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 authors study the phenomenon of neural collapse (NC) under several variants of the layer-peeled model. Since features from modern networks are the outcome of some non-negative activation functions, such as ReLU, the paper first considers the case of non-negative features and shows that label smoothing also produce...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors study the phenomenon of neural collapse (NC) under several variants of the layer-peeled model. Since features from modern networks are the outcome of some non-negative activation functions, such as ReLU, the paper first considers the case of non-negative features and shows that label smoothing also...
The paper presents a Maximum Entropy Information Bottleneck (MEIB), which is a different take on the information bottleneck compared to the well-known variational information bottleneck. I liked very much the idea, both the simplicity and the motivation behind it. I also found very intriguing the property of MEIB be...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper presents a Maximum Entropy Information Bottleneck (MEIB), which is a different take on the information bottleneck compared to the well-known variational information bottleneck. I liked very much the idea, both the simplicity and the motivation behind it. I also found very intriguing the property of...
This paper aims to verify the existence of vocoder fingerprints by training a classifier to identify the sources of generated audio waveforms. The authors analyzed the distinguishability of different vocoders from four aspects: (1) vocoders with different architectures, (2) vocoders trained on different datasets, (3) v...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper aims to verify the existence of vocoder fingerprints by training a classifier to identify the sources of generated audio waveforms. The authors analyzed the distinguishability of different vocoders from four aspects: (1) vocoders with different architectures, (2) vocoders trained on different dataset...
This paper focuses on few-shot anomaly detection where only a few normal samples are available in training. The new method leverages contrastive learning to transfer pre-trained from the source domain to target domains supported by few-shot samples. In addition, the instance positive pair loss is used to tight up norma...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper focuses on few-shot anomaly detection where only a few normal samples are available in training. The new method leverages contrastive learning to transfer pre-trained from the source domain to target domains supported by few-shot samples. In addition, the instance positive pair loss is used to tight ...
This paper is about multivariate time series forecasting with structure learning. Existing works usually assume the graph structure of the multiple time series is given or learned by the node similarity. However, in some applications, the relationship between time series can be much more complicated and the graph stru...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper is about multivariate time series forecasting with structure learning. Existing works usually assume the graph structure of the multiple time series is given or learned by the node similarity. However, in some applications, the relationship between time series can be much more complicated and the gr...
This draft considers the problem of online learning with label shift in the presence of additional conditional shift. In addition to the change in the class-priors $\Pr[y]$, the posterior probability $\Pr[x\|y]$ can also change over time. Based on the previous work of Wu et al. (2021), the authors propose three heuristi...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This draft considers the problem of online learning with label shift in the presence of additional conditional shift. In addition to the change in the class-priors $\Pr[y]$, the posterior probability $\Pr[x\|y]$ can also change over time. Based on the previous work of Wu et al. (2021), the authors propose three ...
This work focused on adversarial MDPs with switching costs. The author proposed the SEEDs algorithms and provided a $O(T^{2/3})$ regret guarantee for both known and unknown transitions. In addition, the author also provided a theoretical guarantee for the lower regret bound, which suggests the proposed SEEDs algorithms...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work focused on adversarial MDPs with switching costs. The author proposed the SEEDs algorithms and provided a $O(T^{2/3})$ regret guarantee for both known and unknown transitions. In addition, the author also provided a theoretical guarantee for the lower regret bound, which suggests the proposed SEEDs al...
This work proposes to replace hand-written prompts used in CLIP style models with GPT-generated prompts for zero-shot image classification. The authors show that their approach (CuPL) of using LLMs to generate diverse text prompts, can improve classification accuracy while removing the need for hand-crafted prompts. St...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work proposes to replace hand-written prompts used in CLIP style models with GPT-generated prompts for zero-shot image classification. The authors show that their approach (CuPL) of using LLMs to generate diverse text prompts, can improve classification accuracy while removing the need for hand-crafted pro...
The work focuses on learning representation for offline reinforcement learning. The authors propose a representation learning method based on cloning the behaviour under which the offline dataset was collected. The main motivations of the work is to (1) reduce the difficulty of learning a policy from limited high-dimen...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The work focuses on learning representation for offline reinforcement learning. The authors propose a representation learning method based on cloning the behaviour under which the offline dataset was collected. The main motivations of the work is to (1) reduce the difficulty of learning a policy from limited hi...
The authors are interested in the long-term engagement of users in recommender systems, which is less studied than immediate engagement. To optimize this long-term engagement in sequential recommendation, they resort reinforcement learning (RL). However, directly applying RL methods to recommender systems can be expens...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors are interested in the long-term engagement of users in recommender systems, which is less studied than immediate engagement. To optimize this long-term engagement in sequential recommendation, they resort reinforcement learning (RL). However, directly applying RL methods to recommender systems can b...
