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 examines data poisoning attacks and defense for joint language-vision model (CLIP) in a retrieval setting. Expanding on Carlini et al. 2022, the authors propose attacks and defenses on both modalities instead of just vision signal and show vulnerability in both. The paper proposes three attacks, all try to l...	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 examines data poisoning attacks and defense for joint language-vision model (CLIP) in a retrieval setting. Expanding on Carlini et al. 2022, the authors propose attacks and defenses on both modalities instead of just vision signal and show vulnerability in both. The paper proposes three attacks, all ...
The paper proposes a Blurring Diffusion Model. It is a type of diffusion model based on heat dissipation, or blurring. The model obtains better performance compared with some SOTA models (e.g., Denoising Diffusion, IHDM). Strength: (1) It makes sense for combining heat dissipation and blurring into the diffusion model....	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: The paper proposes a Blurring Diffusion Model. It is a type of diffusion model based on heat dissipation, or blurring. The model obtains better performance compared with some SOTA models (e.g., Denoising Diffusion, IHDM). Strength: (1) It makes sense for combining heat dissipation and blurring into the diffusio...
This paper proposed a new pipeline that includes 3 stages for detecting out-of-distribution data. First, a graph is generated for an image by using its output from a pre-trained object-detection network. Secondly, the graph-kernel-based method is used to generate the whole-graph embedding from the graph. Finally, given...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposed a new pipeline that includes 3 stages for detecting out-of-distribution data. First, a graph is generated for an image by using its output from a pre-trained object-detection network. Secondly, the graph-kernel-based method is used to generate the whole-graph embedding from the graph. Finall...
This paper considers the optimization of the logistic regression problem with separable data assumption. They observe that when the iteration variable is far from zero, the smoothness parameter decreases (i.e., the function is smoother), which allows more aggressively long step size. They prove the linear convergence o...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers the optimization of the logistic regression problem with separable data assumption. They observe that when the iteration variable is far from zero, the smoothness parameter decreases (i.e., the function is smoother), which allows more aggressively long step size. They prove the linear conve...
This paper presents a novel adversarial framework to train driving policies under learned traffic flow in simulation. It first extends traffic flow learned with reinforcement learning by exposing each other’s intrinsic social value orientations (SVO). Next, it uses this exposed SVO to train adversarial agents as well a...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents a novel adversarial framework to train driving policies under learned traffic flow in simulation. It first extends traffic flow learned with reinforcement learning by exposing each other’s intrinsic social value orientations (SVO). Next, it uses this exposed SVO to train adversarial agents a...
This manuscript introduces a framework to model data generation where the many bias-inducing factors are allowed for an exploration of the bias inheritance mechanism. Through the parametric framework, the authors analyze the data imbalance problem, investigate the various sources of bias, and propose a novel mitigation...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This manuscript introduces a framework to model data generation where the many bias-inducing factors are allowed for an exploration of the bias inheritance mechanism. Through the parametric framework, the authors analyze the data imbalance problem, investigate the various sources of bias, and propose a novel mi...
This paper studies the covariance structures of each convolution filter in deep learning models for classification problems. It builds a parametric model for these structures and proposes to directly sample a multivariate Gaussian distribution as an initialization for a family of CNNs. Numerical results show that in so...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies the covariance structures of each convolution filter in deep learning models for classification problems. It builds a parametric model for these structures and proposes to directly sample a multivariate Gaussian distribution as an initialization for a family of CNNs. Numerical results show th...
This paper presents an efficient model, PMLP, which achieves the GNN-level generalization and accuracy at the cost of training an MLP (which is much cheaper than training a GNN). Specifically, the training of PMLP is identical to that of a normal MLP, where edge information in the graph is completely ignored. During in...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper presents an efficient model, PMLP, which achieves the GNN-level generalization and accuracy at the cost of training an MLP (which is much cheaper than training a GNN). Specifically, the training of PMLP is identical to that of a normal MLP, where edge information in the graph is completely ignored. D...
This work is looking into the multi-tenant federated learning problem, which is considering training multiple federated learning tasks at the same time. The goal of this work is to minimize power consumption while maximizing the performance of training results. This work proposes using a multiple-task share-model techn...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work is looking into the multi-tenant federated learning problem, which is considering training multiple federated learning tasks at the same time. The goal of this work is to minimize power consumption while maximizing the performance of training results. This work proposes using a multiple-task share-mod...
The paper gives the test risk guarantee on the minimum $\ell_2$ norm zero-loss solution (via combination of Lemma 1 and Theorem 1), which is where overparameterized model converges to via SGD. The bound, when simplified, leads to several interesting observations such as double descent via task diversity. Authors also p...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper gives the test risk guarantee on the minimum $\ell_2$ norm zero-loss solution (via combination of Lemma 1 and Theorem 1), which is where overparameterized model converges to via SGD. The bound, when simplified, leads to several interesting observations such as double descent via task diversity. Author...
