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 presents an early attempt at learning geometric lengths from trails and explores how to transfer the learned geometric knowledge to solve a different task. Besides offering a problem formulation, this paper presents an evaluation framework based on finding evaluation correlations against ground truth lengths...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents an early attempt at learning geometric lengths from trails and explores how to transfer the learned geometric knowledge to solve a different task. Besides offering a problem formulation, this paper presents an evaluation framework based on finding evaluation correlations against ground truth...
This paper proposes to use sinusoidal embeddings to represent m/z and intensity values when modeling metabolomics mass spectrometry data using deep learning models. The idea is sensible though not particularly novel. Such embeddings have been used previously for mass spectrometry data (as referenced in the related wo...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes to use sinusoidal embeddings to represent m/z and intensity values when modeling metabolomics mass spectrometry data using deep learning models. The idea is sensible though not particularly novel. Such embeddings have been used previously for mass spectrometry data (as referenced in the re...
The paper studies how to learn models that satisfy various fairness desiderata (e.g., demographic parity and equalized odds) while satisfying differential privacy. Like prior work, it writes down a loss function encoding some weighted combination of accuracy and fairness constraints and aims to develop a private method...	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 studies how to learn models that satisfy various fairness desiderata (e.g., demographic parity and equalized odds) while satisfying differential privacy. Like prior work, it writes down a loss function encoding some weighted combination of accuracy and fairness constraints and aims to develop a privat...
This paper proposes a novel conformal prediction method for time-series data, called Copula Conformal Prediction for multi-step time series forecasting (CopulaCPTS), for uncertainty quantification. In particular, the paper uses the concept of “copula” from statistics; the copula is a joint CDF on data, which fully capt...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes a novel conformal prediction method for time-series data, called Copula Conformal Prediction for multi-step time series forecasting (CopulaCPTS), for uncertainty quantification. In particular, the paper uses the concept of “copula” from statistics; the copula is a joint CDF on data, which fu...
This paper shows a few shot GAN by embed latent factor into an existing GAN to transfer to a new style. This paper tries to mix pre-trained GAN with few shot learning. Their contribution can be concluded as: -propose a simple procedure to utilize a GAN trained on large-scale source-data to generate samples from a ta...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper shows a few shot GAN by embed latent factor into an existing GAN to transfer to a new style. This paper tries to mix pre-trained GAN with few shot learning. Their contribution can be concluded as: -propose a simple procedure to utilize a GAN trained on large-scale source-data to generate samples f...
This paper proposes a method which combines the benefits of iterative optimization and a learned encoder-decoder for the task of image steganography. Similar to the style of LISTA, the paper learns individual optimization blocks in an E2E fashion while applying the blocks iteratively at inference time. Experiment resul...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a method which combines the benefits of iterative optimization and a learned encoder-decoder for the task of image steganography. Similar to the style of LISTA, the paper learns individual optimization blocks in an E2E fashion while applying the blocks iteratively at inference time. Experime...
The authors propose a new DST algorithm that utilizes only sampled connections to reduce peak memory usage while previous DST algorithms rely on computing gradients of all connections. For some selected models, the proposed method can obtain similar model accuracy using only 4% as a size of sampled connections compared...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors propose a new DST algorithm that utilizes only sampled connections to reduce peak memory usage while previous DST algorithms rely on computing gradients of all connections. For some selected models, the proposed method can obtain similar model accuracy using only 4% as a size of sampled connections ...
The aim of this paper is the generation of molecules with desired chemical and biological properties learning the joint distribution of molecules and properties . In order to achieve this the authors propose an energy based generative model that augments a top down generative model (conditional autoregressive model) wi...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The aim of this paper is the generation of molecules with desired chemical and biological properties learning the joint distribution of molecules and properties . In order to achieve this the authors propose an energy based generative model that augments a top down generative model (conditional autoregressive m...
The paper aims to learn better visual representation by leveraging both contrastive learning (CL) and masked image modeling (MIM). The authors claim that simultaneously performing CL and MIM pertaining will lead to worse results. To mitigate this issue, the paper proposes Layer Grafted Pre-training, which finetunes a p...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper aims to learn better visual representation by leveraging both contrastive learning (CL) and masked image modeling (MIM). The authors claim that simultaneously performing CL and MIM pertaining will lead to worse results. To mitigate this issue, the paper proposes Layer Grafted Pre-training, which finet...
Fourier Neural Operator (Li 2020) has become a popular tool in solving PDEs numerically. The main contribution of the paper is to replace the classic Fourier Layer in FNO with the quantum Fourier layer. The authors argue that the proposed algorithm is provably substantially faster (using quantum hardware) than the clas...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Fourier Neural Operator (Li 2020) has become a popular tool in solving PDEs numerically. The main contribution of the paper is to replace the classic Fourier Layer in FNO with the quantum Fourier layer. The authors argue that the proposed algorithm is provably substantially faster (using quantum hardware) than ...
