review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper studies gradient descent for linear regression under the high dimensional framework. Under certain conditions, the authors study the generalization error in such gradient descent, and show that there exists multiple descent in the generalization error of the linear model during training. The authors also con... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies gradient descent for linear regression under the high dimensional framework. Under certain conditions, the authors study the generalization error in such gradient descent, and show that there exists multiple descent in the generalization error of the linear model during training. The authors ... |
The paper presents an interesting training algorithm for training SNNs from scratch with only one timestep.
Very comprehensive results
Simple yet effective idea
Since the authors use a regularization technique, I am wondering if the authors can shed light on how their method differs from previous temporal BN methods ... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper presents an interesting training algorithm for training SNNs from scratch with only one timestep.
Very comprehensive results
Simple yet effective idea
Since the authors use a regularization technique, I am wondering if the authors can shed light on how their method differs from previous temporal BN ... |
The authors propose a new visual-text learning algorithm for document understanding. In particular, the authors argue that the advantage of proposed algorithm is free of using OCR in the prediction.
* The major question I have is the motivation of the work. It is not clear to me why we would like to discard OCR as it... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a new visual-text learning algorithm for document understanding. In particular, the authors argue that the advantage of proposed algorithm is free of using OCR in the prediction.
* The major question I have is the motivation of the work. It is not clear to me why we would like to discard O... |
This paper introduces a dynamic margin selection (DynaMS) method to dynamically construct the training subset by utilizing the distance from candidate samples to the classification boundary. In addition, a light parameter sharing proxy is designed to reduce the additional computation incurred by the selection. Extensiv... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper introduces a dynamic margin selection (DynaMS) method to dynamically construct the training subset by utilizing the distance from candidate samples to the classification boundary. In addition, a light parameter sharing proxy is designed to reduce the additional computation incurred by the selection. ... |
The paper proposes a mechanism for scaling up the receptive field of the convolutional kernels in CNNs, and for improving the performance without extra FLOPS via sparsity. Positive results are shown for several visual tasks.
The paper proposes a simple way for scaling up the size of the convolutional kernels, and empir... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a mechanism for scaling up the receptive field of the convolutional kernels in CNNs, and for improving the performance without extra FLOPS via sparsity. Positive results are shown for several visual tasks.
The paper proposes a simple way for scaling up the size of the convolutional kernels, a... |
This paper proposes a method to detect adversarial samples in tree ensembles without affecting either the model's structure or its original performance. Since the existing adversarial defense method may affect the model's natural performance, this paper's method enables the users to decided whether to apply defense or ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method to detect adversarial samples in tree ensembles without affecting either the model's structure or its original performance. Since the existing adversarial defense method may affect the model's natural performance, this paper's method enables the users to decided whether to apply def... |
This paper first studies how presence of different kinds of noise affect the discrete speech representations from self-supervised models used for spoken LM. The authors utilize the Levenstien distance between sequences of discrete token IDs obtained from HuBERT as an indicator for the presence of noise. Then, a pseudo-... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper first studies how presence of different kinds of noise affect the discrete speech representations from self-supervised models used for spoken LM. The authors utilize the Levenstien distance between sequences of discrete token IDs obtained from HuBERT as an indicator for the presence of noise. Then, a... |
This paper aims to reduce contextual bias in multi-label classification. To this end, the paper aims to remove the effect of unseen context $C$ from the prediction. Specifically, given a casual structure {$C \to X, C \to Y, X \to Y$}, the usual classifier predicts $P(Y|X)$ which implicitly reflects the unseen $C$. Inst... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper aims to reduce contextual bias in multi-label classification. To this end, the paper aims to remove the effect of unseen context $C$ from the prediction. Specifically, given a casual structure {$C \to X, C \to Y, X \to Y$}, the usual classifier predicts $P(Y|X)$ which implicitly reflects the unseen $... |
The paper makes a simple modification to the parameterized posterior distribution to allow for learning group-representations and the within-group instance representation when they are dependent conditional on the observed features: condition the instance representations on both the features and the group representatio... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
The paper makes a simple modification to the parameterized posterior distribution to allow for learning group-representations and the within-group instance representation when they are dependent conditional on the observed features: condition the instance representations on both the features and the group repre... |
This paper proposed a new multigraph topology for cross-silo federated learning in a decentralized network. Specifically, the proposed method extends the overlay graph with weak connections based on the delay time and improved DPASGD algorithm to process isolated nodes for time reduction.
Strength.
1. Using multigraph... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposed a new multigraph topology for cross-silo federated learning in a decentralized network. Specifically, the proposed method extends the overlay graph with weak connections based on the delay time and improved DPASGD algorithm to process isolated nodes for time reduction.
Strength.
1. Using mu... |
**Note: Score updated from 5 to 8 after author response**
The authors develop methods to train deep transformers that lack skip connections and/or normalization layers. They achieve this through the following theoretically-motivated interventions: modified initialization, bias matrices, and location-dependent scaling.... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
**Note: Score updated from 5 to 8 after author response**
The authors develop methods to train deep transformers that lack skip connections and/or normalization layers. They achieve this through the following theoretically-motivated interventions: modified initialization, bias matrices, and location-dependent ... |
The paper proposes a method to obtain predictors that are counterfactually-invariant to certain observed variables. Authors first show that counterfactual invariance of a predictor with respect to the variables A can be ensured given a set of variables Z that block all paths from A to Y. That is, a counterfactual invar... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper proposes a method to obtain predictors that are counterfactually-invariant to certain observed variables. Authors first show that counterfactual invariance of a predictor with respect to the variables A can be ensured given a set of variables Z that block all paths from A to Y. That is, a counterfactu... |
This work presents a multi-interest retrieval model, the objective of which is to increase the performance of the retrieval stage in a standard two-stage recommendation system. Extensive experiments have been done on various large-scale datasets to show the effectiveness of the proposed approach.
