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 proposed a simple yet effective data augmentation method, which is called LatentAugment. Strength: The proposed LatentAugment is straightforward and reasonable; Compared to some previous methods, the proposed method is efficient in terms of computation cost; Weakness: The performance gain is relatively sm...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposed a simple yet effective data augmentation method, which is called LatentAugment. Strength: The proposed LatentAugment is straightforward and reasonable; Compared to some previous methods, the proposed method is efficient in terms of computation cost; Weakness: The performance gain is relat...
This paper proposes a simple and effective method to accelerate the convergence speed and improve the performance of detection transformers, which only need to randomly initialize different groups of queries and use them to train the decoder separately. The method can be applied to all existing detection transformers a...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a simple and effective method to accelerate the convergence speed and improve the performance of detection transformers, which only need to randomly initialize different groups of queries and use them to train the decoder separately. The method can be applied to all existing detection transf...
This work proposed a knowledge distillation module for protein inverse folding, i.e., generating a protein sequence by its 3D structure. The proposed system recovers more diverse amino acid chains with a lower perplexity and higher recovery rate efficiently. Strengths: 1. The paper overall is easy to follow,...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work proposed a knowledge distillation module for protein inverse folding, i.e., generating a protein sequence by its 3D structure. The proposed system recovers more diverse amino acid chains with a lower perplexity and higher recovery rate efficiently. Strengths: 1. The paper overall is easy to...
In this paper, the authors proposed a method for implicitly learning continuous face-varying dimensions, without the need of asking an annotator to explicitly categorize a person. The result can be used in auditing datasets for diversity. The authors also proposed FAX, a novel dataset of 638,180 face similarity judgmen...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this paper, the authors proposed a method for implicitly learning continuous face-varying dimensions, without the need of asking an annotator to explicitly categorize a person. The result can be used in auditing datasets for diversity. The authors also proposed FAX, a novel dataset of 638,180 face similarity...
This paper proposes CLIP-FLOW, which is a method to improve the performance of optical flow models under the semi-supervised learning setting. The proposed method consists of three orthogonal parts: 1) Semi-supervised Contrastive Flow: an additional contrastive loss for training optical flow models; 2) Coordinate Encod...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes CLIP-FLOW, which is a method to improve the performance of optical flow models under the semi-supervised learning setting. The proposed method consists of three orthogonal parts: 1) Semi-supervised Contrastive Flow: an additional contrastive loss for training optical flow models; 2) Coordina...
The paper studies the meta-learning problem regarding various safe RL tasks that are modeled by constrained Markov decision processes (CMDPs). The authors proposed a CMDP embedded online learning framework to generalize learnt safe policies to other unseen CMDP environments. In theory, the authors proved regret bounds...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies the meta-learning problem regarding various safe RL tasks that are modeled by constrained Markov decision processes (CMDPs). The authors proposed a CMDP embedded online learning framework to generalize learnt safe policies to other unseen CMDP environments. In theory, the authors proved regre...
This paper presents two stage end-to-end image recognition model that is robust against adversarial attacks. The model consists of the semantic segmentation module that tries to divide the objects of interest in the scene into several sub-regions and a part-based recognition model that makes the final outputs. Experime...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents two stage end-to-end image recognition model that is robust against adversarial attacks. The model consists of the semantic segmentation module that tries to divide the objects of interest in the scene into several sub-regions and a part-based recognition model that makes the final outputs. ...
This paper proposes a method to quantify the temporal relationships between frames to effectively analyze temporal relevance for video action models. And then, comprehensive experiments present the effects of various factors on temporal modeling. Finally, based on experimental analysis, this paper answers some importan...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a method to quantify the temporal relationships between frames to effectively analyze temporal relevance for video action models. And then, comprehensive experiments present the effects of various factors on temporal modeling. Finally, based on experimental analysis, this paper answers some ...
This paper considers a specific nonsmooth bilevel problem with a structured nonsmooth regularization term, such as the $\ell_p$ norm and the OSCAR penalty, in the lower-level problem. The authors first recursively smooth the nonsmooth regularization term with a smoothing parameter $\mu^k$ that is set to decrease to $0$...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper considers a specific nonsmooth bilevel problem with a structured nonsmooth regularization term, such as the $\ell_p$ norm and the OSCAR penalty, in the lower-level problem. The authors first recursively smooth the nonsmooth regularization term with a smoothing parameter $\mu^k$ that is set to decreas...
This paper looks at the problem of unsupervised RL where skills must be learned for exploration in an environment without access to rewards. The authors take the approach of considering multiple metrics for guiding this unsupervised learning that have been proposed in the literature while choosing among each of them in...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper looks at the problem of unsupervised RL where skills must be learned for exploration in an environment without access to rewards. The authors take the approach of considering multiple metrics for guiding this unsupervised learning that have been proposed in the literature while choosing among each of...
