ICML 2024 Papers — Page 22
International Conference on Machine Learning · 2610 papers
Run-Time Task Composition with Safety Semantics
Kevin Leahy (Worcester Polytechnic Institute), Zachary Serlin (MIT Lincoln Laboratory)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A Boolean task composition method for safety-aware reinforcement learning is proposed, allowing pre-trained single-task policies to synthesize more complex tasks with zero training at deployment.
RVI-SAC: Average Reward Off-Policy Deep Reinforcement Learning
Yukinari Hisaki (Tokyo Institute of Technology), Isao Ono (Tokyo Institute of Technology)
Reinforcement LearningSequential
🎯 What it does: A method for offline deep reinforcement learning based on the average reward criterion, RVI-SAC, is proposed to address the discrepancy between training objectives and performance metrics caused by the use of discounted reward criteria in existing methods for continuous tasks.
S$\Omega$I: Score-based O-INFORMATION Estimation
Mustapha BOUNOUA, Pietro Michiardi (Eurecom)
Diffusion modelScore-based ModelMultimodality
🎯 What it does: A method for estimating O-information based on score functions, SΩI, is proposed, which can estimate O-information and its gradient in continuous multivariate systems without discrete or Gaussian assumptions.
S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning
Guancheng Wan (Wuhan University), Mang Ye (Wuhan University)
Computational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: By integrating spectral domain filtering, a fast MLP encoder, and spatial positive sample pairs within a contrastive learning framework, the S3GCL method is proposed for self-supervised learning of node representations on graphs with varying levels of homogeneity.
S3O: A Dual-Phase Approach for Reconstructing Dynamic Shape and Skeleton of Articulated Objects from Single Monocular Video
Hao Zhang (University of Illinois Urbana-Champaign), Narendra Ahuja (University of Illinois Urbana-Champaign)
Object DetectionSegmentationPose EstimationOptical FlowVideo
🎯 What it does: This paper proposes a dual-stage method called S3O, which can simultaneously learn visible 3D shapes, variable poses, and underlying skeletal structures using only a single unlabelled monocular video, without the need for preset models, camera poses, or key points.
Safe and Robust Subgame Exploitation in Imperfect Information Games
Zhenxing Ge (Nanjing University), Yang Gao (Nanjing University)
OptimizationSafty and PrivacyReinforcement LearningSequential
🎯 What it does: The concept of 'Adaptation Safety' is proposed, and based on this, an instant subgame search framework OX-Search is designed for safe and efficient opponent exploitation under uncertain opponent models.
Safe Exploration in Dose Finding Clinical Trials with Heterogeneous Participants
Isabel Chien (University of Cambridge), Richard E. Turner (University of Cambridge)
OptimizationDrug DiscoveryBiomedical Data
🎯 What it does: SAFE-T is proposed, an adaptive dose allocation method that balances safety, efficacy, and adapts to patient heterogeneity in early dose-finding trials.
Safe Reinforcement Learning using Finite-Horizon Gradient-based Estimation
Juntao Dai (Zhejiang University), Gang Pan (Zhejiang University)
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: A gradient-based finite time domain constrained estimation method GBE is proposed, and based on this, a safe reinforcement learning algorithm CGPO is designed to handle finite time non-discounted constraints in deep safe RL tasks.
Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models
Yongshuo Zong (University of Edinburgh), Timothy Hospedales (University of Edinburgh)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: A security fine-tuning strategy based on a small secure visual language dataset VLGuard is proposed to enhance the safety of VLLMs without compromising their helpfulness.
Saliency strikes back: How filtering out high frequencies improves white-box explanations
Sabine Muzellec (Brown University), Thomas Serre
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates the sources of high-frequency noise in white-box gradient methods and proposes the FORGrad low-pass filtering technique to enhance their interpretability.
SAM as the Guide: Mastering Pseudo-Label Refinement in Semi-Supervised Referring Expression Segmentation
Danni Yang (Xiamen University), Rongrong Ji (Xiamen University)
Object DetectionSegmentationKnowledge DistillationTransformerImageMultimodality
🎯 What it does: Proposes the SemiRES semi-supervised referring expression segmentation framework, which trains using a small amount of labeled data and a large amount of unlabeled data, and refines pseudo-labels through candidate masks generated by SAM.
