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 tackles the survey bandit setting in which the decision maker has a limited budget of questions to ask a user, observes the answers to these questions (constituting the features of the user), then recommends a treatment and observes the outcome. The questions to ask are chosen with a decision theoretic entrop...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper tackles the survey bandit setting in which the decision maker has a limited budget of questions to ask a user, observes the answers to these questions (constituting the features of the user), then recommends a treatment and observes the outcome. The questions to ask are chosen with a decision theoreti...
In this work, the authors present a unification of goal-conditioned behavior cloning/supervised learning and hindsight relabeling methods (e.g., HER) under the framework of divergence minimization. They show that HER is a special case of a divergence minimization when considered under a goal-conditioned Q-learning appr...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In this work, the authors present a unification of goal-conditioned behavior cloning/supervised learning and hindsight relabeling methods (e.g., HER) under the framework of divergence minimization. They show that HER is a special case of a divergence minimization when considered under a goal-conditioned Q-learn...
This paper focuses on the data poisoning task on multimodal encoders. This paper investigates three types of poisoning attacks first. After that, it studies the effectiveness of attacking visual and linguistic features. In addition, it explores two types of defense mechanisms for defending against the attack on multimo...	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 focuses on the data poisoning task on multimodal encoders. This paper investigates three types of poisoning attacks first. After that, it studies the effectiveness of attacking visual and linguistic features. In addition, it explores two types of defense mechanisms for defending against the attack on...
This work improves Random-feature-based attention through the lens of control variates. The paper develops a more flexible form of control variates, which forms a novel attention mechanism that significantly reduces the approximation gap while maintaining linear complexity. The EVA achieves competitive results on sever...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work improves Random-feature-based attention through the lens of control variates. The paper develops a more flexible form of control variates, which forms a novel attention mechanism that significantly reduces the approximation gap while maintaining linear complexity. The EVA achieves competitive results ...
The paper addresses the task of updating a model by learning a stronger model while ensuring that formerly correctly classified instances will get wrong predictions as seldom as possible, a quantity called a negative flip rate (NFR). The paper starts with a theoretical analysis of using ensembles in this context. It th...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper addresses the task of updating a model by learning a stronger model while ensuring that formerly correctly classified instances will get wrong predictions as seldom as possible, a quantity called a negative flip rate (NFR). The paper starts with a theoretical analysis of using ensembles in this contex...
The authors introduce a simple modification to a metric-learning triplet loss by adding penalty term which encourages the distance between the anchor sample and negative sample to be similar to the distance between the positive sample and the negative sample. Strengths - The authors provide results for interesting ...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors introduce a simple modification to a metric-learning triplet loss by adding penalty term which encourages the distance between the anchor sample and negative sample to be similar to the distance between the positive sample and the negative sample. Strengths - The authors provide results for inte...
The paper proposes ROLLIN, and algorithm to reduce the complexity of learning an optimal policy in a sequence of contextual MDPs, i.e. a set of MDPs sharing everything but the reward function in this work. The main idea of the paper is that if the two contexts are close enough, we can modify the initial state distribut...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes ROLLIN, and algorithm to reduce the complexity of learning an optimal policy in a sequence of contextual MDPs, i.e. a set of MDPs sharing everything but the reward function in this work. The main idea of the paper is that if the two contexts are close enough, we can modify the initial state d...
A multi-scale structure-preserving heterologous image transformation method based on conditional adversarial network learning is proposed in this paper. The experimental results show that the proposed algorithm can better suppress the local structural distortion and has significant advantages in evaluation indicators s...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: A multi-scale structure-preserving heterologous image transformation method based on conditional adversarial network learning is proposed in this paper. The experimental results show that the proposed algorithm can better suppress the local structural distortion and has significant advantages in evaluation indi...
This paper extends the Logit Normalization technique by centering and introducing a threshold gamma in the denominator, which is referred to as NormSoftMax. Experimental results indicate that NormSoftMax works well in a number of DNN models with either croess-entropy loss or attention modules. Strength: NormSoftMa...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper extends the Logit Normalization technique by centering and introducing a threshold gamma in the denominator, which is referred to as NormSoftMax. Experimental results indicate that NormSoftMax works well in a number of DNN models with either croess-entropy loss or attention modules. Strength: No...
While graph neural networks (GNNs) and graph transformers (GTs) have been popular for graph representation learning recently, they have their own limitations: * GNNs, especially those based on message passing, are poor at modelling long-range dependencies and * GTs suffer from quadratic computation complexity. This ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: While graph neural networks (GNNs) and graph transformers (GTs) have been popular for graph representation learning recently, they have their own limitations: * GNNs, especially those based on message passing, are poor at modelling long-range dependencies and * GTs suffer from quadratic computation complexity....
