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
The paper studies sparse training with dynamic sparsity in the context of extremely sparse neural network models, trained from scratch, as function approximators for deep reinforcement learning. Consequently, the paper proposes a new sparse training method for deep reinforcement learning, named the Rigged Reinforcement...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies sparse training with dynamic sparsity in the context of extremely sparse neural network models, trained from scratch, as function approximators for deep reinforcement learning. Consequently, the paper proposes a new sparse training method for deep reinforcement learning, named the Rigged Reinf...
This paper proposes Crossformer, a Transformer-based model utilizing cross-dimension dependency for Multivariate time series forecasting. Besides the time information, Crossformer considers dependency between dimension information. Specifically, Dimension-Segment-Wise (DSW) embedding and Two-Stage Attention (TSA) layer...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes Crossformer, a Transformer-based model utilizing cross-dimension dependency for Multivariate time series forecasting. Besides the time information, Crossformer considers dependency between dimension information. Specifically, Dimension-Segment-Wise (DSW) embedding and Two-Stage Attention (TS...
This paper describes a method to train a speech enhancement system using a non-intrusive speech quality estimator. Many combinations of approaches, architectures, datasets, and training schedules are compared and measured with many metrics. The proposed approach does not do as well as the supervised or semi-supervised ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper describes a method to train a speech enhancement system using a non-intrusive speech quality estimator. Many combinations of approaches, architectures, datasets, and training schedules are compared and measured with many metrics. The proposed approach does not do as well as the supervised or semi-sup...
The paper proposes SWARM, a system for training large language models (LLM) on alternative environments to HPC clusters that comprise of consumer-grade, preemptible, unreliable, and geographically distributed devices. SWARM employs pipeline parallelism to partition the model layers into pipeline stages on the distribut...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes SWARM, a system for training large language models (LLM) on alternative environments to HPC clusters that comprise of consumer-grade, preemptible, unreliable, and geographically distributed devices. SWARM employs pipeline parallelism to partition the model layers into pipeline stages on the d...
This paper studies adaptive server step sizes for federated learning (FL). The authors propose a method called FedExP, which uses a server step size akin to the POCS algorithm from the literature on finding feasible points inside the intersections of sets. The authors use a similar expression to estimate how much one c...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies adaptive server step sizes for federated learning (FL). The authors propose a method called FedExP, which uses a server step size akin to the POCS algorithm from the literature on finding feasible points inside the intersections of sets. The authors use a similar expression to estimate how mu...
This paper proposes a new framework to adapt large video-based models to down-stream tasks with a parameter-accuracy trade-off. It analyzes different PETL techniques and investigates the importance of fine-tuning position of their methods. In order to better transfer prefix-tuning from NLP to vision task, it compares d...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new framework to adapt large video-based models to down-stream tasks with a parameter-accuracy trade-off. It analyzes different PETL techniques and investigates the importance of fine-tuning position of their methods. In order to better transfer prefix-tuning from NLP to vision task, it co...
The paper introduces BTM training to train a collection of Expert LMs that built an ELMFOREST. Each component in the forest can be dynamically added/removed/ensembled (merged) at any time. The idea of the method is to train scalable and parallel individual models that each specialize in a particular domain. The paper p...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper introduces BTM training to train a collection of Expert LMs that built an ELMFOREST. Each component in the forest can be dynamically added/removed/ensembled (merged) at any time. The idea of the method is to train scalable and parallel individual models that each specialize in a particular domain. The...
This paper proposes a loss function to learn from multi-label data with only one-positive annotation. Specially, the proposed loss function considers three factors: i) the existing asymmetric loss; ii) it uses a threshold to differentiate true-positive and true-negative in the unlabeled candidate labels. The self-label...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a loss function to learn from multi-label data with only one-positive annotation. Specially, the proposed loss function considers three factors: i) the existing asymmetric loss; ii) it uses a threshold to differentiate true-positive and true-negative in the unlabeled candidate labels. The se...
