ICML 2023 Papers — Page 14
International Conference on Machine Learning · 1828 papers
Probabilistic Contrastive Learning Recovers the Correct Aleatoric Uncertainty of Ambiguous Inputs
Michael Kirchhof (University of Tübingen), Seong Joon Oh (University of Tübingen)
Data SynthesisRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes a probabilistic contrastive learning framework called MCInfoNCE, which can learn the posterior distribution of the input (i.e., uncertainty) using only contrastive supervision, and proves that it can recover the true posterior (mean and variance) to be rotation invariant in non-injective, randomly generated processes;
Probabilistic Imputation for Time-series Classification with Missing Data
SeungHyun Kim (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerAuto EncoderTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: A probability generative framework based on the MNAR assumption, supnotMIWAE, is proposed for classifying missing data in multivariate time series, and the quality of missing value imputation is improved through ObsDropout regularization.
Probabilistic Unrolling: Scalable, Inverse-Free Maximum Likelihood Estimation for Latent Gaussian Models
Alexander Lin (Harvard University), Demba E. Ba
Recommendation SystemOptimizationTabular
🎯 What it does: This paper proposes a framework called Probabilistic Unrolling to accelerate maximum likelihood estimation using gradient EM in high-dimensional latent Gaussian models (LGM). The core idea is to replace the direct inversion of the covariance matrix with Monte Carlo sampling and to solve the posterior mean and samples using iterative linear solvers (such as conjugate gradient). The iterative process is then backpropagated to obtain more accurate gradients.
Probably Anytime-Safe Stochastic Combinatorial Semi-Bandits
Yunlong Hou (National University of Singapore), Zixin Zhong (University of Alberta)
OptimizationReinforcement LearningTabular
🎯 What it does: Proposed the 'potentially safe at any time' random combination semi-bandit problem, which requires minimizing the loss of cumulative rewards while ensuring that the variance of each step's choice does not exceed a threshold;
Progressive Purification for Instance-Dependent Partial Label Learning
Ning Xu (Southeast University), Xin Geng (Southeast University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A progressive purification method named POP (PrOgressive Purification) is proposed for instance-dependent partial label learning (PLL), which gradually removes incorrect candidate labels based on the current model reliability threshold during each training epoch and retrains the model with the cleaned labels.
Projected Tensor Power Method for Hypergraph Community Recovery
Jinxin Wang (Chinese University of Hong Kong), Anthony Man-Cho So (Chinese University of Hong Kong)
Graph
🎯 What it does: A projection tensor power iteration method (PTPM) is proposed for achieving accurate community recovery in the symmetric d-uniform hypergraph stochastic block model (d-HSBM).
Prometheus: Taming Sample and Communication Complexities in Constrained Decentralized Stochastic Bilevel Learning
Zhuqing Liu (Ohio State University), Jia Liu (Ohio State University)
OptimizationMeta LearningImage
🎯 What it does: An algorithm named Prometheus is proposed to solve constrained decentralized stochastic bilevel optimization problems, aiming to reduce sample and communication complexity.
PromptBoosting: Black-Box Text Classification with Ten Forward Passes
Bairu Hou (University of California Santa Barbara), Yang Zhang (Massachusetts Institute of Technology)
ClassificationTransformerPrompt EngineeringText
🎯 What it does: A black-box text classification method called PROMPTBOOSTING is proposed, which can build a strong classifier with only ten forward propagations.
Prompting Large Language Model for Machine Translation: A Case Study
Biao Zhang (Google DeepMind), Alexandra Birch (University of Edinburgh)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The system evaluates the prompting strategies of large language models in machine translation, including template selection, demonstration example selection, monolingual data utilization, and cross-domain transfer.
Propensity Matters: Measuring and Enhancing Balancing for Recommendation
Haoxuan Li (Peking University), Peng Cui (Tsinghua University)
Recommendation SystemTabular
🎯 What it does: This paper studies the balance issue of propensity estimation in recommendation systems, proposes the BMSE metric to measure propensity balance, and incorporates BMSE regularization into the IPS/DR estimators, resulting in two more robust unbiased loss estimators: IPS-V2 and DR-V2.
Proper Losses for Discrete Generative Models
Dhamma Kimpara (University of Colorado Boulder), Bo Waggoner (University of Colorado Boulder)
GenerationData SynthesisGenerative Adversarial Network
🎯 What it does: The study investigates how to design a 'truly correct' loss function in the evaluation of discrete generative models that can only be sampled.
Proper Scoring Rules for Survival Analysis
Hiroki Yanagisawa (IBM Research)
TabularTime Series
🎯 What it does: This paper conducts theoretical expansion and empirical validation of proper scoring rules in survival analysis, proving the appropriateness of four common rules (Pinball, Logarithmic, Brier, Ranked Probability Score) in survival contexts, and constructs loss functions and evaluation metrics based on these rules.
