review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper provides a modified version of the existing MMVAE model called MMVAE+ for weakly-supervised generative learning with multiple modalities. The paper aims to overcome the trade-off between generative quality and generative coherence by suggesting having separate latent encoding for both shared and private feat... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper provides a modified version of the existing MMVAE model called MMVAE+ for weakly-supervised generative learning with multiple modalities. The paper aims to overcome the trade-off between generative quality and generative coherence by suggesting having separate latent encoding for both shared and priv... |
This paper addresses the curse of the dimensionality problem in multiagent reinforcement learning (MARL), where the state-action space grows exponentially as the number of agents increases. To address this challenge, this paper proposes two novel implementations that exploit the permutation invariance (PI) and permutat... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper addresses the curse of the dimensionality problem in multiagent reinforcement learning (MARL), where the state-action space grows exponentially as the number of agents increases. To address this challenge, this paper proposes two novel implementations that exploit the permutation invariance (PI) and ... |
This paper studies the metrics (or abstractions) on the policy space for a given MDP. The paper defines three pseudo-metrics on the policy space: $d_\pi(\pi,\pi’)$ that measures the distance between the outputs of $\pi$ and $\pi’$ given a state, $d_{P^\pi}(\pi,\pi’)$ that measures the distance between the distributions... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the metrics (or abstractions) on the policy space for a given MDP. The paper defines three pseudo-metrics on the policy space: $d_\pi(\pi,\pi’)$ that measures the distance between the outputs of $\pi$ and $\pi’$ given a state, $d_{P^\pi}(\pi,\pi’)$ that measures the distance between the distr... |
This paper studies the dimensionality collapse in self-supervised learning (SSL). Though the uniformity of the representation is regarded as an important feature for downstream tasks, the existing loss function in SimCLR is insensitive to dimensional collapse. With a theoretical guarantee, the authors propose a Wassers... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper studies the dimensionality collapse in self-supervised learning (SSL). Though the uniformity of the representation is regarded as an important feature for downstream tasks, the existing loss function in SimCLR is insensitive to dimensional collapse. With a theoretical guarantee, the authors propose a... |
The authors work on merge individual models built on different training data sets to obtain a single model. More specifically, the authors propose a dataless knowledge fusion which is guided by weights that minimize prediction differences between the merged model and the individual models.
Strength
1. The proposed se... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors work on merge individual models built on different training data sets to obtain a single model. More specifically, the authors propose a dataless knowledge fusion which is guided by weights that minimize prediction differences between the merged model and the individual models.
Strength
1. The pro... |
This paper presents a transformer-based method for symbolic regression using a new feature extractor and a joint supervised learning mechanism. They evaluated the framework on synthetic datasets and compared it with several baseline methods. Experimental results show good performance of the proposed framework.
Streng... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a transformer-based method for symbolic regression using a new feature extractor and a joint supervised learning mechanism. They evaluated the framework on synthetic datasets and compared it with several baseline methods. Experimental results show good performance of the proposed framework.... |
The paper introduces Wasserstein Globalness, a model-agnostic method for measuring the globalness (the inverse of locality) of explainers.
*Strengths*
* The idea becomes rather clear even for a non-expert, and the paper could be of interest to the explainability community.
* The paper reads well and is easy to follo... | 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 introduces Wasserstein Globalness, a model-agnostic method for measuring the globalness (the inverse of locality) of explainers.
*Strengths*
* The idea becomes rather clear even for a non-expert, and the paper could be of interest to the explainability community.
* The paper reads well and is easy ... |
The work provides their "DEMASKED SMOOTHING" approach, which the authors argue provide defence framework against patch attacks for segmentation models. They evaluate their model on semantic segmentation datasets ADE20k and COCO-Stuff-10K. They use models trained on existing benchmarks, evaluated on various masking sche... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The work provides their "DEMASKED SMOOTHING" approach, which the authors argue provide defence framework against patch attacks for segmentation models. They evaluate their model on semantic segmentation datasets ADE20k and COCO-Stuff-10K. They use models trained on existing benchmarks, evaluated on various mask... |
The paper proposes a new approach for protein-protein contact prediction that is based on monomer data (this has been used in related work) and particularly cuts monomers into two sub-parts and pre-trains the model to merge them back. The results show huge performance gains compared to related works.
(+) The paper is... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a new approach for protein-protein contact prediction that is based on monomer data (this has been used in related work) and particularly cuts monomers into two sub-parts and pre-trains the model to merge them back. The results show huge performance gains compared to related works.
(+) The ... |
This paper considers the application of language models (LM) to mathematics reasoning tasks, noting that many existing datasets only contain one reference solution. This is problematic since many derivations may lead to the same answer, and therefore maximum-likelihood estimation of the LM may lead to overfitting. The ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper considers the application of language models (LM) to mathematics reasoning tasks, noting that many existing datasets only contain one reference solution. This is problematic since many derivations may lead to the same answer, and therefore maximum-likelihood estimation of the LM may lead to overfitti... |
The authors compare multiple training procedures in order to understand which training techniques improve target robustness. They find that generally increasing robustness on the source increases robustness on the target. They also find that techniques involving local lipschitz and contrastive learning generally tran... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors compare multiple training procedures in order to understand which training techniques improve target robustness. They find that generally increasing robustness on the source increases robustness on the target. They also find that techniques involving local lipschitz and contrastive learning genera... |
This paper proposes a variational classifier (VC). Typical machine learning classifiers use sigmoid or softmax to deterministically map last layer feature vector to class label predictions. This paper revisits the MLE, MAP and Bayesian under a variational framework. This work designs a novel objective based on ELBO and... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a variational classifier (VC). Typical machine learning classifiers use sigmoid or softmax to deterministically map last layer feature vector to class label predictions. This paper revisits the MLE, MAP and Bayesian under a variational framework. This work designs a novel objective based on ... |
Sharpness-aware training has been widely adopted for its generalization performance.
This paper provides the theoretical analysis for sharpness-aware minimization (SAM).
This paper categorizes the SAM in three ways, worst-direction, ascent-direction, and average direction.