This paper proposes a self-supervised approach for training a speech enhancement network, given data from an inertial measurement unit (IMU) and audio recorded from an earphone microphone. The IMU captures a nonlinear and lower-frequency version of the user's speech through the bone condution channel. The proposed appr...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a self-supervised approach for training a speech enhancement network, given data from an inertial measurement unit (IMU) and audio recorded from an earphone microphone. The IMU captures a nonlinear and lower-frequency version of the user's speech through the bone condution channel. The propo...
This paper proposes a new knowledge graph reasoning (KGR) model that combines embedding-based and rule-based methods. The authors claim that this is the first attempt to embed entities, relations, and logical rules into a unified space for KGR. ## Strength 1. The idea of embedding both knowledge graphs and logical rule...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a new knowledge graph reasoning (KGR) model that combines embedding-based and rule-based methods. The authors claim that this is the first attempt to embed entities, relations, and logical rules into a unified space for KGR. ## Strength 1. The idea of embedding both knowledge graphs and logi...
This paper analyzes the denoising score matching objective used to train diffusion models, which have recently become the dominating class of deep generative models. The paper argues that at intermediate perturbation levels the denoising score matching objective is particularly noisy and "unstable", because one perturb...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper analyzes the denoising score matching objective used to train diffusion models, which have recently become the dominating class of deep generative models. The paper argues that at intermediate perturbation levels the denoising score matching objective is particularly noisy and "unstable", because one...
This paper focuses on the task of 3D object detection with only a small portion of labeled data. Conventional active learning-based approaches show a promising solution, but they fail to balance the trade-off between point cloud informativeness and box-level annotation costs. To overcome this limitation, this paper joi...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper focuses on the task of 3D object detection with only a small portion of labeled data. Conventional active learning-based approaches show a promising solution, but they fail to balance the trade-off between point cloud informativeness and box-level annotation costs. To overcome this limitation, this p...
This paper proposes an approximation of the hypergeometric distribution that introduces a differentiable hyperparameter. The benefit is that, in machine learning algorithms where data may fall into discrete categories, the number and size of the categories is potentially learnable with gradient methods. The paper illus...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an approximation of the hypergeometric distribution that introduces a differentiable hyperparameter. The benefit is that, in machine learning algorithms where data may fall into discrete categories, the number and size of the categories is potentially learnable with gradient methods. The pap...
The paper studies the problem of reducing the memory consumption of pre-trained GPT-like models via post-training quantization methods. These models often have billions or even hundreds of billions of parameters, making the compression overhead also a big cost. To address this issue, the paper introduces layer-by-layer...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper studies the problem of reducing the memory consumption of pre-trained GPT-like models via post-training quantization methods. These models often have billions or even hundreds of billions of parameters, making the compression overhead also a big cost. To address this issue, the paper introduces layer-...
The paper extended weakly supervised image scene graph generation to videos, and presented a new task of weakly supervised video scene graph generation, where a model must learn from unlocalized scene graphs on a sparse set of video frames. In tandem with the new task, the paper also proposed an interesting solution. T...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper extended weakly supervised image scene graph generation to videos, and presented a new task of weakly supervised video scene graph generation, where a model must learn from unlocalized scene graphs on a sparse set of video frames. In tandem with the new task, the paper also proposed an interesting sol...
In this work, the authors propose Deep Graph Ensemble (DGE) to tackle the additional information available in higher order networks $-$ something which standard Message Passing GNNs methods are unable to process without loss of information. The authors propose to do this by leveraging an ensemble of standard MPNNs - wh...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this work, the authors propose Deep Graph Ensemble (DGE) to tackle the additional information available in higher order networks $-$ something which standard Message Passing GNNs methods are unable to process without loss of information. The authors propose to do this by leveraging an ensemble of standard MP...
This paper tackles the problem of video prediction: given a video, the aim is to predict the future frames. This is a challenging problem as it requires learning the dynamics of the scene content in addition to the content itself. The authors argue that the classical autoregressive paradigm is both inconsistent with ma...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper tackles the problem of video prediction: given a video, the aim is to predict the future frames. This is a challenging problem as it requires learning the dynamics of the scene content in addition to the content itself. The authors argue that the classical autoregressive paradigm is both inconsistent...
In this paper, a lightweight blind source separation model (TDANet) is presented. The model is inspired from the brain’s top-down attention architecture. The separation is carried out using an embedder, a separation and a decoder (with the same parameters for all speakers) on the time samples. The model is consiste...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, a lightweight blind source separation model (TDANet) is presented. The model is inspired from the brain’s top-down attention architecture. The separation is carried out using an embedder, a separation and a decoder (with the same parameters for all speakers) on the time samples. The model is ...