Properties of the Mallows Model Depending on the Number of Alternatives: A Warning for an Experimentalist
Niclas Boehmer (Technische Universität Berlin), Sonja Kraiczy (University of Oxford)
Tabular
🎯 What it does: This study investigates the statistical properties of the classic Mallows model and its normalized variant as the number of candidates increases, providing theoretical proofs and experimental validations; it issues a warning to experimenters, suggesting the use of the normalized model.
Protecting Language Generation Models via Invisible Watermarking
Xuandong Zhao (University of California Santa Barbara), Lei Li (University of California Santa Barbara)
GenerationSafty and PrivacyAdversarial AttackText
🎯 What it does: This paper proposes an invisible sequence watermarking method called GINSEW to protect language generation models from model extraction attacks.
Prototype-oriented unsupervised anomaly detection for multivariate time series
Yuxin Li (Xidian University), Mingyuan Zhou (University of Texas at Austin)
Anomaly DetectionMeta LearningTransformerAuto EncoderTime Series
🎯 What it does: This paper proposes a prototype-based unsupervised multivariate time series anomaly detection method called PUAD, which utilizes a Transformer combined with a probabilistic generative framework, and learns multiple normal patterns through prototypes and Optimal Transport.
Prototype-Sample Relation Distillation: Towards Replay-Free Continual Learning
Nader Asadi (Concordia University), Eugene Belilovsky (Concordia University)
ClassificationKnowledge DistillationRepresentation LearningContrastive LearningImage
🎯 What it does: A continuous learning method is proposed that does not use a replay buffer, utilizing Prototype-Sample Relation Distillation (PRD) to maintain the relative similarity between old class prototypes and new task samples, thereby suppressing forgetting and enhancing adaptability.
ProtST: Multi-Modality Learning of Protein Sequences and Biomedical Texts
Minghao Xu (Mila Quebec AI Institute), Jian Tang (HEC Montreal)
ClassificationRetrievalRepresentation LearningDrug DiscoveryTransformerLarge Language ModelContrastive LearningTextMultimodalityBiomedical Data
🎯 What it does: The ProtST framework is proposed, which performs multimodal pre-training by aligning protein sequences with their biomedical text attribute descriptions to enhance the representational capacity of protein language models.
Provable Benefit of Mixup for Finding Optimal Decision Boundaries
Junsoo Oh (Korea Advanced Institute of Science and Technology), Chulhee Yun (Korea Advanced Institute of Science and Technology)
ClassificationOptimizationData-Centric Learning
🎯 What it does: This study investigates how data augmentation techniques such as Mixup affect the sample complexity of finding the optimal decision boundary in binary linear classification problems.
Provable Data Subset Selection For Efficient Neural Networks Training
Murad Tukan (DataHeroes), Dan Feldman (University of Haifa)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a coreset construction algorithm for RBF neural networks and extends it to a data subset selection method applicable for training arbitrary deep neural networks.
Provable Dynamic Fusion for Low-Quality Multimodal Data
Qingyang Zhang (Tianjin University), Xi Peng (Sichuan University)
ClassificationRecognitionMultimodality
🎯 What it does: A dynamic multimodal fusion method QMF based on uncertainty estimation is proposed, and its superiority is proven from the perspective of generalization error theory.
Provable Multi-instance Deep AUC Maximization with Stochastic Pooling
Dixian Zhu (University of Iowa), Tianbao Yang (Texas A&M University)
ClassificationOptimizationImageBiomedical DataMagnetic Resonance ImagingComputed TomographyStochastic Differential Equation
🎯 What it does: This paper proposes a multi-instance deep AUC maximization algorithm called MIDAM based on Variational Randomized Stochastic Pooling (VRSP), addressing the issue of insufficient GPU memory during large-scale multi-instance batch training.
Provable Reset-free Reinforcement Learning by No-Regret Reduction
Hoai-An Nguyen (Rutgers University), Ching-An Cheng (Microsoft Research)
Reinforcement LearningTabular
🎯 What it does: A general regret-free subtraction framework is proposed, systematically designing reset-free reinforcement learning algorithms, transforming the reset-free reinforcement learning problem into a two-player game.
Provably and Practically Efficient Neural Contextual Bandits
Sudeep Salgia (Cornell University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper studies neural contextual bandits, proposing non-asymptotic error bounds for any smooth activation function and sublinear return upper bounds, and presents the NeuralGCB algorithm based on two-sided confidence bounds;
Provably Convergent Schrödinger Bridge with Applications to Probabilistic Time Series Imputation
Yu Chen (Morgan Stanley), Yuriy Nevmyvaka (Morgan Stanley)
RestorationData SynthesisTransformerTime SeriesSequentialBiomedical DataStochastic Differential Equation
🎯 What it does: This paper proposes an Approximate Projection-based Schrödinger Bridge (aIPF) algorithm and provides a convergence analysis; subsequently, the algorithm is applied to probabilistic time series missing value imputation (CSBI), achieving high-quality imputation for randomly missing positions.
