review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper proposes a method to augment the memory buffer, aiming to increase the data diversity. Specifically, this paper chooses to use differential equations to model the memory transformation. The experimental results on standard continual learning benchmarks show some improvements of the proposed method.
Strength:... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method to augment the memory buffer, aiming to increase the data diversity. Specifically, this paper chooses to use differential equations to model the memory transformation. The experimental results on standard continual learning benchmarks show some improvements of the proposed method.
S... |
The paper proposes a post-training quantization (PTQ) weight/activation quantization method.
The gist is to simultaneously optimize a multiplicative element-wise scale factor applied to the target tensor
and the quantization step size, under the reconstruction loss. The method can be seen as closely related in spirit
t... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a post-training quantization (PTQ) weight/activation quantization method.
The gist is to simultaneously optimize a multiplicative element-wise scale factor applied to the target tensor
and the quantization step size, under the reconstruction loss. The method can be seen as closely related in ... |
This paper implements a photoreceptor model in Keras and uses it as a front end in training a shallow CNN model to predict retinal responses to white noise stimuli. The paper reports that the photoreceptor model allows the model to generalize better to new lighting levels. This observation is demonstrated using monkey ... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper implements a photoreceptor model in Keras and uses it as a front end in training a shallow CNN model to predict retinal responses to white noise stimuli. The paper reports that the photoreceptor model allows the model to generalize better to new lighting levels. This observation is demonstrated using... |
This paper proposed CLIP-PAE, which projects and augments CLIP image embedding and text embedding in different subspaces, and uses embedding in subspace to optimize the cosine similarity for text-guided image manipulation. It's applied to different baseline methods and outperforms them in both quantitative and qualitat... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed CLIP-PAE, which projects and augments CLIP image embedding and text embedding in different subspaces, and uses embedding in subspace to optimize the cosine similarity for text-guided image manipulation. It's applied to different baseline methods and outperforms them in both quantitative and ... |
This paper provides a novel and comprehensive framework for convolutional neural network inference over a fully homomorphic encryption environment. Based on previous work FHE-MP-CNN, they made significant improvements by
1) using a new approximation for the non-linear activation function ReLU
2) an automatic procedure ... | 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 provides a novel and comprehensive framework for convolutional neural network inference over a fully homomorphic encryption environment. Based on previous work FHE-MP-CNN, they made significant improvements by
1) using a new approximation for the non-linear activation function ReLU
2) an automatic pr... |
The paper proposes a topic model that discovers a topic hierarchy and works with word embeddings. Inference in the model is done by using conditional transport. The inference is also linked to the Bayesian inference.
**Strengths**:
* Tackling topic hierarchy discovery
* Rather extensive experiments
* Easy to follow te... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes a topic model that discovers a topic hierarchy and works with word embeddings. Inference in the model is done by using conditional transport. The inference is also linked to the Bayesian inference.
**Strengths**:
* Tackling topic hierarchy discovery
* Rather extensive experiments
* Easy to f... |
The paper takes an interesting approach to training neural models for programming language (PL) translation - use of Intermediary Representations (IR): language-agnostic pseudocode that describes the semantics of the program. The paper proposes to augment code translation with IRs, specifically LLVM IR, to facilitate t... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper takes an interesting approach to training neural models for programming language (PL) translation - use of Intermediary Representations (IR): language-agnostic pseudocode that describes the semantics of the program. The paper proposes to augment code translation with IRs, specifically LLVM IR, to faci... |
The paper studies how to use proper scoring rules to solve the survival analysis problem, where we want to estimate the distribution F(t) of a random variable T using possibly censored data, which are samples of Z = min{T, C} and whether Z = T or C. While any proper scoring rule can be used as a loss function to train ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The paper studies how to use proper scoring rules to solve the survival analysis problem, where we want to estimate the distribution F(t) of a random variable T using possibly censored data, which are samples of Z = min{T, C} and whether Z = T or C. While any proper scoring rule can be used as a loss function t... |
The paper considers the multi-agent human-AI collaboration setting, and the challenge of coordinating with humans. Specifically focusing on the Hanabi benchmark task, the authors propose the piKL3 method which as the following components:
(1) a human model that captures diverse skill levels controlled via a \lambda reg... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper considers the multi-agent human-AI collaboration setting, and the challenge of coordinating with humans. Specifically focusing on the Hanabi benchmark task, the authors propose the piKL3 method which as the following components:
(1) a human model that captures diverse skill levels controlled via a \la... |
This paper describes a method for sequentially assembling Lego bricks into 3D structures resembling objects from training classes. Efficient GPU-based methods are developed to filter out invalid attachment points and to learn to identify high-value placements. Unlike prior work, the method handles a heterogeneous mix o... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper describes a method for sequentially assembling Lego bricks into 3D structures resembling objects from training classes. Efficient GPU-based methods are developed to filter out invalid attachment points and to learn to identify high-value placements. Unlike prior work, the method handles a heterogeneo... |
The paper proposes a generalized multiscale learning framework for learning in temporally structured environments. The basic idea is that each weight in a NN is decomposed into a sum of subweights, each updated with different learning and decay rates, with all components evolving independently as compared to previous m... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper proposes a generalized multiscale learning framework for learning in temporally structured environments. The basic idea is that each weight in a NN is decomposed into a sum of subweights, each updated with different learning and decay rates, with all components evolving independently as compared to pr... |
The paper introduces a novel method for performing model extraction attacks capable of not only retaining a good performance but also capturing the robustness of the target model when using adversarial training. The proposed attack, which can be applied under a very practical (and restrictive) threat model, creates “un... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper introduces a novel method for performing model extraction attacks capable of not only retaining a good performance but also capturing the robustness of the target model when using adversarial training. The proposed attack, which can be applied under a very practical (and restrictive) threat model, cre... |
