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 work addresses the task of multi-modal DG where each modality has to cope with its own domain shift. Specifically, the specific scenario of query-based video segmentation is studied to better advance the generalization ability of the model in the multi-modal situation. The authors observe that actions belonging to...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work addresses the task of multi-modal DG where each modality has to cope with its own domain shift. Specifically, the specific scenario of query-based video segmentation is studied to better advance the generalization ability of the model in the multi-modal situation. The authors observe that actions belo...
The authors address the task of 3D human mesh registration with key focus on registering a complete template to (incomplete) single view point cloud. Authors show better performance than baselines on the registration as well as correspondence prediction task. Strengths: + The task of registering partial/ single view da...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors address the task of 3D human mesh registration with key focus on registering a complete template to (incomplete) single view point cloud. Authors show better performance than baselines on the registration as well as correspondence prediction task. Strengths: + The task of registering partial/ single...
As a standard benchmark for POMDP tasks is desired due to rising research interests, the paper proposed 3 new Gym tasks which are memory-dependent POMDPs. The key features of the environments are (1) "strongly depend on memory". (2) different levels of difficulty. (3) good meta properties (Table 1). The authors cond...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: As a standard benchmark for POMDP tasks is desired due to rising research interests, the paper proposed 3 new Gym tasks which are memory-dependent POMDPs. The key features of the environments are (1) "strongly depend on memory". (2) different levels of difficulty. (3) good meta properties (Table 1). The auth...
The paper studies an offline RL algorithm by utilizing the special policy structure of the behavior policy (which generates the offline data). The paper proposes to define the policy being learned by using a mixture of sub-policies, which is in the form of pi(s, a) = sum_z p(z\|s) pi_sub (a \| s, z), where z is the intro...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies an offline RL algorithm by utilizing the special policy structure of the behavior policy (which generates the offline data). The paper proposes to define the policy being learned by using a mixture of sub-policies, which is in the form of pi(s, a) = sum_z p(z\|s) pi_sub (a \| s, z), where z is t...
This paper studies the adversarial robustness of federated learning, which is an important problem when training the model on the client side. Specially, the authors consider the incompatible dilemma between federated learning and adversarial training, and introduce a mediating slack mechanism to better bridge two para...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper studies the adversarial robustness of federated learning, which is an important problem when training the model on the client side. Specially, the authors consider the incompatible dilemma between federated learning and adversarial training, and introduce a mediating slack mechanism to better bridge ...
The authors propose to use transformers for electric impedance tomography, with ideas that may apply to a broade range of boundary value inverse problems. EIT is known to be very ill-posed (only log stable) so it is a challenging test for any method. The authors frame the problem as inverting the samples of the NtD map...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors propose to use transformers for electric impedance tomography, with ideas that may apply to a broade range of boundary value inverse problems. EIT is known to be very ill-posed (only log stable) so it is a challenging test for any method. The authors frame the problem as inverting the samples of the...
To the authors' knowledge, this work proposes the first algorithm that provably converges to a local min-max equilibrium for smooth nonconvex-nonconcave minimax optimization. Pros: This work is highly novel in both algorithm and proof technique, as elaborated in Novelty below. The convergence is also well supporte...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: To the authors' knowledge, this work proposes the first algorithm that provably converges to a local min-max equilibrium for smooth nonconvex-nonconcave minimax optimization. Pros: This work is highly novel in both algorithm and proof technique, as elaborated in Novelty below. The convergence is also well ...
This work aims to propose a differentiable and scalable k-subset sampling algorithm based on conditional Poisson sampling. The scalability comes from the conditional Poisson sampling scheme where each instance is sampled independently such that the vectorized complexity of the proposed algorithm is independent of the s...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work aims to propose a differentiable and scalable k-subset sampling algorithm based on conditional Poisson sampling. The scalability comes from the conditional Poisson sampling scheme where each instance is sampled independently such that the vectorized complexity of the proposed algorithm is independent ...
