review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper addresses the problem of covariate shift on graphs. This framework is compared with 14 baselines on both synthetic and real-world data.
Strength:
- This paper addresses a rather important and interesting problem.
- Many baselines are provided for comparison.
----------------------------------------------... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper addresses the problem of covariate shift on graphs. This framework is compared with 14 baselines on both synthetic and real-world data.
Strength:
- This paper addresses a rather important and interesting problem.
- Many baselines are provided for comparison.
--------------------------------------... |
The authors propose composing multimodal models with language models
in a process they call "socratic models". The key idea is that the
outputs of VLMs/ALMs can be reformulated in text, which can then be
fed to a LLM via a prompt. While each instantiation of a socratic
model is slightly different, this family of approa... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose composing multimodal models with language models
in a process they call "socratic models". The key idea is that the
outputs of VLMs/ALMs can be reformulated in text, which can then be
fed to a LLM via a prompt. While each instantiation of a socratic
model is slightly different, this family o... |
This paper investigates the problem of recourse generation -- generating plans that allow a user to change the decision of a model to some other option. The paper proposes to incorporate the individuality of user preferences into both training and evaluation. This involves both a new objective function, Expected Minimu... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper investigates the problem of recourse generation -- generating plans that allow a user to change the decision of a model to some other option. The paper proposes to incorporate the individuality of user preferences into both training and evaluation. This involves both a new objective function, Expecte... |
This paper aims at modeling complex dynamical systems into smaller subsystems. It does so by combining neural controlled dynamics(Kidger et al.) with self-attention (Vaswani et al.). Empirical evaluations are performed in multiple regression and link prediction environments from 3-body gravitation, springs, charged par... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper aims at modeling complex dynamical systems into smaller subsystems. It does so by combining neural controlled dynamics(Kidger et al.) with self-attention (Vaswani et al.). Empirical evaluations are performed in multiple regression and link prediction environments from 3-body gravitation, springs, cha... |
This paper studies the SGDA algorithm with shuffling and sampling without replacement for non convex-PL minimax optimization problem. The complexity results show significant improvement over exiting analysis that use sampling with replacement.
Strength:
1. The analysis of sampling without replacement for minimax optim... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies the SGDA algorithm with shuffling and sampling without replacement for non convex-PL minimax optimization problem. The complexity results show significant improvement over exiting analysis that use sampling with replacement.
Strength:
1. The analysis of sampling without replacement for minim... |
This paper establishes a new benchmark based on AlphaFold DB, one of the world’s largest protein structure databases. Moreover, the authors propose a new baseline method called AlphaDesign, which achieves 5% higher recovery than previous methods and about 70 times inference speed-up in designing long protein sequences.... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper establishes a new benchmark based on AlphaFold DB, one of the world’s largest protein structure databases. Moreover, the authors propose a new baseline method called AlphaDesign, which achieves 5% higher recovery than previous methods and about 70 times inference speed-up in designing long protein se... |
The authors present a paper on biases in mainstream commercial speech recognition services from Amazon, Google and Microsoft that:
1. shows how there are statistically significant performance differences in a word information lost (WIL) metric when the speech that is being transcribed comes from a native vs non-native ... | 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 authors present a paper on biases in mainstream commercial speech recognition services from Amazon, Google and Microsoft that:
1. shows how there are statistically significant performance differences in a word information lost (WIL) metric when the speech that is being transcribed comes from a native vs non... |
The paper proposes a framework (called PRESTO) for simultaneously partitioning the data space and learning a separate model in each partition. The framework relies on solving an optimization problem with both discrete variables (data partitioning) and continuous parameters
(model parameters) using a constrained optimi... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper proposes a framework (called PRESTO) for simultaneously partitioning the data space and learning a separate model in each partition. The framework relies on solving an optimization problem with both discrete variables (data partitioning) and continuous parameters
(model parameters) using a constraine... |
The paper proposes the best possible operator for decentralized learning in cooperative multi-agent games. Since the best possible operator is almost impossible to implement in practice, the authors also propose a simplified best possible operator. Experiments are conducted on simple stochastic game, stage game, MPE, M... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes the best possible operator for decentralized learning in cooperative multi-agent games. Since the best possible operator is almost impossible to implement in practice, the authors also propose a simplified best possible operator. Experiments are conducted on simple stochastic game, stage game... |
This paper considers the task of representation learning for graph data. The proposed objective is a contrastive learning objective that is based on using clusters in negative sampling. Empirical analysis compares the proposed approach to a wide variety of representation learning and clustering approaches for graphs.
... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper considers the task of representation learning for graph data. The proposed objective is a contrastive learning objective that is based on using clusters in negative sampling. Empirical analysis compares the proposed approach to a wide variety of representation learning and clustering approaches for g... |
This submission discusses using the technique ClippedGossip in the decentralized optimization setting to counter potential Byzantine agents in the network.
The authors provide convergence guarantee in terms of first-order condition to show the attainment of consensus in the network, and give numerical experiments to ... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This submission discusses using the technique ClippedGossip in the decentralized optimization setting to counter potential Byzantine agents in the network.
The authors provide convergence guarantee in terms of first-order condition to show the attainment of consensus in the network, and give numerical experim... |
The authors propose a new sampling method for training deep neural network. The sampler is based on the sample importance, which is defined as the sample difficulty using the propability. This probability is produced by the early trained model using standard uniform sampling. The following training epoch will use the n... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The authors propose a new sampling method for training deep neural network. The sampler is based on the sample importance, which is defined as the sample difficulty using the propability. This probability is produced by the early trained model using standard uniform sampling. The following training epoch will u... |
The paper presents a deep kernel learning method for Gaussian processes. Deep kernel learning parametrises the covariance function using a neural network, and the proposed method further optimises a set of inducing points in the transformed space as well as employs stochastic gradient descent, allowing for faster infer... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper presents a deep kernel learning method for Gaussian processes. Deep kernel learning parametrises the covariance function using a neural network, and the proposed method further optimises a set of inducing points in the transformed space as well as employs stochastic gradient descent, allowing for fast... |
The paper looks at QA when questions are asked in a given scenario, and they can be answered only if providing the model information about the scenario. In addition, such questions require a high level of reasoning, and thus the model should also be able to infer how conditions interact with each other and find correct... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper looks at QA when questions are asked in a given scenario, and they can be answered only if providing the model information about the scenario. In addition, such questions require a high level of reasoning, and thus the model should also be able to infer how conditions interact with each other and find... |
The paper discusses the problem of adversarial training for non-contrastive self-supervised learning (NC-SSL) approaches like BYOL.
