Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
This paper proposes a way to learn gauge transformer for neural fields applications (especially modeling neural radiance field for novel view synthesis). The main contribution of the work is to propose a unified framework to reason about different NeRF paradigms in the language of gauge transformation and proposes a re...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a way to learn gauge transformer for neural fields applications (especially modeling neural radiance field for novel view synthesis). The main contribution of the work is to propose a unified framework to reason about different NeRF paradigms in the language of gauge transformation and propo...
This paper proposes a method of semi-supervised crowd counting which only leverages a subset of labeled data and learn from the unlabeled data. This method relies on pixel-wise distribution matching and leverage optimal transport in the optimzation process. Strength: Based on Table 1, it seems that this approach achie...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a method of semi-supervised crowd counting which only leverages a subset of labeled data and learn from the unlabeled data. This method relies on pixel-wise distribution matching and leverage optimal transport in the optimzation process. Strength: Based on Table 1, it seems that this approa...
This paper studies the problem of estimating the covariance function for a random process. The random process functional data is sampled on a finite set of knots. Four algorithms are proposed, namely, RANDOM-KNOTS, RANDOM-KNOTS-SPATIAL, B-SPLINE, and B-SPLINE-SPATIAL. These four algorithms are proposed with the main mo...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper studies the problem of estimating the covariance function for a random process. The random process functional data is sampled on a finite set of knots. Four algorithms are proposed, namely, RANDOM-KNOTS, RANDOM-KNOTS-SPATIAL, B-SPLINE, and B-SPLINE-SPATIAL. These four algorithms are proposed with the...
The authors propose a new prompting method, called "Least To Most" prompting. Building upon the success of the "chain-of-thought" prompting method, the authors propose to split the prompting process into two stages: problem reduction, and problem solving. The former focusing on decomposing the task at hand into its con...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose a new prompting method, called "Least To Most" prompting. Building upon the success of the "chain-of-thought" prompting method, the authors propose to split the prompting process into two stages: problem reduction, and problem solving. The former focusing on decomposing the task at hand into...
* Offers an alternative differentiable monotonic alignment generator over EfficientTTS, VITS, FastSpeech 2, non-attentive Tacotron, Parallel Tacotron 2, etc. * Using variational inference, learns a 2-level hierarchical latent representation, which can be used to influence the diversity of speech samples. * Joint-trains...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: * Offers an alternative differentiable monotonic alignment generator over EfficientTTS, VITS, FastSpeech 2, non-attentive Tacotron, Parallel Tacotron 2, etc. * Using variational inference, learns a 2-level hierarchical latent representation, which can be used to influence the diversity of speech samples. * Join...
This work builds a pipeline for closed-loop planning via behavior cloning. Their propose to use only past local context (as opposed to SDV state and context in prior work), and use a transformer to predict future states, which is then turned to actions via LQR controller. They show significant improvements over prior m...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work builds a pipeline for closed-loop planning via behavior cloning. Their propose to use only past local context (as opposed to SDV state and context in prior work), and use a transformer to predict future states, which is then turned to actions via LQR controller. They show significant improvements over...
This paper empirically investigates the phenomena of double descent and offline-online correspondence (the gap between test errors of offline and online learners, as far as I understood) for similarity learning. The authors introduce two types of dataset conversion procedures from data points to pairs and two types of ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper empirically investigates the phenomena of double descent and offline-online correspondence (the gap between test errors of offline and online learners, as far as I understood) for similarity learning. The authors introduce two types of dataset conversion procedures from data points to pairs and two t...
The paper proposes a a data collection strategy combined with human preference capture and a joint objecting learning method to boost the performance of pre-trained dialogue models. The paper is very well-written, contains descriptions of the collected dataset alongwith data analysis, as well as a learning strategy th...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a a data collection strategy combined with human preference capture and a joint objecting learning method to boost the performance of pre-trained dialogue models. The paper is very well-written, contains descriptions of the collected dataset alongwith data analysis, as well as a learning str...