This paper investigated the possibility of decomposing the trust region surrogate over joint policies into the respective surrogates for each individual agent's policies for cooperative multi-agent reinforcement learning. The reported experiment results show that the newly developed DPO algorithm based on the decompose...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper investigated the possibility of decomposing the trust region surrogate over joint policies into the respective surrogates for each individual agent's policies for cooperative multi-agent reinforcement learning. The reported experiment results show that the newly developed DPO algorithm based on the d...
The paper proposes a diversified query strategy for an anomaly detector that can be applied in an active learning setup. The querying strategy is based on querying diverse cluster centers seeded by k-means++. 1. Abstract: "However, correctly identifying anomalies requires an estimate of the fraction of anomalies in the...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper proposes a diversified query strategy for an anomaly detector that can be applied in an active learning setup. The querying strategy is based on querying diverse cluster centers seeded by k-means++. 1. Abstract: "However, correctly identifying anomalies requires an estimate of the fraction of anomalie...
The paper proposes an approach to estimate feature importance. Multiple shallow mask models are trained with different random seeds. These resulting weights and then averaged to produce a global model that can be inferior in accuracy but better identify the global feature importance. I found the proposed approach inte...	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 an approach to estimate feature importance. Multiple shallow mask models are trained with different random seeds. These resulting weights and then averaged to produce a global model that can be inferior in accuracy but better identify the global feature importance. I found the proposed appro...
The authors introduce a method to learn Lagrangian dynamics from images. In line with previous work, the authors simplify the learning problem by: (1) representing the Lagrangian as the difference in kinetic and potential energy (both learned), and (2) using Cartesian coordinates to represent the positions of the syste...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors introduce a method to learn Lagrangian dynamics from images. In line with previous work, the authors simplify the learning problem by: (1) representing the Lagrangian as the difference in kinetic and potential energy (both learned), and (2) using Cartesian coordinates to represent the positions of t...
The paper tackles the problem experiments reproducibility with specific reference to inferential reproducibility which proposes to interpret the variation of performance values as due to the following factors; data characteristics, meta-parameter settings, includng also their interactions. The paper starts from the rat...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper tackles the problem experiments reproducibility with specific reference to inferential reproducibility which proposes to interpret the variation of performance values as due to the following factors; data characteristics, meta-parameter settings, includng also their interactions. The paper starts from...
The paper studies the the phenomenon of "Edge-of-Statility" (EOS) training dynamics and analyses the gradient dynamics on a simple example (the example is simple but the analysis is quite complicated). The EOS of training dynamics, initially observed by [Cohen et al. 2021], refers the phenomenon that when running grad...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper studies the the phenomenon of "Edge-of-Statility" (EOS) training dynamics and analyses the gradient dynamics on a simple example (the example is simple but the analysis is quite complicated). The EOS of training dynamics, initially observed by [Cohen et al. 2021], refers the phenomenon that when runn...
To tackle the problem of joint luminance enhancement and denoising in low-light imaging, the author proposed 1) a new network architecture, UHDFour, that utilizes authors' observations on noise and signal luminance's relation in the amplitude and phase images in the frequency domain, and 2) a new dataset, UHD-LL, that ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: To tackle the problem of joint luminance enhancement and denoising in low-light imaging, the author proposed 1) a new network architecture, UHDFour, that utilizes authors' observations on noise and signal luminance's relation in the amplitude and phase images in the frequency domain, and 2) a new dataset, UHD-L...
The paper proposes to use a planner solving a locally quadratic program to improve the action selection during the rollouts. The better action selection should improve the learning performance and enable the agent to achieve a higher reward with fewer samples. The experiments show this improved sample efficiency and th...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes to use a planner solving a locally quadratic program to improve the action selection during the rollouts. The better action selection should improve the learning performance and enable the agent to achieve a higher reward with fewer samples. The experiments show this improved sample efficienc...
This paper explores extending the idea of fast weight with outer-product based to human-interpretable domains beyond network weight space, and examines the idea with the example of natural image generation. The slow-weight network, now termed Fast Weight Painter now progressively refines the pixel values of a 3D image ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper explores extending the idea of fast weight with outer-product based to human-interpretable domains beyond network weight space, and examines the idea with the example of natural image generation. The slow-weight network, now termed Fast Weight Painter now progressively refines the pixel values of a 3...
The submission provides an extensive empirical study on supervised (SL), unsupervised (autoencoder, AE), and self-supervised (SSL) representation learning algorithms in a distribution shift scenario. The key contribution is comparing these algorithms when the representation is learned on the source task while the class...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The submission provides an extensive empirical study on supervised (SL), unsupervised (autoencoder, AE), and self-supervised (SSL) representation learning algorithms in a distribution shift scenario. The key contribution is comparing these algorithms when the representation is learned on the source task while t...