They propose a modification of the loss function that involves an adversarial attack for a pair of objects that, with high probability, lie in different classes.
Such a method allows to be... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper discusses the problem of adversarial training for non-contrastive self-supervised learning (NC-SSL) approaches like BYOL.
They propose a modification of the loss function that involves an adversarial attack for a pair of objects that, with high probability, lie in different classes.
Such a method allo... |
The focus of this work is learning representations of auditory and visual speech from the underlying raw signals, as opposed to learning a representation from hand-crafted features as is typically done. The experiments include an ablation over design choices. The combination of within-modality modeling (for audio) an... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The focus of this work is learning representations of auditory and visual speech from the underlying raw signals, as opposed to learning a representation from hand-crafted features as is typically done. The experiments include an ablation over design choices. The combination of within-modality modeling (for a... |
The paper relaxes the softmax function to a linear operation with unit simplex constraints. Then for the self-attention-only network with a scalar or vector output, the paper shows that the nonconvex optimization problem could be cast as a convex one. It also demonstrates that the convexified problem has an implicit re... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper relaxes the softmax function to a linear operation with unit simplex constraints. Then for the self-attention-only network with a scalar or vector output, the paper shows that the nonconvex optimization problem could be cast as a convex one. It also demonstrates that the convexified problem has an imp... |
To measure the difference between two samples $A$ and $B$, this paper assumes the samples $A$ and $B$ are endpoints of two different geodesics starting from the same point $P$ in a small neighborhood and approximates their squared geodesic distances using Taylor extension which leads to a local linear regression model... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
To measure the difference between two samples $A$ and $B$, this paper assumes the samples $A$ and $B$ are endpoints of two different geodesics starting from the same point $P$ in a small neighborhood and approximates their squared geodesic distances using Taylor extension which leads to a local linear regressi... |
This paper proposes random weight factorization, a new parametrization for linear neural network layers inspired by weight normalization. It argues that the parametrization reduces the distance between different parameter configurations in the loss landscape and reduces spectral bias, leading to performance improvement... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes random weight factorization, a new parametrization for linear neural network layers inspired by weight normalization. It argues that the parametrization reduces the distance between different parameter configurations in the loss landscape and reduces spectral bias, leading to performance imp... |
The paper proposes a new approach to communication efficient federated learning. They keep the weights of the model fixed after random initialization via a SEED. and then only train the probabilistic mask over the model in collaboration with all the clients. They reduce the communication of requirement to less1bpp duri... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a new approach to communication efficient federated learning. They keep the weights of the model fixed after random initialization via a SEED. and then only train the probabilistic mask over the model in collaboration with all the clients. They reduce the communication of requirement to less1... |
This paper studies asynchronous gradient plays in zero-sum polymatrix games under delayed feedbacks, while significant efforts have been made to understand zero-sum two-player matrix games. They first establish that the last iterate of the entropy-regularized optimistic multiplicative weight updates (OMWU) method conve... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies asynchronous gradient plays in zero-sum polymatrix games under delayed feedbacks, while significant efforts have been made to understand zero-sum two-player matrix games. They first establish that the last iterate of the entropy-regularized optimistic multiplicative weight updates (OMWU) meth... |
The paper presents PD-MORL as a fascinating new method to learn a single policy that can integrate task preferences and cover more of the Pareto front than any baseline algorithm. This differentiates it from previous methods that must learn a new policy for any particular preference. It demonstrates marginally better r... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper presents PD-MORL as a fascinating new method to learn a single policy that can integrate task preferences and cover more of the Pareto front than any baseline algorithm. This differentiates it from previous methods that must learn a new policy for any particular preference. It demonstrates marginally ... |
The authors propose a low-light image enhancement method RetinexUTV. The proposed method can estimate a noise level map and a illumination map by an unfolded total variational network and generate a noise-free reflection map guided by the learned noise level map. Finally, the output can be obtain by multiplying the noi... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The authors propose a low-light image enhancement method RetinexUTV. The proposed method can estimate a noise level map and a illumination map by an unfolded total variational network and generate a noise-free reflection map guided by the learned noise level map. Finally, the output can be obtain by multiplying... |
The paper proposes a new metric called #Circles for chemical spacing measurement. The authors analyzed and compared existing metrics and their metric mathematically, proving that their metric has two good properties. They also conducted experiment to analyze current datasets and molecule generation models by their metr... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a new metric called #Circles for chemical spacing measurement. The authors analyzed and compared existing metrics and their metric mathematically, proving that their metric has two good properties. They also conducted experiment to analyze current datasets and molecule generation models by th... |
The paper considers a threat model where an attacker attacks a subset of the test data of a classification system. To develop a defense, the formulate the learning problem for the classier as a min-max optimization where the loss is minimized over the most adversarial subset of a particular cardinality. The authors sho... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper considers a threat model where an attacker attacks a subset of the test data of a classification system. To develop a defense, the formulate the learning problem for the classier as a min-max optimization where the loss is minimized over the most adversarial subset of a particular cardinality. The aut... |
This paper develops a new domain generalization method by learning a linear robust predictor on top of finetuned features extracted by a deep network (trained with ERM). Authors first performed a simulated study suggesting that the ERM produces features that are informative enough and the failure of deep networks to ou... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper develops a new domain generalization method by learning a linear robust predictor on top of finetuned features extracted by a deep network (trained with ERM). Authors first performed a simulated study suggesting that the ERM produces features that are informative enough and the failure of deep networ... |
This paper proposes a new explanation method that improves the robustness of LIME to variations in perturbation methods and sample size. Technically, it leverages the invariant risk minimization (IRM) framework to design a new method for fitting the approximation model. Specifically, it first defines the environment co... | 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 proposes a new explanation method that improves the robustness of LIME to variations in perturbation methods and sample size. Technically, it leverages the invariant risk minimization (IRM) framework to design a new method for fitting the approximation model. Specifically, it first defines the enviro... |
The paper proposed a post-hoc and instance-level explainer called T-GNNExplainer for temporal graph neural network on continuous-time dynamic graphs. The paper provides experimental evaluations on the effectiveness of the proposed explainer on real-world and synthetic datasets.
