Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

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

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
This paper studied the lower bound of minimizing PL functions. It provided an $\Omega((L/\mu)^{1-a})$ (arbitrary small $a>0$) lower bound for finding the optimal solution. It reveals a fundamental difference between PL functions and strongly convex minimization, also it shows that GD is already nearly optimal for solvi...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studied the lower bound of minimizing PL functions. It provided an $\Omega((L/\mu)^{1-a})$ (arbitrary small $a>0$) lower bound for finding the optimal solution. It reveals a fundamental difference between PL functions and strongly convex minimization, also it shows that GD is already nearly optimal f...
This paper introduces an SSL framework that is capable of learning different concepts in an image while the previous studies can only capture the dominant concepts. To this end, they proposed MC-SSL based on the components from SiT and DINO, i.e., Group Mask Model Learning (GMML) and Patch Concept Learning (PCL) based ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper introduces an SSL framework that is capable of learning different concepts in an image while the previous studies can only capture the dominant concepts. To this end, they proposed MC-SSL based on the components from SiT and DINO, i.e., Group Mask Model Learning (GMML) and Patch Concept Learning (PCL...
The authors present a novel algorithm for recursive neural network processing of sequence inputs. The algorithm combines easy-first parsing techniques with beam search in order to efficiently explore the space of possible latent tree structures during training and inference. They also present soft relaxations of the fr...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors present a novel algorithm for recursive neural network processing of sequence inputs. The algorithm combines easy-first parsing techniques with beam search in order to efficiently explore the space of possible latent tree structures during training and inference. They also present soft relaxations o...
The authors propose a modified version of the OAVI algorithm for computing the generators of the set of polynomials who has a given dataset as roots - called the ideal. They show theoretically that their proposed modified algorithm yields significant savings in computational complexity. They then use these ideals to so...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors propose a modified version of the OAVI algorithm for computing the generators of the set of polynomials who has a given dataset as roots - called the ideal. They show theoretically that their proposed modified algorithm yields significant savings in computational complexity. They then use these idea...
The authors analyze the observation that reformulating regression as classification often yields better performance by better learning high-entropy feature representations. To retain the benefits of both high entropy and ordinality of feature learning, the authors propose an ordinal entropy loss to encourage higher-ent...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The authors analyze the observation that reformulating regression as classification often yields better performance by better learning high-entropy feature representations. To retain the benefits of both high entropy and ordinality of feature learning, the authors propose an ordinal entropy loss to encourage hi...
The paper proposes FedeRiCo, a decentralized federated learning framework that allows clients to collaborate much or little with other participating clients to enhance the performance of the federated learned models. FedeRiCo leverages an EM-style optimization procedure to solve the global model weights and clients' co...	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 proposes FedeRiCo, a decentralized federated learning framework that allows clients to collaborate much or little with other participating clients to enhance the performance of the federated learned models. FedeRiCo leverages an EM-style optimization procedure to solve the global model weights and cli...
The authors deeply dived into CAUSAL REPRESENTATION LEARNING on temporal sequences of observations and argue that temporal sequences of observations may still contain instantaneous causal relations in practice. In order to model the causal structure of the temporal sequences of observations with INSTANTANEOUS AND TEM...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The authors deeply dived into CAUSAL REPRESENTATION LEARNING on temporal sequences of observations and argue that temporal sequences of observations may still contain instantaneous causal relations in practice. In order to model the causal structure of the temporal sequences of observations with INSTANTANEOUS...
The paper proposes a model that learns both features and segmentation without supervision. The method combines contrastive learning, a Markov random field model for the feature maps and segmentation. Learning utilizes two losses; position loss which optimizes the probability of the feature vector at each location relat...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The paper proposes a model that learns both features and segmentation without supervision. The method combines contrastive learning, a Markov random field model for the feature maps and segmentation. Learning utilizes two losses; position loss which optimizes the probability of the feature vector at each locati...
The authors propose a method (ELE) to incorporate information from expert (observation-only) demonstrations into an RL agent by first learning an estimate of "progress" and using this estimate as an intrinsic reward. Progress in this case is defined to be the temporal distance between two states. Results are first sho...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose a method (ELE) to incorporate information from expert (observation-only) demonstrations into an RL agent by first learning an estimate of "progress" and using this estimate as an intrinsic reward. Progress in this case is defined to be the temporal distance between two states. Results are f...
The paper suggests that datasets could be heterogeneous since the label generation process might vary systematically, which the paper calls 'label style'. Based on this factor, it further proposes to condition probabilistic segmentation models on label style and to train the models on datasets containing differing labe...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper suggests that datasets could be heterogeneous since the label generation process might vary systematically, which the paper calls 'label style'. Based on this factor, it further proposes to condition probabilistic segmentation models on label style and to train the models on datasets containing differ...
