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
This paper introduce a new metric (ROSCOE) for the generic step-by-step reasoning task. In particular, ROSCOE includes four fine-grained metrics under four perspectives: semantic alignment, semantic similarity, logical inference and language coherence, where each fine-grained metric further contains a bunch of ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduce a new metric (ROSCOE) for the generic step-by-step reasoning task. In particular, ROSCOE includes four fine-grained metrics under four perspectives: semantic alignment, semantic similarity, logical inference and language coherence, where each fine-grained metric further contains a b...
This paper can be thought of as two (related) papers: 1. A proposed benchmark for learning constraints from demonstrations in continuous (approximately deterministic) domains. Here constraint is taken to be an assertion that sum of unknown random variables (either in expectation or per episode) is less than a particul...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper can be thought of as two (related) papers: 1. A proposed benchmark for learning constraints from demonstrations in continuous (approximately deterministic) domains. Here constraint is taken to be an assertion that sum of unknown random variables (either in expectation or per episode) is less than a ...
This work considers the problem of reconstructing sparse vectors from an underdetermined system of equations when the measurements are commuted. First, it gives a lower bound on the required number of measurements given a certain SNR. Then, it analyses an estimator which gives corresponding upper bounds (depending on t...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This work considers the problem of reconstructing sparse vectors from an underdetermined system of equations when the measurements are commuted. First, it gives a lower bound on the required number of measurements given a certain SNR. Then, it analyses an estimator which gives corresponding upper bounds (depend...
The paper learns a policy to automatically re-mesh the current mesh used in physics simulation, so that delicate regions where rich dynamics happens can leverage more computational power. It uses reinforcement learning to train the policy end-to-end, with the reward being a combination of simulation accuracy and speed....	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper learns a policy to automatically re-mesh the current mesh used in physics simulation, so that delicate regions where rich dynamics happens can leverage more computational power. It uses reinforcement learning to train the policy end-to-end, with the reward being a combination of simulation accuracy an...
The paper proposes to quantize ( two values) the updates of the adam optimizer. The approach leverages the bounded nature of adam updates to design a unbiased quantization scheme - it uses the value of the update to define a bernoulli distribution (+1, -1). They show improvements in speed in several experiments. Som...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes to quantize ( two values) the updates of the adam optimizer. The approach leverages the bounded nature of adam updates to design a unbiased quantization scheme - it uses the value of the update to define a bernoulli distribution (+1, -1). They show improvements in speed in several experiment...
The paper proposes a RL-based approach for online weighted bipartite matching. A key novelty of the paper is to augment the decisions of the RL agent with a classic online algorithm to obtain robustness guarantees. Upon the arrival of an online vertex, the algorithm queries both a base online algorithm (called expert) ...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a RL-based approach for online weighted bipartite matching. A key novelty of the paper is to augment the decisions of the RL agent with a classic online algorithm to obtain robustness guarantees. Upon the arrival of an online vertex, the algorithm queries both a base online algorithm (called ...
This submission proposes a minibatch stochastic three points method, which only requires an approximation of the objective function at each iteration. It is an extension of a previous work of stochastic three points method that requires the exact evaluation of objective functions. Strengthes: Compared to STP, evaluatin...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This submission proposes a minibatch stochastic three points method, which only requires an approximation of the objective function at each iteration. It is an extension of a previous work of stochastic three points method that requires the exact evaluation of objective functions. Strengthes: Compared to STP, e...
The paper studies RL in linear mixture MDP with unknown transition, adversarial loss, and bandit feedback. This paper gives the first algorithm in this setting whose regret scales with sqrt(#episode) and supplements a lower bound, which matches the upper bound for the $d$ and $K$ dependency, where $K$ is #epsiode and...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies RL in linear mixture MDP with unknown transition, adversarial loss, and bandit feedback. This paper gives the first algorithm in this setting whose regret scales with sqrt(#episode) and supplements a lower bound, which matches the upper bound for the $d$ and $K$ dependency, where $K$ is #eps...
