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
To tackle the problem of lacking sufficient ground-truth labels in practical applications, this paper proposed a paradigm to fuse the generative model and programmatic weak supervision method, which are often used separately to tackle this issue in previous literatures. The fusion is mainly achieved by encouraging an a...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: To tackle the problem of lacking sufficient ground-truth labels in practical applications, this paper proposed a paradigm to fuse the generative model and programmatic weak supervision method, which are often used separately to tackle this issue in previous literatures. The fusion is mainly achieved by encourag...
Authors present a de-noising method using auto-encoders for time series data that does not need clean data to be trained. However, the method requires an auxiliary clean signal that is related to the target signal. The method also makes assumptions about the SNR level which has to be high. To alleviate this limitation,...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: Authors present a de-noising method using auto-encoders for time series data that does not need clean data to be trained. However, the method requires an auxiliary clean signal that is related to the target signal. The method also makes assumptions about the SNR level which has to be high. To alleviate this lim...
- The paper explores hyperbolic self paced learning for self supervised learning of action representations. - The key idea is to use hyperbolic geometry to automatically order the training samples from easy to hard. This determines the learning pace thus improving learning. - The framework is borrowed from BYOL, whic...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: - The paper explores hyperbolic self paced learning for self supervised learning of action representations. - The key idea is to use hyperbolic geometry to automatically order the training samples from easy to hard. This determines the learning pace thus improving learning. - The framework is borrowed from BY...
This paper explores the nullspace ( adding an element on this space does not affect the output) for Vision Transformer. The authors first demonstrate that a non-trivial nullspace exists for the patch embedding matrices. This idea is extended to the non-linear layers of the vision transformer, which is learned via simpl...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper explores the nullspace ( adding an element on this space does not affect the output) for Vision Transformer. The authors first demonstrate that a non-trivial nullspace exists for the patch embedding matrices. This idea is extended to the non-linear layers of the vision transformer, which is learned v...
The paper proposes a permutation invariant conditional normalizing flows based on a continuous version of graph normalizing flows. The proposed model is evaluated on the tasks of generating realistic traffic scenes and bounding box prediction. Strengths: 1. The paper is well organized and well written. 2. Experiment r...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: The paper proposes a permutation invariant conditional normalizing flows based on a continuous version of graph normalizing flows. The proposed model is evaluated on the tasks of generating realistic traffic scenes and bounding box prediction. Strengths: 1. The paper is well organized and well written. 2. Expe...
This paper presents a discrete representation learning strategy for audio sequences designed primarily for retrieval. The learning strategy encourages representations to be similar for pairs of similar inputs, where pairs are created by data augmentation of the input audio. The authors compare their approach on two ret...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper presents a discrete representation learning strategy for audio sequences designed primarily for retrieval. The learning strategy encourages representations to be similar for pairs of similar inputs, where pairs are created by data augmentation of the input audio. The authors compare their approach on...
This paper assumes there exist Pareto Subspaces, i.e., weight subspaces where multiple optimal functional solutions lie, and develops a weight-ensembling method named Pareto Manifold Learning that casts multi-task problems as learning an ensemble of single-task predictors by interpolating among members during training....	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper assumes there exist Pareto Subspaces, i.e., weight subspaces where multiple optimal functional solutions lie, and develops a weight-ensembling method named Pareto Manifold Learning that casts multi-task problems as learning an ensemble of single-task predictors by interpolating among members during t...
This paper defines GCN-inspired kernels by deriving GP as the limit of GCN. The authors present some theoretical results such as universality and develop a scalable posterior inference method based on low-rank approximation of covariance matrix. The effectiveness of the proposed method is demonstrated using multiple be...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper defines GCN-inspired kernels by deriving GP as the limit of GCN. The authors present some theoretical results such as universality and develop a scalable posterior inference method based on low-rank approximation of covariance matrix. The effectiveness of the proposed method is demonstrated using mul...
Goal: creating specialised tools for different manipulation goals. Proposed Solution: A designer policy that is conditioned on the task goal + design agnostic controller policy that can perform manipulation using the designed tools. This paradigm has already been proposed in recent work (Yuan et al.) the authors have ...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: Goal: creating specialised tools for different manipulation goals. Proposed Solution: A designer policy that is conditioned on the task goal + design agnostic controller policy that can perform manipulation using the designed tools. This paradigm has already been proposed in recent work (Yuan et al.) the autho...
The paper proposes a weight shared training framework for FL that maintains a primary network at the server and distributes subnetworks to clients. The authors propose algorithms for efficient selection of the subnetworks and aggregation. This allows efficient federated learning and deployment of the models. Strength...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a weight shared training framework for FL that maintains a primary network at the server and distributes subnetworks to clients. The authors propose algorithms for efficient selection of the subnetworks and aggregation. This allows efficient federated learning and deployment of the models. ...
