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 studies the implicit emergence of mapping in simple recurrent blind agents. Removing the visual modality allows authors to reduce the possible strategies to reach the specified goal and to thus isolate the construction of an internal map. As mentioned in the paper, no technical novelty is proposed, but a stu...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the implicit emergence of mapping in simple recurrent blind agents. Removing the visual modality allows authors to reduce the possible strategies to reach the specified goal and to thus isolate the construction of an internal map. As mentioned in the paper, no technical novelty is proposed, b...
This paper introduces an auto-encoder with the normalizing flow bottleneck called AE-FLOW for anomaly detection in medical images. The model combines the benefits of normalizing flow methods for computing the anomaly likelihood of extracted features at image level, and the interpretability of reconstruction-based metho...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces an auto-encoder with the normalizing flow bottleneck called AE-FLOW for anomaly detection in medical images. The model combines the benefits of normalizing flow methods for computing the anomaly likelihood of extracted features at image level, and the interpretability of reconstruction-bas...
This paper proposes to learn multiple parameters (with a cosine distance regularization to emphasize the diversity of the learned parameters). During a testing period, one can use the interpolated parameter to increase model robustness. The authors also prove a couple of important theorems that i) a compressed paramete...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to learn multiple parameters (with a cosine distance regularization to emphasize the diversity of the learned parameters). During a testing period, one can use the interpolated parameter to increase model robustness. The authors also prove a couple of important theorems that i) a compressed ...
The paper studies fundamental limits on the Lp robustness of a classifier. The main result is that under natural assumptions, any classifier on n x n images will be vulnerable to L2 perturbations of size O(sqrt(n)). The paper further shows that the result depends on the chosen bit-depth of images. Strengths: - Interest...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies fundamental limits on the Lp robustness of a classifier. The main result is that under natural assumptions, any classifier on n x n images will be vulnerable to L2 perturbations of size O(sqrt(n)). The paper further shows that the result depends on the chosen bit-depth of images. Strengths: - ...
This work is concerned with vision-language representation learning. Main contribution is a new distillation technique that uses stronger uni-modal encoders as teacher network, to improve the finetuning performance of pre-trained VL models. In specific, the work proposes Adaptive Distillation for Vision-Language (ADV...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work is concerned with vision-language representation learning. Main contribution is a new distillation technique that uses stronger uni-modal encoders as teacher network, to improve the finetuning performance of pre-trained VL models. In specific, the work proposes Adaptive Distillation for Vision-Langu...
This paper studies policy optimization for Markov Games to find the Nash equilibrium. The authors parameterize the policy by directly considering the probability for each state and action. The key contribution of this work is to propose a slow-fast framework with a series of conditions, and prove that any pair of slow ...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies policy optimization for Markov Games to find the Nash equilibrium. The authors parameterize the policy by directly considering the probability for each state and action. The key contribution of this work is to propose a slow-fast framework with a series of conditions, and prove that any pair ...
This paper presents an analysis of regression's reformulation as a classification problem and proposes an ordinal entropy loss to learn high-entropy feature representations which preserve ordinality based on the analysis. The proposed method outperforms the previous regression loss, classification loss, and other compa...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper presents an analysis of regression's reformulation as a classification problem and proposes an ordinal entropy loss to learn high-entropy feature representations which preserve ordinality based on the analysis. The proposed method outperforms the previous regression loss, classification loss, and oth...
The paper studies the subgraph localization problem -- given a target graph and a source graph, compute the alignment of the target graph in the source graph. The problem of alignment is closely related to that of detecting isomorphic subgraph in a source graph. The goal here is to compute an indicator function that id...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper studies the subgraph localization problem -- given a target graph and a source graph, compute the alignment of the target graph in the source graph. The problem of alignment is closely related to that of detecting isomorphic subgraph in a source graph. The goal here is to compute an indicator function...