Besides, they provide explicit bias for each w... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
Sharpness-aware training has been widely adopted for its generalization performance.
This paper provides the theoretical analysis for sharpness-aware minimization (SAM).
This paper categorizes the SAM in three ways, worst-direction, ascent-direction, and average direction.
Besides, they provide explicit bias fo... |
The authors study the problem of causal bandits with (arbitrary) combinatorial action spaces in both fixed-confidence and fixed-budget pure exploration (simple regret) settings. The causal graph is assumed to be known and all variables are binary valued. The authors consider two cases. For the first, the causal mode... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors study the problem of causal bandits with (arbitrary) combinatorial action spaces in both fixed-confidence and fixed-budget pure exploration (simple regret) settings. The causal graph is assumed to be known and all variables are binary valued. The authors consider two cases. For the first, the cau... |
The paper proposes a layer freezing method to reduce computational time for training of ML models. An attention-based layer freezing model takes sampled parameters from a layer as input, and predicts if it should be frozen or not. The layers get dynamically frozen over the training period. Results should significant im... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes a layer freezing method to reduce computational time for training of ML models. An attention-based layer freezing model takes sampled parameters from a layer as input, and predicts if it should be frozen or not. The layers get dynamically frozen over the training period. Results should signif... |
This paper lies along the line of work exploring how to combine BC and offline RL. It takes the perspective of using BC as a state representation learning method, proposing to train an intermediate representation in the offline RL agent via BC (either as pre-training or co-training). Experimental results show that the ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper lies along the line of work exploring how to combine BC and offline RL. It takes the perspective of using BC as a state representation learning method, proposing to train an intermediate representation in the offline RL agent via BC (either as pre-training or co-training). Experimental results show t... |
Predicts the residual between the output of a FastSpeech spectrogram predictor and the ground truth. Reduces the complexity of the problem the diffusion model is performing.
# Strengths
* Straightforward application of diffusion to improve a FastSpeech-based TTS system.
# Weaknesses
* Increases complexity in an unpr... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Predicts the residual between the output of a FastSpeech spectrogram predictor and the ground truth. Reduces the complexity of the problem the diffusion model is performing.
# Strengths
* Straightforward application of diffusion to improve a FastSpeech-based TTS system.
# Weaknesses
* Increases complexity in... |
Post rebuttal
====
Thank the authors for the responses. Since the responses addressed my original concerns, I raised my score. However, I agree with reviewer i4TZ and qY8f that the theoretical and empirical studies are still somewhat simple, and could still be improved.
This paper considers how to obtain a parametr... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Post rebuttal
====
Thank the authors for the responses. Since the responses addressed my original concerns, I raised my score. However, I agree with reviewer i4TZ and qY8f that the theoretical and empirical studies are still somewhat simple, and could still be improved.
This paper considers how to obtain a ... |
The paper proposes an approach for transferring skills for goal-conditioned tasks with different action / observation spaces. The method seems to build on Decoupled Policy Optimization (DePO) to learn a planner for the next state, and an inverse dynamics model to produce the action, but additionally conditions on the d... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes an approach for transferring skills for goal-conditioned tasks with different action / observation spaces. The method seems to build on Decoupled Policy Optimization (DePO) to learn a planner for the next state, and an inverse dynamics model to produce the action, but additionally conditions ... |
The paper analyzes data augmentation in detail for (small) computer vision tasks with (small) networks. In particular, the paper analyzes the following aspects of data augmentation:
* What is the *exchange rate* between augmentations and more data?
* How does augmentation help OOD generalization in comparison to more d... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper analyzes data augmentation in detail for (small) computer vision tasks with (small) networks. In particular, the paper analyzes the following aspects of data augmentation:
* What is the *exchange rate* between augmentations and more data?
* How does augmentation help OOD generalization in comparison t... |
The paper proposes to treat "time" in an energy-based model as a random variable and learns a joint distribution of time and samples. Concretely, this results in a parameterized function that outputs the joint log probability of the samples and time indices. Two methods of training such a model have been presented: a) ... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
The paper proposes to treat "time" in an energy-based model as a random variable and learns a joint distribution of time and samples. Concretely, this results in a parameterized function that outputs the joint log probability of the samples and time indices. Two methods of training such a model have been presen... |
The paper proposed a new way of dense retrieval using a retrieval augmented model. The modeling differs from previous models in the sense that it allows integrating new corpora that are never used during the training. This allows the proposed method to perform zero-shot retrieval nicely on the new domain that is just i... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposed a new way of dense retrieval using a retrieval augmented model. The modeling differs from previous models in the sense that it allows integrating new corpora that are never used during the training. This allows the proposed method to perform zero-shot retrieval nicely on the new domain that i... |
A practical and challenging scenario called label Missing Not At Random (MNAR) is usually ignored in previous works on semi-supervised learning (SSL). In MNAR, the labeled and unlabeled data fall into different class distributions resulting in biased label imputation, which deteriorates the performance of SSL models. I... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
A practical and challenging scenario called label Missing Not At Random (MNAR) is usually ignored in previous works on semi-supervised learning (SSL). In MNAR, the labeled and unlabeled data fall into different class distributions resulting in biased label imputation, which deteriorates the performance of SSL m... |
This paper proposes a method for self-supervised representation learning of LIDAR point clouds, to initialise training of standard supervised methods for LIDAR 3D object detection and point cloud segmentation. The paper has two main contributions:
The self-supervised representation is learnt via contrastive learning o... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a method for self-supervised representation learning of LIDAR point clouds, to initialise training of standard supervised methods for LIDAR 3D object detection and point cloud segmentation. The paper has two main contributions:
The self-supervised representation is learnt via contrastive le... |
FIGARO is a model trained on a seq2seq task, where the input is either a human-interpretable high-level description of the music or a learned sequence of VQ-VAE tokens. Output is a symbolic music composition in the proposed REMI+ format, an extension to REMI that adds multi-instrument and multi-time signature capabilit... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
FIGARO is a model trained on a seq2seq task, where the input is either a human-interpretable high-level description of the music or a learned sequence of VQ-VAE tokens. Output is a symbolic music composition in the proposed REMI+ format, an extension to REMI that adds multi-instrument and multi-time signature c... |
This paper presents behavior pathway identification approach that can distinguish a certain user behavior thread from other random actions. Specifically, by applying Gumbel-Softmax on Pathway prediction functions, the proposed approach is able to augment query representations of the multi-head attention model such that... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents behavior pathway identification approach that can distinguish a certain user behavior thread from other random actions. Specifically, by applying Gumbel-Softmax on Pathway prediction functions, the proposed approach is able to augment query representations of the multi-head attention model s... |
This paper introduces a novel approach for the estimation of continuous time state space dynamical models.
