review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
The paper proposed a reinforcement learning based approach for automatically predicting graph augmentations for a graph neural network (GNN) classification problem. The authors argue that label invariance (data augmentations that do not affect labels) is an important, unsolved problem for GNN, and propose a technique t... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposed a reinforcement learning based approach for automatically predicting graph augmentations for a graph neural network (GNN) classification problem. The authors argue that label invariance (data augmentations that do not affect labels) is an important, unsolved problem for GNN, and propose a tec... |
In this work, the authors present DeepDFA, a dataflow analysis-guided graph learning framework and embedding that use program semantic features for vulnerability detection. In experiment, they show DeepDFA ranked first in recall, first in generalizing over unseen projects, and second in F1 among all the state-of-the-ar... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this work, the authors present DeepDFA, a dataflow analysis-guided graph learning framework and embedding that use program semantic features for vulnerability detection. In experiment, they show DeepDFA ranked first in recall, first in generalizing over unseen projects, and second in F1 among all the state-o... |
Goal: The paper addresses the brittleness challenge with semantic parsing based QA systems. For some questions models don’t produce bad logical forms that are not executable, which means they get no answer.
Method: Use a direct answer generator. Instead of using a separate answer generator, you can use the same “read... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Goal: The paper addresses the brittleness challenge with semantic parsing based QA systems. For some questions models don’t produce bad logical forms that are not executable, which means they get no answer.
Method: Use a direct answer generator. Instead of using a separate answer generator, you can use the sa... |
This paper is a combination of two previous works. One is about logical (boolean) compositional mechanisms for multitask learning, and the other is about logical formalisms for temporal skills, both in the context of typical reinforcement learning problems. The idea is to lift the first (composing behaviors in a struct... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper is a combination of two previous works. One is about logical (boolean) compositional mechanisms for multitask learning, and the other is about logical formalisms for temporal skills, both in the context of typical reinforcement learning problems. The idea is to lift the first (composing behaviors in ... |
The authors consider the problem of crafting adversarial attacks that are transferable across different deep neural networks architectures (DNNs). In order to solve this problem the authors propose to craft the attacks on Bayesian neural networks (BNNs) trained with Gaussian posterior approximations. On a set of experi... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors consider the problem of crafting adversarial attacks that are transferable across different deep neural networks architectures (DNNs). In order to solve this problem the authors propose to craft the attacks on Bayesian neural networks (BNNs) trained with Gaussian posterior approximations. On a set o... |
The paper presents a loss function called ASLoss as an adjunct for classification loss to address the confusion issue under data insufficiency and class imbalance cases. The ASLoss can be decomposed into inner aggregation, outer separation and boundary constraint, which contraints the representation to have some invari... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper presents a loss function called ASLoss as an adjunct for classification loss to address the confusion issue under data insufficiency and class imbalance cases. The ASLoss can be decomposed into inner aggregation, outer separation and boundary constraint, which contraints the representation to have som... |
This paper explores the computational features that result from a neural network with weights of fixed polarity. The authors demonstrate through proofs in constrained settings, and experiments in both constrained and more general settings, that networks with fixed and appropriately initialized weight polarity learn fas... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper explores the computational features that result from a neural network with weights of fixed polarity. The authors demonstrate through proofs in constrained settings, and experiments in both constrained and more general settings, that networks with fixed and appropriately initialized weight polarity l... |
This paper performs an analysis of distributional reinforcement learning using a regularization perspective. Specifically, under the assumption of a particular decomposition of the value distribution, they argue that distributional RL provides a risk-sensitive entropy regularization when approximated via the neural fit... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper performs an analysis of distributional reinforcement learning using a regularization perspective. Specifically, under the assumption of a particular decomposition of the value distribution, they argue that distributional RL provides a risk-sensitive entropy regularization when approximated via the ne... |
This work studies front-door and back-door adjustment using tensor product of learned kernels, generalizing Singh et al (2021) which employed fixed-form kernels. The authors establish consistency for the estimators, and demonstrate improved performance on IHDP and a synthetic high-dimensional benchmark.
---
Post-reb... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This work studies front-door and back-door adjustment using tensor product of learned kernels, generalizing Singh et al (2021) which employed fixed-form kernels. The authors establish consistency for the estimators, and demonstrate improved performance on IHDP and a synthetic high-dimensional benchmark.
---
... |
This paper proposed a method that extends weighted least squares to deep neural networks for hyperparameter selection in unsupervised domain adaptation. In theory, the authors show that the target error of the proposed algorithm is asymptotically not worse than twice the error of the unknown optimal aggregation. Numeri... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed a method that extends weighted least squares to deep neural networks for hyperparameter selection in unsupervised domain adaptation. In theory, the authors show that the target error of the proposed algorithm is asymptotically not worse than twice the error of the unknown optimal aggregation... |
This work studies Inter-silo record-level DP in cross-silo federated learning, where all communications to the server should satisfy item-level DP. The authors investigate convex optimization and design algorithms that achieve near-optimal performance for both convex and strongly convex cases. With amplification by shu... | 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 work studies Inter-silo record-level DP in cross-silo federated learning, where all communications to the server should satisfy item-level DP. The authors investigate convex optimization and design algorithms that achieve near-optimal performance for both convex and strongly convex cases. With amplificatio... |
This paper proposes a supervised learning framework that utilizes the philosophy of counterfactual learning to enhance the node classification performance. Specifically, this paper utilizes a counterfactual transformation module to construct sub-ego-graphs that are positively correlated to the label information, and ex... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a supervised learning framework that utilizes the philosophy of counterfactual learning to enhance the node classification performance. Specifically, this paper utilizes a counterfactual transformation module to construct sub-ego-graphs that are positively correlated to the label information... |
The authors argue that our ability to optimize a model trained with DP-SGD is governed by a quantity they call TAN (the total amount of noise), specifically $\Sigma^2 = 2 \sigma^2 N^2/(B^2 S)$, where $\Sigma$ is the TAN, $\sigma$ is the scale of the added Gaussian noise, $N$ is the dataset size, $B$ is the batch size a... | 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 argue that our ability to optimize a model trained with DP-SGD is governed by a quantity they call TAN (the total amount of noise), specifically $\Sigma^2 = 2 \sigma^2 N^2/(B^2 S)$, where $\Sigma$ is the TAN, $\sigma$ is the scale of the added Gaussian noise, $N$ is the dataset size, $B$ is the batc... |
1. This paper presents an information-theoretic view of SSL methods showing that current methods happen to maximize meaningful information-theoretic quantities.
