ICLR 2023 Papers — Page 12
International Conference on Learning Representations · 1573 papers
Plateau in Monotonic Linear Interpolation --- A "Biased" View of Loss Landscape for Deep Networks
Xiang Wang (Duke University), Rong Ge (Duke University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the phenomenon of 'plateau' that occurs during the training of deep networks when linear interpolation is applied. It reveals that the root cause is mainly related to the differences in the initialization scale, depth of the network, and various final layer biases. By analyzing a simplified r-homogeneous weight network, it proves that gradient flow can create bias differences among different categories on balanced datasets, and it explains that the categories learned later often have the largest bias. Experiments validate the theoretical predictions and further explore the impact of bias interpolation methods, initialization scale, and network depth on the length of the plateau.
PLOT: Prompt Learning with Optimal Transport for Vision-Language Models
Guangyi Chen (Carnegie Mellon University), Kun Zhang
ClassificationDomain AdaptationOptimizationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: By learning multiple descriptive prompts for each category within the CLIP framework and using optimal transport (Sinkhorn algorithm) to achieve fine-grained matching between visual features and text prompts, the issue of feature redundancy caused by the convergence of a single prompt is avoided.
Policy Expansion for Bridging Offline-to-Online Reinforcement Learning
Haichao Zhang (Horizon Robotics), Haonan Yu (Horizon Robotics)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes a Policy Expansion framework, which freezes the policy obtained from offline training and combines it with newly learned policies to form a policy set. It achieves the transition from offline to online reinforcement learning through value-guided adaptive combination, realizing the retention of offline knowledge and the synergy of online exploration.
Policy Pre-training for Autonomous Driving via Self-supervised Geometric Modeling
Penghao Wu (Shanghai AI Laboratory), Yu Qiao (Shanghai AI Laboratory)
Pose EstimationDepth EstimationAutonomous DrivingConvolutional Neural NetworkReinforcement LearningContrastive LearningVideo
🎯 What it does: A PPGeo pre-training framework is proposed, utilizing self-supervised geometric modeling (estimating depth, camera intrinsics, and pose) in a two-stage training process, ultimately allowing the visual encoder to predict future poses based solely on a single frame, thereby learning visual representations highly relevant to driving decisions.
Policy-Based Self-Competition for Planning Problems
Jonathan Pirnay (Technical University of Munich), Dominik Gerhard Grimm
OptimizationReinforcement LearningTabular
🎯 What it does: A self-competitive planning method based on Gumbel AlphaZero, GAZ Play-to-Plan, is proposed, which enhances the search quality of single-agent planning tasks by utilizing historical policy duels instead of a single scalar benchmark.
POPGym: Benchmarking Partially Observable Reinforcement Learning
Steven Morad, Amanda Prorok
Recurrent Neural NetworkReinforcement LearningBenchmark
🎯 What it does: Proposed the POPGym framework, which provides 15 multi-level, randomly generated partially observable reinforcement learning (POMDP) environments, and implements 13 memory model baselines, followed by large-scale experimental comparisons based on this benchmark.
Population-size-Aware Policy Optimization for Mean-Field Games
Pengdeng Li (Nanyang Technological University), Bo An (Nanyang Technological University)
OptimizationReinforcement LearningTabular
🎯 What it does: This study proposes a Population-size-aware Policy Optimization (PAPO) method that can generate efficient strategies in finite agent games with varying population sizes, bridging finite agent games and infinite agent (Mean-Field) games by analyzing the evolution of strategies with population size.
Post-hoc Concept Bottleneck Models
Mert Yuksekgonul (Stanford University), James Zou (Stanford University)
Explainability and InterpretabilityData-Centric LearningContrastive LearningImageMultimodality
🎯 What it does: This paper proposes the Post-Concept Bottleneck Model (PCBM), which transforms any pre-trained network into an interpretable concept bottleneck model and restores performance through residual modules.
Powderworld: A Platform for Understanding Generalization via Rich Task Distributions
Kevin Frans (Massachusetts Institute of Technology), Phillip Isola (Massachusetts Institute of Technology)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningWorld ModelPhysics Related
🎯 What it does: A GPU-based powder game physics engine called Powderworld is proposed, and various task distributions are defined on it to study the generalization ability of reinforcement learning and world modeling.
PowerQuant: Automorphism Search for Non-Uniform Quantization
Edouard YVINEC, Kevin Bailly (Datakalab)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: A data-independent non-uniform quantization method called PowerQuant has been developed, which utilizes power functions to automate the search for quantization parameters while maintaining multiplicative invariance, thereby improving the accuracy of low-bit-width networks.
Pre-training via Denoising for Molecular Property Prediction
Sheheryar Zaidi (DeepMind), Jonathan Godwin (DeepMind)
Drug DiscoveryGraph Neural NetworkTransformerScore-based ModelAuto EncoderGraph
🎯 What it does: A self-supervised pre-training method based on denoising 3D molecular structures is proposed, and the pre-trained model is used for molecular property prediction.
