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 introduces a the adversary-aware partial label learning problem to protect the data privacy. The novel adversary-aware loss function, together with an immature teacher within momentum disambiguation algorithm, has achieved state of-the-art performance and proven to be a provable classifier. Pros: * This pape...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces a the adversary-aware partial label learning problem to protect the data privacy. The novel adversary-aware loss function, together with an immature teacher within momentum disambiguation algorithm, has achieved state of-the-art performance and proven to be a provable classifier. Pros: * T...
The authors exploit the properties of adversarial examples to design weighted augmented CE loss to improve transferability. Then, they add two parameters, i.e., fuzzy scaling and fuzzy domain to eliminate some fuzzy region in the searching. Thus, adversarial example with higher transferability is more likely to be foun...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors exploit the properties of adversarial examples to design weighted augmented CE loss to improve transferability. Then, they add two parameters, i.e., fuzzy scaling and fuzzy domain to eliminate some fuzzy region in the searching. Thus, adversarial example with higher transferability is more likely to...
This paper proposed an adapt scheme for sharpness-aware minimization (SAM), in which the algorithm simply perform ERM when landscape is flat and perform SAM which landscape is sharp. This method is based on couple of approximation strategies. The author also establish the theoretical guarantee for their proposed algori...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposed an adapt scheme for sharpness-aware minimization (SAM), in which the algorithm simply perform ERM when landscape is flat and perform SAM which landscape is sharp. This method is based on couple of approximation strategies. The author also establish the theoretical guarantee for their propose...
This work shows new results on convex formulations of neural networks. The goal is to formulate the nonconvex optimization problem as minimum norm problems, derive the convex dual problem and study whether the duality gap is zero. If it is zero, then the dual can be solved via convex optimization tools with theoretical...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This work shows new results on convex formulations of neural networks. The goal is to formulate the nonconvex optimization problem as minimum norm problems, derive the convex dual problem and study whether the duality gap is zero. If it is zero, then the dual can be solved via convex optimization tools with the...
The authors analyze an approach to improve the playing strength of human-imitating chess engines, while trying to maintain the human-like play. The basic method is to use transfer learning (one-step curriculum learning) by retraining an existing human-like model with data from better players or a conventional chess eng...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors analyze an approach to improve the playing strength of human-imitating chess engines, while trying to maintain the human-like play. The basic method is to use transfer learning (one-step curriculum learning) by retraining an existing human-like model with data from better players or a conventional c...
This paper introduces a geometric regularization technique for nerf training under a sparse view setting. The regularization is based on cross-view warping between seen/unseen views using rendered depth, where the warped image is utilized for supervision in feature space. Pros: 1. The overall idea is tidy and easy t...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a geometric regularization technique for nerf training under a sparse view setting. The regularization is based on cross-view warping between seen/unseen views using rendered depth, where the warped image is utilized for supervision in feature space. Pros: 1. The overall idea is tidy an...
Positional embedding is an important component in transformers. Moreover, generalization ability to longer sequences with proper positional embedding is also hot topic. Current paper proposes a new positional embedding conditioned on the input (PEG) and not only position itself. This is done vie convolutional layer and...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Positional embedding is an important component in transformers. Moreover, generalization ability to longer sequences with proper positional embedding is also hot topic. Current paper proposes a new positional embedding conditioned on the input (PEG) and not only position itself. This is done vie convolutional l...
The authors examine in depth the learning-sticking problem in early training of DNNs from the perspective of learning dynamics. They offer a novel explanation for what often happens to network weights and inner activations during the problem - which they call "the temporary feature collapse" (TFC) phenomenon They prov...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors examine in depth the learning-sticking problem in early training of DNNs from the perspective of learning dynamics. They offer a novel explanation for what often happens to network weights and inner activations during the problem - which they call "the temporary feature collapse" (TFC) phenomenon T...
The authors proposed a rendering method for a neural radiance field. When marching on each light ray, the authors chose the sampling points using the idea of importance sampling; the samples are drawn in accordance with the contributions to the final rendered color. The method was designed so as to reduce the number of...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors proposed a rendering method for a neural radiance field. When marching on each light ray, the authors chose the sampling points using the idea of importance sampling; the samples are drawn in accordance with the contributions to the final rendered color. The method was designed so as to reduce the n...
This paper provides experiments suggesting that there is only a very small correlation between performance of a neural classifier on ImageNet and its performance (after fine-tuning) on datasets that are not web-scraped. The numerical experiment seem to be well conducted. Yet, the results of the paper are not very surp...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper provides experiments suggesting that there is only a very small correlation between performance of a neural classifier on ImageNet and its performance (after fine-tuning) on datasets that are not web-scraped. The numerical experiment seem to be well conducted. Yet, the results of the paper are not v...
