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
Most of the recent work on language models for solving math word problems relies on prompting engineering. Instead, this paper studies how to finetune the language models for solving math word problems with outcome-based and process-based approaches. The paper develops two metrics: race error rate and final-answer err...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: Most of the recent work on language models for solving math word problems relies on prompting engineering. Instead, this paper studies how to finetune the language models for solving math word problems with outcome-based and process-based approaches. The paper develops two metrics: race error rate and final-an...
This paper considers the HiP-MDPs where a set of low-dimensional hidden parameters models the variations in tasks, this setting is naturally applicable to settings with transfers. The main contribution of this paper is the theoretical analysis on the regret upper bounds of model and policy transfer algorithms, as well ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper considers the HiP-MDPs where a set of low-dimensional hidden parameters models the variations in tasks, this setting is naturally applicable to settings with transfers. The main contribution of this paper is the theoretical analysis on the regret upper bounds of model and policy transfer algorithms, ...
The authors propose a technique based on topological data analysis to: 1. estimate pointwise the intrinsic dimensionality of the data manifold in the local neighbourhood, and 2. assess if the local neighbourhood resembles a Euclidean space with the same dimensionality. In this way, singularities on the data manifold ca...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors propose a technique based on topological data analysis to: 1. estimate pointwise the intrinsic dimensionality of the data manifold in the local neighbourhood, and 2. assess if the local neighbourhood resembles a Euclidean space with the same dimensionality. In this way, singularities on the data man...
The paper analyzes parallel neural networks for nonparametric regression (including approximation error and regression error analysis). Suppose the regression function has nonhomogeneous regularities, e.g., Besov functions or functions of bounded variation. Parallel neural networks trained with weight decay achieves ne...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper analyzes parallel neural networks for nonparametric regression (including approximation error and regression error analysis). Suppose the regression function has nonhomogeneous regularities, e.g., Besov functions or functions of bounded variation. Parallel neural networks trained with weight decay ach...
This paper first formally analyze the vulnerability of optimal policies trained by safe RL algorithms under observational disturbance. By defining state-adversarial safe RL, this paper theoretically shows these optimal policies are vulnerable under observational adversarial attacks. Based on these analyses, this paper ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper first formally analyze the vulnerability of optimal policies trained by safe RL algorithms under observational disturbance. By defining state-adversarial safe RL, this paper theoretically shows these optimal policies are vulnerable under observational adversarial attacks. Based on these analyses, thi...
This paper proposes to ensemble multiple pre-trained models with different architecture by estimating and thresholding p-value for out-of-distribution detection. Experimental results show the effectiveness of the proposed method. Strengths + Theoretical analysis on the p-value looks interesting. + The proposed metho...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to ensemble multiple pre-trained models with different architecture by estimating and thresholding p-value for out-of-distribution detection. Experimental results show the effectiveness of the proposed method. Strengths + Theoretical analysis on the p-value looks interesting. + The propos...
The paper presents density sketches, which is a cheap and practical way to reduce data in a streaming setting., DS can keep a succint representation of data and can sample unseen data on the fly from this succint represesentation. This is done by reducing KDE into a problem related to histograms. The integrated mean sq...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper presents density sketches, which is a cheap and practical way to reduce data in a streaming setting., DS can keep a succint representation of data and can sample unseen data on the fly from this succint represesentation. This is done by reducing KDE into a problem related to histograms. The integrated...
This paper proposes semantic uncertainty, a new uncertainty estimation metric that operates on the meaning space of natural language sentences. Semantic uncertainty accounts for the invariance of the meaning of sentences against surface syntactic or linguistic styles. Formally, semantic uncertainty of an input context ...	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 proposes semantic uncertainty, a new uncertainty estimation metric that operates on the meaning space of natural language sentences. Semantic uncertainty accounts for the invariance of the meaning of sentences against surface syntactic or linguistic styles. Formally, semantic uncertainty of an input ...
The paper investigates grokking, i.e., the abrupt phase change in the performance of transformers when trained on simple algorithmic tasks, through the lens of mechanistic interpretability. It uses the insights gained to propose new progress measures that vary smoothly throughout training and thus provide new insights ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper investigates grokking, i.e., the abrupt phase change in the performance of transformers when trained on simple algorithmic tasks, through the lens of mechanistic interpretability. It uses the insights gained to propose new progress measures that vary smoothly throughout training and thus provide new i...
This paper proposes a method to make prompt proposals from repository-level context and add concatenate the prompt proposals with normal context to LLM in order to feed more related information to achieve better generations. The paper explores different sources of repository level prompts, as well as prompt context typ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a method to make prompt proposals from repository-level context and add concatenate the prompt proposals with normal context to LLM in order to feed more related information to achieve better generations. The paper explores different sources of repository level prompts, as well as prompt con...