In this paper, the author explores the direction of how to utilize SE(3)-equivariance as a novel inductive bias to facilitate sample efficiency. Specifically, it focuses on the same robotic manipulation task as previous work, Neural Descriptor Field. The authors derive an energy function based on the representation of ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the author explores the direction of how to utilize SE(3)-equivariance as a novel inductive bias to facilitate sample efficiency. Specifically, it focuses on the same robotic manipulation task as previous work, Neural Descriptor Field. The authors derive an energy function based on the representa... |
The authors propose a predict-and-search framework for solving mixed-integer linear programming (MILP) problems. Specifically, they first predict the solution distributions, and then search for near-optimal solutions within a trust region constructed from the prediction. Experiments demonstrate that the proposed framew... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The authors propose a predict-and-search framework for solving mixed-integer linear programming (MILP) problems. Specifically, they first predict the solution distributions, and then search for near-optimal solutions within a trust region constructed from the prediction. Experiments demonstrate that the propose... |
The authors present a novel time-varying importance weight estimator that can detect gradual shifts in the distribution of data. They evaluate their methods on synthetic image classification and reinforcement learning tasks that exhibit distribution shifts and show that the proposed method largely outperforms the stand... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The authors present a novel time-varying importance weight estimator that can detect gradual shifts in the distribution of data. They evaluate their methods on synthetic image classification and reinforcement learning tasks that exhibit distribution shifts and show that the proposed method largely outperforms t... |
In recent years, goal conditioned reinforcement learning approaches have received a lot of attention in the literature. However, curricula for selecting goals during training have predominantly been based on heuristics, which still lead to poor sample-efficiency in realistic scenarios. The paper proposes using curricul... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In recent years, goal conditioned reinforcement learning approaches have received a lot of attention in the literature. However, curricula for selecting goals during training have predominantly been based on heuristics, which still lead to poor sample-efficiency in realistic scenarios. The paper proposes using ... |
The authors present an interesting discussion and MEG classification results in SL vs. GL settings. The submission is not a good match for ICLR, and it would interest the applied neuroscience community. The technical contribution of a relatively standard/off-the-shelf WaveNet to MEG is the major weakness of the paper. ... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors present an interesting discussion and MEG classification results in SL vs. GL settings. The submission is not a good match for ICLR, and it would interest the applied neuroscience community. The technical contribution of a relatively standard/off-the-shelf WaveNet to MEG is the major weakness of the... |
The authors proposed a so-called Crossword Puzzle method to find the optimal criteria for network pruning. The key idea is to find a so-called Fabulous Coordinate which satisfies three key properties for pruning. The authors validated their method on ImageNet and show that they can compress ResNet family by 50% without... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The authors proposed a so-called Crossword Puzzle method to find the optimal criteria for network pruning. The key idea is to find a so-called Fabulous Coordinate which satisfies three key properties for pruning. The authors validated their method on ImageNet and show that they can compress ResNet family by 50%... |
The paper under consideration proposes a method for inverting stochastic diffusion models. Based on this method the authors come up with an (conditional) unpaired image-to-image translation approach called “CycleDiffusion”. The validity of the proposed methodology is verified by several applications.
Strength:
1) The p... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper under consideration proposes a method for inverting stochastic diffusion models. Based on this method the authors come up with an (conditional) unpaired image-to-image translation approach called “CycleDiffusion”. The validity of the proposed methodology is verified by several applications.
Strength:
... |
The paper proposes a method to learn equivariant representations that are steerable by a learned function in representation space. They empirically demonstrate the method as an improvement to ImageNet classification, where the pre-trained model performs better in transfer, retrieval, corruption, OOD detection settings.... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a method to learn equivariant representations that are steerable by a learned function in representation space. They empirically demonstrate the method as an improvement to ImageNet classification, where the pre-trained model performs better in transfer, retrieval, corruption, OOD detection s... |
The paper introduces a new notion of coverability, and shows its sufficiency for efficient online reinforcement learning. This new notion is compared with existing concepts from both offline and online RL. For offline RL, the authors show other weaker notions of coverage are not sufficient for efficient exploration. Fo... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper introduces a new notion of coverability, and shows its sufficiency for efficient online reinforcement learning. This new notion is compared with existing concepts from both offline and online RL. For offline RL, the authors show other weaker notions of coverage are not sufficient for efficient explora... |
This paper studies the use of higher order gradient methods for multi-agent RL with high dimensional state space. It shows that existing methods can lead to miscoordination among agents. A hierarchical reasoning algorithm is proposed. Experimental results are presented to show the applicability of the method.
Streng... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the use of higher order gradient methods for multi-agent RL with high dimensional state space. It shows that existing methods can lead to miscoordination among agents. A hierarchical reasoning algorithm is proposed. Experimental results are presented to show the applicability of the method.
... |
The authors translate diffusion models to the protein structure domain to design new protein scaffolds and motifs.
The diffusion approach seems interesting and a promising way to explore sequence and structure space.
Section 5.1 - Two works are cited to justify using AlphaFold2 as the gold standard of accuracy of tho... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors translate diffusion models to the protein structure domain to design new protein scaffolds and motifs.
The diffusion approach seems interesting and a promising way to explore sequence and structure space.
Section 5.1 - Two works are cited to justify using AlphaFold2 as the gold standard of accurac... |
The paper presents a transformer-based conditional model to generate multitrack symbolic music. The model utilizes a "description to sequence" learning scheme. The authors show that the model produces more desirable musical sequences when human-interpretable features are incorporated during training.
Strengths:
- No... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a transformer-based conditional model to generate multitrack symbolic music. The model utilizes a "description to sequence" learning scheme. The authors show that the model produces more desirable musical sequences when human-interpretable features are incorporated during training.