Strength
+ The introd... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work presents a multi-interest retrieval model, the objective of which is to increase the performance of the retrieval stage in a standard two-stage recommendation system. Extensive experiments have been done on various large-scale datasets to show the effectiveness of the proposed approach.
Strength
+ Th... |
The paper addresses the problem of client drift in Federated Learning, a problem well-known in the literature to diminish the efficacy of FL methods. The authors propose to use a decoupling of features into low and high-level ones, they show that by grouping and sharing the low-level features among clients, gradient di... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper addresses the problem of client drift in Federated Learning, a problem well-known in the literature to diminish the efficacy of FL methods. The authors propose to use a decoupling of features into low and high-level ones, they show that by grouping and sharing the low-level features among clients, gra... |
The paper describes a technique to create a robot control agent that can control a manipulator to execute a variety of tasks specified in the form of a multi-model prompt, combining text and visual tokens.
The paper also introduces a large scale benchmark for problems focusing on the properties of the system VIMA-Ben... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper describes a technique to create a robot control agent that can control a manipulator to execute a variety of tasks specified in the form of a multi-model prompt, combining text and visual tokens.
The paper also introduces a large scale benchmark for problems focusing on the properties of the system ... |
This paper develops a robust MARL framework that considers state uncertainty. The paper first formulates the MARL problem with state uncertainty as MG-SPA and develops theoretical contributions, including the new solution concept of Robust equilibrium and the equilibrium's existence. Then, based on the Bellman equation... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper develops a robust MARL framework that considers state uncertainty. The paper first formulates the MARL problem with state uncertainty as MG-SPA and develops theoretical contributions, including the new solution concept of Robust equilibrium and the equilibrium's existence. Then, based on the Bellman ... |
This paper studies the convergence of gradient descent on the matrix sensing problem in both the overparametrized and underparametrized regimes, with the latter one being completely new compared to existing studies. The (matrix) RIP property is assumed so that the problem is well-conditioned from the statistical perspe... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the convergence of gradient descent on the matrix sensing problem in both the overparametrized and underparametrized regimes, with the latter one being completely new compared to existing studies. The (matrix) RIP property is assumed so that the problem is well-conditioned from the statistica... |
This paper proposes a function approximation framework for mapping and exploration in RL. The function approximation scheme uses multiple local models that predict the next observation, and new local models get created when a prediction error is above a user-defined threshold. The local model used at each step is deter... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper proposes a function approximation framework for mapping and exploration in RL. The function approximation scheme uses multiple local models that predict the next observation, and new local models get created when a prediction error is above a user-defined threshold. The local model used at each step ... |
This paper proposed a new framework, named ReaKE for learning representations for chemicals to better predict chemical reactions
The main contributions of this work are three folds as claimed by the authors: (1) chemical synthesis KG of reactants, products and; (2) contrastive learning strategies in KG triple level and... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposed a new framework, named ReaKE for learning representations for chemicals to better predict chemical reactions
The main contributions of this work are three folds as claimed by the authors: (1) chemical synthesis KG of reactants, products and; (2) contrastive learning strategies in KG triple l... |
The paper addresses the problem of sample efficiency in RL by proposing a new episodic control approach. Episodic control consists of maintaining a table-like structure that can easily be accessed to return values associated with states/state-action pairs. The paper proposes NECSA, which introduces several novelties wi... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper addresses the problem of sample efficiency in RL by proposing a new episodic control approach. Episodic control consists of maintaining a table-like structure that can easily be accessed to return values associated with states/state-action pairs. The paper proposes NECSA, which introduces several nove... |
This paper proposes a method to learn an embedding space, which leverages the human-designed taxonomy, for robotics tasks. It extends Gaussian Process Latent Variable Models (GPLVM) to hyperbolic latent spaces and proposes GPHLVM. The motivation for using hyperbolic space is that distances grow exponentially when movin... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a method to learn an embedding space, which leverages the human-designed taxonomy, for robotics tasks. It extends Gaussian Process Latent Variable Models (GPLVM) to hyperbolic latent spaces and proposes GPHLVM. The motivation for using hyperbolic space is that distances grow exponentially wh... |
DivDis tackles the problem of underspecification by generating diverse hypotheses and then disambiguating to choose the best one. Diversification is performed by training multiple heads to yield different predictions on an unlabeled target distribution. Disambiguation is performed by first actively querying to obtain l... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
DivDis tackles the problem of underspecification by generating diverse hypotheses and then disambiguating to choose the best one. Diversification is performed by training multiple heads to yield different predictions on an unlabeled target distribution. Disambiguation is performed by first actively querying to ... |
Authors present a sequential communication framework to address the relative overgeneralization problem in multi-agent reinforcement learning and test it against a number of communication-free and communication-based baselines. Performance figures drawn against the number of training steps show higher average rewards i... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Authors present a sequential communication framework to address the relative overgeneralization problem in multi-agent reinforcement learning and test it against a number of communication-free and communication-based baselines. Performance figures drawn against the number of training steps show higher average r... |
This paper combines a mixture of experts approach with NeRFs for large-scale neural scene rendering. The key idea is to learn a gating network which uses the positional embedding to choose which NeRF model will be queried for the density and color of a given point. All NeRFs share the same head. The gating network need... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper combines a mixture of experts approach with NeRFs for large-scale neural scene rendering. The key idea is to learn a gating network which uses the positional embedding to choose which NeRF model will be queried for the density and color of a given point. All NeRFs share the same head. The gating netw... |
This paper investigates the landscape causes of collapse in self-supervised learning.
Strength
1. Existng work have conflict opinion about the collapse in SSL. This work try to fill this gap by studying the geometry of the SSL.
2. This work shows that the interplay between data variation and data augmentation determ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates the landscape causes of collapse in self-supervised learning.
Strength
1. Existng work have conflict opinion about the collapse in SSL. This work try to fill this gap by studying the geometry of the SSL.
