review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This work uses a diffusion model as an expressive generative model for behavior cloning. They then turn this into a offline RL algorithm with model improvement by learning a Q-function and performing a form of importance sampling to rejection sample action samples with high value by sampling from the behavior clone mod... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This work uses a diffusion model as an expressive generative model for behavior cloning. They then turn this into a offline RL algorithm with model improvement by learning a Q-function and performing a form of importance sampling to rejection sample action samples with high value by sampling from the behavior c... |
This paper proposes a new metric, called susceptibility, along with training acc to select good models trained on datasets with label noise. Authors observe that models with high training accuracy and low susceptibility will lead to higher acc on the clean test set. The authors also provide a convergence analysis of th... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new metric, called susceptibility, along with training acc to select good models trained on datasets with label noise. Authors observe that models with high training accuracy and low susceptibility will lead to higher acc on the clean test set. The authors also provide a convergence analys... |
This paper proposes a new functional form for neural scaling laws that corrects for two weaknesses in previous functions: (1) that they could only model strict monotonic behaviour, and (2) they could not express inflection points. The function is intuitive and well-designed and outperforms previous scaling law function... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a new functional form for neural scaling laws that corrects for two weaknesses in previous functions: (1) that they could only model strict monotonic behaviour, and (2) they could not express inflection points. The function is intuitive and well-designed and outperforms previous scaling law ... |
This paper provides a framework for domain adaptation in graph neural networks. Algorithmically, the framework relies on addition of two regularizer terms to the loss function. The construction of the regularizer terms is motivated from theoretical results derived using the theory of optimal transport-based domain adap... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper provides a framework for domain adaptation in graph neural networks. Algorithmically, the framework relies on addition of two regularizer terms to the loss function. The construction of the regularizer terms is motivated from theoretical results derived using the theory of optimal transport-based dom... |
This paper applies knowledge distillation to CLIP for image classification and finds that student performance increases when using CLIP as a teacher despite the teacher being a larger size. Some initial explanatory answers are proposed.
Strengths:
- The experiments that are presented seem to be well executed.
- The pro... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper applies knowledge distillation to CLIP for image classification and finds that student performance increases when using CLIP as a teacher despite the teacher being a larger size. Some initial explanatory answers are proposed.
Strengths:
- The experiments that are presented seem to be well executed.
-... |
This paper presents coresets for a robust version of the $k$-clustering problem where the cost of clustering is taken over the dataset after removing $m$ outliers. The coreset size is polynomial in $k$ the number of clusters and linear in $m$, the number of outliers. The paper also shows a lower bound on the coreset si... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper presents coresets for a robust version of the $k$-clustering problem where the cost of clustering is taken over the dataset after removing $m$ outliers. The coreset size is polynomial in $k$ the number of clusters and linear in $m$, the number of outliers. The paper also shows a lower bound on the co... |
This work develops a deep network for assessing diabetic retinopathy (DR) in retina fundus images by implementing a multi-scale attention mechanism, and a brand-new loss function.
Experimental results show that the effectiveness of the developed method.
Strengths:
1. This work develops a CNN based on a multi-scale att... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work develops a deep network for assessing diabetic retinopathy (DR) in retina fundus images by implementing a multi-scale attention mechanism, and a brand-new loss function.
Experimental results show that the effectiveness of the developed method.
Strengths:
1. This work develops a CNN based on a multi-s... |
The composition of the encoder and decoder mapping of standard Gaussian variational autoencoders (VAE) is known to "explain away" high frequency image components as independent Gaussian noise. The authors propose to address the resulting blur in reconstructed and generated images directly, by adding a blur kernel to t... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The composition of the encoder and decoder mapping of standard Gaussian variational autoencoders (VAE) is known to "explain away" high frequency image components as independent Gaussian noise. The authors propose to address the resulting blur in reconstructed and generated images directly, by adding a blur ker... |
The paper studies the problem of computing an approximate Nash equilibrium in congestion games from offline game data.
The authors consider three different types of data, depending on whether rewards are observed at the facility level, player level, or game level, and analyze their corresponding hardness in terms of re... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper studies the problem of computing an approximate Nash equilibrium in congestion games from offline game data.
The authors consider three different types of data, depending on whether rewards are observed at the facility level, player level, or game level, and analyze their corresponding hardness in ter... |
The paper suggests that an important way to improve sample efficiency in reinforcement learning comes from bootstrapping from accurate target estimates. Although multistep methods such as TD(lambda) offer a bias-variance trade-off, the authors argue that such algorithms fail to exploit the underlying graph structure of... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper suggests that an important way to improve sample efficiency in reinforcement learning comes from bootstrapping from accurate target estimates. Although multistep methods such as TD(lambda) offer a bias-variance trade-off, the authors argue that such algorithms fail to exploit the underlying graph stru... |
This paper combines the REP-UCB and RF-UCRL to obtain an improved sample complexity for reward-free reinforcement learning under the structural assumption on Low-rank MDPs.
### Strength:
* The idea is simple and easy to follow.
* The interpretation of the lower bound is interesting.
### Weakness:
* The proof of the l... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper combines the REP-UCB and RF-UCRL to obtain an improved sample complexity for reward-free reinforcement learning under the structural assumption on Low-rank MDPs.
### Strength:
* The idea is simple and easy to follow.
* The interpretation of the lower bound is interesting.
### Weakness:
* The proof ... |
This paper demonstrates that for simple algorithmic problems (arithmetic), language models can be taught to split problems into multiple subproblems, which can then be fed back to the LM to be solved independently. With very small models this can achieve great performance on these tasks, and it circumvents the limit in... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper demonstrates that for simple algorithmic problems (arithmetic), language models can be taught to split problems into multiple subproblems, which can then be fed back to the LM to be solved independently. With very small models this can achieve great performance on these tasks, and it circumvents the ... |
This paper proposes an approach, called AIA, to learn algorithm design with the aid of neural networks. We consider the minimum weighted set cover problem (WSCP), one of the NP-hard problems, as an representative example.
Strengths:
This is an important research area, and the paper has many aspects of novelty.
W... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes an approach, called AIA, to learn algorithm design with the aid of neural networks. We consider the minimum weighted set cover problem (WSCP), one of the NP-hard problems, as an representative example.
Strengths:
This is an important research area, and the paper has many aspects of nov... |
The authors propose algorithms for federated learning (decentralized data) alongside the presence of a centralized dataset (termed "mixed federated learning"). The paper presents 3 algorithms based on gradient exchange between the clients and the server, benchmarks the corresponding performance on 3 tasks (simple class... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors propose algorithms for federated learning (decentralized data) alongside the presence of a centralized dataset (termed "mixed federated learning"). The paper presents 3 algorithms based on gradient exchange between the clients and the server, benchmarks the corresponding performance on 3 tasks (simp... |
The paper extends an SDP relaxation for k-means clustering (Gaussian mixture modeling) from isotropic clusters (with identical scaled identity covariances) to non-isotropic clusters, where each cluster can have its own covariance matrix. When these cluster covariances are unknown (which is most often the case) the pap... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper extends an SDP relaxation for k-means clustering (Gaussian mixture modeling) from isotropic clusters (with identical scaled identity covariances) to non-isotropic clusters, where each cluster can have its own covariance matrix. When these cluster covariances are unknown (which is most often the case)... |
This paper proposes the sample complexity and error bound for MAML applied on linear meta-learning problem. While the features are iid and the tasks apply a subset of features, one can learn the useful features and apply them for downstream learning. The overfitting would not hurt generalization.
Although the model is ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proposes the sample complexity and error bound for MAML applied on linear meta-learning problem. While the features are iid and the tasks apply a subset of features, one can learn the useful features and apply them for downstream learning. The overfitting would not hurt generalization.
Although the m... |
In this paper authors propose to leverage entropy semiring and alignment entropy to improve the performance of neural speech recognition, via regularization or distillation. Experimental results show its effectiveness. There are also open-source contributions based on this work.
Strength: The idea of regularization or ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper authors propose to leverage entropy semiring and alignment entropy to improve the performance of neural speech recognition, via regularization or distillation. Experimental results show its effectiveness. There are also open-source contributions based on this work.
Strength: The idea of regulariza... |
In this paper, the authors investigate a problem in clustering by locality-sensitive hashing for similarity search. They introduce a new method called the Polar Code Nearest Neighbor (PCNN) method that uses the polar codes to maintain a number of clusters in a high-dimensional embedding space. By utilizing the list-dec... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
In this paper, the authors investigate a problem in clustering by locality-sensitive hashing for similarity search. They introduce a new method called the Polar Code Nearest Neighbor (PCNN) method that uses the polar codes to maintain a number of clusters in a high-dimensional embedding space. By utilizing the ... |
This manuscript proposes Corrupted Image Modeling, which is a self-supervised learning framework for generic architectures, e.g., CNN and ViT. Specifically, the proposed method reconstructs or predicts the original image from a generated image whose are partially reconstructed from [MASK] tokens. The extensive experime... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This manuscript proposes Corrupted Image Modeling, which is a self-supervised learning framework for generic architectures, e.g., CNN and ViT. Specifically, the proposed method reconstructs or predicts the original image from a generated image whose are partially reconstructed from [MASK] tokens. The extensive ... |
This paper proposes a new metric for measuring fairness of language
models using text toxicity level and perplexity. It shows that the new
metric correlates well with other gender-specific metrics in the
literature. Using this new metric, a comprehensive study of 24 models
is performed, along an analysis how the depth/... | 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 metric for measuring fairness of language
models using text toxicity level and perplexity. It shows that the new
metric correlates well with other gender-specific metrics in the
literature. Using this new metric, a comprehensive study of 24 models
is performed, along an analysis how th... |
Authors introduce a method to learn strides by backpropagation.
Strengths:
* Learning strides by backpropagation is still an open question as previous work that address this task suffer from a significant increase in computation cost compared to standard strided convolutions.
Weaknesses:
* The paper needs a complete ... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
Authors introduce a method to learn strides by backpropagation.
Strengths:
* Learning strides by backpropagation is still an open question as previous work that address this task suffer from a significant increase in computation cost compared to standard strided convolutions.
Weaknesses:
* The paper needs a c... |
This work presents a novel neural-networks-based algorithm to compute optimal transport maps and plans for both strong and weak transport costs. The problem is converted to a minmax optimization one using noise outsourcing idea. The work proves that the proposed neural-networks are universal approximators of transport ... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This work presents a novel neural-networks-based algorithm to compute optimal transport maps and plans for both strong and weak transport costs. The problem is converted to a minmax optimization one using noise outsourcing idea. The work proves that the proposed neural-networks are universal approximators of tr... |
The paper proposes a method for EEG reconstruction based on CNN.
Strengths:
- Performs well according to reported results
Weaknesses:
- The method is fairly simple simple based a CNN with mirrorred layers that are referred to as encoder and decoder
- There are just two baselines and its surprising that such a simple ... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper proposes a method for EEG reconstruction based on CNN.
Strengths:
- Performs well according to reported results
Weaknesses:
- The method is fairly simple simple based a CNN with mirrorred layers that are referred to as encoder and decoder
- There are just two baselines and its surprising that such a... |
This paper tried to train the deep transformer without skip connection and/or normalisation layers.
This paper tried to train the deep transformer directly by combining parameter initialisation, bias matrices, and location-dependent rescaling in the signal propagation view.
The performance indicates that though the pro... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper tried to train the deep transformer without skip connection and/or normalisation layers.
This paper tried to train the deep transformer directly by combining parameter initialisation, bias matrices, and location-dependent rescaling in the signal propagation view.
The performance indicates that though... |
In this work, the authors study Byzantine-robust distributed learning on heterogeneous data and propose a new method called linear scalarization, where derailed clients are penalized via a trade-off vector. The proposed method is empirically compared with existing methods on several datasets.
The proposed method, calle... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
In this work, the authors study Byzantine-robust distributed learning on heterogeneous data and propose a new method called linear scalarization, where derailed clients are penalized via a trade-off vector. The proposed method is empirically compared with existing methods on several datasets.