Predicting Cellular Responses with Variational Causal Inference and Refined Relational Information
Yulun Wu (University of California), George Karypis (Amazon)
Graph Neural NetworkAuto EncoderBiomedical Data
🎯 What it does: A framework based on graph variational Bayesian causal inference is proposed to predict gene expression in single cells under unobserved interference.
Predictive Inference with Feature Conformal Prediction
Jiaye Teng (Tsinghua University), Yang Yuan (Tsinghua University)
ClassificationSegmentationConvolutional Neural NetworkImageTabular
🎯 What it does: Proposes the Feature Conformal Prediction (Feature CP) framework, which performs distributed confidence interval inference in the semantic feature space of deep models rather than the traditional output space, and provides corresponding non-conformity scores and confidence interval estimation/detection methods.
Predictor-corrector algorithms for stochastic optimization under gradual distribution shift
Subha Maity (University of Michigan), Yuekai Sun (University of Michigan)
Object TrackingOptimizationTabularTime Series
🎯 What it does: A Predictor-Corrector algorithm is proposed to solve stochastic optimization problems with smooth changes over time.
Preference Transformer: Modeling Human Preferences using Transformers for RL
Changyeon Kim (Korea Advanced Institute of Science and Technology), Kimin Lee (Google Research)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes the Preference Transformer, which models human preferences for behavior trajectories using a Transformer, learning non-Markovian reward functions and automatically assigning importance weights.
Preserving Pre-trained Features Helps Calibrate Fine-tuned Language Models
Guande He (Tsinghua University), Jun Zhu (Tsinghua University)
Knowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningText
🎯 What it does: This paper studies the issue of pre-trained language models (PLMs) losing good calibration of uncertainty (over-confidence) during the fine-tuning process and explores how maintaining pre-trained features can enhance the calibration performance of fine-tuned models.
Priors, Hierarchy, and Information Asymmetry for Skill Transfer in Reinforcement Learning
Sasha Salter (University of Oxford), Ingmar Posner (University of Oxford)
Reinforcement LearningSequential
🎯 What it does: By combining hierarchical KL regularization with learnable attention gates (information asymmetry) in multi-task reinforcement learning, a balance between expressive capability and transfer performance in continuous task transfer is achieved.
Private Federated Learning Without a Trusted Server: Optimal Algorithms for Convex Losses
Andrew Lowy (University of Southern California), Meisam Razaviyayn (University of Southern California)
OptimizationFederated LearningSafty and PrivacyTabularSequential
🎯 What it does: This paper studies cross-island federated learning under untrusted servers, proposing cross-island record-level differential privacy (ISRL-DP) and providing an optimal algorithm.
Proactive Multi-Camera Collaboration for 3D Human Pose Estimation
Hai Ci (Peking University), Yizhou Wang (Peking University)
Pose EstimationRecurrent Neural NetworkReinforcement LearningSimultaneous Localization and MappingVideo
🎯 What it does: This study proposes an active multi-camera collaboration framework based on multi-agent reinforcement learning for 3D human pose estimation in dynamic crowds.
Probabilistically Robust Recourse: Navigating the Trade-offs between Costs and Robustness in Algorithmic Recourse
Martin Pawelczyk (University of Tuebingen), Himabindu Lakkaraju (Harvard University)
OptimizationRecurrent Neural NetworkReinforcement LearningTabular
🎯 What it does: Proposes the Probabilistically Robust Recourse (PROBE) framework, allowing users to specify the desired recourse failure rate (r), and generates low-cost and robust recourse through closed-form expressions and gradient optimization;
Programmatically Grounded, Compositionally Generalizable Robotic Manipulation
Renhao Wang (Tsinghua University), Yang Gao (Shanghai Qi Zhi Institute)
Robotic IntelligenceReinforcement LearningVision Language ModelText
🎯 What it does: A modular robotic manipulation framework called PROGRAMPORT is constructed based on programmatic semantic parsing, utilizing the pre-trained CLIP visual-language model for fine-grained visual semantic attribution, and executing robotic arm actions through executable programs.
Progress measures for grokking via mechanistic interpretability
Neel Nanda, Jacob Steinhardt (University of California)
OptimizationExplainability and InterpretabilityTransformerTabular
🎯 What it does: By reverse engineering the learning algorithm of a small transformer in the modular addition task, it was found that the model actually implements Fourier multiplication: mapping the input to two-dimensional rotation vectors and using trigonometric identities to perform addition, then obtaining the correct modular result through the decoder.
Progressive Mix-Up for Few-Shot Supervised Multi-Source Domain Transfer
Ronghang Zhu (University of Georgia), Sheng Li (University of Virginia)
Domain AdaptationMeta LearningConvolutional Neural NetworkImage
🎯 What it does: A knowledge transfer method under multi-source domains with very few target samples (few-shot) is proposed—Progressive Mix-up (P-Mixup).
Progressive Prompts: Continual Learning for Language Models
Anastasia Razdaibiedina (University of Toronto), Amjad Almahairi (Meta AI)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes Progressive Prompts, a method for continuous learning by learning soft prompts for each task and sequentially concatenating them to the input while keeping the model frozen.
