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 studies the domain adaptation when the target domain is unlabeled. A black-box predictor trained on the source domain is employed to assign noisy/pseudo labels to the unlabeled target domain data. In this way, the problem is transformed to learning with noisy labels. The noisy label problem is further solved ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper studies the domain adaptation when the target domain is unlabeled. A black-box predictor trained on the source domain is employed to assign noisy/pseudo labels to the unlabeled target domain data. In this way, the problem is transformed to learning with noisy labels. The noisy label problem is further...
The paper proposes to formulate some algorithmic losses (e.g. sorting) as bi-level optimization problems and solve the inner problem with second-order optimization. It reports performance gains over regular optimization Strengths: * The performance improvements are consistent across the experiments. Weaknesses: * I fi...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes to formulate some algorithmic losses (e.g. sorting) as bi-level optimization problems and solve the inner problem with second-order optimization. It reports performance gains over regular optimization Strengths: * The performance improvements are consistent across the experiments. Weaknesses...
This paper aims at avoiding computing radiance in less contributive parts by reparameterizing the sampling algorithm. This can help decrease the number of evaluations to MLPs. This sounds reasonable. But the experiments do not meet the authors’ expectations. First, decreasing the MLP evaluations should indeed improve e...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper aims at avoiding computing radiance in less contributive parts by reparameterizing the sampling algorithm. This can help decrease the number of evaluations to MLPs. This sounds reasonable. But the experiments do not meet the authors’ expectations. First, decreasing the MLP evaluations should indeed i...
The authors use spiking neural networks for image classification. They use time-to-first-spike (TTFS) coding. After a neuron has fired, any eventual subsequent spike is blocked. Training is done with a surrogate gradient. STRENGTHS: TTFS is appealing because it consumes few spikes, therefore little energy WEAKNESSES:...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The authors use spiking neural networks for image classification. They use time-to-first-spike (TTFS) coding. After a neuron has fired, any eventual subsequent spike is blocked. Training is done with a surrogate gradient. STRENGTHS: TTFS is appealing because it consumes few spikes, therefore little energy WEA...
This work proposes to use monotone and orthogonal operator to build implicit graph neural network. The adopted approach is well build upon monotone operator theory. The new operators allows more stable (via orthogonal operator) or expressive (via monotone operator) implicit GNN models with provable convergence. The dis...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work proposes to use monotone and orthogonal operator to build implicit graph neural network. The adopted approach is well build upon monotone operator theory. The new operators allows more stable (via orthogonal operator) or expressive (via monotone operator) implicit GNN models with provable convergence....
The authors show that across a range of tasks, the performance of Adam improves more as the batch size rises than the performance of SGD. In the full batch limit, Adam continues to significantly outperform SGD on language tasks. Based on this observation, they argue that the benefits of Adam are best studied in the ful...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors show that across a range of tasks, the performance of Adam improves more as the batch size rises than the performance of SGD. In the full batch limit, Adam continues to significantly outperform SGD on language tasks. Based on this observation, they argue that the benefits of Adam are best studied in...
This paper showed several interesting findings: 1. showed that the scaling laws fits well to the empirical test-cross entropy performance of a model at the end of the training. 2. showed that not all the fitted coefficients of the scaling laws vary that much when we change the sampling rate. Specifically, alpha and L ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper showed several interesting findings: 1. showed that the scaling laws fits well to the empirical test-cross entropy performance of a model at the end of the training. 2. showed that not all the fitted coefficients of the scaling laws vary that much when we change the sampling rate. Specifically, alph...
This paper provides a very comprehensive empirical analysis of predicting the test set performance on NLP tasks, and reveals several interesting properties like shape based metrics are generally better than scale based ones, and some fitting strategies (for empirical spectral densities of a weight matric) like E-TPL co...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper provides a very comprehensive empirical analysis of predicting the test set performance on NLP tasks, and reveals several interesting properties like shape based metrics are generally better than scale based ones, and some fitting strategies (for empirical spectral densities of a weight matric) like ...