This field has gained significant attraction since 2018 with the Neural ODE.
The proposed approach is based on the evaluations on short subsections, and a novel state-derivative normalization.
Strengths:
1. The a... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper introduces a novel approach for the estimation of continuous time state space dynamical models.
This field has gained significant attraction since 2018 with the Neural ODE.
The proposed approach is based on the evaluations on short subsections, and a novel state-derivative normalization.
Strengths:
... |
This paper proposes to tackle the problem of multi-document summarization through synthesis evaluation. Do summarization models synthesize large and varied inputs? The paper studies this question from two angles: movie reviews, where the meta-review is human-written and scores are given (roughly equivalent to sentiment... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes to tackle the problem of multi-document summarization through synthesis evaluation. Do summarization models synthesize large and varied inputs? The paper studies this question from two angles: movie reviews, where the meta-review is human-written and scores are given (roughly equivalent to s... |
This paper studies reinforcement learning across domains in the presence of visual distractions. A visual distraction is a part of the underlying state that is uncontrollable, i.e., its change is independent of agent's action, and does not affect the Q values. The proposed approach uses domain randomization in a Deep Q... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies reinforcement learning across domains in the presence of visual distractions. A visual distraction is a part of the underlying state that is uncontrollable, i.e., its change is independent of agent's action, and does not affect the Q values. The proposed approach uses domain randomization in ... |
This paper shows that training a model with both cross-entropy loss and caption supervision are more robust than the model trained only on cross-entropy loss. It also releases CaptioNet, which is a 100-class subset of Imagenet with caption supervision.
Strength: The paper is clear overall. It studies an important ques... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper shows that training a model with both cross-entropy loss and caption supervision are more robust than the model trained only on cross-entropy loss. It also releases CaptioNet, which is a 100-class subset of Imagenet with caption supervision.
Strength: The paper is clear overall. It studies an import... |
The paper proposes two open intent classification methods based on the label-smoothing method, C-LC, and C-ADB. More specifically, the label smoothing method takes a weighted average of the ground-true label and the uniform 1/k vector. The experiments
evaluates the proposed methods on real-world datasets with previous... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes two open intent classification methods based on the label-smoothing method, C-LC, and C-ADB. More specifically, the label smoothing method takes a weighted average of the ground-true label and the uniform 1/k vector. The experiments
evaluates the proposed methods on real-world datasets with ... |
The authors propose a transformer with memory of reaction-trees for a new multi-step-retrosynthesis benchmark.
## Strengths:
- Propose a new Benchmark for Multi-Step-Retrosynthesis
- Propose Metro: Memory-Enhanced Transformer for RetrOsynthetic planning by extending Transformer with an additional memory module
- ... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a transformer with memory of reaction-trees for a new multi-step-retrosynthesis benchmark.
## Strengths:
- Propose a new Benchmark for Multi-Step-Retrosynthesis
- Propose Metro: Memory-Enhanced Transformer for RetrOsynthetic planning by extending Transformer with an additional memory modul... |
This paper is proposing a model for continual learning based on subnetwork isolation and training. Specifically, the authors build on top of SupSup, a prior method relying on a randomly initialized backbone where task-specific subnetworks are discovered and isolated, showcasing good performances on a stream of tasks to... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper is proposing a model for continual learning based on subnetwork isolation and training. Specifically, the authors build on top of SupSup, a prior method relying on a randomly initialized backbone where task-specific subnetworks are discovered and isolated, showcasing good performances on a stream of ... |
This paper proposed a neural network based solver which estimates an explicit transport map rather than just using OT losses. Specifically, the authors introduced a weak OT formulation so as to find the potentially stochastic optimal transport map. The resulting objective becomes a minimax optimization which is learnin... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper proposed a neural network based solver which estimates an explicit transport map rather than just using OT losses. Specifically, the authors introduced a weak OT formulation so as to find the potentially stochastic optimal transport map. The resulting objective becomes a minimax optimization which is... |
This paper introduce a new benchmark for evaluating the generalization ability of instructional action understanding models. The main idea is to collect existing steps in training dataset and re-assembling them as new task, then collect corresponding videos. Authors also propose an effective casual-based method to imp... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduce a new benchmark for evaluating the generalization ability of instructional action understanding models. The main idea is to collect existing steps in training dataset and re-assembling them as new task, then collect corresponding videos. Authors also propose an effective casual-based metho... |
This paper targets the domain generalization task. Based on the observation that fine-tuning may introduce gradient bias and hurt generalization ability, the paper estimates unobservable gradients that reduce potential risks in unseen domains. The main requirement is that there is a pre-trained model. Experimental resu... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper targets the domain generalization task. Based on the observation that fine-tuning may introduce gradient bias and hurt generalization ability, the paper estimates unobservable gradients that reduce potential risks in unseen domains. The main requirement is that there is a pre-trained model. Experimen... |
This paper introduces a new data structure to support lossless compression of neural networks and show the effectiveness and efficiency of inference workflows in computer vision tasks.
Strengthes:
+ Lossless compression method can be widely adopted in various inference workflows without considering the drawbacks o... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper introduces a new data structure to support lossless compression of neural networks and show the effectiveness and efficiency of inference workflows in computer vision tasks.
