arXivSub Start free trial

NeurIPS 2023 Papers — Page 25

Conference on Neural Information Processing Systems · 3218 papers

Reducing Blackwell and Average Optimality to Discounted MDPs via the Blackwell Discount Factor

Julien Grand-Clément (HEC Paris), Marek Petrik (University of New Hampshire)

OptimizationReinforcement Learning

🎯 What it does: This paper introduces the Blackwell discount factor γ_bw, proving that any discount γ > γ_bw leads to an optimal strategy that is both Blackwell optimal and average optimal. It also provides a computable upper bound, completing a generalization of the reduction from Blackwell/average optimality to discount optimality, and offers a new polynomial-time algorithm.

Reducing Shape-Radiance Ambiguity in Radiance Fields with a Closed-Form Color Estimation Method

Qihang Fang (Chinese Academy of Sciences), Liefeng Bo (Alibaba Group)

GenerationData SynthesisNeural Radiance FieldImage

🎯 What it does: Given a density field and training images with known camera poses, this paper proposes a closed-form solution to estimate the color field of the scene, thereby achieving the separation of shape fields and radiance fields during rendering.

Reference-Based POMDPs

Edward Kim (Australian National University), Hanna Kurniawati (Australian National University)

Reinforcement Learning

🎯 What it does: A reference-based POMDP model is proposed, which adds a KL penalty for the reference stochastic policy on the basis of traditional POMDP, allowing the solution to be transformed into an expectation calculation, thereby avoiding the enumeration of all actions at each belief node.

REFINE: A Fine-Grained Medication Recommendation System Using Deep Learning and Personalized Drug Interaction Modeling

Suman Bhoi (National University of Singapore), Ngiap Chuan Tan (SingHealth Polyclinics)

Recommendation SystemDrug DiscoveryGraph Neural NetworkTransformerTabularSequentialBiomedical DataElectronic Health Records

🎯 What it does: A fine-grained drug recommendation system called REFINE is proposed, which generates personalized and safe drug combinations by utilizing the trends of drug dosages and laboratory results from multiple patient visits, as well as the severity of drug interactions.

Refined Mechanism Design for Approximately Structured Priors via Active Regression

Christos Boutsikas (Purdue University), Paritosh Verma (Purdue University)

Optimization

🎯 What it does: A framework for designing an active learning mechanism based on Random Linear Algebra (RLA) is proposed, aimed at approximating unknown high-dimensional prior distributions in multi-item auctions, and constructing approximately optimal and feasible mechanisms based on this.

Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans

Kyowoon Lee (Ulsan National Institute of Science and Technology), Jaesik Choi (Korea Advanced Institute of Science and Technology)

Robotic IntelligenceReinforcement LearningDiffusion modelTabularStochastic Differential Equation

🎯 What it does: Improved planning based on diffusion models by detecting and correcting infeasible trajectories to enhance execution reliability.

Reflexion: language agents with verbal reinforcement learning

Noah Shinn (Northeastern University), Shunyu Yao (Princeton University)

Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought

🎯 What it does: Proposes the Reflexion framework, which uses self-reflective text from language models as a 'semantic gradient' in the experience buffer to guide subsequent decisions and generation.

RegBN: Batch Normalization of Multimodal Data with Regularization

MORTEZA GHAHREMANI, Christian Wachinger (Technical University of Munich)

ClassificationOptimizationConvolutional Neural NetworkTransformerMultimodalityAlzheimer's Disease

🎯 What it does: To address the performance degradation caused by confounding factors and inter-modal dependencies in multimodal neural networks, a parameter-free regularization batch normalization method called RegBN is proposed, which normalizes low-level and high-level features before multimodal fusion.

Regression with Cost-based Rejection

Xin Cheng (Chongqing University), Lei Feng (Nanyang Technological University)

Tabular

🎯 What it does: This paper proposes and studies the framework of Regression with Cost-based Rejection (RcR), which addresses the issue of whether to reject outputs based on given rejection costs during model predictions.

Regret Matching+: (In)Stability and Fast Convergence in Games

Gabriele Farina (Massachusetts Institute of Technology), Haipeng Luo (University of Southern California)

OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular

🎯 What it does: Theoretical and experimental analysis of Regret Matching+ (RM+) and its predictive version reveals potential instability, leading to slow convergence in multi-player games. Two stabilization methods (restarting and removing the area near the origin) are proposed, and based on this, new algorithms such as stable/smooth predictive RM+, conceptual RM+, and extrapolated gradient RM+ are designed; constant or O(T^{1/4}) individual or social regret upper bounds are provided in multi-player normal-form games and extensive-form games, and the performance of the algorithms is validated on matrix games and four types of extensive-form games.

Regret Minimization via Saddle Point Optimization

Johannes Kirschner (University of Alberta), Csaba Szepesvari

OptimizationReinforcement LearningSequential

🎯 What it does: An ANYTIME-E2D algorithm based on average constraint decision-estimation coefficients is proposed for sequential decision-making problems under structured observations.

Regret-Optimal Model-Free Reinforcement Learning for Discounted MDPs with Short Burn-In Time

Xiang Ji (Princeton University), Gen Li (Chinese University of Hong Kong)

Reinforcement Learning

🎯 What it does: In the infinite-horizon discounted Markov decision process, a model-free Q-SlowSwitch-Adv algorithm is designed and analyzed, proving that it can achieve cumulative regret optimality (matching the theoretical lower bound) when the sample size is sufficient, with low space (O(SA)) and low computational (O(T)) complexity, and a short burn-in period (T≥SA·poly(1-γ)).