Strength... |
This paper studies the computation of wave functions in Kohn-Sham Density Functional Theory (KS-DFT). Differently from conventional works, the authors use neural network (NN) models with a stochastic optimization method in machine learning to tackle the problem. The problem is formulated as the minimization problem of ... | 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 computation of wave functions in Kohn-Sham Density Functional Theory (KS-DFT). Differently from conventional works, the authors use neural network (NN) models with a stochastic optimization method in machine learning to tackle the problem. The problem is formulated as the minimization pro... |
The authors apply GFlowNets to the problem of scheduling operations in a computational graph on homogeneous parallel hardware. The evaluate conditional GFlowNets for the first time with the computational graph being the conditioning input and the schedule being the output. They also innovate by conditioning on the rewa... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors apply GFlowNets to the problem of scheduling operations in a computational graph on homogeneous parallel hardware. The evaluate conditional GFlowNets for the first time with the computational graph being the conditioning input and the schedule being the output. They also innovate by conditioning on ... |
This paper introduces a machine learning based heuristic method to adaptively choose an important parameter of ADMM when applying to solve quadratic programming problems. The proposed method added a temporal component via a gated recurrent unit, which allows the model to incorporate consecutive information from previou... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper introduces a machine learning based heuristic method to adaptively choose an important parameter of ADMM when applying to solve quadratic programming problems. The proposed method added a temporal component via a gated recurrent unit, which allows the model to incorporate consecutive information from... |
This paper provides an alternative to the softmax + cross entropy loss widely used in text classification tasks. The proposed method, named adaptive sparse softmax (AS-Softmax), is inspired by the sparse-softmax (Sun et al., 2021) and changes the training objective by including a binary term that, when multiplied by a ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper provides an alternative to the softmax + cross entropy loss widely used in text classification tasks. The proposed method, named adaptive sparse softmax (AS-Softmax), is inspired by the sparse-softmax (Sun et al., 2021) and changes the training objective by including a binary term that, when multipli... |
This work proposes a new type of embeddings of text. Instead of relying on the output of the neural network, the proposed embedding is generated by the actual weights of the network, when it is tuned with the specific content/text. The work shows that the embeddings generated this way capture semantic differences betwe... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes a new type of embeddings of text. Instead of relying on the output of the neural network, the proposed embedding is generated by the actual weights of the network, when it is tuned with the specific content/text. The work shows that the embeddings generated this way capture semantic differenc... |
This paper introduces a unified analysis for accelerated perceptrons via reduction to a certain minmax game. Improved rates are then achieved by designing the dynamics of the min and max player. The analysis is simple and easily captures several prior works as special cases, while also easily enabling the authors to de... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper introduces a unified analysis for accelerated perceptrons via reduction to a certain minmax game. Improved rates are then achieved by designing the dynamics of the min and max player. The analysis is simple and easily captures several prior works as special cases, while also easily enabling the autho... |
The paper claims to prove representation bottlenecks of ConvNets considering the capacity of representing different frequency components of an input sample. It claims to introduce the rule of the forward propagation of intermediate-layer spectrum maps, as equivalent to the forward propagation of feature maps through a ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper claims to prove representation bottlenecks of ConvNets considering the capacity of representing different frequency components of an input sample. It claims to introduce the rule of the forward propagation of intermediate-layer spectrum maps, as equivalent to the forward propagation of feature maps th... |
The paper introduces a new neural network technique for survival analysis termed the neural frailty model, which in turn is inspired by the literature on frailty models. The authors provide a theoretical analysis for the convergence rate of the conditional distributions of the event data and demonstrate state-of-the-ar... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper introduces a new neural network technique for survival analysis termed the neural frailty model, which in turn is inspired by the literature on frailty models. The authors provide a theoretical analysis for the convergence rate of the conditional distributions of the event data and demonstrate state-o... |
This paper works on text-guided and image-guided image translations. The authors propose a guided diffusion method that tries to 1) maintain the structure of the source image, and 2) generate semantic content that matches the guidance. Specifically, contrastive loss, CLIP loss, semantic style loss and divergence loss i... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper works on text-guided and image-guided image translations. The authors propose a guided diffusion method that tries to 1) maintain the structure of the source image, and 2) generate semantic content that matches the guidance. Specifically, contrastive loss, CLIP loss, semantic style loss and divergenc... |
This paper proposes a mathematical framework for working with graphs. The framework is based on two main tools: the multiparameter persistence and persistent homology. This build a framework for topological data analysis that is able to use the data attributes on the nodes. The experiments show promising results with s... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes a mathematical framework for working with graphs. The framework is based on two main tools: the multiparameter persistence and persistent homology. This build a framework for topological data analysis that is able to use the data attributes on the nodes. The experiments show promising result... |
This paper proposes a bespoke CapsNet model called TabCaps that is tailored to tabular data processing. This seems to be the first effort to develop a bottom-up approach specifically for tabular data whereby all features are considered as a vector (per the properties of capsules) which are then transformed by a multiva... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a bespoke CapsNet model called TabCaps that is tailored to tabular data processing. This seems to be the first effort to develop a bottom-up approach specifically for tabular data whereby all features are considered as a vector (per the properties of capsules) which are then transformed by a... |
The proposed paper introduces a novel method for robot motion planing based on a wave propagation model. The introduced model generates path solutions based on a non-linear first-order PDE (Eikonal Equation). The method is validated for motion planning tasks in cluttered 3D environments (Gibson). The discussed results ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The proposed paper introduces a novel method for robot motion planing based on a wave propagation model. The introduced model generates path solutions based on a non-linear first-order PDE (Eikonal Equation). The method is validated for motion planning tasks in cluttered 3D environments (Gibson). The discussed ... |
This paper concentrates on fair machine learning (ML), specifically studying the "sub-optimality" caused when the design phase of a fair estimator is forbidden to have access to the sensitive attributes which it should protect in order to provide fairness. To this end, the authors take the setting where the design phas... | 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 concentrates on fair machine learning (ML), specifically studying the "sub-optimality" caused when the design phase of a fair estimator is forbidden to have access to the sensitive attributes which it should protect in order to provide fairness. To this end, the authors take the setting where the des... |
The manuscript proposed a Generative Communication Model by individually modeling the signals (by an S-VAE) and the communication sent at each graph node (by an RNN+MLP). Experiment on synthetic data shows that the model can successfully recover the designed communication between regions. Experiment on real (HCP) fMRI ... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The manuscript proposed a Generative Communication Model by individually modeling the signals (by an S-VAE) and the communication sent at each graph node (by an RNN+MLP). Experiment on synthetic data shows that the model can successfully recover the designed communication between regions. Experiment on real (HC... |
The paper addresses the question of whether misspecification of human behavior models for IRL can lead to detrimental effects on reward inference accuracy, and provide theoretical guarantees showing that while adversarial situations can be constructed such that small human behavior models lead to "catastrophic" inferen... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper addresses the question of whether misspecification of human behavior models for IRL can lead to detrimental effects on reward inference accuracy, and provide theoretical guarantees showing that while adversarial situations can be constructed such that small human behavior models lead to "catastrophic"... |
The paper proposes a non-probabilistic approach to use autoencoder for learning disentanglement in unsupervised manner. It also proposes a new metric to quantify the disentanglement.