2. This work shows that the interplay between data variation and data augmentatio... |
This paper investigated the characteristics of Mixture of Experts (MoE) models when their expert layers are quantized. The authors revealed that the expert layers have more evenly distributed data than dense layers, and thus they are more robust to quantization. Motivated by this observation, the authors applied low-bi... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper investigated the characteristics of Mixture of Experts (MoE) models when their expert layers are quantized. The authors revealed that the expert layers have more evenly distributed data than dense layers, and thus they are more robust to quantization. Motivated by this observation, the authors applie... |
The authors consider the large-scale implementation of 'neural kernels', kernel
methods that are derived from the architecture of a specific neural network and
the correspondence between randomly-initialized neural networks and gaussian
processes. In contrast to existing large-scale kernel methods for use with
kernels ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The authors consider the large-scale implementation of 'neural kernels', kernel
methods that are derived from the architecture of a specific neural network and
the correspondence between randomly-initialized neural networks and gaussian
processes. In contrast to existing large-scale kernel methods for use with
... |
The paper presents an investigation of how to network reinforcement learning models to accomplish symbolic innovation tasks. The paper analyzes several different, common network topologies as a means of sharing experiences between DQN learners as well as other existing methods that either share experiences or gradients... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper presents an investigation of how to network reinforcement learning models to accomplish symbolic innovation tasks. The paper analyzes several different, common network topologies as a means of sharing experiences between DQN learners as well as other existing methods that either share experiences or g... |
This paper examines the federated learning problem when updates are made asynchronously and communication is compressed. The authors present a new variant of the classic federated averaging (FedAvg) algorithm, namely QuAFL, that supports both asynchronous communication with compression. In some parameter regimes, they ... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper examines the federated learning problem when updates are made asynchronously and communication is compressed. The authors present a new variant of the classic federated averaging (FedAvg) algorithm, namely QuAFL, that supports both asynchronous communication with compression. In some parameter regime... |
The authors address the problem of network pruning under a fixed training budget. Earlier work on network pruning relied on computationally expensive iterative training & pruning regimes. The authors show that when operating under a fixed training regime, a lot of savings can be found by eschewing complex learning ra... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors address the problem of network pruning under a fixed training budget. Earlier work on network pruning relied on computationally expensive iterative training & pruning regimes. The authors show that when operating under a fixed training regime, a lot of savings can be found by eschewing complex lea... |
This paper proves the Strong Lottery Ticket Hypothesis (SLTH) in the context of equivariant neural networks. The SLTH states that for every network f, and every sufficiently overparametrized, randomly-initialized network h, there is a pruning of h that recovers f.
Prior work: In the context of fully-connected networks... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proves the Strong Lottery Ticket Hypothesis (SLTH) in the context of equivariant neural networks. The SLTH states that for every network f, and every sufficiently overparametrized, randomly-initialized network h, there is a pruning of h that recovers f.
Prior work: In the context of fully-connected ... |
The paper discusses the embedding part of learning-based routing problems. That is, some existing learning-based solvers tried to embed problem instances by coordinates of points (to be visited by TSP for example), but the paper is investigating whether or not the coordinate-based embedding is poisonous for DL models.
... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper discusses the embedding part of learning-based routing problems. That is, some existing learning-based solvers tried to embed problem instances by coordinates of points (to be visited by TSP for example), but the paper is investigating whether or not the coordinate-based embedding is poisonous for DL ... |
This paper proposed a self-supervised learning method to train a slimmable network. The self-supervised pre-trained model can be used to fit different computing resource for downstream tasks. The basic idea is to sample sub-networks following slimmable network and train these networks using a self-supervised learning m... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposed a self-supervised learning method to train a slimmable network. The self-supervised pre-trained model can be used to fit different computing resource for downstream tasks. The basic idea is to sample sub-networks following slimmable network and train these networks using a self-supervised le... |
This paper proposed UPGen, a unified pre-trained model for both representation learning and generation, which is an encoder-only model based on masked Transformer for multimodal tasks. They show the potential of the masked token prediction model which can be directly used to generate images and language by iteratively ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed UPGen, a unified pre-trained model for both representation learning and generation, which is an encoder-only model based on masked Transformer for multimodal tasks. They show the potential of the masked token prediction model which can be directly used to generate images and language by iter... |
This paper proposes a simple semiparametric few-shot prompting model for open-domain question answering. For the retriever component of the model, the model uses google search instead of Wikipedia. To account for the long length of documents, each doc is broken into a list of paragraphs where each paragraph is a sequen... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a simple semiparametric few-shot prompting model for open-domain question answering. For the retriever component of the model, the model uses google search instead of Wikipedia. To account for the long length of documents, each doc is broken into a list of paragraphs where each paragraph is ... |
This paper proposes a new challenge for AI -- Mobile Construction. In Mobile Construction an agent is initialized in an environment and tasked with building a given structure. The challenge of the task is that it is long horizon and the agent must operate from local observations.
There are 3 versions of the task that ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a new challenge for AI -- Mobile Construction. In Mobile Construction an agent is initialized in an environment and tasked with building a given structure. The challenge of the task is that it is long horizon and the agent must operate from local observations.
There are 3 versions of the ta... |
Authors point out the pros and cons of two existing methods for table compression. Authors demonstrate that combining hashing and clustering based algorithms provides the best of both worlds. Authors prove that this technique works rigorously in the least-square setting.
Strengths:
1) Authors give thorough literature ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
Authors point out the pros and cons of two existing methods for table compression. Authors demonstrate that combining hashing and clustering based algorithms provides the best of both worlds. Authors prove that this technique works rigorously in the least-square setting.