Main concern. A catalog of experiments to find the best combination of positional encoders for any explored type of graphs. The most successful encoding (powers of adjacency matrices) is not scalable unless large graphs are previously sampled. The argument that the powers of adjacencies for small r=1,2 beats full atten...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Main concern. A catalog of experiments to find the best combination of positional encoders for any explored type of graphs. The most successful encoding (powers of adjacency matrices) is not scalable unless large graphs are previously sampled. The argument that the powers of adjacencies for small r=1,2 beats fu...
The paper combines data replay and EWC into a single objective to tackle continual learning of tasks. Results confirm the benefit of combining an experience replay buffer to replay data from old tasks, and functional priors such as EWC. Authors perform several ablations on split CIFAR100, split mini imagenet and imag...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper combines data replay and EWC into a single objective to tackle continual learning of tasks. Results confirm the benefit of combining an experience replay buffer to replay data from old tasks, and functional priors such as EWC. Authors perform several ablations on split CIFAR100, split mini imagenet ...
The paper proposes to enhance the transferability of adversarial examples by doing adversarial attacks using multiple data augmentation parallelly. Pros: 1. The paper study a novel problem of how data augmentation affects adversarial attacks' effectiveness. 2. The work considers multiple model architectures to test...	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 to enhance the transferability of adversarial examples by doing adversarial attacks using multiple data augmentation parallelly. Pros: 1. The paper study a novel problem of how data augmentation affects adversarial attacks' effectiveness. 2. The work considers multiple model architectures...
This paper proposes a hyper-graph neural network model, ED-HNN, that can model hypergraph diffusion process. They claimed that ED-HNN shows superiority in processing heterophilic hypergraphs and constructing deep models. Strength: - 1. The presentation is clear. - 2. The overall experiments show the superiorit...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a hyper-graph neural network model, ED-HNN, that can model hypergraph diffusion process. They claimed that ED-HNN shows superiority in processing heterophilic hypergraphs and constructing deep models. Strength: - 1. The presentation is clear. - 2. The overall experiments show the su...
This paper introduces two sources of noise applied during training to flatten the loss landscape and make weight more resilient to device variability: Hardware-simulated variation (HSA, gaussian noise which scales with the width of the weight distribution of each layer) and Gradient-ascent variation (GAV, added noise i...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper introduces two sources of noise applied during training to flatten the loss landscape and make weight more resilient to device variability: Hardware-simulated variation (HSA, gaussian noise which scales with the width of the weight distribution of each layer) and Gradient-ascent variation (GAV, added...
This work derives a new lower bound for the class of first-order methods applied to minimize $L$-smooth functions satisfying the Polyak-Łojasiewicz condition with parameter $\mu$. In particular, the authors derive $\Omega\left(\left(\frac{L}{\mu}\right)^{1 - \alpha}\right)$ lower bound for finding $\varepsilon$-solutio...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This work derives a new lower bound for the class of first-order methods applied to minimize $L$-smooth functions satisfying the Polyak-Łojasiewicz condition with parameter $\mu$. In particular, the authors derive $\Omega\left(\left(\frac{L}{\mu}\right)^{1 - \alpha}\right)$ lower bound for finding $\varepsilon$...
This paper proposes a general framework Functional Risk Minimization (FRM), a general framework for scalable training objectives which results in better performance in small experiments in regression and reinforcement learning. FRM model each data point (x_i, y_i) as coming from its own function f_\theta_i. The authors...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a general framework Functional Risk Minimization (FRM), a general framework for scalable training objectives which results in better performance in small experiments in regression and reinforcement learning. FRM model each data point (x_i, y_i) as coming from its own function f_\theta_i. The...
This paper proposes a new backup operator that exploits potential graph structure in the observed transitions. The purpose of this backup operator is to increase data efficiency by allowing information to be backed-up further along state transitions and providing better value estimates. Their results demonstrate, rathe...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a new backup operator that exploits potential graph structure in the observed transitions. The purpose of this backup operator is to increase data efficiency by allowing information to be backed-up further along state transitions and providing better value estimates. Their results demonstrat...
In this work, the authors propose a pretraining diffusion model for defending against 3D adversarial point clouds. They also shows the limitation of robust training for defense due to gradient obfuscation. The results have shown better qualitative performance than SOTA methods. Pros: +) The idea is clear and easy to ...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this work, the authors propose a pretraining diffusion model for defending against 3D adversarial point clouds. They also shows the limitation of robust training for defense due to gradient obfuscation. The results have shown better qualitative performance than SOTA methods. Pros: +) The idea is clear and ...
This paper tackles the problem of quantity-quality trade-off in pseudo-labeling based semi-supervised learning. Specifically, the authors develop a truncated Gaussian function to weight samples based on their confidence, which can be viewed as a soft version of the confidence threshold. They also enhance the utilizatio...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper tackles the problem of quantity-quality trade-off in pseudo-labeling based semi-supervised learning. Specifically, the authors develop a truncated Gaussian function to weight samples based on their confidence, which can be viewed as a soft version of the confidence threshold. They also enhance the ut...
