Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
The paper proposes an approach to perform code infilling that supports multiple insertion points. The approach is by defining an infilling format that describes the missing code locations and the code to be infilled. The training approach is by the traditional autoregressive next token prediction loss. This paper indee...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes an approach to perform code infilling that supports multiple insertion points. The approach is by defining an infilling format that describes the missing code locations and the code to be infilled. The training approach is by the traditional autoregressive next token prediction loss. This pap...
This paper proposes a distributed model-based RL algorithm building on MuZero/EfficientZero which is designed to reduce the wall-clock time required to train the agent. While EfficientZero is sample efficient, it still takes a long time to train. This work identifies and addresses several system-level inefficiencies, r...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a distributed model-based RL algorithm building on MuZero/EfficientZero which is designed to reduce the wall-clock time required to train the agent. While EfficientZero is sample efficient, it still takes a long time to train. This work identifies and addresses several system-level inefficie...
In this paper, the authors tackle adversarial training. They pay more attention to the valley region of the loss landscape of DNNs. This is where collaborative examples exists. Collaborative examples fall within a bounded region of a benign example while having very small loss values. The authors first show that such e...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors tackle adversarial training. They pay more attention to the valley region of the loss landscape of DNNs. This is where collaborative examples exists. Collaborative examples fall within a bounded region of a benign example while having very small loss values. The authors first show tha...
The paper introduces a novel Neural Process (NP) variant which uses martingale posterior distributions to account for epistemic uncertainty. Martingale posteriors are a recent generalization of Bayesian inference proposed by Fong et al. (2021), in which valid posteriors are specified entirely in terms of the predictive...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper introduces a novel Neural Process (NP) variant which uses martingale posterior distributions to account for epistemic uncertainty. Martingale posteriors are a recent generalization of Bayesian inference proposed by Fong et al. (2021), in which valid posteriors are specified entirely in terms of the pr...
This paper presents an approach for adaptively allocating fine-tuning parameters during transfer of pre-trained models to downstream tasks. The authors propose an SVD inspired decomposition of the adapter matrices and develop various importance scores to assess which triplets in the SVD decomposition are removable. Th...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents an approach for adaptively allocating fine-tuning parameters during transfer of pre-trained models to downstream tasks. The authors propose an SVD inspired decomposition of the adapter matrices and develop various importance scores to assess which triplets in the SVD decomposition are remov...
The paper studies convergence of compressed SGD to a first-order stationary point. The paper considers unbiased compressors and improves state-of-the-art convergence rate and communication complexity for both finite-sum and stochastic settings. The algorithms presented in the paper differ in update rules and their requ...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper studies convergence of compressed SGD to a first-order stationary point. The paper considers unbiased compressors and improves state-of-the-art convergence rate and communication complexity for both finite-sum and stochastic settings. The algorithms presented in the paper differ in update rules and th...
This paper studies the bias propagation in federated learning, which is an important issue in terms of the fairness. The authors conducted a range of experiments to understand the potential mechanism of bias propagation and present several interesting insights. Strength (1) It is a first attempt to systematically anal...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper studies the bias propagation in federated learning, which is an important issue in terms of the fairness. The authors conducted a range of experiments to understand the potential mechanism of bias propagation and present several interesting insights. Strength (1) It is a first attempt to systematica...
This paper proposes a minimax game model of backdoor attacks between attackers and defenders in the federated learning (FL) paradigm. Based on the analysis of this model, it uses a reverse-engineering technique to defend against backdoor attacks in the FL process. It provides theoretical analysis and experimental resu...	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 minimax game model of backdoor attacks between attackers and defenders in the federated learning (FL) paradigm. Based on the analysis of this model, it uses a reverse-engineering technique to defend against backdoor attacks in the FL process. It provides theoretical analysis and experimen...
This paper studies a video-language model trained using only a single frame, finding that is can do quite well on many existing benchmarks. The paper then proposes new benchmarks on something-something, where more temporal understanding is required. The paper is well written and easy to understand. The experiments are ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper studies a video-language model trained using only a single frame, finding that is can do quite well on many existing benchmarks. The paper then proposes new benchmarks on something-something, where more temporal understanding is required. The paper is well written and easy to understand. The experime...
This paper improves upon prior approaches that generates furnitures. The proposed method is more controllable, able to generate / complete a scene from either object-level conditioning (think rendered, physically placed objects) or attribute-level conditioning (think word, stylized language description of the scene). T...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper improves upon prior approaches that generates furnitures. The proposed method is more controllable, able to generate / complete a scene from either object-level conditioning (think rendered, physically placed objects) or attribute-level conditioning (think word, stylized language description of the s...