Strengths:
1) Authors give thorough lit... |
This paper poisons unlabeled training data of contrastive learning (CL) to reduce the test accuracy of linear probing. The imperceptible poisons are iteratively optimized to minimize the CL objective function. After the poisons are generated, the authors add them to the training set and run the target CL algorithm from... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper poisons unlabeled training data of contrastive learning (CL) to reduce the test accuracy of linear probing. The imperceptible poisons are iteratively optimized to minimize the CL objective function. After the poisons are generated, the authors add them to the training set and run the target CL algori... |
This paper proposes a way of combining NERF-based volumetric renderer and eulerian fluid simulation to reconstruct fluid dynamics from videos. Given synthesized fluid videos, it can use NERF to estimate the volume of fluid from the images and then simulate it with a differentiable Euler simulator with a ConvNet as the ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a way of combining NERF-based volumetric renderer and eulerian fluid simulation to reconstruct fluid dynamics from videos. Given synthesized fluid videos, it can use NERF to estimate the volume of fluid from the images and then simulate it with a differentiable Euler simulator with a ConvNet... |
The authors propose a probabilistic framework to model modular architectures for the problem of continual learning (CL). They divide the problem of choosing modules in two parts: perceptual transfer (PT), and non-perceptual transfer (NT). In PT, the first l layers are assumed to be pre-trained and the model must choose... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors propose a probabilistic framework to model modular architectures for the problem of continual learning (CL). They divide the problem of choosing modules in two parts: perceptual transfer (PT), and non-perceptual transfer (NT). In PT, the first l layers are assumed to be pre-trained and the model mus... |
This paper proposes a modification of the Go-Explore algorithm to remove the need for hardcoded state representations (which are used in the original version of the algorithm). Specifically, this work proposes to learn an embedding using an auxiliary loss based on predicting the number of timesteps connecting two obser... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a modification of the Go-Explore algorithm to remove the need for hardcoded state representations (which are used in the original version of the algorithm). Specifically, this work proposes to learn an embedding using an auxiliary loss based on predicting the number of timesteps connecting t... |
This paper examines training a specific neural network (namely parallel neural net) using weight decay and shows that the error rate matches the minimax bound for a specific class of functions (i.e., functions with bounded variation).
It looks like a quite interesting effort and the result also appears to be signific... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper examines training a specific neural network (namely parallel neural net) using weight decay and shows that the error rate matches the minimax bound for a specific class of functions (i.e., functions with bounded variation).
It looks like a quite interesting effort and the result also appears to be ... |
In this paper, an approach to controllable image editing with large-scale text-to-image diffusion models is proposed.
More specifically, an approach is presented that first analyzes the cross-attention (and self-attention) layers in the U-net backbone of state-of-the-art text-image diffusion models. It is then shown th... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
In this paper, an approach to controllable image editing with large-scale text-to-image diffusion models is proposed.
More specifically, an approach is presented that first analyzes the cross-attention (and self-attention) layers in the U-net backbone of state-of-the-art text-image diffusion models. It is then ... |
This paper gives new methods for privacy loss accounting in differentially private (DP) algorithms, and illustrates the improvements of its methods over prior work through experimental results. The experiments are mainly performed via the case of running private gradient descent (DP-GD) in various learning tasks. As th... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper gives new methods for privacy loss accounting in differentially private (DP) algorithms, and illustrates the improvements of its methods over prior work through experimental results. The experiments are mainly performed via the case of running private gradient descent (DP-GD) in various learning task... |
Machine Learning models tend to take shortcuts when learning, taking advantage of dataset biases (or spurious correlations) within the data features, that make the task considerably easier to solve. However, these correlations are, indeed, nothing but spurious, and they may not be present during testing, making the mod... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
Machine Learning models tend to take shortcuts when learning, taking advantage of dataset biases (or spurious correlations) within the data features, that make the task considerably easier to solve. However, these correlations are, indeed, nothing but spurious, and they may not be present during testing, making... |
At a high level, this paper is an empirical study of a collection of RL algorithms over a collection of environments.
The algorithms are reimplemented versions of:
* Multiple Quality Diversity (QD) methods, which are evolutionary and do not use exact backpropagation.
* Information Theory-Augmented RL methods, which inv... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
At a high level, this paper is an empirical study of a collection of RL algorithms over a collection of environments.
The algorithms are reimplemented versions of:
* Multiple Quality Diversity (QD) methods, which are evolutionary and do not use exact backpropagation.
* Information Theory-Augmented RL methods, w... |
This paper studies the optimization and generalization of learning over-parameterized implicit neural networks, where only hidden layers are trained. The authors proved global convergence of gradient descent and provide a generalization bound that is initialization sensitive.
Strength:
* This paper proves the global ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the optimization and generalization of learning over-parameterized implicit neural networks, where only hidden layers are trained. The authors proved global convergence of gradient descent and provide a generalization bound that is initialization sensitive.
Strength:
* This paper proves the... |
This paper analyzes the cold posterior effect in Bayesian neural networks. This effect is simply that these models perform better at test time if the posterior is sharpened artificially. The authors argue that there are several works form the literature analyzing the causes of this, with mixed results. In this ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper analyzes the cold posterior effect in Bayesian neural networks. This effect is simply that these models perform better at test time if the posterior is sharpened artificially. The authors argue that there are several works form the literature analyzing the causes of this, with mixed results. ... |
This paper analyzes the impact of data continuity on deep learning models, and propose Lipschitz regularizer which can improve performance of different models based on their preferred data continuity.
The paper has a good motivation, which is reasonable and straightforward;
The proposed regularizer is simple and gener... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper analyzes the impact of data continuity on deep learning models, and propose Lipschitz regularizer which can improve performance of different models based on their preferred data continuity.
The paper has a good motivation, which is reasonable and straightforward;
The proposed regularizer is simple a... |
The authors proposed an autoregressive diffusion model for the graph generation. A node-absorbing diffusion process was introduced. For the forward diffusion a diffusion ordering network was suggested, and for the reverse diffusion a denoising network was designed. These two networks can be trained jointly with a simpl... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The authors proposed an autoregressive diffusion model for the graph generation. A node-absorbing diffusion process was introduced. For the forward diffusion a diffusion ordering network was suggested, and for the reverse diffusion a denoising network was designed. These two networks can be trained jointly with... |
This paper proposes a framework based on denoising diffusion model for solving inverse problem in image processing. The main contribution is the characterization of the forward process for non-linear models which include many well known image formation models. The proposed method has the specificity to both be easier t... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper proposes a framework based on denoising diffusion model for solving inverse problem in image processing. The main contribution is the characterization of the forward process for non-linear models which include many well known image formation models. The proposed method has the specificity to both be ... |
In this paper, the authors propose a KD algorithm that can improve the generalizability of students through perturbation of the teacher's output distribution. The PTLoss proposed in this paper is obtained by expressing a KL-based loss function through a Maclaurin series and then perturbing the terms of the preceding or... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
In this paper, the authors propose a KD algorithm that can improve the generalizability of students through perturbation of the teacher's output distribution. The PTLoss proposed in this paper is obtained by expressing a KL-based loss function through a Maclaurin series and then perturbing the terms of the prec... |
In this paper the authors theoretically investigate the representation function of a class of neural networks, which as far as I know is novel, that they call intra-linked ReLU DNN.
