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 a variant of the stochastic contextual bandits problem, where contextual information is not given but can be collected by asking a finite set of questions. The authors refer to this problem setting as survey bandit. The goal is to learn a policy that asks a series of questions that give the maximum...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies a variant of the stochastic contextual bandits problem, where contextual information is not given but can be collected by asking a finite set of questions. The authors refer to this problem setting as survey bandit. The goal is to learn a policy that asks a series of questions that give the...
The paper considers gradient descent algorithms for computing a parameter of the optimal policy according to a possibly unstable trajectory of a Markov decision process. It is shown that under technically strong assumptions such as continuity, convexity, and smoothness, the rates of convergence will be linear, assuming...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper considers gradient descent algorithms for computing a parameter of the optimal policy according to a possibly unstable trajectory of a Markov decision process. It is shown that under technically strong assumptions such as continuity, convexity, and smoothness, the rates of convergence will be linear, ...
This work focuses on the misspecification setting for both linear contextual bandits and linear MDPs. For the linear bandits, it proposes the DS-OFUL algorithm. It derives an upper bound accordingly and also a lower bound. For the linear MDPs, it devises the DS-LSVI with a similar design and also provides un upper bou...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work focuses on the misspecification setting for both linear contextual bandits and linear MDPs. For the linear bandits, it proposes the DS-OFUL algorithm. It derives an upper bound accordingly and also a lower bound. For the linear MDPs, it devises the DS-LSVI with a similar design and also provides un u...
This paper proposed a self-supervised framework for category-level 3D object pose estimation. A novel category surface embedding module is proposed which can help establishing dense correspondences within a single instance, between two different instances of the same category, and instances in different time frames. Ex...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposed a self-supervised framework for category-level 3D object pose estimation. A novel category surface embedding module is proposed which can help establishing dense correspondences within a single instance, between two different instances of the same category, and instances in different time fr...
Inducing points have traditionally been heavily relied on to alleviate the cubic computational complexity associated with Gaussian process training and inference. There have been various different ways in which inducing points have been used in the literature, but they have nearly always been limited to either the inpu...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: Inducing points have traditionally been heavily relied on to alleviate the cubic computational complexity associated with Gaussian process training and inference. There have been various different ways in which inducing points have been used in the literature, but they have nearly always been limited to either ...
This work focuses on the robust reinforcement learning paradigm, whereby an agent is trained to be robust to worst case variations in an environment. The paper proposes an approach called FARR, whereby the adversary (proposing variations) is penalized if the environment proposed is not solvable. The paper sets out with...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work focuses on the robust reinforcement learning paradigm, whereby an agent is trained to be robust to worst case variations in an environment. The paper proposes an approach called FARR, whereby the adversary (proposing variations) is penalized if the environment proposed is not solvable. The paper sets ...
This paper studies mean field games (MFG) using a notion of correlated equilibrium and proposes an inverse reinforcement learning (IRL) method. The authors mean field games where a representative agent solves a finite horizon MDP, and globally, we look for a Nash equilibrium. They introduce a notion of correlated equil...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies mean field games (MFG) using a notion of correlated equilibrium and proposes an inverse reinforcement learning (IRL) method. The authors mean field games where a representative agent solves a finite horizon MDP, and globally, we look for a Nash equilibrium. They introduce a notion of correlat...
- The paper proposes a zero-shot Knowledge Distillation ("ZSKD") approach (i.e., does not require any samples from teacher's training data). - Similar to DFME (Truong et al.) and MAZE (Karyiappa et al.), the approach consists of two phases per epoch: (a) training a generator to produce synthetic labels; and (b) trainin...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: - The paper proposes a zero-shot Knowledge Distillation ("ZSKD") approach (i.e., does not require any samples from teacher's training data). - Similar to DFME (Truong et al.) and MAZE (Karyiappa et al.), the approach consists of two phases per epoch: (a) training a generator to produce synthetic labels; and (b)...