The paper presents a modification of the NGNN framework of Zhang & Li which uses the embeddings of subgraphs around each node as node embeddings for an outer global message passing loop to DAGs (by defining a total ordering of possible subgraphs and selecting the largest for each node) in order to apply the architectu...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper presents a modification of the NGNN framework of Zhang & Li which uses the embeddings of subgraphs around each node as node embeddings for an outer global message passing loop to DAGs (by defining a total ordering of possible subgraphs and selecting the largest for each node) in order to apply the ar...
The paper focuses on the robustness of motion forecasting models and makes two main contributions: 1. additional labels for the Waymo Open Motion Dataset with causal labels, 2. evaluation of the robustness of several state of the art models using these "causal" labels. The causal labels include agents whose presence in...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper focuses on the robustness of motion forecasting models and makes two main contributions: 1. additional labels for the Waymo Open Motion Dataset with causal labels, 2. evaluation of the robustness of several state of the art models using these "causal" labels. The causal labels include agents whose pre...
This paper proposes a new method to measure image downscaling algorithms. The core idea is to upsample the downsampled image using blind/one-to-many generative-model-based super-resolution methods and compare the upsampled image with the original image. The similarity will be used as the final result. The authors show ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new method to measure image downscaling algorithms. The core idea is to upsample the downsampled image using blind/one-to-many generative-model-based super-resolution methods and compare the upsampled image with the original image. The similarity will be used as the final result. The autho...
The goal of this paper is to try to learn a goal-conditioned policy which is robust to goal-switching, i.e. when the goal given to the policy changes in the middle of the episode. The proposed approach uses a causal transformer, and proposes masking the goal randomly to ensure that the model (GMT) is able to predict th...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The goal of this paper is to try to learn a goal-conditioned policy which is robust to goal-switching, i.e. when the goal given to the policy changes in the middle of the episode. The proposed approach uses a causal transformer, and proposes masking the goal randomly to ensure that the model (GMT) is able to pr...
The authors proposed TSM that combines time series modeling method HiPPO and neural PDE solver. This proposed method has higher accuracy than existing models in 2-D Navier Stokes equation, and has lower inference latency. The authors generally identified these characteristics in various forcing and Reynolds number envi...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors proposed TSM that combines time series modeling method HiPPO and neural PDE solver. This proposed method has higher accuracy than existing models in 2-D Navier Stokes equation, and has lower inference latency. The authors generally identified these characteristics in various forcing and Reynolds num...
The authors proposed a novel problem setting of adversary aware partial label learning (PLL) and a novel solution including an adversary aware loss and immature teacher within momentum (ITWM) to solve it. Theoretical analysis are presented and some empirical results are reported. Strengths: 1. The problem setting is a...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors proposed a novel problem setting of adversary aware partial label learning (PLL) and a novel solution including an adversary aware loss and immature teacher within momentum (ITWM) to solve it. Theoretical analysis are presented and some empirical results are reported. Strengths: 1. The problem sett...
This work extends a recent approach to train MLPs with teacher GNNs to produce models capable of performing well on relational data while preserving low inference latency by omitting belief propagation. To motivate the paper, the authors identify two major limitations to existing MLPs trained with GNNs 1) loss of posit...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work extends a recent approach to train MLPs with teacher GNNs to produce models capable of performing well on relational data while preserving low inference latency by omitting belief propagation. To motivate the paper, the authors identify two major limitations to existing MLPs trained with GNNs 1) loss ...
The paper utilized Normalizing Flows models, which are trainable bijective mappings, composed of fully invertible layers. The core idea is to obtain a bounded support in the latent space by design via shifting the base distribution space (that of the latent variables) from a normal distribution to a bounded uniform o...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper utilized Normalizing Flows models, which are trainable bijective mappings, composed of fully invertible layers. The core idea is to obtain a bounded support in the latent space by design via shifting the base distribution space (that of the latent variables) from a normal distribution to a bounded u...
This paper proposes a framework for treatment effect estimation that consists of two components. The first is to balance representations by minimizing group distance and maximizing individual propensity confusion (i.e., PDIG). The second is to use both pre-balancing and post-balancing representations for outcome predic...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a framework for treatment effect estimation that consists of two components. The first is to balance representations by minimizing group distance and maximizing individual propensity confusion (i.e., PDIG). The second is to use both pre-balancing and post-balancing representations for outcom...
This paper studies the problem of linear regression under DP constraint and adversarial corruption. The authors propose a kind of variant of DP-SGD with a full-batch GD to improve sample complexity and adaptive clipping to guarantee robustness. Their theoretical results can improve the SOTA sample complexity without ad...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper studies the problem of linear regression under DP constraint and adversarial corruption. The authors propose a kind of variant of DP-SGD with a full-batch GD to improve sample complexity and adaptive clipping to guarantee robustness. Their theoretical results can improve the SOTA sample complexity wi...