The paper proposed a GAN-based video generative model that is based upon StyleGAN-V while producing videos with significantly fewer artifacts. The paper first outlined several common artifacts seen in the videos produced by StyleGAN-V, and then provided an extensive set of targeted solutions: 1. Alias-free techniques f...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The paper proposed a GAN-based video generative model that is based upon StyleGAN-V while producing videos with significantly fewer artifacts. The paper first outlined several common artifacts seen in the videos produced by StyleGAN-V, and then provided an extensive set of targeted solutions: 1. Alias-free tech...
The manuscript proposed an adaptive data weighting algorithm for 3D medical image segmentation. The proposed algorithm dynamically adjusts weights between supervised loss and self-supervised consistency loss during model training. The weights are adjusted by the values of loss terms on-the-fly. And the self-supervised ...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The manuscript proposed an adaptive data weighting algorithm for 3D medical image segmentation. The proposed algorithm dynamically adjusts weights between supervised loss and self-supervised consistency loss during model training. The weights are adjusted by the values of loss terms on-the-fly. And the self-sup...
This paper discusses the fundamental problem for certification using piece-wise linear activations like ReLU. Lipschitz-based certification is mainly investigated and advocated with smooth activations. ### Strength The authors present an interesting theoretical discussion regarding Lipschitz-based certification with r...	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 discusses the fundamental problem for certification using piece-wise linear activations like ReLU. Lipschitz-based certification is mainly investigated and advocated with smooth activations. ### Strength The authors present an interesting theoretical discussion regarding Lipschitz-based certificatio...
This paper follows BGRL, and notices that it generalizes poorly due to a lack of negative examples. They propose a non-contrastive framework for link prediction named T-BGRL that uses cheap “negative” samples to improve generalization. Experiments show that T-BGRL improves BGRL’s inductive performance in 5/6 datasets ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper follows BGRL, and notices that it generalizes poorly due to a lack of negative examples. They propose a non-contrastive framework for link prediction named T-BGRL that uses cheap “negative” samples to improve generalization. Experiments show that T-BGRL improves BGRL’s inductive performance in 5/6 d...
This paper applies VisionTransformer for Multivariate Time-Series Classification (VitMTSC) model that learns latent features from raw time-series data for classification tasks and is applicable to large-scale time-series data with millions of data samples of variable lengths. Their results show that it's can obtain nea...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper applies VisionTransformer for Multivariate Time-Series Classification (VitMTSC) model that learns latent features from raw time-series data for classification tasks and is applicable to large-scale time-series data with millions of data samples of variable lengths. Their results show that it's can ob...
This paper introduces a framework for analyzing the regret of risk-aware RL policy with coherent risk measures. The framework is based on an episodic finite-horizon Markov decision process (MDP) instead of the MDP with a discounted factor and infinite horizon. Under this framework, this paper proposes a Risk-Aware Uppe...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper introduces a framework for analyzing the regret of risk-aware RL policy with coherent risk measures. The framework is based on an episodic finite-horizon Markov decision process (MDP) instead of the MDP with a discounted factor and infinite horizon. Under this framework, this paper proposes a Risk-Aw...
This paper investigates an interesting problem, whether text-to-image generation model can help image recognition through synthetic data. The problem is interesting, and could be potentially useful in real world applications. The paper conduct different experiments to investigate how do the synthetic data from generat...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper investigates an interesting problem, whether text-to-image generation model can help image recognition through synthetic data. The problem is interesting, and could be potentially useful in real world applications. The paper conduct different experiments to investigate how do the synthetic data from...
Authors propose novel theoretical framework based on Girsanov formula to learn sampler based on reversing diffusion process from non-normalised density function. Strength: novel theoretical approach to sampling from non-normalised density Weakness: absent experimental comparison with Langevin Monte-Carlo methods as th...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: Authors propose novel theoretical framework based on Girsanov formula to learn sampler based on reversing diffusion process from non-normalised density function. Strength: novel theoretical approach to sampling from non-normalised density Weakness: absent experimental comparison with Langevin Monte-Carlo metho...
