review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
The authors propose a new framework called GIANT (for Geometric InformAtioN boTtleneck) to provide results’ interpretations/explanations of a collaborative -filtering-based recommendation system.
This framework requires first the computation of a user-item interaction graph.
Each node is represented by LightGCN enco... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a new framework called GIANT (for Geometric InformAtioN boTtleneck) to provide results’ interpretations/explanations of a collaborative -filtering-based recommendation system.
This framework requires first the computation of a user-item interaction graph.
Each node is represented by Light... |
This paper proposes a method for Distillation-aware Network Architecture Search (DaNAS). The main component of the method is a “distillation-aware meta accuracy prediction model” which maps a (student architecture, dataset, teacher network) tuple to a prediction of the accuracy of the student on the particular dataset ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a method for Distillation-aware Network Architecture Search (DaNAS). The main component of the method is a “distillation-aware meta accuracy prediction model” which maps a (student architecture, dataset, teacher network) tuple to a prediction of the accuracy of the student on the particular ... |
This paper offers a new perspective to solve the behavior homogeneity issue in current joint policy optimization approaches. The paper points out the connections between policy gradient in multi-agent joint policy optimization and mutual information (MI) maximization followed by the drawbacks of MI maximization. The pa... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper offers a new perspective to solve the behavior homogeneity issue in current joint policy optimization approaches. The paper points out the connections between policy gradient in multi-agent joint policy optimization and mutual information (MI) maximization followed by the drawbacks of MI maximization... |
This work proposes a new attention mechanism for graph transformers aiming at handling graphs with large number of nodes.
More precisely, the attention mechanism is three-fold:
- Local attention, applied to 1-hop neighbors of a given node.
- Global attention via virtual nodes.
- Sparse attention based on an expander g... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposes a new attention mechanism for graph transformers aiming at handling graphs with large number of nodes.
More precisely, the attention mechanism is three-fold:
- Local attention, applied to 1-hop neighbors of a given node.
- Global attention via virtual nodes.
- Sparse attention based on an ex... |
In this paper, authors proposed a multivariate time-series forecasting model (SpectraNet) that unified the forecasting and interpolation problem. Specifically, the model first infers the optimal latent vector on the reference window by minimizing the reconstruction error, then the model generates the full predictions v... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, authors proposed a multivariate time-series forecasting model (SpectraNet) that unified the forecasting and interpolation problem. Specifically, the model first infers the optimal latent vector on the reference window by minimizing the reconstruction error, then the model generates the full predi... |
This paper is concerned with learning causal relationship between latent variables of a deep generative model, used to model high-dimensional complex data such as images. This is an important and difficult topic in causal inference. The paper starts with background about VAEs, SEMs and then proceeds to describing their... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper is concerned with learning causal relationship between latent variables of a deep generative model, used to model high-dimensional complex data such as images. This is an important and difficult topic in causal inference. The paper starts with background about VAEs, SEMs and then proceeds to describi... |
The paper proposes a method to identify anomalous subsequences in a long time series using a graph neural network. The core idea is to use two different graph representations for fixed length subsequences from the time series, one which captures the semantic distance (based on different distance measures) and the secon... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a method to identify anomalous subsequences in a long time series using a graph neural network. The core idea is to use two different graph representations for fixed length subsequences from the time series, one which captures the semantic distance (based on different distance measures) and t... |
This paper proposes a method for training in MARL setting under hybrid communication setting (decentralized vs fully centralized) by using a autoregressive model for predicting next joint observations. Experiment are provided on some toy domains with 2 decentralized value based baselines.
Strength
1. The paper is well ... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a method for training in MARL setting under hybrid communication setting (decentralized vs fully centralized) by using a autoregressive model for predicting next joint observations. Experiment are provided on some toy domains with 2 decentralized value based baselines.
Strength
1. The paper ... |
The authors test a neural network (NN) architecture based on vision transformers on the brain-score competition. When using adversarial training (gradient attack) and rotated data augmentation, the NN reaches state of the art performance for area V4. In contrast with other best performing models, their model is getting... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors test a neural network (NN) architecture based on vision transformers on the brain-score competition. When using adversarial training (gradient attack) and rotated data augmentation, the NN reaches state of the art performance for area V4. In contrast with other best performing models, their model is... |
The paper studies the problem of symbolic regression given the graph structure. It generalizes a previous approach (Cranmer et al. 2020) by additionally learning the formula skeleton. The method is seperated into two steps: searching the Pareto-optimal message passing flows and component-wise symbolic regression. Comp... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper studies the problem of symbolic regression given the graph structure. It generalizes a previous approach (Cranmer et al. 2020) by additionally learning the formula skeleton. The method is seperated into two steps: searching the Pareto-optimal message passing flows and component-wise symbolic regressi... |
This paper proposes an approach for continual learning using a subspace of policies (collection of ‘anchor’ parameters). Given a new task, the control policy either uses a linear interpolation of existing anchor parameters, or introduces a new anchor, and then linearly interpolates in the full set. This causes the numb... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes an approach for continual learning using a subspace of policies (collection of ‘anchor’ parameters). Given a new task, the control policy either uses a linear interpolation of existing anchor parameters, or introduces a new anchor, and then linearly interpolates in the full set. This causes ... |
In this paper, the authors propose a text-only adaptation method for RNN-Ts, allowing to make these models better for a new domain than the one represented in the training data. While existing methods require a change in the RNN-T architecture, a full retraining, or introduce some latency, the proposed approach allows ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the authors propose a text-only adaptation method for RNN-Ts, allowing to make these models better for a new domain than the one represented in the training data. While existing methods require a change in the RNN-T architecture, a full retraining, or introduce some latency, the proposed approach... |
This paper presents the multi-label knowledge distillation to address the issue of multiple semantic labels in multi-label learning scenarios. MKD and LED are introduced for network learning. Experiments on several datasets demonstrate its superiority.
Strength:
(1) The organization is good, which makes the paper easy... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents the multi-label knowledge distillation to address the issue of multiple semantic labels in multi-label learning scenarios. MKD and LED are introduced for network learning. Experiments on several datasets demonstrate its superiority.
