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 is a theoretical work on the analysis of the generalization error in federated learning with heterogeneity of the losses and partial participation of the clients in the training process. The authors address the question: "Would the unparticipating clients benefit from the model trained by participating clients?". ...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This is a theoretical work on the analysis of the generalization error in federated learning with heterogeneity of the losses and partial participation of the clients in the training process. The authors address the question: "Would the unparticipating clients benefit from the model trained by participating cli...
This paper proposed a strategy to make the temperature learnable. Strength: This paper proposed a simple strategy to make the temperature learnable. The proposed method simply changes two lines of the training codes to improve the performance. Weakness: 1. This paper is badly written and hard to follow. 2. The key id...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposed a strategy to make the temperature learnable. Strength: This paper proposed a simple strategy to make the temperature learnable. The proposed method simply changes two lines of the training codes to improve the performance. Weakness: 1. This paper is badly written and hard to follow. 2. Th...
This paper presents a novel network architecture called Substructure-Atom Cross Attention (SACA) for molecular graphs that effectively combines Transformers and GNNs. Substructural patterns are important in molecular tasks, as we see in traditional chemoinformatics fingerprints like ECFP and fragmentation like BRICS/RE...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a novel network architecture called Substructure-Atom Cross Attention (SACA) for molecular graphs that effectively combines Transformers and GNNs. Substructural patterns are important in molecular tasks, as we see in traditional chemoinformatics fingerprints like ECFP and fragmentation like ...
This work introduces robust Universal Adversarial Perturbations (UAPs). The objective of robust UAPs is to increase the resilience of UAPs against image transformations. Robust UAPs are obtained by incorporating transformation functions into the UAP generation process. Evaluation on CIFAR10 and ImageNet demonstrate the...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This work introduces robust Universal Adversarial Perturbations (UAPs). The objective of robust UAPs is to increase the resilience of UAPs against image transformations. Robust UAPs are obtained by incorporating transformation functions into the UAP generation process. Evaluation on CIFAR10 and ImageNet demonst...
1. This paper proposed a two-stage method that reasons over sets of objects or statements. 2. The paper performs empirical evaluation on the BIG-bench benchmark and shows improvements over fewshot GPT3 Strengths: 1) Experimental results suggest the proposed method has an improvement over few-shot GPT. Weaknesses: 1) T...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: 1. This paper proposed a two-stage method that reasons over sets of objects or statements. 2. The paper performs empirical evaluation on the BIG-bench benchmark and shows improvements over fewshot GPT3 Strengths: 1) Experimental results suggest the proposed method has an improvement over few-shot GPT. Weakness...
This work focus on adaptive frameworks for bilevel optimization. It first introduce BiAdam that achieves $\tilde{O}(\epsilon^{-4})$ complexity. Then it introduces BiAdam with variance reduction and achieves $\tilde{O}(\epsilon^{-3})$ complexity. Both results matches state-of-the-art complexities. strengths: a) Th...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This work focus on adaptive frameworks for bilevel optimization. It first introduce BiAdam that achieves $\tilde{O}(\epsilon^{-4})$ complexity. Then it introduces BiAdam with variance reduction and achieves $\tilde{O}(\epsilon^{-3})$ complexity. Both results matches state-of-the-art complexities. strengths: ...
This work introduces a diffusion based approach for unconditional generation of realistic protein structures. It is claimed that the generated structures are similar to naturally-occuring proteins in complexity and structural patterns. [+] The representation of the protein backbone as a sequence of angles should be a ...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work introduces a diffusion based approach for unconditional generation of realistic protein structures. It is claimed that the generated structures are similar to naturally-occuring proteins in complexity and structural patterns. [+] The representation of the protein backbone as a sequence of angles shou...
This paper considers the problem of correlation testing in the federated framework compatible with secure aggregation. They motivate the problem of correlation testing with several applications and also give a relevant hospitals based example for the federated setting. They propose an algorithm with two simple observat...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper considers the problem of correlation testing in the federated framework compatible with secure aggregation. They motivate the problem of correlation testing with several applications and also give a relevant hospitals based example for the federated setting. They propose an algorithm with two simple ...