This paper provides a detailed and comprehensive theoretical analysis to the problem that existing DRO approaches do not bring significant performance gain over ERM. The analysis focus on the GRW framework that minimizes the weighted empirical risk. The authors show that for linear models and wide neural networks, the ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper provides a detailed and comprehensive theoretical analysis to the problem that existing DRO approaches do not bring significant performance gain over ERM. The analysis focus on the GRW framework that minimizes the weighted empirical risk. The authors show that for linear models and wide neural networ...
This paper provides a theoretical analysis on the robustness of explainers. Here, the robustness is measured as the difference between the two point $x$ and $x'$ in a close neighborhood. If explanations for $x$ and $x'$ largely differ, the explainer is deemed to be not robust. For the analysis, the authors adopted the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper provides a theoretical analysis on the robustness of explainers. Here, the robustness is measured as the difference between the two point $x$ and $x'$ in a close neighborhood. If explanations for $x$ and $x'$ largely differ, the explainer is deemed to be not robust. For the analysis, the authors adop...
The paper focuses on the problem of combining replay models and regularization techniques for task-incremental continual learning (TIL). In the words of the authors, "we design a prior that provably gives better gradient reconstructions by utilizing two types of replay and a quadratic weight-regularizer". The concept o...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper focuses on the problem of combining replay models and regularization techniques for task-incremental continual learning (TIL). In the words of the authors, "we design a prior that provably gives better gradient reconstructions by utilizing two types of replay and a quadratic weight-regularizer". The c...
In this paper authors propose to use two separate networks, one that is trained to detect hate speech and the other that is trained to detect sarcasm. The final sarcasm detection score is then a combination of the outputs of both networks. Strengths: - Good idea to try to separate out hate from sarcasm via separate ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper authors propose to use two separate networks, one that is trained to detect hate speech and the other that is trained to detect sarcasm. The final sarcasm detection score is then a combination of the outputs of both networks. Strengths: - Good idea to try to separate out hate from sarcasm via s...
In this paper, the authors study the performance of transformer models on downstream tasks as the total computational budget is decreased. This process, known as cramming in the paper, turns the problem of training these enormous language models in a new direction from the typical scenario used in industrial labs that...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors study the performance of transformer models on downstream tasks as the total computational budget is decreased. This process, known as cramming in the paper, turns the problem of training these enormous language models in a new direction from the typical scenario used in industrial l...
The paper proposes new algorithms for CCA and Partial Least Squares problems based on game-theoretic formulations, following previous work by Gemp et al. The proposed algorithm relaxes some constraints and works well empirically. While the proposed algorithm seems to be promising empirically, its description and motiva...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper proposes new algorithms for CCA and Partial Least Squares problems based on game-theoretic formulations, following previous work by Gemp et al. The proposed algorithm relaxes some constraints and works well empirically. While the proposed algorithm seems to be promising empirically, its description an...
This paper proposes a new out-of-distribution detection method that is based on supervised contrastive learning in a hyperspherical space. The core component of the proposed method is a novel regularizer that encourages the separation of class centroids. The resulting hyperspherical space is highly effective in distan...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a new out-of-distribution detection method that is based on supervised contrastive learning in a hyperspherical space. The core component of the proposed method is a novel regularizer that encourages the separation of class centroids. The resulting hyperspherical space is highly effective i...
In this paper, the authors focus on the seeded graph matching problem, which is a variant of the graph matching problem with several pre-matched pairs as initial seeds. In existing works, the seeds are hardly considered especially in supervised graph matching works. They propose to apply the GNN jointly over both the t...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors focus on the seeded graph matching problem, which is a variant of the graph matching problem with several pre-matched pairs as initial seeds. In existing works, the seeds are hardly considered especially in supervised graph matching works. They propose to apply the GNN jointly over bo...