In this paper, the authors improve multi-view representation learning by splitting documents and aligning viewer tokens along with different parts. The experiments conducted on three datasets demonstrate that the proposed CAMVR performs better than MVR. The analysis also shows that CAMVR has better interpretability. S...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this paper, the authors improve multi-view representation learning by splitting documents and aligning viewer tokens along with different parts. The experiments conducted on three datasets demonstrate that the proposed CAMVR performs better than MVR. The analysis also shows that CAMVR has better interpretabi...
This paper proposes an end-to-end color quantization transformer that is able to discover the color naming system under the need of perception and machine. The framework contains two branches, respectively for structure mining and color localization. The proposed CQFormer quantizes the color space of the input image an...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes an end-to-end color quantization transformer that is able to discover the color naming system under the need of perception and machine. The framework contains two branches, respectively for structure mining and color localization. The proposed CQFormer quantizes the color space of the input ...
This paper proposes group-wise verifiable coded computing with coding techniques and group-wise verification to tackle adversarial workers and the effects of stragglers. The authors demonstrate the performance of group-wise verifiable coded computing in term of overall processing time and verification time. (1)Accordin...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes group-wise verifiable coded computing with coding techniques and group-wise verification to tackle adversarial workers and the effects of stragglers. The authors demonstrate the performance of group-wise verifiable coded computing in term of overall processing time and verification time. (1)...
The paper presents generalization problems of the network in performing the low-level vision task, i.e. authors perform a study by performing the different deraining experiments and convey that when trained with a less complex background, networks learn the less complex element in the image content and degradation to r...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents generalization problems of the network in performing the low-level vision task, i.e. authors perform a study by performing the different deraining experiments and convey that when trained with a less complex background, networks learn the less complex element in the image content and degradat...
This paper focuses on learning a robust deepfake detector against adversarial attacks. The authors propose a method named Disjoint Deepfake Detection (D3). The method is motivated by the observation that redundancy exists in the frequency space, which means fewer signals are enough to make correct predictions. Based on...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper focuses on learning a robust deepfake detector against adversarial attacks. The authors propose a method named Disjoint Deepfake Detection (D3). The method is motivated by the observation that redundancy exists in the frequency space, which means fewer signals are enough to make correct predictions. ...
This paper presents a neural program-based method for an abstract reasoning (ARC from Chollet 2019) style task. The model contains an object-centric representation layer (based on slot attention I believe), a controller (encoding the examples and predicting the program), and an executor (which takes the program represe...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a neural program-based method for an abstract reasoning (ARC from Chollet 2019) style task. The model contains an object-centric representation layer (based on slot attention I believe), a controller (encoding the examples and predicting the program), and an executor (which takes the program...
This paper introduces a PLL approach named ML-PLL to address a more challenging setting–competitive label noise. It addresses the competitive noisy labels by following components: label correction, prediction network training with mutual learning, and transformation based label association learning, which makes couple ...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper introduces a PLL approach named ML-PLL to address a more challenging setting–competitive label noise. It addresses the competitive noisy labels by following components: label correction, prediction network training with mutual learning, and transformation based label association learning, which makes...
The paper proposes a latent variable model to capture concept-specific laws from images/videos. The laws are encoded as Latent Random Functions (specifically, Neural Processes), and can be composed with one another. Strengths: - Multiple datasets with different focus - Illustration of compositionalilty - Demonstration...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The paper proposes a latent variable model to capture concept-specific laws from images/videos. The laws are encoded as Latent Random Functions (specifically, Neural Processes), and can be composed with one another. Strengths: - Multiple datasets with different focus - Illustration of compositionalilty - Demon...
This paper addresses the problem of reward-free reinforcement learning under (possibly mismatching) cost constraints in both the exploration phase and the planning phase. Crucially, the paper shows how to exploit the knowledge of a safe baseline policy to achieve zero constraints violations while matching the sample co...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper addresses the problem of reward-free reinforcement learning under (possibly mismatching) cost constraints in both the exploration phase and the planning phase. Crucially, the paper shows how to exploit the knowledge of a safe baseline policy to achieve zero constraints violations while matching the s...
The paper propose a method for explaining intermediate activations of a neural network using concepts. In a nutshell, f(x) and g(x) are the networks that take activations x from an intermediate layer as input and respectively predict class and concept. The relevance of a concept c to a class y is measured using the pro...	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 propose a method for explaining intermediate activations of a neural network using concepts. In a nutshell, f(x) and g(x) are the networks that take activations x from an intermediate layer as input and respectively predict class and concept. The relevance of a concept c to a class y is measured using...