The paper proposes a generic model to learn intrinsic and extrinsic factors from various contexts using a contrastive learning component, and a disentangling component. Experimental results on real-world datasets demonstrate the effectiveness of the method. Additionally, theoretical proofs are provided for the refined ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a generic model to learn intrinsic and extrinsic factors from various contexts using a contrastive learning component, and a disentangling component. Experimental results on real-world datasets demonstrate the effectiveness of the method. Additionally, theoretical proofs are provided for the ...
The paper uses image translation models to generate diverse biased views of images to debias models. For this, the paper observes that generative models find it easier to latch onto biases instead of the signal, which is consistent with the observation made by previous works for classification models. As such, when ask...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper uses image translation models to generate diverse biased views of images to debias models. For this, the paper observes that generative models find it easier to latch onto biases instead of the signal, which is consistent with the observation made by previous works for classification models. As such, ...
The paper proposes a novel Reinforcement Learning based unsupervised method for the task of goal planning for agent based exploration in a non task-centric paradigm. # Strengths: - (1) Readability. Overall, the readability is fair. The main ideas are well conveyed, justified and articulated while mostly providing suf...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a novel Reinforcement Learning based unsupervised method for the task of goal planning for agent based exploration in a non task-centric paradigm. # Strengths: - (1) Readability. Overall, the readability is fair. The main ideas are well conveyed, justified and articulated while mostly provi...
The paper presents a structural learning algorithm for concurrently learning the multi-task learning architecture and its parameters. The multitask learning is performed by the creation and removal of neurons based on local similarity. The proposed method is validated on the Cityscapes and NYUv2 datasets for five diffe...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper presents a structural learning algorithm for concurrently learning the multi-task learning architecture and its parameters. The multitask learning is performed by the creation and removal of neurons based on local similarity. The proposed method is validated on the Cityscapes and NYUv2 datasets for fi...
This paper proposes a new algorithm to solve the problem of recovering, as much as possible, the original graphical causal structure at the causal timescale from the derived graphical structure at a measurement timescale, where measurements are made every u number of time steps for an unknown u. The new algorithm is ba...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes a new algorithm to solve the problem of recovering, as much as possible, the original graphical causal structure at the causal timescale from the derived graphical structure at a measurement timescale, where measurements are made every u number of time steps for an unknown u. The new algorit...
In this paper, authors propose a new augmentation based on graph neural networks for graph contrastive learning. They decompose the graph neural networks as two parts: the diffusion part and the transformation part. Based on these, they propose weightPrune for the transformation and randMP for the diffusion part. They ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: In this paper, authors propose a new augmentation based on graph neural networks for graph contrastive learning. They decompose the graph neural networks as two parts: the diffusion part and the transformation part. Based on these, they propose weightPrune for the transformation and randMP for the diffusion par...
The paper proposes a novel algorithm to deal with Byzantine resilient decentralized optimization for data training. + The algorithm dose not rely on additional graph connectivity + The performance is better than some existing algorithms - The algorithm does not guarantee accuracy of the optimization process - The ide...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper proposes a novel algorithm to deal with Byzantine resilient decentralized optimization for data training. + The algorithm dose not rely on additional graph connectivity + The performance is better than some existing algorithms - The algorithm does not guarantee accuracy of the optimization process -...
This paper presents a graph-based method to solve point-cloud instance segmentation. The main idea is to train an attention-based deep network that predicts the probability of each pair of points belonging to the same instance. The first step of the proposed method builds a similarity matrix on the point cloud. Each en...	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 graph-based method to solve point-cloud instance segmentation. The main idea is to train an attention-based deep network that predicts the probability of each pair of points belonging to the same instance. The first step of the proposed method builds a similarity matrix on the point cloud....