The authors present a method to automatically locate referents in scenes using a combination of embodied gesture signals and natural language descriptions. Their key contribution is the implementation of the "virtual touch-line", which is the extended line connecting the eye and the fingertip to localize objects in sce...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors present a method to automatically locate referents in scenes using a combination of embodied gesture signals and natural language descriptions. Their key contribution is the implementation of the "virtual touch-line", which is the extended line connecting the eye and the fingertip to localize object...
This paper presents a multiscale PDE surrogate, where a coarse-grained numerical model and a neural operator are added. The neural operator corrects for the errors which the numerical model makes. The paper promises grid independent solutions, and quasi-linear runtime complexity. Learning multiscale PDE surrogates is...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents a multiscale PDE surrogate, where a coarse-grained numerical model and a neural operator are added. The neural operator corrects for the errors which the numerical model makes. The paper promises grid independent solutions, and quasi-linear runtime complexity. Learning multiscale PDE surro...
In this paper, the mixed-precision layer problem is formulated as a traditional NP hard problem and the problem can be solved by low cost methods without fine-tuning. The experimental results show that the proposed method is better than the current SOTA method. Strengths: 1. Based on the total differential calculati...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: In this paper, the mixed-precision layer problem is formulated as a traditional NP hard problem and the problem can be solved by low cost methods without fine-tuning. The experimental results show that the proposed method is better than the current SOTA method. Strengths: 1. Based on the total differential c...
This paper studies if explicitly modeling the behavior policy is beneficial in offline RL. Focusing on offline RL methods based on value function penalizations, the authors argue that multiple previous methods all design the penalty in a way that reduces the chance of overestimating the value of out-of-support state-ac...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies if explicitly modeling the behavior policy is beneficial in offline RL. Focusing on offline RL methods based on value function penalizations, the authors argue that multiple previous methods all design the penalty in a way that reduces the chance of overestimating the value of out-of-support ...
This paper focuses on addressing the over-reliance on contextual biases, and proposes a new method based on the attention mechanism. It draws inspiration from the causal intervention, and leverages attention together with feature sampling/shuffling to realize the function. Experimental results show that the proposed me...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper focuses on addressing the over-reliance on contextual biases, and proposes a new method based on the attention mechanism. It draws inspiration from the causal intervention, and leverages attention together with feature sampling/shuffling to realize the function. Experimental results show that the pro...
The paper presents a self-attention based meta learning framework for the learned Gaussian ESs and tests the performance on neuroevolution tasks. The metaBBO module includes steps like Meta-sampling, Inner loop search, meta normalize, and Meta updating. Limited evaluation results on standard datasets were provided. S...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper presents a self-attention based meta learning framework for the learned Gaussian ESs and tests the performance on neuroevolution tasks. The metaBBO module includes steps like Meta-sampling, Inner loop search, meta normalize, and Meta updating. Limited evaluation results on standard datasets were provi...
This paper aims to explore the problem of data set imbalance in face identification systems. They specifically focus on imbalance with respect to gender presentation. Aiming at the imbalances(both in terms of identities or images per identity) in the train set and the test set, the author conducted several experiments ...	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 aims to explore the problem of data set imbalance in face identification systems. They specifically focus on imbalance with respect to gender presentation. Aiming at the imbalances(both in terms of identities or images per identity) in the train set and the test set, the author conducted several expe...
This article is interested in the so-called "Inversion effect" in human vision: the fact that although humans are as a rule excellent at recognising and discriminating faces, performance falls to very low levels when faces are seen upside down. This is an interesting effect as we are generally apt at recognising object...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This article is interested in the so-called "Inversion effect" in human vision: the fact that although humans are as a rule excellent at recognising and discriminating faces, performance falls to very low levels when faces are seen upside down. This is an interesting effect as we are generally apt at recognisin...
This paper use an information-theoretic framework to analyze the generalization ability of hypotheses and learning algorithm of unsupervised domain adapation problem. Strength: 1、The authors rigorously prove some generalization error bound. One of the bounds is associated only with the learning algorithm, which is very...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper use an information-theoretic framework to analyze the generalization ability of hypotheses and learning algorithm of unsupervised domain adapation problem. Strength: 1、The authors rigorously prove some generalization error bound. One of the bounds is associated only with the learning algorithm, which...
This paper investigates an improved version of Chain-of-thought prompting by leveraging self consistency and shows empirical evidence that their self consistency prompting indeed improve reasoning tasks performance compared with the original chain-of-thoguht prompting. Strength: 1. The self-consistency method is ver...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper investigates an improved version of Chain-of-thought prompting by leveraging self consistency and shows empirical evidence that their self consistency prompting indeed improve reasoning tasks performance compared with the original chain-of-thoguht prompting. Strength: 1. The self-consistency metho...