This paper proposes a reward shaping scheme for MARL towards fixing QMIX, an existing algorithm suffering from suboptimality in non-monotonic settings. The method uses target shaping by predicting reward and next state uncertainty. 1. There are some important works missing in the related works which need to be discuss...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a reward shaping scheme for MARL towards fixing QMIX, an existing algorithm suffering from suboptimality in non-monotonic settings. The method uses target shaping by predicting reward and next state uncertainty. 1. There are some important works missing in the related works which need to be...
This paper mainly studies the problem of multivariate time series forecasting. The authors argue there are three basic challenges and accordingly propose a multi-view time-series graph structure representation method. Extensive results show their improved forecasting performances on some datasets as well as the improve...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: This paper mainly studies the problem of multivariate time series forecasting. The authors argue there are three basic challenges and accordingly propose a multi-view time-series graph structure representation method. Extensive results show their improved forecasting performances on some datasets as well as the...
The paper motivates the study of the relative sizes of regions of attraction of different equilibria, then tackles the case of symmetric, 2x2 coordination games. For this class of games, some partial/preliminary results are given. I found the paper to be well written and easy to read. It is certainly convincing in term...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper motivates the study of the relative sizes of regions of attraction of different equilibria, then tackles the case of symmetric, 2x2 coordination games. For this class of games, some partial/preliminary results are given. I found the paper to be well written and easy to read. It is certainly convincing...
This submission focuses on training teacher models with the goal of making them successful in knowledge distillation. The first part of the paper presents necessary conditions under which a teacher model minimizing the empirical loss computed with one-hot labels captures the Bayes classifier probabilities on the traini...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This submission focuses on training teacher models with the goal of making them successful in knowledge distillation. The first part of the paper presents necessary conditions under which a teacher model minimizing the empirical loss computed with one-hot labels captures the Bayes classifier probabilities on th...
This manuscript recommends an ensemble-like approach for semi-supervised pose estimation. A teacher-student setting where self-supervised pseudo-labels are created by the teacher for unlabeled data is extended to use two teacher-student pairs. The discrepancy between the two teachers is used to assess the quality of th...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This manuscript recommends an ensemble-like approach for semi-supervised pose estimation. A teacher-student setting where self-supervised pseudo-labels are created by the teacher for unlabeled data is extended to use two teacher-student pairs. The discrepancy between the two teachers is used to assess the quali...
This paper proposes a new method for learning policies first from offline dataset and then from a limited number of interactions with the real environment. The special feature of the method is that it learns a pseudometric that corresponds to the similarity between states in the environment. Then, when during online in...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a new method for learning policies first from offline dataset and then from a limited number of interactions with the real environment. The special feature of the method is that it learns a pseudometric that corresponds to the similarity between states in the environment. Then, when during o...
This paper deals with class imbalance from the data augmentation perspective. It is based on and motivated by an analysis on the relationship between data augmentation degree and performance per class. Strengths - Data augmentation is an interesting dimension to study class imbalance in general. - The proposed m...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper deals with class imbalance from the data augmentation perspective. It is based on and motivated by an analysis on the relationship between data augmentation degree and performance per class. Strengths - Data augmentation is an interesting dimension to study class imbalance in general. - The pr...
The paper proposes Contrastive Representation Ensemble and Aggregation for Multimodal FL (CreamFL) to exploit multimodal data from clients in FL settings in a privacy driven world. CreamFL trains large models from clients with heterogeneous architectures and multiple modalities. To fuse multimodal representations, a gl...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes Contrastive Representation Ensemble and Aggregation for Multimodal FL (CreamFL) to exploit multimodal data from clients in FL settings in a privacy driven world. CreamFL trains large models from clients with heterogeneous architectures and multiple modalities. To fuse multimodal representatio...
This paper tries to find empirically validated reasoning for the cause of high variance in fair deep learning, where the fairness metric of interest is average odds (AO), i.e., the average disparity between true and false positive rates. This is generally attributed to randomness (what they call non-determinism) in tra...	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 tries to find empirically validated reasoning for the cause of high variance in fair deep learning, where the fairness metric of interest is average odds (AO), i.e., the average disparity between true and false positive rates. This is generally attributed to randomness (what they call non-determinism...
The paper proposes a novel multi-agent RL environment which extends existing game Agar with multiple teams, thereby enabling M x N style games. A single agent in the game manages up to 16 growing cloned balls, which need to be increased in size to gain more rewards. The resulting action space has a discrete- (type of a...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a novel multi-agent RL environment which extends existing game Agar with multiple teams, thereby enabling M x N style games. A single agent in the game manages up to 16 growing cloned balls, which need to be increased in size to gain more rewards. The resulting action space has a discrete- (t...
