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 combines the machine learning technique within the traditional column generation method for solving combinatorial optimization problems. Specifically, the authors propose to use machine learning algorithms to predict high quality columns that belong to an optimal integer solution. Here the machine learning m...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper combines the machine learning technique within the traditional column generation method for solving combinatorial optimization problems. Specifically, the authors propose to use machine learning algorithms to predict high quality columns that belong to an optimal integer solution. Here the machine le...
This paper presented a new prompt learning approach for vision-language models, namely PLOT. The main idea of this paper is to learn multiple prompts per class and use optimal transport to match them with visual features of images. The whole optimization process contains inner and outer loops to optimize the matching a...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper presented a new prompt learning approach for vision-language models, namely PLOT. The main idea of this paper is to learn multiple prompts per class and use optimal transport to match them with visual features of images. The whole optimization process contains inner and outer loops to optimize the ma...
This article introduces wide graph neural networks (WGNNs). The structure of WGNNs is inspired by an analysis of oversmoothing in linear GNNs. Experiments are presented in which WGNNs achieve SOTA in several benchmark tasks, especially in the setting of heterphilic graphs. Strengths: In my view, the main strengths of ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This article introduces wide graph neural networks (WGNNs). The structure of WGNNs is inspired by an analysis of oversmoothing in linear GNNs. Experiments are presented in which WGNNs achieve SOTA in several benchmark tasks, especially in the setting of heterphilic graphs. Strengths: In my view, the main stren...
This paper focuses on the prediction of protein thermostability and the design of more thermostable proteins. After introducing a new benchmark to train and assess thermostability prediction models, authors describe a contrastive-learning framework to impart protein structure information on representations learned by l...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper focuses on the prediction of protein thermostability and the design of more thermostable proteins. After introducing a new benchmark to train and assess thermostability prediction models, authors describe a contrastive-learning framework to impart protein structure information on representations lear...
This paper studies two methods (modeling label noise and semi-supervised methods) for handling datasets with noisy labels, and the causal structure strongly influences which of the methods perform better. The paper additionally proposes a method for finding the causal structure. The paper studies an interesting problem...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper studies two methods (modeling label noise and semi-supervised methods) for handling datasets with noisy labels, and the causal structure strongly influences which of the methods perform better. The paper additionally proposes a method for finding the causal structure. The paper studies an interesting...
This paper provides an empirical study of representational harms in pre-trained language models. The authors consider safety scores derived from a Mann-Whitney U-test, and compute such safety scores across a range of different models and marginalized demographics, finding that PTLMs have a tendency to show representati...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper provides an empirical study of representational harms in pre-trained language models. The authors consider safety scores derived from a Mann-Whitney U-test, and compute such safety scores across a range of different models and marginalized demographics, finding that PTLMs have a tendency to show repr...
This paper studies the machine learning problems with OOD data (that is, out-of-distribution data). The paper models OOD data by introducing a new variable called spurious attribute, of which the conditional distribution is different across training and testing data. The paper proposes to use CSV (conditional spurious ...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the machine learning problems with OOD data (that is, out-of-distribution data). The paper models OOD data by introducing a new variable called spurious attribute, of which the conditional distribution is different across training and testing data. The paper proposes to use CSV (conditional s...
This paper proposes a method for continual learning in text-image modeling. They propose a formulation that uses a historical and main neural networks whose parameters are used interchangeably between the two networks during parameter optimization as a weighed average. In experiments, the authors compare favorably agai...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a method for continual learning in text-image modeling. They propose a formulation that uses a historical and main neural networks whose parameters are used interchangeably between the two networks during parameter optimization as a weighed average. In experiments, the authors compare favora...
Based on the finding of Non Robust Features (NRF) in the previous work (Ilyas et al., 2019), this paper tries to explain the phenomenon of Catastrophic Overfitting (CO) in FGSM-based adversarial training. It categorizes NRFs into different types and tried to design experiments to show that the two commonly observed ph...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Based on the finding of Non Robust Features (NRF) in the previous work (Ilyas et al., 2019), this paper tries to explain the phenomenon of Catastrophic Overfitting (CO) in FGSM-based adversarial training. It categorizes NRFs into different types and tried to design experiments to show that the two commonly obs...
The authors study approximation and learning of functions from Besov spaces by transformer networks. The results show that transformer networks can avoid curse of dimensionality. Almost minimax optimal rates are presented. Transformer networks are not well understood theoretically. This paper provides some interesting...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors study approximation and learning of functions from Besov spaces by transformer networks. The results show that transformer networks can avoid curse of dimensionality. Almost minimax optimal rates are presented. Transformer networks are not well understood theoretically. This paper provides some int...