2. The paper also presents a generalization bound on downstream tasks based on the above information-theoretical quantities.
3. Crucial to these results is the... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
1. This paper presents an information-theoretic view of SSL methods showing that current methods happen to maximize meaningful information-theoretic quantities.
2. The paper also presents a generalization bound on downstream tasks based on the above information-theoretical quantities.
3. Crucial to these result... |
The paper proposes to use implicit differentiation in value iteration networks, allowing for a deeper network structure/more iterations of planning. This is because the implicit gradient has a constant complexity with respect to the number of planning iterations.
**strengths**
* The paper identifies and fixes an issue... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes to use implicit differentiation in value iteration networks, allowing for a deeper network structure/more iterations of planning. This is because the implicit gradient has a constant complexity with respect to the number of planning iterations.
**strengths**
* The paper identifies and fixes ... |
This paper focused on exploring and extending linear-relaxation-based sample-wise robustness certification to evaluate the robustness of trained neural networks against Universal Perturbations (UPs). The authors pointed out an interesting intersection between universal adversarial perturbations (UAPs) and backdoor trig... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper focused on exploring and extending linear-relaxation-based sample-wise robustness certification to evaluate the robustness of trained neural networks against Universal Perturbations (UPs). The authors pointed out an interesting intersection between universal adversarial perturbations (UAPs) and backd... |
This paper proposes a novel out-of-distribution (OOD) detection algorithm with “store-and-compare” fashion: store the average pattern in the penultimate layer for each class, and during test, compute the inner product for the test blob with the stored pattern for the predicted class, and this inner product will be thre... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a novel out-of-distribution (OOD) detection algorithm with “store-and-compare” fashion: store the average pattern in the penultimate layer for each class, and during test, compute the inner product for the test blob with the stored pattern for the predicted class, and this inner product will... |
The authors propose a new bilinear knowledge graph embedding model named UniBi by learning a unique identity element in a group. The main idea is to constrain the norm and spectral radius of entity and relation embeddings, respectively. The authors provide thorough analyses to show that UniBi has theoretical superiorit... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose a new bilinear knowledge graph embedding model named UniBi by learning a unique identity element in a group. The main idea is to constrain the norm and spectral radius of entity and relation embeddings, respectively. The authors provide thorough analyses to show that UniBi has theoretical su... |
This paper proposes a new diffusion-based T2I models to improve the compositional skills. The linguistic structures are incorporated with the diffusion guidance process based on the controllable properties of manipulating cross-attention layers in diffusion-based T2I models. The model is built based on the stable diffu... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a new diffusion-based T2I models to improve the compositional skills. The linguistic structures are incorporated with the diffusion guidance process based on the controllable properties of manipulating cross-attention layers in diffusion-based T2I models. The model is built based on the stab... |
This paper proposes a novel model named BSTT for sleep staging. BSST integrates the transformer and Bayesian relation inference to simultaneously model spatial-temporal relations and effectively capture the spatial-temporal features of the brain. Bayesian relation inference, the core component of BBST, comes in two for... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a novel model named BSTT for sleep staging. BSST integrates the transformer and Bayesian relation inference to simultaneously model spatial-temporal relations and effectively capture the spatial-temporal features of the brain. Bayesian relation inference, the core component of BBST, comes in... |
This paper proposes a method for training RL agents in settings where the environment distribution is long-tailed, i.e. an agent must solve N tasks, but some of them occurs much more frequently than others. Examples of such a setting include the Zipfian RL environments introduced by [1]. This paper proposes a method th... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a method for training RL agents in settings where the environment distribution is long-tailed, i.e. an agent must solve N tasks, but some of them occurs much more frequently than others. Examples of such a setting include the Zipfian RL environments introduced by [1]. This paper proposes a m... |
This paper introduces two methods, FEDCVAE-ENS and FEDCVAE-KD, for one-shot federated learning with extreme data heterogeneity among clients. Once client models are trained using conditional variational auto-encoder (CVAE) architectures, the decoders, alongside label distributions of all clients, are uploaded to the se... | 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 introduces two methods, FEDCVAE-ENS and FEDCVAE-KD, for one-shot federated learning with extreme data heterogeneity among clients. Once client models are trained using conditional variational auto-encoder (CVAE) architectures, the decoders, alongside label distributions of all clients, are uploaded t... |
The paper studies provable sim-to-real transfer in the linear quadratic Gaussian (LQG) setting. The paper proposes a robust learning algorithm for the simulation part and derives provable guarantees without assuming access to real world data. In order to obtain the guarantees, the authors propose a novel history clippi... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies provable sim-to-real transfer in the linear quadratic Gaussian (LQG) setting. The paper proposes a robust learning algorithm for the simulation part and derives provable guarantees without assuming access to real world data. In order to obtain the guarantees, the authors propose a novel histor... |
The paper proposes a new model for unsupervised object-centric 3D scene understanding, called MORF. Unlike previous methods, it leverages 2D image segmentation masks from the pretrained optical flow method EISEN. These are used to initialized a slot based representation augmented by pixel features (as in PixelNeRF), wh... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes a new model for unsupervised object-centric 3D scene understanding, called MORF. Unlike previous methods, it leverages 2D image segmentation masks from the pretrained optical flow method EISEN. These are used to initialized a slot based representation augmented by pixel features (as in PixelN... |
This paper proposes an approach to applying the Entropy-based Logic Explained Network (E-LEN) to generate global explanations for a GNN from computed local explanations of the GNN, where both local and global explanations are expressed by sub-graphs. Experimental results on three datasets demonstrate good performance 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 proposes an approach to applying the Entropy-based Logic Explained Network (E-LEN) to generate global explanations for a GNN from computed local explanations of the GNN, where both local and global explanations are expressed by sub-graphs. Experimental results on three datasets demonstrate good perfo... |
This paper proposes phase2vec - a self-supervised framework for learning representations of dynamical systems in 2D. The main technical contribution is to use a 2D convolutional network for embedding dynamical system velocities, and formulating a reconstruction loss based on the data creation process to train the netwo... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes phase2vec - a self-supervised framework for learning representations of dynamical systems in 2D. The main technical contribution is to use a 2D convolutional network for embedding dynamical system velocities, and formulating a reconstruction loss based on the data creation process to train t... |
This paper systematically studies different tricks for FGSM adversarial training (FGSM-AT) on CIFAR-10/100. They found that there are three simple tricks that overcome the catastrophic overfitting in FGSM-AT: Data Initialization, Network Structure, and Optimization.