The paper explains the new framework in detail and compares against other autoencoder models on a range of disentanglement metrics.
At t... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a non-probabilistic approach to use autoencoder for learning disentanglement in unsupervised manner. It also proposes a new metric to quantify the disentanglement.
The paper explains the new framework in detail and compares against other autoencoder models on a range of disentanglement metri... |
The authors tackle the current limitation that most existing NAS methods suffer from, which are unsatisfactory generalizability and stability, such as generating a dominant number of skip connections or poor test performance. To address this, the authors propose a transferability-encouraging tri-level optimization fram... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors tackle the current limitation that most existing NAS methods suffer from, which are unsatisfactory generalizability and stability, such as generating a dominant number of skip connections or poor test performance. To address this, the authors propose a transferability-encouraging tri-level optimizat... |
The paper focuses on the problem of imitation learning in stochastic environments. The stochasticity in the environment is characterized by the parameter $\xi$, that is independent of the expert/agent.
The goal of this work is to build imitation agent that achieve distributional realism — match the expert distribution... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper focuses on the problem of imitation learning in stochastic environments. The stochasticity in the environment is characterized by the parameter $\xi$, that is independent of the expert/agent.
The goal of this work is to build imitation agent that achieve distributional realism — match the expert dist... |
The paper proposes Gradient Annealing (GA) and AutoSparse, two complementary approaches to achieve a towards optimal sparsity-accuracy trade-off during the training of sparse neural networks.
Strength:
* Interesting idea with a small level of novelty
* The proposed method seems to be able to slightly improve the perfor... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes Gradient Annealing (GA) and AutoSparse, two complementary approaches to achieve a towards optimal sparsity-accuracy trade-off during the training of sparse neural networks.
Strength:
* Interesting idea with a small level of novelty
* The proposed method seems to be able to slightly improve th... |
The paper provides a theoretical study on the difficulty of learning an unknown ReLU network with a fixed number of layers by another ReLU network with the same number of layers. While previous attempts usually look at NP-hardness of the optimization process of this task, even allowing for known data distribution, the ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper provides a theoretical study on the difficulty of learning an unknown ReLU network with a fixed number of layers by another ReLU network with the same number of layers. While previous attempts usually look at NP-hardness of the optimization process of this task, even allowing for known data distributi... |
This paper provides a gradient based approach to causal discovery in the presence of unobserved confounding. By making assumptions on the structure of the graph along with additive noise assumptions - identification results for non-linear SCMs are provides. Since the conditional independencies involves learning functio... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper provides a gradient based approach to causal discovery in the presence of unobserved confounding. By making assumptions on the structure of the graph along with additive noise assumptions - identification results for non-linear SCMs are provides. Since the conditional independencies involves learning... |
This paper enhances two previous algorithms for non-convex finite sum minimization, SARAH and L-SVRG, by adding a diagonal preconditioner from Jahani et al. The authors prove that the preconditioned methods, which they call Scaled SARAH and Scaled L-SVRG, both converge. The convergence bounds are better than those of p... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper enhances two previous algorithms for non-convex finite sum minimization, SARAH and L-SVRG, by adding a diagonal preconditioner from Jahani et al. The authors prove that the preconditioned methods, which they call Scaled SARAH and Scaled L-SVRG, both converge. The convergence bounds are better than th... |
This paper focuses on multi-user reinforcement learning with low-rank rewards. This paper considers the problem where all the users play in the MDPs with the same dynamics but different reward functions. Exploiting the low-rank property of the reward matrix, the authors design two novel sampling algorithms for tabular ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper focuses on multi-user reinforcement learning with low-rank rewards. This paper considers the problem where all the users play in the MDPs with the same dynamics but different reward functions. Exploiting the low-rank property of the reward matrix, the authors design two novel sampling algorithms for ... |
The authors address the problem of defending against adversarial audio attacks using audio purification based on recently popular diffusion models for recovering clean audio. Overall the idea is quite interesting and the experimental evaluation shows strong performance of the proposed idea.
Strengths:
1. Excellent ide... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors address the problem of defending against adversarial audio attacks using audio purification based on recently popular diffusion models for recovering clean audio. Overall the idea is quite interesting and the experimental evaluation shows strong performance of the proposed idea.
Strengths:
1. Excel... |
This paper studies how to scale up the training/inference of kernel network. It uses data center to reduce the memory and inference cost. It proposes to use inexact stochastic approximation to approximate functional gradient and the preconditioner. Such stochastic gradient allows us to use data augmentation during the ... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper studies how to scale up the training/inference of kernel network. It uses data center to reduce the memory and inference cost. It proposes to use inexact stochastic approximation to approximate functional gradient and the preconditioner. Such stochastic gradient allows us to use data augmentation dur... |
This paper argues that for transformers going deeper isn’t always better, and they show this by proving that wide single layer transformers sometimes outperform deeper transformers in 4 NLP classification tasks, with no pre-training.
The paper reports results on an interesting dimension: for different numbers of atten... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper argues that for transformers going deeper isn’t always better, and they show this by proving that wide single layer transformers sometimes outperform deeper transformers in 4 NLP classification tasks, with no pre-training.