The paper proposed a new method for complex and large action space. The proposed method uses listwise RL for the retrieval task which is further used for action selection. Strengths: - The paper motivates the problem very well. The abstract and introduction are highly aligned with the problem targeted. - The paper us...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposed a new method for complex and large action space. The proposed method uses listwise RL for the retrieval task which is further used for action selection. Strengths: - The paper motivates the problem very well. The abstract and introduction are highly aligned with the problem targeted. - The ...
The authors aim to propose a framework for unsupervised domain adaptation of time series by doing a weird combination of existing models without good justification/motivation for what they introduce. Specifically, they combine the model introduced in “domain adversarial training of neural networks” by a type of contra...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors aim to propose a framework for unsupervised domain adaptation of time series by doing a weird combination of existing models without good justification/motivation for what they introduce. Specifically, they combine the model introduced in “domain adversarial training of neural networks” by a type o...
This paper focuses on refining the semantic prototypes in ZSL. The authors top a V2SM network to the generator, mapping visual features to a semantic-related space. On the other hand, they transfer the semantic embeddings by another VOPE module. The outputs of V2SM and VOPE are aligned by cosine similarity and l2 losse...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper focuses on refining the semantic prototypes in ZSL. The authors top a V2SM network to the generator, mapping visual features to a semantic-related space. On the other hand, they transfer the semantic embeddings by another VOPE module. The outputs of V2SM and VOPE are aligned by cosine similarity and ...
The authors investigate the classification problem with the penalization of the difference in conditional accuracy. They provide a tight characterization between the empirical DCA and variance of 0-1 losses. By combining it with the result of Xie et al., they derive the group-wise distributionally robust optimization f...	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 investigate the classification problem with the penalization of the difference in conditional accuracy. They provide a tight characterization between the empirical DCA and variance of 0-1 losses. By combining it with the result of Xie et al., they derive the group-wise distributionally robust optimi...
This paper describes an approach for pose estimation for objects with symmetries. The approach leverages SO(3)-equivariance to predict distributions over 3D rotations from single images. Image features are obtained from a pre-trained ResNet. These features are orthographically projected to the sphere where the features...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper describes an approach for pose estimation for objects with symmetries. The approach leverages SO(3)-equivariance to predict distributions over 3D rotations from single images. Image features are obtained from a pre-trained ResNet. These features are orthographically projected to the sphere where the ...
The paper tackles the problem of endowing Transformers with the ability to encode information about the past via recurrence. The proposed architecture can leverage the recurrent connections to improve the sample efficiency while maintaining expressivity due to the use of self-attention. Strengths: - The paper is easy...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper tackles the problem of endowing Transformers with the ability to encode information about the past via recurrence. The proposed architecture can leverage the recurrent connections to improve the sample efficiency while maintaining expressivity due to the use of self-attention. Strengths: - The paper...
The authors in this paper discuss the pros and cons of how variance and bias affect gradient estimator, to better understand how gradients are approximated by brains. The authors argue that while bias and variance generally affect training data learning negatively, some amount of bias, variance can be beneficial for ge...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The authors in this paper discuss the pros and cons of how variance and bias affect gradient estimator, to better understand how gradients are approximated by brains. The authors argue that while bias and variance generally affect training data learning negatively, some amount of bias, variance can be beneficia...
This paper presents a GNN-based framework for graph-level anomaly detection, GmapAD. Specifically, GmapAD encapsulate graph structures and node features into node representations and selects the best subset of informative nodes in all possible nodes in graphs for mapping graphs into the new feature space. Then normal a...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a GNN-based framework for graph-level anomaly detection, GmapAD. Specifically, GmapAD encapsulate graph structures and node features into node representations and selects the best subset of informative nodes in all possible nodes in graphs for mapping graphs into the new feature space. Then ...
The paper advocates for an offline training mechanism for memristor devices and proposes a customized training approach (employing noise-aware training and batch normalization) for such devices. The results across a set of small-sized NNs show that using BN and this hardware-restriction-aware training improves the fina...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper advocates for an offline training mechanism for memristor devices and proposes a customized training approach (employing noise-aware training and batch normalization) for such devices. The results across a set of small-sized NNs show that using BN and this hardware-restriction-aware training improves ...
The paper under review proposes a method of selecting a single sound source from a mixture of sounds in a video via a text description of the visual component of the video. The system can be trained on unlabeled data, aka unsupervised training. This is a novel configuration of using a pre-trained audio-visual correspon...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper under review proposes a method of selecting a single sound source from a mixture of sounds in a video via a text description of the visual component of the video. The system can be trained on unlabeled data, aka unsupervised training. This is a novel configuration of using a pre-trained audio-visual c...