I think this work is a valuable contribution to the r... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper systematically studies different tricks for FGSM adversarial training (FGSM-AT) on CIFAR-10/100. They found that there are three simple tricks that overcome the catastrophic overfitting in FGSM-AT: Data Initialization, Network Structure, and Optimization.
I think this work is a valuable contribution ... |
This work proposes a hazard estimator which models the unobserved heterogeneity in individuals using frailty models. They propose two function approximation schemes where in one they assume the covariate effect and time effect to be multiplicative and in the other no assumptions are made. They also provide theoretical ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This work proposes a hazard estimator which models the unobserved heterogeneity in individuals using frailty models. They propose two function approximation schemes where in one they assume the covariate effect and time effect to be multiplicative and in the other no assumptions are made. They also provide theo... |
This paper proposes a novel approach to generalized policy iteration (GPI) exploiting the inherent connection between policy evaluation and policy improvement. It firstly shows that the policy evaluation and policy improvement steps are connected in the case of parameter sharing between the value function and the polic... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a novel approach to generalized policy iteration (GPI) exploiting the inherent connection between policy evaluation and policy improvement. It firstly shows that the policy evaluation and policy improvement steps are connected in the case of parameter sharing between the value function and t... |
This paper studies the question of why trained deep convolutional networks are insensitive to image perturbations such as small translation and rotation. Through a set of studies on CIFAR-10, the paper argues that this is mostly due to spatial pooling and channel pooling, and the two components can be further decoupled... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the question of why trained deep convolutional networks are insensitive to image perturbations such as small translation and rotation. Through a set of studies on CIFAR-10, the paper argues that this is mostly due to spatial pooling and channel pooling, and the two components can be further d... |
The authors consider data distributions of participating and unparticipating clients are drawn
from a meta-distribution and use classic Rademacher complexity- and local Rademacher complexity-based generalisation bounds to develop excess risk bounds for unparticipating clients and fast semi-excess risk bounds for parti... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors consider data distributions of participating and unparticipating clients are drawn
from a meta-distribution and use classic Rademacher complexity- and local Rademacher complexity-based generalisation bounds to develop excess risk bounds for unparticipating clients and fast semi-excess risk bounds f... |
This paper presents a method for Speech Enhancement (SE) applied to Automatic Speech Recognition (ASR). In the proposed approach, the SE model is trained to enhance noisy speech and to keep improving recognition performance of the acoustic model. This is accomplished by training the SE model with multi-task learning, w... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a method for Speech Enhancement (SE) applied to Automatic Speech Recognition (ASR). In the proposed approach, the SE model is trained to enhance noisy speech and to keep improving recognition performance of the acoustic model. This is accomplished by training the SE model with multi-task lea... |
This work proposed a label correction method to tackle the label noise in the segmentation task. They propose an algorithm to correct the Markov label noise.
Strength:
The Markov model for segmentation label noise considers spatial correlation.
Weakness:
The spatial correlation proposed by the authors contradicts t... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposed a label correction method to tackle the label noise in the segmentation task. They propose an algorithm to correct the Markov label noise.
Strength:
The Markov model for segmentation label noise considers spatial correlation.
Weakness:
The spatial correlation proposed by the authors contr... |
This paper investigates applying differentially private stochastic gradient descent (DP-SGD) for training models towards achieving high robustness against distributional shift. Experiments are done on datasets with covariate, label, or subpopulation shift using various noise multipliers in DP-SGD. The resulting perform... | 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 investigates applying differentially private stochastic gradient descent (DP-SGD) for training models towards achieving high robustness against distributional shift. Experiments are done on datasets with covariate, label, or subpopulation shift using various noise multipliers in DP-SGD. The resulting... |
This paper introduces GSS, a general purpose sequence model which leverages gated units and trains significantly faster as shown on several language modeling benchmarks. This simple-to-implement alternative to S4 and DSS which trains 2-3 times faster, and is competitive with Transformer-based baselines on long-range la... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper introduces GSS, a general purpose sequence model which leverages gated units and trains significantly faster as shown on several language modeling benchmarks. This simple-to-implement alternative to S4 and DSS which trains 2-3 times faster, and is competitive with Transformer-based baselines on long-... |
This paper proposes an adaptation of the ViT based CLIP model for video-text alignment that learns the spatio-temporal structure of the videos during the alignment. The authors start by analysis the problems with current approaches that straight forwardly apply CLIP on video for video-text alignment. They find that dat... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes an adaptation of the ViT based CLIP model for video-text alignment that learns the spatio-temporal structure of the videos during the alignment. The authors start by analysis the problems with current approaches that straight forwardly apply CLIP on video for video-text alignment. They find ... |
The paper lacks clarity, so I reproduce here my best understanding of its contents.
The paper proposes to search for the optimum bit width (or more generally, numerical precision) for each layer in a
neural network during post-training quantization. To this end, the authors assume quantization induces an additive
pric... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper lacks clarity, so I reproduce here my best understanding of its contents.