This paper introduces a new explainable graph attention network model. The highlight of this model is its multi-channel explanability. Along with outputing node predictions, the model also generates extra edges and nodes importance scores across preset channels. To along these importance scores with multi-explanation c...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces a new explainable graph attention network model. The highlight of this model is its multi-channel explanability. Along with outputing node predictions, the model also generates extra edges and nodes importance scores across preset channels. To along these importance scores with multi-expla...
This paper proposes an efficient design of Transformer-based models (PatchTST) for multivariate time series forecasting and self-supervised representation learning. It segments time series into patches following similar philosophy of ViT and assume channels are independent. Extensive experiments on 8 eight datasets sho...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes an efficient design of Transformer-based models (PatchTST) for multivariate time series forecasting and self-supervised representation learning. It segments time series into patches following similar philosophy of ViT and assume channels are independent. Extensive experiments on 8 eight data...
This paper proposes a new post-processing algorithm (FaiREE) to satisfy group fairness constraints in classification with finite-sample and distribution-free theoretical guarantees. FaiREE can be adapted to satisfy various group fairness notions, e.g., Equality of Opportunity, Equalized Odds, and Demographic Parity. Sy...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a new post-processing algorithm (FaiREE) to satisfy group fairness constraints in classification with finite-sample and distribution-free theoretical guarantees. FaiREE can be adapted to satisfy various group fairness notions, e.g., Equality of Opportunity, Equalized Odds, and Demographic Pa...
The paper targets a novel {inr2vec} problem setting: to squeeze the weights of a learned implicit neural representation (INR) into compact latent codes for various downstream tasks. By training an encoder-decoder to recover the learned implicit function represented by the INR, the encoded feature can be fed into learni...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper targets a novel {inr2vec} problem setting: to squeeze the weights of a learned implicit neural representation (INR) into compact latent codes for various downstream tasks. By training an encoder-decoder to recover the learned implicit function represented by the INR, the encoded feature can be fed int...
A universal approximator is proposed in this paper using an oscillator NODE (ONODE) whose trajectories can "jump". It was shown that the ONODE is computationally efficient. A number of experiments (including classification tasks and extrapolation on time series) were conducted to validate the model. Strength The m...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: A universal approximator is proposed in this paper using an oscillator NODE (ONODE) whose trajectories can "jump". It was shown that the ONODE is computationally efficient. A number of experiments (including classification tasks and extrapolation on time series) were conducted to validate the model. *Strength...
Bias creeps into ML models trained from human behavioural data, which can usually be mitigated against but (typically) with labeled data for that purpose. This paper proposes a meta-algorithm for debiasing representations called Unsupervised Locality-based Proxy Label assignment. It is evaluated over five datasets, fro...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Bias creeps into ML models trained from human behavioural data, which can usually be mitigated against but (typically) with labeled data for that purpose. This paper proposes a meta-algorithm for debiasing representations called Unsupervised Locality-based Proxy Label assignment. It is evaluated over five datas...
This paper proposes a Group Masked Model Learning (GMML), a Self-Supervised Learning (SSL) mechanism for pretraining vision transformers with the ability to extract the contextual information present in all the concepts in an image. This is achieved by manipulating randomly groups of connected tokens, ensuingly coverin...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a Group Masked Model Learning (GMML), a Self-Supervised Learning (SSL) mechanism for pretraining vision transformers with the ability to extract the contextual information present in all the concepts in an image. This is achieved by manipulating randomly groups of connected tokens, ensuingly...
The authors propose FedPAC for personalized federated learning, which aims to learn a better representation by leveraging global semantic knowledge. Comprehensive experimental results multiple benchmarks demonstrate the effectiveness of the proposed framework. Strength The structure of the proposed FedPAC is reasonabl...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The authors propose FedPAC for personalized federated learning, which aims to learn a better representation by leveraging global semantic knowledge. Comprehensive experimental results multiple benchmarks demonstrate the effectiveness of the proposed framework. Strength The structure of the proposed FedPAC is r...
CLIP has shown remarkable performance on zero-shot visual recognition. However, adversarial examples still greatly affect CLIP's performance. This work propose a text-guided contrastive adversarial training loss to adopt CLIP to attain adversarial robustness for the datasets that are not seen during adversarial trainin...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: CLIP has shown remarkable performance on zero-shot visual recognition. However, adversarial examples still greatly affect CLIP's performance. This work propose a text-guided contrastive adversarial training loss to adopt CLIP to attain adversarial robustness for the datasets that are not seen during adversarial...
This paper extends the visual actionable affordance learning method of Where2Act (Mo et. al. 2021) for dual-gripper manipulation. To deal with the quadratic complexity of dual action space, the authors propose to sequence action predictions from two grippers and conditioning the second on the first. This reduces the ac...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper extends the visual actionable affordance learning method of Where2Act (Mo et. al. 2021) for dual-gripper manipulation. To deal with the quadratic complexity of dual action space, the authors propose to sequence action predictions from two grippers and conditioning the second on the first. This reduce...