This paper discusses the impact of multiple objectives. Pareto optimality guarantees are discussed, and actor-critic formulation is provided. The paper analyses the dynamics of the induced value functions resultant from policy alterations in MORL problems. The analysis seems good. The proposed algorithm mainly adds ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper discusses the impact of multiple objectives. Pareto optimality guarantees are discussed, and actor-critic formulation is provided. The paper analyses the dynamics of the induced value functions resultant from policy alterations in MORL problems. The analysis seems good. The proposed algorithm main...
The paper presents a theoretical analysis, hoping to show that one-step RL is equivalent to a certain type of critic regularization. Strength: - The general idea of discovering equivalence between methods, which look like arbitrary choices, is quite interesting (and can be useful as it clarifies false expectations fo...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper presents a theoretical analysis, hoping to show that one-step RL is equivalent to a certain type of critic regularization. Strength: - The general idea of discovering equivalence between methods, which look like arbitrary choices, is quite interesting (and can be useful as it clarifies false expecta...
This paper proposes a classifier that performs well under the shift of class proportions. The approach involves a distributionally robust formulation (a min-max problem) where the weight vector is constrained in a $f$-divergence ball of radius $\delta$. The optimal post-hoc classifier is shown to be a reweighed classif...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a classifier that performs well under the shift of class proportions. The approach involves a distributionally robust formulation (a min-max problem) where the weight vector is constrained in a $f$-divergence ball of radius $\delta$. The optimal post-hoc classifier is shown to be a reweighed...
This paper proposes a retrieval augmented diffusion model for the task of text to image synthesis. The idea is to condition the diffusion model on the external knowledge base in addition to the text embeddings that are used in text-to-image diffusion models. Conditioning on the retrieved nearest neighbor embeddings hap...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a retrieval augmented diffusion model for the task of text to image synthesis. The idea is to condition the diffusion model on the external knowledge base in addition to the text embeddings that are used in text-to-image diffusion models. Conditioning on the retrieved nearest neighbor embedd...
This paper presents an approach to use MAML-based meta learning to learn a neural operator that approximates complex PDE-based equations with different physical parameters. Instead of optimizing all parameters of the base model during inner optimization (like vanilla MAML) or only the last layer (like ANIL), this paper...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents an approach to use MAML-based meta learning to learn a neural operator that approximates complex PDE-based equations with different physical parameters. Instead of optimizing all parameters of the base model during inner optimization (like vanilla MAML) or only the last layer (like ANIL), th...
The paper introduces a method to analyze the vulnerabilities of deep reinforcement learning policies. Experiments are conducted to illustrate the non-robust features, and show how adversarial attack techniques and robust training affect these features. Some questions or suggestions about the paper: 1. The authors sh...	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 introduces a method to analyze the vulnerabilities of deep reinforcement learning policies. Experiments are conducted to illustrate the non-robust features, and show how adversarial attack techniques and robust training affect these features. Some questions or suggestions about the paper: 1. The au...
The paper explores different embedding extraction techniques in Graph Neural Networks (GNNs), and proposes Graph-connected Network (GraNet) layers, claiming that this approach improves the accuracy compared to traditional GNNs. - Weaknesses: -- it's not clear to me what is the main goal or contribution of the paper. ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper explores different embedding extraction techniques in Graph Neural Networks (GNNs), and proposes Graph-connected Network (GraNet) layers, claiming that this approach improves the accuracy compared to traditional GNNs. - Weaknesses: -- it's not clear to me what is the main goal or contribution of the...
This paper studied the problem of solving federated compositional pairwise risk minimization. This problem has several applications in AUROC maximization with a pairwise loss, partial AUROC maximization with a compositional loss, etc. The main challenge of solving this problem stems from the non-decomposability of the ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper studied the problem of solving federated compositional pairwise risk minimization. This problem has several applications in AUROC maximization with a pairwise loss, partial AUROC maximization with a compositional loss, etc. The main challenge of solving this problem stems from the non-decomposability...