This paper tackles the problem of source-free domain adaptation (SFDA), where a pre-trained source model is adapted using unlabeled target domain data without accessing any source domain data. While previous works mainly focused on cluster assumption in the feature space, the authors propose a different perspective. Th...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper tackles the problem of source-free domain adaptation (SFDA), where a pre-trained source model is adapted using unlabeled target domain data without accessing any source domain data. While previous works mainly focused on cluster assumption in the feature space, the authors propose a different perspec...
This paper studies the task of generalization in RL tasks defined by contextual MDPs (CMDPs). In contrast to meta-RL tasks, this paper focuses on tasks where the agent should learn a policy that performs well on all MDPs instantiated by a CMDP. The paper argues that exploration is key to improving the performance of RL...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the task of generalization in RL tasks defined by contextual MDPs (CMDPs). In contrast to meta-RL tasks, this paper focuses on tasks where the agent should learn a policy that performs well on all MDPs instantiated by a CMDP. The paper argues that exploration is key to improving the performan...
This paper contains two parts. The first part discusses properties of optimal adversarial estimators w.r.t. convex loss and 0-1 loss, including their connection. The second part discusses adversarial lazy training of one-layer wide network, and demonstrates a convergence of population adversarial loss. Strength: Overal...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper contains two parts. The first part discusses properties of optimal adversarial estimators w.r.t. convex loss and 0-1 loss, including their connection. The second part discusses adversarial lazy training of one-layer wide network, and demonstrates a convergence of population adversarial loss. Strength...
A contrastive audio-visual masked autoencoder model is introduced in the paper. The training of self-supervised model is done with both contrastive and masked-autoencoder losses. The model architecture is transformer and only unmasked inputs are kept as input for the encoder. There is a separate encoder for each modal...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: A contrastive audio-visual masked autoencoder model is introduced in the paper. The training of self-supervised model is done with both contrastive and masked-autoencoder losses. The model architecture is transformer and only unmasked inputs are kept as input for the encoder. There is a separate encoder for ea...
The paper considers the problem of bilevel optimization for vertical federated learning. Using a zeroth-order estimator to locally approximate the Jacobian matrix, and adopting GradPerturb algorithm for providing label privacy, it shows convergences in the rate of $O({1}{\sqrt{K}})$ under the nonconvex-stronglyconvex ...	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 considers the problem of bilevel optimization for vertical federated learning. Using a zeroth-order estimator to locally approximate the Jacobian matrix, and adopting GradPerturb algorithm for providing label privacy, it shows convergences in the rate of $O({1}{\sqrt{K}})$ under the nonconvex-strongly...
The paper studies the problem of differential private machine learning (DP ML) and propose to aggreagate the checkpoint (intermediate parameters during training) for test time inference. On the experimental sides, the paper conduct experiments over standard dataset like CIFAR10 and stackoverflow and their methods impro...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper studies the problem of differential private machine learning (DP ML) and propose to aggreagate the checkpoint (intermediate parameters during training) for test time inference. On the experimental sides, the paper conduct experiments over standard dataset like CIFAR10 and stackoverflow and their metho...
The authors present an unsupervised approach, based on recent ideas called 'closed-loop transcription', to extract feature encodings that perform well on both discriminative and generative tasks, on CIFAR-10, CIFAR-100 and Tiny Imagenet data sets. Often there is a tradeoff between generative vs discriminative goals, ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors present an unsupervised approach, based on recent ideas called 'closed-loop transcription', to extract feature encodings that perform well on both discriminative and generative tasks, on CIFAR-10, CIFAR-100 and Tiny Imagenet data sets. Often there is a tradeoff between generative vs discriminative...
This paper proposes a probabilistic method to combine graph generation and contrastive learning. A key innovation is to factorize the graph into several subgraphs. Each subgraph represents a facet of interactions in a sub-community. The learned latent representation for each node is also factorized into sub-representat...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a probabilistic method to combine graph generation and contrastive learning. A key innovation is to factorize the graph into several subgraphs. Each subgraph represents a facet of interactions in a sub-community. The learned latent representation for each node is also factorized into sub-rep...