Provably Efficient Offline Reinforcement Learning with Perturbed Data Sources
Chengshuai Shi (University of Virginia), Jing Yang (Pennsylvania State University)
Reinforcement LearningTabular
🎯 What it does: The research focuses on the target task MDP where the data comes from multiple sources of randomly perturbed versions of offline reinforcement learning, and proposes the HetPEVI algorithm for effective learning.
Provably Efficient Representation Learning with Tractable Planning in Low-Rank POMDP
Jiacheng Guo (Princeton University), Xuezhou Zhang (Princeton University)
OptimizationRepresentation LearningTabularBenchmark
🎯 What it does: A representation learning algorithm utilizing maximum likelihood estimation and optimistic planning is proposed to address the efficient learning and planning problem of low-rank POMDPs.
Provably Invariant Learning without Domain Information
Xiaoyu Tan (INF Technology), Yuan Qi (Artificial Intelligence Innovation and Incubation Institute Fudan University)
Domain AdaptationRepresentation LearningImageTabularTime Series
🎯 What it does: Developed a framework for invariant learning called TIVA, which automatically identifies auxiliary variables unrelated to the target and performs environment partitioning, thereby learning models that rely solely on invariant features;
Provably Learning Diverse Features in Multi-View Data with Midpoint Mixup
Muthu Chidambaram (Duke University), Rong Ge (Duke University)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper studies the feature learning capability of Midpoint Mixup in multi-view data and demonstrates through theoretical and experimental evidence that it can learn all relevant features more fully than traditional ERM on a two-layer CNN.
Provably Learning Object-Centric Representations
Jack Brady (Max Planck Institute for Intelligent Systems), Wieland Brendel (Max Planck Institute for Intelligent Systems)
Data SynthesisRepresentation LearningAuto EncoderImage
🎯 What it does: Under unsupervised conditions, it is studied and proven that object-centered representations can be identified when specific generative processes (reversibility, compositionality, and irreducibility) are satisfied.
Proximal Causal Learning of Conditional Average Treatment Effects
Erik Sverdrup (Stanford University), Yifan Cui (Zhejiang University)
Biomedical DataElectronic Health Records
🎯 What it does: A P-learner based on proximal causal inference is proposed, using a two-stage loss function with double robust scores and cross-fitting to estimate the conditional average treatment effect in the presence of unmeasured confounding.
Pruning via Sparsity-indexed ODE: a Continuous Sparsity Viewpoint
Zhanfeng Mo (Nanyang Technological University), Sinno Pan
OptimizationConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: A continuous sparse perspective pruning framework SpODE based on sparse index ODE is proposed, along with its PSO algorithm. It utilizes the ODE of sparse evolution over time to guide the precise update of the sparse mask, achieving an efficient one-time pruning strategy.
PWSHAP: A Path-Wise Explanation Model for Targeted Variables
Lucile Ter-Minassian (University of Oxford), Christopher C. Holmes
Explainability and InterpretabilityAdversarial AttackTabular
🎯 What it does: For binary treatment variables in machine learning black-box models given a causal graph (DAG), a local explanation method is provided that decomposes Shapley values into contributions along different causal paths.
Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Stella Biderman (EleutherAI), Oskar van der Wal (Institute for Logic, Language and Computation)
TransformerLarge Language ModelText
🎯 What it does: A set of 16 decoder-only autoregressive language models (Pythia) ranging in size from 70M to 12B has been constructed and publicly released, along with 154 checkpoints, a complete training data loader, and training code, supporting fine-grained experiments on training dynamics, scaling, bias, memorization, and frequency effects; case studies were also conducted on the impact of gender bias, memorization, and pre-training word frequency on task performance.
Q-Flow: Generative Modeling for Differential Equations of Open Quantum Dynamics with Normalizing Flows
Owen M Dugan (Massachusetts Institute of Technology), Marin Soljacic (Massachusetts Institute of Technology)
GenerationData SynthesisOptimizationComputational EfficiencyFlow-based ModelPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a new method for simulating the dynamics of quantum systems called Q-Flow, which transforms the evolution of the density matrix into a high-dimensional partial differential equation of the Husimi Q function and solves it using a regularized flow model.
Q-learning Decision Transformer: Leveraging Dynamic Programming for Conditional Sequence Modelling in Offline RL
Taku Yamagata (University of Bristol), Raul Santos-Rodriguez
TransformerReinforcement LearningSequential
🎯 What it does: This paper proposes an offline reinforcement learning framework called QDT, which combines Q-learning and Decision Transformer. It utilizes the value backtracking of Q-learning to re-label the return-to-go in offline data, addressing the issues of DT's lack of stitching capability and the instability of Q-learning in sparse/long-horizon tasks.