This paper proposes a novel graph convolutional network to address the over-smoothing and heterophily issue. The proposed model is equipped with a trainable depth parameter $d$ and a selection of the spectrum of the graph signal. Based on such modeling, this paper discovers a connection between negative GCN depth and g...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a novel graph convolutional network to address the over-smoothing and heterophily issue. The proposed model is equipped with a trainable depth parameter $d$ and a selection of the spectrum of the graph signal. Based on such modeling, this paper discovers a connection between negative GCN dep...
This work points out an interesting connection between Lovasz theta function and contrastive learning. More concretely, the negative term in the InfoNCE loss corresponds to the Lovasz theta problem, with zero weights. From this connection, the authors are able to extend the InfoNCE loss to a broader class and the exper...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This work points out an interesting connection between Lovasz theta function and contrastive learning. More concretely, the negative term in the InfoNCE loss corresponds to the Lovasz theta problem, with zero weights. From this connection, the authors are able to extend the InfoNCE loss to a broader class and t...
This paper propose BINDER, a training-free neural-symbolic framework that (1) maps the task input to a program using a pre-trained language model, where the model has to decide which part in the input can be converted to a target programming language and corresponding tasks API calls for further extension; (2) adopts a...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper propose BINDER, a training-free neural-symbolic framework that (1) maps the task input to a program using a pre-trained language model, where the model has to decide which part in the input can be converted to a target programming language and corresponding tasks API calls for further extension; (2) ...
The paper studies the benefits of ensembling for self-supervised learning methods. The paper proposes to ensemble the “non-representation” part of the SSL models, and proposes different weighting schemes. Extensive experiments are conducted to analyze and evaluate the proposed ensemble method and weighting schemes, whi...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper studies the benefits of ensembling for self-supervised learning methods. The paper proposes to ensemble the “non-representation” part of the SSL models, and proposes different weighting schemes. Extensive experiments are conducted to analyze and evaluate the proposed ensemble method and weighting sche...
This paper focuses on convergence analysis of adversarial training for finite-width two-layer Leaky ReLU networks. Convergence guarantees are given for both inner-loop PGD-style adversarial attack (in deterministic form) and end-to-end adversarial training with imperfect attack (in expectation form). To ease the analys...	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 focuses on convergence analysis of adversarial training for finite-width two-layer Leaky ReLU networks. Convergence guarantees are given for both inner-loop PGD-style adversarial attack (in deterministic form) and end-to-end adversarial training with imperfect attack (in expectation form). To ease th...
This paper introduces cyclophobic reinforcement learning, a negative intrinsic reward for exploration that incentivizes the agent not to revisit previously visited states. This intrinsic reward is computed over different cropped views of the environment by the agent and these are combined to compute the final reward. T...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper introduces cyclophobic reinforcement learning, a negative intrinsic reward for exploration that incentivizes the agent not to revisit previously visited states. This intrinsic reward is computed over different cropped views of the environment by the agent and these are combined to compute the final r...
The current paper under submission studies linear regression and gives an algorithm that is "more practical" than the previous approaches. It is based on the approach of Alabi et al. to a large degree and uses known techniques in DP literature to achieve its objective. The algorithm is more efficient, does not require ...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The current paper under submission studies linear regression and gives an algorithm that is "more practical" than the previous approaches. It is based on the approach of Alabi et al. to a large degree and uses known techniques in DP literature to achieve its objective. The algorithm is more efficient, does not ...
This paper proposes incorporating model-specific prior knowledge into optimizers by modifying the gradients according to a set of model-specific hyper-parameters. The proposed methods are employed on VGG architectures using SGD and AdamW optimizers for image classification and segmentation datasets. The paper introduce...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes incorporating model-specific prior knowledge into optimizers by modifying the gradients according to a set of model-specific hyper-parameters. The proposed methods are employed on VGG architectures using SGD and AdamW optimizers for image classification and segmentation datasets. The paper i...