They define intra-linked ReLU DNN as a modification of a standard feed-foward ReLU multi-layer perceptron, and they prove variations of th... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
In this paper the authors theoretically investigate the representation function of a class of neural networks, which as far as I know is novel, that they call intra-linked ReLU DNN.
They define intra-linked ReLU DNN as a modification of a standard feed-foward ReLU multi-layer perceptron, and they prove variatio... |
The paper focuses on collaborative writing based on pretrained LMs. In particular, it proposes a collaborative LM that writes drafts, adds suggestions, proposes edits, and provides explanations. To training is performed on edits and cited documents (at the next step) that is assumed to be available a priori as well as ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper focuses on collaborative writing based on pretrained LMs. In particular, it proposes a collaborative LM that writes drafts, adds suggestions, proposes edits, and provides explanations. To training is performed on edits and cited documents (at the next step) that is assumed to be available a priori as ... |
This paper proposes an aggregation framework (set representation) which can lead to both injective and continuous GNN node embedding even if raw node features are from conintuous space rather than countable space.
## Strength
- This paper shows that there is no way to continuously embed a $M$-size set whose elements a... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes an aggregation framework (set representation) which can lead to both injective and continuous GNN node embedding even if raw node features are from conintuous space rather than countable space.
## Strength
- This paper shows that there is no way to continuously embed a $M$-size set whose el... |
Connections between neurons in the brain are almost universally exclusively excitatory or inhibitory, which seemingly reduces the representational capacity of biological networks. Inspired by this observation, and the assumption that the brain has been optimized throughout evolution, the authors propose that by freezin... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
Connections between neurons in the brain are almost universally exclusively excitatory or inhibitory, which seemingly reduces the representational capacity of biological networks. Inspired by this observation, and the assumption that the brain has been optimized throughout evolution, the authors propose that by... |
This work targets to segmenting scenes into objects and parts under an out-of-distribution setting. In an out-of-distribution setting, the statistical distribution of the testing example is different from that of the training examples. To handle the problem, the authors build a Generating Fast and Slow Network (GFS-Net... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This work targets to segmenting scenes into objects and parts under an out-of-distribution setting. In an out-of-distribution setting, the statistical distribution of the testing example is different from that of the training examples. To handle the problem, the authors build a Generating Fast and Slow Network ... |
This paper firstly attempts at training a model with cross-domain BN statistics. Through untwining NS and AP in dual BN, it demonstrates that what makes it effective lies in two sets of AP instead of disentangled NS (as claimed in prior work).
It points out a hidden flaw in prior work for visualizing the NS to highligh... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper firstly attempts at training a model with cross-domain BN statistics. Through untwining NS and AP in dual BN, it demonstrates that what makes it effective lies in two sets of AP instead of disentangled NS (as claimed in prior work).
It points out a hidden flaw in prior work for visualizing the NS to ... |
The proposed method adopts randomized CNNs ensemble and uses it as descriptor extractors and detects keypoints. Comparing to the state-of-the-art requires training specifically with supervised or self-supervised learning, the method doesn't need any of that and the author claims to have high robustness against illumina... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The proposed method adopts randomized CNNs ensemble and uses it as descriptor extractors and detects keypoints. Comparing to the state-of-the-art requires training specifically with supervised or self-supervised learning, the method doesn't need any of that and the author claims to have high robustness against ... |
In this paper, the authors propose a method for non-linear generative communication from human brain functional data, addressing three problems with connectivity approaches by explicitly modeling the directionality of communication, finding communication at each timestep, and encouraging sparsity in communication. They... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
In this paper, the authors propose a method for non-linear generative communication from human brain functional data, addressing three problems with connectivity approaches by explicitly modeling the directionality of communication, finding communication at each timestep, and encouraging sparsity in communicati... |
The paper extends the Wasserstein autoencoder framework for the unconditional generative model to the setting of having structural constraints e.g., conditional independence on the latent variables (representation, or factors). In greater detail, the authors consider three examples of having additional information to t... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper extends the Wasserstein autoencoder framework for the unconditional generative model to the setting of having structural constraints e.g., conditional independence on the latent variables (representation, or factors). In greater detail, the authors consider three examples of having additional informat... |
This paper considers the multi-agent reinforcement learning policy evaluation (MARL-PE) problem, where $N$ agents collaborate to evaluate the value function of the global states for a target policy. It focuses on how to analyze the communication cost among agents when using the local temporal-difference (TD) learning m... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper considers the multi-agent reinforcement learning policy evaluation (MARL-PE) problem, where $N$ agents collaborate to evaluate the value function of the global states for a target policy. It focuses on how to analyze the communication cost among agents when using the local temporal-difference (TD) le... |
This paper aims to solve the aspect ratio gap between base and novel classes in few-shot object detection. The authors present a very simple CoRPN method, which only uses multiple classifiers in RPN to generate more diverse anchors. To evaluate the effects of the aspect ratio gap, the authors propose ARShift benchmarks... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper aims to solve the aspect ratio gap between base and novel classes in few-shot object detection. The authors present a very simple CoRPN method, which only uses multiple classifiers in RPN to generate more diverse anchors. To evaluate the effects of the aspect ratio gap, the authors propose ARShift be... |
The paper addresses the problem of measuring to what degree the the function is invariant to certain transformations. The authors suggest to use the Lie Derivative to measure such a quantity. Then Local equivariance error (LEE) is introduced. The paper demonstrates how different parts of various models contribute to th... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper addresses the problem of measuring to what degree the the function is invariant to certain transformations. The authors suggest to use the Lie Derivative to measure such a quantity. Then Local equivariance error (LEE) is introduced. The paper demonstrates how different parts of various models contribu... |