This work centers around the relationship between sharpness (scaling invariance) and generalization in DNNs under the PAC-Bayesian framework. This work decomposes the scale and connectivity parameters, derive the related connectivity tangent kernel (CTK), and builds the PAC-Bayesian generalization bounds. The trace of ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work centers around the relationship between sharpness (scaling invariance) and generalization in DNNs under the PAC-Bayesian framework. This work decomposes the scale and connectivity parameters, derive the related connectivity tangent kernel (CTK), and builds the PAC-Bayesian generalization bounds. The t...
This paper proposes an early stopping strategy for deep image prior. This is achieved by using an efficient ES strategy that consistently detects near-peak performance across several CI tasks and DIP variants. The paper provides high level intuition, theoretical proof and multiple use cases to demonstrate the proposed...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes an early stopping strategy for deep image prior. This is achieved by using an efficient ES strategy that consistently detects near-peak performance across several CI tasks and DIP variants. The paper provides high level intuition, theoretical proof and multiple use cases to demonstrate the ...
This paper proposes to use random sampling for samples in clustering classes to avoid inter-class conflict and a random dropout mechanism for features to generate compact features. The method is easy to understand and implement. The experiments about using VL models for zero-shot/transfer learning tasks are exhaustive....	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to use random sampling for samples in clustering classes to avoid inter-class conflict and a random dropout mechanism for features to generate compact features. The method is easy to understand and implement. The experiments about using VL models for zero-shot/transfer learning tasks are exh...
The paper studies the effect of self-supervised pre-training for Vision Transformer based RL agents. It shows the effect of temporal order verification and VICReg on 10 Atari games. The proposed methods improves final return on Atari games. [Strengths] - The paper is easy to follow. - The proposed method makes intuitiv...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies the effect of self-supervised pre-training for Vision Transformer based RL agents. It shows the effect of temporal order verification and VICReg on 10 Atari games. The proposed methods improves final return on Atari games. [Strengths] - The paper is easy to follow. - The proposed method makes ...
This paper proposes a novel, adaptive strategy to drop classes from the dataset during training based on the f1 score of the model. They show that their approach, Progressive Data Dropout (PDD), is able to achieve similar test/val accuracies while using significantly lower data and time to train on image classification...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes a novel, adaptive strategy to drop classes from the dataset during training based on the f1 score of the model. They show that their approach, Progressive Data Dropout (PDD), is able to achieve similar test/val accuracies while using significantly lower data and time to train on image classi...
The paper is well motivated, citing the rather broad literature on learning logical and probabilistic rules for KG completion, and proposing a method that can learn non-chain-like rules. The method makes sense. They show good results on several KG completion datasets, comparing against relevant baselines. Love the exam...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper is well motivated, citing the rather broad literature on learning logical and probabilistic rules for KG completion, and proposing a method that can learn non-chain-like rules. The method makes sense. They show good results on several KG completion datasets, comparing against relevant baselines. Love ...
The paper is fundamentally a PDE analysis result involving the regularity of solutions to (nonlinear) variational problems with respect to the Barron norm when the function involved in the variational problem (to within an epsilon approximation in the H^1 seminorm) satisfies a Holder-like condition (with respect to the...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper is fundamentally a PDE analysis result involving the regularity of solutions to (nonlinear) variational problems with respect to the Barron norm when the function involved in the variational problem (to within an epsilon approximation in the H^1 seminorm) satisfies a Holder-like condition (with respec...
The authors prove a new identifiability result that provides conditions under which maximally sparse base-predictors yield disentangled representations. Based on this theoretical result, they propose a practical approach to learn disentangled representations based on a sparsity-promoting bi-level optimization problem. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors prove a new identifiability result that provides conditions under which maximally sparse base-predictors yield disentangled representations. Based on this theoretical result, they propose a practical approach to learn disentangled representations based on a sparsity-promoting bi-level optimization p...