The proposed metho... |
In this paper, the authors focus on benchmarking the fairness issue in the field of medical images, taking different algorithms, datasets and models into consideration. They first choose 9 different datasets, covering the different categories of medical images, and then implement 11 different algorithms that aim to mit... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
In this paper, the authors focus on benchmarking the fairness issue in the field of medical images, taking different algorithms, datasets and models into consideration. They first choose 9 different datasets, covering the different categories of medical images, and then implement 11 different algorithms that ai... |
The paper uses a NeRF backbone with CLIP-based text guidance to synthesize 3D objects. Although most components are borrowed from its prior work, Dream Fields, the paper focuses on the sampling algorithm, and shows superior performance on both qualitative and quantitative results.
Strengths:
1. The visual quality is ob... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper uses a NeRF backbone with CLIP-based text guidance to synthesize 3D objects. Although most components are borrowed from its prior work, Dream Fields, the paper focuses on the sampling algorithm, and shows superior performance on both qualitative and quantitative results.
Strengths:
1. The visual quali... |
This paper studies contrastive learning from a theoretical perspective. Previous works on this topic showed the benefits of contrastive learning, assuming the pretraining loss is minimized over all sets of functions with no specific form. In this work, it is assumed that the pretraining loss is minimized by certain cla... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies contrastive learning from a theoretical perspective. Previous works on this topic showed the benefits of contrastive learning, assuming the pretraining loss is minimized over all sets of functions with no specific form. In this work, it is assumed that the pretraining loss is minimized by cer... |
In this work, the authors introduce and study an interesting and pragmatic variant of the popular Correlation Clustering (Minimizing Disagreements) problem. In this problem, we are given a limited budget on the number of queries allowed to make to an expensive oracle that given an edge returns the true labeling of the ... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
In this work, the authors introduce and study an interesting and pragmatic variant of the popular Correlation Clustering (Minimizing Disagreements) problem. In this problem, we are given a limited budget on the number of queries allowed to make to an expensive oracle that given an edge returns the true labeling... |
This paper studies the ability of GNNs to solve linear programming problems. It is quite systematic and proves that within a well defined class of GNNs, and given a set of LPs, there exist networks (with specified weights) which can distinguish pairs of LP problems within the set with different feasibility and objecti... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the ability of GNNs to solve linear programming problems. It is quite systematic and proves that within a well defined class of GNNs, and given a set of LPs, there exist networks (with specified weights) which can distinguish pairs of LP problems within the set with different feasibility and... |
Proposes a modification of the data-augmentation approach to self-supervised learning where the parameters of the transformation used in data-augmentation can be used by the network to predict the latent variables of the un-augmented input. This is advantageous when there exists a set of input features (such as color ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
Proposes a modification of the data-augmentation approach to self-supervised learning where the parameters of the transformation used in data-augmentation can be used by the network to predict the latent variables of the un-augmented input. This is advantageous when there exists a set of input features (such a... |
The authors propose a deep learning model for the task of subgraph counting. This is a challenging algorithmic problem and a well-known #P problem with applications in biology and social science. Deep learning approaches for subgraph counting have been proposed before, but the model presented by the authors improves bo... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a deep learning model for the task of subgraph counting. This is a challenging algorithmic problem and a well-known #P problem with applications in biology and social science. Deep learning approaches for subgraph counting have been proposed before, but the model presented by the authors imp... |
The paper describes Multi-Agent Joint-Predictive representations (MAJOR), which is a self-supervised learning mechanism that allows the MARL system to learn policies in a data-efficient manner.
In MAJOR, observations obtained by individual agents are treated as a masked sequence for representation learning, i.e., maske... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper describes Multi-Agent Joint-Predictive representations (MAJOR), which is a self-supervised learning mechanism that allows the MARL system to learn policies in a data-efficient manner.
In MAJOR, observations obtained by individual agents are treated as a masked sequence for representation learning, i.e... |
This paper studies the convergence performance of FedAvg for training (pyramidal topology) overparameterized deep networks. Specifically, under certain assumptions, the authors design a special initialization strategy for FedAvg and prove linear convergence rate of FedAvg on training such overparameterized models witho... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies the convergence performance of FedAvg for training (pyramidal topology) overparameterized deep networks. Specifically, under certain assumptions, the authors design a special initialization strategy for FedAvg and prove linear convergence rate of FedAvg on training such overparameterized mode... |
The paper proposes a new analysis of Wasserstein Autoencoders and claims the new analysis reveals a learning objective form that naturally can be optimized without the addition of ``ad-hoc" penalties. The paper describes this anaylsis and presents a full derivation of their construction in three prototypical generative... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The paper proposes a new analysis of Wasserstein Autoencoders and claims the new analysis reveals a learning objective form that naturally can be optimized without the addition of ``ad-hoc" penalties. The paper describes this anaylsis and presents a full derivation of their construction in three prototypical ge... |
The paper presents FedEP, a federated learning approach based on expectation propagation in which a global inference task is constructed as local inference tasks. The authors also present several ways to scale the model for modern neural networks. The authors compared their method to baselines on CIFAR-100, StackOverfl... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The paper presents FedEP, a federated learning approach based on expectation propagation in which a global inference task is constructed as local inference tasks. The authors also present several ways to scale the model for modern neural networks. The authors compared their method to baselines on CIFAR-100, Sta... |
The paper asks a very pertinent question as to how we can incorporate knowledge of conserved quantities when solving PDEs via nets. The suggested method is quite unclear.
The main strength of the paper is that it raises and important question and that it has tried some serious benchmark tests. Also interestingly this ... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper asks a very pertinent question as to how we can incorporate knowledge of conserved quantities when solving PDEs via nets. The suggested method is quite unclear.