The paper is motivated by recent findings in the field of knowledge distillation that the teacher should learn the true label distribution on the inputs rather than the one-hot labels with the best performance. The authors propose to take advantage of the Lipschitz regularization and consistency regularization to train...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper is motivated by recent findings in the field of knowledge distillation that the teacher should learn the true label distribution on the inputs rather than the one-hot labels with the best performance. The authors propose to take advantage of the Lipschitz regularization and consistency regularization ...
This paper studies the efficient training of large language models on heterogeneous and preemptive instances. It proposes SWARM parallelism and finds that training can be made communication-efficient with SWAM parallelism. Evaluation of a 13B model shows that the system achieves good training throughput and convergence...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper studies the efficient training of large language models on heterogeneous and preemptive instances. It proposes SWARM parallelism and finds that training can be made communication-efficient with SWAM parallelism. Evaluation of a 13B model shows that the system achieves good training throughput and con...
The purpose of this paper is to present an experimental analysis of factors that influence generalization for data exhibiting chaotic dynamics. For this purpose, the authors built a configurable and extensible model evaluation framework (Validyna). A number of experiments were constructed and run using Validyna. **Wea...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The purpose of this paper is to present an experimental analysis of factors that influence generalization for data exhibiting chaotic dynamics. For this purpose, the authors built a configurable and extensible model evaluation framework (Validyna). A number of experiments were constructed and run using Validyna...
This paper studied an important bandit problem: how to deal with historical data. This paper proposed a meta-algorithm to handle computation and storage issues. The problem this paper studied is important and interesting. The meta-algorithm is simple and intuitive. I feel before we talk about computational efficiency ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studied an important bandit problem: how to deal with historical data. This paper proposed a meta-algorithm to handle computation and storage issues. The problem this paper studied is important and interesting. The meta-algorithm is simple and intuitive. I feel before we talk about computational eff...
The paper proposes pFedKT to address the Non-IID data issue in federated learning. pFedKT is based on personalized FL with knowledge transfer. During local training, it utilizes a hypernetwork to generate a local model. The local model is updated with a contrastive loss as regularization, which limits the distance betw...	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 proposes pFedKT to address the Non-IID data issue in federated learning. pFedKT is based on personalized FL with knowledge transfer. During local training, it utilizes a hypernetwork to generate a local model. The local model is updated with a contrastive loss as regularization, which limits the dista...
This paper follows a line of work on learning to mimic algorithms with neural networks to allow for better out-of-distribution generalization. Following recent advancement, the authors aim to train a GNN to simulate the intermediate steps of graph algorithms, using intermediate states of the true algorithm as supervisi...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper follows a line of work on learning to mimic algorithms with neural networks to allow for better out-of-distribution generalization. Following recent advancement, the authors aim to train a GNN to simulate the intermediate steps of graph algorithms, using intermediate states of the true algorithm as s...
Pretrained deep representations from large datasets have a plethora of applications across vision and natural language. Understanding how information is encoded across the dimensions in the representation is important to design better algorithms for the respective downstream tasks. This paper discusses that the learn...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Pretrained deep representations from large datasets have a plethora of applications across vision and natural language. Understanding how information is encoded across the dimensions in the representation is important to design better algorithms for the respective downstream tasks. This paper discusses that t...
The authors introduce an algorithm capable of learning very general Markov network structures, from general data-types including continuous, discrete and mixed. # Strengths - The speed of the algorithm is very impressive and exciting. Further details need to be provided of course, of the experimental setup but the exp...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors introduce an algorithm capable of learning very general Markov network structures, from general data-types including continuous, discrete and mixed. # Strengths - The speed of the algorithm is very impressive and exciting. Further details need to be provided of course, of the experimental setup but...
This paper studies the problem of (text/image) guided (single) image editing using a pre-trained diffusion model. Specifically, it proposed a novel diffusion-based image translation method by disentangling style and content representations. Borrowed from the disentangling technique introduced recently (Tumanyan et al.,...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper studies the problem of (text/image) guided (single) image editing using a pre-trained diffusion model. Specifically, it proposed a novel diffusion-based image translation method by disentangling style and content representations. Borrowed from the disentangling technique introduced recently (Tumanyan...
The paper studies the Teacher-Student Framework, in the case when the teacher is potentially sub-optimal. Learning here is based on an ensemble off-policy method. The core concept is to develop a teach intervention function that is based on the estimated sub-optimality of student actions with respect to the teacher's v...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies the Teacher-Student Framework, in the case when the teacher is potentially sub-optimal. Learning here is based on an ensemble off-policy method. The core concept is to develop a teach intervention function that is based on the estimated sub-optimality of student actions with respect to the tea...