Progressive Voronoi Diagram Subdivision Enables Accurate Data-free Class-Incremental Learning
Chunwei Ma (University at Buffalo), Jinhui Xu (University at Buffalo)
ClassificationImage
🎯 What it does: This paper proposes a novel data-free incremental learning framework called iVoro, which is based on the stepwise construction of Voronoi diagrams. It achieves high-accuracy class incremental learning without storing old data and while maintaining a constant model capacity.
Progressively Compressed Auto-Encoder for Self-supervised Representation Learning
Jin Li (Shanghai Jiao Tong University), Qi Tian (Huawei Cloud)
Object DetectionRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: This paper proposes the Progressively Compressed Auto-Encoder (PCAE), which significantly reduces the number of reconstruction targets and improves training efficiency by layer-wise dropping redundant tokens during self-supervised pre-training, retaining only the information necessary for reconstruction.
Projective Proximal Gradient Descent for Nonconvex Nonsmooth Optimization: Fast Convergence Without Kurdyka-Lojasiewicz (KL) Property
Yingzhen Yang (Arizona State University), Ping Li (LinkedIn Ads)
OptimizationImage
🎯 What it does: Proposes Projected Proximal Gradient Descent (PPGD) to solve non-convex non-smooth optimization problems with piecewise convex regularization, providing a new convergence analysis that does not rely on KŁ properties;
Prompt-to-Prompt Image Editing with Cross-Attention Control
Amir Hertz (Google Research), Daniel Cohen-or
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: By injecting cross-attention maps of the original image into the text-conditioned diffusion model, local or global editing of images can be achieved solely based on text prompts;
Promptagator: Few-shot Dense Retrieval From 8 Examples
Zhuyun Dai (Google Research), Ming-Wei Chang (Google Research)
RetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Using a small number (≤8) of labeled query-document examples, we generate prompt-based queries with the help of a large language model (FLAN 137B), thereby synthesizing a massive amount of training samples for any retrieval task without traditional large-scale annotated data, ultimately training a dual-encoder retriever and a cross-attention re-ranker.
Prompting GPT-3 To Be Reliable
Chenglei Si (University of Maryland), Lijuan Wang (Microsoft)
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This study proposes various prompting strategies to enhance GPT-3's performance across four reliability dimensions (generalization, bias, fairness, calibration, and factuality).
Proposal-Contrastive Pretraining for Object Detection from Fewer Data
Quentin Bouniot (Universite Paris-Saclay), Amaury Habrard (Universite Jean Monnet Saint-Etienne)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: ProSeCo is proposed, an unsupervised pre-training method for Transformer object detectors that utilizes a large number of candidate boxes for contrastive learning and incorporates positional information.
Protein Representation Learning by Geometric Structure Pretraining
Zuobai Zhang (Mila - Quebec Artificial Intelligence Institute), Jian Tang (HEC Montreal)
Representation LearningProtein Structure PredictionGraph Neural NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: Designed and pre-trained a graph neural network based on protein 3D structures, GearNet, enhancing protein representation through multi-view contrastive learning and self-supervised mask prediction.
Protein Representation Learning via Knowledge Enhanced Primary Structure Reasoning
Hong-Yu Zhou (University of Hong Kong), Yizhou Yu (University of Hong Kong)
Representation LearningDrug DiscoveryProtein Structure PredictionTransformerLarge Language ModelBiomedical Data
🎯 What it does: Pre-training on protein sequences utilizes relational and attribute words from knowledge graphs, employing cross-modal attention for token-level knowledge probing of each amino acid, thereby learning more semantically rich protein representations.
Protein Sequence and Structure Co-Design with Equivariant Translation
Chence Shi (BioGeometry), Jian Tang (HEC Montréal)
Protein Structure PredictionBiomedical Data
🎯 What it does: This paper proposes a new framework for the joint design of protein sequences and structures called PROTSEED, which directly generates target sequences and structures from random initialization using a deformable iterative translation approach.
Proto-Value Networks: Scaling Representation Learning with Auxiliary Tasks
Jesse Farebrother (McGill University), Marc G Bellemare
Representation LearningConvolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: A new auxiliary task framework called Proto-Value Networks (PVN) is proposed, which learns rich state representations by predicting value through the successor measure of random policies.
Prototypical Calibration for Few-shot Learning of Language Models
Zhixiong Han (Microsoft Research), Furu Wei (Microsoft Research)
ClassificationTransformerLarge Language ModelText
🎯 What it does: A prototype-based calibration method is provided for few-shot learning in GPT-style language models, which estimates the prototype clusters for each category and adjusts the decision boundaries to improve classification accuracy.
Provable Defense Against Geometric Transformations
Rem Yang (University of Illinois Urbana Champaign), Gagandeep Singh (University of Illinois Urbana Champaign)
ClassificationAutonomous DrivingImage
🎯 What it does: A deterministic verifiable geometric robustness training framework (CGT) is proposed, along with a GPU-accelerated geometric robustness verifier (FGV), achieving a complete process that directly embeds geometric robustness guarantees during training.