The main strength of the paper is that it raises and important question and that it has tried some serious benchmark tests. Also interesting... |
The paper proposes an exact and scalable sampling algorithm for Gaussian processes when using Matern correlation functions. The authors make use of the recent kernel packets formalism that enables a sparse representation of the covariance matrix and reduced computational complexity. Numerical results that show the adva... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes an exact and scalable sampling algorithm for Gaussian processes when using Matern correlation functions. The authors make use of the recent kernel packets formalism that enables a sparse representation of the covariance matrix and reduced computational complexity. Numerical results that show ... |
This paper proposes to use spatial-coordinate neural networks as an alternative to the spatial representations used in traditional solvers for time-dependent partial differential equations. This work reformulates the PINN-like PDE solvers in a time-dependent way. It calculates the spatial derivatives as in PINNs at eac... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes to use spatial-coordinate neural networks as an alternative to the spatial representations used in traditional solvers for time-dependent partial differential equations. This work reformulates the PINN-like PDE solvers in a time-dependent way. It calculates the spatial derivatives as in PINN... |
This paper proposes an active learning-based approach for knowledge distillation. It assumes that the teacher model does not have access to a fully labelled dataset either, and there is a cost associated to every call of the teacher model to produce soft labels for the unlabelled data. The aim is to minimize the total ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes an active learning-based approach for knowledge distillation. It assumes that the teacher model does not have access to a fully labelled dataset either, and there is a cost associated to every call of the teacher model to produce soft labels for the unlabelled data. The aim is to minimize th... |
This paper proposes an embedding clustering regularization technique to mitigate the posterior collapse
problem in neural topic models. The regularization technique forces the topic embeddings to be the
centers of separately aggregated clusters of word embeddings. As a result, distinct clusters of topic-
word embedding... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes an embedding clustering regularization technique to mitigate the posterior collapse
problem in neural topic models. The regularization technique forces the topic embeddings to be the
centers of separately aggregated clusters of word embeddings. As a result, distinct clusters of topic-
word e... |
This paper proposes to introduce the set permutation invariance constraint in training self-supervised learning approaches. The paper also investigates the effect of using different aggregation function that encourages the permutation invariance among images. Empirical results demonstrates certain benefits of the propo... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes to introduce the set permutation invariance constraint in training self-supervised learning approaches. The paper also investigates the effect of using different aggregation function that encourages the permutation invariance among images. Empirical results demonstrates certain benefits of t... |
The paper aims to evaluate the few-shot capabilities of a neural network on the worst-case subset of a dataset. The main motivation is to showcase the fragile ability of neural networks to memorize spurious statistical cues in the dataset, leading to poor generalization. To find the worst-case subset, the authors propo... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper aims to evaluate the few-shot capabilities of a neural network on the worst-case subset of a dataset. The main motivation is to showcase the fragile ability of neural networks to memorize spurious statistical cues in the dataset, leading to poor generalization. To find the worst-case subset, the autho... |
The authors conduct a study of structured pruning methods on natural language generation tasks. They demonstrate that the prior art do not improve significantly over the naive random pruning baseline. Then the authors attempt to understand the results through two measures that they introduce: sensitivity and uniqueness... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors conduct a study of structured pruning methods on natural language generation tasks. They demonstrate that the prior art do not improve significantly over the naive random pruning baseline. Then the authors attempt to understand the results through two measures that they introduce: sensitivity and un... |
This paper proposes a new approach to unsupervised meta-learning by drawing inspiration from recent self-supervised learning methods. Specifically, they employ an architecture similar to MoCo, where one branch is updated with momentum updates of the other branch, which is updated ‘online’. To form each task for meta-le... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a new approach to unsupervised meta-learning by drawing inspiration from recent self-supervised learning methods. Specifically, they employ an architecture similar to MoCo, where one branch is updated with momentum updates of the other branch, which is updated ‘online’. To form each task for... |
The paper is about how to learn offline environments that can communicate between agents in multi-agent environments. Each agent learns a function that generates messages from their observations, based on the messages in the datasets. The communication function is learned with a multi-head attention structure, and in t... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper is about how to learn offline environments that can communicate between agents in multi-agent environments. Each agent learns a function that generates messages from their observations, based on the messages in the datasets. The communication function is learned with a multi-head attention structure, ... |
The paper discusses the problem of detecting graphs that have been generated algorithmically. The paper discusses four different types of scenarios from the simplest (where the model knows all the possible details about the generation process) to the hardest (where nothing is known in advance).
The classifier used is ... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper discusses the problem of detecting graphs that have been generated algorithmically. The paper discusses four different types of scenarios from the simplest (where the model knows all the possible details about the generation process) to the hardest (where nothing is known in advance).
The classifier ... |
This paper presented a new self-supervised pre-training method for sequential decision-making tasks. Specifically, a new general pre-training-finetuning pipeline named SMART (Self-supervised multi-task pertaining with control transformer) was proposed with a Control Transformer (CT) and a control-centric pre-training o... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presented a new self-supervised pre-training method for sequential decision-making tasks. Specifically, a new general pre-training-finetuning pipeline named SMART (Self-supervised multi-task pertaining with control transformer) was proposed with a Control Transformer (CT) and a control-centric pre-tr... |
The paper provides sharp generalization bounds for the full-batch Gradient Descent algorithm on smooth losses. It builds upon the stability argument and risk decomposition. It derives a generalization error specifically for nonconvex, convex and strongly convex cases with smoothness assumption.
Strength:
The paper pro... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper provides sharp generalization bounds for the full-batch Gradient Descent algorithm on smooth losses. It builds upon the stability argument and risk decomposition. It derives a generalization error specifically for nonconvex, convex and strongly convex cases with smoothness assumption.