The paper reports results on an interesting dimension: for different numbers ... |
This paper presents Adaptive Subgoal Search (AdaSubS), an approach to learning-driven search that uses proposed subgoals and a novel "verifier" network to make more direct and rapid progress towards the target state. The approach builds upon recent progress in this space, notably the BF-kSubS, by proposing subgoals tho... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents Adaptive Subgoal Search (AdaSubS), an approach to learning-driven search that uses proposed subgoals and a novel "verifier" network to make more direct and rapid progress towards the target state. The approach builds upon recent progress in this space, notably the BF-kSubS, by proposing subg... |
This paper proposes a framework for compositional structure modeling in control tasks, that is a modification of the typical Markov Decision Process (MDP). The framework, called Entity-Factored Markov Decision Process provides several insights into developing robot control policy architectures for compositional general... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a framework for compositional structure modeling in control tasks, that is a modification of the typical Markov Decision Process (MDP). The framework, called Entity-Factored Markov Decision Process provides several insights into developing robot control policy architectures for compositional... |
The paper deals with using NAS to search for attackable (image) classification architectures/models and attack generators. The generated attacks have to be image-dependent to make it difficult to fix them. To make this problem tracktable, the authors propose a proxy metric which does not require cumbersome training, bu... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper deals with using NAS to search for attackable (image) classification architectures/models and attack generators. The generated attacks have to be image-dependent to make it difficult to fix them. To make this problem tracktable, the authors propose a proxy metric which does not require cumbersome trai... |
This paper elucidates the importance of instance-dependent transformation in augmenting data and proposes InstaAug that learns the input-dependent transformation which could preserve the information of original input while maintaining sufficient transformation diversity. The invariance module $\phi$ is jointly trained ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper elucidates the importance of instance-dependent transformation in augmenting data and proposes InstaAug that learns the input-dependent transformation which could preserve the information of original input while maintaining sufficient transformation diversity. The invariance module $\phi$ is jointly ... |
This paper presents an SSL method to learn from noisy labeled data. Instead of the previous method, which is based on instance-level SSL, this method is a set level. Authors proposed to learn M models (with parameters $\teta^'_i=(1:M))$ ) wherein in each of the models, the labels of two classes are corrupted. Finally,... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents an SSL method to learn from noisy labeled data. Instead of the previous method, which is based on instance-level SSL, this method is a set level. Authors proposed to learn M models (with parameters $\teta^'_i=(1:M))$ ) wherein in each of the models, the labels of two classes are corrupted. ... |
The authors handles multivariate forecasting problem by utilizing a univariate forecasting model to predict individual dimensions of multivariate setting. They validate the proposed model with its baseline (Transformer-based model, Informer, Autoformer) on benchmark datasets and the experimental results show that the p... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors handles multivariate forecasting problem by utilizing a univariate forecasting model to predict individual dimensions of multivariate setting. They validate the proposed model with its baseline (Transformer-based model, Informer, Autoformer) on benchmark datasets and the experimental results show th... |
In this paper, the authors modeled the data structure with acyclic directed mixed graphs (ADMGs). Under the condition where latent confounding is present, a new autoregressive flow-based approach is proposed to learn the causal and functional relations of the given data. In the experiments. the authors validate compet... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
In this paper, the authors modeled the data structure with acyclic directed mixed graphs (ADMGs). Under the condition where latent confounding is present, a new autoregressive flow-based approach is proposed to learn the causal and functional relations of the given data. In the experiments. the authors validat... |
This paper proposes a method to perform adversarial training with adaptive epsilon per training sample. Several heuristics are proposed to determine epsilon per sample. The benefit is improved accuracy-robustness trade-off.
Strength: This paper is working on a meaningful problem. Finding adaptive epsilon per training s... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper proposes a method to perform adversarial training with adaptive epsilon per training sample. Several heuristics are proposed to determine epsilon per sample. The benefit is improved accuracy-robustness trade-off.
Strength: This paper is working on a meaningful problem. Finding adaptive epsilon per tr... |
The authors proposed to replace the cross-entropy loss with the MCR2 loss in federated learning. They provided a theoretical analysis of learning convergence of the global model using the MCR2 loss. They also empirically show the convergence of the loss and the orthogonality of the class-wise subspaces learnt.
Strengt... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors proposed to replace the cross-entropy loss with the MCR2 loss in federated learning. They provided a theoretical analysis of learning convergence of the global model using the MCR2 loss. They also empirically show the convergence of the loss and the orthogonality of the class-wise subspaces learnt. ... |
In this paper, the authors consider the potential approximability of solutions to PDEs by neural networks. This topic has attracted significant interest in the past few years and a key question, as noted by the authors, is to understand which class of PDEs give rise to solutions that can be efficiently represented by a... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
In this paper, the authors consider the potential approximability of solutions to PDEs by neural networks. This topic has attracted significant interest in the past few years and a key question, as noted by the authors, is to understand which class of PDEs give rise to solutions that can be efficiently represen... |
The paper proposes a vector quantized autoencoder that minimizes the Wasserstein distance between the encoded samples and the code word. In short, the loss function is simply the quantized autoencoding loss plus a penalty term that measures the Wasserstein distance between the empirical distribution of the encoded samp... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a vector quantized autoencoder that minimizes the Wasserstein distance between the encoded samples and the code word. In short, the loss function is simply the quantized autoencoding loss plus a penalty term that measures the Wasserstein distance between the empirical distribution of the enco... |
This paper proposes a system to predict types in untyped or partially typed code. The authors propose to treat type prediction as a code completion problem, which can be solved using transformer models trained for code. However, just applying a transformer model "out of the box" would fail because there can be caller-c... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a system to predict types in untyped or partially typed code. The authors propose to treat type prediction as a code completion problem, which can be solved using transformer models trained for code. However, just applying a transformer model "out of the box" would fail because there can be ... |
Scheduling operations in a computation graph on parallel hardware is NP-hard. This makes ML approaches such as RL attractive. However, the high cost of evaluation necessitates the use of proxy reward models, which can be overfitted to with reward maximization, resulting in poor performance on actual hardware. This is a... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Scheduling operations in a computation graph on parallel hardware is NP-hard. This makes ML approaches such as RL attractive. However, the high cost of evaluation necessitates the use of proxy reward models, which can be overfitted to with reward maximization, resulting in poor performance on actual hardware. T... |
This submission provides new stability guarantees and transferability bounds for general graph convolutional networks. The analysis framework can be applied to both directed and undirected graphs based on recent advanced tools. Specifically, node-level and edge-level perturbations with some new properties are presented... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This submission provides new stability guarantees and transferability bounds for general graph convolutional networks. The analysis framework can be applied to both directed and undirected graphs based on recent advanced tools. Specifically, node-level and edge-level perturbations with some new properties are p... |
This paper proposes a new paradigm – adding a computational physics engine to aid language modeling on physical reasoning tasks.