The paper proposes to search for the optimum bit width (or more generally, numerical precision) for each layer in a
neural network during post-training quantization. To this end, the authors assume quantization induces an addit... |
This paper proposes a mask-based approach to obtain subnetworks for specific tasks in continual learning (CL). It is an improvement of the previous method SupSup, which finds fixed supermask for each task to alleviate forgetting. This paper proposes to perform exclusive and nonoverlapping subnetwork training to avoid c... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a mask-based approach to obtain subnetworks for specific tasks in continual learning (CL). It is an improvement of the previous method SupSup, which finds fixed supermask for each task to alleviate forgetting. This paper proposes to perform exclusive and nonoverlapping subnetwork training to... |
This paper proposes two methods, Experienced Momentum (EM) and Precise Nesterov momentum (PN), to improve the transferability of adversarial attacks. Specifically, the proposed method starts with a few iterations to accumulate the gradient, and uses it as the initial momentum. PN considers the gradient of the current d... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes two methods, Experienced Momentum (EM) and Precise Nesterov momentum (PN), to improve the transferability of adversarial attacks. Specifically, the proposed method starts with a few iterations to accumulate the gradient, and uses it as the initial momentum. PN considers the gradient of the c... |
The paper proposes a method for finding the most likely transition between two states in high-dimensional configuration spaces (for chemical compounds, typically). It does so on building on recent advances in the Schrödinger bridge problem (SBP) and stochastic optimal control (SOC).
It claims to show equivalence betwe... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a method for finding the most likely transition between two states in high-dimensional configuration spaces (for chemical compounds, typically). It does so on building on recent advances in the Schrödinger bridge problem (SBP) and stochastic optimal control (SOC).
It claims to show equivalen... |
The paper proposes smoothness/robustness of explainers of blackbox classifiers as a desirable objective. It gives theoretical bounds for this measure of robustness, (called explainer astuteness) in terms of the Lipschitz constant of the learned blackbox model. Theorems relating the two are given for three types of expl... | 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 proposes smoothness/robustness of explainers of blackbox classifiers as a desirable objective. It gives theoretical bounds for this measure of robustness, (called explainer astuteness) in terms of the Lipschitz constant of the learned blackbox model. Theorems relating the two are given for three types... |
This paper mainly focuses on few-shot learning. Different from the previous methods pursuing good overall accuracy, the authors study the performance on hard few-shot tasks. Specifically they propose a novel algorithm based on constrained optimization to find difficult support samples given query samples. Such a method... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper mainly focuses on few-shot learning. Different from the previous methods pursuing good overall accuracy, the authors study the performance on hard few-shot tasks. Specifically they propose a novel algorithm based on constrained optimization to find difficult support samples given query samples. Such ... |
This paper notices that poisoned and clean "domains" do not help classify one another. Therefore, identifying the subset of the data that does not help classify the rest can defend against poisoning attacks. The paper also proposes a computationally feasible approach to this defense strategy and tries it on several d... | 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 notices that poisoned and clean "domains" do not help classify one another. Therefore, identifying the subset of the data that does not help classify the rest can defend against poisoning attacks. The paper also proposes a computationally feasible approach to this defense strategy and tries it on s... |
This paper presents a novel approach to abstraction in Reinforcement Learning in discrete action spaces by construction a homomorphic representation through equivalent effect state-action pairs. The approach constructs forward and backward models which allow it to reduce the complexity of the value function representat... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents a novel approach to abstraction in Reinforcement Learning in discrete action spaces by construction a homomorphic representation through equivalent effect state-action pairs. The approach constructs forward and backward models which allow it to reduce the complexity of the value function rep... |
In this paper, the authors studied the implicit bias of gradient flow and gradient descent for two-layer leaky-ReLU networks with nearly orthogonal data. It is shown that gradient flow will asymptotically converge to a network with rank at most 2 and linear decision boundary. Such a network is also approximately a max-... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors studied the implicit bias of gradient flow and gradient descent for two-layer leaky-ReLU networks with nearly orthogonal data. It is shown that gradient flow will asymptotically converge to a network with rank at most 2 and linear decision boundary. Such a network is also approximatel... |
This paper studies the question of whether transfer learning (USL) is superior to meta-learning (MAML) for few-shot classification (multi-task), as suggested by recent studies.
It shows that the two methods have identical performance when task diversity is low, while their performance is different when task diversity i... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the question of whether transfer learning (USL) is superior to meta-learning (MAML) for few-shot classification (multi-task), as suggested by recent studies.
It shows that the two methods have identical performance when task diversity is low, while their performance is different when task div... |
The paper extends an existing approach REDQ for dealing with Q value overestimation in (stochastic environment) single agent RL setting to value based multi agent RL (MARL). Experiments are provided on particle environment and Starcraft domains.
1. The related works is incomplete. The paper lacks discussion about the r... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper extends an existing approach REDQ for dealing with Q value overestimation in (stochastic environment) single agent RL setting to value based multi agent RL (MARL). Experiments are provided on particle environment and Starcraft domains.
1. The related works is incomplete. The paper lacks discussion abo... |
The paper seeks to enforce the group fairness metric "Difference of Conditional Accuracy", a generalization of the equalized odds criterion, and points out an interesting connection between optimizing this metric and (class-wise) distributionally robust optimization (DRO). Specifically, they show that the criterion ca... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper seeks to enforce the group fairness metric "Difference of Conditional Accuracy", a generalization of the equalized odds criterion, and points out an interesting connection between optimizing this metric and (class-wise) distributionally robust optimization (DRO). Specifically, they show that the crit... |
In maximum entropy RL, we use a soft Bellman optimality equations where the standard max operation is replaced by the log-sum-exp operation. Just as researchers have proposed to use in-sample max instead of max over the entire action space, this paper propose to use in-sample softmax to improve the softmax operation, s... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In maximum entropy RL, we use a soft Bellman optimality equations where the standard max operation is replaced by the log-sum-exp operation. Just as researchers have proposed to use in-sample max instead of max over the entire action space, this paper propose to use in-sample softmax to improve the softmax oper... |
The paper proposes to improve inference in Markov Logic Networks through the use of deep models. In particular, Mean Field inference is implemented as layers which pass messages to compute summations more efficiently by exploiting logical structure in the MLN.
