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 provides a mid level vision feedback module as an Add-On for CNN's. Authors argue that such mid-level feedback properties improve CNN performance (though perhaps for the wrong reasons), and also argue that this idea is worth pursuing given neuroscience/perceptual psychology motivated ideas. See below for Mai...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper provides a mid level vision feedback module as an Add-On for CNN's. Authors argue that such mid-level feedback properties improve CNN performance (though perhaps for the wrong reasons), and also argue that this idea is worth pursuing given neuroscience/perceptual psychology motivated ideas. See below...
This paper introduces a new method for jointly learning a predictor and an interpretability model. The method offers a general framework that has LEX, INVASE, and REAL-X as special cases. The paper also introduces two more datasets synthetically created from FashionMNIST and CelebA with ground truth selected features. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper introduces a new method for jointly learning a predictor and an interpretability model. The method offers a general framework that has LEX, INVASE, and REAL-X as special cases. The paper also introduces two more datasets synthetically created from FashionMNIST and CelebA with ground truth selected fe...
This paper proposes to use key points as internal representations of scenes and model forward dynamics with graph interaction networks. The authors used a multi-object manipulation testbed for evaluating the proposed Keypoint Interaction Network (KINet). They show that KINet can achieve similar results for forward pred...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes to use key points as internal representations of scenes and model forward dynamics with graph interaction networks. The authors used a multi-object manipulation testbed for evaluating the proposed Keypoint Interaction Network (KINet). They show that KINet can achieve similar results for forw...
The paper deals with the topic of mutual information (MI) maximization in multi-view self-supervised learning (SSL). The paper takes an information-theoretic perspective and shows that many current self-supervised learning methods maximize a lower bound on the MI between the representations of different views---a resul...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper deals with the topic of mutual information (MI) maximization in multi-view self-supervised learning (SSL). The paper takes an information-theoretic perspective and shows that many current self-supervised learning methods maximize a lower bound on the MI between the representations of different views--...
This paper focuses on addressing unsupervised anomaly detection via disentangled conditional VAE. The new architecture combines three core components: beta-VAE, CVAE and the principle of TC. The authors claim that the new method improves the disentanglement of latent features, and the ability to detect anomalies. Mu...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper focuses on addressing unsupervised anomaly detection via disentangled conditional VAE. The new architecture combines three core components: beta-VAE, CVAE and the principle of TC. The authors claim that the new method improves the disentanglement of latent features, and the ability to detect anomal...
This paper presents a neural network brain encoding analysis comparing the performance of Transformer architectures trained on language modeling objectives versus those trained on a narrative summarization objective. The authors argue that the narrative summarization task is a proxy for "deeper understanding," and that...	Recommendation: 8: accept, good paper	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper presents a neural network brain encoding analysis comparing the performance of Transformer architectures trained on language modeling objectives versus those trained on a narrative summarization objective. The authors argue that the narrative summarization task is a proxy for "deeper understanding," ...
This paper proposes a method for learning a Bregman divergence using neural networks. A Bregman divergence is the divergence defined using a convex function, and any convex function has its corresponding Bregman divergence. The proposed method learns the divergence by representing the convex function by an Input Convex...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a method for learning a Bregman divergence using neural networks. A Bregman divergence is the divergence defined using a convex function, and any convex function has its corresponding Bregman divergence. The proposed method learns the divergence by representing the convex function by an Inpu...
In high dimensional setting, it is typically impossible to approximate the solution $\rho_t$ to a Fokker Planck Equation (FPE) with conventional grid-based methods. A standard method consists in considering the associated SDE a simulate many trajectories from this SDE to collect statistics of the solution at any futur...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: In high dimensional setting, it is typically impossible to approximate the solution $\rho_t$ to a Fokker Planck Equation (FPE) with conventional grid-based methods. A standard method consists in considering the associated SDE a simulate many trajectories from this SDE to collect statistics of the solution at a...