SAM-E: Leveraging Visual Foundation Model with Sequence Imitation for Embodied Manipulation
Junjie Zhang (Tsinghua University), Xuelong Li (Shanghai Artificial Intelligence Laboratory)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerPrompt EngineeringImageMultimodality
🎯 What it does: This paper proposes the SAM-E model, which combines the Segment Anything visual foundation model with a multi-view Transformer, and designs a multi-channel heatmap prediction for long sequence actions based on this, achieving efficient and scalable 3D robotic manipulation.
SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention
Romain Ilbert (Huawei Noah's Ark Lab), Ievgen Redko
TransformerTime Series
🎯 What it does: Proposes the SAMformer model to address the underfitting and training instability issues of Transformer in multivariate long-term time series forecasting.
Sample as you Infer: Predictive Coding with Langevin Dynamics
Umais Zahid (Huawei Technologies Research and Development), Zafeirios Fountas (Huawei Technologies Research and Development)
GenerationData SynthesisAuto EncoderImageStochastic Differential Equation
🎯 What it does: This paper proposes the Langevin Predictive Coding (LPC) algorithm for training deep generative models, combining the predictive coding framework, Langevin sampling, warm-starting the inference network, and lightweight preconditioning.
Sample Average Approximation for Conditional Stochastic Optimization with Dependent Data
Yafei Wang (University of Alberta), Linglong Kong (University of Alberta)
OptimizationTabularTime Series
🎯 What it does: This paper studies the sample average approximation (SAA) method for conditional stochastic optimization (CSO) problems under the condition of β-mixing (weak dependence) data, providing a unified bias bound, finite sample guarantees, and asymptotic consistency.
Sample Complexity Bounds for Estimating Probability Divergences under Invariances
Behrooz Tahmasebi (Massachusetts Institute of Technology), Stefanie Jegelka (Technical University of Munich and Massachusetts Institute of Technology)
Tabular
🎯 What it does: The study estimates the sample complexity of various divergence metrics (1-Wasserstein distance, Sobolev IPM, MMD, and L₂, L∞ density estimation) on probability distributions with Lie group invariance, and provides new upper bounds.
Sample-Efficient Multiagent Reinforcement Learning with Reset Replay
Yaodong Yang (Chinese University of Hong Kong), Pheng-Ann Heng
Reinforcement Learning
🎯 What it does: The MARR algorithm is proposed, which enhances sample efficiency in multi-agent reinforcement learning through high replay ratio training in a parallel environment.
Sample-Efficient Robust Multi-Agent Reinforcement Learning in the Face of Environmental Uncertainty
Laixi Shi (California Institute of Technology), Adam Wierman (California Institute of Technology)
OptimizationReinforcement Learning
🎯 What it does: A distributed robust Markov game (RMG) framework is proposed to address environmental uncertainties, and a distributed robust Nash value iteration (DR-NVI) algorithm is designed to achieve approximate learning of robust NE/CE/CCE.
Sample-specific Masks for Visual Reprogramming-based Prompting
Chengyi Cai (University of Melbourne), Feng Liu (University of Melbourne)
ClassificationRecognitionConvolutional Neural NetworkPrompt EngineeringImage
🎯 What it does: By generating a dedicated three-channel mask for each image, the method of visual reprogramming is improved, replacing the traditional shared mask;
Sampling in Unit Time with Kernel Fisher-Rao Flow
Aimee Maurais (Massachusetts Institute of Technology), Youssef Marzouk (Massachusetts Institute of Technology)
Flow-based ModelMultimodality
🎯 What it does: A gradient-independent sampling framework based on Fisher-Rao gradient flow is proposed—Kernel Fisher-Rao Flow (KFRFlow), which can transfer the reference distribution to the target distribution within a unit time.
Sampling is as easy as keeping the consistency: convergence guarantee for Consistency Models
Junlong Lyu (Huawei), Shoubo Feng (Huawei)
OptimizationKnowledge DistillationScore-based ModelOrdinary Differential Equation
🎯 What it does: A convergence guarantee for consistency models has been proposed, proving that both single-step and multi-step sampling can achieve controllable errors in Wasserstein-2 distance and total variation (TV) under minimal assumptions, which only require L2 error, finite second moments of data distribution, and Lipschitz conditions on the score function.