This paper proposes a new method for training Neural ODEs by employing techniques of synchronization and homotopy optimization. In the task of learning physical systems, the authors have shown that it is possible to learn appropriately for long duration data. S1. This paper is the first example of introducing synchroni...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a new method for training Neural ODEs by employing techniques of synchronization and homotopy optimization. In the task of learning physical systems, the authors have shown that it is possible to learn appropriately for long duration data. S1. This paper is the first example of introducing s...
In this work, authors hypothesize that a failure mode of neural networks corresponds to the case where models depend on features in the null-space of the training risk. That is, representations learned through ERM preserve non-negligible variance along directions that do not affect the training risk itself. Authors the...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this work, authors hypothesize that a failure mode of neural networks corresponds to the case where models depend on features in the null-space of the training risk. That is, representations learned through ERM preserve non-negligible variance along directions that do not affect the training risk itself. Aut...
This paper proposes a scheme to improve open-ended generation from autoregressive models that is based on manipulating the token-level distribution of the model during ancestral sampling. Specifically, this scheme proposes pruning off the tail at each step like other popular decoding methods like Nucleus sampling and t...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a scheme to improve open-ended generation from autoregressive models that is based on manipulating the token-level distribution of the model during ancestral sampling. Specifically, this scheme proposes pruning off the tail at each step like other popular decoding methods like Nucleus sampli...
The paper studies VideoQA through the lenses of causality where it attempts to break the spurious correlations caused by biases (e.g., linguistic biases, visual biases etc.) when predicting answers. It proposes to identify these biases by forcing VideoQA models to respond to unanswerable questions obtained by pairing v...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper studies VideoQA through the lenses of causality where it attempts to break the spurious correlations caused by biases (e.g., linguistic biases, visual biases etc.) when predicting answers. It proposes to identify these biases by forcing VideoQA models to respond to unanswerable questions obtained by p...
The paper proposes an autoencoder architecture whose encoder and decoder are constructed based on the convolutional sparse coding generative model. The encoder is a convolutional sparse coding layer solving lasso (sparse coding problem) by unrolling FISTA (fast iterative thresholding algorithm). The decoder is dictated...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper proposes an autoencoder architecture whose encoder and decoder are constructed based on the convolutional sparse coding generative model. The encoder is a convolutional sparse coding layer solving lasso (sparse coding problem) by unrolling FISTA (fast iterative thresholding algorithm). The decoder is ...
This paper introduces a connection between GFlowNets and variational inference (VI) algorithms for hierarchical variational models (HVMs), demonstrating that special cases of training HVMs via VI are equivalent to training GFlowNets via the trajectory balance (TB) objective. This is interesting because GFlowNets can be...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper introduces a connection between GFlowNets and variational inference (VI) algorithms for hierarchical variational models (HVMs), demonstrating that special cases of training HVMs via VI are equivalent to training GFlowNets via the trajectory balance (TB) objective. This is interesting because GFlowNet...
The paper proposes an algorithm to extract object-centric representations from feature maps inspired by Slot Attention. The main insight is to represent slots using a mixture of Gaussians. This comes from the observation that Slot Attention can be viewed as performing soft k-means clustering on features maps, and mix...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes an algorithm to extract object-centric representations from feature maps inspired by Slot Attention. The main insight is to represent slots using a mixture of Gaussians. This comes from the observation that Slot Attention can be viewed as performing soft k-means clustering on features maps,...
The paper reasons about the relationship between SimCLR, a recent self-supervised learning method, and tSNE, a popular dimensionality technique for data visualization. At the intersection of these two methods, the paper proposed a new method, based on self-supervised learning, for high-dimensional data visualization (t...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper reasons about the relationship between SimCLR, a recent self-supervised learning method, and tSNE, a popular dimensionality technique for data visualization. At the intersection of these two methods, the paper proposed a new method, based on self-supervised learning, for high-dimensional data visualiz...