This paper proposes a new masked image modeling method(BEIT v2). A new Vector-Quantized Knowledge Distillation helps the BEIT v2 explore the high-level semantics. Meanwhile, this paper introduces a patch aggregation strategy to enhance global semantic representation. Experiments on image classification and semantic seg...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a new masked image modeling method(BEIT v2). A new Vector-Quantized Knowledge Distillation helps the BEIT v2 explore the high-level semantics. Meanwhile, this paper introduces a patch aggregation strategy to enhance global semantic representation. Experiments on image classification and sema...
This paper focuses on merging individual models on the parameter space without access to the respective training data. The individual models could come from different domains or even different tasks, their knowledge could be fused by merging without any training, and the resulting merging model could perform relatively...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper focuses on merging individual models on the parameter space without access to the respective training data. The individual models could come from different domains or even different tasks, their knowledge could be fused by merging without any training, and the resulting merging model could perform re...
This paper proposes an approach to uncertainty estimation for image-to-image tasks based on continuous masking. The approach can work with arbitrary networks, and distance metrics between images and is guaranteed to yield distances within a small threshold with high probability. Experiments are conducted on image color...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes an approach to uncertainty estimation for image-to-image tasks based on continuous masking. The approach can work with arbitrary networks, and distance metrics between images and is guaranteed to yield distances within a small threshold with high probability. Experiments are conducted on ima...
This work discovers a concrete example on the WILDS-Camelyon17 dataset that ID performance is negatively correlated to OOD performance. A theoretical example based on a simple linear model is also presented to show the trade-off between ID and OOD performance. The authors explain why past studies missed such a negative...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work discovers a concrete example on the WILDS-Camelyon17 dataset that ID performance is negatively correlated to OOD performance. A theoretical example based on a simple linear model is also presented to show the trade-off between ID and OOD performance. The authors explain why past studies missed such a ...
The authors study phase transitions in linear regression with permuted labels for both the oracle setting (in which the regression coefficients are assumed to be known) and the non oracle setting (in which both the permutation and the regression vector are considered as unknown). The main result of the paper is an esti...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors study phase transitions in linear regression with permuted labels for both the oracle setting (in which the regression coefficients are assumed to be known) and the non oracle setting (in which both the permutation and the regression vector are considered as unknown). The main result of the paper is...
The paper investigates the case when deterministic policies are learned in a GAN framework. The authors describe occuring instabilities and attribute them to exploding gradients. # Strenghts * The authors try to thorougly understand existing algorithms and investigate their weaknesses, which is an interesting and val...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper investigates the case when deterministic policies are learned in a GAN framework. The authors describe occuring instabilities and attribute them to exploding gradients. # Strenghts * The authors try to thorougly understand existing algorithms and investigate their weaknesses, which is an interesting...
This work fits the single-hidden-layer neural network to data generated by a teacher network with Gaussian parameters. Conclusion is drawn from experimental results that SGD with automated width selection has both sample and query complexity scaling linearly with the input dimension and width, which complies with the i...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work fits the single-hidden-layer neural network to data generated by a teacher network with Gaussian parameters. Conclusion is drawn from experimental results that SGD with automated width selection has both sample and query complexity scaling linearly with the input dimension and width, which complies wi...
This paper considers non-linear causal discovery under the same setting as Rolland et al., (2022). The authors proposed to replace the scoring function in the SCORE method proposed by Rolland et al., (2022) with diffusion probabilistic models (DPMs). This allows them to update the learned Hessian without re-training th...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper considers non-linear causal discovery under the same setting as Rolland et al., (2022). The authors proposed to replace the scoring function in the SCORE method proposed by Rolland et al., (2022) with diffusion probabilistic models (DPMs). This allows them to update the learned Hessian without re-tra...