The paper proposes approach that enables simple hyperparameter search in scenarios where either label information is delayed or when the training on full dataset is too costly. The paper proses to use an auxiliary neural network to predict missing labels and a method to search for probable dataset to select a subset o...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes approach that enables simple hyperparameter search in scenarios where either label information is delayed or when the training on full dataset is too costly. The paper proses to use an auxiliary neural network to predict missing labels and a method to search for probable dataset to select a ...
This work strives to learn a visual-textual embedding space from weakly label image-text pairs from the internet. In a typical contrastive learning with cosine-similarity, the noise in training data prevents the deep network to produce a solution with alignment-uniformity, and the network achieves a suboptimal system ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work strives to learn a visual-textual embedding space from weakly label image-text pairs from the internet. In a typical contrastive learning with cosine-similarity, the noise in training data prevents the deep network to produce a solution with alignment-uniformity, and the network achieves a suboptimal...
The paper proposes a method to improve indoor localization based on combination of RSSI and accelerometer data from waerable device. The localization data is used to produce features that serve as inputs for a classification model aiming to recognize when patients with Parkinson disease are off their medication. The an...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper proposes a method to improve indoor localization based on combination of RSSI and accelerometer data from waerable device. The localization data is used to produce features that serve as inputs for a classification model aiming to recognize when patients with Parkinson disease are off their medication...
The paper uses a mixture of encoder-decoder transformers for code understanding and generation tasks. It in addition uses multiple pretraining objectives to train the mixture. The following pretraining objectives are standard and have been explored in previous works -- in particular (1) span denoising, (2) causal langu...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper uses a mixture of encoder-decoder transformers for code understanding and generation tasks. It in addition uses multiple pretraining objectives to train the mixture. The following pretraining objectives are standard and have been explored in previous works -- in particular (1) span denoising, (2) caus...
The paper proposes a version of LISTA where the update parameters can be made specific to each data sample even when the sensing matrix is not the same among samples. The paper then also proposes a variational method to dictionary learning to be integrated with the former approach. The paper considers a novel setting ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a version of LISTA where the update parameters can be made specific to each data sample even when the sensing matrix is not the same among samples. The paper then also proposes a variational method to dictionary learning to be integrated with the former approach. The paper considers a novel ...
This paper proposes a new method for searching for optimal perturbation radii for adversarial training, which is shown to be more efficient than existing works. Strength - Well written. - Propose an efficient approach for searching for adaptive perturbation radii for each data point. It has better performance thant ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a new method for searching for optimal perturbation radii for adversarial training, which is shown to be more efficient than existing works. Strength - Well written. - Propose an efficient approach for searching for adaptive perturbation radii for each data point. It has better performanc...
This paper advocates a hybrid approach between inherent interpretability and post-hoc explainability for predictive models trained on image data. The aim is to blur the line between post hoc explanations of a Black-Box and constructing interpretable models, aiming to keep the high performance and flexibility of blackbo...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper advocates a hybrid approach between inherent interpretability and post-hoc explainability for predictive models trained on image data. The aim is to blur the line between post hoc explanations of a Black-Box and constructing interpretable models, aiming to keep the high performance and flexibility of...
The paper explores the reasoning capabilities of large language models (LLMs). They carry out a comprehensive evaluation of 46 reasoning tasks to show that language models perform fairly well on single step reasoning problems, but suffer at multi-step reasoning problems. To that end, the paper proposes a new algorithm ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper explores the reasoning capabilities of large language models (LLMs). They carry out a comprehensive evaluation of 46 reasoning tasks to show that language models perform fairly well on single step reasoning problems, but suffer at multi-step reasoning problems. To that end, the paper proposes a new al...
The paper studies the task of learning continuous dynamical systems from data using a hybrid (machine learning + PDE solvers) approach and compares the performance to purely data-driven methods. In particular, authors consider a regularly occurring in practice setting where only a partial knowledge of the governing equ...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper studies the task of learning continuous dynamical systems from data using a hybrid (machine learning + PDE solvers) approach and compares the performance to purely data-driven methods. In particular, authors consider a regularly occurring in practice setting where only a partial knowledge of the gover...
The authors try to prove that the performance of graph convolutions placed in different combinations among the layers of a neural network is mutually similar for all combinations of the placement. ## Strength The analysis is comprehensive. ## Weaknesses 1. The assumption of the analysis is over-simplified. 2. Some cla...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors try to prove that the performance of graph convolutions placed in different combinations among the layers of a neural network is mutually similar for all combinations of the placement. ## Strength The analysis is comprehensive. ## Weaknesses 1. The assumption of the analysis is over-simplified. 2. ...
This paper combines two rendering modules for synthesizing novel views and novel poses of human bodies from sparse multi-view images. The first one is a body representation based on neural radiance fields and SMPL, extracting pixel-to-image features from each view, and attaching features to SMPL vertices. In this way,...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper combines two rendering modules for synthesizing novel views and novel poses of human bodies from sparse multi-view images. The first one is a body representation based on neural radiance fields and SMPL, extracting pixel-to-image features from each view, and attaching features to SMPL vertices. In t...