The submission titled "Effective passive membership inference attacks in federated learning against overparameterized models" describes a well-motivated black-box membership inference attack that can be instantiated in a federated learning setting, i.e. this is an inference from user update/gradient to dataset membersh...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The submission titled "Effective passive membership inference attacks in federated learning against overparameterized models" describes a well-motivated black-box membership inference attack that can be instantiated in a federated learning setting, i.e. this is an inference from user update/gradient to dataset ...
In this work, the authors theoretically quantify the maximization bias in generalized linear models with Gaussian distributions. In practice, the maximization bias often manifests itself in the calibration of online advertising recommendation system. The authors propose a variance-adjusting de-biasing meta-algorithm th...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this work, the authors theoretically quantify the maximization bias in generalized linear models with Gaussian distributions. In practice, the maximization bias often manifests itself in the calibration of online advertising recommendation system. The authors propose a variance-adjusting de-biasing meta-algo...
This paper proposed a joint training framework of three components - instrument recognition, transcription and source separation. The instrument recognition module is optional and can be replaced by human inputs. The joint training of the transcription and source separation module was shown to be beneficial. The author...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed a joint training framework of three components - instrument recognition, transcription and source separation. The instrument recognition module is optional and can be replaced by human inputs. The joint training of the transcription and source separation module was shown to be beneficial. Th...
This paper proposed a self-supervised visual backbone pretraining method for reward learning in control. The key idea is to formulate representation learning from egocentric videos as an offline goal-conditioned reinforcement learning problem, whose dual-form indicated a way to implement Value-Implicit Pre-training (VI...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposed a self-supervised visual backbone pretraining method for reward learning in control. The key idea is to formulate representation learning from egocentric videos as an offline goal-conditioned reinforcement learning problem, whose dual-form indicated a way to implement Value-Implicit Pre-trai...
This paper studies preference-based offline reinforcement learning where we only have access to the preferences over offline trajectories. Compared to the standard offline RL setting, PbRL doesn't assume the reward function is available, which is more realistic. To solve the problem, the authors propose an iterative al...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies preference-based offline reinforcement learning where we only have access to the preferences over offline trajectories. Compared to the standard offline RL setting, PbRL doesn't assume the reward function is available, which is more realistic. To solve the problem, the authors propose an iter...
This paper proposes a method for learning from data following a mixture of distributions whose input-output relationships are different. The proposed method maintains multiple hypothesis functions accounting for those different distributions and minimizes the loss between the label and the closest output among those gi...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a method for learning from data following a mixture of distributions whose input-output relationships are different. The proposed method maintains multiple hypothesis functions accounting for those different distributions and minimizes the loss between the label and the closest output among ...
In this paper, the authors propose a new loss and evaluation metric based on persistence theory for matching different image segmentations. The proposed loss can be quite different from the usual Wasserstein distance between persistence diagrams, and is rather based on the algebraic foundations of persistence theory, w...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In this paper, the authors propose a new loss and evaluation metric based on persistence theory for matching different image segmentations. The proposed loss can be quite different from the usual Wasserstein distance between persistence diagrams, and is rather based on the algebraic foundations of persistence t...
The paper suggests a novel approach to extend neural networks during training using the functional gradient. The approach starts with an impoverished network and iteratively adds neurons to the network making sure to follow the true gradient direction. The authors mathematically derive an optimal technique to grow the ...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper suggests a novel approach to extend neural networks during training using the functional gradient. The approach starts with an impoverished network and iteratively adds neurons to the network making sure to follow the true gradient direction. The authors mathematically derive an optimal technique to g...
This paper generalizes the recently proposed S4 model with an input dependent time constant used in state transition matrix and conduct thorough experiments to illustrate its benefits. Strengths: - Input dependent transition mechanism in state space models is a novel, important and natural extention of the original S4 ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper generalizes the recently proposed S4 model with an input dependent time constant used in state transition matrix and conduct thorough experiments to illustrate its benefits. Strengths: - Input dependent transition mechanism in state space models is a novel, important and natural extention of the orig...
The authors proposes a new framework that tries to combine neural perception and symbolic reasoning. In general, this framework consists of three components: (1) Neural Solver (NS) (2) Mask Predictor (MP) and (3) Logical Solver (LS). The NS computes an initial solution for the input tasks, then the MP predicts the in...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors proposes a new framework that tries to combine neural perception and symbolic reasoning. In general, this framework consists of three components: (1) Neural Solver (NS) (2) Mask Predictor (MP) and (3) Logical Solver (LS). The NS computes an initial solution for the input tasks, then the MP predict...