The time complexity analysis is performed to show that the... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes to improve inference in Markov Logic Networks through the use of deep models. In particular, Mean Field inference is implemented as layers which pass messages to compute summations more efficiently by exploiting logical structure in the MLN.
The time complexity analysis is performed to show ... |
An unlearning algorithm for graph neural networks is proposed in the paper. The paper tries to find the difference between the upweighted model and the original model by utilizing a hessian-based approximation, which can be solved by investigating the corresponding linear system with conjugate gradient methods. The fou... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
An unlearning algorithm for graph neural networks is proposed in the paper. The paper tries to find the difference between the upweighted model and the original model by utilizing a hessian-based approximation, which can be solved by investigating the corresponding linear system with conjugate gradient methods.... |
This paper presents a new finding in subspace by constructing a Compressed Parameter Subspaces (CPS). It is found that a
geometric structure representing distance-regularized parameters mapped to a set of train-time distributions can maximize average accuracy over a broad range of distribution shifts.
Strength
+ The s... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents a new finding in subspace by constructing a Compressed Parameter Subspaces (CPS). It is found that a
geometric structure representing distance-regularized parameters mapped to a set of train-time distributions can maximize average accuracy over a broad range of distribution shifts.
Strength
... |
The paper addresses a problem arising from the generality of source-based conventional batch normalization (CBN) or test-based batch normalization (TBN), with the former biasing the architecture to source distribution and the latter suffering from inaccuracies for small test batch sizes. The paper proposes combining CB... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper addresses a problem arising from the generality of source-based conventional batch normalization (CBN) or test-based batch normalization (TBN), with the former biasing the architecture to source distribution and the latter suffering from inaccuracies for small test batch sizes. The paper proposes comb... |
The paper addresses the evaluation of force fields (FF) for atomistic simulations.
The authors propose benchmark datasets and metrics based on molecular dynamics(MD) simulation, a major use case of FFs.
When state-of-the-art machine learning(ML)-based FFs are evaluated on the benchmark datasets, conventional force ac... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper addresses the evaluation of force fields (FF) for atomistic simulations.
The authors propose benchmark datasets and metrics based on molecular dynamics(MD) simulation, a major use case of FFs.
When state-of-the-art machine learning(ML)-based FFs are evaluated on the benchmark datasets, conventional ... |
The paper proposes RS and MI, two permutation-based image encryption algorithms designed to render images unrecognizable to humans, yet still enable image classification and object detection respectively, by downstream vision transformers. The first step of both encryption algorithms is to partition images in to N equa... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes RS and MI, two permutation-based image encryption algorithms designed to render images unrecognizable to humans, yet still enable image classification and object detection respectively, by downstream vision transformers. The first step of both encryption algorithms is to partition images in t... |
The paper is about evaluation of open-ended generation of LMs, in particular the Mauve (pillutla et al.) metric.
Evaluation of open ended text generation takes a distributional format. It is framed in this paper Divergence(p_w, q_w). This task has two sub-tasks, (1) Density estimation of p_w and q_w (when necessary).... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper is about evaluation of open-ended generation of LMs, in particular the Mauve (pillutla et al.) metric.
Evaluation of open ended text generation takes a distributional format. It is framed in this paper Divergence(p_w, q_w). This task has two sub-tasks, (1) Density estimation of p_w and q_w (when nec... |
In summary, this paper first observes that there exists a historical training stage where the model has a higher OOD detection performance than the final well-trained one. After that, this paper aims to backtrack the model to that stage. To achieve this, this paper proposes a method called UM to forget those "atypical ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In summary, this paper first observes that there exists a historical training stage where the model has a higher OOD detection performance than the final well-trained one. After that, this paper aims to backtrack the model to that stage. To achieve this, this paper proposes a method called UM to forget those "a... |
This paper proposes a surprising new method for keypoint matching across image pairs. The idea is essentially to use randomly-initialized CNNs to generate the features. The random features are first preprocessed slightly to get local peaks, and then the features which do not have mutual neighbors are discarded. Finally... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a surprising new method for keypoint matching across image pairs. The idea is essentially to use randomly-initialized CNNs to generate the features. The random features are first preprocessed slightly to get local peaks, and then the features which do not have mutual neighbors are discarded.... |
This paper studies adversarial attacks on sequential decision-making policies, with a focus on *statistically undetectable* attacks. The authors assume exact knowledge of the world model, and introduce a novel class of adversarial attacks called illusory attacks, which are consistant with the world dynamics and thus mo... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies adversarial attacks on sequential decision-making policies, with a focus on *statistically undetectable* attacks. The authors assume exact knowledge of the world model, and introduce a novel class of adversarial attacks called illusory attacks, which are consistant with the world dynamics and... |
1, Adopt importance sampling to select the masked patches with richer semantic information for reconstruction, instead of random sampling as done in previous MIM works.
2, Propose a new contrastive loss that aligns the tokens of the vision transformer extracted from the selected masked patches and the remaining ones.
3... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
1, Adopt importance sampling to select the masked patches with richer semantic information for reconstruction, instead of random sampling as done in previous MIM works.