QAS-Bench: Rethinking Quantum Architecture Search and A Benchmark
Xudong Lu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
BenchmarkPhysics Related
🎯 What it does: A unified benchmark for quantum architecture search (QAS) is proposed, divided into two types of tasks: QC regeneration and unitary matrix approximation;
QASA: Advanced Question Answering on Scientific Articles
Yoonjoo Lee (KAIST), Moontae Lee (LG AI Research)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed the QASA benchmark and a three-step QA method to achieve deep reasoning in scientific papers.
Quantifying Human Priors over Social and Navigation Networks
Gecia Bravo-Hermsdorff (University of London)
Graph Neural NetworkGraph
🎯 What it does: The author designed an MCMCP experiment to allow participants to infer hidden relationships under partial graph information, thereby quantifying human prior knowledge of social and navigation graph structures.
Quantifying the Knowledge in GNNs for Reliable Distillation into MLPs
Lirong Wu (Westlake University), Stan Z. Li (Westlake University)
Knowledge DistillationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a GNN-to-MLP knowledge distillation method based on reliable knowledge points, utilizing the invariance of information entropy under noise perturbation to measure the reliability of GNN knowledge, and providing additional supervision for MLP training through reliability sampling.
Quantifying the Variability Collapse of Neural Networks
Jing Xu (Tsinghua University), Haoxiong Liu (Tsinghua University)
Domain AdaptationOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes and evaluates a new variability collapse indicator, VCI, to quantify the variability collapse of the last layer features of neural networks and studies its relationship with transfer performance.
Quantile Credit Assignment
Thomas Mesnard (DeepMind), Remi Munos (DeepMind)
Recurrent Neural NetworkReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: Two novel credit assignment algorithms are proposed—Quantile Credit Assignment (QCA) and Hindsight QCA (HQCA), which utilize quantiles of the return distribution to estimate the level of 'luck', thereby achieving more accurate credit assignment for agent behavior.
Quantitative Universal Approximation Bounds for Deep Belief Networks
Julian Sieber (Zalando Ireland Limited), Johann Gehringer (Imperial College London)
🎯 What it does: This paper studies the approximation capability of Deep Belief Networks (DBN) for continuous probability density functions and provides an upper bound on quantization error as a function of the number of hidden layer units.
Quantized Distributed Training of Large Models with Convergence Guarantees
Ilia Markov (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
TransformerLarge Language ModelText
🎯 What it does: A QSDP (Quantized Sharded Data-Parallel) training framework is proposed, which can quantize weights and gradients while maintaining convergence guarantees, eliminating the communication bottleneck of FSDP and achieving efficient large-scale language model training.
Quantum 3D Graph Learning with Applications to Molecule Embedding
Ge Yan (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Drug DiscoveryGraph Neural NetworkGraphPhysics Related
🎯 What it does: Using quantum circuits for node embedding of molecular 3D graphs, a complete quantum 3D graph learning framework is proposed.
Quantum Lower Bounds for Finding Stationary Points of Nonconvex Functions
Chenyi Zhang (Stanford University), Tongyang Li (Peking University)
OptimizationPhysics Related
🎯 What it does: This paper systematically studies the query lower bounds of quantum algorithms in finding ε-approximate stationary points of non-convex functions, proving that the quantum lower bounds match the classical lower bounds for both p-th order derivatives and stochastic gradient input models, indicating that there is no quantum speedup in these settings.
Quantum Policy Gradient Algorithm with Optimized Action Decoding
Nico Meyer (Fraunhofer Institute for Integrated Circuits), Michael J. Hartmann (Friedrich-Alexander University Erlangen-Nuremberg)
Reinforcement LearningPhysics Related
🎯 What it does: A reinforcement learning policy gradient algorithm based on Variational Quantum Circuits (VQC) is designed, and an optimized action decoding (post-processing) scheme is proposed to improve the quality of action selection;
Quantum Ridgelet Transform: Winning Lottery Ticket of Neural Networks with Quantum Computation
Hayata Yamasaki (University of Tokyo), Sho Sonoda (RIKEN AIP)
OptimizationComputational EfficiencyTime SeriesPhysics Related
🎯 What it does: This paper proposes and implements the Quantum Ridgelet Transform (QRT), which efficiently completes the high-dimensional discrete Ridgelet transform in linear time using quantum computing, and applies this transform as a subroutine to find sparse trainable subnetworks (Lottery Tickets) to approximate a given function.
Quantum Speedups for Zero-Sum Games via Improved Dynamic Gibbs Sampling
Adam Bouland (Stanford University), Kevin Tian (Microsoft Research)
OptimizationPhysics Related
🎯 What it does: A quantum algorithm is designed to compute ε-approximate Nash equilibria in m × n zero-sum games, and its time complexity is provided.
QuantumDARTS: Differentiable Quantum Architecture Search for Variational Quantum Algorithms
Wenjie Wu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationNeural Architecture SearchImageGraphPhysics Related
🎯 What it does: A differentiable quantum architecture search method called QuantumDARTS is designed to automatically generate parameterized quantum circuits (PQC) for variational quantum algorithms (VQE, QNN, etc.).