*Strength:*
The idea proposed by this paper is pretty novel and effective;
This paper is well written and easy to understand;
This paper provides a new benchmark to test physic related... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes a new paradigm – adding a computational physics engine to aid language modeling on physical reasoning tasks.
*Strength:*
The idea proposed by this paper is pretty novel and effective;
This paper is well written and easy to understand;
This paper provides a new benchmark to test physic... |
The paper considers the problem of learning fair representations containing little or no information about the protected variable. In the absence of prior knowledge about downstream tasks, it requires the learned representations to be both fair and discriminative enough, which turns out to be a trade-off and may incur ... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper considers the problem of learning fair representations containing little or no information about the protected variable. In the absence of prior knowledge about downstream tasks, it requires the learned representations to be both fair and discriminative enough, which turns out to be a trade-off and ma... |
The paper develops a new algorithm for gene finding which is compatible with embeddings obtained from deep and transfer learning.
The paper argues the prior models for gene prediction “lack the flexibility to incorporate deep learning representation learning techniques” as a potential drawback. This sets the motivatio... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper develops a new algorithm for gene finding which is compatible with embeddings obtained from deep and transfer learning.
The paper argues the prior models for gene prediction “lack the flexibility to incorporate deep learning representation learning techniques” as a potential drawback. This sets the m... |
This paper proposes the multi-explanation graph attention network (MEGAN), which is an attention-based self-explaining model for graph regression and classification. MEGAN features multiple explanation channels independent of the task specifications. The edge explanations are given as the edge importance tensor, which ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes the multi-explanation graph attention network (MEGAN), which is an attention-based self-explaining model for graph regression and classification. MEGAN features multiple explanation channels independent of the task specifications. The edge explanations are given as the edge importance tensor... |
The paper introduced a more detailed prompt for LLM math reasoning, and study their impact on task composition and assisting in solving math problems in a recent dataset. The proposed method shows very high accuracy in simple tasks and starts to fall short when the tasks are more complex. One of the challenges is the m... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper introduced a more detailed prompt for LLM math reasoning, and study their impact on task composition and assisting in solving math problems in a recent dataset. The proposed method shows very high accuracy in simple tasks and starts to fall short when the tasks are more complex. One of the challenges ... |
The authors propose, Lifting Contrastive Learning (LiftedCL), a feature representation learning approach based on contrastive learning with primary focus on the task of 2D/3D human pose estimation from a single image. Authors further use kinematic chain space (KCS) layer and a discriminator to regularise the skeleton. ... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The authors propose, Lifting Contrastive Learning (LiftedCL), a feature representation learning approach based on contrastive learning with primary focus on the task of 2D/3D human pose estimation from a single image. Authors further use kinematic chain space (KCS) layer and a discriminator to regularise the sk... |
This paper proposes Hierarchical Prompting that follows coarse-to-fine semantic structure. The authors predefine $M$ learnable prompts (coarse classes) and inject them into the Transformer layers. The $M$ learnable prompts are used to predict the coarse-class prototype $S$ and are optimized by a softmax loss $\mathcal{... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes Hierarchical Prompting that follows coarse-to-fine semantic structure. The authors predefine $M$ learnable prompts (coarse classes) and inject them into the Transformer layers. The $M$ learnable prompts are used to predict the coarse-class prototype $S$ and are optimized by a softmax loss $\... |
This paper introduces a novel framework for cross-domain graph pre-training using fewer training samples. The paper first demonstrates the phenomenon of the “curse of big data” - more training and graph datasets do not always lead to better performance in downstream node and graph classification tasks. Next, it present... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper introduces a novel framework for cross-domain graph pre-training using fewer training samples. The paper first demonstrates the phenomenon of the “curse of big data” - more training and graph datasets do not always lead to better performance in downstream node and graph classification tasks. Next, it... |
This paper studies the over-smoothing problem in graph neural networks and proposes a new architecture called deeply-supervised GNN (DSGNN) as a solution. Inspired by deeply-supervised nets (Lee et al., 2015), DSGNN has a classifier attached to each hidden layer and gets supervision during training. During inference, D... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the over-smoothing problem in graph neural networks and proposes a new architecture called deeply-supervised GNN (DSGNN) as a solution. Inspired by deeply-supervised nets (Lee et al., 2015), DSGNN has a classifier attached to each hidden layer and gets supervision during training. During infe... |
This paper derives a convex formulation for self-attention in transformers. The authors extend the formulation to various transformers architectures and then apply this convex formulation with an additional regularization for training transformers to find a globally optimal solution. The proposed formulation also sugge... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper derives a convex formulation for self-attention in transformers. The authors extend the formulation to various transformers architectures and then apply this convex formulation with an additional regularization for training transformers to find a globally optimal solution. The proposed formulation al... |
This paper designs a new notion “almost sure support stability (a.s. support stability)” to measure the learned model’s stability to the disturbance in the training data. The authors then give a generalization bound based on this stability measure. The authors also show that ReLU network is a.s. locally smooth and a.s.... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper designs a new notion “almost sure support stability (a.s. support stability)” to measure the learned model’s stability to the disturbance in the training data. The authors then give a generalization bound based on this stability measure. The authors also show that ReLU network is a.s. locally smooth ... |
This paper proposes a novel protein sequence pre-training method, Knowledge-exploited Auto-encoder for Proteins (KeAP). Instead of injecting biomedical knowledge by directly applying knowledge embedding constraints as OntoProtein, KeAP employ a cross-attention mechanism to attentively inject relevant relation and attri... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a novel protein sequence pre-training method, Knowledge-exploited Auto-encoder for Proteins (KeAP). Instead of injecting biomedical knowledge by directly applying knowledge embedding constraints as OntoProtein, KeAP employ a cross-attention mechanism to attentively inject relevant relation a... |
The paper propose a reparametrization of the L1 loss to efficiently solve L1 penalized problems.