* This paper explores the mechanism behind the generative adversarial imitation learning, which provide an important conclusion that instability is caused by deterministic policies, rather than GANs. * It provides some existing methods relieve exploding gradients, but at the expense of “non-confidence”, and ST-GAIL ha...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: * This paper explores the mechanism behind the generative adversarial imitation learning, which provide an important conclusion that instability is caused by deterministic policies, rather than GANs. * It provides some existing methods relieve exploding gradients, but at the expense of “non-confidence”, and ST...
The paper focuses on set-level self-supervised learning (SLSSL) and proposes a new method for this problem with the same name SLSSL. In order to eliminate the bad effects of corrupted labels, SLSSL augments a minibatch by corrupting labels and obtains a augmented model. Then SLSSL enforces the consistency between outpu...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper focuses on set-level self-supervised learning (SLSSL) and proposes a new method for this problem with the same name SLSSL. In order to eliminate the bad effects of corrupted labels, SLSSL augments a minibatch by corrupting labels and obtains a augmented model. Then SLSSL enforces the consistency betwe...
This paper modifies two aspects of the FID model (retrieval-augmented text generation) in section 3: (1) the authors truncate the passages to speed up the model (2) they modify the explainability component by using a ranking task. Results on KILT show a substantial improvement over the FID model. Strengths: simple mod...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper modifies two aspects of the FID model (retrieval-augmented text generation) in section 3: (1) the authors truncate the passages to speed up the model (2) they modify the explainability component by using a ranking task. Results on KILT show a substantial improvement over the FID model. Strengths: si...
This paper formulates a problem of decentralized online bandit optimization on directed graphs, in which the graph structure dictates the order of the players and how players observe the actions of other players. Moreover, all players can observe a joint bandit reward, which is a linear combination of the reward of eac...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper formulates a problem of decentralized online bandit optimization on directed graphs, in which the graph structure dictates the order of the players and how players observe the actions of other players. Moreover, all players can observe a joint bandit reward, which is a linear combination of the rewar...
This paper presents a novel method to generate 3D assets from text descriptions. The proposed method utilizes a pretrained diffusion model to generate a denoised 2D image from text and noisy input. The noisy input is rendered with a NerF model and known noise. The NerF model is trained by minimizing the noise residual ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper presents a novel method to generate 3D assets from text descriptions. The proposed method utilizes a pretrained diffusion model to generate a denoised 2D image from text and noisy input. The noisy input is rendered with a NerF model and known noise. The NerF model is trained by minimizing the noise r...
The paper describes multiple tricks to push the performance and of the self-supervised Barlow Twins step in a block-wise setting. Most prominently, it is reported than a block-wise training of a ResNet architecture with Barlow twins leads to a high top-1 accuracy of a down-stream classifier (70+% which is one percent a...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper describes multiple tricks to push the performance and of the self-supervised Barlow Twins step in a block-wise setting. Most prominently, it is reported than a block-wise training of a ResNet architecture with Barlow twins leads to a high top-1 accuracy of a down-stream classifier (70+% which is one p...
The paper considers achieving the zero knowledge fair decision tree learning algorithms to guarantee confidentiality both the model and training data. An extension to ensembles of trees is also proposed. Strength: 1. The idea of considering confidential proof of fairness for training is interesting. 2. They design an...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper considers achieving the zero knowledge fair decision tree learning algorithms to guarantee confidentiality both the model and training data. An extension to ensembles of trees is also proposed. Strength: 1. The idea of considering confidential proof of fairness for training is interesting. 2. They d...
Two unbiased stochastic proximal solvers for learning graph equilibrium models are proposed. They are inspired by the stochastic proximal gradient descent method and its variance reduction variant (called USP and USP-VR solver). Both provide considerable computational speed-ups in comparison with the original solvers. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Two unbiased stochastic proximal solvers for learning graph equilibrium models are proposed. They are inspired by the stochastic proximal gradient descent method and its variance reduction variant (called USP and USP-VR solver). Both provide considerable computational speed-ups in comparison with the original s...
The authors address the problem of federated learning in the scenario where clients have differently defined label sets (assuming a super set of labels and a subset with and without partial overlap), but without requiring cross-client data relabeling. Further they consider also the setting where only a small portion of...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors address the problem of federated learning in the scenario where clients have differently defined label sets (assuming a super set of labels and a subset with and without partial overlap), but without requiring cross-client data relabeling. Further they consider also the setting where only a small po...