R-U-SURE? Uncertainty-Aware Code Suggestions By Maximizing Utility Across Random User Intents
Daniel D. Johnson (Google Research), Christian Walder (Google Research)
OptimizationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper proposes the R-U-SURE system, which uses sampling from generative models as a proxy for user intent. By combining minimum Bayes risk, dual decomposition, and decision diagrams, it provides uncertainty annotations for code suggestions, enhancing the developer experience in uncertain contexts.
RACE: Improve Multi-Agent Reinforcement Learning with Representation Asymmetry and Collaborative Evolution
Pengyi Li (Tianjin University), Xian Fu (Tianjin University)
Reinforcement LearningAgentic AISequentialBenchmark
🎯 What it does: Proposes the RACE framework, which combines evolutionary algorithms and multi-agent reinforcement learning to enhance collaborative learning through asymmetric representation and co-evolution.
Raising the Cost of Malicious AI-Powered Image Editing
Hadi Salman (Massachusetts Institute of Technology), Aleksander Madry (Massachusetts Institute of Technology)
GenerationAdversarial AttackDiffusion modelImage
🎯 What it does: This paper proposes a method to 'immunize' images by adding imperceptible adversarial perturbations, preventing large-scale diffusion models from performing realistic edits, thereby increasing the cost of malicious AI image editing.
Random Classification Noise does not defeat All Convex Potential Boosters Irrespective of Model Choice
Yishay Mansour (Tel Aviv University), Robert Williamson
ClassificationOptimizationTabular
🎯 What it does: This paper proves that the root cause of the failure of all convex boosters due to random classification noise is the linear separator model, rather than the loss function itself, through the theory of convex latent functions.
Random Grid Neural Processes for Parametric Partial Differential Equations
Arnaud Vadeboncoeur (University of Cambridge), Omer Deniz Akyildiz (Imperial College London)
Convolutional Neural NetworkTabularPhysics Related
🎯 What it does: This paper proposes a probabilistic physics-informed deep model based on Random Grid Neural Processes (RGNP), which can simultaneously learn the forward and inverse mappings of parameterized partial differential equations (PDEs) and quantify the uncertainty of the solution fields and parameters.
Random Matrix Analysis to Balance between Supervised and Unsupervised Learning under the Low Density Separation Assumption
Vasilii Feofanov (Huawei Noah's Ark Lab), Aladin Virmaux (Huawei Noah's Ark Lab)
ClassificationHyperparameter SearchTabular
🎯 What it does: A linear semi-supervised classification framework QLDS based on the low-density separation hypothesis is proposed, along with an explicit closed-form solution.
Random Shuffle Transformer for Image Restoration
Jie Xiao (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationTransformerImage
🎯 What it does: Proposes a shuffle strategy to achieve global non-local interactions in local window Transformers, constructing the ShuffleFormer model;
Random Teachers are Good Teachers
Felix Sarnthein (ETH Zurich), Thomas Hofmann (ETH Zurich)
Knowledge DistillationRepresentation LearningImage
🎯 What it does: In a setting without labels and data augmentation, self-distillation is performed using a randomly initialized teacher model to observe the improvement of the student model in representation learning.
Randomized Gaussian Process Upper Confidence Bound with Tighter Bayesian Regret Bounds
Shion Takeno (Nagoya Institute of Technology), Masayuki Karasuyama (Nagoya Institute of Technology)
Optimization
🎯 What it does: This paper studies a randomized variant of Gaussian Process Upper Confidence Bound (GP-UCB) in Bayesian optimization, proposes an improved IRGP-UCB algorithm, and provides tighter upper bounds for its Bayesian Cumulative Reward (BCR) and Bayesian Simple Reward (BSR);
Randomized Schur Complement Views for Graph Contrastive Learning
Vignesh Kothapalli (New York University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A data augmentation method for graph contrastive learning based on randomized Schur complement (rLap) is proposed, utilizing random node elimination and unbiased Clique approximation to preserve the random walk probabilities of the original graph.
RankMe: Assessing the Downstream Performance of Pretrained Self-Supervised Representations by Their Rank
Quentin Garrido (Meta AI), Yann LeCun (New York University)
Representation LearningHyperparameter SearchContrastive LearningImage
🎯 What it does: RankMe is proposed, an unlabeled evaluation metric based on the effective rank of the JE-SSL embedding matrix, used for predicting downstream task performance and hyperparameter selection.
Reachability-Aware Laplacian Representation in Reinforcement Learning
Kaixin Wang (National University of Singapore), Xinchao Wang (National University of Singapore)
Reinforcement LearningGraph
🎯 What it does: A new state representation method RA-LapRep is proposed in reinforcement learning to address the issue of mismatch between distance and reachability in traditional Laplace representation.