This paper proposes a faster algorithm, DMCMC, for integrating the reverse SDE/ODE associated with diffusion models. Unlike many existing works, DMCMC samples in the joint space of (x, sigma). Gibbs sampling is adopted for sampling in the joint space. Sampling x \| sigma is just ordinary Langevin dynamics, and sampling ...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposes a faster algorithm, DMCMC, for integrating the reverse SDE/ODE associated with diffusion models. Unlike many existing works, DMCMC samples in the joint space of (x, sigma). Gibbs sampling is adopted for sampling in the joint space. Sampling x \| sigma is just ordinary Langevin dynamics, and s...
This paper presented CBP-QSNN, which applies the CBP algorithm proposed by Kim & Jeong in a recent study of quantized DNN as a general post-training method to quantized SNN. The authors began by introducing key elements of CBP, including the Lagrangian objective function and the weight-constraint function cs(w) in it, ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper presented CBP-QSNN, which applies the CBP algorithm proposed by Kim & Jeong in a recent study of quantized DNN as a general post-training method to quantized SNN. The authors began by introducing key elements of CBP, including the Lagrangian objective function and the weight-constraint function cs(w)...
The paper aims to learn multiple RL tasks with the same network. SNIP is used to prune all but the 5% most "relevant" parameters of the network, yielding a mask for every task. Tasks can share parameters, and therefore all tasks have to be trained simultaneously. The authors show that SNIP depends strongly on the data ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper aims to learn multiple RL tasks with the same network. SNIP is used to prune all but the 5% most "relevant" parameters of the network, yielding a mask for every task. Tasks can share parameters, and therefore all tasks have to be trained simultaneously. The authors show that SNIP depends strongly on t...
This paper proposes a new architecture for Neural Processes named Versatile Neural Process (VNP). This new architecture is build as an improvement over previous Attentive Neural Process (ANP) with two main modifications: 1. The context points are pre-processed by a encoder module built using set convolutions and self-a...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes a new architecture for Neural Processes named Versatile Neural Process (VNP). This new architecture is build as an improvement over previous Attentive Neural Process (ANP) with two main modifications: 1. The context points are pre-processed by a encoder module built using set convolutions an...
This paper presents a balanced sampling data augmentation technique, for addressing the multi-domain and long-tailed learning problems in which the number of samples in different classes is imbalanced and potential domain shift. First, the class-specific and domain-specific representations are decoupled by instance-nor...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a balanced sampling data augmentation technique, for addressing the multi-domain and long-tailed learning problems in which the number of samples in different classes is imbalanced and potential domain shift. First, the class-specific and domain-specific representations are decoupled by inst...
The paper presents an interesting training algorithm for training SNNs from scratch. It uses a combination of skip connections and BN tools to get better accuracy on image recognition tasks. + Very comprehensive results + Simple yet effective idea - Since the authors use a BN technique, I am wondering if the authors...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The paper presents an interesting training algorithm for training SNNs from scratch. It uses a combination of skip connections and BN tools to get better accuracy on image recognition tasks. + Very comprehensive results + Simple yet effective idea - Since the authors use a BN technique, I am wondering if the...
Based Gaussian process, this paper develops a zeroth-order optimization algorithm which requires fewer function queries to estimate the gradient than previous zeroth-order methods. Additionally, a so-called dynamic virtual update schemes is incorporated. Theoretically, the proposed method is shown to obtain gradient wi...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: Based Gaussian process, this paper develops a zeroth-order optimization algorithm which requires fewer function queries to estimate the gradient than previous zeroth-order methods. Additionally, a so-called dynamic virtual update schemes is incorporated. Theoretically, the proposed method is shown to obtain gra...
This paper tackles an imbalanced semi-supervised learning (SSL) problem by proposing a new way to construct the pseudo labels of unlabeled samples, coined INlier Pseudo-Labeling (INPL). Unlike the previous approaches, which use the softmax confidence from the training classifier, INPL instead uses the energy score calc...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper tackles an imbalanced semi-supervised learning (SSL) problem by proposing a new way to construct the pseudo labels of unlabeled samples, coined INlier Pseudo-Labeling (INPL). Unlike the previous approaches, which use the softmax confidence from the training classifier, INPL instead uses the energy sc...
