Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
The paper proposes Neural-Symbolic Recursive Machine (NSR) as a means for neural networks to learn compositional rules through 3 modules, one for processing perception, syntactic parsing and semantic reasoning. These modules made up the Grounded Symbol System (GSS) and are jointly learned through a deduction-abduction ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper proposes Neural-Symbolic Recursive Machine (NSR) as a means for neural networks to learn compositional rules through 3 modules, one for processing perception, syntactic parsing and semantic reasoning. These modules made up the Grounded Symbol System (GSS) and are jointly learned through a deduction-ab...
The standard GAE estimator uses a lambda weighted average of the k-step returns to estimate the advantage, i.e., it uses $\lambda(r_0 + \gamma V(s_1)) + \lambda^2 (r_0 + \gamma r_1 + \gamma^2 V(s_2) + ...)$ (in a normalized fashion). In this standard method, the sequence continues until the terminal state or until $\in...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The standard GAE estimator uses a lambda weighted average of the k-step returns to estimate the advantage, i.e., it uses $\lambda(r_0 + \gamma V(s_1)) + \lambda^2 (r_0 + \gamma r_1 + \gamma^2 V(s_2) + ...)$ (in a normalized fashion). In this standard method, the sequence continues until the terminal state or un...
The paper shows that randomized smoothing in floating point arithmetic is susceptible to adversarial attacks. * Strength The paper shows that randomized smoothing in floating point arithmetic, which has not been an obvious attack angle. * Weakness The proposed solution requires i.i.d. random numbers, which in pract...	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 shows that randomized smoothing in floating point arithmetic is susceptible to adversarial attacks. * Strength The paper shows that randomized smoothing in floating point arithmetic, which has not been an obvious attack angle. * Weakness The proposed solution requires i.i.d. random numbers, which ...
The paper presents a novel generative task of generating textual descriptions from music. They focus on a small dataset of classical music recordings paired with expert descriptions and present comparative experiments using various pre-trained language models vs their method with a novel topology-preservation loss. Str...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents a novel generative task of generating textual descriptions from music. They focus on a small dataset of classical music recordings paired with expert descriptions and present comparative experiments using various pre-trained language models vs their method with a novel topology-preservation l...
This paper proposed a new partical-based variational inference method by minimization the functional gradient of KL with a regularization where preconditioning seems to be crucial for the performance. The efficiency compared to SVGD is illustrated on several standard benchmark data sets. Relations to previous work were...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposed a new partical-based variational inference method by minimization the functional gradient of KL with a regularization where preconditioning seems to be crucial for the performance. The efficiency compared to SVGD is illustrated on several standard benchmark data sets. Relations to previous w...
To improve the attack-defense gap for robust classifiers, the paper suggests using a sample-specific attack-based flag, Counter-Attack(CA), that tells the user if there are adversarial samples within a $\epsilon$-ball of the given sample. In the case of a perfect attack, the authors show that Counter-Attack also provi...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: To improve the attack-defense gap for robust classifiers, the paper suggests using a sample-specific attack-based flag, Counter-Attack(CA), that tells the user if there are adversarial samples within a $\epsilon$-ball of the given sample. In the case of a perfect attack, the authors show that Counter-Attack al...
This article introduces a vectorization---called PersGril---of $2$-parameter-based persistence, akin to the well-known persistence landscapes routinely used for 1-persistent homology. They prove that their vectorization is stable (1-Lipschitz for the infinite norm) with respect to the inputs functions $\mathcal{X} \to ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This article introduces a vectorization---called PersGril---of $2$-parameter-based persistence, akin to the well-known persistence landscapes routinely used for 1-persistent homology. They prove that their vectorization is stable (1-Lipschitz for the infinite norm) with respect to the inputs functions $\mathcal...
The author proposed to a new geo-encoding method called GeoVeX. The idea is to train an autoencoder on the OSM data projected to H3. Experiments show that GeoVeX outperforms baselines like Hex2Vec and Space2Vec. Strength - The paper is clearly written and experiments demonstrated GeoVeX's improvement. Weakness - The ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The author proposed to a new geo-encoding method called GeoVeX. The idea is to train an autoencoder on the OSM data projected to H3. Experiments show that GeoVeX outperforms baselines like Hex2Vec and Space2Vec. Strength - The paper is clearly written and experiments demonstrated GeoVeX's improvement. Weaknes...