This paper proposes Learning to Split (ls), a novel algorithm that automatically detects a potential bias in datasets. More specifically, ls consists of a Splitter and a Predictor. The Splitter learns to divide the dataset into a train split and a test split. Then, the Predictor is trained using the divided train split... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes Learning to Split (ls), a novel algorithm that automatically detects a potential bias in datasets. More specifically, ls consists of a Splitter and a Predictor. The Splitter learns to divide the dataset into a train split and a test split. Then, the Predictor is trained using the divided tra... |
The paper proposes an architecture that can leverage external knowledge to make better decisions in RL problems. The architecture implements an attention mechanism that dynamically chooses to attend to an internal knowledge mapping or external ones. It is trained end-to-end via the Gumbel-softmax trick. Experiments are... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes an architecture that can leverage external knowledge to make better decisions in RL problems. The architecture implements an attention mechanism that dynamically chooses to attend to an internal knowledge mapping or external ones. It is trained end-to-end via the Gumbel-softmax trick. Experim... |
This paper analyses model-based RL methods and studies the impact of the model-based rollout horizon H. The main finding of the paper is that longer horizon do not guarantee better results (they have diminishing returns, and may even be detrimental). This is even true for oracle models (perfect models!), so it is not c... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper analyses model-based RL methods and studies the impact of the model-based rollout horizon H. The main finding of the paper is that longer horizon do not guarantee better results (they have diminishing returns, and may even be detrimental). This is even true for oracle models (perfect models!), so it ... |
In this submission, the authors proposed a method to partition neural response variability within and across brain areas.
****Pros:
- The paper is generally well-written and, for the most part, clear.
- Understanding inter-area communications represents an interesting problem in neuroscience. This is generally a less... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
In this submission, the authors proposed a method to partition neural response variability within and across brain areas.
****Pros:
- The paper is generally well-written and, for the most part, clear.
- Understanding inter-area communications represents an interesting problem in neuroscience. This is generall... |
The paper proposes a new Transformer architecture for better holistic explainability. The architecture generates explanations that reflect all the model components instead of only the attention layers. To do so, the authors take explicit designs for each of the Transformer modules so that they can be summarized by a si... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a new Transformer architecture for better holistic explainability. The architecture generates explanations that reflect all the model components instead of only the attention layers. To do so, the authors take explicit designs for each of the Transformer modules so that they can be summarized... |
In this paper, authors carry out an extensive empirical assessment of a popular evaluation metric for generative models of natural images: the Fréchet Inception Distance. In particular, experiments are designed to indicate properties of the distance when used on data that diverges somehow from ImageNet-1k, used to pre-... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
In this paper, authors carry out an extensive empirical assessment of a popular evaluation metric for generative models of natural images: the Fréchet Inception Distance. In particular, experiments are designed to indicate properties of the distance when used on data that diverges somehow from ImageNet-1k, used... |
The authors propose an anti-symmetric GNN to tackle the oversmoothing phenomenon in GNNs.
The authors suggest an ODE based approach and analyze their model A-DGN and show that it is non-dissipative, i.e., feature/energy preserving and therefore suggest that it can alleviate the oversmoothing phenomenon. Several experi... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose an anti-symmetric GNN to tackle the oversmoothing phenomenon in GNNs.
The authors suggest an ODE based approach and analyze their model A-DGN and show that it is non-dissipative, i.e., feature/energy preserving and therefore suggest that it can alleviate the oversmoothing phenomenon. Severa... |
In the paper, the author proposed a new Transformer variant with linear complexity, which is called Waveformer. In each Waveformer layer, the input will first project to the coefficient space using a forward discrete wavelet transform. Then a linearized attention operation is applied via random features. Finally, the o... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In the paper, the author proposed a new Transformer variant with linear complexity, which is called Waveformer. In each Waveformer layer, the input will first project to the coefficient space using a forward discrete wavelet transform. Then a linearized attention operation is applied via random features. Finall... |
In this work, the authors propose a set of changes to the AlphaFold2 architecture to improve its performance on antibody folding. Most importantly, they show how to replace the MSA component of AlphaFold2 with a much faster encoder approach. Consequently, the proposed approach achieves competitive or better performance... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this work, the authors propose a set of changes to the AlphaFold2 architecture to improve its performance on antibody folding. Most importantly, they show how to replace the MSA component of AlphaFold2 with a much faster encoder approach. Consequently, the proposed approach achieves competitive or better per... |
This paper proposed an unsupervised deep learning method for single image denoising, which only takes a set of noisy images for model training. The basic idea is combining the Recorrupted-to-Recorrupted loss with a sign flipping scheme and a pseudo supervised loss on self-generated samples. The experiments on both synt... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
This paper proposed an unsupervised deep learning method for single image denoising, which only takes a set of noisy images for model training. The basic idea is combining the Recorrupted-to-Recorrupted loss with a sign flipping scheme and a pseudo supervised loss on self-generated samples. The experiments on b... |
The paper studies a multiagent POMDP and proposes a communication mechanism for the agents to exchange information about their decision-making. In the process, the agents have the same objective, aiming to find a joint policy that maximizes their utility. The authors introduced a communication mechanism for the agents,... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies a multiagent POMDP and proposes a communication mechanism for the agents to exchange information about their decision-making. In the process, the agents have the same objective, aiming to find a joint policy that maximizes their utility. The authors introduced a communication mechanism for the... |
The paper studies the problem of dictionary learning for graphs. It proposed an improved graph dictionary learning algorithm based on a robust Gromov-Wasserstein discrepancy (RGWD). It also provided some theoretical results and numerical results.