This paper proposed a new way of acceleration diffusion based generative models that combines MCMC on a augmented space with reverse-S/ODE integrators. More specifically, a Gibbs sampler procedure is introduced to traverse the augmented space based on a pre-trained classifier. Numerical experiments demonstate the effec...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposed a new way of acceleration diffusion based generative models that combines MCMC on a augmented space with reverse-S/ODE integrators. More specifically, a Gibbs sampler procedure is introduced to traverse the augmented space based on a pre-trained classifier. Numerical experiments demonstate t...
This paper proposes a novel adaptive depth network by inducing a training strategy without the need of additional intermediate gate/classifier. It use a skip-aware BN, with a cost of 0.07% parameter increasing, but reduces the inference cost and achieve better results comparing with the non-adaptive baselines. Pro: 1) ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a novel adaptive depth network by inducing a training strategy without the need of additional intermediate gate/classifier. It use a skip-aware BN, with a cost of 0.07% parameter increasing, but reduces the inference cost and achieve better results comparing with the non-adaptive baselines. ...
This paper proves pointwise convergence of q-replicator dynamics to NE and corresponding bounds on average price of anarchy, generalizing previous works. Strengths: The results are solid. The motivation and proof ideas are well explained. Weaknesses: I'm not familiar with this specific topic studied in this paper, ...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves pointwise convergence of q-replicator dynamics to NE and corresponding bounds on average price of anarchy, generalizing previous works. Strengths: The results are solid. The motivation and proof ideas are well explained. Weaknesses: I'm not familiar with this specific topic studied in this...
The paper introduces a novel task-incremental approach that diminishes the gradients based on their importance (SPG). Specifically, the importance is calculated as the normalized gradients on the current and past tasks. For the current task, the gradients are taken w.r.t. the log-likelihood of the data. For past tasks...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper introduces a novel task-incremental approach that diminishes the gradients based on their importance (SPG). Specifically, the importance is calculated as the normalized gradients on the current and past tasks. For the current task, the gradients are taken w.r.t. the log-likelihood of the data. For pa...
The paper provides a unified setting among generative models. In particular it offers a perspective on diffusion models (DM) that cast them in a framework more similar to GANs, which admit a Gaussian formulation of the latent space. This point of view is then leveraged in three settings. Contributions: 1. provide a f...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper provides a unified setting among generative models. In particular it offers a perspective on diffusion models (DM) that cast them in a framework more similar to GANs, which admit a Gaussian formulation of the latent space. This point of view is then leveraged in three settings. Contributions: 1. pro...
This paper presents graph contrastive learning approach with counterfactual views. Specifically, the authors propose a counterfactual inference objective to generate task-oriented causal views that will be helpful in determining positive and negative samples. Experimental results show superiority over existing graph co...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper presents graph contrastive learning approach with counterfactual views. Specifically, the authors propose a counterfactual inference objective to generate task-oriented causal views that will be helpful in determining positive and negative samples. Experimental results show superiority over existing ...
The paper proposes improved sampling for training program surrogates. The proposed sampling scheme takes into account both data distribution and sample complexity of different paths. A language suitable for application of the proposed sampling scheme is introduced, and the scheme is evaluated on a graphics program. Str...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: The paper proposes improved sampling for training program surrogates. The proposed sampling scheme takes into account both data distribution and sample complexity of different paths. A language suitable for application of the proposed sampling scheme is introduced, and the scheme is evaluated on a graphics prog...
This paper proposes a channel pruning method which aims to enhance global consistence among samples for network channels. It claims integrating both static and dynamic information into the formulation. Comparison on small datasets demonstrate the effectiveness of the proposed methods. Strong points + The bayesian b...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a channel pruning method which aims to enhance global consistence among samples for network channels. It claims integrating both static and dynamic information into the formulation. Comparison on small datasets demonstrate the effectiveness of the proposed methods. Strong points + The ba...