The authors propose FEDDC, a new scheme for training federated learning systems. The main idea is that at each communication round a client can either receive the model of another random client (daisy chaining round) or an aggregated model (aggregation round). The authors show convergence rate and generalization bound ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose FEDDC, a new scheme for training federated learning systems. The main idea is that at each communication round a client can either receive the model of another random client (daisy chaining round) or an aggregated model (aggregation round). The authors show convergence rate and generalizatio...
In this paper, the authors transform the model hijacking attack into a more general multimodal settings, where the hijacking and original tasks are performed on data of different modalities. Specifically, they focus on the setting where an adversary implements a natural language processing (NLP) hijacking task into an ...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this paper, the authors transform the model hijacking attack into a more general multimodal settings, where the hijacking and original tasks are performed on data of different modalities. Specifically, they focus on the setting where an adversary implements a natural language processing (NLP) hijacking task ...
This paper proposed a weakly supervised method for oriented object detection by leveraging only horizontal bounding boxes. The network is able to learn to predict the rotated boxes by building the transition functions between a WS branch on original images and a SS branch on rotated images. By building upon the FCOS d...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed a weakly supervised method for oriented object detection by leveraging only horizontal bounding boxes. The network is able to learn to predict the rotated boxes by building the transition functions between a WS branch on original images and a SS branch on rotated images. By building upon th...
This paper introduces a new approach (MDPO) to model-based decentralized reinforcement learning by leveraging a latent variable function $\psi(z\|o)$ to help learn the observation transition $P_i(o_i'\|o_i,a_i,z_i)$ and reward $R_i(o_i,a_i,z_i)$ models. The main problem this paper tries to address is that the assumptio...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper introduces a new approach (MDPO) to model-based decentralized reinforcement learning by leveraging a latent variable function $\psi(z\|o)$ to help learn the observation transition $P_i(o_i'\|o_i,a_i,z_i)$ and reward $R_i(o_i,a_i,z_i)$ models. The main problem this paper tries to address is that the a...
The paper proposes a new goal selection algorithm for training curricula of goal-oriented reinforcement learning agents. The algorithm, called query the agent, estimates epistemic uncertainty of the agent in the state space and sets goals in areas with higher epistemic uncertainty. The algorithm is evaluated on a set o...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a new goal selection algorithm for training curricula of goal-oriented reinforcement learning agents. The algorithm, called query the agent, estimates epistemic uncertainty of the agent in the state space and sets goals in areas with higher epistemic uncertainty. The algorithm is evaluated on...
This paper considers supervised learning in the setting where the data is drawn in a 2 stage process: first, a domain is selected, second, a batch of data is drawn from this domain. The observations include the batch index, but not the underlying domain the data was collected from. The learning problem considered here ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper considers supervised learning in the setting where the data is drawn in a 2 stage process: first, a domain is selected, second, a batch of data is drawn from this domain. The observations include the batch index, but not the underlying domain the data was collected from. The learning problem consider...
The paper proposes a method for an ILP-style rule-learning problem by using gradient descent over differentiable approximations to the various connectives in the language under consideration. The method is empirically tested on a suite of ILP tasks, the CLUTRR benchmark, and some knowledge graph completion tasks. The m...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper proposes a method for an ILP-style rule-learning problem by using gradient descent over differentiable approximations to the various connectives in the language under consideration. The method is empirically tested on a suite of ILP tasks, the CLUTRR benchmark, and some knowledge graph completion task...
This paper addresses the problem of the aspect-ratio shift between the base and novel classes in few-shot object detection resulting in the poor box proposals of the RPN in a two-stage detector. They propose using multiple classifiers instead of one as in the current RPN. Their experiments show significant improvements...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper addresses the problem of the aspect-ratio shift between the base and novel classes in few-shot object detection resulting in the poor box proposals of the RPN in a two-stage detector. They propose using multiple classifiers instead of one as in the current RPN. Their experiments show significant impr...