This paper focuses on the out-of-distribution detection task and proposes a new method called CIDER, which trains the model by optimizing a dispersion loss and a compactness loss. The compactness loss encourages samples to be close to their class prototypes while the dispersion loss encourages large angular distance am...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper focuses on the out-of-distribution detection task and proposes a new method called CIDER, which trains the model by optimizing a dispersion loss and a compactness loss. The compactness loss encourages samples to be close to their class prototypes while the dispersion loss encourages large angular dis...
The paper provides a procedure for building better surrogate networks for transformer models fine-tuned to a specific downstream task that allows for faster MPC for private sharing with clients while preserving the model's performance (defined by model accuracy). To build a better surrogate, the authors propose a two-s...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper provides a procedure for building better surrogate networks for transformer models fine-tuned to a specific downstream task that allows for faster MPC for private sharing with clients while preserving the model's performance (defined by model accuracy). To build a better surrogate, the authors propose...
This paper aims to tackle the partially observable few-shot learning problem where useful features are only contained in some of the views of a data instance. It develops a product of experts model to map different views to the representations with different uncertainty, which can capture the importance of different fe...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper aims to tackle the partially observable few-shot learning problem where useful features are only contained in some of the views of a data instance. It develops a product of experts model to map different views to the representations with different uncertainty, which can capture the importance of diff...
This paper introduces a method for extracting compact and interpretable vector maps from input camera and lidar data. The method consists of three key components: a feature extractor, a map element predictor and a polyline generator. The feature extractor uses a CNN to extract features from the images and the point clo...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a method for extracting compact and interpretable vector maps from input camera and lidar data. The method consists of three key components: a feature extractor, a map element predictor and a polyline generator. The feature extractor uses a CNN to extract features from the images and the p...
The paper focuses on the empirical study of zero-shot RL, whose goal is to learn a set of policies and/or representations during training so that they can be adapted to solve unseen tasks in a zero-shot manner. Specifically, the authors focus on the successor features (SFs) framework and forward-backward (FB) represent...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper focuses on the empirical study of zero-shot RL, whose goal is to learn a set of policies and/or representations during training so that they can be adapted to solve unseen tasks in a zero-shot manner. Specifically, the authors focus on the successor features (SFs) framework and forward-backward (FB) r...
In this paper, the authors introduce a methodology for pareto front sub-network identification from a supernetwork. This is done in a two step process, by first conducting fair subsampling of subnetworks for Nf epochs, followed by Pareto-Rank training for Np epochs. Sub-networks that are furthest from the pareto front...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: In this paper, the authors introduce a methodology for pareto front sub-network identification from a supernetwork. This is done in a two step process, by first conducting fair subsampling of subnetworks for Nf epochs, followed by Pareto-Rank training for Np epochs. Sub-networks that are furthest from the pare...
The authors propose a novel adversarial attack strategy for discrete-time graph neural networks (GNNs). The objective is to perturb the sequence of graphs to decrease accuracy on prediction tasks such as dynamic link prediction and node classification. To preserve temporal dynamics of the graph to avoid detection, the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors propose a novel adversarial attack strategy for discrete-time graph neural networks (GNNs). The objective is to perturb the sequence of graphs to decrease accuracy on prediction tasks such as dynamic link prediction and node classification. To preserve temporal dynamics of the graph to avoid detecti...
The paper presents a novel representation learning method inspired by energy-based generative models, which can be combined with different restoration tasks as self-supervised surrogate tasks. Specifically, a neural network-based energy function is learned which assigns a high energy to corrupted images, and a low ener...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper presents a novel representation learning method inspired by energy-based generative models, which can be combined with different restoration tasks as self-supervised surrogate tasks. Specifically, a neural network-based energy function is learned which assigns a high energy to corrupted images, and a ...
Inspired by Regularized Lottery Ticket Hypothesis, which hypothesizes that smooth subnetworks exist within a dense network, this paper propose Soft-SubNetworks (SoftNet), an incremental learning strategy that preserves the learned class knowledge and learns the newer ones. The SoftNet jointly learns the model weights a...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Inspired by Regularized Lottery Ticket Hypothesis, which hypothesizes that smooth subnetworks exist within a dense network, this paper propose Soft-SubNetworks (SoftNet), an incremental learning strategy that preserves the learned class knowledge and learns the newer ones. The SoftNet jointly learns the model w...
This paper introduces a sequential node matching scheme for graph matching, via deep reinforcement learning. The proposed scheme differs from the majority of existing works that obtain the whole matching in one shot. The main effectiveness of the proposed method seems to lie in handling outliers. Experiments on both...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a sequential node matching scheme for graph matching, via deep reinforcement learning. The proposed scheme differs from the majority of existing works that obtain the whole matching in one shot. The main effectiveness of the proposed method seems to lie in handling outliers. Experiments...