Strengthes:
+ Lossless compression method can be widely adopted in various inference workflows without considering the dra... |
This paper proposes a method, TECO, for the task of video prediction. TECO learns compressed representations to effectively and efficiently condition on long-form videos by modeling long-term dependencies.In order to achieve that, they leverage the MaskGit prior for dynamic prediction and propose a few new architecture... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper proposes a method, TECO, for the task of video prediction. TECO learns compressed representations to effectively and efficiently condition on long-form videos by modeling long-term dependencies.In order to achieve that, they leverage the MaskGit prior for dynamic prediction and propose a few new arch... |
This paper studies the excess risk of two-layer ReLU neural networks training by two-stage gradient based algorithms with spherical data iid spread on the unit sphere. Under such neural networks training, the authors demonstrate that, two layer neural networks are able to achieve the minimax $O(1/n)$ rate in a certain ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the excess risk of two-layer ReLU neural networks training by two-stage gradient based algorithms with spherical data iid spread on the unit sphere. Under such neural networks training, the authors demonstrate that, two layer neural networks are able to achieve the minimax $O(1/n)$ rate in a ... |
The paper proposes a variance reduction method based on the STORM algorithm. Not like STORM, their algorithm does not require access to specific problem parameters such as the Lipschitz constant and the gradient bound. They prove that this algorithm will converge at the rate of $O(T^{-1/3})$ under slightly weaker assum... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper proposes a variance reduction method based on the STORM algorithm. Not like STORM, their algorithm does not require access to specific problem parameters such as the Lipschitz constant and the gradient bound. They prove that this algorithm will converge at the rate of $O(T^{-1/3})$ under slightly weak... |
This paper studies the problem of how model accuracy relates to sample efficiency a class of model-based reinforcement learning algorithms, that is, model-based value expansion algorithms. The paper conducts experiments to show that the improvement of using the oracle dynamics model over a learned dynamics model is not... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the problem of how model accuracy relates to sample efficiency a class of model-based reinforcement learning algorithms, that is, model-based value expansion algorithms. The paper conducts experiments to show that the improvement of using the oracle dynamics model over a learned dynamics mode... |
This paper introduced a new type of neural layer, which is inspired by the well-known Taylor expansion. In particular, for each layer, the learning model is represented by a Takyor summation within two orders. To address the high dimensionality of the Hassien tensor, the low-rank Tucker decomposition is taken into acco... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper introduced a new type of neural layer, which is inspired by the well-known Taylor expansion. In particular, for each layer, the learning model is represented by a Takyor summation within two orders. To address the high dimensionality of the Hassien tensor, the low-rank Tucker decomposition is taken i... |
This paper presents an approach that generalizes symbolic regression
to graph-structured physical mechanisms. As opposed to classical
Symbolic Regression, this work assumes that X and Y in y=F(x) can be
both represented as graphs. The method is based on a two-level
optimization procedure where first the formula skeleto... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper presents an approach that generalizes symbolic regression
to graph-structured physical mechanisms. As opposed to classical
Symbolic Regression, this work assumes that X and Y in y=F(x) can be
both represented as graphs. The method is based on a two-level
optimization procedure where first the formula... |
In this paper, the authors proposed a new approach(with 3 different loss functions) to jointly estimate the geographical ranges of tens of thousands of different species simultaneously. At the same time, the authors develop a series of experiments under different settings to justify the effectiveness and robustness of ... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this paper, the authors proposed a new approach(with 3 different loss functions) to jointly estimate the geographical ranges of tens of thousands of different species simultaneously. At the same time, the authors develop a series of experiments under different settings to justify the effectiveness and robust... |
Authors state that the efficiency of traditional active learning (AL) should be significantly increased using approaches from continual learning (CL), when we consider past datapoints already used for training in AL as previous tasks. Authors test several standard CL approaches and newly proposed CL methods specificall... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
Authors state that the efficiency of traditional active learning (AL) should be significantly increased using approaches from continual learning (CL), when we consider past datapoints already used for training in AL as previous tasks. Authors test several standard CL approaches and newly proposed CL methods spe... |
The paper presents collaborative GAN, which aims to better balance the workload between individual generators. Specifically, a pre-trained classifier is introduced, and the training strategy is modified by adding a total variation distance based on WGAN-GP. Some experiments were conducted on MNIST dataset.
1) The nov... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
The paper presents collaborative GAN, which aims to better balance the workload between individual generators. Specifically, a pre-trained classifier is introduced, and the training strategy is modified by adding a total variation distance based on WGAN-GP. Some experiments were conducted on MNIST dataset.
1)... |
In this paper, author introduce a probabilistic framework called contrastive learning with edge partitioning (CLEP) that integrates generative modeling and graph contrastive learning. CLEP models edge generation by cumulative latent node interactions over multiple mutually independent hidden communities.
**Weaknesses**... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
In this paper, author introduce a probabilistic framework called contrastive learning with edge partitioning (CLEP) that integrates generative modeling and graph contrastive learning. CLEP models edge generation by cumulative latent node interactions over multiple mutually independent hidden communities.
**Weak... |
This paper pointed out the shortcomings in the existing approaches for continual learning, mainly catastrophic forgetting of the previous task during new task adoption. It proposes per-iteration continual evaluation with new metrics that enable measuring worst-case performance. This work empirically studied five existi... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper pointed out the shortcomings in the existing approaches for continual learning, mainly catastrophic forgetting of the previous task during new task adoption. It proposes per-iteration continual evaluation with new metrics that enable measuring worst-case performance. This work empirically studied fiv... |
This paper first points out that aggressive augmentations may worsen representation and lead to sub-optimal performance. To mitigate this issue, they propose a novel contrastive learning framework (E-GCL) that employs the equivariance principle to implement cross-graph discrimination. To this end, the authors propose i... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper first points out that aggressive augmentations may worsen representation and lead to sub-optimal performance. To mitigate this issue, they propose a novel contrastive learning framework (E-GCL) that employs the equivariance principle to implement cross-graph discrimination. To this end, the authors p... |
The authors introduce the combination shift problem towards generalization - especially when the goal is to learn with as few combinations are observed in data. Towards this goal, the author extend the definition of disentaglement beyond of invariances to group actions - highlighting the necessity to exploit the equiva... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The authors introduce the combination shift problem towards generalization - especially when the goal is to learn with as few combinations are observed in data. Towards this goal, the author extend the definition of disentaglement beyond of invariances to group actions - highlighting the necessity to exploit th... |
The paper tries to accelerate the inference of BNNs by:
1- Quantizing the BatchNorm layer into 8-bits
2- Simplifying the deployment of BN layer non ResNet like networks
3- And proposing an optimized assembly implementation of the binary convolution
Weaknesses: The paper lacks novelty for a conference like ICLR. Quanti... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper tries to accelerate the inference of BNNs by:
1- Quantizing the BatchNorm layer into 8-bits
2- Simplifying the deployment of BN layer non ResNet like networks
3- And proposing an optimized assembly implementation of the binary convolution
Weaknesses: The paper lacks novelty for a conference like ICLR... |
The authors provide a variational inference optimizer on the manifold of positive-definite matrices with all its implementation details and several experiments with results that are, in a wide set of experiments, surprisingly analogous to sampling methods.