Regularity as Intrinsic Reward for Free Play

Cansu Sancaktar (Max Planck Institute for Intelligent Systems), Georg Martius (Max Planck Institute for Intelligent Systems)

Robotic IntelligenceReinforcement Learning from Human FeedbackGraph Neural NetworkReinforcement LearningWorld ModelGraph

🎯 What it does: A new intrinsic reward called Regularity Reward (RaIR) is proposed, which encourages agents to explore structured and orderly behavior patterns by minimizing the entropy of state descriptions. This is combined with model uncertainty (ensemble disagreement) to guide robots in constructing tower-like and symmetrical structures during the untasked 'free play' phase; subsequently, zero-shot downstream assembly tasks are performed on the same model.

Regularization properties of adversarially-trained linear regression

Antonio H. Ribeiro, Thomas B. Schön (Uppsala University)

OptimizationAdversarial AttackTabular

🎯 What it does: This paper studies adversarial training in linear regression and analyzes its regularization properties, revealing the equivalence relationships between adversarial training and minimum norm interpolators, Lasso, Ridge, and square root Lasso in different parameter ranges.

Regularized Behavior Cloning for Blocking the Leakage of Past Action Information

Seokin Seo (KAIST), Kee-Eung Kim (KAIST)

Autonomous DrivingOptimizationReinforcement LearningSequential

🎯 What it does: This study investigates the dangers of behavior cloning leaking past action information when using observational history and proposes a regularization framework based on conditional independence.

Regularizing Neural Networks with Meta-Learning Generative Models

Shin'ya Yamaguchi (NTT), Hisashi Kashima (Kyoto University)

ClassificationData SynthesisMeta LearningGenerative Adversarial NetworkImage

🎯 What it does: This paper studies an improved method for generative data augmentation and proposes a Meta-Generative Regularization (MGR) strategy, which consists of two parts: Pseudo-Consistency Regularization (PCR) and Meta-Pseudo Sampling (MPS), aimed at enhancing the performance of classification models in small sample scenarios.

Rehearsal Learning for Avoiding Undesired Future

Tian Qin (Nanjing University), Zhi-Hua Zhou (Nanjing University)

OptimizationReinforcement LearningTabular

🎯 What it does: This paper proposes a 'rehearsal learning' framework that utilizes a Structured Replay Model (SRM) to make executable decisions in a limited interaction environment to avoid adverse futures.

ReHLine: Regularized Composite ReLU-ReHU Loss Minimization with Linear Computation and Linear Convergence

Ben Dai (Chinese University of Hong Kong), Yixuan Qiu (Shanghai University of Finance and Economics)

OptimizationTabular

🎯 What it does: A general algorithm named ReHLine is proposed to minimize the regularized empirical risk minimization (ERM) problem that includes convex piecewise linear quadratic (PLQ) loss functions and linear constraints, capable of efficiently handling various common losses such as SVM, Huber, quantile regression, etc.;

Reinforcement Learning with Fast and Forgetful Memory

Steven Morad (University of Cambridge), Amanda Prorok (University of Cambridge)

Recurrent Neural NetworkReinforcement LearningSequentialBenchmark

🎯 What it does: Proposes the Fast and Forgetful Memory (FFM) model as a pluggable alternative to RNNs, achieving faster training and higher rewards in reinforcement learning through a memory mechanism;

Reinforcement Learning with Simple Sequence Priors

Tankred Saanum (Max Planck Institute for Biological Cybernetics), Eric Schulz (Max Planck Institute for Biological Cybernetics)

TransformerReinforcement LearningSequential

🎯 What it does: Introducing sequence compression priors in reinforcement learning encourages agents to produce compressible and predictable action sequences, thereby accelerating learning, enhancing returns, and improving robustness to noise.

Reinforcement-Enhanced Autoregressive Feature Transformation: Gradient-steered Search in Continuous Space for Postfix Expressions

Dongjie Wang (University of Central Florida), Yanjie Fu (Arizona State University)

OptimizationRecurrent Neural NetworkReinforcement LearningTabular

🎯 What it does: The MOAT framework is proposed, transforming the automatic feature transformation (AFT) problem into continuous optimization. It collects high-quality training data through reinforcement learning, uses suffix expressions to represent transformation sequences, and performs gradient search and beam search in continuous embedding space to generate the optimal feature space.

Reining Generalization in Offline Reinforcement Learning via Representation Distinction

Yi Ma (Tianjin University), Zhaopeng Meng (Tianjin University)

Reinforcement LearningTabular

🎯 What it does: This study investigates the problem of overgeneralization in offline reinforcement learning and proposes the Representation Distinction (RD) method, which suppresses generalization by distinguishing the source of samples in the representation space (dataset samples vs. samples generated by the learning policy), thereby improving the performance of offline RL.

Relative Entropic Optimal Transport: a (Prior-aware) Matching Perspective to (Unbalanced) Classification

Liangliang Shi (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)

ClassificationRepresentation LearningImage

🎯 What it does: This paper views classification as a matching problem and proposes Relative Entropy Regularized Optimal Transport (RE-OT) and its inverse variant for long-tail classification and representation learning.