This paper addresses the SSDA problem from the perspective of causality learning. It involves two step for learning, one using GAN and data augmentation to generate intermediate data and second using two semi-supervised learner to do cross-supervision for the aim of debiasing. The effectiveness of the proposed method i...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper addresses the SSDA problem from the perspective of causality learning. It involves two step for learning, one using GAN and data augmentation to generate intermediate data and second using two semi-supervised learner to do cross-supervision for the aim of debiasing. The effectiveness of the proposed ...
This paper proposes to regularize neural networks using isometric properties. The paper presents some intriguing ideas along with some preliminary results that would be worth exploring further. Strength: The idea is relatively straightforward and intuitive. Some preliminary experimental results are provided. Weakness...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes to regularize neural networks using isometric properties. The paper presents some intriguing ideas along with some preliminary results that would be worth exploring further. Strength: The idea is relatively straightforward and intuitive. Some preliminary experimental results are provided. ...
This paper propose an approach to address the problem of imperfect reward under offline settings. The main idea is to formulate the problem as a bi-level optimization problem, where the upper problem is to minimize the gap between the stationary distribution of the optimal policy defined by the imperfect reward, and ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper propose an approach to address the problem of imperfect reward under offline settings. The main idea is to formulate the problem as a bi-level optimization problem, where the upper problem is to minimize the gap between the stationary distribution of the optimal policy defined by the imperfect rewa...
The paper proposes ZONE, a computational framework of curriculum task generation for reinforcement learning based on a concept from developmental psychology, i.e., zone of proximal development (ZPD). The target of the framework is to measure and schedule difficulty and progression of a student model. The approach is co...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes ZONE, a computational framework of curriculum task generation for reinforcement learning based on a concept from developmental psychology, i.e., zone of proximal development (ZPD). The target of the framework is to measure and schedule difficulty and progression of a student model. The approa...
This paper incorporates the resisual algorithm (RA) with SAC-N used in offline RL setting, and finds that the additional RA objective can not only improve the SOTA performance, but also reduce the number of ensemble functions for a more efficient computation. Strength 1. The paper empirically shows that adding RA to SA...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper incorporates the resisual algorithm (RA) with SAC-N used in offline RL setting, and finds that the additional RA objective can not only improve the SOTA performance, but also reduce the number of ensemble functions for a more efficient computation. Strength 1. The paper empirically shows that adding ...
The authors present a self-training approach for detecting novel objects in an open-world setting using off-the-shelf monocular depth and normal estimators. The presented method improves over prior work on open-world settings on COCO and ADE20K. Strengths: - Intuitive, novel method that outperforms prior work on impo...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors present a self-training approach for detecting novel objects in an open-world setting using off-the-shelf monocular depth and normal estimators. The presented method improves over prior work on open-world settings on COCO and ADE20K. Strengths: - Intuitive, novel method that outperforms prior work...
The authors investigate an important and open question: how does the the learning rule (gradient descent, Feedback alignment, Direct Feedback Alignment, error-modulated Hebbian learning, Gated linear networks) influence the learning dynamics during neural network training. The authors focus on wide neural networks, and...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The authors investigate an important and open question: how does the the learning rule (gradient descent, Feedback alignment, Direct Feedback Alignment, error-modulated Hebbian learning, Gated linear networks) influence the learning dynamics during neural network training. The authors focus on wide neural netwo...
The paper proposes an interesting approach to predicting freezing layers during training by training a meta-predictor. The attention based meta-predictor takes in weight history and predicts if a layer should be frozen or not. The predictor is also pretrained on a dataset generated using the CKA similarities between a ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes an interesting approach to predicting freezing layers during training by training a meta-predictor. The attention based meta-predictor takes in weight history and predicts if a layer should be frozen or not. The predictor is also pretrained on a dataset generated using the CKA similarities be...