The paper presents a novel architecture for graph neural networks called MeGraph. MeGraph builds on the Select-Reduce-Connect framework for graph pooling (Grattarola et al., 2022). In particular, the Select function of a pooling method implicitly induces a bipartite "inter-graph" between the input and output graph. B...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper presents a novel architecture for graph neural networks called MeGraph. MeGraph builds on the Select-Reduce-Connect framework for graph pooling (Grattarola et al., 2022). In particular, the Select function of a pooling method implicitly induces a bipartite "inter-graph" between the input and output ...
The paper proposes a theoretical analysis of SGMs that extends Chen et al. (2022c) and the bad set approach from (Lee et al., 2022a). Compared to Chen et al. the analysis is much more general in a number of ways, and compared to Lee et al. the analysis relaxes an LSI assumption and is applicable to a larger class of g...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper proposes a theoretical analysis of SGMs that extends Chen et al. (2022c) and the bad set approach from (Lee et al., 2022a). Compared to Chen et al. the analysis is much more general in a number of ways, and compared to Lee et al. the analysis relaxes an LSI assumption and is applicable to a larger cl...
This article proposed a distill multiple steps method to decrease the number of required steps to speed up iterative non-autoregressive transformer. Strengths: 1. The training of the distilled student model obtained improved training efficiency with a certain translation quality. 2. The proposed approach gained impr...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This article proposed a distill multiple steps method to decrease the number of required steps to speed up iterative non-autoregressive transformer. Strengths: 1. The training of the distilled student model obtained improved training efficiency with a certain translation quality. 2. The proposed approach gai...
This paper proposes an end-to-end visual learning method for 3D robotic manipulation. The authors construct an SE(3)-equivariant energy-based model to learn end-to-end from limited demonstrations without prior knowledge. The proposed Equivariant Descriptor Fields could be generalized to unseen poses, instances, and tar...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes an end-to-end visual learning method for 3D robotic manipulation. The authors construct an SE(3)-equivariant energy-based model to learn end-to-end from limited demonstrations without prior knowledge. The proposed Equivariant Descriptor Fields could be generalized to unseen poses, instances,...
This paper proposes a defense against backdoor attacks to FL. The assumption is that the system has access to some clean validation dataset, which can be used to filter backdoored model updates from malicious clients via analyzing the output layer distribution. Some evaluation is performed to show the effectiveness of ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a defense against backdoor attacks to FL. The assumption is that the system has access to some clean validation dataset, which can be used to filter backdoored model updates from malicious clients via analyzing the output layer distribution. Some evaluation is performed to show the effective...
In this paper, the authors studied the optimization and generalization of two-layer ReLU network when the first layer neuron rotate little. Three regimes are considered, including the near initialization regime, Neural Collapse regime and non-rotate regime. The results show that low test error could be achieved and hav...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors studied the optimization and generalization of two-layer ReLU network when the first layer neuron rotate little. Three regimes are considered, including the near initialization regime, Neural Collapse regime and non-rotate regime. The results show that low test error could be achieved...
The paper aims at learning whether a certain text is visualizable or not as generative models such as DALLE-2 are becoming very popular. To this effort, the authors curate a dataset called Text Imageability Dataset (TIMED) which contains 3620 sentences with label for whether it is a visual or non-visual text. The autho...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper aims at learning whether a certain text is visualizable or not as generative models such as DALLE-2 are becoming very popular. To this effort, the authors curate a dataset called Text Imageability Dataset (TIMED) which contains 3620 sentences with label for whether it is a visual or non-visual text. T...
This paper works on image restoration and introduces a multi-branch dynamic selective frequency (MDSF) module with learned kernels to generate high and low-frequency features and a multi-branch compact selective frequency (MCSF) module to enhance the features with an attention mechanism. The proposed method achieves se...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper works on image restoration and introduces a multi-branch dynamic selective frequency (MDSF) module with learned kernels to generate high and low-frequency features and a multi-branch compact selective frequency (MCSF) module to enhance the features with an attention mechanism. The proposed method ach...