Strengths:
1. Graph-level dictionary learning is a challenging problem a... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper studies the problem of dictionary learning for graphs. It proposed an improved graph dictionary learning algorithm based on a robust Gromov-Wasserstein discrepancy (RGWD). It also provided some theoretical results and numerical results.
Strengths:
1. Graph-level dictionary learning is a challenging p... |
This paper studies deep neural network through the lens of "class interference". Class interference corresponds to difficulty for a pair of classes, where one class and another class are hard to distinguish for the neural network. More specifically the cross-class test of generalization matrix (CCTM) is used to measure... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies deep neural network through the lens of "class interference". Class interference corresponds to difficulty for a pair of classes, where one class and another class are hard to distinguish for the neural network. More specifically the cross-class test of generalization matrix (CCTM) is used to... |
This paper explores a static alternative pruning method for dynamic pruning methods. They propose channel attention-based learn-to-rank algorithm and channel attention prior among all sample-specific channel saliencies. A Bayesian-based regularization is further introduced to enhance the performance.
Weaknesses
1. The... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper explores a static alternative pruning method for dynamic pruning methods. They propose channel attention-based learn-to-rank algorithm and channel attention prior among all sample-specific channel saliencies. A Bayesian-based regularization is further introduced to enhance the performance.
Weaknesse... |
This paper gives an improved one-shot coreset selection method for classification problems. A one-shot coreset is a subset of the training set, and the goal is to find a coreset/subset of a given size such that the training error on the coreset is minimized.
The paper first studies the ability of the coreset to “cover... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper gives an improved one-shot coreset selection method for classification problems. A one-shot coreset is a subset of the training set, and the goal is to find a coreset/subset of a given size such that the training error on the coreset is minimized.
The paper first studies the ability of the coreset t... |
This paper proposes an interpretable method for Natural Language Inference: first, the premise and hypothesis sentences are splitted into “phrases” (or sub-sentences) with a custom heuristic method, then a pre-trained sentence-BERT is used with cosine-similarity to align phrases from the premise to phrases from the hyp... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes an interpretable method for Natural Language Inference: first, the premise and hypothesis sentences are splitted into “phrases” (or sub-sentences) with a custom heuristic method, then a pre-trained sentence-BERT is used with cosine-similarity to align phrases from the premise to phrases from... |
the paper proposes a more general form of kernel function that is typically used in temporal/spatio-temporal point process, by considering an absolute time-dependent component in addition to the relative spatial-time inputs. In addition, authors made another contribution by proposing a more efficient approach to ensur... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
the paper proposes a more general form of kernel function that is typically used in temporal/spatio-temporal point process, by considering an absolute time-dependent component in addition to the relative spatial-time inputs. In addition, authors made another contribution by proposing a more efficient approach ... |
While existing zero-shot NAS algorithms can usually achieve competitive performance in practice, their zero-shot estimators are unfortunately observed to perform inconsistently in different tasks (e.g., search spaces and datasets), which is even not comparable to the most straightforward one, i.e., the number of parame... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
While existing zero-shot NAS algorithms can usually achieve competitive performance in practice, their zero-shot estimators are unfortunately observed to perform inconsistently in different tasks (e.g., search spaces and datasets), which is even not comparable to the most straightforward one, i.e., the number o... |
This paper introduces a new model called CPINN, which is an agent based approach to the machine learning solution of PDEs. CPINN is an adversarial approach where the discriminator and PINN, the existing approach to solve PDEs using neural networks, plays a game in order to improve the solution precision of PDEs. The au... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper introduces a new model called CPINN, which is an agent based approach to the machine learning solution of PDEs. CPINN is an adversarial approach where the discriminator and PINN, the existing approach to solve PDEs using neural networks, plays a game in order to improve the solution precision of PDEs... |
In this paper, the authors empirically investigate the effects of data augmentation on model performance. They first define a performance metric called "exchange rates", measuring the effects of data augmentation in terms of the size of extra real data required for the same performance. From various perspectives such a... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
In this paper, the authors empirically investigate the effects of data augmentation on model performance. They first define a performance metric called "exchange rates", measuring the effects of data augmentation in terms of the size of extra real data required for the same performance. From various perspective... |
This paper suggests that the features learned by ERM are "good enough" for out-of-distribution (OOD) generalization---they just need to be used in the "right way". This in turns suggest a shift in focus for OOD generalization, from feature learning to _robust regression_. Towards this end, a new objective called DARE i... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper suggests that the features learned by ERM are "good enough" for out-of-distribution (OOD) generalization---they just need to be used in the "right way". This in turns suggest a shift in focus for OOD generalization, from feature learning to _robust regression_. Towards this end, a new objective calle... |
This paper is dedicated to the problem of spurious correlations. The work systematically studies fundamental questions such as how many spuriously correlated training points are necessary for a neural net to get biased towards learning it. Specifically, it investigates the domain in which spurious correlations are rare... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper is dedicated to the problem of spurious correlations. The work systematically studies fundamental questions such as how many spuriously correlated training points are necessary for a neural net to get biased towards learning it. Specifically, it investigates the domain in which spurious correlations ... |
This paper presents new algorithms for learning (two-layer) multi-layer perceptrons with ReLU activations. The approach uses the Burer-Monteiro factorization, which yields a nonconvex optimization problem that (under certain low-rank conditions) does not admit spurious local optima. A collection of other related result... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper presents new algorithms for learning (two-layer) multi-layer perceptrons with ReLU activations. The approach uses the Burer-Monteiro factorization, which yields a nonconvex optimization problem that (under certain low-rank conditions) does not admit spurious local optima. A collection of other relate... |
This paper presents a framework based on Slot Attention (for object-centric representation learning) and Transformer (for reasoning) to tackle the task of RAVEN's Progression Matrices. The input to the network is composed of 8 images as the context, and 8 images as candidate answers to be filled in as the ninth image. ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a framework based on Slot Attention (for object-centric representation learning) and Transformer (for reasoning) to tackle the task of RAVEN's Progression Matrices. The input to the network is composed of 8 images as the context, and 8 images as candidate answers to be filled in as the ninth... |
This paper studies the problem of online reinforcement learning with adversarial losses and switching costs, where there is a cost for switching the policy during learning. The main result is an algorithm that achieves $\tilde{O}(T^{2/3})$ regret and switching cost, for both the setting of known and unknown transitions... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the problem of online reinforcement learning with adversarial losses and switching costs, where there is a cost for switching the policy during learning. The main result is an algorithm that achieves $\tilde{O}(T^{2/3})$ regret and switching cost, for both the setting of known and unknown tra... |
The paper claims in the abstract that they develop a method for exploration that is "theoretically well motivated, and comes with zero computational cost while leading to significant sample efficiency gains in deep reinforcement learning training". The claim about their experiments is that the "technique improves the h... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper claims in the abstract that they develop a method for exploration that is "theoretically well motivated, and comes with zero computational cost while leading to significant sample efficiency gains in deep reinforcement learning training". The claim about their experiments is that the "technique improv... |
To alleviate the computational overhead of the planning-based RL algorithms, this paper proposes an Offline RL method Trajectory Autoencoding Planner (TAP), which learns and plans in the compact latent action space.