Sampling-based Multi-dimensional Recalibration
Youngseog Chung (Carnegie Mellon University), Jeff Schneider (Carnegie Mellon University)
TabularPhysics Related
🎯 What it does: Research on probability prediction calibration and recalibration methods for multidimensional regression.
SAPG: Split and Aggregate Policy Gradients
Jayesh Singla (Carnegie Mellon University), Deepak Pathak (Carnegie Mellon University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A new on-policy reinforcement learning algorithm SAPG is proposed, which splits large-scale parallel environments into multiple subsets and learns policies separately, then aggregates the data from each subset to the leader policy using importance sampling, achieving higher convergence performance in large batch data scenarios.
Sarah Frank-Wolfe: Methods for Constrained Optimization with Best Rates and Practical Features
Aleksandr Beznosikov (Innopolis University), Gauthier Gidel (Universite de Montreal)
OptimizationTabular
🎯 What it does: Two random finite-sum algorithms based on Frank-Wolfe and SARAH are proposed for finite-sum minimization problems with structured constraints;
SaVeR: Optimal Data Collection Strategy for Safe Policy Evaluation in Tabular MDP
Subhojyoti Mukherjee (University of Wisconsin Madison), Robert D Nowak
Recommendation SystemOptimizationSafty and PrivacyReinforcement LearningTabular
🎯 What it does: A method is proposed to actively collect data to achieve low variance evaluation of target policies while satisfying safety constraints.
Scalable AI Safety via Doubly-Efficient Debate
Jonah Brown-Cohen (Google DeepMind), Georgios Piliouras (Google DeepMind)
Safty and PrivacyAgentic AI
🎯 What it does: A "Doubly Efficient Debate" framework is proposed, designing several interactive proof protocols that allow two polynomial-time AI provers to reliably verify computation tasks dependent on black-box human evaluations with only a constant number of human judgments required.
Scalable and Flexible Causal Discovery with an Efficient Test for Adjacency
Alan Nawzad Amin (New York University), Andrew Gordon Wilson (New York University)
OptimizationExplainability and InterpretabilityGraph Neural NetworkGraphTabular
🎯 What it does: This paper proposes a scalable differentiable adjacency testing method (DAT) and a graph learning framework based on DAT, called DAT-Graph, for efficiently and reliably learning causal graphs from large-scale complex system data.
Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers
Katherine Crowson (Stability AI), Enrico Shippole
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes the Hourglass Diffusion Transformer (HDiT), a transformer architecture that can be directly trained in pixel space and supports high resolutions (e.g., 1024×1024), reducing computational complexity from O(n²) to O(n);
Scalable Multiple Kernel Clustering: Learning Clustering Structure from Expectation
Weixuan Liang (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
Gaussian SplattingTabular
🎯 What it does: This paper proposes a scalable multi-kernel clustering framework SMKC, which is based on the low-rank structure of the expected kernel matrix under the Gaussian distribution assumption, and learns the clustering structure through anchor point sampling and rank-k approximation.
Scalable Online Exploration via Coverability
Philip Amortila (University of Illinois), Akshay Krishnamurthy (Microsoft Research)
Reinforcement LearningSequential
🎯 What it does: A new exploration objective, L1-Coverage, is proposed, which can efficiently generate a set of policies that cover the entire state space in reward-agnostic reinforcement learning, supporting optimal policy learning for any subsequent reward function.
Scalable Pre-training of Large Autoregressive Image Models
Alaaeldin El-Nouby (Apple), Armand Joulin (Google DeepMind)
ClassificationRecognitionGenerationTransformerLarge Language ModelImage
🎯 What it does: Proposed and trained a large-scale autoregressive image model (AIM), transferring the autoregressive pre-training approach of large language models to the visual domain;
Scalable Safe Policy Improvement for Factored Multi-Agent MDPs
Federico Bianchi (University of Verona), Alessandro Farinelli (University of Verona)
OptimizationReinforcement LearningBenchmark
🎯 What it does: A scalable safe policy improvement algorithm for factored multi-agent Markov decision processes (FV-MCTS-SPIBB) is proposed, which generates safe improved policies by using the constrained Max-Plus/Var-El method in MCTS.
Scalable Wasserstein Gradient Flow for Generative Modeling through Unbalanced Optimal Transport
Jaemoo Choi (Seoul National University), Myungjoo Kang (Seoul National University)
GenerationData SynthesisOptimizationComputational EfficiencyFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: A scalable Wasserstein gradient flow generative model based on semi-dual JKO decomposition (S-JKO) is proposed.