Strength:
The p... |
The paper proposes a framework they call as "MultiViz" aimed at visualizing the internal working of multimodal models. In my opinion, the major contribution from the paper is breaking down multimodal interpretability into 4 components: (1) unimodal contributions, (2) cross-modal interactions, (3) multimodal-representat... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a framework they call as "MultiViz" aimed at visualizing the internal working of multimodal models. In my opinion, the major contribution from the paper is breaking down multimodal interpretability into 4 components: (1) unimodal contributions, (2) cross-modal interactions, (3) multimodal-rep... |
This paper studies a contextual bandit problem with **stationary** context and **fixed** action set. At each step, the environment randomly selects a context $x$, then the player needs to choose an action $a$ that maximizes $f(\mu(x)a)$ where $f(\cdot)$ is a known concave function and $\mu$ is an unknown matrix-valued ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies a contextual bandit problem with **stationary** context and **fixed** action set. At each step, the environment randomly selects a context $x$, then the player needs to choose an action $a$ that maximizes $f(\mu(x)a)$ where $f(\cdot)$ is a known concave function and $\mu$ is an unknown matrix... |
This paper presents an empirical assessment of variations in behaviour of different representation learning paradigms when the underlying training sample is corrupted with noise. Authors focused on image classification tasks, and verified gap in performances post corruptions. They further evaluated against which kinds ... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper presents an empirical assessment of variations in behaviour of different representation learning paradigms when the underlying training sample is corrupted with noise. Authors focused on image classification tasks, and verified gap in performances post corruptions. They further evaluated against whic... |
The paper proposes a predictor-based neural architecture search with N-sized mixed batch, in which the performances of k architectures are evaluated by training from scratch and the performances of the rest N-k ones are predicted by the predictor. Besides, it studies the impact of predictors on NAS theoretically and em... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a predictor-based neural architecture search with N-sized mixed batch, in which the performances of k architectures are evaluated by training from scratch and the performances of the rest N-k ones are predicted by the predictor. Besides, it studies the impact of predictors on NAS theoreticall... |
This paper proposes a new clustering-based non-parameteric way of estimating the state visitation density, and uses it to construct a reward to encourage more exploration of less-visited states. The paper also proposes a way of converting the state input into some embedding space. Overall, both ideas seem to have some ... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a new clustering-based non-parameteric way of estimating the state visitation density, and uses it to construct a reward to encourage more exploration of less-visited states. The paper also proposes a way of converting the state input into some embedding space. Overall, both ideas seem to ha... |
This paper studies the problem of robust constrained reinforcement learning (RL) where the underlying model has uncertainty. The goal is to ensure that the constraints are satisfied for the worst-case MDP in a set of MDPs, while at the same time to maximize the reward over the uncertainty set. A robust primal-dual algo... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the problem of robust constrained reinforcement learning (RL) where the underlying model has uncertainty. The goal is to ensure that the constraints are satisfied for the worst-case MDP in a set of MDPs, while at the same time to maximize the reward over the uncertainty set. A robust primal-d... |
This paper leverages double descent for scientific data analysis, with face-based social behavior as a case study.
The organization is clear and easy to follow.
My concerns regarding this paper are as below.
1. Limited novelty. The main contributions of this paper are not enough for ICLR.
2. Some related works are mis... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper leverages double descent for scientific data analysis, with face-based social behavior as a case study.
The organization is clear and easy to follow.
My concerns regarding this paper are as below.
1. Limited novelty. The main contributions of this paper are not enough for ICLR.
2. Some related works... |
The authors present a model for adaptive super resolution on mobile devices. Their approach employs a one-shot neural architecture search to generate alternatives sharing weights. For inference, they introduce an incremental adaptation method. Their model aims to mantain a steady frame rate with a low memory footprint ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors present a model for adaptive super resolution on mobile devices. Their approach employs a one-shot neural architecture search to generate alternatives sharing weights. For inference, they introduce an incremental adaptation method. Their model aims to mantain a steady frame rate with a low memory fo... |
The paper proposes a human-in-the-loop framework that first tries to find most distinguishable samples for each algorithm, and then let human summarize observable patterns from those samples.
Strength: The paper provides several applications of this method, which seems interesting.
Weakness: It is unclear how exactly... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes a human-in-the-loop framework that first tries to find most distinguishable samples for each algorithm, and then let human summarize observable patterns from those samples.
Strength: The paper provides several applications of this method, which seems interesting.
Weakness: It is unclear how... |
This work defines predictive heterogeneity to quantify the heterogeneity in the data which influences the predictive performance. It proposes an algorithm which quantifies the mentioned heterogeneity and use that in examples to show how it can be helpful in understanding the subtleties of the data.
The paper seems inte... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This work defines predictive heterogeneity to quantify the heterogeneity in the data which influences the predictive performance. It proposes an algorithm which quantifies the mentioned heterogeneity and use that in examples to show how it can be helpful in understanding the subtleties of the data.
The paper se... |
The paper proposes a GMoE model for solving classification under domain shifts, specifically in the DG setting. The paper presents results on 8 benchmarks and shows promising performance.
Strengths
- I liked the idea of applying mixture of experts for solving the problem of DG. I resonate with the authors, that it is... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a GMoE model for solving classification under domain shifts, specifically in the DG setting. The paper presents results on 8 benchmarks and shows promising performance.
Strengths
- I liked the idea of applying mixture of experts for solving the problem of DG. I resonate with the authors, th... |
Downloading pre-trained backbones from third-party platforms and deploying them in various downstream tasks is now standard practice. This poses security risks such as backdoor attacks. The authors investigate an interesting question: Do we need labels for backdoor defense? To that end, the authors propose using unsupe... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Downloading pre-trained backbones from third-party platforms and deploying them in various downstream tasks is now standard practice. This poses security risks such as backdoor attacks. The authors investigate an interesting question: Do we need labels for backdoor defense? To that end, the authors propose usin... |
Two interesting machine learning models that are simultaneously of theoretical and practical interest are Deep Equilibrium Models (DEQs) and well-behaved infinite-width neural networks trained using gradient flow (Neural tangent kernel, NTK). The first is theoretically interesting, because it can *sometimes* be used to... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
Two interesting machine learning models that are simultaneously of theoretical and practical interest are Deep Equilibrium Models (DEQs) and well-behaved infinite-width neural networks trained using gradient flow (Neural tangent kernel, NTK). The first is theoretically interesting, because it can *sometimes* be... |
The paper suggests a state space model to handle both high and low dimensional observation spaces with highly nonlinear observation model, such as animated images of physical systems. The internal state representation is probabilistic, taking account uncertainties. The model is a nonlinear extension of the Hammersein-W... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper suggests a state space model to handle both high and low dimensional observation spaces with highly nonlinear observation model, such as animated images of physical systems. The internal state representation is probabilistic, taking account uncertainties. The model is a nonlinear extension of the Hamm... |
The paper proposes to structure the latent space of a deep neural network following the Bohr model of atoms. In this latent space, each datapoint is to be embedded as an 'atom', i.e. a set of particles represented by their position and their charge.