Strength:
(1) The organization is good, which makes the pa... |
This paper leverages contraction theory to improve the robustness of neural ordinary differential equation (ODE). Directly obtaining contraction property needs to regularize the Jacobian matrix, which is computationally heavy. The authors theoretically prove that penalizing the weight matrix can be a contractivity-prom... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper leverages contraction theory to improve the robustness of neural ordinary differential equation (ODE). Directly obtaining contraction property needs to regularize the Jacobian matrix, which is computationally heavy. The authors theoretically prove that penalizing the weight matrix can be a contractiv... |
The paper proposes LS4 which is a generative model for sequences inspired by the deep state space model (SSM) S4. LS4 performs latent space evolution following a state space ODE and is trained via sequence VAE objectives. State-of-the-art results on selected datasets for continuous-time latent generative models are rep... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
The paper proposes LS4 which is a generative model for sequences inspired by the deep state space model (SSM) S4. LS4 performs latent space evolution following a state space ODE and is trained via sequence VAE objectives. State-of-the-art results on selected datasets for continuous-time latent generative models... |
This submission studies the optimal activation function in a two-layer random features model that minimizes a weighted sum of the test error and the model sensitivity measured by a Sobolev norm. The analysis assumes the proportional limit and spherical data; in this setting both the test error and the sensitivity have ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This submission studies the optimal activation function in a two-layer random features model that minimizes a weighted sum of the test error and the model sensitivity measured by a Sobolev norm. The analysis assumes the proportional limit and spherical data; in this setting both the test error and the sensitivi... |
The authors propose HeatFlow, a framework based on the heat equation on a Riemannian manifold, to interpreting a representation of givn deep learning model. Especially, they focus to analyze the multi-sale behvariors of the model, which cannot be revealed by using a naive gradient-based model interpretation approach. T... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors propose HeatFlow, a framework based on the heat equation on a Riemannian manifold, to interpreting a representation of givn deep learning model. Especially, they focus to analyze the multi-sale behvariors of the model, which cannot be revealed by using a naive gradient-based model interpretation app... |
The authors propose a Hamiltonian based feature learning in GNNs to overcome the oversmoothing and oversquashing phenomena that are evident in many existing GNNs.
A rigorous theoretical discussion is provided and several experiments are conducted.
Strengths:
- The authors provide a rigorous theoretical discussion.
-... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors propose a Hamiltonian based feature learning in GNNs to overcome the oversmoothing and oversquashing phenomena that are evident in many existing GNNs.
A rigorous theoretical discussion is provided and several experiments are conducted.
Strengths:
- The authors provide a rigorous theoretical discu... |
This paper proposes a simple method for visualizing the inductive bias of a supervised learning algorithm. The method involves meta-learning the labels for a dataset, such that the learning algorithm is able to easily generalize when trained on a subset of those labels. For example, if the inputs are 1D, the meta-learn... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes a simple method for visualizing the inductive bias of a supervised learning algorithm. The method involves meta-learning the labels for a dataset, such that the learning algorithm is able to easily generalize when trained on a subset of those labels. For example, if the inputs are 1D, the me... |
This paper presents a novel method for training image denoising techniques, called NERDS. The paper makes the observation that a noisy image can be downsampled to produce a pseudo-clean image, as the downsampling operation removes noise. Based on the noise statistics between the psudeo-clean image and the original im... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a novel method for training image denoising techniques, called NERDS. The paper makes the observation that a noisy image can be downsampled to produce a pseudo-clean image, as the downsampling operation removes noise. Based on the noise statistics between the psudeo-clean image and the ori... |
This paper extends the simple linear scalarization method in multi-objective optimization to the GflowNets setting. Two variants of GflowNets are proposed and extensive empirical experiments are provided, which verifies the effectiveness of them.
Strengths:
+ The idea of extending gflow nets to the multi-objective sett... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper extends the simple linear scalarization method in multi-objective optimization to the GflowNets setting. Two variants of GflowNets are proposed and extensive empirical experiments are provided, which verifies the effectiveness of them.
Strengths:
+ The idea of extending gflow nets to the multi-object... |
In this work, the authors study if a GPT-based language model, trained on a set of board game move data, is capable of learning a "world model" of the game. To investigate this:
1. Given the representation generated by the LM, representing a game state, the authors try to use a probe (either a linear one or a non-linea... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
In this work, the authors study if a GPT-based language model, trained on a set of board game move data, is capable of learning a "world model" of the game. To investigate this:
1. Given the representation generated by the LM, representing a game state, the authors try to use a probe (either a linear one or a n... |
This paper studies the convergence of off-policy algorithms, which learn a policy that optimizes the average-reward rate using data generated by some other non-controlled policy, for weakly-communicating MDPs, the most general set of MDPs that allow for a learning algorithm that can find under experience a policy of op... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the convergence of off-policy algorithms, which learn a policy that optimizes the average-reward rate using data generated by some other non-controlled policy, for weakly-communicating MDPs, the most general set of MDPs that allow for a learning algorithm that can find under experience a poli... |
The paper proposed an approach to improve zero-shot prediction results by making a language model predict the instruction (or measure likelihood) given the label and the input, rather than having the model predict the label instead. The paper demonstrates the effectiveness on multiple datasets within the BIG-Bench benc... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposed an approach to improve zero-shot prediction results by making a language model predict the instruction (or measure likelihood) given the label and the input, rather than having the model predict the label instead. The paper demonstrates the effectiveness on multiple datasets within the BIG-Be... |
This paper introduces a novel approach to the problem of human pose transfer, based on Vision Transformers (ViT). The authors propose a system with two encoding and one decoding Vision Transformers. The decoder consists of two branches, a warping branch that predicts the optical flow for warping the source image to the... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduces a novel approach to the problem of human pose transfer, based on Vision Transformers (ViT). The authors propose a system with two encoding and one decoding Vision Transformers. The decoder consists of two branches, a warping branch that predicts the optical flow for warping the source imag... |
The authors propose an extension to the existing SplitFed algorithm to address the high communication costs imposed by transmitting gradients during each backprop. They propose using product quantization to reduce the communication costs, followed by a Taylor series correct to correct the returned gradients from the se... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors propose an extension to the existing SplitFed algorithm to address the high communication costs imposed by transmitting gradients during each backprop. They propose using product quantization to reduce the communication costs, followed by a Taylor series correct to correct the returned gradients fro... |