Scale-Free Image Keypoints Using Differentiable Persistent Homology
Giovanni Barbarani (Politecnico di Torino), Carlo Masone (Politecnico di Torino)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented a keypoint detector MorseDet based on differentiable persistent homology, proposing an unsupervised loss function centered on persistence and boundary similarity, achieving scale-invariant and learnable keypoint detection.
Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training
Yechan Kim (KAIST), Dongsu Han (KAIST)
Mixture of ExpertsText
🎯 What it does: The ES-MoE method is proposed, allowing the training of large-scale Mixture-of-Experts (MoE) models that exceed GPU memory limits with a limited number of GPUs.
Scaling Down Deep Learning with MNIST-1D
Samuel James Greydanus (Oregon State University), Dmitry Kobak (University of Tübingen)
ClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkContrastive LearningImageTime SeriesBenchmark
🎯 What it does: A low-dimensional, programmatically generated benchmark dataset called MNIST-1D is proposed, which is used to validate various deep learning phenomena in a fast experimental environment.
Scaling Exponents Across Parameterizations and Optimizers
Katie E Everett, Jeffrey Pennington (Google DeepMind)
OptimizationHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: This paper systematically studies the scale indices of different parameterization methods and optimizers during the width expansion of large-scale neural networks, proposes a new perspective on alignment, and provides corresponding theoretical derivations and empirical validations.
Scaling Laws for Fine-Grained Mixture of Experts
Jan Ludziejewski (University of Warsaw), Sebastian Jaszczur (University of Warsaw)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper constructs a fine-grained Mixture of Experts (MoE) model by introducing the granularity hyperparameter and derives an expansion law that includes granularity, model size, and training data volume based on extensive experiments.
Scaling Laws for the Value of Individual Data Points in Machine Learning
Ian Connick Covert, James Zou (Stanford University)
Data-Centric LearningConvolutional Neural NetworkGaussian SplattingImageText
🎯 What it does: Proposed and validated the scale law of individual data points: the marginal contribution of each sample decays logarithmically with the size of the dataset.
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
Patrick Esser (Stability AI), Robin Rombach (Stability AI)
GenerationData SynthesisTransformerDiffusion modelRectified FlowAuto EncoderImageTextMultimodality
🎯 What it does: This paper constructs a model that performs excellently in high-resolution text-to-image synthesis tasks by improving the Rectified Flow (RF) training sampling method and proposing a new multimodal Transformer architecture.
Scaling Tractable Probabilistic Circuits: A Systems Perspective
Anji Liu (University of California), Guy Van den Broeck (University of California)
Computational EfficiencyImageText
🎯 What it does: This paper presents PyJuice, an efficient GPU implementation for probabilistic circuits (PC) that can significantly accelerate both forward and backward computations while reducing memory usage.
Scene Graph Generation Strategy with Co-occurrence Knowledge and Learnable Term Frequency
HyeongJin Kim, Byoung Chul Ko (Keimyung University)
Object DetectionGenerationGraph Neural NetworkImage
🎯 What it does: This paper proposes a method that simultaneously utilizes object co-occurrence knowledge (CooK) and a learnable TF-IDF layer to improve relationship inference in scene graph generation (SGG).
SceneCraft: An LLM Agent for Synthesizing 3D Scenes as Blender Code
Ziniu Hu (California Institute of Technology), Alireza Fathi (Google DeepMind)
GenerationData SynthesisOptimizationAI Code AssistantTransformerLarge Language ModelTextMeshRetrieval-Augmented Generation
🎯 What it does: Developed SceneCraft, an LLM-based Blender code generator that converts natural language descriptions into executable 3D scene scripts.
SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models
Xiaoxuan Wang (University of California), Wei Wang (University of California)
Large Language ModelPrompt EngineeringTextMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: This study constructs the SCIBENCH benchmark to evaluate the capabilities of large language models in solving university-level problems in physics, chemistry, and mathematics.
Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation
Mingyuan Zhou (University of Texas at Austin), Hai Huang (Google)
GenerationKnowledge DistillationDiffusion modelScore-based ModelImage
🎯 What it does: Distill the pre-trained diffusion model into a single-step generator, achieving exponential FID reduction using a data-free distillation method.