The paper studies an interesting problem on how to automatically generate augmented data for learning invairance representations. The paper proposes a MCMC based methods, and then tested the results in some datasets. The paper seems to be on a novel track of goals, but needs further development. - strength - the ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper studies an interesting problem on how to automatically generate augmented data for learning invairance representations. The paper proposes a MCMC based methods, and then tested the results in some datasets. The paper seems to be on a novel track of goals, but needs further development. - strength ...
The paper provides a method for meta-learning a preconditioner for gradient descent. The authors motivate the method from the mirror descent perspective, and use a convergence rate bound as a training objective. The authors parameterize the preconditioner in the K-FAC format [1]. Empirically, the method achieves good r...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper provides a method for meta-learning a preconditioner for gradient descent. The authors motivate the method from the mirror descent perspective, and use a convergence rate bound as a training objective. The authors parameterize the preconditioner in the K-FAC format [1]. Empirically, the method achieve...
The paper presents a hierarchy of latent representations weighted by a series of monotonically increasing hyper-parameters, which compose an information bottleneck. The authors empirically demonstrate their idea on dSprites and Shapes3D to show the proposed method is able to learn disentangled representations while pre...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper presents a hierarchy of latent representations weighted by a series of monotonically increasing hyper-parameters, which compose an information bottleneck. The authors empirically demonstrate their idea on dSprites and Shapes3D to show the proposed method is able to learn disentangled representations w...
The authors proposed to use the total amount of noise (TAN) to determine the privacy budget in Renyi DP. They observe scaling laws with TAN for DP-SGD which can be used to reduce the computational cost for hyper-parameter tuning. By the hyper-parameter tuning, the authors greatly improve the state-of-the-art on neural ...	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 proposed to use the total amount of noise (TAN) to determine the privacy budget in Renyi DP. They observe scaling laws with TAN for DP-SGD which can be used to reduce the computational cost for hyper-parameter tuning. By the hyper-parameter tuning, the authors greatly improve the state-of-the-art on...
This paper introduces an offline imitation learning algorithm. The proposed algorithm consists of two parts: (1) to train a world model using demonstration of any quality and (2) to train a policy using $\lambda$-TD where the reward function is computed as Equation (7) (i.e. intrinsic reward). Finally, the authors empi...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper introduces an offline imitation learning algorithm. The proposed algorithm consists of two parts: (1) to train a world model using demonstration of any quality and (2) to train a policy using $\lambda$-TD where the reward function is computed as Equation (7) (i.e. intrinsic reward). Finally, the auth...
The paper proposes a new coded distributed gradient descent approach to mitigate the effect of stragglers in the distributed computation of gradients. The authors combine prior schemes on coded distributed gradient computation with repetition of unfinished tasks from straggling nodes across computation rounds with the ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper proposes a new coded distributed gradient descent approach to mitigate the effect of stragglers in the distributed computation of gradients. The authors combine prior schemes on coded distributed gradient computation with repetition of unfinished tasks from straggling nodes across computation rounds w...
This paper addresses the issue with representing (esp. heterophilic, power law) graphs by factorising the adjacency matrix into a combination of homophilic and heterophilic factors that are interpretable. Lower bounds based on logisitic principal components analysis (LPCA) is improved removing assumptions on the max de...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper addresses the issue with representing (esp. heterophilic, power law) graphs by factorising the adjacency matrix into a combination of homophilic and heterophilic factors that are interpretable. Lower bounds based on logisitic principal components analysis (LPCA) is improved removing assumptions on th...
The paper studies the effect of convolutional neural networks encoding position information due to padding. A new metric for measuring and visualizing the encoded positional information named Position-information Pattern from Padding (PPP) was proposed. The paper then discusses some of the disadvantages of the existing...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper studies the effect of convolutional neural networks encoding position information due to padding. A new metric for measuring and visualizing the encoded positional information named Position-information Pattern from Padding (PPP) was proposed. The paper then discusses some of the disadvantages of the ...