This paper builds connections between existing diffusion models such as DDPM with serial reproduction paradigm. One major conclusion is that, current diffusion models can be explained as a natural consequence of this connection correspondence. Simulations on the MNIST dataset shows that the connection can be utilized f...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper builds connections between existing diffusion models such as DDPM with serial reproduction paradigm. One major conclusion is that, current diffusion models can be explained as a natural consequence of this connection correspondence. Simulations on the MNIST dataset shows that the connection can be ut...
The paper proposes a new kind of tensor decomposition (TD) framework for non-negative tensors which is inspired by many-body interactions in physics. Unlike most traditional TDs which are hard to fit since they correspond to non-convex optimization problems, the proposed framework corresponds to a convex optimization...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The paper proposes a new kind of tensor decomposition (TD) framework for non-negative tensors which is inspired by many-body interactions in physics. Unlike most traditional TDs which are hard to fit since they correspond to non-convex optimization problems, the proposed framework corresponds to a convex opti...
In this paper, the authors proposed the spike-efficiency measurement for spiking neural networks and showed the relationship between spike-efficiency and the heterogeneity. They also performed some empirical experiments to demonstrate the significance of neuronal and synaptic heterogeneity in reducing spiking activity....	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: In this paper, the authors proposed the spike-efficiency measurement for spiking neural networks and showed the relationship between spike-efficiency and the heterogeneity. They also performed some empirical experiments to demonstrate the significance of neuronal and synaptic heterogeneity in reducing spiking a...
This paper first analyzes the relationship between ensemble and mixup and empirically shows that both have some similarities in terms of their decision boundary --- there is some uncertainty region for mixup and ensemble, that is not seen in vanilla training. Then, the paper proposes a Test-Time mixup technique that al...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper first analyzes the relationship between ensemble and mixup and empirically shows that both have some similarities in terms of their decision boundary --- there is some uncertainty region for mixup and ensemble, that is not seen in vanilla training. Then, the paper proposes a Test-Time mixup technique...
This paper inspects MIM pre-training approaches and finds MIM essentially helps the model to learn better middle-order interactions between patches. Motivated by that, this paper proposes a novel MIM-based method, dubbed A$^2$MIM, that works well for both (small-sized) ConvNets & ViTs. Extensive experiments are conduct...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This paper inspects MIM pre-training approaches and finds MIM essentially helps the model to learn better middle-order interactions between patches. Motivated by that, this paper proposes a novel MIM-based method, dubbed A$^2$MIM, that works well for both (small-sized) ConvNets & ViTs. Extensive experiments are...
The paper proposes an approach to augment a ML model's training process when ground truth explanations (in this case, subsets of relevant features) are available. The approach is based on encouraging a model's learned representation to be similar with the full input and with the subset of relevant features. --- Strengt...	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 paper proposes an approach to augment a ML model's training process when ground truth explanations (in this case, subsets of relevant features) are available. The approach is based on encouraging a model's learned representation to be similar with the full input and with the subset of relevant features. ---...
The paper provides a convergence analysis of Neural TD Learning with a projection onto a ball of fixed radius around the initial point. The paper provides some new results on the convergence of TD learnign with neural network. The results seem interesting. Weaknesses of the paper are summarized as follows: 1) The de...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper provides a convergence analysis of Neural TD Learning with a projection onto a ball of fixed radius around the initial point. The paper provides some new results on the convergence of TD learnign with neural network. The results seem interesting. Weaknesses of the paper are summarized as follows: 1...
This work tries to develop an upper bound of the robust risk and design a new algorithm for adversarial training called ARoW which minimizes a surrogate version of the developed upper bound. Strength: 1. Well-written 2. Claim both theoretical and empirical contributions Weakness: 1. (1) I guess Equation A.1 should...	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 work tries to develop an upper bound of the robust risk and design a new algorithm for adversarial training called ARoW which minimizes a surrogate version of the developed upper bound. Strength: 1. Well-written 2. Claim both theoretical and empirical contributions Weakness: 1. (1) I guess Equation A....
The paper studies the problem of simultaneously predicting times and location of multiple (fixed number of) future events with a neural spatio-temporal point process model. The proposed model is based on an encoder-decoder transformer architecture, and the distribution of the future events is modeled by **Strengths:...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper studies the problem of simultaneously predicting times and location of multiple (fixed number of) future events with a neural spatio-temporal point process model. The proposed model is based on an encoder-decoder transformer architecture, and the distribution of the future events is modeled by **St...
The authors propose TDPM, a truncated diffusion probabilistic model that essentially skips the diffusion steps by truncating the start/end of the process, stopping at an implicit non-gaussian distribution which can be sampled from another generative model. The goal is to reduce the number of steps without compromising ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The authors propose TDPM, a truncated diffusion probabilistic model that essentially skips the diffusion steps by truncating the start/end of the process, stopping at an implicit non-gaussian distribution which can be sampled from another generative model. The goal is to reduce the number of steps without compr...