This paper studies how to speed up cross-silo federated learning, in which different silos need to have multiple peer-to-peer communication/synchronization at each round. Following previous works, this paper proposes a new efficient communication topology based on multigraph. Extensive experiments on multiple datasets ...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper studies how to speed up cross-silo federated learning, in which different silos need to have multiple peer-to-peer communication/synchronization at each round. Following previous works, this paper proposes a new efficient communication topology based on multigraph. Extensive experiments on multiple d...
The paper presents a new architecture for generative modeling that makes use of hyperbolic geometry. Specifically, the builds on Chen et al 2022 to propose a generative framework that combines both Autoencoders and GANs. Additionally, the paper also a presents a new method for splitting and concatenating vectors in the...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper presents a new architecture for generative modeling that makes use of hyperbolic geometry. Specifically, the builds on Chen et al 2022 to propose a generative framework that combines both Autoencoders and GANs. Additionally, the paper also a presents a new method for splitting and concatenating vector...
This paper focuses on developing a new type of embeddings for texts (sentences or documents, but not words). The proposed method's basic idea is to consider actual weights in the trained models (seem to assume language models). Then, the method computes the weight difference between the original and micro-tuned ones (...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper focuses on developing a new type of embeddings for texts (sentences or documents, but not words). The proposed method's basic idea is to consider actual weights in the trained models (seem to assume language models). Then, the method computes the weight difference between the original and micro-tune...
Wav2vec 2.0 is widespread among speech community and used extensively to pre-train speech models for later fin-tuning on the speech recognition task. The downsides of wav2vec 2.0 is necessity of huge resources to get state-of-the-art on Librispeech, hard to tune as there are a lot of hyper-parameters and also contrasti...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: Wav2vec 2.0 is widespread among speech community and used extensively to pre-train speech models for later fin-tuning on the speech recognition task. The downsides of wav2vec 2.0 is necessity of huge resources to get state-of-the-art on Librispeech, hard to tune as there are a lot of hyper-parameters and also c...
This paper presents a decoding algorithm that reduces (racial, gender) bias and toxicity in the generated outputs of language models. The proposed algorithm consists of token-level and context-dependent components that modify the bias terms of selected or all the layers of a language model at decoding time to suppress ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper presents a decoding algorithm that reduces (racial, gender) bias and toxicity in the generated outputs of language models. The proposed algorithm consists of token-level and context-dependent components that modify the bias terms of selected or all the layers of a language model at decoding time to s...
This paper considers offline reinforcement learning in the presence of unobserved confounding. There are two observations. First, including the "expert recommended" action (from the behavioral policy) as part of the state is helpful, because this action was informed by the unobserved confounder. Second, if there are pr...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers offline reinforcement learning in the presence of unobserved confounding. There are two observations. First, including the "expert recommended" action (from the behavioral policy) as part of the state is helpful, because this action was informed by the unobserved confounder. Second, if ther...
This paper analyzes classifier-guided sampling of diffusion models and observes that applying higher-order methods for accelerated sampling from the model does not work well in the guidance scenario. The root cause for that is that the additional guidance term defined by the classifier makes the generative ODE harder t...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper analyzes classifier-guided sampling of diffusion models and observes that applying higher-order methods for accelerated sampling from the model does not work well in the guidance scenario. The root cause for that is that the additional guidance term defined by the classifier makes the generative ODE ...
This paper presents analysis and empirical findings of distilling knowledge from large image-text foundational models (e.g. CLIP). The key observations are 1) resistance to the capacity gap of the teacher/student models, 2) the image-text models are not overconfident on the wrong labels. Strength: * The problem of dis...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper presents analysis and empirical findings of distilling knowledge from large image-text foundational models (e.g. CLIP). The key observations are 1) resistance to the capacity gap of the teacher/student models, 2) the image-text models are not overconfident on the wrong labels. Strength: * The proble...