Using these values, the authors can then compute losses to make sure t... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes to structure the latent space of a deep neural network following the Bohr model of atoms. In this latent space, each datapoint is to be embedded as an 'atom', i.e. a set of particles represented by their position and their charge.
Using these values, the authors can then compute losses to mak... |
In this work, authors show that the batch size affects the inference results of deep neural network models.
In the empirical study, authors studied bidirectional encoder representations from transformers (BERT) and generative pre-trained transformer (GPT) natural language processing (NLP) models, and the super-resolu... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
In this work, authors show that the batch size affects the inference results of deep neural network models.
In the empirical study, authors studied bidirectional encoder representations from transformers (BERT) and generative pre-trained transformer (GPT) natural language processing (NLP) models, and the supe... |
The paper proposes to perform data bootstrapping by using a performant language model generate prompts that specify puzzles and finetune models such data (after filtering for correctness).
Pros:
This method has some potential to augment data in order to improve model's general code understanding, generation and reason... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes to perform data bootstrapping by using a performant language model generate prompts that specify puzzles and finetune models such data (after filtering for correctness).
Pros:
This method has some potential to augment data in order to improve model's general code understanding, generation an... |
Existing graph transformers have succeeded in different applications but require cumbersome architecture designs and tuning by experienced engineers. This paper proposed a novel search framework for graph transformers, automatically searching within the designed unified search space for optimal architecture designs. Th... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
Existing graph transformers have succeeded in different applications but require cumbersome architecture designs and tuning by experienced engineers. This paper proposed a novel search framework for graph transformers, automatically searching within the designed unified search space for optimal architecture des... |
This paper provides another viewpoint of investigating the robustness of DNNs. The motivation is that when a model is training, the decision boundary w.r.t. the model is dynamically changing over epoch time, so that the distance of training data points and the decision boundary varies. When the distance gets closer, th... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper provides another viewpoint of investigating the robustness of DNNs. The motivation is that when a model is training, the decision boundary w.r.t. the model is dynamically changing over epoch time, so that the distance of training data points and the decision boundary varies. When the distance gets cl... |
Authors proposed BFReg-NN as a general deep learning model and designed its architecture based on the regulatory relations and hierarchical relations among genes, proteins and pathways. Incorporating the biological knowledge into the network architecture design, authors tried to break the “black-box” nature of neural n... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
Authors proposed BFReg-NN as a general deep learning model and designed its architecture based on the regulatory relations and hierarchical relations among genes, proteins and pathways. Incorporating the biological knowledge into the network architecture design, authors tried to break the “black-box” nature of ... |
This paper presents a novel framework that utilizes the annotation disagreements in human-annotated benchmarks and human-annotated preference to improve the model training in scenarios where annotation records are provided.
Strength:
- The idea is simple yet effective.
- The paper is well-motivated and well-writte... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a novel framework that utilizes the annotation disagreements in human-annotated benchmarks and human-annotated preference to improve the model training in scenarios where annotation records are provided.
Strength:
- The idea is simple yet effective.
- The paper is well-motivated and wel... |
In this paper, the authors formalize the relationship between perfect security and couplings of distributions, and propose the first instance of a steganography method that has the benefits of non-trivial efficiency and perfect security guarantees. Experiments with GPT-2 and WaveRNN demonstrate the effectiveness of the... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the authors formalize the relationship between perfect security and couplings of distributions, and propose the first instance of a steganography method that has the benefits of non-trivial efficiency and perfect security guarantees. Experiments with GPT-2 and WaveRNN demonstrate the effectivenes... |
The paper addresses the extension of existing fairness notions in ML such as Demographic Parity, Equal Opportunity, and Calibration to settings where the both the target and sensitive attributes are potentially multivariate and continuous.
It proposes a metric (FairCOCCO) based on the cross covariance operator over r... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper addresses the extension of existing fairness notions in ML such as Demographic Parity, Equal Opportunity, and Calibration to settings where the both the target and sensitive attributes are potentially multivariate and continuous.
It proposes a metric (FairCOCCO) based on the cross covariance operato... |
The paper mainly explores the issue of “overgeneralization” in unsupervised anomaly detection and proposed a bio-inspired solution, termed “Random Forgetting Twin Memory (RFTM).” The new model did not change the fundamental structure of existing learning models, e.g., autoencoder, and work in a plug-and-play fashion. I... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper mainly explores the issue of “overgeneralization” in unsupervised anomaly detection and proposed a bio-inspired solution, termed “Random Forgetting Twin Memory (RFTM).” The new model did not change the fundamental structure of existing learning models, e.g., autoencoder, and work in a plug-and-play fa... |
This paper proposes a modified version of MCTS that encourages exploration and incorporates uncertainty into the planning tree. The authors compare the performance of the proposed method against MuZero in two environments: Slide and Mountain Car.
Strength:
1. This paper is well-written and easy to follow.
2. This propo... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a modified version of MCTS that encourages exploration and incorporates uncertainty into the planning tree. The authors compare the performance of the proposed method against MuZero in two environments: Slide and Mountain Car.
Strength:
1. This paper is well-written and easy to follow.
2. Th... |
This work proposes Pseudoinverse-Guided Diffusion Models, which uses problem-agnostic diffusion models to reach the empirical performance of problem-specific ones. The proposed approach directly estimates conditional scores without additional training. In particular, it can address inverse problems with noisy and non-l... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This work proposes Pseudoinverse-Guided Diffusion Models, which uses problem-agnostic diffusion models to reach the empirical performance of problem-specific ones. The proposed approach directly estimates conditional scores without additional training. In particular, it can address inverse problems with noisy a... |
The paper proposes a joint learning on node classification and attribute augmentation on heterogeneous graphs. The attribute augmentation is based on the aggregations among one-hop neighbors weighted by attentions. Experiments show the proposed approach results in marginal improvements over the baseline methods.