Strength: Rigurous, state-of-the-art results, style.
Weakness... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The authors provide a variational inference optimizer on the manifold of positive-definite matrices with all its implementation details and several experiments with results that are, in a wide set of experiments, surprisingly analogous to sampling methods.
Strength: Rigurous, state-of-the-art results, style.
... |
This paper proposes the Lagrangian Schrodinger bridge (LSB) problem, a framework to recover the population dynamics from temporal data. LSB is a stochastic optimal control problem, which enables the authors to leverage the structure and propose the training objective (7). This method is evaluated on synthetic and singl... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes the Lagrangian Schrodinger bridge (LSB) problem, a framework to recover the population dynamics from temporal data. LSB is a stochastic optimal control problem, which enables the authors to leverage the structure and propose the training objective (7). This method is evaluated on synthetic a... |
Different from the common inpainting methods, this paper propose a novel image inpainting model tending to fill the missing region with new visual instances instead of filling in the background content. The author first use two transformer-based networks to inpaint the background segmentation map and foreground segment... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
Different from the common inpainting methods, this paper propose a novel image inpainting model tending to fill the missing region with new visual instances instead of filling in the background content. The author first use two transformer-based networks to inpaint the background segmentation map and foreground... |
The paper proposes to learn spatial embedding for coordinates based on H3 hexagons and convolutional autoencoders using the OSM’s tag data.
S1. The problem is important and using convolutional autoencoders in building global spatial embeddings seems to be novel.
W1. The method is very similar to Hex2Vec (both based o... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes to learn spatial embedding for coordinates based on H3 hexagons and convolutional autoencoders using the OSM’s tag data.
S1. The problem is important and using convolutional autoencoders in building global spatial embeddings seems to be novel.
W1. The method is very similar to Hex2Vec (both... |
the paper investigates possible causes of catastrophic overfitting (CO), a well known state in adversarial training (AT) of robust model which leads to the collapse of model robustness during later stages of AT. The authors propose to augment the data with easy to classify DCT patterns during analisis, in order to gain... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
the paper investigates possible causes of catastrophic overfitting (CO), a well known state in adversarial training (AT) of robust model which leads to the collapse of model robustness during later stages of AT. The authors propose to augment the data with easy to classify DCT patterns during analisis, in order... |
This paper presents a novel deep-learning model capable of inferring the underlying causal structure. The method does this by treating the process as a black-box function which is estimated through a network that maps observational and interventional data to a causal diagram (the output). They demonstrate their approac... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a novel deep-learning model capable of inferring the underlying causal structure. The method does this by treating the process as a black-box function which is estimated through a network that maps observational and interventional data to a causal diagram (the output). They demonstrate their... |
This paper considers the problem of permuted linear regression. Here, you are given $Y \in \mathbb R^{n \times m}, X \in \mathbb R^{n \times p}$ such that $Y = \Pi X B + W$ where $\Pi$ is an unknown permutation matrix, $B \in \mathbb R^{p \times m}$ is the signal matrix, $X \in \mathbb R^{n \times p}$ is the set of Gau... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper considers the problem of permuted linear regression. Here, you are given $Y \in \mathbb R^{n \times m}, X \in \mathbb R^{n \times p}$ such that $Y = \Pi X B + W$ where $\Pi$ is an unknown permutation matrix, $B \in \mathbb R^{p \times m}$ is the signal matrix, $X \in \mathbb R^{n \times p}$ is the se... |
This paper proposed a new framework, namely interactive portrait harmonization, to allow users to select a reference region in the background image to guide the harmonization. The key to this framework is a luminance matching loss to ensure the luminance consistency between the selected reference region and the foregro... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper proposed a new framework, namely interactive portrait harmonization, to allow users to select a reference region in the background image to guide the harmonization. The key to this framework is a luminance matching loss to ensure the luminance consistency between the selected reference region and the... |
In this paper, the authors generalize the differentiable causal discovery method to the setting in the presence of latent variables. Compared to related work, the method can tackle non-linear data. They provide both theoretical and empirical results.
Strength:
1. The paper is clearly written.
2. The authors give a det... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
In this paper, the authors generalize the differentiable causal discovery method to the setting in the presence of latent variables. Compared to related work, the method can tackle non-linear data. They provide both theoretical and empirical results.
Strength:
1. The paper is clearly written.
2. The authors gi... |
Traditional training-set attacks manipulate model predictions by inserting adversarial instances directly into the training set. This paper proposes a novel threat model for training-set attacks whereby the adversarial training instances are created during the training process itself via a malicious augmentation schem... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
Traditional training-set attacks manipulate model predictions by inserting adversarial instances directly into the training set. This paper proposes a novel threat model for training-set attacks whereby the adversarial training instances are created during the training process itself via a malicious augmentati... |
The paper proposes a group federated learning algorithm with formal convergence/BGL guarantees.
-Privacy leakage/secrecy of the proposed algorithm is not discussed. It is unclear how much privacy leakage occurs. The proposed algorithm requires the exchange of additional information "r" per round, so one *must* discuss ... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a group federated learning algorithm with formal convergence/BGL guarantees.
-Privacy leakage/secrecy of the proposed algorithm is not discussed. It is unclear how much privacy leakage occurs. The proposed algorithm requires the exchange of additional information "r" per round, so one *must* ... |
This paper makes a few contributions: using a newly published weight generation model HyperTransformer to generate model weights for incremental task/class learning; modifying the classic SCE loss with an existing prototype loss (from a few-shot learning paper).