Relax, it doesn’t matter how you get there: A new self-supervised approach for multi-timescale behavior analysis

Mehdi Azabou (Georgia Institute of Technology), Eva L Dyer

Representation LearningRobotic IntelligenceConvolutional Neural NetworkVideo

🎯 What it does: A self-supervised multi-scale behavior representation learning framework BAMS is proposed, which can simultaneously capture short-term dynamics and long-term trends at different time scales.

Reliable learning in challenging environments

Nina Balcan, Dravyansh Sharma (Carnegie Mellon University)

Domain AdaptationOptimizationAdversarial AttackTabular

🎯 What it does: This paper studies the problem of designing reliable learners in challenging testing environments, particularly in the context of adversarial testing attacks and natural distribution shifts, and proposes a reliable learner with provable optimal guarantees.

Reliable Off-Policy Learning for Dosage Combinations

Jonas Schweisthal (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)

OptimizationDrug DiscoveryReinforcement LearningBiomedical DataElectronic Health Records

🎯 What it does: A reliable off-policy learning framework is proposed to address the dose combination problem in personalized medicine, capable of estimating joint dose effects, detecting limited overlap regions, and optimizing individual dose schemes under these constraints.

ReMaX: Relaxing for Better Training on Efficient Panoptic Segmentation

Shuyang Sun (University of Oxford), Liang-Chieh Chen (Google Research)

Object DetectionSegmentationTransformerImage

🎯 What it does: By introducing Relaxation (ReMask and ReClass) to the mask and class predictions of the mask transformer during the training phase, the convergence speed and final performance of the model on the efficient panoptic segmentation task are improved.

Removing Hidden Confounding in Recommendation: A Unified Multi-Task Learning Approach

Haoxuan Li (Peking University), Peng Wu (Beijing Technology and Business University)

Recommendation SystemVideo

🎯 What it does: A unified multi-task learning framework is proposed in the recommendation system, utilizing a small amount of unbiased data to calibrate propensity scores and error estimates, eliminating biases caused by hidden confounding.

Repetition In Repetition Out: Towards Understanding Neural Text Degeneration from the Data Perspective

Huayang Li (Nara Institute of Science and Technology), Yixuan Su (University of Cambridge)

GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data

🎯 What it does: This study investigates the fundamental causes of the degeneration phenomenon in neural text generation, finding a high correlation between the proportion of repeated words in the training data and the rate of generated repetitions. It proposes a method to suppress the model's reliance on repetitions by randomly masking repeated n-grams in the attention mechanism (Repetition Dropout).

Replicability in Reinforcement Learning

Amin Karbasi (Yale University), Felix Zhou (Yale University)

Reinforcement Learning

🎯 What it does: This paper studies reproducibility as an algorithmic property in reinforcement learning, proposing a reproducible algorithm in the discounted tabular MDP environment and providing an analysis of sample and time complexity.

Replicable Clustering

Hossein Esfandiari (Google Research), Felix Zhou (Yale University)

OptimizationTabular

🎯 What it does: This paper designs a replicable algorithm for statistical clustering problems, providing replicable approximate solutions for k-medians, k-means, and k-centers.

Replicable Reinforcement Learning

ERIC EATON, Jessica Sorrell (University of Pennsylvania)

Reinforcement Learning

🎯 What it does: A formal reproducibility framework is proposed for the first time in reinforcement learning, and two reproducible algorithms are implemented: Reproducible Value Iteration under Parallel Sampling (rPVI) and Reproducible R-max with Exploration (RepRMAX).

RePo: Resilient Model-Based Reinforcement Learning by Regularizing Posterior Predictability

Chuning Zhu (University of Washington), Abhishek Gupta (University of Washington)

Robotic IntelligenceReinforcement LearningVideo

🎯 What it does: A visual model-based reinforcement learning method named RePo is proposed, which learns to compress latent representations that are predictable for task-relevant information and robust to irrelevant variations in dynamic and noisy environments, and provides an adaptation mechanism during testing.

Representation Equivalent Neural Operators: a Framework for Alias-free Operator Learning

Francesca Bartolucci (Delft University of Technology), Rima Alaifari (Seminar for Applied Mathematics, ETH Zurich)

Convolutional Neural NetworkTime Series

🎯 What it does: A framework for Representationally Equivalent Neural Operators (ReNO) is proposed, which systematically quantifies and eliminates aliasing errors in the discretization process of operators.

Representation Learning via Consistent Assignment of Views over Random Partitions

Thalles Santos Silva (University of Campinas), Adín Ramírez Rivera (University of Oslo)

RetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningImage

🎯 What it does: This paper proposes CARP - a self-supervised clustering method based on random partitioning for consistent view assignment, aimed at representation learning of visual features.

Representational Strengths and Limitations of Transformers

Clayton Sanford (Columbia University), Matus Telgarsky (New York University)

Transformer

🎯 What it does: This paper studies the expressive power of the Transformer self-attention mechanism and proposes the sparse average task (q-SA) as an analytical tool. It proves that the Transformer can approximate this task with only logarithmic levels of embedding dimensions, while ordinary feedforward networks and recurrent networks require polynomial levels of width/state. It also discusses the Pair-Match and Triple-Match tasks, demonstrating that the Transformer can efficiently handle binary matching, while ternary matching requires additional structures or higher-order attention.