Experiments are proposed on standard L1-penalized optimization problems and sparsity deep learning (network compression).
Strength
The application to network compression seems interesting
Weaknesses
- IMO the paper clearl... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The paper propose a reparametrization of the L1 loss to efficiently solve L1 penalized problems.
Experiments are proposed on standard L1-penalized optimization problems and sparsity deep learning (network compression).
Strength
The application to network compression seems interesting
Weaknesses
- IMO the pape... |
This paper extends the context-dependent gating (XDG) approach to avoid catastrophic forgetting in life-long learning, i.e. learning a new task might overwrite the weights learned for another task. XDG randomly disabled 20% of the neurons in each layer to ensure the creation of distinct neural ensembles for different t... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper extends the context-dependent gating (XDG) approach to avoid catastrophic forgetting in life-long learning, i.e. learning a new task might overwrite the weights learned for another task. XDG randomly disabled 20% of the neurons in each layer to ensure the creation of distinct neural ensembles for dif... |
The authors consider the regret minimization problem for infinite-horizon average-reward Markov decision process (MDP). They propose the Reward-Weighted Posterior Sampling of Policy (RWPSP) a Thompson sampling inspired algorithm. Precisely RWPSP maintains a posterior distribution over policies with is updated with the... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors consider the regret minimization problem for infinite-horizon average-reward Markov decision process (MDP). They propose the Reward-Weighted Posterior Sampling of Policy (RWPSP) a Thompson sampling inspired algorithm. Precisely RWPSP maintains a posterior distribution over policies with is updated ... |
This paper finds that relational reasoning is a key component of mathematical reasoning, whether using natural language or abstract symbols as indicated by our experiments on the GSM8K and the unit conversion tasks. Training the models with relational abstraction can outperform models trained using numerical expression... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper finds that relational reasoning is a key component of mathematical reasoning, whether using natural language or abstract symbols as indicated by our experiments on the GSM8K and the unit conversion tasks. Training the models with relational abstraction can outperform models trained using numerical ex... |
The paper presents Diffusion-QL, an offline RL method that can be seen as an actor-critic method, where the actor is a generative diffusion models, and the critic is a standard Q-function. The diffusion model is trained with supervised learning to imitate the data distribution. The main benefit of diffusion models is t... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper presents Diffusion-QL, an offline RL method that can be seen as an actor-critic method, where the actor is a generative diffusion models, and the critic is a standard Q-function. The diffusion model is trained with supervised learning to imitate the data distribution. The main benefit of diffusion mod... |
This paper proposes a TopoZero framework to improve structure alignment for common space learning methods. Based on two weaknesses of HSVA (trains model with local structure and loses some high-dimensional structure information), the authors propose a Topology-guided sampling strategy and a Topology Alignment Module. E... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a TopoZero framework to improve structure alignment for common space learning methods. Based on two weaknesses of HSVA (trains model with local structure and loses some high-dimensional structure information), the authors propose a Topology-guided sampling strategy and a Topology Alignment M... |
This paper provides a new analysis for implicit regularization for the heavy ball momentum gradient descent. The paper shows that the momentum updates for gradient descent and its stochastic counterpart are closer to the modified regularized gradient flow than previously known. Their analysis is based on studying a pie... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper provides a new analysis for implicit regularization for the heavy ball momentum gradient descent. The paper shows that the momentum updates for gradient descent and its stochastic counterpart are closer to the modified regularized gradient flow than previously known. Their analysis is based on studyi... |
The paper explores an under-investigated problem, that of automatically discovering geometric quantities that indicate the potential success or failure of fitting tasks. These include passing an object through a tube-like container, putting an object in a drawer or a sphere, etc. The authors use the term eigen-lengths ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper explores an under-investigated problem, that of automatically discovering geometric quantities that indicate the potential success or failure of fitting tasks. These include passing an object through a tube-like container, putting an object in a drawer or a sphere, etc. The authors use the term eigen-... |
This work proposes a method to accelerate both training and inference in Graph Neural Networks by quantizing both the
weights and all the intermediate results (features, errors, gradients) to 8 bits. The proposed training and inference
pipeline can be executed on integer-only hardware.
To deal with the high dynamic ra... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This work proposes a method to accelerate both training and inference in Graph Neural Networks by quantizing both the
weights and all the intermediate results (features, errors, gradients) to 8 bits. The proposed training and inference
pipeline can be executed on integer-only hardware.
To deal with the high dy... |
I found this paper fairly hard to understand, so I will write a summary that’s fairly different from the author’s presentation that would have been easier for me to understand, and the authors can tell me if my summary is inaccurate.
There have been several agents that play multiplayer online battle arena (MOBA) games... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
I found this paper fairly hard to understand, so I will write a summary that’s fairly different from the author’s presentation that would have been easier for me to understand, and the authors can tell me if my summary is inaccurate.