The paper empirically investigates whether pretrained large language models can have abstraction capability -- whether the model can be aware of the grammar instead of memorizing surface word patterns, taking from a transferability perspective between tasks of different characteristics. The paper also reports the effe...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper empirically investigates whether pretrained large language models can have abstraction capability -- whether the model can be aware of the grammar instead of memorizing surface word patterns, taking from a transferability perspective between tasks of different characteristics. The paper also reports ...
This paper proposes a framework unifying multiple approaches for out-of-distribution generalization, namely, (i) data-augmentation, (ii) enforcing distributional invariances and (iii) invariant risk minimization. The framework is centered around the assumption that both the training and testing domains are "Causally In...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a framework unifying multiple approaches for out-of-distribution generalization, namely, (i) data-augmentation, (ii) enforcing distributional invariances and (iii) invariant risk minimization. The framework is centered around the assumption that both the training and testing domains are "Cau...
This paper considers different spiking neuron models and explores hyperparameters on multiple tasks in order to facilitate efficient training via the surrogate gradient method. The study focuses primarily on the shape of the surrogate gradient function. They find that low dampening, high sharpness and low tail-flatness...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper considers different spiking neuron models and explores hyperparameters on multiple tasks in order to facilitate efficient training via the surrogate gradient method. The study focuses primarily on the shape of the surrogate gradient function. They find that low dampening, high sharpness and low tail-...
Authors propose Action Adaptive Policy (AAP) to adapt to unseen effects of actions during inference of an embodied AI navigation task. Experiments show that AAP is effective at adapting to unseen action outcomes at test time, outperforming a set of strong baselines by a significant margin, and works with disabled actio...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: Authors propose Action Adaptive Policy (AAP) to adapt to unseen effects of actions during inference of an embodied AI navigation task. Experiments show that AAP is effective at adapting to unseen action outcomes at test time, outperforming a set of strong baselines by a significant margin, and works with disabl...
The paper uses a hierarchical attention based neural networ to learn the solution operator associated to multiscale PDEs. In addition, the second contribution is the sobolev type norm used as the loss function, giving more weight to the higher frequencies of the target. I have mixed feelings about this paper. I have a...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper uses a hierarchical attention based neural networ to learn the solution operator associated to multiscale PDEs. In addition, the second contribution is the sobolev type norm used as the loss function, giving more weight to the higher frequencies of the target. I have mixed feelings about this paper. ...
The authors propose to initialize ResNets with ReLU activations with exact dynamical isometry by combining ideas from delta-orthogonal initialization and looks-linear initialization. The proposed initialization scheme guarantees that every residual block is an exact orthogonal map. A small empirical evaluation shows th...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors propose to initialize ResNets with ReLU activations with exact dynamical isometry by combining ideas from delta-orthogonal initialization and looks-linear initialization. The proposed initialization scheme guarantees that every residual block is an exact orthogonal map. A small empirical evaluation ...
This work solves a problem of sample-efficient experimental design with teaming of a human expert and AI. Since an AI starts to solve the experimental design from scratch, a human expert can help the AI to accelerate the design process. To fully utilize such knowledge from human experts and AI system, the authors pro...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This work solves a problem of sample-efficient experimental design with teaming of a human expert and AI. Since an AI starts to solve the experimental design from scratch, a human expert can help the AI to accelerate the design process. To fully utilize such knowledge from human experts and AI system, the aut...
This paper suggested that colored noise for motor exploration in RL, in particular Pink noise ($\beta$ = 1) is a better choice than using Gaussian white noise and OU noise in many control tasks. The method is simple and versatile for online RL, and effective on the being tested environments. The main contribution is th...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper suggested that colored noise for motor exploration in RL, in particular Pink noise ($\beta$ = 1) is a better choice than using Gaussian white noise and OU noise in many control tasks. The method is simple and versatile for online RL, and effective on the being tested environments. The main contributi...