The paper studies the problem of estimating individual treatment effects (ITE) under unobserved confounding using binary instruments. The paper starts with discussing conditions under which the ITE is identifiable. It then proceeds with proposing a two-stage meta learner (MRIV) that estimates the ITE. In the first stag...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper studies the problem of estimating individual treatment effects (ITE) under unobserved confounding using binary instruments. The paper starts with discussing conditions under which the ITE is identifiable. It then proceeds with proposing a two-stage meta learner (MRIV) that estimates the ITE. In the fi...
The authors proposed a three-player GAN-based model called PGASL for class-inbalanced data. The model was tested against several existing models for class-imbalanced data using two text datasets and showed superior performance. Strength - The proposed strategy seems straightforward. Weaknesses - Why was the model t...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The authors proposed a three-player GAN-based model called PGASL for class-inbalanced data. The model was tested against several existing models for class-imbalanced data using two text datasets and showed superior performance. Strength - The proposed strategy seems straightforward. Weaknesses - Why was the...
This paper considers the problem of representation learning for bag-of-words count data, trying to understand the usefulness of recent self-supervised learning, including reconstruction objectives (Eq 1) and contrastive objectives (Eq 2). The key question is: why might these approaches be better than previous ones? The...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers the problem of representation learning for bag-of-words count data, trying to understand the usefulness of recent self-supervised learning, including reconstruction objectives (Eq 1) and contrastive objectives (Eq 2). The key question is: why might these approaches be better than previous o...
The paper proposes a modification of TRADES, one of the most popular algorithms to obtain adversarially robust classifiers, to improve its performance: in particular, the regularization term to achieve robustness is weighted to penalize more the training examples which are less robust. In the experimental evaluation on...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes a modification of TRADES, one of the most popular algorithms to obtain adversarially robust classifiers, to improve its performance: in particular, the regularization term to achieve robustness is weighted to penalize more the training examples which are less robust. In the experimental evalu...
This article presents a new loss (KFIoU) approximating the Skew Intersection over Union Loss (SkewIoU) for the rotated object detection problem. This loss is composed of a scale-insensitive center point loss, and a second distance-insensitive term using Gaussian modeling and Kalman filtering. Compared to other Gaussian...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This article presents a new loss (KFIoU) approximating the Skew Intersection over Union Loss (SkewIoU) for the rotated object detection problem. This loss is composed of a scale-insensitive center point loss, and a second distance-insensitive term using Gaussian modeling and Kalman filtering. Compared to other ...
The paper proposes a progressive data dropout (PDD) framework that gradually discards the majority of the samples for already-accurate classes, in order to speedup the training. There are other design choices including warmup, residue and swapout. Experiments on MNIST, CIFAR, SVHN and ImageNet are performed. Strengths:...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper proposes a progressive data dropout (PDD) framework that gradually discards the majority of the samples for already-accurate classes, in order to speedup the training. There are other design choices including warmup, residue and swapout. Experiments on MNIST, CIFAR, SVHN and ImageNet are performed. St...
The submission presents an algorithm for learning decision trees with logistic regression models at the leaf nodes. It initially considers soft splits and estimation of the model using expectation maximization before proceeding to more interpretable hard splits and a stochastic expectation maximization algorithm for le...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The submission presents an algorithm for learning decision trees with logistic regression models at the leaf nodes. It initially considers soft splits and estimation of the model using expectation maximization before proceeding to more interpretable hard splits and a stochastic expectation maximization algorith...
This paper studies offline reinforcement learning with heterogeneous data sources. The paper propose a new algorithm applying pessimism to handle the randomness from both the sample and the source. It proves the sample efficiency of such an algorithm. The theoretical analysis is based on tabular and linear MDPs. Streng...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies offline reinforcement learning with heterogeneous data sources. The paper propose a new algorithm applying pessimism to handle the randomness from both the sample and the source. It proves the sample efficiency of such an algorithm. The theoretical analysis is based on tabular and linear MDPs...
This work presents a novel method for the co-design of 1D sequence and 3D structures of antibody proteins that can bind to the target antigen. The method proposes to include the target antigen and the light chain of the antibody as additional conditions, and designs a novel E(3)-equivariant network to predict the amino...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work presents a novel method for the co-design of 1D sequence and 3D structures of antibody proteins that can bind to the target antigen. The method proposes to include the target antigen and the light chain of the antibody as additional conditions, and designs a novel E(3)-equivariant network to predict t...
This paper proposed a replay buffer strategy to help model-based RL models effectively adapt to local changes in the environment. When local changes happen, the traditional first-in-first-out (FIFO) data buffer will interfere with the model training due to the out-of-data samples still being stored in the buffer. To fi...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposed a replay buffer strategy to help model-based RL models effectively adapt to local changes in the environment. When local changes happen, the traditional first-in-first-out (FIFO) data buffer will interfere with the model training due to the out-of-data samples still being stored in the buffe...