Strength:
It is a good baseline method for using the we... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper makes a few contributions: using a newly published weight generation model HyperTransformer to generate model weights for incremental task/class learning; modifying the classic SCE loss with an existing prototype loss (from a few-shot learning paper).
Strength:
It is a good baseline method for usin... |
This paper provides some insights on the vulnerability of l_{infty} adversarial trained model. The paper identifies inequality phenomenon occurs during l_{infty} adversarial training which is quantified by Gini index. To show the later, the paper proposes two methods: inductive noise and occlusion to demonstrate the vu... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper provides some insights on the vulnerability of l_{infty} adversarial trained model. The paper identifies inequality phenomenon occurs during l_{infty} adversarial training which is quantified by Gini index. To show the later, the paper proposes two methods: inductive noise and occlusion to demonstrat... |
This paper proposes a methodology to edit memory in a transformer model which can be used to modify or add the model's knowledge.
Pros:
- Knowledge editing seems quite intriguing especially for transformer models which seems like a black box. This seems to be a step in the right direction of model interpretability.
... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a methodology to edit memory in a transformer model which can be used to modify or add the model's knowledge.
Pros:
- Knowledge editing seems quite intriguing especially for transformer models which seems like a black box. This seems to be a step in the right direction of model interpretabi... |
This paper proposes to integrate depth signal into self-supervised learning methods as an additional signal in 2 ways. The first method simply concatenates the depth channel onto the RGB image and trains the SSL methods with RGBD input. The second method uses the single-view view synthesis methods to generate new views... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes to integrate depth signal into self-supervised learning methods as an additional signal in 2 ways. The first method simply concatenates the depth channel onto the RGB image and trains the SSL methods with RGBD input. The second method uses the single-view view synthesis methods to generate n... |
This paper proposes a novel memory-efficient training method Auxiliary Activation Learning, which reduces the amount of data to be stored in memory without additional computations. The experimental results illustrate that Auxiliary Activation Learning can significantly reduce the memory cost on computer vision and NLP ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a novel memory-efficient training method Auxiliary Activation Learning, which reduces the amount of data to be stored in memory without additional computations. The experimental results illustrate that Auxiliary Activation Learning can significantly reduce the memory cost on computer vision ... |
This paper studies the robust overfitting of one-step adversarial training and finds that robust overfitting is related to the increase of abnormal adversarial examples generated for training. Abnormal adversarial examples have even lower losses after adding the adversarial perturbations, which can be a reason for cau... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the robust overfitting of one-step adversarial training and finds that robust overfitting is related to the increase of abnormal adversarial examples generated for training. Abnormal adversarial examples have even lower losses after adding the adversarial perturbations, which can be a reason... |
This work proposes three different ways to regularize GFlowNet training with Wasserstein distance. Implementations of the regularization term and an upper bound of it are derived for tractable training. The path regularization is claimed to be capable of generating more diverse candidates via maximizing the Wasserstein... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposes three different ways to regularize GFlowNet training with Wasserstein distance. Implementations of the regularization term and an upper bound of it are derived for tractable training. The path regularization is claimed to be capable of generating more diverse candidates via maximizing the Was... |
This paper studies a notion coined as "benign memorization", which refers to the phenomenon that when training with standard data augmentation, a deep neural network could fit to complete random label assignment on the training set, yet still learns representations that shows surprisingly good discrimination power unde... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies a notion coined as "benign memorization", which refers to the phenomenon that when training with standard data augmentation, a deep neural network could fit to complete random label assignment on the training set, yet still learns representations that shows surprisingly good discrimination po... |
This paper proposes a communication-efficient distributed gradient clipping algorithm for federated learning, which is called EPISODE. The algorithm works particularly well with the heterogenous data and under the nonconvex and relaxed smoothness setting. The novelty consists in two techniques: episodic gradient clippi... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes a communication-efficient distributed gradient clipping algorithm for federated learning, which is called EPISODE. The algorithm works particularly well with the heterogenous data and under the nonconvex and relaxed smoothness setting. The novelty consists in two techniques: episodic gradien... |
This paper proposes a combinatorial node labeling framework, which solves CO problems by labeling the nodes. The model is based on graph attention networks and trained by reinforcement learning. Graph coloring and minimum vertex cover are studied. Experiments prove the performance of the proposed framework.
Strength
- ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a combinatorial node labeling framework, which solves CO problems by labeling the nodes. The model is based on graph attention networks and trained by reinforcement learning. Graph coloring and minimum vertex cover are studied. Experiments prove the performance of the proposed framework.
Str... |
This paper mainly focuses on a practical scenario of FL where the labels might be marked by different criteria over different centers/nodes. A new learning strategy is proposed for dealing with the mismatches of the labels, and it can be deployed with existing FL methods. Theoretical analysis shows that the proposed Fe... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper mainly focuses on a practical scenario of FL where the labels might be marked by different criteria over different centers/nodes. A new learning strategy is proposed for dealing with the mismatches of the labels, and it can be deployed with existing FL methods. Theoretical analysis shows that the pro... |
Remarking that language models can assign high probability to low “quality” outputs, this paper proposes “sequence likelihood calibration” (SLiC) as a solution. With SLiC, “decoding heuristics become unnecessary” and “quality significantly improves.” The claims are empirically supported by experiments on abstractive su... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Remarking that language models can assign high probability to low “quality” outputs, this paper proposes “sequence likelihood calibration” (SLiC) as a solution. With SLiC, “decoding heuristics become unnecessary” and “quality significantly improves.” The claims are empirically supported by experiments on abstra... |
This paper propose a DP-SGD based learning algorithm to achieve synthetic data generation. The proposed method achieve DP to protect the privacy of the training data.
1. This paper simply combines existing methods (DP-SGD and GPT-2). The novel and the technical contribution to the community is limited.
2. The model p... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper propose a DP-SGD based learning algorithm to achieve synthetic data generation. The proposed method achieve DP to protect the privacy of the training data.