This paper proposes novel client sampling strategies to accelerate the convergence of federated averaging methods with partial client participation. The idea is to determine the sampling strategies to minimize variance from the worst-case convergence bounds of the methods. Designing optimal client sampling strategies...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes novel client sampling strategies to accelerate the convergence of federated averaging methods with partial client participation. The idea is to determine the sampling strategies to minimize variance from the worst-case convergence bounds of the methods. Designing optimal client sampling st...
This paper studies the contradiction in the existing multimodal robustness literatures. Intuitively, multimodal models are supposed to be more robust than unimodal model as extra redundancy is provided by multimodal data. However, some existing works find that multimodal integration may be more vulnerable to attack, no...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper studies the contradiction in the existing multimodal robustness literatures. Intuitively, multimodal models are supposed to be more robust than unimodal model as extra redundancy is provided by multimodal data. However, some existing works find that multimodal integration may be more vulnerable to at...
This paper designs an Actor-Critic to learn the optimal mix-up strategy for performing Unsupervised Continual Learning (UCL). The corresponding framework, named AUDR, enables adaptively rectifying the update direction in UCL according to its performance under different mixtures of memory data and incoming data. AUDR in...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper designs an Actor-Critic to learn the optimal mix-up strategy for performing Unsupervised Continual Learning (UCL). The corresponding framework, named AUDR, enables adaptively rectifying the update direction in UCL according to its performance under different mixtures of memory data and incoming data....
This paper studies multi-fidelity AutoML algorithms and proposes an improvement over ASHA. Inspired by the observation that there might exist a cross-over point of the learning curves of different configurations where their ranking swaps, the author propose to leverage this instability of ranking to progressively allo...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper studies multi-fidelity AutoML algorithms and proposes an improvement over ASHA. Inspired by the observation that there might exist a cross-over point of the learning curves of different configurations where their ranking swaps, the author propose to leverage this instability of ranking to progressiv...
Update after rebuttal I am very happy to see a very extensive and detailed authors' rebuttal with a number of additional evaluations and results. Most of my questions have been answered and my criticism has been addressed to a sufficient degree. Though there could be some detailed follow-up discussions, I think ove...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: Update after rebuttal I am very happy to see a very extensive and detailed authors' rebuttal with a number of additional evaluations and results. Most of my questions have been answered and my criticism has been addressed to a sufficient degree. Though there could be some detailed follow-up discussions, I t...
This paper considers node features, graph structure, and node labels as three views, and designs three loss components correspondingly. Moreover, the latent $F$ and the model parameters $\Theta$ are proposed to be optimized alternatingly to achieve efficiency and scalability. Strength: 1. The proposed multi-view loss...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper considers node features, graph structure, and node labels as three views, and designs three loss components correspondingly. Moreover, the latent $F$ and the model parameters $\Theta$ are proposed to be optimized alternatingly to achieve efficiency and scalability. Strength: 1. The proposed multi-v...
The paper implements an implementation of the policy iteration algorithm using trained LLM. The strength of this paper lies in using an LLM without any training. This is also a weakness, at its present form, the paper feels not that significant. I enjoyed the paper a lot. The presented utilized an LLM and, without a...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper implements an implementation of the policy iteration algorithm using trained LLM. The strength of this paper lies in using an LLM without any training. This is also a weakness, at its present form, the paper feels not that significant. I enjoyed the paper a lot. The presented utilized an LLM and, w...
This paper focuses on hyper-parameter optimization for the federated learning (FL- HPO) problem which is important because the choice of the HPs can have a dramatic impact on the performance. However, there are still some challenges in this area, including the unavailability of the whole dataset, high communication cos...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper focuses on hyper-parameter optimization for the federated learning (FL- HPO) problem which is important because the choice of the HPs can have a dramatic impact on the performance. However, there are still some challenges in this area, including the unavailability of the whole dataset, high communica...