This paper is in the realm of generalizable reward learning and the aim is to learn value-discriminative visual representation from a diverse set of human manipulation demonstrations by contrasting conditioned on goals. The learned representation is used for constructing reward as similarity in embedded space between a...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper is in the realm of generalizable reward learning and the aim is to learn value-discriminative visual representation from a diverse set of human manipulation demonstrations by contrasting conditioned on goals. The learned representation is used for constructing reward as similarity in embedded space b...
## Overview In this paper, PIXEL, a pixel-based encoder of language is introduced to circumvent the vocabulary bottleneck problem. PIXEL is built on ViT-MAE and is pretrained on the same corpus as BERT. Specifically, the authors proposed to render text onto blank images and patchify raw pixels into image tokens and mas...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: ## Overview In this paper, PIXEL, a pixel-based encoder of language is introduced to circumvent the vocabulary bottleneck problem. PIXEL is built on ViT-MAE and is pretrained on the same corpus as BERT. Specifically, the authors proposed to render text onto blank images and patchify raw pixels into image tokens...
The authors propose using using hitting times, commute times and resistive embeddings as features in a GNN. They show that using using this makes the network more powerful than 1-WL. They also show that this can be approximated by sketching techniques in large graphs. Then they show that empirically these networks do b...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose using using hitting times, commute times and resistive embeddings as features in a GNN. They show that using using this makes the network more powerful than 1-WL. They also show that this can be approximated by sketching techniques in large graphs. Then they show that empirically these netwo...
This paper studies where (weak) prior learning can and can’t work in unsupervised inverse problems, by considering dictionary learning in a conventional model as well as convolutional dictionary learning. Inpainting and deblurring asks are addressed in this work. Weakness: this paper is roughly an extension of the wor...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies where (weak) prior learning can and can’t work in unsupervised inverse problems, by considering dictionary learning in a conventional model as well as convolutional dictionary learning. Inpainting and deblurring asks are addressed in this work. Weakness: this paper is roughly an extension of...
The authors consider the definition of unlearning developed in previous work. The paper claims that these definitions do not satisfy the true meaning of unlearning as data could be added back in at later points in time. To this end, the propose that the new definition should be stateless and algorithms for deletion and...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors consider the definition of unlearning developed in previous work. The paper claims that these definitions do not satisfy the true meaning of unlearning as data could be added back in at later points in time. To this end, the propose that the new definition should be stateless and algorithms for dele...
An important research question in neural computing is the interplay between Batch Normalization (BN) and learning. In this vein, this paper studies the influence of BN on gradient norm across the layers. BN causes an exponential growth with depth for the norm of a gradient in the first layer. When studying this phenom...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: An important research question in neural computing is the interplay between Batch Normalization (BN) and learning. In this vein, this paper studies the influence of BN on gradient norm across the layers. BN causes an exponential growth with depth for the norm of a gradient in the first layer. When studying thi...
This work introduced a new setting of few-shot supervised mulit-source domain adaptation with a few labeled target samples and labeled multiple source domain data. Under this new setting, they proposed a progressive mix-up method for few-shot supervised multi-source domain transfer. Specifically, it creates an intermed...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work introduced a new setting of few-shot supervised mulit-source domain adaptation with a few labeled target samples and labeled multiple source domain data. Under this new setting, they proposed a progressive mix-up method for few-shot supervised multi-source domain transfer. Specifically, it creates an ...
This paper proposes to learn a new exploration technique to learn a symbolic model of an environment. This model is then used to speed up goal directed exploration. The proposed technique is then evaluated against Q-learning with epsilon-greedy and R-max as well as a hierarchical RL method. Experiments are done on bloc...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes to learn a new exploration technique to learn a symbolic model of an environment. This model is then used to speed up goal directed exploration. The proposed technique is then evaluated against Q-learning with epsilon-greedy and R-max as well as a hierarchical RL method. Experiments are done...
This paper proposes to find salient features of images of the target class in the encoded latent space with an auto-encoder, perturb the source image in its encoded representations with a learned mask, and decode the representation to form poisoned data for dirty-label data poisoning. Mask learning is performed with a ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes to find salient features of images of the target class in the encoded latent space with an auto-encoder, perturb the source image in its encoded representations with a learned mask, and decode the representation to form poisoned data for dirty-label data poisoning. Mask learning is performed...
This paper proposed an automated data augmentation method for fair graph representation learning. This seems to be the first automated graph data augmentation method that focuses on learning fair representations on graph data. The authors designed the Graphair framework with four components: edge augmentation, feature ...	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 proposed an automated data augmentation method for fair graph representation learning. This seems to be the first automated graph data augmentation method that focuses on learning fair representations on graph data. The authors designed the Graphair framework with four components: edge augmentation, ...