This paper proposes an approach for uncertainty estimation in free-form text generation called "semantic entropy". The method proposed in this paper is unsupervised and can directly be applied to off-the-shelf language models. Semantic entropy provides a better uncertainty estimation than standard entropy, scales bette...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes an approach for uncertainty estimation in free-form text generation called "semantic entropy". The method proposed in this paper is unsupervised and can directly be applied to off-the-shelf language models. Semantic entropy provides a better uncertainty estimation than standard entropy, scal...
This paper tackles the communication bottleneck in the scalability of federated learning. A linearity constraint on the decompression operator is introduced, together with some popular scalar quantizations to to reconcile security and communication efficiency. The proposed algorithm reports the best results on LEAF ben...	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 tackles the communication bottleneck in the scalability of federated learning. A linearity constraint on the decompression operator is introduced, together with some popular scalar quantizations to to reconcile security and communication efficiency. The proposed algorithm reports the best results on ...
This paper proposes to first recognize influential visual parameters by its averaged gradients and then tune those parameters. In order to improve adaptation capability, the paper also propose to switch to a low-rank reparameterization of the weight matrix that contains enough sensitive parameters. The paper evaluated ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes to first recognize influential visual parameters by its averaged gradients and then tune those parameters. In order to improve adaptation capability, the paper also propose to switch to a low-rank reparameterization of the weight matrix that contains enough sensitive parameters. The paper ev...
The author studies the batchnorm's negative effect when training with quantized gradient. To address the problem, the author proposes to add an additional loss to control the variance of batchnorm. Empirical results show improvement when quantizing weight, activation and gradient. Strength: 1. The observation of the ba...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The author studies the batchnorm's negative effect when training with quantized gradient. To address the problem, the author proposes to add an additional loss to control the variance of batchnorm. Empirical results show improvement when quantizing weight, activation and gradient. Strength: 1. The observation o...
The paper proposed a non-Gaussian process regression which allows to model GPs with time changes and thus accounts for non-Gaussian behaviours such as heavy tails. The model is constructed using a latent transformed input space, and the Levy process is used to model the random evolution of the latent transformation. Fu...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper proposed a non-Gaussian process regression which allows to model GPs with time changes and thus accounts for non-Gaussian behaviours such as heavy tails. The model is constructed using a latent transformed input space, and the Levy process is used to model the random evolution of the latent transforma...
This paper considers the problem of inserting combinatorial solvers into neural network architectures. Since combinatorial solvers have discrete outputs the gradient is zero or non-existent, meaning that an alternative "gradient" must be used on the backwards pass in order to provide directional information to update t...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper considers the problem of inserting combinatorial solvers into neural network architectures. Since combinatorial solvers have discrete outputs the gradient is zero or non-existent, meaning that an alternative "gradient" must be used on the backwards pass in order to provide directional information to ...
The paper studies the fairness attack and defense problem. The authors propose a black-box attack against fair clustering algorithms that works by perturbing a small percentage of samples’ protected group memberships. They also proposed a defense algorithm, named Consensus Fair Clustering (CFC), that utilizes consensus...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper studies the fairness attack and defense problem. The authors propose a black-box attack against fair clustering algorithms that works by perturbing a small percentage of samples’ protected group memberships. They also proposed a defense algorithm, named Consensus Fair Clustering (CFC), that utilizes c...
The authors propose a new solution for the task of Sketch healing, where corrupted or unfinished sketches are reconstructed without the corruption. This task is commonly accomplished through representation learning. The manuscript describes a method focused on ‘stable representations’, where complete and corrupted sket...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The authors propose a new solution for the task of Sketch healing, where corrupted or unfinished sketches are reconstructed without the corruption. This task is commonly accomplished through representation learning. The manuscript describes a method focused on ‘stable representations’, where complete and corrup...
The paper proposes a method to mitigate the problem of over-conservative policy when optimizing under the worst-case state adversary. Under certain conditions, it proves a tighter bound of the average-case return and worst-case return under state adversary. A two-level update algorithm is proposed to automatically tune...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a method to mitigate the problem of over-conservative policy when optimizing under the worst-case state adversary. Under certain conditions, it proves a tighter bound of the average-case return and worst-case return under state adversary. A two-level update algorithm is proposed to automatica...
This work proposes INR-GAN-based text-to-image generation method, where model weights are modulated with the additional module called hypernetwork and RGB value is predicted given pixel coordinate as input. To make a model to reflect the textual condition better, the authors propose a new method of generating modulatin...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This work proposes INR-GAN-based text-to-image generation method, where model weights are modulated with the additional module called hypernetwork and RGB value is predicted given pixel coordinate as input. To make a model to reflect the textual condition better, the authors propose a new method of generating m...