Provable Memorization Capacity of Transformers
Junghwan Kim (University of Michigan), Barzan Mozafari (University of Michigan)
ClassificationRecognitionTransformerTextSequential
🎯 What it does: Prove the memory capacity of the Transformer model under finite precision, providing theoretical upper limits for tasks such as sequence-to-sequence mapping, sequence classification, and language modeling, and experimentally validating on the NER and MNLI datasets.
Provable Robustness against Wasserstein Distribution Shifts via Input Randomization
Aounon Kumar (University of Maryland), Soheil Feizi (University of Maryland)
ClassificationDomain AdaptationAdversarial AttackImage
🎯 What it does: A distribution-level robustness framework based on input randomization is proposed, which provides a provable lower bound on accuracy for distribution shifts constrained by Wasserstein distance.
Provable Sim-to-real Transfer in Continuous Domain with Partial Observations
Jiachen Hu (Peking University), Liwei Wang (Peking University)
Domain AdaptationRobotic IntelligenceReinforcement Learning
🎯 What it does: The research addresses the simulation-to-real transfer problem in continuous domains with partial observations, providing theoretical guarantees.
Provably Auditing Ordinary Least Squares in Low Dimensions
Ankur Moitra (Massachusetts Institute of Technology), Dhruv Rohatgi (Massachusetts Institute of Technology)
Tabular
🎯 What it does: The study provides a provable audit and estimation of the stability of the ordinary least squares regression model in low dimensions (constant dimensions), specifically whether the coefficient signs can change after removing the minimum number of samples.
Provably Efficient Lifelong Reinforcement Learning with Linear Representation
Sanae Amani (University of California), Ching-An Cheng (Microsoft Research)
Reinforcement Learning
🎯 What it does: An algorithm for lifelong reinforcement learning, UCBlvd, is proposed, achieving sublinear cumulative loss and sublinear planning calls on adversarial task sequences.
Provably Efficient Risk-Sensitive Reinforcement Learning: Iterated CVaR and Worst Path
Yihan Du (Tsinghua University), Longbo Huang (Tsinghua University)
Reinforcement Learning
🎯 What it does: Proposes an iterative CVaR reinforcement learning framework and designs corresponding regret minimization and optimal policy identification algorithms, studying its optimal lower bounds.
Pruning Deep Neural Networks from a Sparsity Perspective
Enmao Diao (Duke University), Vahid Tarokh (Duke University)
CompressionOptimizationConvolutional Neural NetworkImage
🎯 What it does: A new sparsity measure called PQ Index (PQI) is proposed, and based on this, an adaptive pruning algorithm (SAP) is developed. The PQI dynamically determines the pruning ratio at each step, achieving a higher compression rate while maintaining accuracy.
Pseudo-label Training and Model Inertia in Neural Machine Translation
Benjamin Hsu (Amazon Web Services Artificial Intelligence Labs), Georgiana Dinu (Amazon Web Services Artificial Intelligence Labs)
TransformerText
🎯 What it does: This paper systematically studies the impact of Pseudo-Label Training (PLT) on the stability (model inertia) of neural machine translation (NMT) models, demonstrating that PLT can significantly enhance the model's robustness to input perturbations and output consistency during retraining; it also explores the relationship between this effect and the simplification of training data distribution.
Pseudoinverse-Guided Diffusion Models for Inverse Problems
Jiaming Song, Jan Kautz
RestorationSuper ResolutionDiffusion modelScore-based ModelImage
🎯 What it does: A pseudo-inverse guided diffusion model (Π GDM) is proposed, which directly estimates conditional scores through a problem-agnostic diffusion model combined with the pseudo-inverse of the measurement model, addressing various inverse problems.
Pushing the Accuracy-Group Robustness Frontier with Introspective Self-play
Jeremiah Zhe Liu (Google Research), Deepak Ramachandran (Google Research)
Data-Centric LearningTabular
🎯 What it does: This paper proposes the Introspective Self-play (ISP) method, which incorporates a self-reflection task into model training, enabling the model to identify and predict whether samples come from minority groups, thereby improving uncertainty estimation under data bias and the sampling rate of tail samples in active learning.
Pushing the Limits of Fewshot Anomaly Detection in Industry Vision: Graphcore
Guoyang Xie (University of Surrey), Feng Zheng (Southern University of Science and Technology)
Anomaly DetectionGraph Neural NetworkImage
🎯 What it does: A few-shot industrial visual anomaly detection framework based on graph neural networks, GraphCore, is proposed, using rotation-invariant visual features (VIIF) as anomaly measurement features.
PV3D: A 3D Generative Model for Portrait Video Generation
Zhongcong Xu, Mike Zheng Shou
GenerationData SynthesisGenerative Adversarial NetworkVideo
🎯 What it does: A PV3D model is proposed, which can learn from monocular 2D videos to generate multi-view consistent, dynamic 3D portrait videos.