This paper evaluates the impact of using language and vision pretrained transformers as an initialization for a Decision Transformer style model used in offline RL settings. It provide methods for improving the transferability of the pretrained models and demonstrate that pretraining consistently matches or outperform...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper evaluates the impact of using language and vision pretrained transformers as an initialization for a Decision Transformer style model used in offline RL settings. It provide methods for improving the transferability of the pretrained models and demonstrate that pretraining consistently matches or ou...
The paper proposes a new way to measure the 'importance' of a neuron, by re-parameterizing the scaling parameters of the batchnorm layers. The proposed method is to use a sigmoid on the scaling parameter, to ensure that it lies in [0,1], and the value of the sigmoid reflects the importance. Using this value (value clos...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a new way to measure the 'importance' of a neuron, by re-parameterizing the scaling parameters of the batchnorm layers. The proposed method is to use a sigmoid on the scaling parameter, to ensure that it lies in [0,1], and the value of the sigmoid reflects the importance. Using this value (va...
This paper proposes ReAct; a novel framework for prompting large language models (LLMs) on tasks that require explicit reasoning and/or acting in an environment. Driven by recent work in plugging in LLMs into the main loop of a reasoning problem (e.g., fact verification or multi-hop question answering), or embodied pla...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes ReAct; a novel framework for prompting large language models (LLMs) on tasks that require explicit reasoning and/or acting in an environment. Driven by recent work in plugging in LLMs into the main loop of a reasoning problem (e.g., fact verification or multi-hop question answering), or embo...
This paper presents an empirical study of the adversarial robustness of dynamic neural networks (DyNNs) with early-exits. The authors find that DyNNs are more robust than SDNNs, and DyNNs can be used to generate adversarial samples efficiently. The authors also proposes a novel adversarial attack method specifically de...	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 presents an empirical study of the adversarial robustness of dynamic neural networks (DyNNs) with early-exits. The authors find that DyNNs are more robust than SDNNs, and DyNNs can be used to generate adversarial samples efficiently. The authors also proposes a novel adversarial attack method specifi...
The paper considers the representation learning problem in domain generalization setting with fairness constraints (DP, EO). The paper presents upper and lower bounds for accuracy, and also an upper bound for fairness violation. Empirically, the paper aims to solve a minmax problem to account for fairness when learning...	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 considers the representation learning problem in domain generalization setting with fairness constraints (DP, EO). The paper presents upper and lower bounds for accuracy, and also an upper bound for fairness violation. Empirically, the paper aims to solve a minmax problem to account for fairness when ...
The paper suggests a set of selection criteria to training points in each AL step to avoid the so-called catastrophic forgetting which degrades model performance. A variety of heuristics is proposed some motivated by prior work and 2 new ones proposed by the authors (CAL-SD and CAL_SDS2). Strengths 1. The problem tackl...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper suggests a set of selection criteria to training points in each AL step to avoid the so-called catastrophic forgetting which degrades model performance. A variety of heuristics is proposed some motivated by prior work and 2 new ones proposed by the authors (CAL-SD and CAL_SDS2). Strengths 1. The probl...
The paper argues that when the number of tasks are increased in a continual learning setting, catastrophic forgetting ceases to be a problem and a simple SGD training with gradient masking keeps on accumulating knowledge. Towards this, the authors propose a framework, called SCoLe, where the authors repetitively sample...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper argues that when the number of tasks are increased in a continual learning setting, catastrophic forgetting ceases to be a problem and a simple SGD training with gradient masking keeps on accumulating knowledge. Towards this, the authors propose a framework, called SCoLe, where the authors repetitivel...
This paper proposes a framework for federated learning by adjusting the frequency of communication between agents and the server with an adaptive quantization scheme. Specifically, the authors combine two quantization schemes, namely, the adaptive quantization rule (AdaQuantFL) and lazily aggregated quantization (LAQ)....	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 proposes a framework for federated learning by adjusting the frequency of communication between agents and the server with an adaptive quantization scheme. Specifically, the authors combine two quantization schemes, namely, the adaptive quantization rule (AdaQuantFL) and lazily aggregated quantizatio...