This work proposes to apply a post-hoc learning process for correcting label (class-prior) shifts between training and test distribution. It utilizes a labelled validation set that has the same label prior distribution as the test data and optimizes a delta-constrained objective (Eq.(1) or (2)). The delta quantifies ho...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work proposes to apply a post-hoc learning process for correcting label (class-prior) shifts between training and test distribution. It utilizes a labelled validation set that has the same label prior distribution as the test data and optimizes a delta-constrained objective (Eq.(1) or (2)). The delta quant...
This paper tries to explain why the adaptive methods converge faster than (S)GD. Specifically, this paper first defines a measure called $R_{med}^{OPT}$ which can be viewed as a stable version of the condition number. This paper then empirically observes that over tasks where Adam achieves empirical success, Adam and...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper tries to explain why the adaptive methods converge faster than (S)GD. Specifically, this paper first defines a measure called $R_{med}^{OPT}$ which can be viewed as a stable version of the condition number. This paper then empirically observes that over tasks where Adam achieves empirical success, ...
The paper presents a method for privacy-preserving neural network inference using homomorphic encryption. The main contributation is optimizing the convoluational layer to fit the SIMD structure of homomorphic encryption, including reducing the number of non-zero weights and re-training. The core idea is interesting a...	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 method for privacy-preserving neural network inference using homomorphic encryption. The main contributation is optimizing the convoluational layer to fit the SIMD structure of homomorphic encryption, including reducing the number of non-zero weights and re-training. The core idea is inter...
This paper introduces a new data set for deduplication historical news articles, and evaluates several baseline algorithms on it, finding that neural methods perform very well compared to n-gram-based baselines. Strengths The paper is well-written and introduces a resource and a baseline evaluation that will be of int...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a new data set for deduplication historical news articles, and evaluates several baseline algorithms on it, finding that neural methods perform very well compared to n-gram-based baselines. Strengths The paper is well-written and introduces a resource and a baseline evaluation that will b...
The paper studies the properties of sharpness aware minimization through a second order Taylor approximation, in the setting where the perturbation radius tends to zero. Under many strong assumptions, some sort of equivalence (in the limit) between different notions of sharpness-aware-minimization and the spectra of th...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies the properties of sharpness aware minimization through a second order Taylor approximation, in the setting where the perturbation radius tends to zero. Under many strong assumptions, some sort of equivalence (in the limit) between different notions of sharpness-aware-minimization and the spect...
In this paper, the authors proposed to leverage variational inference to learn the underlying distribution of the prompts with or without image feature as the condition. The effectiveness of the proposed method is verified on several benchmarks. Pros: 1. The proposed idea is intuitive and straightforward. 2. The improv...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors proposed to leverage variational inference to learn the underlying distribution of the prompts with or without image feature as the condition. The effectiveness of the proposed method is verified on several benchmarks. Pros: 1. The proposed idea is intuitive and straightforward. 2. Th...
This paper studies re-using the intermediate checkpoints during DP training for two purposes: 1. improve the accuracy; 2. construct reasonable uncertainty quantification. The contribution is on the methodology and empirical results. Strength: This paper is clearly presented. The contribution is solid and the experiment...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies re-using the intermediate checkpoints during DP training for two purposes: 1. improve the accuracy; 2. construct reasonable uncertainty quantification. The contribution is on the methodology and empirical results. Strength: This paper is clearly presented. The contribution is solid and the ex...
The paper advocates to use robust kernel density estimation in the self-attention mechanism to mitigate the issue of data contamination. In particular, the idea is to down-weight the weight of bad samples in the estimation process. The authors conduct empirical experiments to demonstrate the effectiveness of the propos...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper advocates to use robust kernel density estimation in the self-attention mechanism to mitigate the issue of data contamination. In particular, the idea is to down-weight the weight of bad samples in the estimation process. The authors conduct empirical experiments to demonstrate the effectiveness of th...
This paper post-processes the representation of deep neural networks from a physical perspective. This process considers not only the relationship between different components within a sample point, but also interaction between sample pairs. The proposed method: - promotes the discrete nature of sample points in a dee...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper post-processes the representation of deep neural networks from a physical perspective. This process considers not only the relationship between different components within a sample point, but also interaction between sample pairs. The proposed method: - promotes the discrete nature of sample points ...
The authors proposed a novel approach to improve worst-group performance, CROIS, that works well with reduced/limited group annotations on training/val set. The idea is simple yet effective: obtain good feature extractors on group-unlabeled data with ERM loss, and only retrain the last layer on group-labeled data with ...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors proposed a novel approach to improve worst-group performance, CROIS, that works well with reduced/limited group annotations on training/val set. The idea is simple yet effective: obtain good feature extractors on group-unlabeled data with ERM loss, and only retrain the last layer on group-labeled da...
