generati...
This paper proposed to use denoising as pretraining for SE(3)-invariant neural network models. The main idea is to add some noise to the coordinates, and use a score model to predict the added noise. The pretrained model are used in several different tasks, including molecule property prediction, force field and bindin...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposed to use denoising as pretraining for SE(3)-invariant neural network models. The main idea is to add some noise to the coordinates, and use a score model to predict the added noise. The pretrained model are used in several different tasks, including molecule property prediction, force field an...
This paper presents an approach for real-time demoire on mobile devices. Demoireing is an interesting (and hard) problem of removing moire patterns from images, which can easily appear due to high frequency patterns in the scene interfering with the camera color filter array. It is hard because the moire patterns spa...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents an approach for real-time demoire on mobile devices. Demoireing is an interesting (and hard) problem of removing moire patterns from images, which can easily appear due to high frequency patterns in the scene interfering with the camera color filter array. It is hard because the moire patt...
This paper proposes to model time series data as discretized observations from an underlying continuous stochastic process using a diffusion framework. The main contribution of the paper is to extend the method proposed in [1] and [2] to the case where a stochastic noise process is used for the forward process. Based o...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposes to model time series data as discretized observations from an underlying continuous stochastic process using a diffusion framework. The main contribution of the paper is to extend the method proposed in [1] and [2] to the case where a stochastic noise process is used for the forward process....
The paper explored the style-sensitive channels for effective style synthesis and proposed Normalization Perturbation for robust object detection. Their method generalizes well under some OOD detection datasets, like Foggy Cityscapes. Strength - The proposed Normalization Perturbation (NP) method could synthesize vario...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper explored the style-sensitive channels for effective style synthesis and proposed Normalization Perturbation for robust object detection. Their method generalizes well under some OOD detection datasets, like Foggy Cityscapes. Strength - The proposed Normalization Perturbation (NP) method could synthesi...
This paper studies the biases in the VCR dataset. It first identifies two important bias problems, by exploiting which significantly better accuracy can be obtained than the random guess. The authors then propose to synthesize counterfactual text and image data to debias the VCR dataset. An intra-sample contrastive lea...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: This paper studies the biases in the VCR dataset. It first identifies two important bias problems, by exploiting which significantly better accuracy can be obtained than the random guess. The authors then propose to synthesize counterfactual text and image data to debias the VCR dataset. An intra-sample contras...
This paper proposes an approach for GNNs to jointly prune and sub-sample neighboring nodes (importance sampling) and provides theoretical and empirical evaluation to show the improvements in sample complexity and converge rate. The paper provides a theoretical generalization analysis of GNN training with sparsified d...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proposes an approach for GNNs to jointly prune and sub-sample neighboring nodes (importance sampling) and provides theoretical and empirical evaluation to show the improvements in sample complexity and converge rate. The paper provides a theoretical generalization analysis of GNN training with spar...
The authors present an extension to convolutional networks (ConvNets) tailored to simulate the behavior of hierarchical transformers, aiming to provide an alternative to visional transformers (ViT) with a significantly reduced number of trainable parameters. The architectural variations in the proposed model focused on...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors present an extension to convolutional networks (ConvNets) tailored to simulate the behavior of hierarchical transformers, aiming to provide an alternative to visional transformers (ViT) with a significantly reduced number of trainable parameters. The architectural variations in the proposed model fo...
In this paper, a new method of sample-based NAS is proposed, Falcon. The key idea is to represent the model design space as a graph, where nodes correspond to separate designs and an edge corresponds to a minimal change in one of the design parameters. The problem of architecture search is thus translated to a black-bo...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this paper, a new method of sample-based NAS is proposed, Falcon. The key idea is to represent the model design space as a graph, where nodes correspond to separate designs and an edge corresponds to a minimal change in one of the design parameters. The problem of architecture search is thus translated to a ...
Authors are the first to show that zero-shot image recognition models built on top of CLIP are still susceptible to adversarial attacks, and that standard adversarial training objective, while effective at preventing adversarial attacks, destroys the rich image-language capability of CLIP, making the defended model use...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: Authors are the first to show that zero-shot image recognition models built on top of CLIP are still susceptible to adversarial attacks, and that standard adversarial training objective, while effective at preventing adversarial attacks, destroys the rich image-language capability of CLIP, making the defended m...