The paper considers the problem of adversarial robustness. For this problem the paper investigates the usefulness of randomized ensemble classifiers (REC) where one classifier is randomly selected from the ensemble during inference. The main motivation behind considering RECs over deterministic ensembles is that the f... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper considers the problem of adversarial robustness. For this problem the paper investigates the usefulness of randomized ensemble classifiers (REC) where one classifier is randomly selected from the ensemble during inference. The main motivation behind considering RECs over deterministic ensembles is th... |
The paper presents a theoretical analysis of GANs and shows how GANs can learn hierarchically generated distributions. Here hierarchical generation means that the distribution comes from an unknown generator that follows a forward super-resolution structure where each successive layer gives a higher resolution image (s... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper presents a theoretical analysis of GANs and shows how GANs can learn hierarchically generated distributions. Here hierarchical generation means that the distribution comes from an unknown generator that follows a forward super-resolution structure where each successive layer gives a higher resolution ... |
The paper present two data augmentation strategies, Gandalf and LabelMix, that adds label features or the instance-label interpolations as training instances in extreme multi-label classification (XMC) training. Empirical results show that the proposed Gandalf method is able improve the performance of various XMC model... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper present two data augmentation strategies, Gandalf and LabelMix, that adds label features or the instance-label interpolations as training instances in extreme multi-label classification (XMC) training. Empirical results show that the proposed Gandalf method is able improve the performance of various X... |
This work studies the attention distribution of an attacked Transformer models. Based on their findings, the authors propose a trojan attention loss to enhance the attack efficiency. The authors conducted a comprehensive study on this method, showing the effectiveness of their approach.
Strength:
- The method is well m... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work studies the attention distribution of an attacked Transformer models. Based on their findings, the authors propose a trojan attention loss to enhance the attack efficiency. The authors conducted a comprehensive study on this method, showing the effectiveness of their approach.
Strength:
- The method i... |
This paper aims to improve the performance of personalized federated learning, and, for which, the authors propose two knowledge transfer schemes. In particular, the historical knowledge learned in the local clients is transferred from the hypernetwork, which stores the knowledge of previous local models, to the curren... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper aims to improve the performance of personalized federated learning, and, for which, the authors propose two knowledge transfer schemes. In particular, the historical knowledge learned in the local clients is transferred from the hypernetwork, which stores the knowledge of previous local models, to th... |
*Disclaimer: I am a vision person and an emergency reviewer*
This manuscript studies the problem of named entity recognition (similar to object detection) and relation extraction (similar to human object interaction detection) from French texts transcripted by off-the-shelf OCR models. The method firstly classifies te... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
*Disclaimer: I am a vision person and an emergency reviewer*
This manuscript studies the problem of named entity recognition (similar to object detection) and relation extraction (similar to human object interaction detection) from French texts transcripted by off-the-shelf OCR models. The method firstly class... |
This paper mainly focuses on how to improve the current learnable graph data augmentation strategy (GDA). The authors are motivated by the empirical observation that the previous GDA method only works in an early stage of the training process. Therefore, the authors design a new data augmentation strategy using reinfor... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper mainly focuses on how to improve the current learnable graph data augmentation strategy (GDA). The authors are motivated by the empirical observation that the previous GDA method only works in an early stage of the training process. Therefore, the authors design a new data augmentation strategy using... |
The paper proposes a collaborative symmetricity exploitation framework to solve the decoupling capacitor placement problem. The framework leverages action-permutation symmetry in offline imitation learning to augment data and to improve generalization capability. It can outperform iterative online search methods like G... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a collaborative symmetricity exploitation framework to solve the decoupling capacitor placement problem. The framework leverages action-permutation symmetry in offline imitation learning to augment data and to improve generalization capability. It can outperform iterative online search method... |
This paper studies the problem of how to compute $L_2$-heavy hitters in the sliding window model under differential privacy. The main contribution is giving the first differential private algorithm for this problem using sublinear working space.
Strength:
1. The $L_2$-heavy hitter problem in the sliding window model i... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper studies the problem of how to compute $L_2$-heavy hitters in the sliding window model under differential privacy. The main contribution is giving the first differential private algorithm for this problem using sublinear working space.
Strength:
1. The $L_2$-heavy hitter problem in the sliding window... |
The paper considers the problem of non-linear optimization with combinatorial constraints. A surrogate model based approach is proposed where the key idea is to learn a linear model of the underlying objective which is optimized via combinatorial solvers (for e.g. mixed integer program solvers SCIP) to generate a solut... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper considers the problem of non-linear optimization with combinatorial constraints. A surrogate model based approach is proposed where the key idea is to learn a linear model of the underlying objective which is optimized via combinatorial solvers (for e.g. mixed integer program solvers SCIP) to generate... |
This papertargets two issues in graph similarity computation including the node number awareness issue and the inference speed issue. The authors first analyze and show that the underlying reason of the first issue is the graph pooling. Then, the Different Attention is proposed to get the model aware of the node number... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This papertargets two issues in graph similarity computation including the node number awareness issue and the inference speed issue. The authors first analyze and show that the underlying reason of the first issue is the graph pooling. Then, the Different Attention is proposed to get the model aware of the nod... |
To alleviate the problem of multi-modality, the paper introduces a training objective that could consider multiple translations in the directed acyclic graph. The method is based on a fuzzy alignment between reference and the directed acyclic graph based on n-gram matching. Authors explore an alignment score to measur... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
To alleviate the problem of multi-modality, the paper introduces a training objective that could consider multiple translations in the directed acyclic graph. The method is based on a fuzzy alignment between reference and the directed acyclic graph based on n-gram matching. Authors explore an alignment score t... |
This paper provides a new result on certified robustness for graph matching. Its based on randomized smoothing (Cohen 2019), but the authors use a correlation matrix based on the graph information to construct a joint Gaussian distribution for smoothing (vs the standard single Gaussian distribution). The experimental... | 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 new result on certified robustness for graph matching. Its based on randomized smoothing (Cohen 2019), but the authors use a correlation matrix based on the graph information to construct a joint Gaussian distribution for smoothing (vs the standard single Gaussian distribution). The expe... |
The authors try to solve the problem of inverting a deterministic dynamics model when query access to the said model (being able to query the next state given the current state and the current action). The performance of the trained inverse model is evaluated on a distribution of reference state trajectories.