Strength:
1. The paper proposes an inte... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposed a post-hoc and instance-level explainer called T-GNNExplainer for temporal graph neural network on continuous-time dynamic graphs. The paper provides experimental evaluations on the effectiveness of the proposed explainer on real-world and synthetic datasets.
Strength:
1. The paper proposes... |
This paper proposed an oriented object detector that used horizontal box annotations. The proposed method applied weakly- and self-supervised learning for predicting the angle of the object. The experimental results on standard benchmarks show the effectiveness of the proposed method.
## Strength:
### (1) The proposed ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed an oriented object detector that used horizontal box annotations. The proposed method applied weakly- and self-supervised learning for predicting the angle of the object. The experimental results on standard benchmarks show the effectiveness of the proposed method.
## Strength:
### (1) The p... |
This paper proposes a test-time batch normalization method to tackle domain shifts. The core idea is to learn interpolating vectors to combine running statistics and test-time batch statistics in batch normalization layers. The authors also propose to use a gradient distance score to initialize and regularize the inter... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a test-time batch normalization method to tackle domain shifts. The core idea is to learn interpolating vectors to combine running statistics and test-time batch statistics in batch normalization layers. The authors also propose to use a gradient distance score to initialize and regularize t... |
The authors present DISSECT, a method for estimating cell fraction and gene expression from bulk RNA sequencing data. The primary innovation in DISSECT is the introduction of a regularization term during training that encourages similar expression profiles between simulated bulk data (drawn from scRNA-seq data) and rea... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors present DISSECT, a method for estimating cell fraction and gene expression from bulk RNA sequencing data. The primary innovation in DISSECT is the introduction of a regularization term during training that encourages similar expression profiles between simulated bulk data (drawn from scRNA-seq data)... |
The paper analyzes the statistical efficiency of the score estimate obtained by the score matching objective and compares it with the maximum likelihood based approach. Specifically, relation between the loss optimized in score matching J has been compared to the KL divergence between the true distribution p and the es... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper analyzes the statistical efficiency of the score estimate obtained by the score matching objective and compares it with the maximum likelihood based approach. Specifically, relation between the loss optimized in score matching J has been compared to the KL divergence between the true distribution p an... |
The paper proposes a new objective for the problem of few-shot image synthesis based on extending implicit maximum likelihood estimation (IMLE). Specifically, they derive a data-adaptive IMLE objective for training the implicit generative model. Experiment results on six few-shot image synthesis datasets, namely Grumpy... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper proposes a new objective for the problem of few-shot image synthesis based on extending implicit maximum likelihood estimation (IMLE). Specifically, they derive a data-adaptive IMLE objective for training the implicit generative model. Experiment results on six few-shot image synthesis datasets, namel... |
Visual imitation learning algorithm using patch rewards is proposed. For patch rewards, input images from agent and expert are decomposed into small patches and classified by using multiple patch-wise discriminator. The output values of patch-wise discriminators are postprocessed (by $h$) and aggregated (by $Aggr$, e.g... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Visual imitation learning algorithm using patch rewards is proposed. For patch rewards, input images from agent and expert are decomposed into small patches and classified by using multiple patch-wise discriminator. The output values of patch-wise discriminators are postprocessed (by $h$) and aggregated (by $Ag... |
The paper proposes to use the energy function of modern Hopfield networks to detect out-of-distribution datapoints.
The premise of Hopfield networks is to place plausible data points at the local minima of the energy function. The authors propose to use this property for detecting datapoint that are located far away fr... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes to use the energy function of modern Hopfield networks to detect out-of-distribution datapoints.
The premise of Hopfield networks is to place plausible data points at the local minima of the energy function. The authors propose to use this property for detecting datapoint that are located far... |
This paper studies large-scale neural rendering. To tackle the challenges of lacking universal scene decomposition, not learnable decomposition procedure, and independent sub-networks optimization, the paper proposes Switch-NeRF, a new end-to-end large-scale NeRF with learning-based scene decomposition. It designs a Sp... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper studies large-scale neural rendering. To tackle the challenges of lacking universal scene decomposition, not learnable decomposition procedure, and independent sub-networks optimization, the paper proposes Switch-NeRF, a new end-to-end large-scale NeRF with learning-based scene decomposition. It desi... |
This paper proposes Contrastive Backdoor Defense (CBD), a novel backdoor defense that can exploit unlabeled data to remove potential backdoors in a provided model. It first generates untargeted adversarial examples via a contrastive loss. It then proposes to finetune the model using a Backdoor-to-Standard Pulling loss ... | 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 proposes Contrastive Backdoor Defense (CBD), a novel backdoor defense that can exploit unlabeled data to remove potential backdoors in a provided model. It first generates untargeted adversarial examples via a contrastive loss. It then proposes to finetune the model using a Backdoor-to-Standard Pulli... |
**Background**. This work focuses on the problem of so-called *neural set* functions, where the input is a set, the task could be clustering, classification etc., and the map from a set to an output is a neural network.
Since the input is a set, there are certain invariances that are relevant for this case, e.g., the... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
**Background**. This work focuses on the problem of so-called *neural set* functions, where the input is a set, the task could be clustering, classification etc., and the map from a set to an output is a neural network.