The authors propose the utilization of the concept of Fragmentation-and-Recall to solve spatial and reinforcement learning problems. In the former case, which they call FarMap, they address the use case of exploration in sub-map based SLAM and in the latter case, which they call FarCuriosity, the problem of catastrophi...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The authors propose the utilization of the concept of Fragmentation-and-Recall to solve spatial and reinforcement learning problems. In the former case, which they call FarMap, they address the use case of exploration in sub-map based SLAM and in the latter case, which they call FarCuriosity, the problem of cat...
The authors present study of the role of multi-agent topology on innovation towards goal of clarifying which social network structures are optimal for which innovation tasks, and which properties of experience sharing improve multi-level innovation. For multi-level hierarchical problem setting (WordCraft), three differ...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The authors present study of the role of multi-agent topology on innovation towards goal of clarifying which social network structures are optimal for which innovation tasks, and which properties of experience sharing improve multi-level innovation. For multi-level hierarchical problem setting (WordCraft), thre...
The paper proposes two tricks to optimize the training of slot attention. First, it initializes the query with learnable embedding instead of sampling from a learnable Gaussian distribution. Second, it applies bi-level optimization to the training. In practice, the slot binding process serves as the optimization of an ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper proposes two tricks to optimize the training of slot attention. First, it initializes the query with learnable embedding instead of sampling from a learnable Gaussian distribution. Second, it applies bi-level optimization to the training. In practice, the slot binding process serves as the optimizatio...
One of the key challenges in learning from private data sources in a collaborative manner as in federated learning is the data heterogeneity (non-IIDness) among data shards. This causes drift among local models which slows down the convergences or entails a poor generalization on local distributions. This issue is furt...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: One of the key challenges in learning from private data sources in a collaborative manner as in federated learning is the data heterogeneity (non-IIDness) among data shards. This causes drift among local models which slows down the convergences or entails a poor generalization on local distributions. This issue...
This paper proposes a domain dual branch network for domain generalization. The first contribution is to empirically verify which layers should be shared between domain and target classification. Then, authors propose a new augmentation strategy for domain augmentation. Experiments using ResNet-18 demonstrates its effe...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a domain dual branch network for domain generalization. The first contribution is to empirically verify which layers should be shared between domain and target classification. Then, authors propose a new augmentation strategy for domain augmentation. Experiments using ResNet-18 demonstrates ...
In this paper, the authors propose a novel personalized method for FCO via the forward-backward envelope. Personalized models are firstly updated, then the local models are updated with Second-order optimization. After the local update, the global update is done via FedAVG. Convergence results of the proposed algorithm...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: In this paper, the authors propose a novel personalized method for FCO via the forward-backward envelope. Personalized models are firstly updated, then the local models are updated with Second-order optimization. After the local update, the global update is done via FedAVG. Convergence results of the proposed a...
This paper exposes the hidden topology behind a pretrained BERT language model by training a CRF-VAE, which describes latent states and their relationships. It shows that many of the discovered states are linguistically meaningful. State transitions are also interpretable and can be followed to construct sentences. Str...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper exposes the hidden topology behind a pretrained BERT language model by training a CRF-VAE, which describes latent states and their relationships. It shows that many of the discovered states are linguistically meaningful. State transitions are also interpretable and can be followed to construct senten...
This paper proposes a version of hyperbolic GCN based Lorentz model without resorting to the tangent space. Some of the layer operations exist in prior work. Experiments are conducted on standard small-scale citation networks and tree datasets. There are recently many efforts on removing the tangent space of hyperboli...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a version of hyperbolic GCN based Lorentz model without resorting to the tangent space. Some of the layer operations exist in prior work. Experiments are conducted on standard small-scale citation networks and tree datasets. There are recently many efforts on removing the tangent space of h...
This work aims to examine the quality of different types of input embeddings for GNNs, and design a model architecture by combining both GNNs and non-GNN models (which are referred to as unconnected models in the paper). The authors provide empirical results to demonstrate the importance of having a high-quality input ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work aims to examine the quality of different types of input embeddings for GNNs, and design a model architecture by combining both GNNs and non-GNN models (which are referred to as unconnected models in the paper). The authors provide empirical results to demonstrate the importance of having a high-qualit...