Q-Pensieve: Boosting Sample Efficiency of Multi-Objective RL Through Memory Sharing of Q-Snapshots
Wei Hung (National Yang Ming Chiao Tung University), Xi Liu (Meta AI)
Reinforcement Learning
🎯 What it does: The Q-Pensieve algorithm is proposed, which achieves policy-level knowledge sharing by using Q-snapshots in multi-objective reinforcement learning, significantly improving sample efficiency.
QAID: Question Answering Inspired Few-shot Intent Detection
Asaf Yehudai (IBM Israel Research Lab), Boaz Carmeli (IBM Israel Research Lab)
ClassificationRetrievalTransformerContrastive LearningText
🎯 What it does: Reformulate the few-shot intent detection task as a question-answer retrieval task, using the user's utterance as the question and the intent name as the answer, employing a dual-encoder and token-level late interaction for matching.
Quality-Similar Diversity via Population Based Reinforcement Learning
Shuang Wu (Tencent AI Lab), Yang Wei
Reinforcement LearningSequential
🎯 What it does: The Quality-Similarity Diversity (QSD) problem is proposed, aiming to generate a diverse set of strategies at different quality levels.
QuAnt: Quantum Annealing with Learnt Couplings
Marcel Seelbach Benkner (Universitat Siegen), Vladislav Golyanik (MPI for Informatics)
OptimizationMeta LearningContrastive LearningPoint CloudGraph
🎯 What it does: This paper proposes a method to regress QUBO matrices using neural networks, solving them with quantum annealing, and achieving end-to-end backpropagation through contrastive loss, thus eliminating the need for manual derivation of QUBO;
Quantifying and Mitigating the Impact of Label Errors on Model Disparity Metrics
Julius Adebayo (Prescient Design / Genentech), Bobbie Chern (Meta Inc.)
ClassificationData-Centric LearningSupervised Fine-TuningTabular
🎯 What it does: This study investigates the impact of label errors on model fairness metrics and proposes a label error identification and automatic correction method based on influence functions.
Quantifying Memorization Across Neural Language Models
Nicholas Carlini (Google Research), Chiyuan Zhang (Google Research)
GenerationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The paper quantifies the memorization degree of large language models under different scales, data redundancy, and context lengths through sampling of the model training set and prefix prompting.
Quantile Risk Control: A Flexible Framework for Bounding the Probability of High-Loss Predictions
Jake Snell, Richard Zemel (Columbia University)
ClassificationObject DetectionImageText
🎯 What it does: This paper proposes a distribution-free Quantile Risk Control (QRC) framework that provides strict upper bounds for quantile risk measures (such as VaR, CVaR, interval VaR) for predictive models, and can simultaneously select and control multiple models.
Quantized Compressed Sensing with Score-Based Generative Models
Xiangming Meng (University of Tokyo), Yoshiyuki Kabashima (University of Tokyo)
RestorationCompressionDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes a method for recovering compressed sensing signals that have been quantized (including 1-bit, 2-bit, 3-bit, etc.) using a pre-trained score-based generative model (SGM), referred to as QCS-SGM.
Quasi-optimal Reinforcement Learning with Continuous Actions
Yuhan Li (University of Illinois Urbana Champaign), Ruoqing Zhu (University of Illinois Urbana Champaign)
Recommendation SystemSafty and PrivacyReinforcement LearningBiomedical Data
🎯 What it does: A continuous action space reinforcement learning algorithm based on a quasi-optimal Bellman operator and kernel embedding is proposed to find an approximately optimal policy in safety-sensitive scenarios such as medical dosage recommendations.
Random Laplacian Features for Learning with Hyperbolic Space
Tao Yu, Christopher De Sa
ClassificationRepresentation LearningGraph Neural NetworkTextGraph
🎯 What it does: This paper proposes a random feature-based mapping called HyLa, which uses hyperplane waves in the hypersphere (Poincaré ball) as feature mappings. The hypersphere is then embedded into Euclidean space and directly input into standard Euclidean graph networks for learning.
RandProx: Primal-Dual Optimization Algorithms with Randomized Proximal Updates
Laurent Condat (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
Optimization
🎯 What it does: A randomized primal-dual algorithm, RandProx, is proposed to solve large-scale convex optimization problems that include both smooth and non-smooth functions, potentially accompanied by linear operators.
Rarity Score : A New Metric to Evaluate the Uncommonness of Synthesized Images
Jiyeon Han (Kim Jaechul Graduate School of AI), Jaesik Choi (NAVER AI Lab)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A rarity score is proposed, which can assess the rarity of a single generated image, generative models, and similar datasets within the distribution of authenticity.
Re-calibrating Feature Attributions for Model Interpretation
Peiyu Yang (University of Western Australia), Ajmal Saeed Mian
Explainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Recalibrated the path integral-based feature attribution method to make the explanations more theoretically aligned and no longer require taking absolute values.