There have been several agents that play multiplayer online battle arena (MOB... |
This paper proposed an implicit sampler by minimizing the Fisher divergence and then propose to improve it by using the annealing technique. Specifically, the author proposed an alternative objective such that minimizing it is equivalent to minimizing the Fisher divergence. Later, the author tried to solve the distant ... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper proposed an implicit sampler by minimizing the Fisher divergence and then propose to improve it by using the annealing technique. Specifically, the author proposed an alternative objective such that minimizing it is equivalent to minimizing the Fisher divergence. Later, the author tried to solve the ... |
The authors propose a method to quantify and estimate how various components of graph neural network models will be affected by node/edge removals in the training set. In particular, the authors derive influence functions for the Simple Graph Convolution (SGC) architecture. The authors provide rigorous theoretical resu... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The authors propose a method to quantify and estimate how various components of graph neural network models will be affected by node/edge removals in the training set. In particular, the authors derive influence functions for the Simple Graph Convolution (SGC) architecture. The authors provide rigorous theoreti... |
This paper presents a Variational Classification (VC), which generalizes the softmax classifier, and the authors claim that it mirrors the relationship between the variational auto-encoder and the deterministic auto-encoder. The main contribution is to include the latent variable to softmax classifier and then VC objec... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper presents a Variational Classification (VC), which generalizes the softmax classifier, and the authors claim that it mirrors the relationship between the variational auto-encoder and the deterministic auto-encoder. The main contribution is to include the latent variable to softmax classifier and then ... |
The paper proposes a novel text representation named Polynomial Band. Cross-scale attention is used to enhance the feature of text regions. The fitting loss function is optimized to adapt to scene text detection task. The results on two benchmarks show the effectiveness.
Strength
1. The paper proposes Polynomial Band, ... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a novel text representation named Polynomial Band. Cross-scale attention is used to enhance the feature of text regions. The fitting loss function is optimized to adapt to scene text detection task. The results on two benchmarks show the effectiveness.
Strength
1. The paper proposes Polynomia... |
The paper addresses the video prediction problem where the model needs to predict the future frames (or tokens) conditioned on the first frame (context). This paper specifically focuses on addressing the high memory cost and inference efficiency. The method builds upon VQGAN to encode frames into quantized tokens and l... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The paper addresses the video prediction problem where the model needs to predict the future frames (or tokens) conditioned on the first frame (context). This paper specifically focuses on addressing the high memory cost and inference efficiency. The method builds upon VQGAN to encode frames into quantized toke... |
The authors propose MixQuant, a simple mixed precision search algorithm. MixQaunt evaluates the quantization error per layer (thus assumes independence) and compares the relative increase of lower bit-widths wrt the 8 bit error. They combine their search algorithm with BRECQ to improve the overall quantization accuracy... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The authors propose MixQuant, a simple mixed precision search algorithm. MixQaunt evaluates the quantization error per layer (thus assumes independence) and compares the relative increase of lower bit-widths wrt the 8 bit error. They combine their search algorithm with BRECQ to improve the overall quantization ... |
Authors propose a primal heuristic for solving MIP, following much of Neural Diving strategy from Nair et al (2021). Instead of fixing predicted variables from ML model, authors propose to solve a modified MIP problem which constrains the solution space to be the L1 ball around the predicted solution. Authors demonstra... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
Authors propose a primal heuristic for solving MIP, following much of Neural Diving strategy from Nair et al (2021). Instead of fixing predicted variables from ML model, authors propose to solve a modified MIP problem which constrains the solution space to be the L1 ball around the predicted solution. Authors d... |
This paper proposed a secure querying scheme for heterogeneous federated learning (HFL). HFL is a setting where clients in collaborative learning will train a model with different model architectures. Hence the global model cannot be directly aggregated/averaged from local client models. The proposed GuardHFL can query... | 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 secure querying scheme for heterogeneous federated learning (HFL). HFL is a setting where clients in collaborative learning will train a model with different model architectures. Hence the global model cannot be directly aggregated/averaged from local client models. The proposed GuardHFL c... |
This paper proposed using value bootstrapping to augment the return data for learning decision transformers or RvS (RL via supervised learning) in general. This approach, which performs explicit bootstrapping / stitching, allows mapping optimal actions to optimal return-to-gos that do not appear in the actual experienc... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposed using value bootstrapping to augment the return data for learning decision transformers or RvS (RL via supervised learning) in general. This approach, which performs explicit bootstrapping / stitching, allows mapping optimal actions to optimal return-to-gos that do not appear in the actual e... |
1. The authors propose to study a problem that they call combination shift: a simplified version of domain generalization in which all environments and labels are available at test time, but not necessarily all combinations.
2. Higgins et al [2018] define disentanglement based on equivariance wrt to the action of a pro... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
1. The authors propose to study a problem that they call combination shift: a simplified version of domain generalization in which all environments and labels are available at test time, but not necessarily all combinations.
2. Higgins et al [2018] define disentanglement based on equivariance wrt to the action ... |
This manuscript did two things. One is to propose a new dataset that collects agent trajectories from multiple independent agents with different preferences on the objectives. This dataset contains trajectories from both well-trained agents and semi-trained agents. HalfCheetah is a typical example where actions that ar... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This manuscript did two things. One is to propose a new dataset that collects agent trajectories from multiple independent agents with different preferences on the objectives. This dataset contains trajectories from both well-trained agents and semi-trained agents. HalfCheetah is a typical example where actions... |
The authors offer a method of estimating the grouping loss, which fills in some shortcomings of common measures of network calibration. They then use their new metric to measure the grouping loss of a variety of different architectures and settings.
## Strengths
- The method is rigorously defined
- The results are in... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The authors offer a method of estimating the grouping loss, which fills in some shortcomings of common measures of network calibration. They then use their new metric to measure the grouping loss of a variety of different architectures and settings.
## Strengths
- The method is rigorously defined
- The result... |
This paper proposes a method called V-IP (Variational Information Pursuit) that does a multi-step prediction to improve interpretability instead of doing a one-pass prediction like other neural nets do. In each step, only a small sets of features (i.e. "query set" called in the paper) are revealed and the goal is to ma... | 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 a method called V-IP (Variational Information Pursuit) that does a multi-step prediction to improve interpretability instead of doing a one-pass prediction like other neural nets do. In each step, only a small sets of features (i.e. "query set" called in the paper) are revealed and the goal ... |
AR models are very common in many domains, especially language models. Maximum likelihood is the widely used approach for training these models. However, ML training of AR models causes some other issues such as exposure bias. Techniques such as scheduled sampling have been introduced before to address these issues.
T... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
AR models are very common in many domains, especially language models. Maximum likelihood is the widely used approach for training these models. However, ML training of AR models causes some other issues such as exposure bias. Techniques such as scheduled sampling have been introduced before to address these is... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.