1. This paper simply combines existing methods (DP-SGD and GPT-2). The novel and the technical contribution to the community is limited.
2. The... |
The paper proposes an object-centric representation learning model that targets large-scale real-world datasets. To do so, they propose to learn the object-centric representations by reconstructing the features of a pre-trained encoder (DINO). The hypothesis is that the features pre-trained on large-scale datasets cont... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes an object-centric representation learning model that targets large-scale real-world datasets. To do so, they propose to learn the object-centric representations by reconstructing the features of a pre-trained encoder (DINO). The hypothesis is that the features pre-trained on large-scale datas... |
This paper proposes an explainable recommender system based on variational autoencoders (VAE). The main idea is to model user/item textual reviews with an autoencoder whose latent space is regularized by user/item cluster-membership distributions. The latter are derived from the user-item interaction data by first clus... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes an explainable recommender system based on variational autoencoders (VAE). The main idea is to model user/item textual reviews with an autoencoder whose latent space is regularized by user/item cluster-membership distributions. The latter are derived from the user-item interaction data by fi... |
This work highlights the limitations of L-inf adversarial training - models trained using Linf-AT (specifically at large perturbation bounds) tend to rely heavily on very few features, when compared to normally trained models. This makes them vulnerable to attacks that perturb very few pixels, or occlusion based attack... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This work highlights the limitations of L-inf adversarial training - models trained using Linf-AT (specifically at large perturbation bounds) tend to rely heavily on very few features, when compared to normally trained models. This makes them vulnerable to attacks that perturb very few pixels, or occlusion base... |
This paper proposes a reinforcement learning based algorithm to estimate long-term effect for a class of nonstationary problems. Empirical results in both synthetic and real datasets show the potential of the proposed algorithm.
The paper studies a practical and important problem: estimate long-term effect under nonsta... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a reinforcement learning based algorithm to estimate long-term effect for a class of nonstationary problems. Empirical results in both synthetic and real datasets show the potential of the proposed algorithm.
The paper studies a practical and important problem: estimate long-term effect unde... |
The paper proposes to use pretrained protein representations for predicting whether two given proteins interact (protein-protein interaction, PPI). Specifically, it proposes to use pretrained embeddings from the encoder part of the OmegaFold protein structure prediction model. For the trained prediction network, a GNN ... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes to use pretrained protein representations for predicting whether two given proteins interact (protein-protein interaction, PPI). Specifically, it proposes to use pretrained embeddings from the encoder part of the OmegaFold protein structure prediction model. For the trained prediction network... |
This paper has an interesting finding for improving generalization ability of image deraining. By reducing the background images in training data, the deraining model will tend to learn background context instead of learning rain streaks, preventing deraining model overfitting to rain streaks patterns. The finding is i... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper has an interesting finding for improving generalization ability of image deraining. By reducing the background images in training data, the deraining model will tend to learn background context instead of learning rain streaks, preventing deraining model overfitting to rain streaks patterns. The find... |
This paper considers the problem of reconstructing a small-depth ReLU network with queries. They give results for depth 2 and depth 3 networks under some general position assumptions (and additional assumptions for depth 3). They solve the exact reconstruction problem where the goal is to reconstruct the parameters of ... | 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 considers the problem of reconstructing a small-depth ReLU network with queries. They give results for depth 2 and depth 3 networks under some general position assumptions (and additional assumptions for depth 3). They solve the exact reconstruction problem where the goal is to reconstruct the parame... |
This paper addresses the problem of subgraph localization (SGL). The goal is to identify a good fit of a target subgraph within a larger source graph. This problem is related to the subgraph isomorphism problem and is NP-hard. The authors establish a connection between the SGL problem and the additive inverse eigen val... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper addresses the problem of subgraph localization (SGL). The goal is to identify a good fit of a target subgraph within a larger source graph. This problem is related to the subgraph isomorphism problem and is NP-hard. The authors establish a connection between the SGL problem and the additive inverse e... |
The paper aims to propose a memory-efficient deep equilibrium model (DEQ) for solving inverse problems. However, after reading the paper, I am not sure if I have a clear understanding of what are the key contributions of the proposed LUSER model. It seems to me that the key difference between LUSER and prior DEQ models... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper aims to propose a memory-efficient deep equilibrium model (DEQ) for solving inverse problems. However, after reading the paper, I am not sure if I have a clear understanding of what are the key contributions of the proposed LUSER model. It seems to me that the key difference between LUSER and prior DE... |
The paper aimed to explore if the variance of losses should be always penalized when learning with noisy labels. The paper suggests adding a variance regularization to the loss, which will encourage the variance of losses. Consequently, when calculating the gradients of the proposed loss, if the loss of an example is ... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper aimed to explore if the variance of losses should be always penalized when learning with noisy labels. The paper suggests adding a variance regularization to the loss, which will encourage the variance of losses. Consequently, when calculating the gradients of the proposed loss, if the loss of an exam... |
1. The efficiency of Actionable Recourse Summaries (AReS) is improved, and a method of Global & Efficient Counterfactual Explanations (GLOBE-CE) is designed to address the stability issues flexibly.
2. The authors provide a mathematical analysis of categorical feature translations and use it in their presented method.
... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
1. The efficiency of Actionable Recourse Summaries (AReS) is improved, and a method of Global & Efficient Counterfactual Explanations (GLOBE-CE) is designed to address the stability issues flexibly.