This work propose a simple fine tuning method for unsupervised word alignment based on multilingual contextualized pre-trained language model, e.g., BERT. This method is based on the similarity score of the word representations in two languages and employs fine tuning by 1) iteratively augmenting code-switched training...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work propose a simple fine tuning method for unsupervised word alignment based on multilingual contextualized pre-trained language model, e.g., BERT. This method is based on the similarity score of the word representations in two languages and employs fine tuning by 1) iteratively augmenting code-switched ...
To address imbalanced label distributions in node classification tasks, the author suggested an over-sampling-based technique. To obtain unlabelled nodes rather than artificial ones, it uses unsupervised approaches in the embedding space. Additionally, new samples are simultaneously chosen using geometric confidence ra...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: To address imbalanced label distributions in node classification tasks, the author suggested an over-sampling-based technique. To obtain unlabelled nodes rather than artificial ones, it uses unsupervised approaches in the embedding space. Additionally, new samples are simultaneously chosen using geometric confi...
During the search, the neural architecture search (NAS) algorithm uses the validation accuracy of the trained candidate model for feedback. As training a candidate every time is expensive, performance predictors, which take the architecture as input and output the validation accuracy, are used. This paper proposes a tr...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: During the search, the neural architecture search (NAS) algorithm uses the validation accuracy of the trained candidate model for feedback. As training a candidate every time is expensive, performance predictors, which take the architecture as input and output the validation accuracy, are used. This paper propo...
This paper proposes a decentralized asynchronous stochastic first-order algorithm DADAO to minimize the sum of strongly convex functions over time-varying decentralized networks. This paper develops the theoretical performance guarantee of a related ODE system and tests the empirical performance by numerical simulation...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a decentralized asynchronous stochastic first-order algorithm DADAO to minimize the sum of strongly convex functions over time-varying decentralized networks. This paper develops the theoretical performance guarantee of a related ODE system and tests the empirical performance by numerical si...
The authors proposed to leverage Doob's h-transform to learn diffusion models within a constraint domain, including product spaces of any type, such as discrete, categorical, and their mix. Various experiments are conducted to demonstrate the interesting applications of such an algorithm. Pros: 1. The use of Doob'...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The authors proposed to leverage Doob's h-transform to learn diffusion models within a constraint domain, including product spaces of any type, such as discrete, categorical, and their mix. Various experiments are conducted to demonstrate the interesting applications of such an algorithm. Pros: 1. The use ...
This paper introduces a method for representation learning of multivariate time-series, which consists in the combination of Contrastive learning (for encoding the data in a compact latent space) and Self Organizing Maps (for better interpretability of the latent space). The use of interpretable unsupervised techniques...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper introduces a method for representation learning of multivariate time-series, which consists in the combination of Contrastive learning (for encoding the data in a compact latent space) and Self Organizing Maps (for better interpretability of the latent space). The use of interpretable unsupervised te...
This paper studies preserving differential privacy in when releasing information of streaming models. The authors propose a new DP mechanism that can release a set of sufficient statistics which can be used to estimate any symmetric norm of the data. This paper studies a very important problem, i.e., releasing the nor...	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 preserving differential privacy in when releasing information of streaming models. The authors propose a new DP mechanism that can release a set of sufficient statistics which can be used to estimate any symmetric norm of the data. This paper studies a very important problem, i.e., releasing...
The current research piece introduce combines existing pretrained language models as DeBERTa and ELECTRA into a single model. In addition, authors introduce GDES, a technique that the shared embeddings from the generator and discriminative models interfere with each other by avoiding propagating the gradients due to th...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The current research piece introduce combines existing pretrained language models as DeBERTa and ELECTRA into a single model. In addition, authors introduce GDES, a technique that the shared embeddings from the generator and discriminative models interfere with each other by avoiding propagating the gradients d...
This paper investigates applying knowledge distillation to federated learning. It summarizes many related works to give a set of common building blocks and experiments with different choices for each block on ResNet20/56 + Cifar10/100 to see their impact on the accuracy. Strength: - There is a good effort to summarize ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper investigates applying knowledge distillation to federated learning. It summarizes many related works to give a set of common building blocks and experiments with different choices for each block on ResNet20/56 + Cifar10/100 to see their impact on the accuracy. Strength: - There is a good effort to su...
This paper proposes a miniGPT-like transformer architecture for learning the world model of RL agents in POMDP environments. This World Model's training data is derived from the present policy model's interplay with the real world. The input and output images are then represented in the VQGAN-style and utilized to trai...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a miniGPT-like transformer architecture for learning the world model of RL agents in POMDP environments. This World Model's training data is derived from the present policy model's interplay with the real world. The input and output images are then represented in the VQGAN-style and utilized...