This paper proposes ModelAngelo, a two-step pipeline to predict the protein structure using cryo-EM maps and protein sequences as input. To do this, ModelAngelo first employs a 3D CNN to predict positions of alpha carbons in protein sequences taking cryo-EM maps as input. Apart from this, they train a separate graph ne...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes ModelAngelo, a two-step pipeline to predict the protein structure using cryo-EM maps and protein sequences as input. To do this, ModelAngelo first employs a 3D CNN to predict positions of alpha carbons in protein sequences taking cryo-EM maps as input. Apart from this, they train a separate ...
The paper presents a method to estimate 3D rotations of an object from images using classical physics. The paper is focused on predicting satellite rotations. It uses Hamiltonian representation of rotational motion. With this representation, motion trajectory is differentiable everywhere. This is not true for rotation...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper presents a method to estimate 3D rotations of an object from images using classical physics. The paper is focused on predicting satellite rotations. It uses Hamiltonian representation of rotational motion. With this representation, motion trajectory is differentiable everywhere. This is not true for ...
The authors introduce a novel method for Preference-based Reinforcement Learning, which targets the problem of non-Markovian rewards. Namely, that human trajectory preferences cannot be assumed to be based on a sum of Markovian rewards, as expected in common RL and PbRL settings. The method learns non-Markovian rewards...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors introduce a novel method for Preference-based Reinforcement Learning, which targets the problem of non-Markovian rewards. Namely, that human trajectory preferences cannot be assumed to be based on a sum of Markovian rewards, as expected in common RL and PbRL settings. The method learns non-Markovian...
This paper proposes to initialize each linear layer of a deep neural network to be identity so as to accelerate training. Strengths: 1. The experiments are done on many datasets. 2. The code is attached for reproducibility. Weaknesses: 1. It has been shown (Thereom 5 in [1]) that the identity initialization may not l...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes to initialize each linear layer of a deep neural network to be identity so as to accelerate training. Strengths: 1. The experiments are done on many datasets. 2. The code is attached for reproducibility. Weaknesses: 1. It has been shown (Thereom 5 in [1]) that the identity initialization m...
This paper presents a novel method for decoding images from FMRI, using a hierarchical VAE. First, a VAE is trained on images from imagenet. Then the encoder is removed, and a mapping is learned from the images to the latent variables of the VAE. The VAE is divided into lower and higher level areas (i.e., it's hierarch...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a novel method for decoding images from FMRI, using a hierarchical VAE. First, a VAE is trained on images from imagenet. Then the encoder is removed, and a mapping is learned from the images to the latent variables of the VAE. The VAE is divided into lower and higher level areas (i.e., it's ...
The paper proposes a way of defining the constraints in constrained RL via Scenario-based Programming. They demonstrate the approach on the problem of mapless navigation in simulation as well as on a real Turtlebot. Strengths: - The problem is interesting as defining the reward function and constraints can be difficul...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a way of defining the constraints in constrained RL via Scenario-based Programming. They demonstrate the approach on the problem of mapless navigation in simulation as well as on a real Turtlebot. Strengths: - The problem is interesting as defining the reward function and constraints can be ...
This paper presents an imitation learning-based task and motion planning (TAMP) method for active exploration and topological mapping of unknown indoor environments. Being metric-free, topological mapping allows for greater computational efficiency with respect to metric-based approaches. The method is composed of two ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents an imitation learning-based task and motion planning (TAMP) method for active exploration and topological mapping of unknown indoor environments. Being metric-free, topological mapping allows for greater computational efficiency with respect to metric-based approaches. The method is composed...