This paper proposes a method for more efficient video transformer networks based on linear attention. The main contribution is using a context-gating like mechanism to re-weight the features before attention. The approach is evaluated on kinetics and something-something datasets. The paper is well written, and the expe...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a method for more efficient video transformer networks based on linear attention. The main contribution is using a context-gating like mechanism to re-weight the features before attention. The approach is evaluated on kinetics and something-something datasets. The paper is well written, and ...
The authors present an approach to utilize a convolutional neural network (CLEEGN) with limited novelty (there are so many such approaches on the market already) and super limited evaluation on a tiny EEG dataset. Due to little originality and minimal assessment, the paper is not suitable for the ICLR. The application ...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The authors present an approach to utilize a convolutional neural network (CLEEGN) with limited novelty (there are so many such approaches on the market already) and super limited evaluation on a tiny EEG dataset. Due to little originality and minimal assessment, the paper is not suitable for the ICLR. The appl...
This paper describes the method for training the speech enhancements model using only real noisy speech (as opposed to the typical use of labeled clean and noisy speech). The paper describes the use of a Quality Predictor (Q) to guide the training of the speech enhancement model. The main contributions of the paper ar...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper describes the method for training the speech enhancements model using only real noisy speech (as opposed to the typical use of labeled clean and noisy speech). The paper describes the use of a Quality Predictor (Q) to guide the training of the speech enhancement model. The main contributions of the ...
This paper presents a learning algorithm for first price auction with budget constraint. The authors consider an artificial discounted scenario and measure the regret of a learning algorithm as the sum of the suboptimality gap. Their propose two algorithms: the first always observe the highest bid of the others while t...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper presents a learning algorithm for first price auction with budget constraint. The authors consider an artificial discounted scenario and measure the regret of a learning algorithm as the sum of the suboptimality gap. Their propose two algorithms: the first always observe the highest bid of the others...
This paper studies dimension-wise disentangled representations for downstream applications. Through extensive experiments, the authors conclude that disentanglement is not a necessity for achieving good performance in downstream tasks, and general-purpose representation learning methods could achieve better (or at leas...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies dimension-wise disentangled representations for downstream applications. Through extensive experiments, the authors conclude that disentanglement is not a necessity for achieving good performance in downstream tasks, and general-purpose representation learning methods could achieve better (or...
This paper aims to address two issues with prior work on the task of hierarchical time series forecasting (HSTF): (i) the assumption of rigid coherency (i.e. forecasts enforce the time-series values of datasets to satisfy the underlying hierarchical constraints strictly) and (ii) the lack of calibrated forecasts. To a...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper aims to address two issues with prior work on the task of hierarchical time series forecasting (HSTF): (i) the assumption of rigid coherency (i.e. forecasts enforce the time-series values of datasets to satisfy the underlying hierarchical constraints strictly) and (ii) the lack of calibrated forecast...
The work starts with using marginal likelihood as an alternative objective to validation metric for HPO, which has advantages in the small data region and potentially better generalization performance. It then talked about how to compute marginal likelihood and proposed an efficient method that only requires a single t...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The work starts with using marginal likelihood as an alternative objective to validation metric for HPO, which has advantages in the small data region and potentially better generalization performance. It then talked about how to compute marginal likelihood and proposed an efficient method that only requires a ...
The work examines the problem of continuous-time discrete diffusion models, and the paper mainly focuses on 1) extending the existing discrete-time discrete diffusion models to the continuous-time via stochastic jump processes, 2) extending the score function to generic categorical variables and deriving a score-matchi...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The work examines the problem of continuous-time discrete diffusion models, and the paper mainly focuses on 1) extending the existing discrete-time discrete diffusion models to the continuous-time via stochastic jump processes, 2) extending the score function to generic categorical variables and deriving a scor...
The script studies the minimum width of neural networks to achieve universal approximation. Similar topics are also extensively studied in Park et al. 2022. The presentation needs to be improved. For instance, many terminologies such as C-UAP are not defined. 1. I doubt the significance of the script compared with the...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The script studies the minimum width of neural networks to achieve universal approximation. Similar topics are also extensively studied in Park et al. 2022. The presentation needs to be improved. For instance, many terminologies such as C-UAP are not defined. 1. I doubt the significance of the script compared ...
This paper presents an Ordered GNN for handling the heterophily problem, which attempts to improve the updating process in the message passing scheme. The paper studies an important problem, i.e., how to extend current GNNs to heterophilic graphs. This paper presents an Ordered GNN for handling the heterophily problem,...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper presents an Ordered GNN for handling the heterophily problem, which attempts to improve the updating process in the message passing scheme. The paper studies an important problem, i.e., how to extend current GNNs to heterophilic graphs. This paper presents an Ordered GNN for handling the heterophily ...