The core... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors try to solve the problem of inverting a deterministic dynamics model when query access to the said model (being able to query the next state given the current state and the current action). The performance of the trained inverse model is evaluated on a distribution of reference state trajectories.
... |
This paper presents a new reinforcement learning architecture for learning robot behaviors from pixels.
Strenghs:
- very important and broad problem to be tackling
- well-motivated, general approach
- detailed description of the algorithm, including hyperparameters
- excellent set of ablations
Weaknesses:
- Some of th... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents a new reinforcement learning architecture for learning robot behaviors from pixels.
Strenghs:
- very important and broad problem to be tackling
- well-motivated, general approach
- detailed description of the algorithm, including hyperparameters
- excellent set of ablations
Weaknesses:
- So... |
This paper investigates how sensitive neural networks are with respect to rare spurious correlations from three perspectives. First, studying how many training points with the spurious pattern would cause noticeable spurious correlations in the synthetic and real datasets. Second, studying how rare spurious correlation... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper investigates how sensitive neural networks are with respect to rare spurious correlations from three perspectives. First, studying how many training points with the spurious pattern would cause noticeable spurious correlations in the synthetic and real datasets. Second, studying how rare spurious cor... |
Based on the fact that adversarial perturbations are mainly on high-frequency part, this paper propose a regularization term to align the frequency response of natural and adversarial samples.
strength:
+ as far as I know, the proposed method that aligns frequency domain response is novel.
+ in numerical experiments,... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
Based on the fact that adversarial perturbations are mainly on high-frequency part, this paper propose a regularization term to align the frequency response of natural and adversarial samples.
strength:
+ as far as I know, the proposed method that aligns frequency domain response is novel.
+ in numerical expe... |
The paper applies predictive coding (PC)—an alternative to backpropagation—to graph neural networks, and attempts to show an improvement in robustness to adversarial attacks. The authors describe the predictive-coding scheme for graph convolutional networks (GCNs), and evaluate robustness against several forms of adver... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper applies predictive coding (PC)—an alternative to backpropagation—to graph neural networks, and attempts to show an improvement in robustness to adversarial attacks. The authors describe the predictive-coding scheme for graph convolutional networks (GCNs), and evaluate robustness against several forms ... |
This paper studies a single-input single-output (SISO) linear system that has a similar sequential input-output structure as the recurrent neural network (RNN). The ground-truth labels are generated by a teacher which is assumed to be a similar linear system. The authors model gradient descent (GD) with small step size... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies a single-input single-output (SISO) linear system that has a similar sequential input-output structure as the recurrent neural network (RNN). The ground-truth labels are generated by a teacher which is assumed to be a similar linear system. The authors model gradient descent (GD) with small s... |
This paper advances the theory of neural causal models and prove that they can recover the counterfactuals (L3 in Pearl's ladder of causality). The authors also provide new identification results and Algorithm 1 that uses supervised learning techniques to tell if the causal prior constraints are sufficient for identifi... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper advances the theory of neural causal models and prove that they can recover the counterfactuals (L3 in Pearl's ladder of causality). The authors also provide new identification results and Algorithm 1 that uses supervised learning techniques to tell if the causal prior constraints are sufficient for ... |
The authors focus on a challenging scenario called MNAR where the labeled and unlabeled data fall into different class distributions, which is different from traditional semi-supervised learning. In order to address the problem, the authors introduce a new technique called Pseudo-Rectifying Guidance (PRG). Using Markov... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The authors focus on a challenging scenario called MNAR where the labeled and unlabeled data fall into different class distributions, which is different from traditional semi-supervised learning. In order to address the problem, the authors introduce a new technique called Pseudo-Rectifying Guidance (PRG). Usin... |
This paper proposes a method for learning the “achievement” (i.e. subtask) structure of an environment from offline data, with sparse rewards indicating when achievements are attained. First, an achievement representation is learned using a contrastive-learning type procedure over the transitions where achievements wer... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a method for learning the “achievement” (i.e. subtask) structure of an environment from offline data, with sparse rewards indicating when achievements are attained. First, an achievement representation is learned using a contrastive-learning type procedure over the transitions where achievem... |
This paper proposes a method called “SWARM parallelism” to allow model-parallel training of large models on heterogeneous clusters (e.g., a mix of weak and strong GPUs, unreliable compute nodes or networks with uneven bandwidth).
Strength:
- The problem this paper is trying to address is important.
- Overall the pape... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a method called “SWARM parallelism” to allow model-parallel training of large models on heterogeneous clusters (e.g., a mix of weak and strong GPUs, unreliable compute nodes or networks with uneven bandwidth).
Strength:
- The problem this paper is trying to address is important.
- Overall ... |
1. This paper studies the effect of label error on a model’s group-based disparity metrics. Differences in terms of calibration error can be observed for the minority (smallest) group and the majority (largest) group.
2. They also propose an approach (influence function) to estimate how changing a single training input... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
1. This paper studies the effect of label error on a model’s group-based disparity metrics. Differences in terms of calibration error can be observed for the minority (smallest) group and the majority (largest) group.