Since the input is a set, there are certain invariances that are relevant for this case, e... |
This paper proposes a domain adaptation method in test time. To do so, this paper proposes a semi-supervised slot-centric approach that combines slot attention and object scene representation transformer. The model adapts to a single test sample without supervision at test time. It is optimized for the self-supervised ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a domain adaptation method in test time. To do so, this paper proposes a semi-supervised slot-centric approach that combines slot attention and object scene representation transformer. The model adapts to a single test sample without supervision at test time. It is optimized for the self-sup... |
This paper aims to have a compression-aware minimizer that can compress dense DNN into a sparse one, and most importantly, without much performance drop. Inspired by the sharpness-aware minimizer, the authors leverage the flat minima to find the weights' coordinates that are stable to the perturbation. With the price o... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper aims to have a compression-aware minimizer that can compress dense DNN into a sparse one, and most importantly, without much performance drop. Inspired by the sharpness-aware minimizer, the authors leverage the flat minima to find the weights' coordinates that are stable to the perturbation. With the... |
The article defines a measure of Maximum Mean Discrepancy (MMD) for Graph Fourier features. it discusses how one could use the Laplacian of a graph suited to multivariate data so as to compute a GF-MMD. Some empirical results and applications are shown in an extensive numerical section.
Strength
- The proposed method... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The article defines a measure of Maximum Mean Discrepancy (MMD) for Graph Fourier features. it discusses how one could use the Laplacian of a graph suited to multivariate data so as to compute a GF-MMD. Some empirical results and applications are shown in an extensive numerical section.
Strength
- The propose... |
The paper regards generative retrieval that given a query generates important information about text documents from the parameter space of a trained model instead of relying on fixed-sized document embeddings which might be characterized by limited expressiveness. Such a generative model interacts with the parameters o... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper regards generative retrieval that given a query generates important information about text documents from the parameter space of a trained model instead of relying on fixed-sized document embeddings which might be characterized by limited expressiveness. Such a generative model interacts with the para... |
Motivated by the performance gap between self-supervised learning and its counter supervised model, this work provides a one-stage self-supervised small model pretraining protocol. The slimmable network idea is used to get the representations of one augmentation $x_1$. Weights are shared between the set of slimmable en... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
Motivated by the performance gap between self-supervised learning and its counter supervised model, this work provides a one-stage self-supervised small model pretraining protocol. The slimmable network idea is used to get the representations of one augmentation $x_1$. Weights are shared between the set of slim... |
In this paper, the author proposes a RGB SLAM model with neural implicit map representation. In contrast to previous competitors, e.g., iMap and NICE-SLAM requiring depth maps as input, the proposed approach takes only RGB image sequences as input and leverages the hierarchical feature volumes with more levels to boost... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the author proposes a RGB SLAM model with neural implicit map representation. In contrast to previous competitors, e.g., iMap and NICE-SLAM requiring depth maps as input, the proposed approach takes only RGB image sequences as input and leverages the hierarchical feature volumes with more levels ... |
The paper unifies several approximation to backpropagation through the energy-based models view.
### Strengths
The paper unifies several backprop approximation under the same energy-based framework.
This framework leads to different approximations.
### Weaknesses
I don’t understand what the take-home message is. P... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper unifies several approximation to backpropagation through the energy-based models view.
### Strengths
The paper unifies several backprop approximation under the same energy-based framework.
This framework leads to different approximations.
### Weaknesses
I don’t understand what the take-home messa... |
This paper introduces a new explainability framework for multi-variate time series data. This new explainability framework is a feature removal model. The authors motivate their approach by the fact that the observation of same feature over subsequent time steps is not independent and that feature can have varying leve... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper introduces a new explainability framework for multi-variate time series data. This new explainability framework is a feature removal model. The authors motivate their approach by the fact that the observation of same feature over subsequent time steps is not independent and that feature can have vary... |
This paper proposes an ensemble based label-propagation method for classification in the presence of label noise. In the method, the authors emphasize the importance of generating pseudo-labels only for datapoints a model has not seen during training. Empirical comparisons against existing algorithms are limited but pr... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes an ensemble based label-propagation method for classification in the presence of label noise. In the method, the authors emphasize the importance of generating pseudo-labels only for datapoints a model has not seen during training. Empirical comparisons against existing algorithms are limite... |
The paper develops a new model for olfactory receptors.
Positives:
- Tackling a new and interesting problem of predicting receptor-odorant binding.
- Proposing a new model involving GNN and BERT embedding for olfactory receptors.
- Thorough ablation study of the model.
- Gathering a new dataset for olfactory receptors.... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper develops a new model for olfactory receptors.
Positives:
- Tackling a new and interesting problem of predicting receptor-odorant binding.
- Proposing a new model involving GNN and BERT embedding for olfactory receptors.
- Thorough ablation study of the model.
- Gathering a new dataset for olfactory re... |
The paper proposes parameter efficient tuning for vision transformers. The method starts with large adapters and gradually iteratively reduces their size with pruning. The paper also introduces a scoring function that weights the importance of neurons. The method outperforms existing approaches with higher accuracy and... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes parameter efficient tuning for vision transformers. The method starts with large adapters and gradually iteratively reduces their size with pruning. The paper also introduces a scoring function that weights the importance of neurons. The method outperforms existing approaches with higher accu... |
The paper proposed a model to do multi-hop question-answering over Knowledge Graphs. The core model responsible for relevance is trained once and shared between the initial retrieval of the subgraph and the detailed reasoning phases. A key claim of the paper is that by unifying the model for the two phases, we can get ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposed a model to do multi-hop question-answering over Knowledge Graphs. The core model responsible for relevance is trained once and shared between the initial retrieval of the subgraph and the detailed reasoning phases. A key claim of the paper is that by unifying the model for the two phases, we ... |
This paper proposed a new training framework for a language model that represents text in a lexicon space. The main motivation is that the large-scale retrieval (first stage retrieval) requires a high retrieval performance and a low latency. The dense representation approach could not work smoothly with the indexing te... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed a new training framework for a language model that represents text in a lexicon space. The main motivation is that the large-scale retrieval (first stage retrieval) requires a high retrieval performance and a low latency. The dense representation approach could not work smoothly with the ind... |
In this paper, the authors proposed ways to do safe RL in an unknown stochastic environment. Specifically, they model the problem in hard constraints instead of the standard CMDP constraints and derived a novel barrier function method to incorporate the hard constraints into the RL algorithm. The authors manage to show... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In this paper, the authors proposed ways to do safe RL in an unknown stochastic environment. Specifically, they model the problem in hard constraints instead of the standard CMDP constraints and derived a novel barrier function method to incorporate the hard constraints into the RL algorithm. The authors manage... |
The paper proposes a new measure for internal clustering evaluation. To compute this measure, one first computes the differential entropy for each cluster (assuming that the elements are sampled from a multivariate normal distribution) and then subtracts the differential entropy for the cluster centers from the average... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes a new measure for internal clustering evaluation. To compute this measure, one first computes the differential entropy for each cluster (assuming that the elements are sampled from a multivariate normal distribution) and then subtracts the differential entropy for the cluster centers from the... |
This paper studies the performance of the RL algorithms within the low data regime. Extensive experiments are carried out to demonstrate how well the performance profiles influence the sample complexity.