This paper proposes an hardware-aware NAS on a superset that preserves the Pareto ranking (accuracy and latency). It demonstrates better accuracy-latency tradeoff when compared to SOTA approaches. Strength * Good trade-off between latency and accuracy is achieved using the proposed approach comparing to the SOTA. Weak...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes an hardware-aware NAS on a superset that preserves the Pareto ranking (accuracy and latency). It demonstrates better accuracy-latency tradeoff when compared to SOTA approaches. Strength * Good trade-off between latency and accuracy is achieved using the proposed approach comparing to the SOT...
The paper proposes a learning framework that aims to address the change in the number of agents in an environment as a student agent is trained through a curriculum of tasks as proposed by a separate teacher agent. The teacher agent is modeled as a contextual bandit that aims to resolve the inherent non-stationarity in...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a learning framework that aims to address the change in the number of agents in an environment as a student agent is trained through a curriculum of tasks as proposed by a separate teacher agent. The teacher agent is modeled as a contextual bandit that aims to resolve the inherent non-station...
This paper proposes a new asynchronous decentralized training algorithm with local updates, named SWIFT. Compared to previous works in this direction, the authors removed the bounded delay assumption in analysis, and hence, obtained better theoretical results. In order to ensure the convergence, a new technique is intr...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a new asynchronous decentralized training algorithm with local updates, named SWIFT. Compared to previous works in this direction, the authors removed the bounded delay assumption in analysis, and hence, obtained better theoretical results. In order to ensure the convergence, a new technique...
This paper proposes a progressive distillation strategy for object detection and instance segmentation models, i.e., multiple teacher models transfer knowledge to a student model in a sequential manner. A heuristic method is also given to choose the order of teacher models and its efficacy is demonstrated by experiment...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a progressive distillation strategy for object detection and instance segmentation models, i.e., multiple teacher models transfer knowledge to a student model in a sequential manner. A heuristic method is also given to choose the order of teacher models and its efficacy is demonstrated by ex...
The paper introduces a method for offline reinforcement learning that involves fitting a more expressive density model and an episodic planning technique. The method demonstrates decent results on D4RL, a standard benchmark for offline reinforcement learning. # Strength * The approach is simple and produces strong res...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces a method for offline reinforcement learning that involves fitting a more expressive density model and an episodic planning technique. The method demonstrates decent results on D4RL, a standard benchmark for offline reinforcement learning. # Strength * The approach is simple and produces st...
This work proposes zeroth-order perturbed gradient descent algorithm which is the first zeroth order algorithm to the authors' knowledge that finds second-order stationary point with only $2m$ function evaluations per iteration for any $m\in [1,d]$ ($d$ is the dimensionality), in contrast to $\Omega(d)$ per iteration i...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This work proposes zeroth-order perturbed gradient descent algorithm which is the first zeroth order algorithm to the authors' knowledge that finds second-order stationary point with only $2m$ function evaluations per iteration for any $m\in [1,d]$ ($d$ is the dimensionality), in contrast to $\Omega(d)$ per ite...
*Paper Summary* The authors suggest a structural representation model for the task of phylogenetic inference by the utilization of graph neural networks architectures. For the initial node attributes, they propose the construction of raw features based on the Dirichlet energy minimization. Using message passing ste...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: *Paper Summary* The authors suggest a structural representation model for the task of phylogenetic inference by the utilization of graph neural networks architectures. For the initial node attributes, they propose the construction of raw features based on the Dirichlet energy minimization. Using message pas...
The paper considers studies dynamics of mini-batch SGD using Kolmogorov complexity. They define a notion of accuracy discrepancy as a KL-divergence between accuracy of the model on previous batches in the current epoch and on the last batch. Using the fact that the random strings used for generating the epoch’s permuta...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper considers studies dynamics of mini-batch SGD using Kolmogorov complexity. They define a notion of accuracy discrepancy as a KL-divergence between accuracy of the model on previous batches in the current epoch and on the last batch. Using the fact that the random strings used for generating the epoch’s...