This paper targets at few-shot classification. Motivated by Label Smooth, it designs LP-FT-FB to impose equivariance on the feature extractor to address the over-fitting problem. It first train a randomly initialized linear classifier on novel samples with FBR. Then the pretrained feature extractor and the classifier a...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper targets at few-shot classification. Motivated by Label Smooth, it designs LP-FT-FB to impose equivariance on the feature extractor to address the over-fitting problem. It first train a randomly initialized linear classifier on novel samples with FBR. Then the pretrained feature extractor and the clas...
In this paper the authors address the problem of graph representation learning where edges are associated with text. In doing so the authors propose to combine pretrained networks with GNN with local network aggregation to inject text representation of nodes to its neighbors. The authors conduct experiments on various ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper the authors address the problem of graph representation learning where edges are associated with text. In doing so the authors propose to combine pretrained networks with GNN with local network aggregation to inject text representation of nodes to its neighbors. The authors conduct experiments on ...
This paper proposes to learn a multi-step inverse model [Lamb et al.] to learn a representation that ignores exogenous information (i.e., uncontrollable information) for offline RL. In addition, this paper introduces several temporally-correlated and diverse visual distractors on top of the v-d4rl dataset to investigat...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes to learn a multi-step inverse model [Lamb et al.] to learn a representation that ignores exogenous information (i.e., uncontrollable information) for offline RL. In addition, this paper introduces several temporally-correlated and diverse visual distractors on top of the v-d4rl dataset to in...
This work studies the impact of various sub-tokenization methods for neural language models trained on source code. Across a range of tasks and datasets, it demonstrates a number of actionable insights, such as that UnigramLM trained models perform better than BPE ones, and that certain types of tokenization (especiall...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work studies the impact of various sub-tokenization methods for neural language models trained on source code. Across a range of tasks and datasets, it demonstrates a number of actionable insights, such as that UnigramLM trained models perform better than BPE ones, and that certain types of tokenization (e...
In this paper, the authors devise a two-step conversion framework for ANN-SNN conversion. Unlike previous methods which consider the ANN-SNN conversion error, this work considers the error between source ANN and altered ANN as well as altered ANN with SNN. The idea is generally sound and simple to understand. Yet in my...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: In this paper, the authors devise a two-step conversion framework for ANN-SNN conversion. Unlike previous methods which consider the ANN-SNN conversion error, this work considers the error between source ANN and altered ANN as well as altered ANN with SNN. The idea is generally sound and simple to understand. Y...
This paper studies the problem of optimizing for aggregation weights assigned to client model updates along with weight decay so as to aid generalization. In particular, the authors take the viewpoint of seeing the aggregation and server side update analogous to that of mini-batch SGD in centralized training and propos...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies the problem of optimizing for aggregation weights assigned to client model updates along with weight decay so as to aid generalization. In particular, the authors take the viewpoint of seeing the aggregation and server side update analogous to that of mini-batch SGD in centralized training an...
The paper has considered non-uniformly sampled graphs from a metric-probability space and developed methods to estimate the unknown sampling density from those graphs. Also, the authors have experimentally tested the model and approach on synthetic and real-world graphs. Strength: The topic seems quite interesting and ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper has considered non-uniformly sampled graphs from a metric-probability space and developed methods to estimate the unknown sampling density from those graphs. Also, the authors have experimentally tested the model and approach on synthetic and real-world graphs. Strength: The topic seems quite interest...
The authors consider the problem of learning to solve constraint satisfaction problems. This entails finding values of a set of variables that satisfy given constraints. The authors propose using a transformer model, extended with recurrence, to perform this variable assignment. Existing work has shown that transforme...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors consider the problem of learning to solve constraint satisfaction problems. This entails finding values of a set of variables that satisfy given constraints. The authors propose using a transformer model, extended with recurrence, to perform this variable assignment. Existing work has shown that tr...
The paper proposes a trainable objective less ESD as the extra calibration loss jointly optimized with the NLL loss during training. It is claimed as a binning-free objective lease without need to tune any additional parameters. The experiments were extensively conducted across three architectures (MLPs, CNNs, and Tran...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes a trainable objective less ESD as the extra calibration loss jointly optimized with the NLL loss during training. It is claimed as a binning-free objective lease without need to tune any additional parameters. The experiments were extensively conducted across three architectures (MLPs, CNNs, ...
This paper proposes a preconditioned update to stochastic gradient methods， based upon Hutchinson’s approach to approximating the diagonal of the Hessian. Strength: introducing a new method to improve the convergence of the stochastic method Weaknesses: (1) The theoretical complexity is only for the smooth objective...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a preconditioned update to stochastic gradient methods， based upon Hutchinson’s approach to approximating the diagonal of the Hessian. Strength: introducing a new method to improve the convergence of the stochastic method Weaknesses: (1) The theoretical complexity is only for the smooth o...