This work proposes ResGrad to improve the existing TTS models. Specifically, it uses a DFM to learn to generate the residual between an existing TTS model’s output and the groundtruth spectrogram. The residual is then added back to the TTS model’s output to get the refined output. Experiments on three datasets (LJSpeec...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work proposes ResGrad to improve the existing TTS models. Specifically, it uses a DFM to learn to generate the residual between an existing TTS model’s output and the groundtruth spectrogram. The residual is then added back to the TTS model’s output to get the refined output. Experiments on three datasets ...
* The authors propose a method for predicting the Remaining Useful Life (RUL) of components. They combine an LSTM-based autoencoder to extract features and a Bi-LSTM network to predict RUL. They validate the approach in one dataset and make an extensive test of hyperparameters and execution setups. In the end, they com...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: * The authors propose a method for predicting the Remaining Useful Life (RUL) of components. They combine an LSTM-based autoencoder to extract features and a Bi-LSTM network to predict RUL. They validate the approach in one dataset and make an extensive test of hyperparameters and execution setups. In the end, ...
The paper proposes to learn a domain-invariant representation based on the probabilistic framework by mapping each data point into probabilistic embeddings. For this, a probabilistic MMD and a probabilistic CSA loss are proposed. The paper proposes some new methodology and shows promising results on medical datasets. B...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes to learn a domain-invariant representation based on the probabilistic framework by mapping each data point into probabilistic embeddings. For this, a probabilistic MMD and a probabilistic CSA loss are proposed. The paper proposes some new methodology and shows promising results on medical dat...
This paper studies the VCRed regularization in self-supervised learning algorithms. The authors show that VCReg enforces pairwise independence between the features of the learned representation. Detailed analysis is provided. Strengths: 1: This paper studies the VCRed regularization in self-supervised learning algori...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies the VCRed regularization in self-supervised learning algorithms. The authors show that VCReg enforces pairwise independence between the features of the learned representation. Detailed analysis is provided. Strengths: 1: This paper studies the VCRed regularization in self-supervised learnin...
This paper proposes a scalable offline policy pre-training approach based on natural language instructions. It enables automatic augmentation of pre-training data with large language models to relabel and chain across trajectories. It empirically shows that the proposed approach outperforms prior pre-training approac...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a scalable offline policy pre-training approach based on natural language instructions. It enables automatic augmentation of pre-training data with large language models to relabel and chain across trajectories. It empirically shows that the proposed approach outperforms prior pre-training...
The paper makes the following improvements based on Vision Transformer (ViT). 1. The paper adds a Transformer Block to the existing ViT to predict coarse categories. 2. After obtaining the coarse category results, a simple self-attention mechanism is used to obtain the coarse category weights. 3. Finally, the coarse ca...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper makes the following improvements based on Vision Transformer (ViT). 1. The paper adds a Transformer Block to the existing ViT to predict coarse categories. 2. After obtaining the coarse category results, a simple self-attention mechanism is used to obtain the coarse category weights. 3. Finally, the c...
This paper proposes to introduce two additional parameters – density and size scale in the box embedding for fine-grained entity typing. The experiments on three benchmarks show the proposed approach achieved some marginal improvements over the considered baselines. Strength: The authors introduced density and size sc...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes to introduce two additional parameters – density and size scale in the box embedding for fine-grained entity typing. The experiments on three benchmarks show the proposed approach achieved some marginal improvements over the considered baselines. Strength: The authors introduced density and...
Cache-timing Attack (CTA) endangers privacy and security of data. However, traditional detection methods are not extendable. This work tackles CTA and detection from game-theoretic perspective by proposing a deep multi-agent reinforcement learning (MARL) approach with transformer for attackers and detectors. Strength: ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Cache-timing Attack (CTA) endangers privacy and security of data. However, traditional detection methods are not extendable. This work tackles CTA and detection from game-theoretic perspective by proposing a deep multi-agent reinforcement learning (MARL) approach with transformer for attackers and detectors. St...