The key points are:
* (1) TAP achieves both *spatial abstraction* and *temporal abstraction* by appl... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
To alleviate the computational overhead of the planning-based RL algorithms, this paper proposes an Offline RL method Trajectory Autoencoding Planner (TAP), which learns and plans in the compact latent action space.
The key points are:
* (1) TAP achieves both *spatial abstraction* and *temporal abstraction*... |
The authors propose a graph neural network architecture that uses vectorized neurons, extensive hidden feature mixing operations, and global/local reference frame transformations to produce protein embeddings that are sensitive to or "aware of" residue-specific orientations within the larger macromolecule. They first i... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a graph neural network architecture that uses vectorized neurons, extensive hidden feature mixing operations, and global/local reference frame transformations to produce protein embeddings that are sensitive to or "aware of" residue-specific orientations within the larger macromolecule. They... |
This paper is primarily concerned with **identifying spurious attributes** learned by NNs when learning a specific dataset and **correcting them** during fine-tuning. To this end, they propose mechanistic similarity, a concept that checks whether two pre-trained models with low loss are functionally similar. Based on t... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper is primarily concerned with **identifying spurious attributes** learned by NNs when learning a specific dataset and **correcting them** during fine-tuning. To this end, they propose mechanistic similarity, a concept that checks whether two pre-trained models with low loss are functionally similar. Ba... |
This paper presented an explanation model on motif of the graph. The motif of choice is generated with domain knowledge. Following by an attention model, the attention weights is utilized for the explanation for the GNN from the motif. Compared with state of the art, the method showed high performance. The human readab... | 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 presented an explanation model on motif of the graph. The motif of choice is generated with domain knowledge. Following by an attention model, the attention weights is utilized for the explanation for the GNN from the motif. Compared with state of the art, the method showed high performance. The huma... |
This paper is working on self-supervised articulated object pose estimation at a category level. The objective is to estimate the canonical shape, articulated structure, and transformation that assembles all canonical shapes together. The supervision comes from the overall reconstruction. Experiments are carried out on... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper is working on self-supervised articulated object pose estimation at a category level. The objective is to estimate the canonical shape, articulated structure, and transformation that assembles all canonical shapes together. The supervision comes from the overall reconstruction. Experiments are carrie... |
The paper studies the problem of meta Bayesian optimization (BO), which aims to warm-start the BO process by exploiting knowledge from related tasks. In this paper, the authors propose warm-starting the acquisition function, which takes the form of a classifier in the likelihood-free BO setting. Gradient boosting is fu... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper studies the problem of meta Bayesian optimization (BO), which aims to warm-start the BO process by exploiting knowledge from related tasks. In this paper, the authors propose warm-starting the acquisition function, which takes the form of a classifier in the likelihood-free BO setting. Gradient boosti... |
The paper presents the analysis of characterizing invariance learning with lowest worst-case loss as a special case of partial transportability tasks. The paper introduces the notion of "selection diagram" and considers domain generalization problems from the perspective of transportability analysis. Theoretical analys... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper presents the analysis of characterizing invariance learning with lowest worst-case loss as a special case of partial transportability tasks. The paper introduces the notion of "selection diagram" and considers domain generalization problems from the perspective of transportability analysis. Theoretica... |
This paper proposes a deep study on positional encoding for transformer models. They visualize the positional attentions for various models and point out that the current learned positional encodings have two important properties: locality and symmetry. Based on the two properties, they design handcrafted positional en... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a deep study on positional encoding for transformer models. They visualize the positional attentions for various models and point out that the current learned positional encodings have two important properties: locality and symmetry. Based on the two properties, they design handcrafted posit... |
The authors propose a theoretical take on the well-established intuition that RL-based molecular optimization approaches essentially "attack" locally-overfit predictor functions and result in bad, unrealistic molecular designs on account of these "biases" in the predictive model. They split this bias into a misspecific... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a theoretical take on the well-established intuition that RL-based molecular optimization approaches essentially "attack" locally-overfit predictor functions and result in bad, unrealistic molecular designs on account of these "biases" in the predictive model. They split this bias into a mis... |
This paper proposed ED-HNN, a GNN for hypergraphs inspired by the algorithm for solving hypergraph diffusion (optimization problem on hypergraphs). ED-HNN approximates the gradient-based optimization algorithm for hypergraph diffusion by a message passing on a bipartite graph that is equivalent to the original hypergra... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed ED-HNN, a GNN for hypergraphs inspired by the algorithm for solving hypergraph diffusion (optimization problem on hypergraphs). ED-HNN approximates the gradient-based optimization algorithm for hypergraph diffusion by a message passing on a bipartite graph that is equivalent to the original ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.