The paper proposes to use an iterative decoding scheme for entropy model based on masked image modeling to improve video compression. This scheme has the advantage that a fixed autoregressive order is not used and thus leading to a better modeling of the latents. Further, SWIN transformer based model is used to improve...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes to use an iterative decoding scheme for entropy model based on masked image modeling to improve video compression. This scheme has the advantage that a fixed autoregressive order is not used and thus leading to a better modeling of the latents. Further, SWIN transformer based model is used to...
The authors present a meta-learning framework to infer a system's or circuit's inductive bias. In their work, the authors claim that their method connects architectural design choices to function space features. This work demonstrates the usability of the method in inferring the inductive bias for relatively simple mod...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors present a meta-learning framework to infer a system's or circuit's inductive bias. In their work, the authors claim that their method connects architectural design choices to function space features. This work demonstrates the usability of the method in inferring the inductive bias for relatively si...
This paper proposes to train one-time-step SNNs with a Hoyer regularizer and Hoyer spike layer. The proposed Hoyer spike layer uses adaptive threshold based on the Hoyer extremum of membrane potentials, and a Hoyer regularization for membrane potentials is added in the loss function. Experiments on static image classif...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper proposes to train one-time-step SNNs with a Hoyer regularizer and Hoyer spike layer. The proposed Hoyer spike layer uses adaptive threshold based on the Hoyer extremum of membrane potentials, and a Hoyer regularization for membrane potentials is added in the loss function. Experiments on static image...
This paper is basically an implementation of the prospective configuration algoithm (Yuhang Song, et al, Inferring neural activity before plasticity: A foundation for learning beyond backpropagation) where each layer of a heirarchical Gaussian generative model (with the further assumption that the covariance matrix of ...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper is basically an implementation of the prospective configuration algoithm (Yuhang Song, et al, Inferring neural activity before plasticity: A foundation for learning beyond backpropagation) where each layer of a heirarchical Gaussian generative model (with the further assumption that the covariance ma...
The paper proposes a new method to build robust FL. The paper performs theoretical analysis and evaluation on multiple datasets and baselines. Strength + A new robust FL method is proposed. + Theoretical analysis is performed on the proposed method to show its robustness. Weakness - Some baseline attacks are not...	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 method to build robust FL. The paper performs theoretical analysis and evaluation on multiple datasets and baselines. Strength + A new robust FL method is proposed. + Theoretical analysis is performed on the proposed method to show its robustness. Weakness - Some baseline attacks...
This paper focuses on the analysis of hyperparameters in binary neural networks (BNN). The hyperparameters could explain the BNNs with latent real-valued weights during training. However, the magnitude of binary weights is not interpretable with these hyperparameters yet. This paper provides an interpretation of these ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper focuses on the analysis of hyperparameters in binary neural networks (BNN). The hyperparameters could explain the BNNs with latent real-valued weights during training. However, the magnitude of binary weights is not interpretable with these hyperparameters yet. This paper provides an interpretation o...
This paper introduces voxurf - a method for surface reconstruction that combines explicit voxel representations with SDF-based volume rendering. On top of naive combination of existing techniques, the key ideas are: employing a coarse-to-fine strategy, such that coarse geometry is estimated first; a dual color network,...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper introduces voxurf - a method for surface reconstruction that combines explicit voxel representations with SDF-based volume rendering. On top of naive combination of existing techniques, the key ideas are: employing a coarse-to-fine strategy, such that coarse geometry is estimated first; a dual color ...
This work introduces PaLI, a new large-scale vision-language pretraining model with multilingual enhancement. The architecture follows the previous scheme that leverages language pretraining as the main component which takes in vision and language feature tokens. The main contributions of this work are three folds: A n...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work introduces PaLI, a new large-scale vision-language pretraining model with multilingual enhancement. The architecture follows the previous scheme that leverages language pretraining as the main component which takes in vision and language feature tokens. The main contributions of this work are three fo...
This paper combines continuous time state-space models with transformer themes. The method is applied to large language and speech modelling tasks, and is compared to some baselines. The method appears to perform well. Some analysis of the temporal characteristics of the learned model are explored. As far as I can ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper combines continuous time state-space models with transformer themes. The method is applied to large language and speech modelling tasks, and is compared to some baselines. The method appears to perform well. Some analysis of the temporal characteristics of the learned model are explored. As far a...
This paper introduces Make-A-Video, a text-video generation approach trained based on a text-image model. The core idea of Make-A-Video is to take use of learned text-vision correlation from well-trained text-image model to accelerate the learning of text-video generation. It also claims that no paired text-video data ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper introduces Make-A-Video, a text-video generation approach trained based on a text-image model. The core idea of Make-A-Video is to take use of learned text-vision correlation from well-trained text-image model to accelerate the learning of text-video generation. It also claims that no paired text-vid...