Reproducibility in Multiple Instance Learning: A Case For Algorithmic Unit Tests

Edward Raff (Booz Allen Hamilton), James Holt (Laboratory for Physical Sciences)

ClassificationAnomaly DetectionGraph Neural NetworkTabular

🎯 What it does: An algorithmic unit test for multi-instance learning (MIL) models is proposed, using synthetic data to verify whether the model adheres to the basic assumptions of MIL and to assess the compliance of existing deep MIL models.

Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone

Zeyinzi Jiang (Alibaba Group), Jingren Zhou (Alibaba Group)

OptimizationComputational EfficiencyTransformerImage

🎯 What it does: The Res-Tuning framework is proposed, decoupling the tuner from the backbone to achieve flexible and efficient parameterized fine-tuning, and a memory-efficient Res-Tuning-Bypass is designed.

Resetting the Optimizer in Deep RL: An Empirical Study

Kavosh Asadi (Amazon), Shoham Sabach (Amazon)

OptimizationReinforcement LearningSequential

🎯 What it does: This paper studies the optimization of the value function in deep reinforcement learning as a series of changing optimization problems, and proposes a simple strategy to reset the internal state of the optimizer (such as the first and second moment estimates of Adam) each time the target network is updated; experiments show that this reset strategy can significantly improve the performance of Rainbow and other RL algorithms.

Residual Alignment: Uncovering the Mechanisms of Residual Networks

Jianing Li (University of Toronto), Vardan Papyan (University of Toronto)

ClassificationOptimizationConvolutional Neural NetworkImage

🎯 What it does: This paper systematically studies the residual Jacobian matrix of residual networks, discovering and quantifying the phenomenon of Residual Alignment (RA), and proving its relevance to neural folding.

Residual Q-Learning: Offline and Online Policy Customization without Value

Chenran Li (University of California Berkeley), Wei Zhan (Toyota Research Institute)

Autonomous DrivingReinforcement LearningSequential

🎯 What it does: A residual Q-learning framework is proposed, which customizes the policy offline and online using existing imitation strategies without knowing its reward function.

Resilient Constrained Learning

Ignacio Hounie (University of Pennsylvania), Luiz F. O. Chamon (University of Stuttgart)

OptimizationFederated LearningImage

🎯 What it does: A framework of 'elastic constraint learning' is proposed, which automatically adjusts the degree of constraint relaxation during the training process to achieve a balance between the target loss and multiple constraints (such as fairness, robustness, and the distribution heterogeneity in federated learning).

Resilient Multiple Choice Learning: A learned scoring scheme with application to audio scene analysis

Victor Letzelter (Valeo), Gaël Richard (Telecom Paris)

ClassificationRecognitionAnomaly DetectionRecurrent Neural NetworkMultimodalityAudio

🎯 What it does: A multi-choice learning variant named rMCL is proposed, which addresses the issues of overconfidence and hypothesis collapse in regression tasks using a learnable scoring mechanism.

ResMem: Learn what you can and memorize the rest

Zitong Yang (Stanford University), Sanjiv Kumar (Google Research)

ClassificationRecognitionConvolutional Neural NetworkTransformerSupervised Fine-TuningImageTextRetrieval-Augmented Generation

🎯 What it does: This paper proposes a two-stage learning framework called ResMem (Residual-Memorization), which first trains a base deep network (DeepNet) to perform initial fitting on the training samples, and then uses a k-Nearest-Neighbor (k-NN) regressor to explicitly memorize the residuals of the network, with the final prediction being the sum of the base network output and the k-NN residual.

Resolving the Tug-of-War: A Separation of Communication and Learning in Federated Learning

Junyi Li (University of Maryland), Heng Huang (University of Maryland)

OptimizationFederated LearningImage

🎯 What it does: The FedSep framework is proposed, which separates the communication layer from the learning layer in federated learning and achieves dual-layer optimization through decoding/encoding mapping.

ResoNet: Noise-Trained Physics-Informed MRI Off-Resonance Correction

Alfredo De Goyeneche (University of California), Michael Lustig (University of California)

RestorationSegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging

🎯 What it does: This paper proposes a physics-driven unrolled deep learning framework that utilizes synthetic noise images, random field maps, fat-water ratios, and coil sensitivity data to train the network for deblurring frequency shifts and fat-water separation in non-Cartesian spiral MRI acquisitions.

Response Length Perception and Sequence Scheduling: An LLM-Empowered LLM Inference Pipeline

Zangwei Zheng (National University of Singapore), Yang You (National University of Singapore)

TransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: A sequence scheduling scheme based on LLM for predicting response length and grouping by length is proposed to significantly improve LLM inference throughput.

Responsible AI (RAI) Games and Ensembles

Yash Gupta (Carnegie Mellon University), Pradeep Kumar Ravikumar

OptimizationImage

🎯 What it does: A unified 'Responsible AI (RAI) Game' framework is proposed, which reduces various responsible AI objectives such as fairness, robustness, and subgroup balance to a minimax loss problem over a set of predefined distributions. Two types of solving algorithms are provided: game-based online learning methods and boosting-based greedy methods.

ResShift: Efficient Diffusion Model for Image Super-resolution by Residual Shifting

Zongsheng Yue (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)

RestorationSuper ResolutionTransformerDiffusion modelImage

🎯 What it does: An efficient diffusion model called ResShift is designed for image super-resolution tasks, compressing the sampling steps from the common thousands to just 15 steps.