This paper studies text-supervised semantic segmentation. The authors propose multi-View Consistent learning (ViewCo) to introduce the correspondence among multiple augmented views of the same image. The experimental results on serval datasets demonstrate the effectiveness of the proposed method. Strength + Compared to...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This paper studies text-supervised semantic segmentation. The authors propose multi-View Consistent learning (ViewCo) to introduce the correspondence among multiple augmented views of the same image. The experimental results on serval datasets demonstrate the effectiveness of the proposed method. Strength + Com...
The paper provides a theoretical generalization analysis to show that jointly applying sparsification methods on both the graph edges (network topology) and neurons (model) of a graph neural network (GNN) makes training more efficient in terms of sample complexity (required number of known labels) and convergence rate...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper provides a theoretical generalization analysis to show that jointly applying sparsification methods on both the graph edges (network topology) and neurons (model) of a graph neural network (GNN) makes training more efficient in terms of sample complexity (required number of known labels) and converge...
If I understand this paper correctly, the authors introduce a way how to optimize for integrating “feint” actions into the solutions of stochastic-game-like interactions to promote diversity. They do so by gradually building up the criterion function from three aspects they consider the most important, and refer to the...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: If I understand this paper correctly, the authors introduce a way how to optimize for integrating “feint” actions into the solutions of stochastic-game-like interactions to promote diversity. They do so by gradually building up the criterion function from three aspects they consider the most important, and refe...
The paper introduces a new multi-modal VAE model which includes modality-specific latent variables, as well as shared latent variables with the latter encoded through a mixture-of-experts model. The new variational bound allows to utilize private latent variables that do not hinder cross-modal coherence and generation...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The paper introduces a new multi-modal VAE model which includes modality-specific latent variables, as well as shared latent variables with the latter encoded through a mixture-of-experts model. The new variational bound allows to utilize private latent variables that do not hinder cross-modal coherence and ge...
The paper proposes a first of its kind in-processing learning framework FairGBM for training GBDT, without affecting its performance, constrained by fairness. The paper tries to advanced the area of FairML, limiting risks of unfair or biased ML systems and aims to establish a gold standard method. The authors provide e...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes a first of its kind in-processing learning framework FairGBM for training GBDT, without affecting its performance, constrained by fairness. The paper tries to advanced the area of FairML, limiting risks of unfair or biased ML systems and aims to establish a gold standard method. The authors p...
This submission proposed to conduct post-training quantization by learning element-wise scaling parameters $\boldsymbol{S}$. Pre-quantized parameters are divided by the scaling parameters and then rounded to integer. Basically, the proposed method is a subtle re-training for pre-trained parameters for quantization. Wea...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This submission proposed to conduct post-training quantization by learning element-wise scaling parameters $\boldsymbol{S}$. Pre-quantized parameters are divided by the scaling parameters and then rounded to integer. Basically, the proposed method is a subtle re-training for pre-trained parameters for quantizat...
The paper introduces a self-supervised visual pretraining technique called CIM: Corrupted Image Modeling. The main idea is to randomly select patches and replace them with plausible alternatives, instead of using a MASK token. An enhancer network tries to solve pretext tasks on these augmented images (generating all or...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper introduces a self-supervised visual pretraining technique called CIM: Corrupted Image Modeling. The main idea is to randomly select patches and replace them with plausible alternatives, instead of using a MASK token. An enhancer network tries to solve pretext tasks on these augmented images (generatin...
This paper proposes a method to interpret out-of-distribution detection using learnt high-level concepts. Strengths - To my knowledge, this is the first work to study the notion of concept-based explanations for OOD detection. - The idea of LDA on the concept layer is clever and simple. - The proposed metrics are easy ...	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 proposes a method to interpret out-of-distribution detection using learnt high-level concepts. Strengths - To my knowledge, this is the first work to study the notion of concept-based explanations for OOD detection. - The idea of LDA on the concept layer is clever and simple. - The proposed metrics a...