2. They also propose an approach (influence function) to estimate how changing a single traini... |
This paper extends the recent certified robustness theory with a simple procedure that randomizes the input of the model within a transformation space. It abandons the $l_p$ distance change assumption, and does not need any restrictive assumptions or global Lipshitz. With the above theoretical analysis, this paper can... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper extends the recent certified robustness theory with a simple procedure that randomizes the input of the model within a transformation space. It abandons the $l_p$ distance change assumption, and does not need any restrictive assumptions or global Lipshitz. With the above theoretical analysis, this p... |
The paper proposes UniMax, a simple language data sampling strategy to provide a more uniform coverage of high-resource languages. This allows the training to be more resistant to overfitting and memorization on low-resource languages since they are repeated excessively on a standard upsampling strategy. The paper also... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes UniMax, a simple language data sampling strategy to provide a more uniform coverage of high-resource languages. This allows the training to be more resistant to overfitting and memorization on low-resource languages since they are repeated excessively on a standard upsampling strategy. The pa... |
This paper presents a new method to estimate the 3D flow and volumetric densities of a moving fluid from a monocular video. The proposed framework consists of two deep networks, one for predicting the volumetric densities from a single image, and one for predicting flow given the input images and predicted densities fr... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper presents a new method to estimate the 3D flow and volumetric densities of a moving fluid from a monocular video. The proposed framework consists of two deep networks, one for predicting the volumetric densities from a single image, and one for predicting flow given the input images and predicted dens... |
The authors consider the recently proposed setting of unsupervised continual learning (UCL), which I understand as generally studying the extent that unlabelled information can be exploited in a continual learning setting leveraging self-supervised learning. The authors build a new approach, AUDR, extending the recent ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors consider the recently proposed setting of unsupervised continual learning (UCL), which I understand as generally studying the extent that unlabelled information can be exploited in a continual learning setting leveraging self-supervised learning. The authors build a new approach, AUDR, extending the... |
### Summary
This research examines the effect of batch size on performance in multi-step reinforcement learning. In supervised-learning experiments, several researchers have found that the variance of neural networks during training has a significant effect on learnability and convergence, since the neura... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
### Summary
This research examines the effect of batch size on performance in multi-step reinforcement learning. In supervised-learning experiments, several researchers have found that the variance of neural networks during training has a significant effect on learnability and convergence, since t... |
This paper presents ULF (Unsupervised Labeling Function) which uses k-fold cross-validation to correct errors in labeling functions (LFs) by using cross-validation to identify confident sample estimates and update LF class label assignments. They explore 2 variants of ULF: a count-based and language model / DeepULF bas... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents ULF (Unsupervised Labeling Function) which uses k-fold cross-validation to correct errors in labeling functions (LFs) by using cross-validation to identify confident sample estimates and update LF class label assignments. They explore 2 variants of ULF: a count-based and language model / Dee... |
The paper proposes 3 methods for permuting the weights of a neural network $A$ to closely match a second neural network $B$. With this procedure, they show that often they achieve linear mode connectivity between the permuted model $A$ and model $B$. They further show that it is possible to merge independently trained ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes 3 methods for permuting the weights of a neural network $A$ to closely match a second neural network $B$. With this procedure, they show that often they achieve linear mode connectivity between the permuted model $A$ and model $B$. They further show that it is possible to merge independently ... |
The paper develops a method of adding gates to hidden units in order to protect their weights from modification during training on new tasks. The goal is to prevent catastrophic forgetting. The method, LXDG, extends a prior method XDG, by (among other things) removing the need for a task label.
(strengths)
The backgr... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The paper develops a method of adding gates to hidden units in order to protect their weights from modification during training on new tasks. The goal is to prevent catastrophic forgetting. The method, LXDG, extends a prior method XDG, by (among other things) removing the need for a task label.
(strengths)
Th... |
This paper studies clean-label availability attacks against unsupervised contrastive learning. The authors first observe that previous poisons, designed to compromise supervised learning, fail to harm contrastive learning algorithms, such as SimCLR and BYOL. Then, the authors propose a new attack method tailored to com... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper studies clean-label availability attacks against unsupervised contrastive learning. The authors first observe that previous poisons, designed to compromise supervised learning, fail to harm contrastive learning algorithms, such as SimCLR and BYOL. Then, the authors propose a new attack method tailore... |
This paper presents a benchmark that touches on an interesting and important topic, musculoskeletal dexterous hand manipulation. The authors further demonstrate that pre-training the policy on 14 manipulation tasks allows for easier fine-tuning in both in-domain and out-of-domain tasks.
Strength:
1. The proposed conta... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper presents a benchmark that touches on an interesting and important topic, musculoskeletal dexterous hand manipulation. The authors further demonstrate that pre-training the policy on 14 manipulation tasks allows for easier fine-tuning in both in-domain and out-of-domain tasks.
Strength:
1. The propos... |
This work focuses on molecular dynamics (MD) simulations. Previous methods use force/energy errors as the evaluation, but based on the authors' analysis, these metrics are not enough.
Specifically, in this paper, the authors first curate four MD datasets based on previous work and design specific evaluation metrics fo... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This work focuses on molecular dynamics (MD) simulations. Previous methods use force/energy errors as the evaluation, but based on the authors' analysis, these metrics are not enough.
Specifically, in this paper, the authors first curate four MD datasets based on previous work and design specific evaluation me... |
This study proposes an innovative strategy to attack reinforcement learning algorithms. Importantly, the attack requires minimal, information about the victim, in contrast to many other approaches that assume much more extensive inside information. Such black box approaches are necessarily less effective but ultimately... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This study proposes an innovative strategy to attack reinforcement learning algorithms. Importantly, the attack requires minimal, information about the victim, in contrast to many other approaches that assume much more extensive inside information. Such black box approaches are necessarily less effective but ul... |
This paper proposes Pseudoinverse-guided Diffusion Models ($\Pi$GDM), that use pretrained diffusion models (problem-agnostic) to solve inverse problems. The proposed $\Pi$GDM works with noisy, non-linear, or even non-differentiable measurements, and demonstrates promising results.
## strengths
- The paper is well-writt... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper proposes Pseudoinverse-guided Diffusion Models ($\Pi$GDM), that use pretrained diffusion models (problem-agnostic) to solve inverse problems. The proposed $\Pi$GDM works with noisy, non-linear, or even non-differentiable measurements, and demonstrates promising results.