Restart Sampling for Improving Generative Processes

Yilun Xu (Massachusetts Institute of Technology), Tommi S. Jaakkola

GenerationData SynthesisDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: A new sampling method called Restart Sampling is proposed, which combines the advantages of ODE and SDE;

Restless Bandits with Average Reward: Breaking the Uniform Global Attractor Assumption

Yige Hong (Carnegie Mellon University), Weina Wang (Carnegie Mellon University)

OptimizationReinforcement Learning

🎯 What it does: This paper proposes a general framework based on simulation and virtual processes called Follow-the-Virtual-Advice (FTVA), which can transform the optimal policy of any single-arm MDP into an approximately optimal policy for the N-armed restless bandit problem, and provides an upper bound on the optimality error under the average reward objective.

ReSync: Riemannian Subgradient-based Robust Rotation Synchronization

Huikang Liu (Shanghai University of Finance and Economics), Anthony Man-Cho So (Chinese University of Hong Kong)

OptimizationPoint Cloud

🎯 What it does: A Riemannian subgradient-based algorithm called ReSync is proposed for the direct solution of the robust rotation synchronization problem, combined with spectral linear initialization.

Retaining Beneficial Information from Detrimental Data for Neural Network Repair

Long-Kai Huang (Tencent AI Lab), Sinno Pan

ClassificationData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage

🎯 What it does: This paper proposes a two-step model repair framework: first, it identifies harmful samples that lead to generalization failure using a small amount of clean data, and then aligns them to a clean distribution using an energy-based model and fine-tunes on the aligned data.

Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face Recognition

Samuel Dooley (University of Maryland), Micah Goldblum (New York University)

RecognitionOptimizationHyperparameter SearchNeural Architecture SearchImage

🎯 What it does: This paper reveals the impact of model architecture and hyperparameters on the bias of face recognition systems through large-scale experiments, and proposes an automatic method using Neural Architecture Search (NAS) + Hyperparameter Optimization (HPO) to find network structures that are both highly accurate and fair, significantly reducing gender and racial bias.

Rethinking Conditional Diffusion Sampling with Progressive Guidance

Anh-Dung Dinh (University of Sydney), Chang Xu (University of Sydney)

GenerationData SynthesisLarge Language ModelDiffusion modelImage

🎯 What it does: An improved classifier guidance method called Progressive Guidance is proposed for the sampling process of diffusion models, aimed at alleviating the issues of diversity compression and adversarial feature construction in traditional classifier guidance.

Rethinking Gauss-Newton for learning over-parameterized models

Michael Arbel (Univ. Grenoble Alpes), Pierre Wolinski (Univ. Grenoble Alpes)

OptimizationTabularStochastic Differential Equation

🎯 What it does: This study investigates the global convergence and implicit bias of the Gauss-Newton method in over-parameterized single hidden layer networks, and tests its generalization performance through synthetic regression task experiments.

Rethinking Incentives in Recommender Systems: Are Monotone Rewards Always Beneficial?

Fan Yao (University of Virginia), Haifeng Xu (University of Chicago)

Recommendation SystemOptimizationTabular

🎯 What it does: This paper studies how to guide content creators to make content creation decisions that are more beneficial to social welfare in a competitive environment by designing reward mechanisms in recommendation systems. It proposes and analyzes the Backward Reward Mechanism (BRM) and its subclass BRCM.

Rethinking Semi-Supervised Imbalanced Node Classification from Bias-Variance Decomposition

Divin Yan (Fudan University), Zengfeng Huang (Fudan University)

ClassificationGraph Neural NetworkContrastive LearningGraph

🎯 What it does: A theoretical framework based on bias-variance decomposition is proposed to address the issue of class imbalance in node classification within graph neural networks, utilizing graph data augmentation to estimate model variance and incorporating a regularization term;

Rethinking Semi-Supervised Medical Image Segmentation: A Variance-Reduction Perspective

Chenyu You (Yale University), James s Duncan

SegmentationContrastive LearningBiomedical DataMagnetic Resonance Imaging

🎯 What it does: A semi-supervised contrastive learning framework called ARCO is proposed based on hierarchical grouping theory to improve the labeling efficiency and model robustness in medical image segmentation.

Rethinking the Backward Propagation for Adversarial Transferability

Xiaosen Wang (Huawei), Kun He (Huazhong University of Science and Technology)

Adversarial AttackImage

🎯 What it does: The attack modifies the backpropagation process to reduce the truncation of gradients caused by nonlinear layers (ReLU, max-pooling), thereby enhancing the transferability of the generated adversarial samples across different models.

Rethinking the Role of Token Retrieval in Multi-Vector Retrieval

Jinhyuk Lee (Google), Vincent Y Zhao

RetrievalComputational EfficiencyTransformerSupervised Fine-TuningText

🎯 What it does: Proposes the XTR model, which improves the three-stage process of multi-vector retrieval by directly scoring the retrieved tokens, eliminating the step of collecting all tokens.

Rethinking Tokenizer and Decoder in Masked Graph Modeling for Molecules

Zhiyuan Liu (National University of Singapore), Tat-Seng Chua (National University of Singapore)

Drug DiscoveryGraph Neural NetworkTransformerAuto EncoderGraph

🎯 What it does: This paper explores the key components in Masked Graph Modeling (MGM) for molecular graphs—graph tokenizers and decoders—and proposes the SimSGT framework.