This paper proposes an interpretation method for object detector. The authors slightly modify the existing Grad-CAM method to make it location-sensitive. In addition, ODAM-Train is also proposed for improved interpretation for overlapping objects. The experiments show that the proposed method provides better interpreta...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an interpretation method for object detector. The authors slightly modify the existing Grad-CAM method to make it location-sensitive. In addition, ODAM-Train is also proposed for improved interpretation for overlapping objects. The experiments show that the proposed method provides better in...
This paper describes an image captioning task to represent language acquisition through its structure. The learner or "speaker" is the image captioner. The feedback is provided via a "listener" that provides feedback in the form of a ground truth caption. The innovation comes from the inclusion, in the speaker model ...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper describes an image captioning task to represent language acquisition through its structure. The learner or "speaker" is the image captioner. The feedback is provided via a "listener" that provides feedback in the form of a ground truth caption. The innovation comes from the inclusion, in the speake...
The authors propose the use of sparse-variational GPs with a deep kernel and a transformer architecture to conduct hyperparameter optimization across different search spaces. They conduct an empirical comparison on three different tasks against most of the state-of-the-art. Strengths - Interesting idea - Many exper...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors propose the use of sparse-variational GPs with a deep kernel and a transformer architecture to conduct hyperparameter optimization across different search spaces. They conduct an empirical comparison on three different tasks against most of the state-of-the-art. Strengths - Interesting idea - Ma...
This paper proposes a generalisation of the family of linear networks (SGC, SSGC, APPNP, etc.) to heterophilous graphs. In contrast to existing methods which work well with homophilous networks, this work studies an objective which can work well with heterophilous graphs. Authors demonstrate the least squares based sol...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a generalisation of the family of linear networks (SGC, SSGC, APPNP, etc.) to heterophilous graphs. In contrast to existing methods which work well with homophilous networks, this work studies an objective which can work well with heterophilous graphs. Authors demonstrate the least squares b...
The work claims there is a trade-off between two properties of representation learning 1. When a model is pre-trained, the learned representation will be useful for downstream tasks with a small number of labels. 2. The representation is useful for different downstream tasks. i.e. the average accuracy on multiple downs...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The work claims there is a trade-off between two properties of representation learning 1. When a model is pre-trained, the learned representation will be useful for downstream tasks with a small number of labels. 2. The representation is useful for different downstream tasks. i.e. the average accuracy on multip...
This paper is a benchmark for out of distribution generalization in object detection. The authors create a benchmark to evaluate out of distribution generalization algorithms specifically for the task of object detection (object classification and bounding box regression), using multiple well known datasets in specific...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper is a benchmark for out of distribution generalization in object detection. The authors create a benchmark to evaluate out of distribution generalization algorithms specifically for the task of object detection (object classification and bounding box regression), using multiple well known datasets in ...
This paper proposes SAGE, a method that extracts semantic-aware global explanations that can be applied to NLP in general and specifically experiments on name entity recognition. ### Strengths: * The proposed method, SAGE, is potentially useful. ### Weaknesses: * The paper may need major revision in its writing. * T...	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 proposes SAGE, a method that extracts semantic-aware global explanations that can be applied to NLP in general and specifically experiments on name entity recognition. ### Strengths: * The proposed method, SAGE, is potentially useful. ### Weaknesses: * The paper may need major revision in its writin...
The paper addresses the problem of unsupervised semantic segmentation from images, which is relatively new in the literature. The methodology is based on saliency detection, clustering of self-supervised features (e.g., DINO) within the salient regions to obtain pseudo-masks, followed by a few iterations of self-traini...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper addresses the problem of unsupervised semantic segmentation from images, which is relatively new in the literature. The methodology is based on saliency detection, clustering of self-supervised features (e.g., DINO) within the salient regions to obtain pseudo-masks, followed by a few iterations of sel...
This paper addresses the problem of inferring individual treatment effects from observational data using representation learning via a discrepancy constraint between the representations of the treatment and control groups. In particular, the authors address the problem of dealing with the mini batch sampling which occu...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper addresses the problem of inferring individual treatment effects from observational data using representation learning via a discrepancy constraint between the representations of the treatment and control groups. In particular, the authors address the problem of dealing with the mini batch sampling wh...
The paper investigates the mechanism by which MLM pre-training may result if fine-tuned models that are more robust to spurious features. The paper proposes a setting where there is a spurious feature that can predict a classification decision with a simple decision boundary and high but non-perfect accuracy, and a ro...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper investigates the mechanism by which MLM pre-training may result if fine-tuned models that are more robust to spurious features. The paper proposes a setting where there is a spurious feature that can predict a classification decision with a simple decision boundary and high but non-perfect accuracy, ...