Authors proposed GeoVEX as a framework for global representation learning in gespatial settings. The apprach leverages the H3 geospatial indexing system and data from Open Street Maps to create an embedding for each location on earth. An autoencoder like network architecture was introduced for the embedding generation ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Authors proposed GeoVEX as a framework for global representation learning in gespatial settings. The apprach leverages the H3 geospatial indexing system and data from Open Street Maps to create an embedding for each location on earth. An autoencoder like network architecture was introduced for the embedding gen...
The paper considers dynamics of stochastic gradient descent (SGD) and relates accuracy with a notion of "accuracy discrepancy". The paper shows that if the "accuracy discrepancy" is large enough, then SGD can find a model with perfect accuracy, while on the other hand if the "accuracy discrepancy" is small, there exist...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper considers dynamics of stochastic gradient descent (SGD) and relates accuracy with a notion of "accuracy discrepancy". The paper shows that if the "accuracy discrepancy" is large enough, then SGD can find a model with perfect accuracy, while on the other hand if the "accuracy discrepancy" is small, the...
This work find that the inherent edge noise can perturb the graph topology and labels, which may reduce link prediction performance. Thus, the authors propose an information bottleneck guided method, namely RGIB. RGIB achieves robustness representation of graphs. The experiments show effectiveness of RGIB. This paper i...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work find that the inherent edge noise can perturb the graph topology and labels, which may reduce link prediction performance. Thus, the authors propose an information bottleneck guided method, namely RGIB. RGIB achieves robustness representation of graphs. The experiments show effectiveness of RGIB. This...
The paper presents a method for answering questions about scenarios -- questions for which there isn’t a fixed answer but is varied depending on additional conditions that are unstated in the text. Method: Given a scenario, a question, and a set of conditions (extracted from an input text) that are to be considered w...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents a method for answering questions about scenarios -- questions for which there isn’t a fixed answer but is varied depending on additional conditions that are unstated in the text. Method: Given a scenario, a question, and a set of conditions (extracted from an input text) that are to be cons...
This paper studies the problem of observationally robust reinforcement learning. In this setting, an agent learns via interacting with an MDP, but is evaluated in a version of this MDP where noise is added to the observed states by applying an (unknown) “noise kernel” $T: X \rightarrow \Delta(X)$; this evaluation envi...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the problem of observationally robust reinforcement learning. In this setting, an agent learns via interacting with an MDP, but is evaluated in a version of this MDP where noise is added to the observed states by applying an (unknown) “noise kernel” $T: X \rightarrow \Delta(X)$; this evaluat...
This paper proposes a defensive mechanism, Cognitive Distillation (CD), that extracts potential trigger patterns (and a mask) responsible for the model's prediction. The paper starts with the observation that, even if the trigger pattern consists of multiple pixels and/or is complex, a few pixel values correspond to ac...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a defensive mechanism, Cognitive Distillation (CD), that extracts potential trigger patterns (and a mask) responsible for the model's prediction. The paper starts with the observation that, even if the trigger pattern consists of multiple pixels and/or is complex, a few pixel values correspo...
This paper incorporates the lateral inhibition of neuron into the Deep network for image classification. Results show the effectiveness of the proposed mechanism. The lateral inhibition is not novel but the idea to explore the usability of the lateral inhibition into artificial neural network is interesting. The key w...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper incorporates the lateral inhibition of neuron into the Deep network for image classification. Results show the effectiveness of the proposed mechanism. The lateral inhibition is not novel but the idea to explore the usability of the lateral inhibition into artificial neural network is interesting. T...
A contrastive learning based self-supervised visual representation learning method is proposed in this paper. Upon observing the positive and negative assignments of training samples, this paper proposes to construct a candidate neighbor set where top candidate neighbors are selected as positive samples. When contribut...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: A contrastive learning based self-supervised visual representation learning method is proposed in this paper. Upon observing the positive and negative assignments of training samples, this paper proposes to construct a candidate neighbor set where top candidate neighbors are selected as positive samples. When c...
This paper studies the exploration problem for MDPs with transition kernel captured by general function classes with bounded covering number. The transition dynamics is assumed to be a low-rank Gaussian RBF kernel with unknown features of states and actions, which captures both linear MDPs and a number of neural nets a...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the exploration problem for MDPs with transition kernel captured by general function classes with bounded covering number. The transition dynamics is assumed to be a low-rank Gaussian RBF kernel with unknown features of states and actions, which captures both linear MDPs and a number of neura...