2, Propose a new contrastive loss that aligns the tokens of the vision transformer extracted from the selected masked patches and the remaining... |
The paper proposes to improve the pseudo-label in semi-supervised learning with a well-calibrated model. To that end, it improves the calibration with Bayesian model averaging. It also provides a theoretical result for the benefit of the calibration in semi-supervised learning for the task of classification. The approa... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes to improve the pseudo-label in semi-supervised learning with a well-calibrated model. To that end, it improves the calibration with Bayesian model averaging. It also provides a theoretical result for the benefit of the calibration in semi-supervised learning for the task of classification. Th... |
This work proposes a noise conditional likelihood-based method for training autoregressive models. Specifically, instead of directly training the model on clean images, the authors propose to train a sequence of autoregressive models on images perturbed with different noise levels using MLE. The authors show with empir... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This work proposes a noise conditional likelihood-based method for training autoregressive models. Specifically, instead of directly training the model on clean images, the authors propose to train a sequence of autoregressive models on images perturbed with different noise levels using MLE. The authors show wi... |
This paper studied fine-grained image recognition under few-shot settings. The authors proposed to learn a dictionary of parts using object parsing first and then train a few-shot learner based on the parsed features. Learning a structured aligned representation is the right way for fine-grained few-shot learning. The ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studied fine-grained image recognition under few-shot settings. The authors proposed to learn a dictionary of parts using object parsing first and then train a few-shot learner based on the parsed features. Learning a structured aligned representation is the right way for fine-grained few-shot learni... |
Existing recourse generation methods optimize for any single objective that is shared across different users not paying attention to a user’s individuality. This might result in impractical recourses due to undesirable feature changes. Different individuals’ preferences may associate different costs to the same feature... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Existing recourse generation methods optimize for any single objective that is shared across different users not paying attention to a user’s individuality. This might result in impractical recourses due to undesirable feature changes. Different individuals’ preferences may associate different costs to the same... |
This paper proposes a backdoor poisoning attack that aims to defeat latent separability-based defenses that purify the training set. The attack has two main components: regularization samples that contain the trigger but assigned to the correct class and asymmetric triggers that are weaker versions of the actual trigge... | 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 backdoor poisoning attack that aims to defeat latent separability-based defenses that purify the training set. The attack has two main components: regularization samples that contain the trigger but assigned to the correct class and asymmetric triggers that are weaker versions of the actua... |
The authors propose a method to learn the time-reversible, nonlinear dynamics of an incompressible fluid. For computational tractability, the authors learn a reduced model over a limited set of initial conditions with tensor precomputation. The authors show that their approach predicts solutions significantly closer to... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a method to learn the time-reversible, nonlinear dynamics of an incompressible fluid. For computational tractability, the authors learn a reduced model over a limited set of initial conditions with tensor precomputation. The authors show that their approach predicts solutions significantly c... |
The authors develop an approach for conditional average treatment effect estimation for interventions that have an arbitrary dimension and type (i.e., could be continuous valued or binary). They do so by extending previous work by Johansson/Shalit on binary treatments.
Strengths:
Because the majority of work in cau... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors develop an approach for conditional average treatment effect estimation for interventions that have an arbitrary dimension and type (i.e., could be continuous valued or binary). They do so by extending previous work by Johansson/Shalit on binary treatments.
Strengths:
Because the majority of wor... |
The manuscript presents a procedure for protein engineering using a model-based reinforcement approach building on the ESM2 language model. The authors demonstrate that the high dimensional ESM2 representation can be mapped to a lower dimensional representation space which is suitable for optimization, and that full am... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The manuscript presents a procedure for protein engineering using a model-based reinforcement approach building on the ESM2 language model. The authors demonstrate that the high dimensional ESM2 representation can be mapped to a lower dimensional representation space which is suitable for optimization, and that... |
The paper proposes a continuous normalizing flow architecture that is invariant to input permutations, which is useful for modelling densities of object _sets_. Authors design an equivariant dynamics function by only modelling point-wise and pair-wise interactions. Using this dynamics function with an invariant base di... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The paper proposes a continuous normalizing flow architecture that is invariant to input permutations, which is useful for modelling densities of object _sets_. Authors design an equivariant dynamics function by only modelling point-wise and pair-wise interactions. Using this dynamics function with an invariant... |
The paper deals with the question of learning Stackelberg equilibria -- games with a leader-follower structure -- an important topic with many relevant applications. The main contribution is to show that, given access to a best-response oracle for the follower, this problem can be reduced to learning in a POMDP consist... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper deals with the question of learning Stackelberg equilibria -- games with a leader-follower structure -- an important topic with many relevant applications. The main contribution is to show that, given access to a best-response oracle for the follower, this problem can be reduced to learning in a POMDP... |
This paper studies the calibration of differentially private learners based on stochastic gradient descent. The paper first observes the miscalibration is due to the per-example gradient clipping. Then, the paper provides differentially private recalibration to reduce calibration errors. The basic idea is to divide 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:
This paper studies the calibration of differentially private learners based on stochastic gradient descent. The paper first observes the miscalibration is due to the per-example gradient clipping. Then, the paper provides differentially private recalibration to reduce calibration errors. The basic idea is to ... |
The paper proposes a new way to incorporate MBconv in a transformers architecture. This in order to propose an efficient architecture for many tasks with good FLOPS- accuracy and parameters-accuracy trade-off.
Strenghts:
- The paper is well written and easy to follow.
- The proposed method is simple and seems to give ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a new way to incorporate MBconv in a transformers architecture. This in order to propose an efficient architecture for many tasks with good FLOPS- accuracy and parameters-accuracy trade-off.
Strenghts:
- The paper is well written and easy to follow.
- The proposed method is simple and seems ... |
This paper proposes *radial neural networks*, neural networks with a non-pointwise nonlinearity that is a scalar rescaling of the vector of activations as a function of its norm. The paper argues that such neural networks are rotation equivariant and proves that they are universal function approximators. It further int... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proposes *radial neural networks*, neural networks with a non-pointwise nonlinearity that is a scalar rescaling of the vector of activations as a function of its norm. The paper argues that such neural networks are rotation equivariant and proves that they are universal function approximators. It fur... |
The paper proposes three changes to vanilla diffusion models: i) a regularization term in the training objective, ii) a source of randomness in the training objective to account for the fact that there is no single solution to the denoising problem iii) a class of more general corruption processes. The paper demonstrat... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The paper proposes three changes to vanilla diffusion models: i) a regularization term in the training objective, ii) a source of randomness in the training objective to account for the fact that there is no single solution to the denoising problem iii) a class of more general corruption processes. The paper de... |
This paper proposed an any-scale balanced proposal for discrete distribution. Motivated by recent findings that non-local proposals improve sampling efficient, the proposed method closes the gaps of the choice of weight function and the accuracy of first order gradient approximation. Comparison to several baseline meth... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposed an any-scale balanced proposal for discrete distribution. Motivated by recent findings that non-local proposals improve sampling efficient, the proposed method closes the gaps of the choice of weight function and the accuracy of first order gradient approximation. Comparison to several basel... |
They propose a novel analysis of the classical label propagation algorithm with an error bound. They also propose a framework to incorporate multiple sources of noisy information.