Reasons for the Superiority of Stochastic Estimators over Deterministic Ones: Robustness, Consistency and Perceptual Quality
Guy Ohayon (Technion), Tomer Michaeli (Technion)
RestorationSuper ResolutionGenerative Adversarial NetworkImage
🎯 What it does: This paper discusses the advantages and disadvantages of stochastic and deterministic image recovery methods, proving that stochastic methods have theoretical advantages in consistency, perceptual quality, and robustness.
Recasting Self-Attention with Holographic Reduced Representations
Mohammad Mahmudul Alam (University of Maryland), James Holt (Laboratory for Physical Sciences)
Anomaly DetectionComputational EfficiencyTransformerTabular
🎯 What it does: Redesign self-attention as a linear attention mechanism based on HRR (Holographic Reduced Representations), forming Hrrformer;
Reconstructive Neuron Pruning for Backdoor Defense
Yige Li (Xidian University), Yu-Gang Jiang (Fudan University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A backdoor defense method called Reconstructive Neuron Pruning (RNP) is proposed, which combines neuron-level unlearning with filter-level recovering to expose and prune backdoor-related neurons, thereby eliminating backdoor attacks.
Recovering Top-Two Answers and Confusion Probability in Multi-Choice Crowdsourcing
Hyeonsu Jeong (KAIST), Hye Won Chung (KAIST)
Convolutional Neural NetworkImage
🎯 What it does: A multi-choice crowdsourcing model is proposed, aiming to simultaneously recover the correct answer, the most confusable answer, and their confusion probabilities; an optimal two-stage inference algorithm is provided.
Recovery Bounds on Class-Based Optimal Transport: A Sum-of-Norms Regularization Framework
Arman Rahbar (Chalmers University of Technology), Hamid Krim (North Carolina State University)
ClassificationOptimizationImage
🎯 What it does: An optimal transport theoretical framework based on class structures is proposed, and the recovery of hidden class structures and efficient solving is achieved by incorporating sum-of-norms (SON) regularization.
ReDi: Efficient Learning-Free Diffusion Inference via Trajectory Retrieval
Kexun Zhang (University of California), Lei Li (University of California)
GenerationRetrievalComputational EfficiencyDiffusion modelImageRetrieval-Augmented GenerationOrdinary Differential Equation
🎯 What it does: This paper proposes REDI, a method that accelerates diffusion model inference by utilizing pre-built trajectory retrieval, allowing it to skip a large number of sampling steps without additional training.
Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMC
Yilun Du (Massachusetts Institute of Technology), Will Sussman Grathwohl
GenerationData SynthesisDiffusion modelScore-based ModelImageText
🎯 What it does: This paper proposes the use of energy-based parameterized diffusion models combined with MCMC sampling to achieve seamless combination generation of pre-trained models.
Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs
Saro Passaro (Meta AI), C. Lawrence Zitnick (Meta AI)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes to map SO(3) convolutions to SO(2) convolutions by aligning the principal axes of node invariants with edge vectors, significantly reducing computational complexity and achieving an efficient SO(3) equivariant graph neural network (eSCN).
Refined Regret for Adversarial MDPs with Linear Function Approximation
Yan Dai (Tsinghua University), Julian Zimmert (Google Research)
OptimizationReinforcement Learning
🎯 What it does: This paper studies learning in adversarial Markov decision processes (MDPs) where the loss function can vary arbitrarily over K rounds. Two algorithms are proposed that improve the regret from the existing O(K^(2/3)) to O(√K).
Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models
Dongjun Kim (KAIST), Il-chul Moon
GenerationData SynthesisDiffusion modelScore-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Based on the pre-trained diffusion model, a discriminator is introduced to supervise the generated trajectories, thereby improving sampling quality.
Reflected Diffusion Models
Aaron Lou (Stanford University), Stefano Ermon (Stanford University)
GenerationData SynthesisDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A reflection diffusion model is proposed, utilizing reflective stochastic differential equations to model in the data-supported domain;
Regions of Reliability in the Evaluation of Multivariate Probabilistic Forecasts
Étienne Marcotte (ServiceNow Research), Nicolas Chapados (ServiceNow Research)
Time SeriesBenchmark
🎯 What it does: This paper systematically evaluates the reliability of commonly used scoring rules in multivariate probabilistic time series forecasting through finite sample power analysis and composite benchmarks.
Regression with Label Permutation in Generalized Linear Model
Guanhua Fang (Fudan University), Ping Li (LinkedIn)
Tabular
🎯 What it does: This paper conducts a theoretical analysis of the label permutation problem in generalized linear models (GLM) for multiple responses and proposes two algorithms that can achieve label recovery.
Regression with Sensor Data Containing Incomplete Observations
Takayuki Katsuki (IBM Research), Takayuki Osogami (IBM Research)
TabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a regression method to address the negative noise bias caused by incomplete observations of sensor labels, with the goal of learning an accurate size prediction function.