This paper proposes a deep reinforcement learning method for partially observable environments named Deep Transformer Q-Networks (DTQN). The key feature of DTQN is that, for a given chunk of history sampled from the replay buffer, a loss is taken over predicted Q-values for all timesteps in the history (rather than j...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a deep reinforcement learning method for partially observable environments named Deep Transformer Q-Networks (DTQN). The key feature of DTQN is that, for a given chunk of history sampled from the replay buffer, a loss is taken over predicted Q-values for all timesteps in the history (rathe...
The paper tackles the problem of statistical inferences from differential private synthetic datasets. The authors use the tools of measurement error models to derive an estimator that is bias-corrected. Experiments are performed that compare the proposed estimator to SSP (sufficient statistic perturbation), which show ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper tackles the problem of statistical inferences from differential private synthetic datasets. The authors use the tools of measurement error models to derive an estimator that is bias-corrected. Experiments are performed that compare the proposed estimator to SSP (sufficient statistic perturbation), whi...
In this work the authors present a data poisoning attack that inserts class-conditional triggers in the data augmentation stage. The threat model the authors consider is: * The defender is oblivious and uses the attacker's malicious data augmentation pipeline * The attacker has arbitrary access to the data augmentati...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this work the authors present a data poisoning attack that inserts class-conditional triggers in the data augmentation stage. The threat model the authors consider is: * The defender is oblivious and uses the attacker's malicious data augmentation pipeline * The attacker has arbitrary access to the data au...
This paper provides new insight to understand GANs based on log density ratios as a gradient flow in the Wasserstein space. Specifically, the bi-level step of adversarial training in GANs is regarded as it first estimates the vector field of the gradient flow (by discriminator), and next the generator is updated to lea...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper provides new insight to understand GANs based on log density ratios as a gradient flow in the Wasserstein space. Specifically, the bi-level step of adversarial training in GANs is regarded as it first estimates the vector field of the gradient flow (by discriminator), and next the generator is update...
This paper presents a new offline-to-online finetuning approach that proposes to maintain both the policy learned offline and the new online policy during online-finetuning. The authors first train the offline policy and Q-function, select the policy from a categorical distribution with the learned Q-values during onli...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents a new offline-to-online finetuning approach that proposes to maintain both the policy learned offline and the new online policy during online-finetuning. The authors first train the offline policy and Q-function, select the policy from a categorical distribution with the learned Q-values dur...
This paper addresses the problem of image copy detection through image retrieval and proposes a method that optimizes jointly both representation learning and approximate similarity search. Large databases rely on indexes for approximate and efficient search, but the challenge comes when copies are significantly modifi...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper addresses the problem of image copy detection through image retrieval and proposes a method that optimizes jointly both representation learning and approximate similarity search. Large databases rely on indexes for approximate and efficient search, but the challenge comes when copies are significantl...
This work introduces a new continual learning method TST. TST considers both preventing catastrophic forgetting (CF) and encouraging knowledge transfer (KT). To prevent CF, TST directly use one existing CL method called SupSup. To encourage KT, TST borrows the idea of gradient-based importance score and applies this sc...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work introduces a new continual learning method TST. TST considers both preventing catastrophic forgetting (CF) and encouraging knowledge transfer (KT). To prevent CF, TST directly use one existing CL method called SupSup. To encourage KT, TST borrows the idea of gradient-based importance score and applies...
The authors consider the problem of reducing the size of hidden layers in convolutional neural networks (CNN). Size reduction is usually achieved by setting an appropriate stride for the convolutional layers. The stride is a hyperparameter and is usually fine-tuned using cross-validation. An alternative approach is Dif...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors consider the problem of reducing the size of hidden layers in convolutional neural networks (CNN). Size reduction is usually achieved by setting an appropriate stride for the convolutional layers. The stride is a hyperparameter and is usually fine-tuned using cross-validation. An alternative approac...