Score-Based Causal Discovery of Latent Variable Causal Models
Ignavier Ng (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Score-based Model
🎯 What it does: A score-based causal discovery framework named SALAD is proposed for identifying linear causal models that include latent variables and their causal relationships.
SCoRe: Submodular Combinatorial Representation Learning
Anay Majee (University of Texas at Dallas), Rishabh K Iyer
Object DetectionRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes the SCoRe framework, treating representation learning as a set problem, and designs a combinatorial loss based on submodular information measures to simultaneously suppress inter-class bias and intra-class variance.
Scribble-Supervised Semantic Segmentation with Prototype-based Feature Augmentation
Guiyang Chan (Hohai University), Bainian Chen (Hohai University)
SegmentationTransformerImage
🎯 What it does: Proposes a prototype-based feature enhancement method for weakly supervised semantic segmentation based on sketch lines;
Second-Order Uncertainty Quantification: A Distance-Based Approach
Yusuf Sale (Ludwig Maximilian University of Munich), Eyke Hüllermeier (Ludwig Maximilian University of Munich)
🎯 What it does: A framework for quantifying predictive uncertainty for second-order probability distributions is proposed, providing a complete set of axioms and constructing total, uncertainty, and certainty quantification metrics that satisfy these axioms based on Wasserstein distance.
See More Details: Efficient Image Super-Resolution by Experts Mining
Eduard Zamfir (University of Würzburg), Radu Timofte (University of Würzburg)
RestorationSuper ResolutionConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: An efficient single-image super-resolution model called SeemoRe is proposed, achieving high-quality reconstruction with low computational cost through expert mining.
Seesaw: Compensating for Nonlinear Reduction with Linear Computations for Private Inference
Fabing Li (Xi'an Jiaotong University), Mingyu Gao (Tsinghua University)
Safty and PrivacyComputational EfficiencyNeural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: A neural architecture search method named Seesaw is proposed, specifically designed for privacy-preserving inference, which reduces the computational cost of nonlinear operations by incorporating more linear computations in the network and reusing nonlinear results.
Seizing Serendipity: Exploiting the Value of Past Success in Off-Policy Actor-Critic
Tianying Ji (Tsinghua University), Huazhe Xu (Tsinghua University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Proposes the Blended Exploitation and Exploration (BEE) operator, which integrates historical best actions with the current policy to improve Q-value estimation and enhance Actor-Critic learning;
Selecting Large Language Model to Fine-tune via Rectified Scaling Law
Haowei Lin (Peking University), Yitao Liang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the scalability of fine-tuning large language models (LLMs) and proposes the Rectified Scaling Law to predict fine-tuning performance, designing an efficient model selection algorithm called AtS based on this law.
Selective Mixup Helps with Distribution Shifts, But Not (Only) because of Mixup
Damien Teney (Idiap Research Institute), Ehsan Abbasnejad (University of Adelaide)
Domain AdaptationImage
🎯 What it does: This paper studies the Selective Mixup method in the context of distribution shift scenarios, revealing that its performance improvement primarily stems from implicit resampling rather than the mixing operation itself, and based on this, designs a new combination of resampling and mixing;
Self-Alignment of Large Language Models via Monopolylogue-based Social Scene Simulation
Xianghe Pang (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A social scene simulator named MATRIX has been designed and implemented, utilizing a single LLM to play multiple roles in the same dialogue, simulating social interactions and consequences resulting from user instructions; subsequently, self-criticism is generated based on these consequences to correct responses, achieving self-alignment of the LLM, and maintaining high inference speed through supervised fine-tuning (SFT) on the simulated data.
Self-attention Networks Localize When QK-eigenspectrum Concentrates
Han Bao (Kyoto University), Ryo Karakida (AIST)
TransformerText
🎯 What it does: This paper studies the relationship between attention localization in self-attention mechanisms and the concentration of eigenvalue spectra of the query-key matrix, revealing that localization can be achieved by minimizing spectral variance, which alleviates two failure modes: rank collapse and entropy collapse.
Self-Attention through Kernel-Eigen Pair Sparse Variational Gaussian Processes
Yingyi Chen (KU Leuven), Johan Suykens
ClassificationOptimizationComputational EfficiencyTransformerImageText
🎯 What it does: By constructing a dual-branch Kernel-Eigen Pair Sparse Variational Gaussian Process (KEP-SVGP) model, Bayesian uncertainty estimation of Transformer self-attention is achieved.