### Strength:
- The paper is well-written and easy to follow. Extensive experiments well support the claimed resul... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the performance of the RL algorithms within the low data regime. Extensive experiments are carried out to demonstrate how well the performance profiles influence the sample complexity.
### Strength:
- The paper is well-written and easy to follow. Extensive experiments well support the claim... |
This paper proposed a method to automatically learn constraints over nodes in a graph, and apply the constraints for spatial-temporal forecasting. The major contribution is the proposed constraint learning framework.
Strength:
1. The motivation for exploring relationships between nodes in terms of constraints is insigh... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed a method to automatically learn constraints over nodes in a graph, and apply the constraints for spatial-temporal forecasting. The major contribution is the proposed constraint learning framework.
Strength:
1. The motivation for exploring relationships between nodes in terms of constraints i... |
This paper proposes a convex alternative to the attention operation in standard Transformers with the benefits of better optimization and interpretability. The analysis is insightful and deep, and can certainly enlighten further research in the community.
[ Strength ]
+ The motivation is clear and important. The inter... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes a convex alternative to the attention operation in standard Transformers with the benefits of better optimization and interpretability. The analysis is insightful and deep, and can certainly enlighten further research in the community.
[ Strength ]
+ The motivation is clear and important. T... |
In this work, the authors proposed a novel neural operator architecture that factorizes the convolution on Fourier space into separate dimensions. Consequentially, the F-FNO model can scale up to a higher number of layers and achieve smaller errors. The paper has a comprehensive numerical study on multiple types of par... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this work, the authors proposed a novel neural operator architecture that factorizes the convolution on Fourier space into separate dimensions. Consequentially, the F-FNO model can scale up to a higher number of layers and achieve smaller errors. The paper has a comprehensive numerical study on multiple type... |
This paper introduces a transformer-based MIMO framework, called MixViT. At training time, two subsets of tokens from two images are taken as input and the model is tasked to predict labels for the two input images. At test time, the same input is passed to the model multiple times with different sourced tokens. The fe... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper introduces a transformer-based MIMO framework, called MixViT. At training time, two subsets of tokens from two images are taken as input and the model is tasked to predict labels for the two input images. At test time, the same input is passed to the model multiple times with different sourced tokens... |
**Update 11-16-22** I acknowledge the revisions made by the authors. I believe the proposed revisions further improve this already-strong submission. I am holding my score at 8 Accept and am willing to further champion this paper's acceptance if needed.
This paper proposes an algorithm for improving temporal alignment... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
**Update 11-16-22** I acknowledge the revisions made by the authors. I believe the proposed revisions further improve this already-strong submission. I am holding my score at 8 Accept and am willing to further champion this paper's acceptance if needed.
This paper proposes an algorithm for improving temporal a... |
In this paper, authors present a method for pretraining a generalisable and resilient SSL model with control transformer for multi-task sequential decision making. They propose a self-supervised and control-centric objective that encourages the transformer-based model to capture control-relevant representation. The eva... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In this paper, authors present a method for pretraining a generalisable and resilient SSL model with control transformer for multi-task sequential decision making. They propose a self-supervised and control-centric objective that encourages the transformer-based model to capture control-relevant representation.... |
In their manuscript, "Predictive coding with approximate Laplace Monte Carlo", the authors propose a modification to the predictive coding method that is designed to improve its performance for multi-layer deep learning style models.
Whereas I usually being with the strengths, I will begin with the weaknesses here. The... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
In their manuscript, "Predictive coding with approximate Laplace Monte Carlo", the authors propose a modification to the predictive coding method that is designed to improve its performance for multi-layer deep learning style models.
Whereas I usually being with the strengths, I will begin with the weaknesses h... |
The authors propose a method to learn representation that is invariant to sensitive groups (or domains as claimed by the authors). This is achieved through a network structure that has a transformer to generate time-wise attention, domain-specific feature-wise attention (to extract features that are specific to each do... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a method to learn representation that is invariant to sensitive groups (or domains as claimed by the authors). This is achieved through a network structure that has a transformer to generate time-wise attention, domain-specific feature-wise attention (to extract features that are specific to... |
This paper studies the convergence of stochastic gradient descent ascend (SGDA) for nonconvex-PL minimax problems, where the data points are randomly shuffled and sampled without replacement during the training. The authors provided strong theoretical convergence of SGDA and confirm the empirical observation that rando... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies the convergence of stochastic gradient descent ascend (SGDA) for nonconvex-PL minimax problems, where the data points are randomly shuffled and sampled without replacement during the training. The authors provided strong theoretical convergence of SGDA and confirm the empirical observation th... |
The paper is about open-vocabulary object detection (OVD), where detection models are trained from a set of base categories with bounding box annotations, as well as a set of image-caption pairs. Like prior works, this paper leverages existing, pre-trained vision & language models (VLM) like CLIP that were trained on l... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper is about open-vocabulary object detection (OVD), where detection models are trained from a set of base categories with bounding box annotations, as well as a set of image-caption pairs. Like prior works, this paper leverages existing, pre-trained vision & language models (VLM) like CLIP that were trai... |
This work presents a interesting approach to achieve group robustness without group label by re-sampling training data based on last-layer gradients, the work is simple, comes with rigorous theoretical justification and extensive experiments.