This paper proposes a framework for Learnable Randomness Injection (LRI) to train inherently interpretable geometric deep-learning models for scientific data. It perturbed data by learnable Bernoulli and Gaussian randomness masks in terms of existence and location importance. Furthermore, four scientific datasets with ...	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 a framework for Learnable Randomness Injection (LRI) to train inherently interpretable geometric deep-learning models for scientific data. It perturbed data by learnable Bernoulli and Gaussian randomness masks in terms of existence and location importance. Furthermore, four scientific datase...
This paper studies the negative impacts of sparse features as a result of Linf adversarial training. The authors find that this sparsity causes the Linf robust models to be more vulnerable to small occlusions and smaller magnitudes of random Gaussian noise in comparison to standard trained models. Strengths: I think t...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper studies the negative impacts of sparse features as a result of Linf adversarial training. The authors find that this sparsity causes the Linf robust models to be more vulnerable to small occlusions and smaller magnitudes of random Gaussian noise in comparison to standard trained models. Strengths: I...
The paper presents comprehensive analyses of self-supervised vision transformers, which provide some new insights about the differences between CL and MIM. For instance, ViTs pre-trained with CL focus on global patterns compared with MIM, collapse into homogeneity while MIM shows more diversity, and reduces the high-fr...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper presents comprehensive analyses of self-supervised vision transformers, which provide some new insights about the differences between CL and MIM. For instance, ViTs pre-trained with CL focus on global patterns compared with MIM, collapse into homogeneity while MIM shows more diversity, and reduces the...
This paper is motivated by the empirical success of retrieval-based classification models and therefore studies the generalization properties of these models. The models that this paper touches on broadly involve models which retrieves similar labeled examples from training data for an instance. The question that this ...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper is motivated by the empirical success of retrieval-based classification models and therefore studies the generalization properties of these models. The models that this paper touches on broadly involve models which retrieves similar labeled examples from training data for an instance. The question th...
This paper introduces a novel method, Mode-Optimized Task Allocation (MOTA), for continual learning with superior backward and forward transfer performance as well as lower average task drift. The paper first identifies the key weakness behind prior regularization-based continual learning methods: they anchor all futu...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper introduces a novel method, Mode-Optimized Task Allocation (MOTA), for continual learning with superior backward and forward transfer performance as well as lower average task drift. The paper first identifies the key weakness behind prior regularization-based continual learning methods: they anchor ...
The authors propose a fully-parametric, deep-learning method for interventional density estimation, called Interventional Normalizing Flows (INFs). INFs provide a properly normalized density estimator. The authors further develop a two-step training procedure with a one-step bias correction for efficient and doubly rob...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The authors propose a fully-parametric, deep-learning method for interventional density estimation, called Interventional Normalizing Flows (INFs). INFs provide a properly normalized density estimator. The authors further develop a two-step training procedure with a one-step bias correction for efficient and do...
This work propose to train TPP in a meta learning framework and treat each sequence as a different task via the neural process. The work defines the context dataset, target input and output for TPP model. Also, the work introduces the cross-attention module to introduce local history matching to learn more informative ...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This work propose to train TPP in a meta learning framework and treat each sequence as a different task via the neural process. The work defines the context dataset, target input and output for TPP model. Also, the work introduces the cross-attention module to introduce local history matching to learn more info...
This work studies the *relationship* between *deep models* and their corresponding *selective prediction and uncertainty estimation performance*. Specifically, this work considers several uncertainty estimation metrics, including AUROC, ECE, AURC and coverage for selective accuracy constraint. Using 523 mod...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This work studies the *relationship* between *deep models* and their corresponding *selective prediction and uncertainty estimation performance*. Specifically, this work considers several uncertainty estimation metrics, including AUROC, ECE, AURC and coverage for selective accuracy constraint. Using...