2. The authors provide a mathematical analysis of categorical feature translations and use it in their presented ... |
This paper learns to discover objects by video prediction. The method called OPPOLE extracts each frame into object states(inferred object 3D position and 1D pose), and background. The method learns the decomposition in an unsupervised method that renders the object representation into image frames and compares it with... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper learns to discover objects by video prediction. The method called OPPOLE extracts each frame into object states(inferred object 3D position and 1D pose), and background. The method learns the decomposition in an unsupervised method that renders the object representation into image frames and compares... |
This paper proposed a probabilistic joint Gaussian mixture model (JGMM) that could simultaneously produce various forms of explanations for DCNN, including proxy models, prototypes, criticisms, influential examples, counter-factual examples and semi-factual examples. In this way, the compatibility and consistency among... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposed a probabilistic joint Gaussian mixture model (JGMM) that could simultaneously produce various forms of explanations for DCNN, including proxy models, prototypes, criticisms, influential examples, counter-factual examples and semi-factual examples. In this way, the compatibility and consisten... |
This paper works on open-vocabulary panoptic segmentation using pretrained CLIP weights. The authors first propose a new panoptic segmentation framework that first do class-agnostic mask proposal using an existing frameworks (MaskRCNN and Mask2Former), and then use the predicted mask to interact with the CLIP visual ba... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper works on open-vocabulary panoptic segmentation using pretrained CLIP weights. The authors first propose a new panoptic segmentation framework that first do class-agnostic mask proposal using an existing frameworks (MaskRCNN and Mask2Former), and then use the predicted mask to interact with the CLIP v... |
The structure of the Hessian of the data log likelihood carries information about topological ordering. Since this information is learned by score based generative models, DPMs carry within them information necessary to reconstruct topological orderings. The paper runs with this insight to obtain a practical algorith... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The structure of the Hessian of the data log likelihood carries information about topological ordering. Since this information is learned by score based generative models, DPMs carry within them information necessary to reconstruct topological orderings. The paper runs with this insight to obtain a practical ... |
This paper tackles the problem of novel category discover (NCD) in the setting where samples have been mis-labelled. Standard NCD gives a model a labelled set of images from closed-set categories, as well as an unlabelled set of images from disjoint categories. The task becomes to learn a classifier which can both reco... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper tackles the problem of novel category discover (NCD) in the setting where samples have been mis-labelled. Standard NCD gives a model a labelled set of images from closed-set categories, as well as an unlabelled set of images from disjoint categories. The task becomes to learn a classifier which can b... |
This paper analyzes the approximation and estimation error of transformers to the target functions of fixed-length in a mixed smooth Besov space and in an anisotropic Besov space. The authors also show that transformer models are capable of avoiding the curse of dimensionality and obtaining almost minimax optimal rate.... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper analyzes the approximation and estimation error of transformers to the target functions of fixed-length in a mixed smooth Besov space and in an anisotropic Besov space. The authors also show that transformer models are capable of avoiding the curse of dimensionality and obtaining almost minimax optim... |
This paper builds from the work in [1] to make efficient ensembles. This paper proposes to build ensembles by using shared templates/layers across ensemble members. Unlike previous work in this domain, they mention that because they can use average weights of different dimensions [1], they can combine models with layer... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper builds from the work in [1] to make efficient ensembles. This paper proposes to build ensembles by using shared templates/layers across ensemble members. Unlike previous work in this domain, they mention that because they can use average weights of different dimensions [1], they can combine models wi... |
This paper investigated the ethics disclosure in the AI papers that use crowdsourcing as a data collection method published in ICLR, NeurIPS, and Springer journals in the past four years. The main finding of the investigation is that compared with the Psychology paper, ethics disclosures are far less common. Also, the ... | 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 investigated the ethics disclosure in the AI papers that use crowdsourcing as a data collection method published in ICLR, NeurIPS, and Springer journals in the past four years. The main finding of the investigation is that compared with the Psychology paper, ethics disclosures are far less common. Al... |
This paper focusing on continuous open temporal graph where new nodes and new target node labels will join with time going on. This field is lack of study, and this paper proposes OTGNet which is far better than similar works.
## Strength
- This paper focues on continuous open temporal graph which is lack of formal st... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper focusing on continuous open temporal graph where new nodes and new target node labels will join with time going on. This field is lack of study, and this paper proposes OTGNet which is far better than similar works.
## Strength
- This paper focues on continuous open temporal graph which is lack of f... |
Fourier Neural Operator (FNO) is a data-based approach for learning an infinite-dimensional mapping, which is helpful for numerically solving PDEs. The paper extends FNO to an evolutional setup. FNO can be used in an evolutional setup too, for which one can learn a mapping that maps the solution at time t to that at ti... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
Fourier Neural Operator (FNO) is a data-based approach for learning an infinite-dimensional mapping, which is helpful for numerically solving PDEs. The paper extends FNO to an evolutional setup. FNO can be used in an evolutional setup too, for which one can learn a mapping that maps the solution at time t to th... |
This paper studies a framework on meta-learning in multi-agent games and considered a range of popular game structures, including zero-sum games, potential games, general-sum multi-player games, and Stackelberg security games. The general game is composed by a sequence of multiple sub-games, where each sub-games have ... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies a framework on meta-learning in multi-agent games and considered a range of popular game structures, including zero-sum games, potential games, general-sum multi-player games, and Stackelberg security games. The general game is composed by a sequence of multiple sub-games, where each sub-gam... |
This paper is about changes that need to be made to error-backpropagation in order to train a leaky-integrate and fire spiking neuron network. A technique that is commonly used is to "smoothen" the spike so that gradients do not collapse onto the two values of 0 or infinity. This paper suggests an adaptive smoother. Th... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper is about changes that need to be made to error-backpropagation in order to train a leaky-integrate and fire spiking neuron network. A technique that is commonly used is to "smoothen" the spike so that gradients do not collapse onto the two values of 0 or infinity. This paper suggests an adaptive smoo... |
This paper proposes an interesting new Hyperspherical Uniformity Gap (HUG) objective that explicitly minimizes within-class variation while maximizing between-class variation. An especially interesting innovation is introducing---instead of canonical linear classifiers---a "learnable proxy" that is minimized directly w... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes an interesting new Hyperspherical Uniformity Gap (HUG) objective that explicitly minimizes within-class variation while maximizing between-class variation. An especially interesting innovation is introducing---instead of canonical linear classifiers---a "learnable proxy" that is minimized di... |
The authors propose a new module named “InLay” that can be plugged into different models across multiple modalities to improve out of distribution performance. The work leverages the idea of “indirection” to learn *indirection representations* for a given sequence of inputs. The rationale behind the proposed method is ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose a new module named “InLay” that can be plugged into different models across multiple modalities to improve out of distribution performance. The work leverages the idea of “indirection” to learn *indirection representations* for a given sequence of inputs. The rationale behind the proposed me... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.