This paper finds that previous methods, which computed margins using the true class and the most confusing class, are insufficient for handling the instance reweighting in the multi-class cases. The authors argue that instances close to the intersection of the decision boundaries should be more important, so they shoul...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper finds that previous methods, which computed margins using the true class and the most confusing class, are insufficient for handling the instance reweighting in the multi-class cases. The authors argue that instances close to the intersection of the decision boundaries should be more important, so th...
The paper proposes blind learning, a novel split learning framework for training transformers. In this framework, the patch embedding, MLP, and loss layers reside at the client, and the transformer blocks reside at the server. During training and inference, the client obfuscates its data and model weights by shufflin...	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 blind learning, a novel split learning framework for training transformers. In this framework, the patch embedding, MLP, and loss layers reside at the client, and the transformer blocks reside at the server. During training and inference, the client obfuscates its data and model weights by ...
Synthetic data generation is a key application of differentially-private generative models---it allows you to build a synthetic dataset that can be used repeatedly without privacy risks for downstream ML or data science use cases. This paper presents an empirical evaluation of different DP generative models on mainly t...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Synthetic data generation is a key application of differentially-private generative models---it allows you to build a synthetic dataset that can be used repeatedly without privacy risks for downstream ML or data science use cases. This paper presents an empirical evaluation of different DP generative models on ...
The paper makes an interesting observation that the error after the bellman optimality operation, instead of following gaussian noise, should in theory follow Gumbel distribution. This motivates the usage of Gumbel regression instead of least square to learn the Q functions, and the paper shows its connection to CQL in...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper makes an interesting observation that the error after the bellman optimality operation, instead of following gaussian noise, should in theory follow Gumbel distribution. This motivates the usage of Gumbel regression instead of least square to learn the Q functions, and the paper shows its connection t...
The paper proposes a novel approach for learning over neural fields that is discretization invariant. The authors propose a mathematical framework and design a network following this definition. They show experimental results on several vision benchmarks. Strengths: - The paper addresses an important problem, the depen...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a novel approach for learning over neural fields that is discretization invariant. The authors propose a mathematical framework and design a network following this definition. They show experimental results on several vision benchmarks. Strengths: - The paper addresses an important problem, t...
The paper presents a conditional generative model based on NeRF-VAE. The differences are that the proposed methods use a set of latent vectors to encode the scenes instead of one and it incorporates multi-view geometry local features for better feature learning. It also applies normalizing flow to model conditional pri...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: The paper presents a conditional generative model based on NeRF-VAE. The differences are that the proposed methods use a set of latent vectors to encode the scenes instead of one and it incorporates multi-view geometry local features for better feature learning. It also applies normalizing flow to model conditi...
In this paper, authors propose to utilize the label hierarchies to boost the performance of CLIP for zero shot classification. The main steps include the generation of the subclasses for each class by either using the GT label hierarchies or by querying GPT-3, then conduct the CLIP via these sub-classes, and finally m...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, authors propose to utilize the label hierarchies to boost the performance of CLIP for zero shot classification. The main steps include the generation of the subclasses for each class by either using the GT label hierarchies or by querying GPT-3, then conduct the CLIP via these sub-classes, and f...
In this work, the graph neural network structure is integrated into the GP framework to construct a kernel function with graph structure, and the low-rank approximation method is used to deduce the efficient training and inference method of GP, aiming to improve the performance and efficiency of graph structure data cl...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: In this work, the graph neural network structure is integrated into the GP framework to construct a kernel function with graph structure, and the low-rank approximation method is used to deduce the efficient training and inference method of GP, aiming to improve the performance and efficiency of graph structure...
This paper considers a hybrid federated learning scenario (both sample and feature levels). To handle heterogeneity in the feature level, the authors adopt an idea from VFL to split the model into blocks. Each model block corresponds to a feature block. To handle heterogeneity in the sample level, the authors use a mod...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper considers a hybrid federated learning scenario (both sample and feature levels). To handle heterogeneity in the feature level, the authors adopt an idea from VFL to split the model into blocks. Each model block corresponds to a feature block. To handle heterogeneity in the sample level, the authors u...
This paper explores the neural collaborative filtering bandits problem. Specifically, the authors first introduce relative groups to formulate groups given a specific content. Then, they use a meta learner and user learners to learn non-linear and linear reward functions. The authors also claim that this is the first w...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper explores the neural collaborative filtering bandits problem. Specifically, the authors first introduce relative groups to formulate groups given a specific content. Then, they use a meta learner and user learners to learn non-linear and linear reward functions. The authors also claim that this is the...