This paper provides some interesting results, but there are some questions needed to be answered.
1. Definition 1 is confused in this paper,... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
They propose a novel analysis of the classical label propagation algorithm with an error bound. They also propose a framework to incorporate multiple sources of noisy information.
This paper provides some interesting results, but there are some questions needed to be answered.
1. Definition 1 is confused in thi... |
The paper proposes Eigen Memory Tree to serve as an efficient data structure for online memory.
Pros: potentially useful.
Cons: I am not convinced ICLR is the right place for this paper. While there is connection to machine learning, the key contribution of a more efficient and practical data structure should be eval... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper proposes Eigen Memory Tree to serve as an efficient data structure for online memory.
Pros: potentially useful.
Cons: I am not convinced ICLR is the right place for this paper. While there is connection to machine learning, the key contribution of a more efficient and practical data structure should... |
This paper studies self-supervised long-tailed learning with the auxiliary of out-of-distribution data by designing a sampling strategy together with a distribution-level supervised contrastive loss. Empirical validation demonstrates the effectiveness of the proposed method on a range of datasets compared with previous... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper studies self-supervised long-tailed learning with the auxiliary of out-of-distribution data by designing a sampling strategy together with a distribution-level supervised contrastive loss. Empirical validation demonstrates the effectiveness of the proposed method on a range of datasets compared with ... |
The work proposes Group Masked Model Learning framework for self supervised learning.
Strength:
The work reports promising results for self supervised learning in low data regime.
Weaknesses:
1. The novelty of the work is confusing as the work looks quite similar to other popular approaches for self supervised lear... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The work proposes Group Masked Model Learning framework for self supervised learning.
Strength:
The work reports promising results for self supervised learning in low data regime.
Weaknesses:
1. The novelty of the work is confusing as the work looks quite similar to other popular approaches for self supervi... |
This paper proposes a method for multi-modal human activity recognition using RGB and IMU data. The main contributions include 1) modeling IMU data using several 2D images following the approach of Wang and Oats (2015) to build a Gramian Angular Field (GAF). The paper then proposes 2) extracting equal sized diagonal ma... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a method for multi-modal human activity recognition using RGB and IMU data. The main contributions include 1) modeling IMU data using several 2D images following the approach of Wang and Oats (2015) to build a Gramian Angular Field (GAF). The paper then proposes 2) extracting equal sized dia... |
The paper proposes a new active learning approach by taking into consideration augmentations during the Active Learning acquisition phase. The paper argues that the existing active learning baselines do not effectively utilize unlabeled samples, particularly during the early stages of acquisition. To fix this, the auth... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper proposes a new active learning approach by taking into consideration augmentations during the Active Learning acquisition phase. The paper argues that the existing active learning baselines do not effectively utilize unlabeled samples, particularly during the early stages of acquisition. To fix this, ... |
This paper provides a tool to design policies for solving tasks for different goals. It provides a general framework, that the downstream policy can then use in a RL framework. The paper provides proof of concept demonstrations for designing policies able to solve different manipulation tasks, conditional on the type o... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper provides a tool to design policies for solving tasks for different goals. It provides a general framework, that the downstream policy can then use in a RL framework. The paper provides proof of concept demonstrations for designing policies able to solve different manipulation tasks, conditional on th... |
The paper proposes to relax the assumption of statistical independence used in many disentangled representation learning works. The paper proposes to only assume that the support of the latent factors’ distribution factorizes, which is a weaker constraint than statistical independence. Specifically, the paper proposes ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes to relax the assumption of statistical independence used in many disentangled representation learning works. The paper proposes to only assume that the support of the latent factors’ distribution factorizes, which is a weaker constraint than statistical independence. Specifically, the paper p... |
This paper investigates deep vs. wide (in terms of the number of attention heads) models with a variety of recent transformer variants. It argues that, compared to their deep and thin counterparts, shallow and wide architectures (1) achieve the same or better performance, (2) has fewer parameters, and (3) run faster. T... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper investigates deep vs. wide (in terms of the number of attention heads) models with a variety of recent transformer variants. It argues that, compared to their deep and thin counterparts, shallow and wide architectures (1) achieve the same or better performance, (2) has fewer parameters, and (3) run f... |
This paper puts forth a methodology to understand the underlying factors (in the model and/or data) that affect blindspot detection methods. They rely on synthetic data wherein the number and nature of blindspots can be controlled. They also propose a new blindspot detection approach, PlaneSpot.
Blindspot detection is ... | 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 puts forth a methodology to understand the underlying factors (in the model and/or data) that affect blindspot detection methods. They rely on synthetic data wherein the number and nature of blindspots can be controlled. They also propose a new blindspot detection approach, PlaneSpot.
Blindspot detec... |
In order to better utilize the training paradigm of CTDE, the authors introduce dependency policy correction and dependency critic correction based on AC architecture to consider the action dependencies among the agents, and also design the corresponding network structure so that more accurate joint policy can be learn... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In order to better utilize the training paradigm of CTDE, the authors introduce dependency policy correction and dependency critic correction based on AC architecture to consider the action dependencies among the agents, and also design the corresponding network structure so that more accurate joint policy can ... |
This paper proposes a framework that computes a compact latent vector for an input implicit neural representation for 3D shape (i.e., encoded with MLP weights). It's trained on a dataset of INRs via reconstruction loss, and is able to obtain a compact embedding from an unseen INR (but still from the same data distribut... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a framework that computes a compact latent vector for an input implicit neural representation for 3D shape (i.e., encoded with MLP weights). It's trained on a dataset of INRs via reconstruction loss, and is able to obtain a compact embedding from an unseen INR (but still from the same data d... |
This paper proposes a method to represent a class as a subspace in the deep learning regime. The contributions of this paper are the formulation of classes as subspaces, the Grassmannian layer to update subspaces, learning the Grassmannian layer using constraint optimization. The core contribution is to represent a cla... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method to represent a class as a subspace in the deep learning regime. The contributions of this paper are the formulation of classes as subspaces, the Grassmannian layer to update subspaces, learning the Grassmannian layer using constraint optimization. The core contribution is to represe... |
Poor generalization on out-of-distribution data is a common problem among machine learning models. In classification tasks, one recently proposed solution is Domain Adversarial Training, which asks the model to predict the domain of the data sample in addition to its class, and encourages the model to learn domain-agno... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
Poor generalization on out-of-distribution data is a common problem among machine learning models. In classification tasks, one recently proposed solution is Domain Adversarial Training, which asks the model to predict the domain of the data sample in addition to its class, and encourages the model to learn dom... |
This work proposes a variant of mean field games, where there is a correlation device, and subsequently introduce the corresponding equilibrium concept.