Re-Imagen: Retrieval-Augmented Text-to-Image Generator
Wenhu Chen (Google Research), William W. Cohen (Google Research)
GenerationRetrievalDiffusion modelImageTextRetrieval-Augmented Generation
🎯 What it does: The paper proposes a retrieval-enhanced text-to-image generation model called Re-Imagen, which generates more realistic images by retrieving relevant image-text pairs from an external multimodal knowledge base as conditions.
Re-parameterizing Your Optimizers rather than Architectures
Xiaohan Ding (Tencent AI Lab), Guiguang Ding (Tsinghua University)
OptimizationHyperparameter SearchConvolutional Neural NetworkImage
🎯 What it does: This paper proposes transferring the model's prior knowledge to the optimizer, designing a Gradient Re-parameterization method and implementing RepOptimizers, further validating its effectiveness on the VGG-style pure network RepOpt-VGG.
Re-weighting Based Group Fairness Regularization via Classwise Robust Optimization
Sangwon Jung (Seoul National University), Taesup Moon (Seoul National University)
OptimizationImageTextMultimodalityTabular
🎯 What it does: A unified distributionally robust optimization framework (FairDRO) is proposed, achieving group fairness;
ReAct: Synergizing Reasoning and Acting in Language Models
Shunyu Yao (Princeton University), Yuan Cao (Google Research)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the ReAct framework, which utilizes large language models to alternately generate thoughts and actions between reasoning and action, supporting interaction with external knowledge bases or environments to complete tasks.
Real-Time Image Demoir$\acute{e}$ing on Mobile Devices
Yuxin Zhang (Xiamen University), Rongrong Ji (Xiamen University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This study proposes a dynamic de-Moiré acceleration method (DDA) for real-time de-Moiré on mobile devices: the image is divided into several patches, and a Moiré prior based on frequency and color is used to measure the complexity of the patches, allocating different scales of CNN sub-networks according to complexity; parameter sharing is achieved through a supernet to avoid parameter inflation caused by multiple networks.
Real-time variational method for learning neural trajectory and its dynamics
Matthew Dowling (Stony Brook University), Il Memming Park (Champalimaud Foundation)
Time Series
🎯 What it does: An online variational Kalman filter (eVKF) is proposed, which achieves real-time inference of neural trajectories and their dynamics, while learning dynamic parameters within the same framework.
Recitation-Augmented Language Models
Zhiqing Sun (Google Research), Denny Zhou (Google Research)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: A 'recite-and-answer' paradigm is proposed, where a large language model recalls relevant knowledge paragraphs from its own weights and then answers questions based on the recalled content, forming the 'RECITE' framework.
Recon: Reducing Conflicting Gradients From the Root For Multi-Task Learning
Guangyuan SHI, Xiao-Ming Wu (Hong Kong Polytechnic University)
OptimizationImage
🎯 What it does: This paper proposes the Recon method, which reduces gradient conflicts in multi-task learning and improves performance by first calculating the task gradient conflict scores for each layer and selecting the most severely conflicted shared layers to convert them into task-specific layers.
Recursive Time Series Data Augmentation
Amine Mohamed Aboussalah (New York University), Chi-Guhn Lee (University of Toronto)
Generative Adversarial NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: A recursive interpolation method (RIM) was designed and validated for time series data augmentation, enhancing the model's generalization ability in data-scarce situations.
Red PANDA: Disambiguating Image Anomaly Detection by Removing Nuisance Factors
Niv Cohen (Hebrew University of Jerusalem), Yedid Hoshen (Hebrew University of Jerusalem)
Anomaly DetectionContrastive LearningImage
🎯 What it does: This paper proposes an anomaly detection method using known irrelevant attribute labels for domain-supervised decomposition (Red PANDA). It learns representations that are insensitive to noise factors through contrastive learning and reconstruction loss, and employs kNN density estimation for anomaly scoring.
Regression with Label Differential Privacy
Badih Ghazi (Google), Chiyuan Zhang (Google)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper proposes a label differential privacy mechanism for regression tasks—Random Response on Bins (RR on Bins) based on prior distribution, and provides a dynamic programming algorithm to solve for the optimal binning.
Relational Attention: Generalizing Transformers for Graph-Structured Tasks
Cameron Diao (Rice University), Ricky Loynd (Microsoft Research)
Graph Neural NetworkTransformerGraph
🎯 What it does: A new relational attention mechanism is proposed, enabling the Transformer to simultaneously handle features of nodes and edges, thus working directly on graph structures;
Relative Behavioral Attributes: Filling the Gap between Symbolic Goal Specification and Reward Learning from Human Preferences
Lin Guan (Arizona State University), Subbarao Kambhampati (Arizona State University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: A Relative Behavior Attribute (RBA) framework is proposed, allowing users to adjust agent behavior through symbolic feedback on attribute strength (such as 'increase step size' or 'decrease softness');
Relative representations enable zero-shot latent space communication
Luca Moschella (Sapienza University of Rome), Emanuele Rodolà (Amazon Web Services)
Representation LearningGraph Neural NetworkTransformerAuto EncoderImageTextGraph
🎯 What it does: This paper proposes a Relative Representation method that maps traditional absolute latent representations into an angle similarity space based on anchor points, enabling zero-shot latent space communication and model stitching across models, datasets, and tasks.