This paper proposes to improve the discrete sequence model for long-term video prediction. The proposed method operates on (patch-wise) quantized representation of videos extracted by VQ-GAN, and models the temporal dynamics through Transformers. To handle the quadratic complexity of Transformers thus enabling training...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposes to improve the discrete sequence model for long-term video prediction. The proposed method operates on (patch-wise) quantized representation of videos extracted by VQ-GAN, and models the temporal dynamics through Transformers. To handle the quadratic complexity of Transformers thus enabling ...
This paper provides some regret bounds for online matrix completion: at each round, a recommendation is made for each of the M users and the reward is observed. The reward distribution is unchanged through time (this is therefore a contextual bandit scenario). In the simplest approach (Sections 3, 4, B and C), the a...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper provides some regret bounds for online matrix completion: at each round, a recommendation is made for each of the M users and the reward is observed. The reward distribution is unchanged through time (this is therefore a contextual bandit scenario). In the simplest approach (Sections 3, 4, B and C...
This paper studies the problem of hypothesis testing in community detection: instead of asking to recover the exact community assignments, the goal is to perform hypothesis testing on them (e.g. asking whether two given nodes belong to the same community or not). The authors introduce a framework for measuring distanc...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the problem of hypothesis testing in community detection: instead of asking to recover the exact community assignments, the goal is to perform hypothesis testing on them (e.g. asking whether two given nodes belong to the same community or not). The authors introduce a framework for measuring...
This paper propose to investigate meta-continual learning scenario where the task boundaries are not known and whether distribution shift occur or not. To solve this challenging problem, they focus on some empirical findings that the loss values or free energies sharply increases when the boundary meets. Thus they intr...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper propose to investigate meta-continual learning scenario where the task boundaries are not known and whether distribution shift occur or not. To solve this challenging problem, they focus on some empirical findings that the loss values or free energies sharply increases when the boundary meets. Thus t...
This paper proposes and analyzes stochastic algorithms that apply to solve a specific distributional robust optimization problem. This problem, Eq. (1), is not a general formulation solved by the existing methods. The uncertainty set of Eq. (1) must be a KL-ball, and the KL constraint in Eq. (1) seems to be too deliber...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes and analyzes stochastic algorithms that apply to solve a specific distributional robust optimization problem. This problem, Eq. (1), is not a general formulation solved by the existing methods. The uncertainty set of Eq. (1) must be a KL-ball, and the KL constraint in Eq. (1) seems to be too...
This paper focuses on the problem of goal conditioned reinforcement learning (GCRL), where the authors propose a new way of sampling goals during the data collection phase. Specifically, the authors propose to sample goals with probability proportional to two factors: the prior distribution of goals and the intermediat...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper focuses on the problem of goal conditioned reinforcement learning (GCRL), where the authors propose a new way of sampling goals during the data collection phase. Specifically, the authors propose to sample goals with probability proportional to two factors: the prior distribution of goals and the int...
In this paper, the authors address the problem of a multi-attribute shift in Domain Generalization. They extensively and theoretically characterize various realizations of canonical causal graph modeling distribution shifts. They theoretically demonstrate that every such shift would entail a different independence cons...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this paper, the authors address the problem of a multi-attribute shift in Domain Generalization. They extensively and theoretically characterize various realizations of canonical causal graph modeling distribution shifts. They theoretically demonstrate that every such shift would entail a different independe...
This paper focuses on the issues in the neural network-based policy in meta reinforcement learning (RL), such as overfitting \& poor generalization ability, difficult/inefficient to deploy with limited computational resources, and poor interpretability. To address those issues, the framework of Contextual Symbolic Poli...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper focuses on the issues in the neural network-based policy in meta reinforcement learning (RL), such as overfitting \& poor generalization ability, difficult/inefficient to deploy with limited computational resources, and poor interpretability. To address those issues, the framework of Contextual Symbo...