Strength:
* Comparing to previous approach, this current paper proposed a di... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This work presents a interesting approach to achieve group robustness without group label by re-sampling training data based on last-layer gradients, the work is simple, comes with rigorous theoretical justification and extensive experiments.
Strength:
* Comparing to previous approach, this current paper propo... |
The paper proposes TINA, an adaptation mechanism that starts from large adapter layers and interactively reduce their capacity by target domain. New dense layers are inject after each block in a VIT-family backbone (after every MLP and every Multi-head Self-attention block). In practice, the authors adopt a pretrained ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes TINA, an adaptation mechanism that starts from large adapter layers and interactively reduce their capacity by target domain. New dense layers are inject after each block in a VIT-family backbone (after every MLP and every Multi-head Self-attention block). In practice, the authors adopt a pre... |
This paper proposed a sampling method for SGD that takes into account the importance of training examples. Specifically, the proposed method trains using uniform sampling for the first E epochs, and then compute importance scores for all training examples based on how accurate the current model’s prediction is. The pro... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposed a sampling method for SGD that takes into account the importance of training examples. Specifically, the proposed method trains using uniform sampling for the first E epochs, and then compute importance scores for all training examples based on how accurate the current model’s prediction is.... |
Spiking Neural Networks mimic the biological nervous systems for ultra-low-power NN execution.
However, it can be sensitive to the weight perturbations that may lead to significant performance drop.
The paper aims to understand the causal relationship between the perturbations and the performance drop to devise a Weigh... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
Spiking Neural Networks mimic the biological nervous systems for ultra-low-power NN execution.
However, it can be sensitive to the weight perturbations that may lead to significant performance drop.
The paper aims to understand the causal relationship between the perturbations and the performance drop to devise... |
This paper analyze the effect of attention on emergent communication. The paper hypothesizes that the attention mechanism will bias agents towards more compositional encodings. Experimental results on a small-scale symbolic game and image reference game derived from Fashion-MNIST demonstrate that attentive speaker and/... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper analyze the effect of attention on emergent communication. The paper hypothesizes that the attention mechanism will bias agents towards more compositional encodings. Experimental results on a small-scale symbolic game and image reference game derived from Fashion-MNIST demonstrate that attentive spea... |
This paper is about training a multi-scale diffusion model in a unified framework. Specifically, it uses a casacaded mutli-scale generation process (from lower-resolution to higher-resolution). Unlike prior work which uses several score networks at each scale level. This paper uses a single network and proposes to use ... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper is about training a multi-scale diffusion model in a unified framework. Specifically, it uses a casacaded mutli-scale generation process (from lower-resolution to higher-resolution). Unlike prior work which uses several score networks at each scale level. This paper uses a single network and proposes... |
The authors introduce a setting and game by which they study the emergence of conversational strategies when eliciting and discovering information. In particular, the authors show that commonly observed (among humans and animals) turn-taking behavior does not automatically arise in agents trained on the game, but that ... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The authors introduce a setting and game by which they study the emergence of conversational strategies when eliciting and discovering information. In particular, the authors show that commonly observed (among humans and animals) turn-taking behavior does not automatically arise in agents trained on the game, b... |
This work focuses on the model poisoning and backdoor attacks in federated learning. It proposes an attack that anticipates and accounts for the entire federated learning pipeline, including behaviors of other clients.
The main contributions can be summarized as:
(1) Proposing a backdoor attack in federated learning ... | 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 focuses on the model poisoning and backdoor attacks in federated learning. It proposes an attack that anticipates and accounts for the entire federated learning pipeline, including behaviors of other clients.
The main contributions can be summarized as:
(1) Proposing a backdoor attack in federated l... |
The paper presents a method for selecting subsets of prompted multitask training data for fine-tuning a language model (Data-Efficient FineTuning, or DEFT). Given a small set of unlabeled task examples, the paper uses nearest neighbors to retrieve a targeted task subset from a larger pool of prompted examples. Using t... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a method for selecting subsets of prompted multitask training data for fine-tuning a language model (Data-Efficient FineTuning, or DEFT). Given a small set of unlabeled task examples, the paper uses nearest neighbors to retrieve a targeted task subset from a larger pool of prompted examples.... |
This paper propose a meta mapper network to tackle the task of multi-modal few-shot learning. Particularly, the proposed meta mapper contextualizes the soft visual prompts with the features of a pre-trained/frozen visual encoder and language models, for adapting the language model to output for the target task. During ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper propose a meta mapper network to tackle the task of multi-modal few-shot learning. Particularly, the proposed meta mapper contextualizes the soft visual prompts with the features of a pre-trained/frozen visual encoder and language models, for adapting the language model to output for the target task.... |
This paper considers a special case of multiclass classification problem. Here, bag level labels and counts of each class are available. The paper proposes a two-stage algorithm to solve the new defined problem.
Strength
- This paper draws attention on considering new problem settings for multiclass classification.
We... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper considers a special case of multiclass classification problem. Here, bag level labels and counts of each class are available. The paper proposes a two-stage algorithm to solve the new defined problem.