Streng... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes a joint learning on node classification and attribute augmentation on heterogeneous graphs. The attribute augmentation is based on the aggregations among one-hop neighbors weighted by attentions. Experiments show the proposed approach results in marginal improvements over the baseline methods... |
The paper proposes a novel idea of generating images in bits level. The proposed method is shown to be effective on selected dataset. The selected datasets are not simple and includes complex dataset such as COCO(in contrast, some of the works only report results on CIFAR and ImageNet, which are simpler than COCO). The... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper proposes a novel idea of generating images in bits level. The proposed method is shown to be effective on selected dataset. The selected datasets are not simple and includes complex dataset such as COCO(in contrast, some of the works only report results on CIFAR and ImageNet, which are simpler than CO... |
This paper studies the languages that emerge between two agents that solve a task collaboratively. The main focus of the paper is on studying whether Harris’ articulation scheme (HAS) holds in the emergent languages. To test for that, the authors resort to an unsupervised segmentation algorithm derived from HAS. At it... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper studies the languages that emerge between two agents that solve a task collaboratively. The main focus of the paper is on studying whether Harris’ articulation scheme (HAS) holds in the emergent languages. To test for that, the authors resort to an unsupervised segmentation algorithm derived from HA... |
This paper proposes regularization terms to promote the contractivity of neural ODEs. It derives a computational efficient regularizer for a special class of neural ODEs and showed empirical robustness comparing with plain neural ODEs.
Strength:
1) Leveraging control tools to enforce contraction properties on neural OD... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proposes regularization terms to promote the contractivity of neural ODEs. It derives a computational efficient regularizer for a special class of neural ODEs and showed empirical robustness comparing with plain neural ODEs.
Strength:
1) Leveraging control tools to enforce contraction properties on n... |
This paper proposes a multi-domain long-tailed recognition method to simultaneously address the two types of distribution shift: subpopulation shift and domain shift. The basic idea is sample augmentation where new samples are generated by representation disentanglement for tailed class. The experimental results on syn... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a multi-domain long-tailed recognition method to simultaneously address the two types of distribution shift: subpopulation shift and domain shift. The basic idea is sample augmentation where new samples are generated by representation disentanglement for tailed class. The experimental result... |
This paper introduce a simple mechanism to allow the training of MIM model tacking the reconstruction difficulty into account. It also introduce deep supervision that can help the diversify the feature representation across different layers. Good performance are reported to justfiy its methods.
Strength:
- The intuiti... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper introduce a simple mechanism to allow the training of MIM model tacking the reconstruction difficulty into account. It also introduce deep supervision that can help the diversify the feature representation across different layers. Good performance are reported to justfiy its methods.
Strength:
- The... |
This paper presents an interesting approach to improving DETR-based models across multiple tasks. The proposed approach mainly consists of the following three steps:
- learn image-adaptive dynamic coefficients for each of the original grouped learned queries,
- perform convex combination, i.e., aggregate, each group o... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents an interesting approach to improving DETR-based models across multiple tasks. The proposed approach mainly consists of the following three steps:
- learn image-adaptive dynamic coefficients for each of the original grouped learned queries,
- perform convex combination, i.e., aggregate, each... |
This paper proposes to guide the formal theorem prover using informal proofs. The key idea is to convert the informal proofs to formal proof sketches by prompting LLM as a few-short learner. On the miniF2F-test dataset, this approach improves the performance of the Sledgehammer + heuristics prover of Isabelle from 20.9... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes to guide the formal theorem prover using informal proofs. The key idea is to convert the informal proofs to formal proof sketches by prompting LLM as a few-short learner. On the miniF2F-test dataset, this approach improves the performance of the Sledgehammer + heuristics prover of Isabelle f... |
This work proposed a model-agnostic framework, Recursion of Thought (RoT), to release the capacity constraint by the maximum size of a single context in language models. RoT teaches a language model to divide and conquer complex problems by recursively creating multiple contexts; therefore, a complex problem could be s... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This work proposed a model-agnostic framework, Recursion of Thought (RoT), to release the capacity constraint by the maximum size of a single context in language models. RoT teaches a language model to divide and conquer complex problems by recursively creating multiple contexts; therefore, a complex problem co... |
This paper provides a method to predict air quality by combining spatial-temporal data and graph neural networks. The contributions of this paper are mainly on integrating HYSPLIT for modeling the dynamic relationships between nodes.
Strength:
1. The topic of air quality prediction is of great social impact.
Weaknesse... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper provides a method to predict air quality by combining spatial-temporal data and graph neural networks. The contributions of this paper are mainly on integrating HYSPLIT for modeling the dynamic relationships between nodes.
Strength:
1. The topic of air quality prediction is of great social impact.
W... |
This paper proposes a new geometry-based knowledge graph embedding method named ExpressivE, which represents entities as points and relations as hyper-parallelograms. Extensive theoretical analyses show that ExpressivE is fully expressive and is capable of capturing various inference patterns. Experiments demonstrate t... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a new geometry-based knowledge graph embedding method named ExpressivE, which represents entities as points and relations as hyper-parallelograms. Extensive theoretical analyses show that ExpressivE is fully expressive and is capable of capturing various inference patterns. Experiments demon... |
This work proposes an extension of model adaptation via meta learning in an online fashion. This paper particularly targets the problem model adaptation in the environment with changing tasks and input distributions. With the ability for fast adaptation, the proposed method called as Fully-Online Meta-Learning (FOML) a... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposes an extension of model adaptation via meta learning in an online fashion. This paper particularly targets the problem model adaptation in the environment with changing tasks and input distributions. With the ability for fast adaptation, the proposed method called as Fully-Online Meta-Learning ... |
Authors propose the “IDEAL” algorithm that can improve the distillation process when there is no data to be used for the distillation process, and also when the teacher provides only the hard-labels (no soft predictions over the labels). IDEAL employs a generator to generate examples on-the-fly in this data-free scenar... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
Authors propose the “IDEAL” algorithm that can improve the distillation process when there is no data to be used for the distillation process, and also when the teacher provides only the hard-labels (no soft predictions over the labels). IDEAL employs a generator to generate examples on-the-fly in this data-fre... |
This paper introduces a method to generate samples from a target domain containing very few shots using a fixed pretrained GAN.