The authors present a neural network based approach for managing the so called tradeoff between interpretability and accuracy in timeseries data by learning a dictionary of discrete representations. They guarantee that 1) only a small number of patterns that can be visualised easily are learnt 2) training a linear clas...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors present a neural network based approach for managing the so called tradeoff between interpretability and accuracy in timeseries data by learning a dictionary of discrete representations. They guarantee that 1) only a small number of patterns that can be visualised easily are learnt 2) training a lin...
In this work, the authors study the impact of different deep learning networks on the fairness of face recognition algorithms. A Neural Architecture Search is performed, and a wide variety of network architectures (with multiple sets of hyperparameters) are employed for the analysis. The Rank Disparity metric is used f...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this work, the authors study the impact of different deep learning networks on the fairness of face recognition algorithms. A Neural Architecture Search is performed, and a wide variety of network architectures (with multiple sets of hyperparameters) are employed for the analysis. The Rank Disparity metric i...
This paper showcases the use of LLM on solving classical planning problems. A dataset based on PDDL syntax which includes 4 tasks is generated for training and evaluation of the model. The proposed model, Plansformer, is fine-tuned on the generated dataset from pre-trained CodeT5 model. The model is evaluated on each p...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper showcases the use of LLM on solving classical planning problems. A dataset based on PDDL syntax which includes 4 tasks is generated for training and evaluation of the model. The proposed model, Plansformer, is fine-tuned on the generated dataset from pre-trained CodeT5 model. The model is evaluated o...
In this paper, the authors investigate the problem of temporal causal discovery. Existing work relies on the absence of instantaneous effects with fixed noise distributions, whereas constraint-based methods require stronger faithfulness assumptions. To alleviate this problem, the authors propose Rhino, which combines v...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: In this paper, the authors investigate the problem of temporal causal discovery. Existing work relies on the absence of instantaneous effects with fixed noise distributions, whereas constraint-based methods require stronger faithfulness assumptions. To alleviate this problem, the authors propose Rhino, which co...
The authors want to explore a very interesting question why the prompting learning of large pretrained language models leads to strong performance in a variety of downstream tasks, especially in zero-shot setups? This paper proposed a novel method to identify the evidence of the model’s task-specific competence in prom...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors want to explore a very interesting question why the prompting learning of large pretrained language models leads to strong performance in a variety of downstream tasks, especially in zero-shot setups? This paper proposed a novel method to identify the evidence of the model’s task-specific competence...
This paper studies contrastive self-supervised learning from the view of SNE. The authors show that InfoNCE loss can be seen as a special case of SNE, by setting the data similarity matrix P according to the data augmentation and embedding matrix Q as the softmax of pairwise similarity. Under the SNE framework, new ins...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This paper studies contrastive self-supervised learning from the view of SNE. The authors show that InfoNCE loss can be seen as a special case of SNE, by setting the data similarity matrix P according to the data augmentation and embedding matrix Q as the softmax of pairwise similarity. Under the SNE framework,...
The paper studies the data-driven low-rank approximation problem on data streams, which has gained attention recently. For a single matrix, some fast algorithms for low-rank approximation are based on sketching, and specifically they choose a sparse sketching matrix with random signs, which is multiplied by the origina...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper studies the data-driven low-rank approximation problem on data streams, which has gained attention recently. For a single matrix, some fast algorithms for low-rank approximation are based on sketching, and specifically they choose a sparse sketching matrix with random signs, which is multiplied by the...
In this paper, the authors analyze the three different message-passing schemes of Graph Neural Networks conditional on the geometric property, namely weak, strong, and pure methods, and empirically study their effect. In this paper, the authors analyze the three message-passing schemes of Graph Neural Networks conditio...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this paper, the authors analyze the three different message-passing schemes of Graph Neural Networks conditional on the geometric property, namely weak, strong, and pure methods, and empirically study their effect. In this paper, the authors analyze the three message-passing schemes of Graph Neural Networks ...