The paper proposes a method for solving the quantum many body problem using neural networks as variational ansatz. In contrast to more standard approaches which model the wave function directly, authors use a score-based model (hence modeling the gradient of the underlying probability distribution associated with the w...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper proposes a method for solving the quantum many body problem using neural networks as variational ansatz. In contrast to more standard approaches which model the wave function directly, authors use a score-based model (hence modeling the gradient of the underlying probability distribution associated wi...
This paper proposes a contrastive learning method to balance the learning of positive and negative tokens in text generation tasks (e.g., language modeling and open-domain dialogue generation tasks). Strength 1. The paper is easy to follow and the idea is intuitive. 2. Several case studies are given which are enc...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This paper proposes a contrastive learning method to balance the learning of positive and negative tokens in text generation tasks (e.g., language modeling and open-domain dialogue generation tasks). Strength 1. The paper is easy to follow and the idea is intuitive. 2. Several case studies are given which...
This paper proposes a new data-augmentation strategy based on curriculum learning for long-tail problems. The key idea is to estimate the appropriate strength of data augmentation needed for each class during training. The proposed method was evaluated on widely used datasets and achieved favorable performance. ## Stre...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new data-augmentation strategy based on curriculum learning for long-tail problems. The key idea is to estimate the appropriate strength of data augmentation needed for each class during training. The proposed method was evaluated on widely used datasets and achieved favorable performance....
This work provides a formal definition for robustness based on learning theoretical terms. Specifically, holomorphicity enables complexity analysis tool to investigate the phenomenon of adversarial examples. In addition, the analysis provides a geometrical interpretation for these phenomena. Strength: It is novel to s...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work provides a formal definition for robustness based on learning theoretical terms. Specifically, holomorphicity enables complexity analysis tool to investigate the phenomenon of adversarial examples. In addition, the analysis provides a geometrical interpretation for these phenomena. Strength: It is no...
An RL algorithm, TAM, is proposed to solve large-scale VRP problems. TAM learns how to split the problem into smaller problems and then solve the smaller problems which are instances of TSP by well-performed heuristic algorithms like LKH3. The splitting problem is defined as an RL problem in which the state of the sys...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: An RL algorithm, TAM, is proposed to solve large-scale VRP problems. TAM learns how to split the problem into smaller problems and then solve the smaller problems which are instances of TSP by well-performed heuristic algorithms like LKH3. The splitting problem is defined as an RL problem in which the state of...
This work provides a convergence guarantee for using a score-based diffusion model to sample from an arbitrary distribution. The method has significantly looser assumptions than previous work, and accounts for three sources of error: (1) L2 score estimation error, (2) discretization of the reverse SDE sampling algorith...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This work provides a convergence guarantee for using a score-based diffusion model to sample from an arbitrary distribution. The method has significantly looser assumptions than previous work, and accounts for three sources of error: (1) L2 score estimation error, (2) discretization of the reverse SDE sampling ...
The paper aims to analyze the Denoising effect provided by GNNs (on corrupted graph data, typically with noisy labels). It recalls the Neumann Graph Convolution (NGC) perspective and how it connects with Graph Signal Denoising (GSD). It defines the High-order Graph Connectivity Factor, \tau, which is a measure of the...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper aims to analyze the Denoising effect provided by GNNs (on corrupted graph data, typically with noisy labels). It recalls the Neumann Graph Convolution (NGC) perspective and how it connects with Graph Signal Denoising (GSD). It defines the High-order Graph Connectivity Factor, \tau, which is a measur...
This paper introduces a new paradigm to cluster incomplete vectors using subspaces as proxies to exploit the geometry of the Grassmannian. The authors leverage this new perspective to develop an algorithm to cluster and complete data in a union of subspaces via a fusion penalty formulation. The analysis with synthetic ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper introduces a new paradigm to cluster incomplete vectors using subspaces as proxies to exploit the geometry of the Grassmannian. The authors leverage this new perspective to develop an algorithm to cluster and complete data in a union of subspaces via a fusion penalty formulation. The analysis with sy...