The authors studied the problem of amino-acid mutation prediction in protein by regarding this problem as solving a denoising problem with lightweight GNN. Several large-scale pre-trained models have previously been proposed and applied to this problem, such as ESM-1v and MSA transformer. The authors proposed a new lig...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors studied the problem of amino-acid mutation prediction in protein by regarding this problem as solving a denoising problem with lightweight GNN. Several large-scale pre-trained models have previously been proposed and applied to this problem, such as ESM-1v and MSA transformer. The authors proposed a...
The paper investigates the problem of auditing fairness automatically, online, and in a black-box manner. It proposes AVOIR to address this problem for multiple fairness metrics. It claims to optimize the previously used adaptive Hoeffding inequality to decrease the sample complexity. AVOIR also comes along a visualisa...	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 investigates the problem of auditing fairness automatically, online, and in a black-box manner. It proposes AVOIR to address this problem for multiple fairness metrics. It claims to optimize the previously used adaptive Hoeffding inequality to decrease the sample complexity. AVOIR also comes along a v...
Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry fo...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. sorry for the late notice. Hi, I'm currently on a medical leave and won't be able to perform ICRL review duties. ...
The paper addresses the learn optimal policy for data augmentation on image classification task. Learning the optimal augmentation policy, which is the latent variable, interesting research problem. The paper proposes a simple via EM-based method (call LatentAugment) to estimate the probability of optimal policy via c...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper addresses the learn optimal policy for data augmentation on image classification task. Learning the optimal augmentation policy, which is the latent variable, interesting research problem. The paper proposes a simple via EM-based method (call LatentAugment) to estimate the probability of optimal poli...
The paper proposes an offline imitation learning (IL) method supported by a word model. The world model is obtained by training on the transitions collected in a dataset, without any direct access to the actual environment. A intrinsic reward is also designed that measures the divergence between expert and agent in the...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes an offline imitation learning (IL) method supported by a word model. The world model is obtained by training on the transitions collected in a dataset, without any direct access to the actual environment. A intrinsic reward is also designed that measures the divergence between expert and agen...
The paper presents a simple improvement to universal adversarial perturbations to make them robust to transformations. The paper motivates empirical experiments via theoretical observations. How does the ability to craft perturbations which are robust under transformations depend on the invariance properties of the vi...	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 presents a simple improvement to universal adversarial perturbations to make them robust to transformations. The paper motivates empirical experiments via theoretical observations. How does the ability to craft perturbations which are robust under transformations depend on the invariance properties o...
This paper introduces a very interesting view of deep latent space. The authors introduce a idea inspired by phisics, and propose a new model regularization method. The new regularization method promotes distance between data points based on they intra-sample properties learned from the task. The idea is validated on s...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper introduces a very interesting view of deep latent space. The authors introduce a idea inspired by phisics, and propose a new model regularization method. The new regularization method promotes distance between data points based on they intra-sample properties learned from the task. The idea is valida...
This paper aims to address one challenge in post-processing fairness mitigation algorithms: they require either distributional assumptions or access to infinite data examples. This paper presents a framework that takes in any score-based black-box classifier and outputs group-wise thresholds to correct for the fairness...	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 aims to address one challenge in post-processing fairness mitigation algorithms: they require either distributional assumptions or access to infinite data examples. This paper presents a framework that takes in any score-based black-box classifier and outputs group-wise thresholds to correct for the ...
This paper reformulates few shot learning about 3D parsing from analogy-driven prediction view. The authors design an Analogical Networks that are comprised of an encoder, retriever and modulator sub-modules. Experiment are performed on benchmark for 3D object segmentation. Strength: 1. All figures and tables are in ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper reformulates few shot learning about 3D parsing from analogy-driven prediction view. The authors design an Analogical Networks that are comprised of an encoder, retriever and modulator sub-modules. Experiment are performed on benchmark for 3D object segmentation. Strength: 1. All figures and tables...
The paper proposes an autoencoder for data residing on a Riemannian manifold alongside a geometric regularize. The reconstruction trivially replaces Euclidean distances with geodesics ones, and the bulk of the paper is concerned with the regularization. I did not fully understand the regulerization scheme (details belo...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes an autoencoder for data residing on a Riemannian manifold alongside a geometric regularize. The reconstruction trivially replaces Euclidean distances with geodesics ones, and the bulk of the paper is concerned with the regularization. I did not fully understand the regulerization scheme (deta...
This paper examines the properties of identity in bilinear models for knowledge encoding. To satisfy the properties, a modified bilinear model is proposed, in which the entity and relation embeddings are normalized. This form a unit ball. It is shown that the desired properties of unity can be met. In the experiments o...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper examines the properties of identity in bilinear models for knowledge encoding. To satisfy the properties, a modified bilinear model is proposed, in which the entity and relation embeddings are normalized. This form a unit ball. It is shown that the desired properties of unity can be met. In the exper...