The paper proposes a fragment-based contrastive learning algorithm for learning the discriminative representations of the molecules. They propose to decompose a molecule into two fragments by breaking a single bound (none ring). The bound is chosen such that two fragments have a similar number of atoms. Then the compl...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper proposes a fragment-based contrastive learning algorithm for learning the discriminative representations of the molecules. They propose to decompose a molecule into two fragments by breaking a single bound (none ring). The bound is chosen such that two fragments have a similar number of atoms. Then t...
This paper studies the targeted attack on time-series forecasting tasks. Specifically, this paper proposes three different targeted attack goals: directional attack, amplitudinal attack, and temporal attack. This paper then designs the corresponding FGSM, PGD, and APGD attacks for these goals. All attacks are evaluated...	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 targeted attack on time-series forecasting tasks. Specifically, this paper proposes three different targeted attack goals: directional attack, amplitudinal attack, and temporal attack. This paper then designs the corresponding FGSM, PGD, and APGD attacks for these goals. All attacks are e...
The authors propose a method to model correlations between neurons in large populations which is based on sum-product networks (SPNs). They show that it outperforms two baselines (pairwise maximum entropy models and restricted Boltzmann machines) on a synthetic dataset and the Allen Brain Observatory neuropixels datase...	Recommendation: 8: accept, good paper	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The authors propose a method to model correlations between neurons in large populations which is based on sum-product networks (SPNs). They show that it outperforms two baselines (pairwise maximum entropy models and restricted Boltzmann machines) on a synthetic dataset and the Allen Brain Observatory neuropixel...
This paper deals with safe exploration in reinforcement learning in which an agent is required to ensure safety during training. The authors present a neuro-symbolic approach called SPICE. based on symbolic weakest preconditions. Empirically, they evaluate their approach on toy benchmarks, and show that it is able to a...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper deals with safe exploration in reinforcement learning in which an agent is required to ensure safety during training. The authors present a neuro-symbolic approach called SPICE. based on symbolic weakest preconditions. Empirically, they evaluate their approach on toy benchmarks, and show that it is a...
The paper presents a neuron-based method to defense against backdoor attacks. They first identify less important neurons in the network and then purify them using a fine-tuning based method. The paper contributes to both the selection of less important neurons and in purifying the network. Compared to the common method...	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 presents a neuron-based method to defense against backdoor attacks. They first identify less important neurons in the network and then purify them using a fine-tuning based method. The paper contributes to both the selection of less important neurons and in purifying the network. Compared to the commo...
This work focused on kernelized contextual bandits with distributed computation and asynchronous communication. The author proposed the Async-KernelUCB algorithm with $\tilde O(\sqrt(T))$ regret and only $O(N^2)$ communication complexity. In addition, both the Synthetic and real-world experiments show that Async-Kernel...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work focused on kernelized contextual bandits with distributed computation and asynchronous communication. The author proposed the Async-KernelUCB algorithm with $\tilde O(\sqrt(T))$ regret and only $O(N^2)$ communication complexity. In addition, both the Synthetic and real-world experiments show that Asyn...
The paper presents an approach for self-supervised training of Transformers based on the idea of masking and predicting. Two objectives have been utilized during the training process. One is patch reconstruction and the other one is named "patch concept classification" to learn different concepts/classes for data token...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper presents an approach for self-supervised training of Transformers based on the idea of masking and predicting. Two objectives have been utilized during the training process. One is patch reconstruction and the other one is named "patch concept classification" to learn different concepts/classes for da...
This paper presents a deep architecture based on RGB and IMU wearable sensors for multi-modal human activity recognition. The authors propose to encode the IMU sensor series into GAF images and then feed the RGB and GAF images into 2D CNNs for classification. Experiments are conducted on three public datasets. Strength...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a deep architecture based on RGB and IMU wearable sensors for multi-modal human activity recognition. The authors propose to encode the IMU sensor series into GAF images and then feed the RGB and GAF images into 2D CNNs for classification. Experiments are conducted on three public datasets. ...