Reliability of CKA as a Similarity Measure in Deep Learning
MohammadReza Davari (Concordia University), Eugene Belilovsky (Concordia University)
ClassificationAnomaly DetectionKnowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The theoretical and experimental analysis of the sensitivity of the similarity metric CKA in deep learning models reveals its impact on subset translation, outliers, and transformations that maintain linear separability, and demonstrates that CKA values can be manipulated through simple transformations while keeping functional behavior unchanged.
REPAIR: REnormalizing Permuted Activations for Interpolation Repair
Keller Jordan (Hive AI), Behnam Neyshabur (Google Research)
Image TranslationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A method called REPAIR is proposed to address the variance collapse problem that occurs during linear interpolation between two aligned networks, significantly reducing performance barriers during the interpolation process.
Reparameterization through Spatial Gradient Scaling
Alexander Detkov (Huawei Kirin Solutions), Di Niu (University of Alberta)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A gradient scale reparameterization method that does not change the network structure is proposed—Spatial Gradient Scaling (SGS), which enhances the model's generalization performance by dynamically adjusting the learning rates at different positions of the convolutional kernel based on the spatial mutual information of the feature maps.
Replay Memory as An Empirical MDP: Combining Conservative Estimation with Experience Replay
Hongming Zhang (University of Alberta), Martin Müller (Institute of Automation Chinese Academy of Sciences)
Reinforcement LearningSequential
🎯 What it does: View experience replay as Experience Replay Memory MDP (RM-MDP), solve its value function through dynamic programming to obtain a conservative estimate ˆQ, and then regularize this estimate with the DQN target value and policy to form the CEER method.
Replicable Bandits
Hossein Esfandiari (Google Research), Grigoris Velegkas (Yale University)
OptimizationReinforcement Learning
🎯 What it does: This paper introduces the concept of replicable strategies and studies the problem of designing replicable strategies in a stochastic multi-armed bandit environment.
Represent to Control Partially Observed Systems: Representation Learning with Provable Sample Efficiency
Lingxiao Wang (Northwestern University), Zhaoran Wang (Northwestern University)
Representation LearningReinforcement Learning
🎯 What it does: A reinforcement learning algorithm named Represent to Control (RTC) is proposed, which can learn and utilize low-dimensional representations to achieve optimal control in partially observable Markov decision processes (POMDPs).
Representation Learning for Low-rank General-sum Markov Games
Chengzhuo Ni (Princeton University), Mengdi Wang (Princeton University)
Representation LearningReinforcement Learning
🎯 What it does: A sample-efficient learning framework GERL_MG2 suitable for general and low-rank Markov games for multi-agent systems is proposed, supporting nonlinear function approximation and achieving scalable deep learning implementations.
Representational Dissimilarity Metric Spaces for Stochastic Neural Networks
Lyndon Duong, Alex H Williams
Representation LearningImage
🎯 What it does: Proposed and implemented stochastic shape metrics for random neural networks, capable of capturing geometric differences in average activation and noise covariance simultaneously;
ResAct: Reinforcing Long-term Engagement in Sequential Recommendation with Residual Actor
Wanqi Xue (Nanyang Technological University), Bo An (Nanyang Technological University)
Recommendation SystemReinforcement LearningAuto EncoderTabularSequential
🎯 What it does: This paper proposes the ResAct method, which first reconstructs online recommendation behavior and then improves it through a residual actor to enhance long-term user engagement in sequential recommendations.
Restricted Strong Convexity of Deep Learning Models with Smooth Activations
Arindam Banerjee (University of Illinois), Misha Belkin
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the convergence of deep learning models with smooth activation functions during the gradient descent optimization process and proposes a new RSC (Restricted Strong Convexity) analysis framework.
Rethinking Graph Lottery Tickets: Graph Sparsity Matters
Bo Hui (Auburn University), Wei-Shinn Ku (Clemson University)
Graph Neural NetworkGraph
🎯 What it does: To address the performance degradation of graph neural networks under high graph sparsity, this paper proposes a sparsification method based on Wasserstein distance auxiliary loss and adversarial min-max optimization, and verifies the transferability of the resulting Graph Lottery Ticket (GLT) in transfer learning.
Rethinking Self-Supervised Visual Representation Learning in Pre-training for 3D Human Pose and Shape Estimation
Hongsuk Choi (Samsung Research America), Kyoung Mu Lee (Seoul National University)
Pose EstimationRepresentation LearningContrastive LearningImage
🎯 What it does: This study investigates pre-training methods for 3D human pose and shape estimation, comparing the effects of self-supervised learning (SSL), 2D annotations, and synthetic data.
Rethinking skip connection model as a learnable Markov chain
Dengsheng Chen (Meituan), Enhua Wu (University of Chinese Academy of Sciences)
ClassificationImage TranslationTransformerImageText
🎯 What it does: View residual networks as learnable Markov chains and propose a 'Penal Connection' mechanism to enable the network to learn more efficient Markov chains, thereby improving training speed and final performance.