This paper presents PONET to fuse multi-modal data for making survival predictions. Specifically, a spare biological pathway-informed embedding network (Han et al., 2018; Elmarakeby et al., 2021) is used to deal with genomic data to reveal meaningful biological interpretations. To capture modality-specific information,...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents PONET to fuse multi-modal data for making survival predictions. Specifically, a spare biological pathway-informed embedding network (Han et al., 2018; Elmarakeby et al., 2021) is used to deal with genomic data to reveal meaningful biological interpretations. To capture modality-specific info...
The paper finds a "circuit" (in this case, a subset of attention heads), C, that is meant to explain how GPT-2 small solves a particular synthetic task named Indirect Object Identification (IOI). This task composes 15 templates of forms like "Then, [B] and [A] went to the [PLACE]. [B] gave a [OBJECT] to [A]". The goal ...	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 finds a "circuit" (in this case, a subset of attention heads), C, that is meant to explain how GPT-2 small solves a particular synthetic task named Indirect Object Identification (IOI). This task composes 15 templates of forms like "Then, [B] and [A] went to the [PLACE]. [B] gave a [OBJECT] to [A]". T...
The adaptive version of the online adaptive topic-aware influence maximization problem is considered in this work, where its primary objective is the spreading of specific content in social networks. The problem is formulated as an infinite-horizon discounted MDP, and a model-based reinforcement algorithmic scheme, cal...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The adaptive version of the online adaptive topic-aware influence maximization problem is considered in this work, where its primary objective is the spreading of specific content in social networks. The problem is formulated as an infinite-horizon discounted MDP, and a model-based reinforcement algorithmic sch...
This paper proposes a new approach to message classification. Specifically: * It is based on SOTA NLP building blocks. * It has a novel technique to infuse metadata. This paper shows that: * Adding metadata increase the performance of their model. * Their ulti-modality building block outperforms other methods. Stre...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a new approach to message classification. Specifically: * It is based on SOTA NLP building blocks. * It has a novel technique to infuse metadata. This paper shows that: * Adding metadata increase the performance of their model. * Their ulti-modality building block outperforms other method...
The focus of this paper is deep machine unlearning. This problem refers to removing the influence of a subset of data from the weights of a trained deep model. For instance, an application of this kind of method is to remove user data to guarantee their "right to be forgotten." Other applications include removing noisy...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The focus of this paper is deep machine unlearning. This problem refers to removing the influence of a subset of data from the weights of a trained deep model. For instance, an application of this kind of method is to remove user data to guarantee their "right to be forgotten." Other applications include removi...
This paper proposes a model for coarse-to-fine point cloud completion. The model first extracts global features from a partial input, then produces a coarse point cloud which is fed to several blocks in a sequential manner, each increasing the number of output points and modifying them. Overall, architecture heavily re...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a model for coarse-to-fine point cloud completion. The model first extracts global features from a partial input, then produces a coarse point cloud which is fed to several blocks in a sequential manner, each increasing the number of output points and modifying them. Overall, architecture he...
This paper proposes a novel method to generate realistic images on the target sparse domain by transferring a large-scale generator from the source domain without retraining the generator. Also, the proposed method can generate diverse target samples. The key idea is to solve an optimization problem on the latent domai...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper proposes a novel method to generate realistic images on the target sparse domain by transferring a large-scale generator from the source domain without retraining the generator. Also, the proposed method can generate diverse target samples. The key idea is to solve an optimization problem on the late...
This paper proposes a new distributional RL algorithm named SinkhornDRL, which proposes to solve the distribution matching problem in distributional RL with Sinkhorn divergence. Sinkhorn divergence in formulation is equivalent to an entropy-regularized version of MMD, and it could be featured as an interpolation betw...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a new distributional RL algorithm named SinkhornDRL, which proposes to solve the distribution matching problem in distributional RL with Sinkhorn divergence. Sinkhorn divergence in formulation is equivalent to an entropy-regularized version of MMD, and it could be featured as an interpolat...