ReTR: Modeling Rendering Via Transformer for Generalizable Neural Surface Reconstruction

Yixun Liang (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)

TransformerPoint Cloud

🎯 What it does: A Transformer-based rendering framework called ReTR is proposed, improving the traditional volume rendering process to achieve generalizable neural surface reconstruction.

Retrieval-Augmented Multiple Instance Learning

Yufei CUI, Antoni B. Chan (City University of Hong Kong)

ClassificationDomain AdaptationImageBiomedical DataMagnetic Resonance Imaging

🎯 What it does: This paper proposes an optimal transport-based retrieval-enhanced multi-instance learning framework, RAM-MIL, aimed at improving weakly supervised classification of pathological images.

RETVec: Resilient and Efficient Text Vectorizer

Elie Bursztein (Google), Alexey Kurakin (Google)

ClassificationComputational EfficiencyRepresentation LearningAdversarial AttackRecurrent Neural NetworkTransformerLarge Language ModelText

🎯 What it does: A new multilingual text vectorizer, RETVec, has been developed, utilizing a custom UTF-8 binary character encoding and an optional lightweight model to embed words into a 256-dimensional space, achieving robustness against misspellings and adversarial attacks.

Reusable Slotwise Mechanisms

Trang Nguyen (Mila - Quebec AI Institute), Yoshua Bengio (Mila - Quebec AI Institute)

Object DetectionObject TrackingTransformerReinforcement LearningVision-Language-Action ModelImageVideo

🎯 What it does: A modular dynamic model called Reusable Slotwise Mechanisms (RSM) is proposed, which utilizes Central Context Information (CCI) for slot communication and serializes updates to object slots through a reusable multilayer perceptron mechanism, achieving high-quality future frame prediction, visual question answering, and action planning.

Reusing Pretrained Models by Multi-linear Operators for Efficient Training

Yu Pan (Huawei Noah's Ark Lab), Qun Liu (Huawei Noah's Ark Lab)

OptimizationComputational EfficiencyTransformerImageText

🎯 What it does: This paper proposes a complete mapping method that uses a multilinear operator (Mango) to map all weights of a small pre-trained model to a larger model, thereby accelerating the training of the large model while maintaining functionality.

RevColV2: Exploring Disentangled Representations in Masked Image Modeling

Qi Han (MEGVII Technology), Xiangyu Zhang (MEGVII Technology)

Object DetectionSegmentationKnowledge DistillationRepresentation LearningTransformerAuto EncoderImage

🎯 What it does: A novel RevColV2 architecture is designed, utilizing symmetric reversible column encoders and decoders in masked image modeling (MIM) pre-training, while keeping the entire autoencoder network intact during pre-training and downstream tasks without being pruned.

Reverse Engineering Self-Supervised Learning

Ido Ben-Shaul (Tel Aviv University), Yann LeCun (New York University)

ClassificationRepresentation LearningContrastive LearningImage

🎯 What it does: Analyzing the clustering behavior of representations learned by SSL at different levels and semantic categories

Reversible and irreversible bracket-based dynamics for deep graph neural networks

Anthony Gruber (Sandia National Laboratories), Nathaniel Trask (University of Pennsylvania)

Graph Neural NetworkAuto EncoderGraphPhysics Related

🎯 What it does: A deep graph neural network architecture based on structure-preserving bracket dynamics is proposed, utilizing external calculus to map graph attention to energy-conserving or dissipative dynamics;

Revisit the Power of Vanilla Knowledge Distillation: from Small Scale to Large Scale

Zhiwei Hao (Beijing Institute of Technology), Yunhe Wang (University of Sydney)

Object DetectionSegmentationKnowledge DistillationConvolutional Neural NetworkTransformerImage

🎯 What it does: This paper re-evaluates the performance of the classic knowledge distillation method (vanilla KD) on large-scale datasets and points out that evaluations on small-scale datasets underestimate its capabilities.

Revisit Weakly-Supervised Audio-Visual Video Parsing from the Language Perspective

Yingying Fan (Wuhan University), Yutian Lin (Wuhan University)

RecognitionSegmentationData-Centric LearningPrompt EngineeringVision Language ModelContrastive LearningVideoMultimodalityAudio

🎯 What it does: A weakly supervised audio-visual video parsing method based on language prompts is proposed, which significantly reduces segment-level noise by constructing various event occurrence prompts and utilizing CLIP/CLAP for segment-level label inference and dynamic weighting.

Revisiting Adversarial Robustness Distillation from the Perspective of Robust Fairness

Xinli Yue (Wuhan University), Lingchen Zhao (Wuhan University)

Knowledge DistillationAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: This paper conducts an in-depth study on the robust fairness issue of student models in adversarial robust distillation (ARD) and proposes a class difficulty-based reweighting strategy called Fair-ARD to enhance the worst-case robustness of student models across different classes.

Revisiting Adversarial Training for ImageNet: Architectures, Training and Generalization across Threat Models

Naman Deep Singh, Matthias Hein (University of Tuebingen)

Adversarial AttackConvolutional Neural NetworkTransformerImage

🎯 What it does: The paper re-evaluates the effectiveness of adversarial training on ImageNet, systematically comparing two modern architectures: ViT and ConvNeXt, and enhancing adversarial robustness through improvements in model stem, pre-training, data augmentation, and training schemes.