This paper presents a method using normalizing flows for offline reinforcement learning (ORL). To alleviate the difference between training and testing datasets, the proposed method maps uniform distributions to action spaces using the flow model. Experimental results present that the proposed method outperforms the st...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents a method using normalizing flows for offline reinforcement learning (ORL). To alleviate the difference between training and testing datasets, the proposed method maps uniform distributions to action spaces using the flow model. Experimental results present that the proposed method outperform...
This work presents a new way to perform multi-step logical reasoning tasks with frozen pre-trained large language models. In particular they first select a few (most often two?) pieces of information (Selection) by scoring each fact in the context with the model’s log-likelihood. Then the retrieved facts are combined a...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This work presents a new way to perform multi-step logical reasoning tasks with frozen pre-trained large language models. In particular they first select a few (most often two?) pieces of information (Selection) by scoring each fact in the context with the model’s log-likelihood. Then the retrieved facts are co...
This paper introduces DI-Nets, a formulation of neural networks that is invariant to discretization of the input space. More specifically, the authors consider building models that take neural fields as input, and unlike other recent work, do not depend on the specific parameterization of that neural field, which is de...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces DI-Nets, a formulation of neural networks that is invariant to discretization of the input space. More specifically, the authors consider building models that take neural fields as input, and unlike other recent work, do not depend on the specific parameterization of that neural field, whi...
The paper proposes building a generative LM (HSN) using a VAE framework to encode sentences into symbols which encode different aspects of language (like topics and sentiments). Using these discrete symbols, the authors claim that we can achieve compositionality of sentence encodings. The proof of concept is demonstra...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes building a generative LM (HSN) using a VAE framework to encode sentences into symbols which encode different aspects of language (like topics and sentiments). Using these discrete symbols, the authors claim that we can achieve compositionality of sentence encodings. The proof of concept is d...
The paper addresses the inference sub-optimality issue of neural processes. To achieve competitive performance of NPs, the paper proposes a surrogate objective of the target log-likelihood in a meta learning setup, and introduces a tractable way to optimize the surrogate objective through variational expectation maximi...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper addresses the inference sub-optimality issue of neural processes. To achieve competitive performance of NPs, the paper proposes a surrogate objective of the target log-likelihood in a meta learning setup, and introduces a tractable way to optimize the surrogate objective through variational expectatio...
In this work, prompt learning is reexamined, and several unexpected findings that defy accepted notions of the prompt are presented. First , random prompts without learning or fine-grained design may likewise function effectively in zero-shot recognition. Second, direct linear classifier fine-tuning performs more effec...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this work, prompt learning is reexamined, and several unexpected findings that defy accepted notions of the prompt are presented. First , random prompts without learning or fine-grained design may likewise function effectively in zero-shot recognition. Second, direct linear classifier fine-tuning performs mo...
The paper focuses on applying spiking neural networks (SNNs) to text classification tasks. Spiking neural networks more closely mimic biological neural networks, with the main difference from typical artificial neural network being a temporal quality: each neuron accumulates a membrane charge over time (across infere...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper focuses on applying spiking neural networks (SNNs) to text classification tasks. Spiking neural networks more closely mimic biological neural networks, with the main difference from typical artificial neural network being a temporal quality: each neuron accumulates a membrane charge over time (acros...
The paper investigates OOD detection, with different ensembles amd dropout methods, of policy and value networks in PPO. Experiments report RL-performance and OOD-ROC curves on 4 Mujoco and 4 Atari benchmarks. All methods decreased perrformance in (almost all) environments, but Masksembles have been reported to work so...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper investigates OOD detection, with different ensembles amd dropout methods, of policy and value networks in PPO. Experiments report RL-performance and OOD-ROC curves on 4 Mujoco and 4 Atari benchmarks. All methods decreased perrformance in (almost all) environments, but Masksembles have been reported to...
This paper provides an approach to pre-train language models by infusing them with logic information to enhance their performance on logical reasoning. They propose to use a "fact" which is extracted from the text itself and propose to use three pre-training tasks 1) Logical connective masking, 2) Logical structure com...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper provides an approach to pre-train language models by infusing them with logic information to enhance their performance on logical reasoning. They propose to use a "fact" which is extracted from the text itself and propose to use three pre-training tasks 1) Logical connective masking, 2) Logical struc...
The authors propose a method for training a classifier on data with Instance-Dependent Noisy Labels (IDN Labels, i.e. where label noise is a function of the instances themselves). Their technique is within a class of techniques that monitor a heuristic during training that measures some notion of noise in labels, and ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors propose a method for training a classifier on data with Instance-Dependent Noisy Labels (IDN Labels, i.e. where label noise is a function of the instances themselves). Their technique is within a class of techniques that monitor a heuristic during training that measures some notion of noise in labe...