This paper proposes a neural diffusion-reaction process model, which can take into account the timestamps and model the temporal embedding of participants behind high-order interactions, compared with existing related tensor factorization methods. - To that end, they build the model upon the ODE framework that can han...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a neural diffusion-reaction process model, which can take into account the timestamps and model the temporal embedding of participants behind high-order interactions, compared with existing related tensor factorization methods. - To that end, they build the model upon the ODE framework that...
This paper study an interesting problem that improves user engagement in sequential recommendations. Considering that current rules of taking a break are manually settled, this paper propose to learning a user dependent and stationary policy to decide when to take a break. Lotka-Volterra dynamics is used to model two p...	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 study an interesting problem that improves user engagement in sequential recommendations. Considering that current rules of taking a break are manually settled, this paper propose to learning a user dependent and stationary policy to decide when to take a break. Lotka-Volterra dynamics is used to mod...
This paper introduces a benchmark for evaluating vision-based dynamics models. The paper shows that common prediction metrics (typically revolving around predicted image quality and not physical plausibility or causal interaction effects) may not necessarily correlate with control performance, and thus proposes an eval...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper introduces a benchmark for evaluating vision-based dynamics models. The paper shows that common prediction metrics (typically revolving around predicted image quality and not physical plausibility or causal interaction effects) may not necessarily correlate with control performance, and thus proposes...
The paper considers the (federated) private model personalization problem, and provides a utility guarantee in the bilinear case. The technical contribution is to prove a utility bound when using DPSGD to learn the shared representation, which improves upon the best known bound (that uses sufficient statistics perturba...	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 considers the (federated) private model personalization problem, and provides a utility guarantee in the bilinear case. The technical contribution is to prove a utility bound when using DPSGD to learn the shared representation, which improves upon the best known bound (that uses sufficient statistics ...
This paper proposes a new neural attention model called "ChordMixer" to enable long-range interactions for sequences with variable lengths. The proposal takes inspiration from the Chord protocol in P2P networks. The ChordMixer consists of a series of ChordMixer blocks where in each block, an input token with $d/M$ dime...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a new neural attention model called "ChordMixer" to enable long-range interactions for sequences with variable lengths. The proposal takes inspiration from the Chord protocol in P2P networks. The ChordMixer consists of a series of ChordMixer blocks where in each block, an input token with $d...
This work proposes an approach for indoor localisation utilising RSSI and accelerometer data based on transformers with CRF layer an alternating loss function. Its contributions are mostly in the specific application (indoor localisation) rather than the methods. S + important ML application + solid analysis of the da...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work proposes an approach for indoor localisation utilising RSSI and accelerometer data based on transformers with CRF layer an alternating loss function. Its contributions are mostly in the specific application (indoor localisation) rather than the methods. S + important ML application + solid analysis o...
The paper studies the limits of the adversarial robustness that can be obtained by combining $M$ different classifiers into one randomized classifier that at inference time, chooses one of the elements with certain probability. Some intermediate lower and upper bounds are presented that illustrate the main intuition be...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The paper studies the limits of the adversarial robustness that can be obtained by combining $M$ different classifiers into one randomized classifier that at inference time, chooses one of the elements with certain probability. Some intermediate lower and upper bounds are presented that illustrate the main intu...
The authors introduce a feature generation and model boosting methodology, dubbed "OpenFE". OpenFE proceeds by 1. Generating a set of possible features, by ways of atomic operations (elementary operations, as well as "grouped" operations that apply mean/min/max operations over subsets of the data corresponding to th...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The authors introduce a feature generation and model boosting methodology, dubbed "OpenFE". OpenFE proceeds by 1. Generating a set of possible features, by ways of atomic operations (elementary operations, as well as "grouped" operations that apply mean/min/max operations over subsets of the data correspondi...