## strengths
- The paper is we... |
This paper presents a scalable multi-model continual meta-learning algorithm. This method associates a cluster of similar tasks with a set of meta-knowledge components instead of one single component in previous approaches. If I understand correctly, “multi-modal” in the title is due to a set of components being used. ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents a scalable multi-model continual meta-learning algorithm. This method associates a cluster of similar tasks with a set of meta-knowledge components instead of one single component in previous approaches. If I understand correctly, “multi-modal” in the title is due to a set of components bein... |
This paper analyzes the generalization performance of a simple model — linear regression. Given that a prior work (Bartlett et. al. 2019) has shown a benign over-fitting of the final iterate of SGD for over-parameterized linear models, this work with milder conditions theoretically shows that good generalization can al... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper analyzes the generalization performance of a simple model — linear regression. Given that a prior work (Bartlett et. al. 2019) has shown a benign over-fitting of the final iterate of SGD for over-parameterized linear models, this work with milder conditions theoretically shows that good generalizatio... |
This work designs a more general IMLE objective by introducing an adaptive neighbourhood radii. The theoretical derivation and the proposed curriculum learning strategy are beautiful and impressive. Some experiments are also conducted to verify the effectiveness of the proposed method.
Strength:
1. The idea of introduc... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This work designs a more general IMLE objective by introducing an adaptive neighbourhood radii. The theoretical derivation and the proposed curriculum learning strategy are beautiful and impressive. Some experiments are also conducted to verify the effectiveness of the proposed method.
Strength:
1. The idea of ... |
This paper proposes a new attention mechanism that depends only on the relative positions of tokens and is independent of the features themselves. Since the attention coefficient depend only on the relative positions, the resulting attention matrix is Toeplitz. This results in a reduction in the number of parameters an... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a new attention mechanism that depends only on the relative positions of tokens and is independent of the features themselves. Since the attention coefficient depend only on the relative positions, the resulting attention matrix is Toeplitz. This results in a reduction in the number of param... |
This paper proposes a new method for inference of the intractable posterior for deep Gaussian processes (DGPs). The idea is to minimize the Stein discrepancy between the approximated posterior distribution which is demonstrated by a generator and the true posterior distribution. The algorithm utilizes a discriminator t... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a new method for inference of the intractable posterior for deep Gaussian processes (DGPs). The idea is to minimize the Stein discrepancy between the approximated posterior distribution which is demonstrated by a generator and the true posterior distribution. The algorithm utilizes a discrim... |
The paper tackles the problem of distraction robustness and out-of-distribution (OOD) generalization in model-free RL. To this end, the paper proposes to use domain randomization combined with some modification to the Q-learning objective. Specifically, the model takes two views, one clean and one noisy. The noisy view... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper tackles the problem of distraction robustness and out-of-distribution (OOD) generalization in model-free RL. To this end, the paper proposes to use domain randomization combined with some modification to the Q-learning objective. Specifically, the model takes two views, one clean and one noisy. The no... |
The paper proposes a value-based RL architecture that is aimed at achieving good generalization, in the settings of "contextualized" MDPs (where the agent is trained on one set of MDPs and tested on a different set sampled from the same distribution). This is achieved by encouraging a more explorative behavior during t... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes a value-based RL architecture that is aimed at achieving good generalization, in the settings of "contextualized" MDPs (where the agent is trained on one set of MDPs and tested on a different set sampled from the same distribution). This is achieved by encouraging a more explorative behavior ... |
The work presented in this paper targets multimodal representation learning, following a line of work that uses a variational autoencoder (VAE) as a basis that is extended to learn a latent representation from multiple modalities, that is observations of the same phenomenon through the lenses of a variety of input sour... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The work presented in this paper targets multimodal representation learning, following a line of work that uses a variational autoencoder (VAE) as a basis that is extended to learn a latent representation from multiple modalities, that is observations of the same phenomenon through the lenses of a variety of in... |
This paper is focused on instance segmentation. Baseline method applise Discrete Cosine Transform to improve the segmentation quality around the object boundary. This paper extends DCT-Mask from image-level to patch-level. Through Figure 4 we could see the quality improvement visually.
The paper introduces PatchDCT, wh... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper is focused on instance segmentation. Baseline method applise Discrete Cosine Transform to improve the segmentation quality around the object boundary. This paper extends DCT-Mask from image-level to patch-level. Through Figure 4 we could see the quality improvement visually.