This paper introduces a new benchmark (SciRepEval) for scientific representation learning consisting of 25 tasks in 4 formats (classification, regression, ranking, and search). It shows that learning a separate document representation for each task format would improve the task performance compared to learning a single...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a new benchmark (SciRepEval) for scientific representation learning consisting of 25 tasks in 4 formats (classification, regression, ranking, and search). It shows that learning a separate document representation for each task format would improve the task performance compared to learning ...
This paper proposes AIM (Adapting Image Models) for video classification with the focus on reducing tunable parameter and training (fine-tuning) time. The main contributions are: 1) using Adapter (Houlsby et al. 2019) to adapt image models (pretrained on large datasets such as ImageNet-21k or CLIP); 2) designing variou...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes AIM (Adapting Image Models) for video classification with the focus on reducing tunable parameter and training (fine-tuning) time. The main contributions are: 1) using Adapter (Houlsby et al. 2019) to adapt image models (pretrained on large datasets such as ImageNet-21k or CLIP); 2) designin...
This paper proposes a new training strategy to improve the performance of DETR-based object detectors. The authors indicate the main reason for the low convergence rate of DETR-based object detectors is in the one-to-one matching and propose a simple strategy of group-wise one-to-many assignment to address. It is a sim...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new training strategy to improve the performance of DETR-based object detectors. The authors indicate the main reason for the low convergence rate of DETR-based object detectors is in the one-to-one matching and propose a simple strategy of group-wise one-to-many assignment to address. It ...
Multi-vector retrieval models can fetch improved results on certain retrieval tasks, while at the same time the approach leads to large search indexes and expensive computation. Thus, this paper proposes ALIGNER, a novel multi-vector retrieval model that learns sparse pairwise alignments between query and document toke...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Multi-vector retrieval models can fetch improved results on certain retrieval tasks, while at the same time the approach leads to large search indexes and expensive computation. Thus, this paper proposes ALIGNER, a novel multi-vector retrieval model that learns sparse pairwise alignments between query and docum...
This paper mainly addresses the following two problems: 1. It introduces an objective function for training generative autoencoders (encoder-decoder networks). The training objective consists of two functions: one represents the reconstruction error, and the other indicates the discrepancy of the latent code distribut...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper mainly addresses the following two problems: 1. It introduces an objective function for training generative autoencoders (encoder-decoder networks). The training objective consists of two functions: one represents the reconstruction error, and the other indicates the discrepancy of the latent code d...
This paper can be regarded as a follow-up of (Ergun et al. 2021) which is published in ICLR2022. The main contribution of this submission is to answer the question: "Is it possible to design a $k$-means and a $k$-medians algorithm that achieve (1 + $\alpha$)-approximate clustering when the predictor is not very accurat...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper can be regarded as a follow-up of (Ergun et al. 2021) which is published in ICLR2022. The main contribution of this submission is to answer the question: "Is it possible to design a $k$-means and a $k$-medians algorithm that achieve (1 + $\alpha$)-approximate clustering when the predictor is not very...
The paper introduces a differentiable rendering algoithm that converts 3D meshes into a collection of 3D gaussians and ray traces the resulting scene. The conversions to volumetric gaussians smoothes the visibility function of the scene, similar to the smooth visibility and blending functions of soft differentiable ras...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper introduces a differentiable rendering algoithm that converts 3D meshes into a collection of 3D gaussians and ray traces the resulting scene. The conversions to volumetric gaussians smoothes the visibility function of the scene, similar to the smooth visibility and blending functions of soft differenti...
This paper considers the robustness of deep classification models, noting that out-of-the-box models can yield arbitrarily high-confidence predictions on out-of-distribution data. As a proposed remedy, “total activation classifiers” (TAC) are introduced to capture “class-dependent patterns” which are claimed to offer a...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper considers the robustness of deep classification models, noting that out-of-the-box models can yield arbitrarily high-confidence predictions on out-of-distribution data. As a proposed remedy, “total activation classifiers” (TAC) are introduced to capture “class-dependent patterns” which are claimed to...