This paper aims to guarantee the performance fairness among different clients in federated learning, while protecting the overall average performance from being sacrificed. The authors propose two algorithms based on variance reduction (in terms of all clients) and semi-variance reduction (in terms of the worst-off cli...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper aims to guarantee the performance fairness among different clients in federated learning, while protecting the overall average performance from being sacrificed. The authors propose two algorithms based on variance reduction (in terms of all clients) and semi-variance reduction (in terms of the worst...
This paper aims to investigate the role of regularization used in existing GNN explainers. The paper analyzes the interpretable GNN model proposed by Miao et al. 2022 by rewriting the training objective function and mapping it into two parts, the feature attribution objective and the feature selection objective. And th...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: This paper aims to investigate the role of regularization used in existing GNN explainers. The paper analyzes the interpretable GNN model proposed by Miao et al. 2022 by rewriting the training objective function and mapping it into two parts, the feature attribution objective and the feature selection objective...
This paper proposes an algorithm that solves bilevel optimization problem in an asynchronous distributed manner. The iteration complexity for the algorithm to obtain $\epsilon$-stationary point is upper bounded by $\mathcal{O}(\frac{1}{\epsilon^2})$. Empirical results show that under asynchronous setting, the proposed ...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes an algorithm that solves bilevel optimization problem in an asynchronous distributed manner. The iteration complexity for the algorithm to obtain $\epsilon$-stationary point is upper bounded by $\mathcal{O}(\frac{1}{\epsilon^2})$. Empirical results show that under asynchronous setting, the p...
This paper proposes a new method to deal with backdoor attacks problems. Existing backdoor attack algorithms could not deal with high corruption ratio well, the proposed method, however, leverage on the noisy label defense algorithm to develop a robust version of backdoor defense. By applying adversarial learning on se...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new method to deal with backdoor attacks problems. Existing backdoor attack algorithms could not deal with high corruption ratio well, the proposed method, however, leverage on the noisy label defense algorithm to develop a robust version of backdoor defense. By applying adversarial learni...
The paper introduces a novel face interaction mechanism for GNN-based rigid body simulators to solve the penetration issue existing when the contact point is far from nodes. The new framework significantly increases the accuracy of contact resolution and enables the simulation of complex shapes. The framework is even m...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper introduces a novel face interaction mechanism for GNN-based rigid body simulators to solve the penetration issue existing when the contact point is far from nodes. The new framework significantly increases the accuracy of contact resolution and enables the simulation of complex shapes. The framework i...
The authors propose an uncertainty aware active learning domain adaptation based on imposing a dirichlet prior over model predictions. This allows them to get a distribution over model predictions, which in turn enables them to compute 2 kinds of uncertainty measures for each target datapoint. 1) target distribution un...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors propose an uncertainty aware active learning domain adaptation based on imposing a dirichlet prior over model predictions. This allows them to get a distribution over model predictions, which in turn enables them to compute 2 kinds of uncertainty measures for each target datapoint. 1) target distrib...
This paper proposes a training-free two-stage method based on Codex model for QA tasks. The first stage uses few-shot learning to prompt Codex and converts a natural language question to a SQL or Python program, where difficult-to-resolve parts are represented as API calls. In the second stage, Codex is prompted again ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a training-free two-stage method based on Codex model for QA tasks. The first stage uses few-shot learning to prompt Codex and converts a natural language question to a SQL or Python program, where difficult-to-resolve parts are represented as API calls. In the second stage, Codex is prompte...
In this paper, the authors claim that existing distance metric learning methods update the distance metric based on local relationships and ignore the more global ones (however, this claim is not right). The authors argue that updating distance metric based on the closest class centers has some drawbacks and they claim...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this paper, the authors claim that existing distance metric learning methods update the distance metric based on local relationships and ignore the more global ones (however, this claim is not right). The authors argue that updating distance metric based on the closest class centers has some drawbacks and th...