Rethinking Symbolic Regression: Morphology and Adaptability in the Context of Evolutionary Algorithms
Kei Sen Fong (National University of Singapore), Mehul Motani (National University of Singapore)
OptimizationTransformerLarge Language ModelTabular
🎯 What it does: A minimal genetic programming symbolic regression algorithm based on Terminal Walking Language Model (TW-MLM) and a variable adaptive fitness function (alternating between MSE and MSEDI) is proposed and implemented, integrated into the traditional GP framework.
Rethinking the Effect of Data Augmentation in Adversarial Contrastive Learning
Rundong Luo (Peking University), Yisen Wang (Peking University)
Representation LearningAdversarial AttackContrastive LearningImage
🎯 What it does: A new dynamic adversarial contrastive learning method, DYNACL, is proposed to address the dilemma of data augmentation intensity in self-supervised adversarial training.
Rethinking the Expressive Power of GNNs via Graph Biconnectivity
Bohang Zhang (Peking University), Di He (Peking University)
Graph Neural NetworkTransformerGraph
🎯 What it does: This paper systematically evaluates the expressive power of existing GNNs from the perspective of graph biconnectivity, proving that most mainstream models cannot identify cut vertices/cut edges, and proposes a distance-based Generalized Distance Weisfeiler–Lehman (GD-WL) framework along with its Transformer implementation Graphormer-GD, which theoretically possesses complete expressive power for all biconnectivity problems.
Retrieval-based Controllable Molecule Generation
Zichao Wang (Rice University), Anima Anandkumar (NVIDIA)
GenerationRetrievalDrug DiscoveryTransformerGraphBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A retrieval-based controllable molecular generation framework, RetMol, is proposed, which guides a pre-trained generative model to generate new molecules that meet target properties using a small number of example molecules that satisfy design conditions through retrieval and information fusion, without the need for task-specific fine-tuning.
Reversible Column Networks
Yuxuan Cai (MEGVII Technology), Xiangyu Zhang (MEGVII Technology)
Object DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a reversible column network (RevCol) - a novel network architecture composed of multiple column sub-networks that achieves feature decoupling at multiple levels through multi-level reversible connections while maintaining information integrity, applicable to CNNs and Transformers.
Revisit Finetuning strategy for Few-Shot Learning to Transfer the Emdeddings
Heng Wang, Yong Li
ClassificationMeta LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A novel fine-tuning strategy LP-FT-FB is proposed, which combines linear probing with Firth bias correction to efficiently and unbiasedly fine-tune pre-trained feature extractors under few samples, thereby enhancing the performance of few-shot learning.
Revisiting adapters with adversarial training
Sylvestre-Alvise Rebuffi (DeepMind), Sven Gowal (DeepMind)
Domain AdaptationAdversarial AttackTransformerImage
🎯 What it does: Achieve co-training of clean images and adversarial samples on Vision Transformer using a minimal amount of domain-specific parameters (such as classification token), and realize a smooth trade-off between clean accuracy and robustness without increasing computational cost through 'adversarial model soup'.
Revisiting Graph Adversarial Attack and Defense From a Data Distribution Perspective
Kuan Li (Chinese Academy of Sciences), Qing He (Chinese Academy of Sciences)
Adversarial AttackGraph Neural NetworkGraph
🎯 What it does: Re-examine graph structure attacks and defenses from the perspective of data distribution, propose a distribution shift formula and theoretical explanation, and provide a series of practical techniques, ultimately improving attack (fast attack) and defense (self-training defense) methods.
Revisiting Intrinsic Reward for Exploration in Procedurally Generated Environments
Kaixin Wang (National University of Singapore), Shuicheng YAN
Reinforcement Learning
🎯 What it does: The intrinsic rewards in procedural generation environments are dissected, systematically evaluating the contributions of lifelong and episodic rewards in exploration, and validating the effects of both through extensive ablation experiments.
Revisiting Populations in multi-agent Communication
Paul Michel (DeepMind), Angeliki Lazaridou (DeepMind)
Recurrent Neural NetworkReinforcement LearningImageTabular
🎯 What it does: Reassessing the population size effect in multi-agent communication, it is found that standard fully connected training leads to co-adaptation between the receiver and multiple senders, resulting in a lack of language uniformity. A partitioned training protocol is proposed, restricting the receiver to co-adapt only with its own sender, and explicitly incorporating the goals of internal communication and mutual intelligibility to promote consistency and composability of group language.
REVISITING PRUNING AT INITIALIZATION THROUGH THE LENS OF RAMANUJAN GRAPH
Duc N.M Hoang (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
Convolutional Neural NetworkImage
🎯 What it does: This paper explores the feasibility of Pruning at Initialization (PaI) during the network initialization phase, analyzing the relationship between the topology and performance of sparse networks using the properties of Ramanujan graphs, and proposes two new metrics, IMDB and NaRC, to evaluate the connectivity and randomness of sparse networks.