Strength
- This paper draws attention on considering new problem settings for multiclass classificat... |
The authors study the highly important problem of transfer learning with DNNs on tabular data. This avenue has huge potential in many practical applications. The problem and DNNs for tabular data are generally underexplored so this contribution could be important for the community. They propose a way to use DNNs for tr... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The authors study the highly important problem of transfer learning with DNNs on tabular data. This avenue has huge potential in many practical applications. The problem and DNNs for tabular data are generally underexplored so this contribution could be important for the community. They propose a way to use DNN... |
This paper proposes a novel zero-shot framework for solving diverse linear image restoration (IR) problems, named Denoising Diffusion Null-Space Model (DDNM). Firstly, it provides detailed theory analysis for using a pre-trained diffusion model to solve linear IR problems. Secondly, the proposed DDNM can well solve tho... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a novel zero-shot framework for solving diverse linear image restoration (IR) problems, named Denoising Diffusion Null-Space Model (DDNM). Firstly, it provides detailed theory analysis for using a pre-trained diffusion model to solve linear IR problems. Secondly, the proposed DDNM can well s... |
This paper proposed a deep learning-based group-level (GL) neural decoding model that can be adapted to different subjects. The key module of the GL model is the learnable subject embeddings. Experiments were conducted on an MEG neural dataset. Experimental results indicated that the proposed GL model can achieve signi... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper proposed a deep learning-based group-level (GL) neural decoding model that can be adapted to different subjects. The key module of the GL model is the learnable subject embeddings. Experiments were conducted on an MEG neural dataset. Experimental results indicated that the proposed GL model can achie... |
This paper develops a new FL solution. In the designed solution, the local clients distill meta knowledge based on local private data and shared knowledge from other clients, the distilled meta knowledge is uploaded to the server for global model training. As the global model training is based on meta knowledge from al... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper develops a new FL solution. In the designed solution, the local clients distill meta knowledge based on local private data and shared knowledge from other clients, the distilled meta knowledge is uploaded to the server for global model training. As the global model training is based on meta knowledge... |
Firstly, in this paper, the authors purpose that ACMP can help solve the over-smoothing problem in the GNN network by keeping the Dirichlet energy of GNN. Then, this paper shows that adding the term repulsive force may cause the particles to be pushed away to infinite, the Dirichlet energy becomes unbounded, and how th... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Firstly, in this paper, the authors purpose that ACMP can help solve the over-smoothing problem in the GNN network by keeping the Dirichlet energy of GNN. Then, this paper shows that adding the term repulsive force may cause the particles to be pushed away to infinite, the Dirichlet energy becomes unbounded, an... |
This work proposes PG-TD, which integrates Transformer with MCTS for code generation. PG-TD works by generating a tree of possible programs, first generating high likelihood sequences and then evaluating them on test problems. It then selects the best and searches for more sequences until it completes the program. The ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This work proposes PG-TD, which integrates Transformer with MCTS for code generation. PG-TD works by generating a tree of possible programs, first generating high likelihood sequences and then evaluating them on test problems. It then selects the best and searches for more sequences until it completes the progr... |
The paper proposes an auxiliary data augmentation technique that can be used on top of several long-tailed recognition methods. The main insight of the paper is that class-wise augmentation actually improves the performance of non-augmented classes. This is an important as well as counter-intuitive observation. The aut... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes an auxiliary data augmentation technique that can be used on top of several long-tailed recognition methods. The main insight of the paper is that class-wise augmentation actually improves the performance of non-augmented classes. This is an important as well as counter-intuitive observation.... |
This paper presents a differentiable easy-first beam search for structure induction, called Beam Tree Recursive Cells (BT-RC). A couple methods are presented to handle the sparse gradient issue in beam search. There are two sources of sparsity: parent composition and the topk filtering of beams. The two groups of metho... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a differentiable easy-first beam search for structure induction, called Beam Tree Recursive Cells (BT-RC). A couple methods are presented to handle the sparse gradient issue in beam search. There are two sources of sparsity: parent composition and the topk filtering of beams. The two groups ... |
This paper considers combining multi-objective optimization with multiple-hypothesis testing. More specifically, they consider optimizing some objectives Q_i's while applying risk control on multiple other objectives such that Q_j\le c with high probability. They consider reducing the dimension of the high dimensional ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper considers combining multi-objective optimization with multiple-hypothesis testing. More specifically, they consider optimizing some objectives Q_i's while applying risk control on multiple other objectives such that Q_j\le c with high probability. They consider reducing the dimension of the high dime... |
The paper analyzes the behavior of CLIP towards various compositions of hand-crafted text prompts and puts out some observations. One of the key observations of the paper is the behavior of CLIP to negative prompts, in which the authors insert a “not” word to the prompt, e.g. “this is not a photo of a dog”. The paper s... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper analyzes the behavior of CLIP towards various compositions of hand-crafted text prompts and puts out some observations. One of the key observations of the paper is the behavior of CLIP to negative prompts, in which the authors insert a “not” word to the prompt, e.g. “this is not a photo of a dog”. The... |
This work proposes a Single-dataset Unified Generalization (SUG) framework for the challenging one-to-many domain generalization task in 3D point clouds. A Multi-grained Sub-domain Alignment (MSA) is proposed for spliting the single dataset into multiple sub-domains and then constraining the learned representations to ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This work proposes a Single-dataset Unified Generalization (SUG) framework for the challenging one-to-many domain generalization task in 3D point clouds. A Multi-grained Sub-domain Alignment (MSA) is proposed for spliting the single dataset into multiple sub-domains and then constraining the learned representat... |
This paper proposed SONew (Sparsified Online Newton), which is an efficient variant of online Newton step (ONS) by using $b$-banded sparsity graph. The time complexity of this method is $O(b^3 n)$, and the regret bound is $O(T^{3/4})$. The authors provide some simple experiments to verify the performance of the propose... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposed SONew (Sparsified Online Newton), which is an efficient variant of online Newton step (ONS) by using $b$-banded sparsity graph. The time complexity of this method is $O(b^3 n)$, and the regret bound is $O(T^{3/4})$. The authors provide some simple experiments to verify the performance of the... |