Revisiting Area Convexity: Faster Box-Simplex Games and Spectrahedral Generalizations

Arun Jambulapati (Simons Institute), Kevin Tian (University of Texas at Austin)

Optimization

🎯 What it does: The application of regional convexity in optimization problems is studied, and a new algorithm is proposed to solve the box-simplex game, along with improvements to the solvers for related subproblems.

Revisiting Implicit Differentiation for Learning Problems in Optimal Control

Ming Xu (Australian National University), Stephen Gould (Australian National University)

OptimizationTime Series

🎯 What it does: A new IDOC method is proposed, which directly solves the trajectory derivatives of constrained discrete-time optimal control problems using the implicit function theorem, avoiding traditional Riccati recursion.

Revisiting Logistic-softmax Likelihood in Bayesian Meta-Learning for Few-Shot Classification

Tianjun Ke (Renmin University of China), Feng Zhou (Renmin University of China)

ClassificationMeta LearningImage

🎯 What it does: This paper redesigns the logistic-softmax likelihood function with a temperature parameter and applies it to a deep kernel-based Gaussian process meta-learning framework, proposing a data augmentation-based analytical mean-field approximate inference method.

Revisiting Scalarization in Multi-Task Learning: A Theoretical Perspective

Yuzheng Hu (University of Illinois), Han Zhao (University of Illinois)

OptimizationRobotic IntelligenceTabular

🎯 What it does: This study investigates whether linear scalarization (i.e., the weighted sum of the losses of each task) in linear multi-task learning (MTL) can fully cover the Pareto front.

Revisiting the Minimalist Approach to Offline Reinforcement Learning

Denis Tarasov (Tinkoff), Sergey Kolesnikov (Tinkoff)

Reinforcement LearningTabular

🎯 What it does: In this study, the authors built upon the TD3+BC minimal offline reinforcement learning algorithm by incorporating a series of implementation detail improvements that have emerged in recent years, resulting in a new offline RL method called ReBRAC.

Revisiting Visual Model Robustness: A Frequency Long-Tailed Distribution View

Zhiyu Lin (Beijing Jiaotong University), Jitao Sang (Peng Cheng Lab)

ClassificationObject DetectionAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: To address the robustness issue of visual models, the author treats the frequency domain as a long-tail distribution, defining and studying the relationship between underfitting of high-frequency components (HFC) and robustness, and proposes a spectral sampling strategy called Balanced Spectral Sampling (BaSS) to enhance the trade-off between model robustness and accuracy.

Reward Finetuning for Faster and More Accurate Unsupervised Object Discovery

Katie Z Luo (Cornell University), Kilian Q Weinberger (Cornell University)

Object DetectionAutonomous DrivingReinforcement LearningPoint Cloud

🎯 What it does: This paper proposes an unsupervised 3D object detection method called DRIFT, which utilizes reward signals for fine-tuning.

Reward Imputation with Sketching for Contextual Batched Bandits

Xiao Zhang (Renmin University of China), Ji-Rong Wen (Renmin University of China)

Recommendation SystemOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular

🎯 What it does: A reward missing compensation method (SPUIR) is proposed for the partial information feedback problem in Contextual Batched Bandits (CBB), which imputes the rewards of unexecuted actions and updates the policy.

Reward Scale Robustness for Proximal Policy Optimization via DreamerV3 Tricks

Ryan Sullivan (University of Maryland), Joseph Suarez (Massachusetts Institute of Technology)

OptimizationReinforcement LearningTabular

🎯 What it does: This study transfers various stability techniques from DreamerV3 to PPO and evaluates their impact on the performance and robustness of PPO.

Reward-agnostic Fine-tuning: Provable Statistical Benefits of Hybrid Reinforcement Learning

Gen Li (Chinese University of Hong Kong), Yuxin Chen (University of Pennsylvania)

Reinforcement LearningTabular

🎯 What it does: This paper studies tabular reinforcement learning in mixed reinforcement learning (RL) and proposes a three-stage mixed RL algorithm that outperforms pure offline RL and pure online RL in terms of sample complexity.

Reward-Directed Conditional Diffusion: Provable Distribution Estimation and Reward Improvement

Hui Yuan (Princeton University), Mengdi Wang (Princeton University)

GenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelScore-based ModelImage

🎯 What it does: This paper proposes a semi-supervised learning framework: first, it learns the reward function using a small amount of data with noisy reward labels, then it generates pseudo-labels for a large amount of unlabeled data, and subsequently trains a reward-conditioned diffusion model with the pseudo-labels to achieve 'reward-guided' generation; theoretical guarantees for this method in reward distribution estimation and reward improvement are also provided.

Rewarded soups: towards Pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards

Alexandre Rame (Sorbonne Université), Matthieu Cord (Valeo)

OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelImageText

🎯 What it does: By separately fine-tuning the same pre-trained model on different proxy rewards through reinforcement learning, and then performing linear interpolation in the weight space, a set of models that can approach Pareto optimality across various rewards is obtained.

Rewiring Neurons in Non-Stationary Environments

Zhicheng Sun (Peking University), Yadong MU

Reinforcement LearningMultimodality

🎯 What it does: In continual reinforcement learning, a method is proposed to reconnect the network through learnable permutations of hidden layer neurons, achieving a synergy between structural plasticity and weight learning.