This work proposes a backdoor attack which does not require modifying the inputs and only requires modifying labels. The way it operates is by trigger selection where a trigger is an object already in the training images. The work also tests against several defenses. I generally like the idea of this paper, and the e...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work proposes a backdoor attack which does not require modifying the inputs and only requires modifying labels. The way it operates is by trigger selection where a trigger is an object already in the training images. The work also tests against several defenses. I generally like the idea of this paper, a...
In this paper, the authors introduced TAMiL, a continual-learning model inspired by the global workspace theory that can learn multiple tasks without catastrophic forgetting by constructing a common representation space across tasks. By combining previous approaches on self-regulated neurogenesis and experience replay,...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this paper, the authors introduced TAMiL, a continual-learning model inspired by the global workspace theory that can learn multiple tasks without catastrophic forgetting by constructing a common representation space across tasks. By combining previous approaches on self-regulated neurogenesis and experience...
The present paper is concerned about measuring how much a set of molecules covers the chemical space. All of the metrics that they consider take a set of molecules $\mathcal{S}$ as input and output a coverage metric. Among many existing heuristically-designed metrics, the authors propose a metric called #Circles, which...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The present paper is concerned about measuring how much a set of molecules covers the chemical space. All of the metrics that they consider take a set of molecules $\mathcal{S}$ as input and output a coverage metric. Among many existing heuristically-designed metrics, the authors propose a metric called #Circle...
This work considers neural operators in the semisupervised setting. It extends Fourier neural operator with a hypernetwork structure, which allows composition in time. The paper provides numerical experiments on 1d Burger equation and 2d and 3d Navier-Stokes equation. The proposed method has a slightly improved error r...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work considers neural operators in the semisupervised setting. It extends Fourier neural operator with a hypernetwork structure, which allows composition in time. The paper provides numerical experiments on 1d Burger equation and 2d and 3d Navier-Stokes equation. The proposed method has a slightly improved...
This is a theoretical paper on an NTK analysis of two-layer ReLU network with a trainable nontrivial initial bias. The motivation of study is that by using an appropriately initialized bias term, the network produces sparse activation at initialization; and if one considers the kernel regime then the sparsity is main...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This is a theoretical paper on an NTK analysis of two-layer ReLU network with a trainable nontrivial initial bias. The motivation of study is that by using an appropriately initialized bias term, the network produces sparse activation at initialization; and if one considers the kernel regime then the sparsity...
The paper proposes a simple complexity-based prompting method for chain-of-thought prompting. The idea is to prompt with more complex questions: The authors use example questions with more reasoning steps (9 per question). When combining the predictions from multiple prompts, they choose the top K predictions with the ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a simple complexity-based prompting method for chain-of-thought prompting. The idea is to prompt with more complex questions: The authors use example questions with more reasoning steps (9 per question). When combining the predictions from multiple prompts, they choose the top K predictions w...
This paper proposes an architecture for solving CSPs based on a transformer network. The transformer makes multiple passes over the inputs, keeping information across different passes in the form of a recurrent state. The network is also augmented with a constraint loss that injects problem knowledge into the training ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an architecture for solving CSPs based on a transformer network. The transformer makes multiple passes over the inputs, keeping information across different passes in the form of a recurrent state. The network is also augmented with a constraint loss that injects problem knowledge into the t...
The paper is concerned with the scalability of training GNNs on large graph datasets where there may not be sufficient memory to store the graph structure. The basic idea is to decompose the graph into smaller components. The paper builds upon the earlier works such as (Chiang et al., 2019), which uses METIS to decompo...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper is concerned with the scalability of training GNNs on large graph datasets where there may not be sufficient memory to store the graph structure. The basic idea is to decompose the graph into smaller components. The paper builds upon the earlier works such as (Chiang et al., 2019), which uses METIS to...