The paper assumes limited label availability and propose a weighted multi-source distillation method, in practice that means distill multiple (diverse) source models trained on different domains, weighing them by their relevance for the target task, assuming such target task is available Strengths: - The paper shows th...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper assumes limited label availability and propose a weighted multi-source distillation method, in practice that means distill multiple (diverse) source models trained on different domains, weighing them by their relevance for the target task, assuming such target task is available Strengths: - The paper ...
This paper considers a stochastic bandit model where the reward function belongs to a class of uniformly bounded functions and the additive noise can be heteroscedastic. A multi-level learning framework is proposed to tackle this general bandit model where this paper designs an algorithm to construct the variance-aware...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers a stochastic bandit model where the reward function belongs to a class of uniformly bounded functions and the additive noise can be heteroscedastic. A multi-level learning framework is proposed to tackle this general bandit model where this paper designs an algorithm to construct the varian...
This paper focuses on CNN compression using low-rank factorizations. In particular, the authors propose a new tensor decomposition called SeKron, which generalizes traditional TT, TR, CP and Tucker decomposition. The authors establish a SVD-based algorithm to decompose a given tensor into the SeKron format. Due to the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper focuses on CNN compression using low-rank factorizations. In particular, the authors propose a new tensor decomposition called SeKron, which generalizes traditional TT, TR, CP and Tucker decomposition. The authors establish a SVD-based algorithm to decompose a given tensor into the SeKron format. Due...
This paper addresses the problem of novel view synthesis from a set of input images. The authors presented a point-based method which is build on top of exsiting works on differentiable point-based rendering. They introduced point sculpting for pruning and adding points to improve the photo-consistency and completness ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper addresses the problem of novel view synthesis from a set of input images. The authors presented a point-based method which is build on top of exsiting works on differentiable point-based rendering. They introduced point sculpting for pruning and adding points to improve the photo-consistency and comp...
The paper presents oblivious sketches for logistic regression with significantly improved sketching sizes both in terms of $d$- the dimension of the data and $\mu$ the measure of complexity of compressing the data. The paper also gives sketches for the $\ell_1$ regression and variance regularized logistic regression. T...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper presents oblivious sketches for logistic regression with significantly improved sketching sizes both in terms of $d$- the dimension of the data and $\mu$ the measure of complexity of compressing the data. The paper also gives sketches for the $\ell_1$ regression and variance regularized logistic regre...
the submission compared convolution neural networks of the same building block and the transformer models on the impact of masked pre-training on downstream tasks including secondary structure predictions and protein design tasks. Results show that there isn't a significant difference between convolutional neural netwo...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: the submission compared convolution neural networks of the same building block and the transformer models on the impact of masked pre-training on downstream tasks including secondary structure predictions and protein design tasks. Results show that there isn't a significant difference between convolutional neur...
The paper studied shape-conditioned 3D molecule generation, which aims to generate molecules with a desirable shape. The author proposed an encoder-decoder architecture, where the encoder can encode both molecular graph representation and molecular shape representation by point clouds. Specifically, the main difference...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper studied shape-conditioned 3D molecule generation, which aims to generate molecules with a desirable shape. The author proposed an encoder-decoder architecture, where the encoder can encode both molecular graph representation and molecular shape representation by point clouds. Specifically, the main di...
This paper studies ‘clone structured cognitive graphs’ (CSCG), a particular type of action-conditioned Hidden Markov Model (HMM) with specific assumptions on the emission function (deterministic emissions and partial observability, where multiple hidden states can generate the same observation). They investigate the ab...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper studies ‘clone structured cognitive graphs’ (CSCG), a particular type of action-conditioned Hidden Markov Model (HMM) with specific assumptions on the emission function (deterministic emissions and partial observability, where multiple hidden states can generate the same observation). They investigat...
This paper proposes to use data programming to address the semi-supervised continuous learning, achieving the performance close to the fully supervised continual learning. Pros: * They consider a more realistic continual learning with only a few labeled data, and adopt a novel technique, data programming, to annotate t...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposes to use data programming to address the semi-supervised continuous learning, achieving the performance close to the fully supervised continual learning. Pros: * They consider a more realistic continual learning with only a few labeled data, and adopt a novel technique, data programming, to an...
A new a probabilistic model, named joint Gaussian mixture model(JGMM), is proposed for post-hoc explanations of deep neural networks. By jointly modeling the latent features of lower and higher layers with a joint GMM, JGMM, trained with a post-hoc EM algorithm, is capable of delivering both global explanations (like P...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: A new a probabilistic model, named joint Gaussian mixture model(JGMM), is proposed for post-hoc explanations of deep neural networks. By jointly modeling the latent features of lower and higher layers with a joint GMM, JGMM, trained with a post-hoc EM algorithm, is capable of delivering both global explanations...