The paper proposes a meta-learning framework for temporal point processes (TPPs). The paper points out a limitation of the existing TPP models - treating all event sequences in the dataset as realization of the same process. The proposed meta-learning approach addresses this limitation and leads to improved performance... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes a meta-learning framework for temporal point processes (TPPs). The paper points out a limitation of the existing TPP models - treating all event sequences in the dataset as realization of the same process. The proposed meta-learning approach addresses this limitation and leads to improved per... |
This paper introduces a new neural ODE-based framework to model irregularly-sampled sequential data. The approach is based on neural control differential equations, and on the observation that attention models can be viewed as integrating a differential equation path. The authors evaluate their approach on a few toy ex... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper introduces a new neural ODE-based framework to model irregularly-sampled sequential data. The approach is based on neural control differential equations, and on the observation that attention models can be viewed as integrating a differential equation path. The authors evaluate their approach on a fe... |
The paper proposes a new probabilistic decoding method for generating a sequence of text from a trained autoregressive model $p(x_{t+1}|x_{1:t})$. The method, IQR-IP, is carried out by first doing top-p and top-k filtering on the possible generations, then computing the quantiles of the remaining tokens. For quartiles ... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a new probabilistic decoding method for generating a sequence of text from a trained autoregressive model $p(x_{t+1}|x_{1:t})$. The method, IQR-IP, is carried out by first doing top-p and top-k filtering on the possible generations, then computing the quantiles of the remaining tokens. For qu... |
This work proposes a cross-modality knowledge distillation framework for multi-view 3D object detection, called BEVDistill. It introduces two types of feature distillation, i.e., dense feature distillation and sparse instance distillation. Experiments prove the effectiveness of the proposed modules on the nuScenes data... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes a cross-modality knowledge distillation framework for multi-view 3D object detection, called BEVDistill. It introduces two types of feature distillation, i.e., dense feature distillation and sparse instance distillation. Experiments prove the effectiveness of the proposed modules on the nuSce... |
The authors of the paper derive an estimator of a term that lower bounds the grouping loss term of a proper scoring rule. They show that this estimator is tighter than baselines and use it to analyze existing models. The derivations seem novel and correct; however, it is unclear to me why having this estimator is inter... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors of the paper derive an estimator of a term that lower bounds the grouping loss term of a proper scoring rule. They show that this estimator is tighter than baselines and use it to analyze existing models. The derivations seem novel and correct; however, it is unclear to me why having this estimator ... |
The paper presents a method for domain generalization. The core assumption is that every domain can be decomposed to some elementary domain (something like a base domains) which are invariant. Every domain is a linear combination of these base domains. With this assumption, we can have elementary domain prediction func... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The paper presents a method for domain generalization. The core assumption is that every domain can be decomposed to some elementary domain (something like a base domains) which are invariant. Every domain is a linear combination of these base domains. With this assumption, we can have elementary domain predict... |
In this work, the authors propose to use a meta learning way to train neural processes aiming at one-step prediction of event time of point processes. The proposed TPPs employ a latent variable layer and attention mechanism to improve the prediction performance. The method was demonstrated with benchmark datasets.
Stre... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
In this work, the authors propose to use a meta learning way to train neural processes aiming at one-step prediction of event time of point processes. The proposed TPPs employ a latent variable layer and attention mechanism to improve the prediction performance. The method was demonstrated with benchmark datase... |
This paper considers chain-of-thought (CoT) in multilingual setup and shows that despite a highly imbalanced pretraining dataset, LLM still learns strong CoT capability in non-English. It extends an existing dataset to multiple languages. It also compares different ways to construct exemplar and CoT.
Strength
It intro... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper considers chain-of-thought (CoT) in multilingual setup and shows that despite a highly imbalanced pretraining dataset, LLM still learns strong CoT capability in non-English. It extends an existing dataset to multiple languages. It also compares different ways to construct exemplar and CoT.
Strength
... |
The paper proposes an optimal transport-based method to learn dictionaries that generalize over datasets (this is a domain adaptation problem). They propose two approaches: one to reconstruct the samples based on the learned barycenters, and another based on an ensemble of classifiers that are learned based on the dict... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes an optimal transport-based method to learn dictionaries that generalize over datasets (this is a domain adaptation problem). They propose two approaches: one to reconstruct the samples based on the learned barycenters, and another based on an ensemble of classifiers that are learned based on ... |
This work trains a generative model (conditional diffusion Transformer) called G.pt over NN checkpoints of supervised + reinforcement learning (RL) tasks such that given a prompt of (initial input parameter vector, target loss/error/return), the generative model outputs an updated parameter vector that achieves the des... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This work trains a generative model (conditional diffusion Transformer) called G.pt over NN checkpoints of supervised + reinforcement learning (RL) tasks such that given a prompt of (initial input parameter vector, target loss/error/return), the generative model outputs an updated parameter vector that achieves... |
The paper studies the computational bottleneck problem of Transformer Neural Processes (TNPs) that results from the quadratic complexity of transformers with respect to the input sequence length. The paper aims to combine the best of the worlds of state-of-the-art expressive power at the expense of high compute cost (T... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper studies the computational bottleneck problem of Transformer Neural Processes (TNPs) that results from the quadratic complexity of transformers with respect to the input sequence length. The paper aims to combine the best of the worlds of state-of-the-art expressive power at the expense of high compute... |
The paper studies Continual Learning (CL) under a specific framework called SCoLe (Scaling Continual Learning). The major difference between SCoLe and CL is the sparse and controllable appearance of previous tasks and data when the number of tasks is very large. In addition, the paper proposes under the SCoLe framewor... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper studies Continual Learning (CL) under a specific framework called SCoLe (Scaling Continual Learning). The major difference between SCoLe and CL is the sparse and controllable appearance of previous tasks and data when the number of tasks is very large. In addition, the paper proposes under the SCoLe ... |
This paper proposes a parameter and data efficient network architecture and the corresponding training procedure for personalization and federated learning. The proposed architecture includes FiLM layers in the pretrained backbones and takes Naive Bayes classifier as the head layer.
The paper is well motivated as both... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a parameter and data efficient network architecture and the corresponding training procedure for personalization and federated learning. The proposed architecture includes FiLM layers in the pretrained backbones and takes Naive Bayes classifier as the head layer.
The paper is well motivated... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.