The paper attempts to provide a theoretical analysis computation in predictive coding networks (PCNs), relating it to the more commonly used backpropagation-based (BP) training of feedforward artificial neural nets (ANNs). Strengths: Understanding the relation between PCN and usual ANN with BP is definitely a worthy g...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The paper attempts to provide a theoretical analysis computation in predictive coding networks (PCNs), relating it to the more commonly used backpropagation-based (BP) training of feedforward artificial neural nets (ANNs). Strengths: Understanding the relation between PCN and usual ANN with BP is definitely a ...
The authors analyze ensembles in the context of selective prediction. The authors show that under loose assumptions, a selective classifier using the probability assigned to the predicted class to decide whether to abstain from predicting can be improved by ensembling $n$ copies of it (differing only in their random se...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors analyze ensembles in the context of selective prediction. The authors show that under loose assumptions, a selective classifier using the probability assigned to the predicted class to decide whether to abstain from predicting can be improved by ensembling $n$ copies of it (differing only in their r...
In this paper, the authors introduce a new library for benchmarking RL methods on NLP tasks, along with a modified version of PPO for on-policy optimization. Results demonstrate that the proposed NLPO algorithm outperforms the PPO baseline according to human assessment. Strength: • This paper builds a cross-task divers...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors introduce a new library for benchmarking RL methods on NLP tasks, along with a modified version of PPO for on-policy optimization. Results demonstrate that the proposed NLPO algorithm outperforms the PPO baseline according to human assessment. Strength: • This paper builds a cross-tas...
This paper proposed a new sparse SSL approach by exploring the correlation between in-training sparsity and SSL. This paper investigates the challenges of the sparsity-induced asymmetric SSL (a.k.a prune-and-regrow) and proposes synchronized contrastive pruning approach. The proposed approach is adaptive to various gra...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper proposed a new sparse SSL approach by exploring the correlation between in-training sparsity and SSL. This paper investigates the challenges of the sparsity-induced asymmetric SSL (a.k.a prune-and-regrow) and proposes synchronized contrastive pruning approach. The proposed approach is adaptive to var...
A method for disentangled representation learning from sequential data is proposed. It is based on the eigendecomposition of an estimation of the Koopman operator of dynamics. The method is demonstrated with standard benchmark datasets. ### Strengths - The idea of using Koopman operator's spectrum for static/dynamic d...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: A method for disentangled representation learning from sequential data is proposed. It is based on the eigendecomposition of an estimation of the Koopman operator of dynamics. The method is demonstrated with standard benchmark datasets. ### Strengths - The idea of using Koopman operator's spectrum for static/d...
This paper studies interpretability for image classification. The authors propose a schema-based architecture (SchemaNet) based on vision transformers, which first extract features from pre-trained backbone and form a so-called ingredient relation graph (IR-Graph) with a feature to graph (Feat2Graph) module. The IR-Gra...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies interpretability for image classification. The authors propose a schema-based architecture (SchemaNet) based on vision transformers, which first extract features from pre-trained backbone and form a so-called ingredient relation graph (IR-Graph) with a feature to graph (Feat2Graph) module. Th...
The paper gives a theoretical analysis of a simple (three layer) vision transformer network. The analysis is focused on training a ViT on a simple distribution of images constructed from label-relevant and label-irrelevant patches. Namely, there is a dictionary of M patches, and each patch in the input image is a noisy...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper gives a theoretical analysis of a simple (three layer) vision transformer network. The analysis is focused on training a ViT on a simple distribution of images constructed from label-relevant and label-irrelevant patches. Namely, there is a dictionary of M patches, and each patch in the input image is...
This paper identifies causally interacting features of high dimensional temporal/spatial data by considering the sparsity of underlying causal mechanisms instead of link sparsity, which can select the critical corresponding features for downstream causal discovery. Theoretical studies are provided, and empirical evalua...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper identifies causally interacting features of high dimensional temporal/spatial data by considering the sparsity of underlying causal mechanisms instead of link sparsity, which can select the critical corresponding features for downstream causal discovery. Theoretical studies are provided, and empirica...
This paper characterizes the identifiability of the label noise transition matrix for the generic case at the instance level based on Kruskal’s identifiability results. The main contribution is that it finds three separate independent and identically distributed (i.i.d.) noisy labels (random variables) are both necessa...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper characterizes the identifiability of the label noise transition matrix for the generic case at the instance level based on Kruskal’s identifiability results. The main contribution is that it finds three separate independent and identically distributed (i.i.d.) noisy labels (random variables) are both...