It analyzes this type of game/equilibrium concept, in particular an existence result (Thm 1).
Additionally, the imitation learning objective for learning the policie... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This work proposes a variant of mean field games, where there is a correlation device, and subsequently introduce the corresponding equilibrium concept.
It analyzes this type of game/equilibrium concept, in particular an existence result (Thm 1).
Additionally, the imitation learning objective for learning the... |
This paper empirically studies fitting single-hidden-layer neural networks where data is also generated by single-hidden-layer ReLU teacher networks. This paper demonstrates that stochastic gradient descent (SGD) with automated width selection attains small expected errors for a specific experiment setting. They fina... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper empirically studies fitting single-hidden-layer neural networks where data is also generated by single-hidden-layer ReLU teacher networks. This paper demonstrates that stochastic gradient descent (SGD) with automated width selection attains small expected errors for a specific experiment setting. T... |
This paper targets the incapability of GNNs on heterophilous graphs and proposes to process the node feature and graph structures separately. Specifically, Multiple node feature prototypes are first generated with only the node feature information. Then, arbitrary GNN is used to obtain a structural view (embedding) of ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper targets the incapability of GNNs on heterophilous graphs and proposes to process the node feature and graph structures separately. Specifically, Multiple node feature prototypes are first generated with only the node feature information. Then, arbitrary GNN is used to obtain a structural view (embedd... |
The paper presents a probabilistic programming language based on extending a prolog-like language with probabilistic facts and equipping a query with probabilistic semantics. The language includes both continuous and discrete random variables and it includes parameterized neural networks.
The authors describe a fairl... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper presents a probabilistic programming language based on extending a prolog-like language with probabilistic facts and equipping a query with probabilistic semantics. The language includes both continuous and discrete random variables and it includes parameterized neural networks.
The authors describe... |
This paper studies strategic classification under graph neural networks. In the model of strategic user behavior, it is assumed that a user plays myopic best-response over a sequence of multiple update rounds. The myopic response mapping means that a user may move their reported features in the direction that best impr... | 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 studies strategic classification under graph neural networks. In the model of strategic user behavior, it is assumed that a user plays myopic best-response over a sequence of multiple update rounds. The myopic response mapping means that a user may move their reported features in the direction that b... |
This paper introduces a greedy context-based meta-testing procedure to tackle the distribution shift issue when performing online adaptation of offline meta-trained policies. The proposed method extends FOCAL with a diverse latent task embedding sampling strategy and a context selection mechanism based on the highest r... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper introduces a greedy context-based meta-testing procedure to tackle the distribution shift issue when performing online adaptation of offline meta-trained policies. The proposed method extends FOCAL with a diverse latent task embedding sampling strategy and a context selection mechanism based on the h... |
The authors propose a novel approach to learn a data augmentation method by formulating the task as invariance-constrained learning problem. They leverage Monte Carlo Markov Chain (MCMC) sampling to solve it, which achieves state-of-the-art results on CIFAR10 and CIFAR100 for specific wide neural network architectures.... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose a novel approach to learn a data augmentation method by formulating the task as invariance-constrained learning problem. They leverage Monte Carlo Markov Chain (MCMC) sampling to solve it, which achieves state-of-the-art results on CIFAR10 and CIFAR100 for specific wide neural network archit... |
The papers proposes to learn the view generation component for contrastive learning for time series data.
The alogithm is simple and includes usage of 5 common augmentations techniques and one new one (TimeDis) applied sequentialy to generate a view.
The loss function is similar to SimCLR and is maximized wrt parameter... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
The papers proposes to learn the view generation component for contrastive learning for time series data.
The alogithm is simple and includes usage of 5 common augmentations techniques and one new one (TimeDis) applied sequentialy to generate a view.
The loss function is similar to SimCLR and is maximized wrt p... |
This paper considers variational inequality (VI) problems where complex constraints exist. They use the framework of the interior point method with log-barriers to handle the constraints. They merge this framework with the ADMM algorithm to provide the final design of the algorithm which is a two-loop algorithm. The ou... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper considers variational inequality (VI) problems where complex constraints exist. They use the framework of the interior point method with log-barriers to handle the constraints. They merge this framework with the ADMM algorithm to provide the final design of the algorithm which is a two-loop algorithm... |
The manuscript introduces a new method for protein design (inverse folding) and a benchmark consisting of a curated set of structures from the AlphaFold Database. The authors demonstrate a substantial improvement compared to a selection of earlier methods.
The paper makes good use of the recently released AlphaFold Dat... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The manuscript introduces a new method for protein design (inverse folding) and a benchmark consisting of a curated set of structures from the AlphaFold Database. The authors demonstrate a substantial improvement compared to a selection of earlier methods.
The paper makes good use of the recently released Alpha... |
This paper aims to design a domain generalization model based on causal balancing. To this end, the authors assume that the observation is causally determined by the label and some latent factors and a random variable. Based on such assumption, the authors have provided many theories to demonstrate that the causal mode... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper aims to design a domain generalization model based on causal balancing. To this end, the authors assume that the observation is causally determined by the label and some latent factors and a random variable. Based on such assumption, the authors have provided many theories to demonstrate that the cau... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.