This paper studies the edge of stability phenomenon. They propose to incorporate the cubic term in the convergence analysis and show that there is a self-stabilization property caused by this cubic term for general nonconvex optimization problems. Moreover, they show that GD is inherently regularizing the sharpness, i....	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper studies the edge of stability phenomenon. They propose to incorporate the cubic term in the convergence analysis and show that there is a self-stabilization property caused by this cubic term for general nonconvex optimization problems. Moreover, they show that GD is inherently regularizing the sharp...
This paper investigates bit-flip trojan attacks on text models with syntactic triggers. Several ideas are proposed for improving the effectiveness of these trojans. In experiments, this increases attack success rate (ASR) while having less of an effect on clean accuracy compared to baselines. Ablations confirm that mod...	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 investigates bit-flip trojan attacks on text models with syntactic triggers. Several ideas are proposed for improving the effectiveness of these trojans. In experiments, this increases attack success rate (ASR) while having less of an effect on clean accuracy compared to baselines. Ablations confirm ...
This paper tries to fix bias in VideoQA models through causally approaches. The authors noticed that VideoQA models tend to just learn the dataset statistics ("How many ..." --> "2") and look to rectify that by forcing the model to answer questions that don't make sense for the input in order to disrupt the spurious co...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper tries to fix bias in VideoQA models through causally approaches. The authors noticed that VideoQA models tend to just learn the dataset statistics ("How many ..." --> "2") and look to rectify that by forcing the model to answer questions that don't make sense for the input in order to disrupt the spu...
This paper studies offline policy interval evaluation for discounted MDPs where the goal is to generate an interval that contains the ground truth with high probability. Specifically, this paper shows that, without the coverage and realizability, there exists a lower bound of the asymptotic bias. Then the paper propose...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies offline policy interval evaluation for discounted MDPs where the goal is to generate an interval that contains the ground truth with high probability. Specifically, this paper shows that, without the coverage and realizability, there exists a lower bound of the asymptotic bias. Then the paper...
This paper argues the existing drawback of OOD learning algorithms is because of lacking sufficient domains. The authors provide a hardness result that all learning algorithms require at least $poly (1/\epsilon)$ number of training domains to achieve an $\epsilon$ excess error. ### Strength - The authors provide theor...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper argues the existing drawback of OOD learning algorithms is because of lacking sufficient domains. The authors provide a hardness result that all learning algorithms require at least $poly (1/\epsilon)$ number of training domains to achieve an $\epsilon$ excess error. ### Strength - The authors provi...
This paper proposes a local model for boundary detection. At the core of the method, it assumes a local Gaussian Markov model in the feature space, where the features can be jointly learnt or derived from a neural network. The model uses a contrastive learning scheme to optimize features and pixel location connectivity...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper proposes a local model for boundary detection. At the core of the method, it assumes a local Gaussian Markov model in the feature space, where the features can be jointly learnt or derived from a neural network. The model uses a contrastive learning scheme to optimize features and pixel location conn...
This paper studies the task of detecting a planted community structure with small clusters (e.g. a planted clique) inside a graph with heterogeneous degrees. More precisely, the authors consider a random graph model such that $$ P(i \sim j) = \theta_i \theta_j P_{\pi(i), \pi(j)} $$ where $\theta$ models the degree het...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the task of detecting a planted community structure with small clusters (e.g. a planted clique) inside a graph with heterogeneous degrees. More precisely, the authors consider a random graph model such that $$ P(i \sim j) = \theta_i \theta_j P_{\pi(i), \pi(j)} $$ where $\theta$ models the de...
This paper proposes to improve diffusion probabilistic models by considering multivariate diffusions. The motivation inherits previous probabilistic modeling methods that use auxiliary variables. By augmenting the diffusion space, it enables mixing between data dimensions and auxiliary dimensions, leading to better-ali...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes to improve diffusion probabilistic models by considering multivariate diffusions. The motivation inherits previous probabilistic modeling methods that use auxiliary variables. By augmenting the diffusion space, it enables mixing between data dimensions and auxiliary dimensions, leading to be...
The submission manifests several tricks in order to potentially improve the original MAE's recipe. The tricks include: 1) addition dropout in the attention layers; 2) a study on the normalization targets; and 3) intermediate mask tokens. The results are slightly improved with the authors' own implementation. Strengths:...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The submission manifests several tricks in order to potentially improve the original MAE's recipe. The tricks include: 1) addition dropout in the attention layers; 2) a study on the normalization targets; and 3) intermediate mask tokens. The results are slightly improved with the authors' own implementation. St...
The authors propose learning embedding in augmentation subspaces via masking, which specifies which dimensions are relevant for which augmentation. In the loss function, the mask is applied to the embeddings before comparison. Also, the loss function includes a L1 norm term that encourages masks to be sparse (with f...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors propose learning embedding in augmentation subspaces via masking, which specifies which dimensions are relevant for which augmentation. In the loss function, the mask is applied to the embeddings before comparison. Also, the loss function includes a L1 norm term that encourages masks to be sparse...