Rewrite Caption Semantics: Bridging Semantic Gaps for Language-Supervised Semantic Segmentation

Yun Xing (Nanyang Technological University), Shijian Lu (Nanyang Technological University)

SegmentationRetrievalTransformerContrastive LearningImageTextMultimodality

🎯 What it does: A concept curation pipeline is proposed, utilizing CLIP to bridge the semantic gap between images and text through vision-driven expansion, text-to-vision guided relevance ranking, and clustering sampling, in order to enhance the zero-shot performance of language-supervised semantic segmentation.

REx: Data-Free Residual Quantization Error Expansion

Edouard YVINEC, Kevin Bailly (Datakalab)

ClassificationObject DetectionSegmentationComputational EfficiencyData-Centric LearningLarge Language ModelImageText

🎯 What it does: A data-free quantization method REx is proposed, which achieves flexible precision-speed trade-offs at different bit widths by recursively expanding quantization residuals and incorporating sparse grouping.

RGMIL: Guide Your Multiple-Instance Learning Model with Regressor

Zhaolong Du (Xidian University), Lin Xiong (Xidian University)

ClassificationRecognitionConvolutional Neural NetworkImage

🎯 What it does: A regression-guided multi-instance learning network (RGMIL) is proposed, utilizing a new aggregator called Regressor-Guided Pooling (RGP) to achieve high-quality instance-level representation and bag-level classification.

RH-BrainFS: Regional Heterogeneous Multimodal Brain Networks Fusion Strategy

Hongting Ye (Southeast University), Yonggui Yuan (Southeast University)

ClassificationGraph Neural NetworkTransformerMultimodalityBiomedical DataMagnetic Resonance Imaging

🎯 What it does: The RH-BrainFS model is proposed, which extracts regional features through brain graph networks and utilizes a Transformer combined with a fusion bottleneck to integrate structural (SC) and functional (FC) brain networks, addressing the issue of regional heterogeneity between the two modalities.

Riemannian Laplace approximations for Bayesian neural networks

Federico Bergamin (Technical University of Denmark), Georgios Arvanitidis (Technical University of Denmark)

TabularOrdinary Differential Equation

🎯 What it does: A Riemannian geometry-based Laplace approximation is proposed to better approximate the weight posterior in Bayesian neural networks;

Riemannian Projection-free Online Learning

Zihao Hu (Georgia Institute of Technology), Jacob Abernethy (Google Research)

Optimization

🎯 What it does: Designed a projection-free online learning algorithm on Riemannian geometry, utilizing separation or linear optimization operators to implement feasibility constraints and provided a sublinear regret upper bound.

Riemannian Residual Neural Networks

Isay Katsman (Yale University), Christopher De Sa (Cornell University)

VideoGraphOrdinary Differential Equation

🎯 What it does: The paper proposes a general Riemannian Residual Neural Network (RResNet), which extends the additive operation of traditional residual networks to any smooth Riemannian manifold and achieves geometrically consistent residual connections through the exponential map.

Riemannian SAM: Sharpness-Aware Minimization on Riemannian Manifolds

Jihun Yun (KAIST), Eunho Yang (KAIST)

OptimizationGraph

🎯 What it does: This paper studies and implements the Sharpness-Aware Minimization (Riemannian SAM) generalized on Riemannian manifolds, providing an example on Lorentz manifolds (Lorentz SAM) and a theoretical explanation for Fisher SAM.

Riemannian stochastic optimization methods avoid strict saddle points

Ya-Ping Hsieh (ETH Zurich), Panayotis Mertikopoulos (University Grenoble Alpes)

Optimization

🎯 What it does: This paper proposes and analyzes a class of Riemannian Robbins-Monro (RRM) stochastic optimization methods, proving that under mild assumptions, it almost surely avoids strict saddle points and only converges to local minima.

Rigorous Runtime Analysis of MOEA/D for Solving Multi-Objective Minimum Weight Base Problems

Anh Viet Do (University of Adelaide), Andrew M. Sutton (University of Minnesota Duluth)

OptimizationGraph

🎯 What it does: This study investigates the multi-objective minimum weight base problem, providing a rigorous runtime analysis of the MOEA/D algorithm and proving that it can find all extreme points in fixed-parameter polynomial time.

Risk-Averse Active Sensing for Timely Outcome Prediction under Cost Pressure

Yuchao Qin (University of Cambridge), Changhee Lee (Chung-Ang University)

Reinforcement LearningTime SeriesBiomedical DataAlzheimer's Disease

🎯 What it does: A risk-averse active perception framework RAS is proposed to achieve timely and accurate outcome predictions under cost pressure.

Risk-Averse Model Uncertainty for Distributionally Robust Safe Reinforcement Learning

James Queeney (Boston University), Mouhacine Benosman (Mitsubishi Electric Research Laboratories)

Safty and PrivacyReinforcement Learning

🎯 What it does: This paper proposes a safe reinforcement learning framework under model uncertainty (RAMU), which processes model distribution with risk aversion, ensuring that the learned policy maintains safety constraints while achieving robust performance during deployment.

RiskQ: Risk-sensitive Multi-Agent Reinforcement Learning Value Factorization

Siqi Shen (Xiamen University), Cheng Wang (Xiamen University)

Reinforcement LearningSequential

🎯 What it does: This paper proposes a risk-sensitive multi-agent reinforcement learning value decomposition framework called RiskQ, and addresses the issue of traditional value decomposition methods being unable to meet the requirements of collaborative decision-making in the face of risk measures (such as VaR and distorted risk measures) by defining the Risk-Sensitive Individual-Global Maximization (RIGM) principle.