Regret Bounds for Markov Decision Processes with Recursive Optimized Certainty Equivalents
Wenhao XU, Xuedong He (Chinese University of Hong Kong)
OptimizationReinforcement Learning
🎯 What it does: A risk-sensitive reinforcement learning framework based on Recursive Optimization of Certainty Equivalent (OCE) is proposed, and a model-based algorithm OCE-VI is designed, providing optimal proofs for upper and lower bounds.
Regret Minimization and Convergence to Equilibria in General-sum Markov Games
Liad Erez (Blavatnik School of Computer Science), Yishay Mansour (Blavatnik School of Computer Science)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: A completely decentralized, communication-free algorithm is proposed to achieve sublinear swap regret and convergence to correlated equilibria in multi-player general-sum Markov games.
Regret-Minimizing Double Oracle for Extensive-Form Games
Xiaohang Tang (University College London), Yaodong Yang (Peking University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: A general Regret-Minimizing Double Oracle (RMDO) framework is proposed, and based on this, a Periodic Double Oracle (PDO) is introduced to address the sample complexity issue of the double oracle method in extensive form games (EFG).
Regularization and Variance-Weighted Regression Achieves Minimax Optimality in Linear MDPs: Theory and Practice
Toshinori Kitamura (University of Tokyo), Yutaka Matsuo (University of Tokyo)
Reinforcement Learning
🎯 What it does: This study investigates the sample complexity of Mirror Descent Value Iteration (MDVI) in linear MDPs and proposes the theoretically optimal Variance-Weighted Least-Squares MDVI (VWLS-MDVI) along with the practical deep RL algorithm Deep Variance Weighting (DVW);
Regularization-free Diffeomorphic Temporal Alignment Nets
Ron Shapira Weber (Ben Gurion University), Oren Freifeld (Ben Gurion University)
ClassificationOptimizationRecurrent Neural NetworkTime Series
🎯 What it does: A non-regularized joint alignment and averaging method for time series is proposed, trained using the Inverse Consistency Average Error (ICAE) loss and its triplet variant.
Regularizing Towards Soft Equivariance Under Mixed Symmetries
Hyunsu Kim (KAIST), Juho Lee (KAIST)
Autonomous DrivingConvolutional Neural NetworkTime Series
🎯 What it does: A regularization-based soft equivariance method (PER) is proposed, which can handle multiple types of approximate symmetries simultaneously.
Reinforcement Learning Can Be More Efficient with Multiple Rewards
Christoph Dann (Google Research), Mehryar Mohri (Courant Institute of Mathematical Sciences)
OptimizationComputational EfficiencyReinforcement LearningAgentic AI
🎯 What it does: This paper studies how to achieve higher sample efficiency in reinforcement learning when using multiple reward functions through action elimination algorithms, and provides a theoretically better instance-dependent regret bound.
Reinforcement Learning from Passive Data via Latent Intentions
Dibya Ghosh (University of California Berkeley), Sergey Levine (University of California Berkeley)
Reinforcement LearningVideo
🎯 What it does: This paper proposes a method to pre-train state representations by learning latent intentions using only passive observation data (without action/reward information), and applies it to downstream reinforcement learning tasks.
Reinforcement Learning in Low-rank MDPs with Density Features
Audrey Huang (University of Illinois Urbana-Champaign), Nan Jiang (University of Illinois Urbana-Champaign)
Reinforcement Learning
🎯 What it does: This paper studies sample-efficient reinforcement learning using state density features in low-rank Markov decision processes (MDPs) and proposes offline and online occupancy estimation algorithms.
Reinforcement Learning with General Utilities: Simpler Variance Reduction and Large State-Action Space
Anas Barakat (ETH Zurich), Niao He (ETH Zurich)
OptimizationReinforcement Learning
🎯 What it does: A single-loop, parameter-free normalized variance reduction strategy gradient algorithm is proposed for solving the general utility optimization problem in reinforcement learning.
Reinforcement Learning with History Dependent Dynamic Contexts
Guy Tennenholtz (Google Research), Craig Boutilier
Recommendation SystemTransformerReinforcement LearningTabular
🎯 What it does: A dynamic context Markov decision process (DCMDP) framework is proposed for modeling history-dependent non-Markov environments, and based on this, a subclass of Logistic DCMDP with aggregated features is introduced.
Relevant Walk Search for Explaining Graph Neural Networks
Ping Xiong (Technische Universität Berlin), Shinichi Nakajima (Technische Universität Berlin)
Explainability and InterpretabilityComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: Proposes a polynomial-time algorithm based on maximum product (max-product) message passing (EMP-neu and AMP-ave) that can efficiently find the most relevant walks in GNN predictions, addressing the exponential search problem of the original GNN-LRP.
Reliable Measures of Spread in High Dimensional Latent Spaces
Anna Marbut, Travis J Wheeler
Representation LearningText
🎯 What it does: This paper proposes new metrics to measure the distribution breadth of data in the latent space of natural language processing models and evaluates the utilization of the model space using these metrics.