This paper investigates the training dynamics in federated deep learning from the perspective of mini-batch SGD. Client coherence and global weight shrinking regularization are introduced in the proposed FedAWO framework. The corresponding parameters are learned in the server side with a proxy dataset. FedAWO’s effecti...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper investigates the training dynamics in federated deep learning from the perspective of mini-batch SGD. Client coherence and global weight shrinking regularization are introduced in the proposed FedAWO framework. The corresponding parameters are learned in the server side with a proxy dataset. FedAWO’s...
I'm increasing my score from 5 to 6 as the authors have addressed several of my concerns. However, I am still not convinced about the utility of this architecture or pretrained weights beyond the specific experimental setting that is considered by the authors because the task someone is performing has a big effect on t...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: I'm increasing my score from 5 to 6 as the authors have addressed several of my concerns. However, I am still not convinced about the utility of this architecture or pretrained weights beyond the specific experimental setting that is considered by the authors because the task someone is performing has a big eff...
This paper considers the online low-rank matrix completion problem for recommendation systems. An expected regret formulation, as the core analysis goal of the proposed problem, is defined and then minimized via the proposed ETC algorithm and a further OCTAL algorithm for special case of rank-one reward matrices. It pr...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers the online low-rank matrix completion problem for recommendation systems. An expected regret formulation, as the core analysis goal of the proposed problem, is defined and then minimized via the proposed ETC algorithm and a further OCTAL algorithm for special case of rank-one reward matrice...
The paper investigates the HBFP parameter search space to further improve efficiency and density in hardware accelerators. They propose training on HBFP6, after experiments making the trade off: lower block size gives more accuracy, but is less HW efficient. They show accuracy booster method that switch number of manti...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper investigates the HBFP parameter search space to further improve efficiency and density in hardware accelerators. They propose training on HBFP6, after experiments making the trade off: lower block size gives more accuracy, but is less HW efficient. They show accuracy booster method that switch number ...
The paper performs a large scale empirical study to investigate whether disentangled representations provide a clear benefit for the final performance on downstream tasks. First, the ground-truth disentangled representation (normalized true factors) are compared to a rotated version of the same representations. The aut...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper performs a large scale empirical study to investigate whether disentangled representations provide a clear benefit for the final performance on downstream tasks. First, the ground-truth disentangled representation (normalized true factors) are compared to a rotated version of the same representations....
This paper presents the factorized Fourier neural operator (F-FNO), which brings a set of techniques on FNO. The first technique is the separable Fourier representation to improve model stability and performance. Second, the improved residual connections. Third, the training strategies are carefully designed. Impressiv...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents the factorized Fourier neural operator (F-FNO), which brings a set of techniques on FNO. The first technique is the separable Fourier representation to improve model stability and performance. Second, the improved residual connections. Third, the training strategies are carefully designed. I...
This paper studies the benign overfitting phenomenon of deep neural models, namely over-parameterized deep neural models fitting the training data can still achieve low generalization error. The authors focus on a specific regime where the model is mildly overparameterized and label noise is prevalent. They show that i...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the benign overfitting phenomenon of deep neural models, namely over-parameterized deep neural models fitting the training data can still achieve low generalization error. The authors focus on a specific regime where the model is mildly overparameterized and label noise is prevalent. They sho...
This paper tackle the task of evaluating active feature acquisition (AFA) agent when the AFA agent cannot always access the complete set of features during decision making and there is distribution shift between training and evaluation stages. Specifically, it considers a subset of AFA problems where the features acqui...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper tackle the task of evaluating active feature acquisition (AFA) agent when the AFA agent cannot always access the complete set of features during decision making and there is distribution shift between training and evaluation stages. Specifically, it considers a subset of AFA problems where the featur...