In the paper, authors analyze the gradients of a maxout network with respect to inputs and parameters. In consequence, the authors improve the training of deep maxout networks in the case of fully-connected and convolutional networks. 1. The introduction paragraph, "Maxout networks," is not clear. 2. The introduction s...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In the paper, authors analyze the gradients of a maxout network with respect to inputs and parameters. In consequence, the authors improve the training of deep maxout networks in the case of fully-connected and convolutional networks. 1. The introduction paragraph, "Maxout networks," is not clear. 2. The introd...
The authors provide the first provably convergent algorithm for bilevel optimization with non-smooth non-Lipschitz lower-level function to their knowledge, via smoothing and penalty techniques. The proposed algorithm is empirically more accurate and efficient than existing state of the arts. Pros: The authors provide t...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors provide the first provably convergent algorithm for bilevel optimization with non-smooth non-Lipschitz lower-level function to their knowledge, via smoothing and penalty techniques. The proposed algorithm is empirically more accurate and efficient than existing state of the arts. Pros: The authors p...
This paper proposes an approach to domain generalisation whereby they use Fourier to generate synthetic images which are then used as a worst-case target domain to improve the model's robustness and therefore its generalisation. This is reminiscent of a domain adaptation approach. The amplitude-based generator that cre...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes an approach to domain generalisation whereby they use Fourier to generate synthetic images which are then used as a worst-case target domain to improve the model's robustness and therefore its generalisation. This is reminiscent of a domain adaptation approach. The amplitude-based generator ...
The submission considers a sample-efficient learning for PSRs. The nature of the problem is similar to (Liu et al. 2022), and previous algorithmic framework is extended from POMDPs to PSRs. This generalization to PSR seems relatively easy since the core test-set is provided a priori, and the construction of confidence ...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The submission considers a sample-efficient learning for PSRs. The nature of the problem is similar to (Liu et al. 2022), and previous algorithmic framework is extended from POMDPs to PSRs. This generalization to PSR seems relatively easy since the core test-set is provided a priori, and the construction of con...
This paper aims to show that using deterministic policy in a generative adversarial imitation learning algorithm is not proper, and results in a significant instability. This paper does the following: - empirical comparisons between SD3-GAIL, TD3-GAIL, TSSG - show a toy domain and prove that PLR approaches 0 as the siz...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper aims to show that using deterministic policy in a generative adversarial imitation learning algorithm is not proper, and results in a significant instability. This paper does the following: - empirical comparisons between SD3-GAIL, TD3-GAIL, TSSG - show a toy domain and prove that PLR approaches 0 as...
This paper introduces Mega, a gated, single-headed attention mechanism that incorporates a (damped, multi-dimensional) moving average. Instead of being computed directly on the hidden state, the query and keys are computed on a EMA-transformed version of it. The gating mechanism is akin to GRUs (with different inputs)....	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper introduces Mega, a gated, single-headed attention mechanism that incorporates a (damped, multi-dimensional) moving average. Instead of being computed directly on the hidden state, the query and keys are computed on a EMA-transformed version of it. The gating mechanism is akin to GRUs (with different ...
The authors propose a theoretical and empirical study on the problem of dimensional collapse due to over-smoothing in graph-regularized MLPs. Then, they propose to regularize the cross-correlation between node features and pooled features to alleviate this problem. The resulting model "Ortho-Reg" surpasses the performa...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose a theoretical and empirical study on the problem of dimensional collapse due to over-smoothing in graph-regularized MLPs. Then, they propose to regularize the cross-correlation between node features and pooled features to alleviate this problem. The resulting model "Ortho-Reg" surpasses the ...
The paper describes sampling (MCMC parlance) as a diffusion problem. In the forward direction, the target distribution diffuses to Gaussian noise, while the reverse direction of interest allows diffusing from noise to the target distribution. As can be seen and is highlighted by the authors, this problem is classical i...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper describes sampling (MCMC parlance) as a diffusion problem. In the forward direction, the target distribution diffuses to Gaussian noise, while the reverse direction of interest allows diffusing from noise to the target distribution. As can be seen and is highlighted by the authors, this problem is cla...
In this manuscript, the authors proposed a method named comfort zone which utilize the SVD to augment data for regression problems. To demonstrate the effectiveness of proposed approach, experiments across different datasets and network architecture are conducted. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this manuscript, the authors proposed a method named comfort zone which utilize the SVD to augment data for regression problems. To demonstrate the effectiveness of proposed approach, experiments across different datasets and network architecture are conducted. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~...
The paper addresses the robustness vs accuracy trade-off that is often observed in adversarial training of neural networks. It proposes the following ensembling technique to overcome this trade-off. The ensemble is made up of two models. The first model is trained using standard adversarial robustness techniques with o...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper addresses the robustness vs accuracy trade-off that is often observed in adversarial training of neural networks. It proposes the following ensembling technique to overcome this trade-off. The ensemble is made up of two models. The first model is trained using standard adversarial robustness technique...