The paper proposes a 1. A new learning setup called Cascade in which the agent is provided a semantic goal and a failed trajectory and the goal of the agent is to predict an intervention for the initial state of a pivot object so that the semantic goal is achieved counterfactually. 2. A new dynamics model based on sem...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper proposes a 1. A new learning setup called Cascade in which the agent is provided a semantic goal and a failed trajectory and the goal of the agent is to predict an intervention for the initial state of a pivot object so that the semantic goal is achieved counterfactually. 2. A new dynamics model base...
This paper studies the problem of multi-objective online learning. In the classic online convex optimization problem, there is a single objective function, and the goal is to find the best action that leads to the best value of the objective function in a sequential decision-making setting. This paper extends the probl...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies the problem of multi-objective online learning. In the classic online convex optimization problem, there is a single objective function, and the goal is to find the best action that leads to the best value of the objective function in a sequential decision-making setting. This paper extends t...
The paper tackle the problem of efficient transfer by leveraging both hierarchy and KL-regularisation. The authors claim that fine-tuning, hierarchical methods and imitation based approaches can all fail and therefore propose a method that combines elements from all of them. The authors perform experiments on robotic o...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper tackle the problem of efficient transfer by leveraging both hierarchy and KL-regularisation. The authors claim that fine-tuning, hierarchical methods and imitation based approaches can all fail and therefore propose a method that combines elements from all of them. The authors perform experiments on r...
This paper studies the capability limits of data poisoning, i.e., the minimum amount of data required to successfully inject backdoor behavior into DNN. In particular, the authors craft a universal adversarial patch as the trigger, use FUS to select samples, and exploit individual consistency to stabilize the performan...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper studies the capability limits of data poisoning, i.e., the minimum amount of data required to successfully inject backdoor behavior into DNN. In particular, the authors craft a universal adversarial patch as the trigger, use FUS to select samples, and exploit individual consistency to stabilize the p...
The paper aims at modeling adversarial attacks which can perturb the input graph sequence on discrete-time dynamic graph models meanwhile preserving the temporal dynamics of the graph.Then they propose a constraint named Temporal Dynamics-Aware Perturbation (TDAP) to make imperceptible perturbations in discrete-time. T...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper aims at modeling adversarial attacks which can perturb the input graph sequence on discrete-time dynamic graph models meanwhile preserving the temporal dynamics of the graph.Then they propose a constraint named Temporal Dynamics-Aware Perturbation (TDAP) to make imperceptible perturbations in discrete...
This paper introduces an Evolving Graph Contrastive Memory (EGCM) framework which effectively models the multiple behaviors in the recommendation. EGCM is comprised of (1) a multi-behavior graph encoder to model the behavior-aware short-term interests of users, (2) a dynamic cross-relational memory network to model cro...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces an Evolving Graph Contrastive Memory (EGCM) framework which effectively models the multiple behaviors in the recommendation. EGCM is comprised of (1) a multi-behavior graph encoder to model the behavior-aware short-term interests of users, (2) a dynamic cross-relational memory network to m...
The paper discussed a few similarity metrics and conducted layer-wise comparisons of VAEs using the similarity metric with different datasets, models and parameters. And conclusions are drawn on the model's disentanglement and posterior collapse based on the results. The authors found that encoder representations are l...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: The paper discussed a few similarity metrics and conducted layer-wise comparisons of VAEs using the similarity metric with different datasets, models and parameters. And conclusions are drawn on the model's disentanglement and posterior collapse based on the results. The authors found that encoder representatio...
The paper proposes a new variant performing supervised contrastive learning with continuous labels. Improvements on four image-based and one EEG regression tasks are demonstrated when using the learned feature space instead of training various baseline algorithms from scratch. The paper also shows improvements in terms...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper proposes a new variant performing supervised contrastive learning with continuous labels. Improvements on four image-based and one EEG regression tasks are demonstrated when using the learned feature space instead of training various baseline algorithms from scratch. The paper also shows improvements ...