Strengths:
- concise and straightforward method for performing the task at hand
- generally well-written and easy to follow
- compelling results for this very-few-shot domain adaptation gene... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper introduces a method to generate samples from a target domain containing very few shots using a fixed pretrained GAN.
Strengths:
- concise and straightforward method for performing the task at hand
- generally well-written and easy to follow
- compelling results for this very-few-shot domain adaptat... |
The paper proposes an algorithm to learn instance-specific augmentation for images, which substantially improves computer vision models over a wide range of scenarios.
Strength:
1. The algorithm is simple and effective.
2. It applies to a wide range of scenarios, including supervised learning and contrastive learning.
... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes an algorithm to learn instance-specific augmentation for images, which substantially improves computer vision models over a wide range of scenarios.
Strength:
1. The algorithm is simple and effective.
2. It applies to a wide range of scenarios, including supervised learning and contrastive le... |
The paper proposes an algorithm for coreset selection with low computation time. The key idea is to approximate a function using RBFNN on a large dataset and then construct coresets for radial basis and Laplicain loss function. Empirically, the authors show that the proposed method can find a small coreset with competi... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes an algorithm for coreset selection with low computation time. The key idea is to approximate a function using RBFNN on a large dataset and then construct coresets for radial basis and Laplicain loss function. Empirically, the authors show that the proposed method can find a small coreset with... |
This paper studies the problem of data heterogeneity in federated learning. It suggests mathematical analysis based on the “worst-case margin” theory to measure the generalization contribution of clients’ FedAvg updates. Based on theoretical results authors propose decoupling the neural network into 2 parts: feature ex... | 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 studies the problem of data heterogeneity in federated learning. It suggests mathematical analysis based on the “worst-case margin” theory to measure the generalization contribution of clients’ FedAvg updates. Based on theoretical results authors propose decoupling the neural network into 2 parts: fe... |
Automated feature engineering is another needed ingredient in 'data science as a service.' Prior work focuses either on feature selection or feature engineering that is independent on the underlying data. The authors propose an RL algorithm that directly transforms data by applying a sequence of unary or binary operati... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
Automated feature engineering is another needed ingredient in 'data science as a service.' Prior work focuses either on feature selection or feature engineering that is independent on the underlying data. The authors propose an RL algorithm that directly transforms data by applying a sequence of unary or binary... |
This work attempts to address the curse of high dimensionality problem for the tabular data. The authors build their model, PLATO, on the well formed insight that auxiliary KG about the feature connections can improve performance as the sparsity induced in the model helps with the generalizations in the cases where d>>... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
This work attempts to address the curse of high dimensionality problem for the tabular data. The authors build their model, PLATO, on the well formed insight that auxiliary KG about the feature connections can improve performance as the sparsity induced in the model helps with the generalizations in the cases w... |
This paper finds the main reason behind performance collapse in DARTS. A new algorithm named lambda-DARTS is proposed, which is able to generate better performance compared with other methods. Specifically, two new regularization terms are proposed to prevent performance collapse by harmonizing operation selection via... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper finds the main reason behind performance collapse in DARTS. A new algorithm named lambda-DARTS is proposed, which is able to generate better performance compared with other methods. Specifically, two new regularization terms are proposed to prevent performance collapse by harmonizing operation selec... |
This paper concerns causal effect estimation when the confounders and treatment are both derived from texts. One issue herein is that when adjusting for the entire text, there is a violation of the overlap assumption which is required for drawing valid causal conclusions. The proposed solution is to use supervised repr... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper concerns causal effect estimation when the confounders and treatment are both derived from texts. One issue herein is that when adjusting for the entire text, there is a violation of the overlap assumption which is required for drawing valid causal conclusions. The proposed solution is to use supervi... |
This paper presents an analysis of existing self-supervised learning methods with vision transformers. Extensive experiments are performed on the off-the-shelf pretrained model with the linear and attentive probes to verify the speculation that contrastive learning is a part-to-whole task and masked image modeling is a... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper presents an analysis of existing self-supervised learning methods with vision transformers. Extensive experiments are performed on the off-the-shelf pretrained model with the linear and attentive probes to verify the speculation that contrastive learning is a part-to-whole task and masked image model... |
The authors present a theoretical analysis of the computational benefits of interneuron-mediated recurrent connections by characterizing the synaptic change dynamics while learning to whiten an input. Specifically, the authors provide analytical solutions to the dynamics of a recurrent neural network without any pointw... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors present a theoretical analysis of the computational benefits of interneuron-mediated recurrent connections by characterizing the synaptic change dynamics while learning to whiten an input. Specifically, the authors provide analytical solutions to the dynamics of a recurrent neural network without an... |
This paper presents a new architecture to satisfy the mini-batch consistency (MBC) property, which is an important property required by set functions in the streaming fashion. Specifically, given a mini-batch consistent function $f$ (e.g., slot set encoder, SSE), the authors prove that for an arbitrary set function $f^... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a new architecture to satisfy the mini-batch consistency (MBC) property, which is an important property required by set functions in the streaming fashion. Specifically, given a mini-batch consistent function $f$ (e.g., slot set encoder, SSE), the authors prove that for an arbitrary set func... |
The authors consider two problems: 1) relating convex surrogate losses with zero-ones losses in adversarial training; 2) bounding the (adversarial) convex risk of shallow ReLU networks.
For the former, the authors' results apply to a very broad setting -- arbitrary perturbation sets and general data distributions. Th... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors consider two problems: 1) relating convex surrogate losses with zero-ones losses in adversarial training; 2) bounding the (adversarial) convex risk of shallow ReLU networks.
For the former, the authors' results apply to a very broad setting -- arbitrary perturbation sets and general data distribut... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.