This paper proposes a method for unsupervised domain adaptation wherein pseudo-labeling, teacher-student approaches, augmentation (mix-up) and MMD based alignment are put together in one framework. Given a target sample and source sample, a mix-up sample is first constructed. The label for the mix-up sample is created ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a method for unsupervised domain adaptation wherein pseudo-labeling, teacher-student approaches, augmentation (mix-up) and MMD based alignment are put together in one framework. Given a target sample and source sample, a mix-up sample is first constructed. The label for the mix-up sample is ...
The authors proposed a new variational inference algorithm for settings where the parameters of the approximate posterior form a Riemannian manifold. Unlike previous methods that relied on various forms approximations, the proposed method is based on an exact formula for the gradient. STRENGTHS: * The paper provides a ...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The authors proposed a new variational inference algorithm for settings where the parameters of the approximate posterior form a Riemannian manifold. Unlike previous methods that relied on various forms approximations, the proposed method is based on an exact formula for the gradient. STRENGTHS: * The paper pro...
This paper propose a new way to combine knowledge distillation with contastive learning, which updates the encoders in a momentum manner to obtain efficient and consistent encoding. Experiments on both classification and detection have been conducted. Strength: 1. Sufficient experiments have been conducted on both clas...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper propose a new way to combine knowledge distillation with contastive learning, which updates the encoders in a momentum manner to obtain efficient and consistent encoding. Experiments on both classification and detection have been conducted. Strength: 1. Sufficient experiments have been conducted on b...
This paper shows the convergence of specific first-order gradient-based algorithms on deep linear networks with Bures-Wasserstein loss, under the assumption that the initial weights are balanced. The author studies both the original non-smooth Bures-Wasserstein loss and a smooth perturbative one. They characterize the ...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper shows the convergence of specific first-order gradient-based algorithms on deep linear networks with Bures-Wasserstein loss, under the assumption that the initial weights are balanced. The author studies both the original non-smooth Bures-Wasserstein loss and a smooth perturbative one. They character...
The paper investigates the generalization of L2O on unseen test cases that substantially differ from the training distribution. It proposes a self-adapted L2O algorithm (M-L2O) incorporated with meta-adaptation. The generalization advantages of M-L2O over out-of-distribution tasks have been theoretically and empiricall...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper investigates the generalization of L2O on unseen test cases that substantially differ from the training distribution. It proposes a self-adapted L2O algorithm (M-L2O) incorporated with meta-adaptation. The generalization advantages of M-L2O over out-of-distribution tasks have been theoretically and em...
The authors study the task of training regression models with the guarantee of label differential privacy (DP). Based on a global prior distribution on label values, which could be obtained privately, they derive a label DP randomization mechanism that is optimal under a given regression loss function. Strength: The au...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors study the task of training regression models with the guarantee of label differential privacy (DP). Based on a global prior distribution on label values, which could be obtained privately, they derive a label DP randomization mechanism that is optimal under a given regression loss function. Strength...
This paper proposes the prior knowledge mechanism for feature distillation, which can fully excavate the distillation potential of big models. Furthermore, the authors propose the dynamic prior knowledge (DPK) to solve the ‘larger models are not always better teachers’ issue, which makes the performance of the student ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes the prior knowledge mechanism for feature distillation, which can fully excavate the distillation potential of big models. Furthermore, the authors propose the dynamic prior knowledge (DPK) to solve the ‘larger models are not always better teachers’ issue, which makes the performance of the ...
This paper proposes a new point cloud registration method considering an unbalanced pair of point clouds. The main assumption is that one of them contains much fewer points than the other and represents only a part of the scene in interest. There are several spatial levels as well as the corresponding matching modules ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new point cloud registration method considering an unbalanced pair of point clouds. The main assumption is that one of them contains much fewer points than the other and represents only a part of the scene in interest. There are several spatial levels as well as the corresponding matching ...
The paper proposes RTD-AE, a topology-based autoencoder that aims to preserve topological information from the data space in latent embeddings. For this, RTD-AE leverages Representation Topology Divergence (RTD), a technique that allows to measure the similarity of VR complexes between two point clouds, here the data s...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes RTD-AE, a topology-based autoencoder that aims to preserve topological information from the data space in latent embeddings. For this, RTD-AE leverages Representation Topology Divergence (RTD), a technique that allows to measure the similarity of VR complexes between two point clouds, here th...
In this work, the authors developed a clever twist to the traditional neural language modeling paradigm by building a model that operates on image patches of the text. It learns to output “masked” patches using an MSE loss. Through this, the model alleviates issues with fixed or non-transferable vocabulary. When fine-t...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this work, the authors developed a clever twist to the traditional neural language modeling paradigm by building a model that operates on image patches of the text. It learns to output “masked” patches using an MSE loss. Through this, the model alleviates issues with fixed or non-transferable vocabulary. Whe...