================================================================= [Summary] This paper proposes an improved version of ReLIC, dubbed ReLICv2, by including views of varying sizes and saliency masking into the ReLIC's training loss, and presents benchmarking results using ResNet encoders in several experiments. ====...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: ================================================================= [Summary] This paper proposes an improved version of ReLIC, dubbed ReLICv2, by including views of varying sizes and saliency masking into the ReLIC's training loss, and presents benchmarking results using ResNet encoders in several experiment...
High level motivation: The authors are interested in understanding what it takes for machine learning systems to reliable generalize to novel circumstances, especially when those are out-of-distribution. Research question: Do insights from the theory of computation (specifically the Chomsky Hierarchy) inform ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: High level motivation: The authors are interested in understanding what it takes for machine learning systems to reliable generalize to novel circumstances, especially when those are out-of-distribution. Research question: Do insights from the theory of computation (specifically the Chomsky Hierarchy)...
An approach is proposed for improving performance when corruption types are known in advance, and when one cannot or is not willing to retrain the underlying model. The approach involves 1. training a small corruption-type classifier that operates on carefully-normalized, frequency-domain input images, 2. a table of ba...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: An approach is proposed for improving performance when corruption types are known in advance, and when one cannot or is not willing to retrain the underlying model. The approach involves 1. training a small corruption-type classifier that operates on carefully-normalized, frequency-domain input images, 2. a tab...
The authors present a multi-task learning framework called prefer-to-classify (P2C), which is based on pair-wise preference learning. Specifically, their work is focused on handling NLP tasks. The paper explains the overall approach, and the process of selecting informative pairs, and how the preferences are collected...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors present a multi-task learning framework called prefer-to-classify (P2C), which is based on pair-wise preference learning. Specifically, their work is focused on handling NLP tasks. The paper explains the overall approach, and the process of selecting informative pairs, and how the preferences are c...
This paper proposes a privacy-preserving class distribution estimation method for dealing with class-imbalanced FL scenarios. S1. The paper is easy to follow. S2. Some theoretical analysis is provided. W1. The assumption of theorem 1 is unrealistic (i.e., the input distribution of all classes is equivalent). While th...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes a privacy-preserving class distribution estimation method for dealing with class-imbalanced FL scenarios. S1. The paper is easy to follow. S2. Some theoretical analysis is provided. W1. The assumption of theorem 1 is unrealistic (i.e., the input distribution of all classes is equivalent). ...
This paper analyzes classes of online games for which classes of algorithms will always converge to a Nash equilibrium. The main result is: for m-strongly monotone games, any no-regret algorithm will converge to a Nash equilibrium, if m satisfies certain conditions. The paper also considers other types of algorithms: f...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper analyzes classes of online games for which classes of algorithms will always converge to a Nash equilibrium. The main result is: for m-strongly monotone games, any no-regret algorithm will converge to a Nash equilibrium, if m satisfies certain conditions. The paper also considers other types of algor...
This paper introduces a method for explicit task clustering as an inductive bias for Meta-RL algorithms operating on a task distribution. The key assumption is heterogeneity within that task distribution, i.e. that the i.i.d. assumption of draws from the task distribution is unrealistic in real Meta-RL setting and that...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper introduces a method for explicit task clustering as an inductive bias for Meta-RL algorithms operating on a task distribution. The key assumption is heterogeneity within that task distribution, i.e. that the i.i.d. assumption of draws from the task distribution is unrealistic in real Meta-RL setting ...
This paper tackles the problem of OoD detection in the case where the distribution of training classes is not-uniform. In other words, the authors tackle OoD detection in the case of long-tail image recognition. The authors first make the observation that a number of existing OoD methods (either implicitly or explicitl...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper tackles the problem of OoD detection in the case where the distribution of training classes is not-uniform. In other words, the authors tackle OoD detection in the case of long-tail image recognition. The authors first make the observation that a number of existing OoD methods (either implicitly or e...