This work proposes a contrastive learning method for self-supervised node-level tasks in graph domain. It involves (1) sampling positive samples from the first order neighbourhoods and (2) kernelizing negative loss to reduce the training time and memory overheads. The main contribution of this paper is presenting a sim...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work proposes a contrastive learning method for self-supervised node-level tasks in graph domain. It involves (1) sampling positive samples from the first order neighbourhoods and (2) kernelizing negative loss to reduce the training time and memory overheads. The main contribution of this paper is presenti...
The paper studies federated optimization under strong convexity and second-order similarity assumptions. Using the idea of trading off local computational cost for less communication, the paper proposes a new algorithm SVRP based on proximal point methods and variance reduction (SVRG). With client sampling, SVRP can ac...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper studies federated optimization under strong convexity and second-order similarity assumptions. Using the idea of trading off local computational cost for less communication, the paper proposes a new algorithm SVRP based on proximal point methods and variance reduction (SVRG). With client sampling, SVR...
The paper "Localized Graph Contrastive Learning" suggests a contrastive learning method (Local-GCL) for graph data, based on the InfoNCE loss. In this setting, each graph node is associated with a feature vector, and the goal is to learn a meaningful representation of the graph nodes (representation quality can be meas...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper "Localized Graph Contrastive Learning" suggests a contrastive learning method (Local-GCL) for graph data, based on the InfoNCE loss. In this setting, each graph node is associated with a feature vector, and the goal is to learn a meaningful representation of the graph nodes (representation quality can...
The paper analyses tree architectures for neural trees ensembles via NTK. Authors provide important insights on how the structure of trees affects NTK and consider depth limit of NTK. Experimental results are based on kernel regression with arising NTK. Strength: -- rigorous theoretical analysis -- comparison of differ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper analyses tree architectures for neural trees ensembles via NTK. Authors provide important insights on how the structure of trees affects NTK and consider depth limit of NTK. Experimental results are based on kernel regression with arising NTK. Strength: -- rigorous theoretical analysis -- comparison o...
This paper proposes to use the "linear probing then fine-tuning" strategy (LP-FT) introduced in [1], as well as the firth bias reduction (FBR) introduced in [2], to learn the adaptation to novel-classes step in few-shot learning. The proposed method is evaluated on a variety of FSL tasks and demonstrate good results in...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to use the "linear probing then fine-tuning" strategy (LP-FT) introduced in [1], as well as the firth bias reduction (FBR) introduced in [2], to learn the adaptation to novel-classes step in few-shot learning. The proposed method is evaluated on a variety of FSL tasks and demonstrate good re...
The Author's investigation reveals critical insights into the gap of uniform convergence for analyzing pre-trained representations, their stochastic nature under gradient descent optimization, what model convergence means to them, and how they might interact with downstream tasks. Authors propose a simple approach whic...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The Author's investigation reveals critical insights into the gap of uniform convergence for analyzing pre-trained representations, their stochastic nature under gradient descent optimization, what model convergence means to them, and how they might interact with downstream tasks. Authors propose a simple appro...
The authors propose a technique to learn a globally Lipschitz function when the data lie near a manifold. They derive a suitable constrained optimization problem together with the associated optimization scheme. Theoretical results are provided which show that under some conditions it is sensible to solve the empirical...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors propose a technique to learn a globally Lipschitz function when the data lie near a manifold. They derive a suitable constrained optimization problem together with the associated optimization scheme. Theoretical results are provided which show that under some conditions it is sensible to solve the e...
In this paper, the authors present a formulation to create a trainable activation function based on a convex combination of other activations. This new mixture approach is intended to work on physics-informed neural networks (PINNd) where higher order derivatives play a role, thus causing problems for some of the piece...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: In this paper, the authors present a formulation to create a trainable activation function based on a convex combination of other activations. This new mixture approach is intended to work on physics-informed neural networks (PINNd) where higher order derivatives play a role, thus causing problems for some of t...