ReLOAD: Reinforcement Learning with Optimistic Ascent-Descent for Last-Iterate Convergence in Constrained MDPs
Ted Moskovitz (University College London), Tom Zahavy (DeepMind)
OptimizationReinforcement LearningSequentialBenchmark
🎯 What it does: This paper proposes a new constrained reinforcement learning framework called ReLOAD, which utilizes Optimistic Mirror Descent (OMD) to achieve Last-Iterate Convergence (LIC) for Constrained Markov Decision Processes (CMDP), and provides a specific implementation based on policy and function approximation.
Reparameterized Policy Learning for Multimodal Trajectory Optimization
Zhiao Huang (University of California San Diego), Hao Su (University of California San Diego)
OptimizationRobotic IntelligenceReinforcement LearningWorld ModelMultimodality
🎯 What it does: In reinforcement learning with high-dimensional continuous action spaces, a method is proposed to generate multimodal trajectory policies through latent variable reparameterization, combined with model learning and state entropy rewards to achieve efficient exploration and optimization.
Repository-Level Prompt Generation for Large Language Models of Code
Disha Shrivastava (Mila), Daniel Tarlow (McGill University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A Repo-Level Prompt Generator (RLPG) framework is proposed, which automatically generates prompts suitable for single-line code completion by utilizing repository structure and cross-file context.
Representation Learning with Multi-Step Inverse Kinematics: An Efficient and Optimal Approach to Rich-Observation RL
Zakaria Mhammedi (Massachusetts Institute of Technology), Alexander Rakhlin (Massachusetts Institute of Technology)
Representation LearningRobotic IntelligenceReinforcement Learning
🎯 What it does: This paper studies a sample-efficient algorithm for reinforcement learning under high-dimensional observations, proposing the MusIK algorithm, which addresses the computational infeasibility, strong statistical assumptions, and insufficient sample complexity of existing algorithms.
Representation-Driven Reinforcement Learning
Ofir Nabati (Technion Institute of Technology), Shie Mannor (Nvidia Research)
Representation LearningReinforcement LearningSequential
🎯 What it does: This paper proposes a reinforcement learning framework called RepRL based on policy representation learning, which maps policies to a linear feature space and utilizes linear multi-armed bandit algorithms for exploration and exploitation.
Representations and Exploration for Deep Reinforcement Learning using Singular Value Decomposition
Yash Chandak (University of Massachusetts), Diana L Borsa
Representation LearningReinforcement LearningMultimodality
🎯 What it does: A method based on Singular Value Decomposition (SVD) is proposed to obtain representations that preserve the transfer structure within the domain and explore in partially observable environments.
Representer Point Selection for Explaining Regularized High-dimensional Models
Che-Ping Tsai (Carnegie Mellon University), Pradeep Kumar Ravikumar
Recommendation SystemExplainability and InterpretabilityText
🎯 What it does: A general high-dimensional representer method is proposed to explain the predictive contributions of high-dimensional regularization models (such as ℓ1 sparse models and nuclear norm low-rank models) to test samples.
Reprogramming Pretrained Language Models for Antibody Sequence Infilling
Igor Melnyk (IBM Research), Devleena Das (Georgia Institute of Technology)
GenerationData SynthesisDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: The study proposes ReprogBert, which transfers the pre-trained English BERT to the antibody sequence filling task through model reprogramming, generating diverse and structurally consistent CDR sequences.
Restoration based Generative Models
Jaemoo Choi (Seoul National University), Myungjoo Kang (Seoul National University)
RestorationGenerationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A restorative generative model (RGM) is proposed through the maximum a posteriori (MAP) objective combined with implicit priors, allowing for flexible selection of the denoising process and eliminating the sampling bottleneck of traditional diffusion models.
Restoration-Degradation Beyond Linear Diffusions: A Non-Asymptotic Analysis For DDIM-type Samplers
Sitan Chen (Harvard University), Alex Dimakis
RestorationOptimizationDiffusion modelStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The non-asymptotic convergence of deterministic diffusion model samplers is studied, and operational interpretations are provided.
Resurrecting Recurrent Neural Networks for Long Sequences
Antonio Orvieto (ETH Zurich), Soham De (DeepMind)
Recurrent Neural NetworkSequential
🎯 What it does: A new Linear Recursive Unit (LRU) is proposed, which, through a series of design improvements (linear recursion, complex diagonalization, exponential parameterization, normalization, etc.), enables traditional RNNs to achieve or even exceed the performance of deep State Space Models (SSM) on long sequence tasks in the Long Range Arena.
Rethink DARTS Search Space and Renovate a New Benchmark
Jiuling Zhang (University of Chinese Academy of Sciences), Zhiming Ding (Institute of Software, Chinese Academy of Sciences)
Neural Architecture SearchImageBenchmark
🎯 What it does: This paper redesigns the DARTS search space, proposing a larger and more challenging LHD space and constructing a multi-condition evaluation benchmark based on it.