The paper introduces Patc... |
This work is in the domain of text generation for long-form articles. The key contribution of this work is to provide additional metadata in the form of named entities to the generative model. Such entities are either provided in advance (oracle) or can also be derived using entity retrieval based on image associated w... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work is in the domain of text generation for long-form articles. The key contribution of this work is to provide additional metadata in the form of named entities to the generative model. Such entities are either provided in advance (oracle) or can also be derived using entity retrieval based on image asso... |
This paper tackles the problem of concept-based explanations for deep neural network OOD detectors. The authors build on Yeh et al. (2020)’s argument on concept completeness and propose to use two metrics: detection (concept) completeness and concept separability. Following such metrics, the authors then design an algo... | 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 tackles the problem of concept-based explanations for deep neural network OOD detectors. The authors build on Yeh et al. (2020)’s argument on concept completeness and propose to use two metrics: detection (concept) completeness and concept separability. Following such metrics, the authors then design... |
This paper introduces deep transformer Q-networks (DTQN) for learning in partially observable environments. The authors train DTQN with a double Q-learning objective using on-policy samples. Given a trajectory of observations till now, observations are encoded using a transformer decoder with a masked attention to gene... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper introduces deep transformer Q-networks (DTQN) for learning in partially observable environments. The authors train DTQN with a double Q-learning objective using on-policy samples. Given a trajectory of observations till now, observations are encoded using a transformer decoder with a masked attention... |
This paper proposes a new method for Bayesian meta learning by learning priors in the function space. The key insight that distinguishes this work from prior literature is that functional inference techniques only make use of the prior score function and do not require a normalized prior density; therefore, instead of ... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a new method for Bayesian meta learning by learning priors in the function space. The key insight that distinguishes this work from prior literature is that functional inference techniques only make use of the prior score function and do not require a normalized prior density; therefore, ins... |
The paper deals with the relevant multi-task RL problem. The authors propose a novel method that follows the "minimum description length" principle, to learn a common structure between the tasks. The aim of the proposed approach is to improve the generalization. The authors provide sample complexity results in the fini... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper deals with the relevant multi-task RL problem. The authors propose a novel method that follows the "minimum description length" principle, to learn a common structure between the tasks. The aim of the proposed approach is to improve the generalization. The authors provide sample complexity results in ... |
This paper (S-NeRF) aims to train a NeRF using images from driving datasets (e.g., nuScenes and Waymo) with a limited number of views (2-6 images). The improvement over Urban-NeRF is that it densifies sparse LiDAR depth for depth supervision, through a reprojection-based confidence map approach. The main claimed advant... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper (S-NeRF) aims to train a NeRF using images from driving datasets (e.g., nuScenes and Waymo) with a limited number of views (2-6 images). The improvement over Urban-NeRF is that it densifies sparse LiDAR depth for depth supervision, through a reprojection-based confidence map approach. The main claime... |
This work proposes a framework for applying off-the-shelf post-training quantization (PTQ) methods for learned image compression (LIC) task. The authors show that the optimum quantization scaling factor may not be discovered by only minimizing the MSE loss for LIC task and provide theoretical insights on this. To allev... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes a framework for applying off-the-shelf post-training quantization (PTQ) methods for learned image compression (LIC) task. The authors show that the optimum quantization scaling factor may not be discovered by only minimizing the MSE loss for LIC task and provide theoretical insights on this. ... |
This paper studies graph representation learning and propose MeGraph that could learn multi-scale graph representation. MeGraph enables information exchange across multi-scale graphs, which distinguishes MeGraph from many recent hierarchical graph neural networks.
To achieve this goal, MeGraph first uses graph pooling ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies graph representation learning and propose MeGraph that could learn multi-scale graph representation. MeGraph enables information exchange across multi-scale graphs, which distinguishes MeGraph from many recent hierarchical graph neural networks.
To achieve this goal, MeGraph first uses graph ... |
The paper "Lower Bounds on the Depth of Integral ReLU Neural Networks via Lattice Polytopes" derives a theoretical result and proves that log_2(n) hidden layers are necessary to compute the maximum of n input numbers. This is achieved by building on previous theoretical work that links neural networks to tropical geome... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper "Lower Bounds on the Depth of Integral ReLU Neural Networks via Lattice Polytopes" derives a theoretical result and proves that log_2(n) hidden layers are necessary to compute the maximum of n input numbers. This is achieved by building on previous theoretical work that links neural networks to tropic... |
The paper uses score matching SDEs for learning inverse problems with physical simulators. Different from common diffusion models, the forward SDE is changed: the drift is given by a physical model and the diffusion term does not grow to infinity as time increases. Training and inference algorithms based on SDEs and pr... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper uses score matching SDEs for learning inverse problems with physical simulators. Different from common diffusion models, the forward SDE is changed: the drift is given by a physical model and the diffusion term does not grow to infinity as time increases. Training and inference algorithms based on SDE... |
This work is after finding proper scoring rules in survival analysis. They have a parameter vector w whose true specification underpins the proofs their provide for the discussed scoring rules being proper. They approximate the parameter vector w using an EM algorithm which can in turn be plugged into their scoring rul... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This work is after finding proper scoring rules in survival analysis. They have a parameter vector w whose true specification underpins the proofs their provide for the discussed scoring rules being proper. They approximate the parameter vector w using an EM algorithm which can in turn be plugged into their sco... |
The authors propose succinct data structures and a new formulate, Runtime-Accessible Sequence, to efficiently utilize new data structures. RAS is provided with two versions (element-wise and block-wise) with synergistic effects with pruning and quantization. The authors provide experimental results to support the claim... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors propose succinct data structures and a new formulate, Runtime-Accessible Sequence, to efficiently utilize new data structures. RAS is provided with two versions (element-wise and block-wise) with synergistic effects with pruning and quantization. The authors provide experimental results to support t... |
This paper propose a new graph sampling mechanism for reducing the computational complexity in full attention of graph transformer. The sampling mechanism takes structural and semantic similarity into consideration. Experimental results show the effectiveness of the proposed method. Additionally proposed positioanl enc... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper propose a new graph sampling mechanism for reducing the computational complexity in full attention of graph transformer. The sampling mechanism takes structural and semantic similarity into consideration. Experimental results show the effectiveness of the proposed method. Additionally proposed positi... |
This paper proposes a new self-supervised learning framework (SSL), which aims to transfer the supervised sparse learning to SSL and reduce the required computational resources during the pretraining stage. Specifically, this paper investigates the correlation between training sparsity and SSL, which embraces the benef... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a new self-supervised learning framework (SSL), which aims to transfer the supervised sparse learning to SSL and reduce the required computational resources during the pretraining stage. Specifically, this paper investigates the correlation between training sparsity and SSL, which embraces t... |
The authors propose MBMF, essentially a feature selection procedure for a second step imputation method. With MBMF the authors aim to recover the m-graph that governs a system's missingness (in the tabular domain). Using the m-graph, the authors recover a variable of interest's markov blanket which are used as features... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The authors propose MBMF, essentially a feature selection procedure for a second step imputation method. With MBMF the authors aim to recover the m-graph that governs a system's missingness (in the tabular domain). Using the m-graph, the authors recover a variable of interest's markov blanket which are used as ... |
In this work, the authors propose an approximated natural-gradient descent for NN via the Legendre transformation.
The authors show that the proposed methods performs similarly as K-FAC and its variants in some medium scale problems (FACES and MNIST).
Strength:
* Re-formulation of natural-gradient descent via the Leg... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this work, the authors propose an approximated natural-gradient descent for NN via the Legendre transformation.