The paper shows that the adoption of contrastive learning can improve the text-image semantic consistency and the quality of synthetic images, and the adoption of style block can enhance the fine-grained details of the images. The experiments by implementing both methods on AttnGAN and SSAGAN shows the effectiveness of...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: The paper shows that the adoption of contrastive learning can improve the text-image semantic consistency and the quality of synthetic images, and the adoption of style block can enhance the fine-grained details of the images. The experiments by implementing both methods on AttnGAN and SSAGAN shows the effectiv...
The paper introduces truncated diffusion models, which truncates the forward process at certain timesteps so that when sampling, only a small portion of the reverse process is needed, which saves sampling time. To map noise to the noisy data at the truncated timestep, a conditional GAN is trained. The paper also builds...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper introduces truncated diffusion models, which truncates the forward process at certain timesteps so that when sampling, only a small portion of the reverse process is needed, which saves sampling time. To map noise to the noisy data at the truncated timestep, a conditional GAN is trained. The paper als...
The paper proposes to use Dense Correlation Fields which is a pyramid of correlation volume of visual features. The correlation takes into account long-term correlation, and high-level semantically correlation feature. The method shows improvement over 2D CNN and 3D CNN baselines by a healthy margin. ### Pros 1. The id...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes to use Dense Correlation Fields which is a pyramid of correlation volume of visual features. The correlation takes into account long-term correlation, and high-level semantically correlation feature. The method shows improvement over 2D CNN and 3D CNN baselines by a healthy margin. ### Pros 1...
This paper proposes a deep neural network to approximate the computation of the hyper-volume, which becomes expensive for a large number of objectives or points. The network has a special architecture to take into account the particularities of the hyper-volume computations such as invariances to permutations. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a deep neural network to approximate the computation of the hyper-volume, which becomes expensive for a large number of objectives or points. The network has a special architecture to take into account the particularities of the hyper-volume computations such as invariances to permut...
The paper explores a method for pretraining an Audio-Visual speech recognition model directly from raw video with both audio and visual signals. It does so by asymmetrically applying two student-teacher networks 1. the audio student learns to predict both audio and visual targets generated by respective teachers, 2. th...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper explores a method for pretraining an Audio-Visual speech recognition model directly from raw video with both audio and visual signals. It does so by asymmetrically applying two student-teacher networks 1. the audio student learns to predict both audio and visual targets generated by respective teacher...
The paper aims to improve the numerical stability of stein mixture in variational inference. To achieve this, the proposed EBLO-within-stein method allows the choice of a class of hierarchical attractives forces indexed by the hyper-parameter $\alpha$ in the Renyi divergence. Different values of $\alpha$ acounts for di...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper aims to improve the numerical stability of stein mixture in variational inference. To achieve this, the proposed EBLO-within-stein method allows the choice of a class of hierarchical attractives forces indexed by the hyper-parameter $\alpha$ in the Renyi divergence. Different values of $\alpha$ acount...
This paper proposes a new method to improve previous (cluster-based) important client sampling methods in federated learning. The new method is motivated by the insight that it would be beneficial to select clients from diverse groups. Convergence analysis are also provided and the authors claim they improve over exist...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a new method to improve previous (cluster-based) important client sampling methods in federated learning. The new method is motivated by the insight that it would be beneficial to select clients from diverse groups. Convergence analysis are also provided and the authors claim they improve ov...
The submission proposes an online learning approach to bidding in repeated stochastic first price auctions with a global budget constraint. The main idea is to approximately learn the distribution of values and other bidders’ top bid, and then solve a DP for the optimal “non-anticipating” strategy assuming values and b...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The submission proposes an online learning approach to bidding in repeated stochastic first price auctions with a global budget constraint. The main idea is to approximately learn the distribution of values and other bidders’ top bid, and then solve a DP for the optimal “non-anticipating” strategy assuming valu...
This paper gives lower bounds for differentially private ERM in the unconstrained and non-euclidean case. They provide a simple blackbox reduction approach to reduce lower bounds in the constrained case to unconstrained cases based on the idea of a Lipschitz extension of a function. They also give lower bounds for both...	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 gives lower bounds for differentially private ERM in the unconstrained and non-euclidean case. They provide a simple blackbox reduction approach to reduce lower bounds in the constrained case to unconstrained cases based on the idea of a Lipschitz extension of a function. They also give lower bounds ...
In this work authors adopt hierarchical classification for OOD detection and based on hierarchy they make predictions at varying level of granularity which could provide enhance examinability w.r.t OOD. This is indeed useful for OOD detection or open-set detection, and authors provide various qualitative & quantitative...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this work authors adopt hierarchical classification for OOD detection and based on hierarchy they make predictions at varying level of granularity which could provide enhance examinability w.r.t OOD. This is indeed useful for OOD detection or open-set detection, and authors provide various qualitative & quan...