Message-passing neural networks (MPNNs) and their extensions (Subgraph MPNNs) have inspired state-of-the-art models for graph data; however, their counting power is still limited and it is known that Subgraph GNNs cannot count cycles with more than $4$ nodes. To boost the counting power, this paper extends the idea of...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: Message-passing neural networks (MPNNs) and their extensions (Subgraph MPNNs) have inspired state-of-the-art models for graph data; however, their counting power is still limited and it is known that Subgraph GNNs cannot count cycles with more than $4$ nodes. To boost the counting power, this paper extends the...
Tasks: Automatic speech recognition (ASR, measure WER), language modeling (measure PPL) Two new models are proposed, which are based on state space models (SSMs). The newly proposed models are: MSSM and Stateformer. The MSSM purely uses the SSM, with residual connections and some layer norm. The Stateformer is like a...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Tasks: Automatic speech recognition (ASR, measure WER), language modeling (measure PPL) Two new models are proposed, which are based on state space models (SSMs). The newly proposed models are: MSSM and Stateformer. The MSSM purely uses the SSM, with residual connections and some layer norm. The Stateformer i...
This paper proposed an energy-based out-of-distribution detection for graph neural network. The contributions include introducing energy-based method to graph out-of-distribution detection, providing some theoretical analysis, presenting energy propagation, and conducting a lot of experiments. Strength: 1. The paper is...	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 proposed an energy-based out-of-distribution detection for graph neural network. The contributions include introducing energy-based method to graph out-of-distribution detection, providing some theoretical analysis, presenting energy propagation, and conducting a lot of experiments. Strength: 1. The ...
The authors propose a method for learning when there is high amount of label noise present, using an iterative bi-level optimization. Their method in an iterative fashion first trains a classifier (predictor), then defines a pseudo-labels for the selector (abstain) classifier using confidence of the trained predictor o...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors propose a method for learning when there is high amount of label noise present, using an iterative bi-level optimization. Their method in an iterative fashion first trains a classifier (predictor), then defines a pseudo-labels for the selector (abstain) classifier using confidence of the trained pre...
The paper provides a reinforcement learning method for subgraph matching problem. Given a query graph, they attempt to find the subgraph from a large graph that matches a query graph. Then the paper compare the underlying method against several baselines on real world datasets. Strengths: + Important problem + Intuitiv...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper provides a reinforcement learning method for subgraph matching problem. Given a query graph, they attempt to find the subgraph from a large graph that matches a query graph. Then the paper compare the underlying method against several baselines on real world datasets. Strengths: + Important problem + ...
The paper introduces a new set of benchmarks dubbed "SMC-Bench" (Sparsity May Cry). The authors correctly remark that the ever-growing research in sparsifying neural networks is often (if not always) evaluated on limited set of benchmarks which are often one of the very well understood, and "relatively easy" datasets. ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper introduces a new set of benchmarks dubbed "SMC-Bench" (Sparsity May Cry). The authors correctly remark that the ever-growing research in sparsifying neural networks is often (if not always) evaluated on limited set of benchmarks which are often one of the very well understood, and "relatively easy" da...
The paper utilizes a method called Adversarial Representation Learning (ARL) as a base. ARL learns a privacy-preserving encoding of sensitive user data before it is shared, but lacks formal guarantees. The authors link the local Lipschitz constant of a neural network in ARL (obfuscation layer) with its local sensitivi...	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 utilizes a method called Adversarial Representation Learning (ARL) as a base. ARL learns a privacy-preserving encoding of sensitive user data before it is shared, but lacks formal guarantees. The authors link the local Lipschitz constant of a neural network in ARL (obfuscation layer) with its local s...
This paper proposes a novel approach to pessimism-based offline RL. Instead of the standard practice of explicitly constructing a lower confidence bound for value functions, this new approach uses perturbed rewards to implicitly quantify the training uncertainty and construct lower confidence bounds for the ground trut...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a novel approach to pessimism-based offline RL. Instead of the standard practice of explicitly constructing a lower confidence bound for value functions, this new approach uses perturbed rewards to implicitly quantify the training uncertainty and construct lower confidence bounds for the gro...