The authors show that the proposed methods performs similarly as K-FAC and its variants in some medium scale problems (FACES and MNIST).
Strength:
* Re-formulation of natural-gradient descent via... |
The paper considers the federated learning problem and proposes to
minimize the difference between two log determinant functions instead of
the cross-entropic loss.
# Weakness
1. The work is not solid. My
judgment is based on the following:
- The background section is unnecessarily long. For example, the
r... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The paper considers the federated learning problem and proposes to
minimize the difference between two log determinant functions instead of
the cross-entropic loss.
# Weakness
1. The work is not solid. My
judgment is based on the following:
- The background section is unnecessarily long. For example, th... |
This paper considers the problem of training neural networks which operate on sets. The main focus is on the mini-batch consistency (MBC) property, which roughly states the following: Let $X$ be a set, and $(X_1, \dots, X_n)$. Given a map $f$ from sets to vectors and an aggregation function $g$, the pair $(f,g)$ are ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper considers the problem of training neural networks which operate on sets. The main focus is on the mini-batch consistency (MBC) property, which roughly states the following: Let $X$ be a set, and $(X_1, \dots, X_n)$. Given a map $f$ from sets to vectors and an aggregation function $g$, the pair $(f,... |
This paper empirically looks at how changing the neural network architecture can impact continual learning performance. For example, if changing width/depth of the network, or adding BatchNorm layers, or adding skip connections, changes how well the network can learn continually. Many experiments are performed on Split... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper empirically looks at how changing the neural network architecture can impact continual learning performance. For example, if changing width/depth of the network, or adding BatchNorm layers, or adding skip connections, changes how well the network can learn continually. Many experiments are performed ... |
This work presents a framework for joint learning scene decomposition and NeRF, targeting NeRF for large-scale scene. It is well-recognized that decomposing NeRF is needed for large-scale scene, i.e., a standalone NeRF model for a specific region, because of the model can be extremely large without decomposition and th... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work presents a framework for joint learning scene decomposition and NeRF, targeting NeRF for large-scale scene. It is well-recognized that decomposing NeRF is needed for large-scale scene, i.e., a standalone NeRF model for a specific region, because of the model can be extremely large without decompositio... |
The paper introduces the _optimal commitment problem_, where an agent must decide, at each time step $t$, whether to continue running the current experiment (i.e., _commit_ to the current experiment), or terminate the current experiment, switching to a more promising experiment (or to no experiment at all, if one does ... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The paper introduces the _optimal commitment problem_, where an agent must decide, at each time step $t$, whether to continue running the current experiment (i.e., _commit_ to the current experiment), or terminate the current experiment, switching to a more promising experiment (or to no experiment at all, if o... |
This paper is concerned with how to pool irregularly spaced meshes for multi-scale estimation in physical simulations. The idea proposed herein is to use so-called bi-strides which "pools nodes by striding every other BFS frontier". The importance lies in the fact that many meshes of important problem exhibit complex t... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper is concerned with how to pool irregularly spaced meshes for multi-scale estimation in physical simulations. The idea proposed herein is to use so-called bi-strides which "pools nodes by striding every other BFS frontier". The importance lies in the fact that many meshes of important problem exhibit c... |
The paper studies a version of correlation clustering where the learner has access to a weak predictor of node similarity in addition to an accurate stronger predictor. The weak model is assumed to have negligible query cost, and can be used in addition to the stronger model. For a free parameter $\epsilon$, he paper p... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper studies a version of correlation clustering where the learner has access to a weak predictor of node similarity in addition to an accurate stronger predictor. The weak model is assumed to have negligible query cost, and can be used in addition to the stronger model. For a free parameter $\epsilon$, he... |
In this study the authors investigate post-training compression of CNN-based image classification neural networks. One method to compress a CNN-based NN is tensor decomposition. Their goal is to learn how much the approximation error of tensor decomposition is predictive of the performance of the compressed CNN-based N... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this study the authors investigate post-training compression of CNN-based image classification neural networks. One method to compress a CNN-based NN is tensor decomposition. Their goal is to learn how much the approximation error of tensor decomposition is predictive of the performance of the compressed CNN... |
This paper studies the behavior of a single ReLU neuron in the gradient descent training process. The main results include that first, a new class of distributions (called O(1)-regular marginals) is proposed to characterize the regularity of the input space distributions. This class includes multivariate Gaussian distr... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the behavior of a single ReLU neuron in the gradient descent training process. The main results include that first, a new class of distributions (called O(1)-regular marginals) is proposed to characterize the regularity of the input space distributions. This class includes multivariate Gaussi... |
The paper aims at the union of manifolds assumption in deep learning, which states that data lies on a disjoint union of manifolds of varying intrinsic dimensions. The authors empirically verify this hypothesis on commonly-used image datasets.
Strength:
1. The union of manifolds assumption is useful in deep learning.
2... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper aims at the union of manifolds assumption in deep learning, which states that data lies on a disjoint union of manifolds of varying intrinsic dimensions. The authors empirically verify this hypothesis on commonly-used image datasets.
Strength:
1. The union of manifolds assumption is useful in deep lea... |
The paper presents a self-supervised speech enhancement method for a scenario when both a microphone recording and a signal from an accelerometer sensor is available. Activity of the latter signal is well correlated with the speech signal activity, while it's also not corrupted by the acoustic noise from the environmen... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
The paper presents a self-supervised speech enhancement method for a scenario when both a microphone recording and a signal from an accelerometer sensor is available. Activity of the latter signal is well correlated with the speech signal activity, while it's also not corrupted by the acoustic noise from the en... |
This paper proposes an ML pipeline to evolve agent topologies for open ended learning by combining previous works EPOET and NEAT. Experiments are done on the 2D bipedal environment.
1. The paper lacks sufficient insights/results/illustrations to understand how the agent topologies are changing, most of the methodology ... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes an ML pipeline to evolve agent topologies for open ended learning by combining previous works EPOET and NEAT. Experiments are done on the 2D bipedal environment.
1. The paper lacks sufficient insights/results/illustrations to understand how the agent topologies are changing, most of the meth... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.