Self-cognitive Denoising in the Presence of Multiple Noisy Label Sources
Yi-Xuan Sun (Ant Group), JUN ZHOU
ClassificationKnowledge DistillationText
🎯 What it does: A multi-source noise label learning framework SDM based on the self-awareness capability of neural networks is proposed, which can simultaneously identify sample-level noise and annotator quality during the training process.
Self-Composing Policies for Scalable Continual Reinforcement Learning
Mikel Malagon (University of the Basque Country), Jose A. Lozano (Basque Center for Applied Mathematics)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: A scalable modular neural network architecture called CompoNet is proposed, which self-combines previously learned policy modules to address the issues of catastrophic forgetting and interference in continual reinforcement learning.
Self-Consistency Training for Density-Functional-Theory Hamiltonian Prediction
He Zhang (Xi'an Jiaotong University), Tie-Yan Liu (Microsoft Research AI for Science)
Graph Neural NetworkTabularPhysics Related
🎯 What it does: A label-free training method based on the DFT self-consistency principle is proposed for directly predicting molecular Hamiltonian matrices.
Self-Correcting Self-Consuming Loops for Generative Model Training
Nate Gillman (Brown University), Chen Sun (Google DeepMind)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: This paper proposes and validates a self-correcting self-consuming loop mechanism in the training of generative models. By using data generated by the previous model in each training generation and performing self-correction, followed by fine-tuning with a mix of real data, the aim is to stabilize model training and prevent collapse.
Self-Driven Entropy Aggregation for Byzantine-Robust Heterogeneous Federated Learning
Wenke Huang (Wuhan University), Bo Du (Wuhan University)
Federated LearningImage
🎯 What it does: A self-driven entropy aggregation method (SDEA) based on random public data is proposed to resist Byzantine attacks in heterogeneous federated learning.
Self-Infilling Code Generation
Lin Zheng (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
GenerationAI Code AssistantLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a self-infilling-based code generation framework that integrates infilling capabilities into autoregressive decoding, achieving a non-monotonic code generation process.
Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models
Zixiang Chen (University of California), Quanquan Gu (University of California)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes a Self-Play Fine-tuning (SPIN) method that iteratively improves model performance using synthetic data generated by a weak LLM without the need for additional human annotations or external rewards.
Self-Rewarding Language Models
Weizhe Yuan (Meta), Jason E Weston
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes Self-Rewarding Language Models, allowing the model to generate and evaluate instruction-following data on its own, utilizing LLM-as-a-Judge for self-reinforcement.
Self-Supervised Coarsening of Unstructured Grid with Automatic Differentiation
Sergei Shumilin (Skolkovo Institute of Science and Technology), Vladimir Vanovskiy (Skolkovo Institute of Science and Technology)
OptimizationComputational EfficiencyGraph Neural NetworkPoint CloudPhysics Related
🎯 What it does: A self-supervised coarsening algorithm based on differentiable physics for unstructured grids is proposed, utilizing k-means, automatic differentiation, and stochastic minimization to reduce the number of grid points while maintaining numerical simulation accuracy.
Self-Supervised Interpretable End-to-End Learning via Latent Functional Modularity
Hyunki Seong (Korea Advanced Institute of Science and Technology), Hyunchul Shim
Autonomous DrivingExplainability and InterpretabilityRobotic IntelligenceConvolutional Neural NetworkTransformerContrastive LearningMultimodality
🎯 What it does: An end-to-end network called MoNet based on functional modularization is proposed, which can learn autonomous navigation decisions under self-supervised conditions without task labels.
SelfIE: Self-Interpretation of Large Language Model Embeddings
Haozhe Chen (Columbia University), Chengzhi Mao (Mila)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The SelfIE framework is proposed, utilizing the decoding capability of LLM itself to achieve open, untrained explanations and controllability of LLM hidden representations by inserting hidden embeddings and guiding the generation of natural language explanations through prompts.
SelfVC: Voice Conversion With Iterative Refinement using Self Transformations
Paarth Neekhara (NVIDIA), Julian McAuley (University of California San Diego)
GenerationData SynthesisTransformerGenerative Adversarial NetworkAudio
🎯 What it does: This paper proposes SelfVC, a text-independent voice conversion framework that iteratively trains through self-generated acoustic transformations, enabling zero-shot, cross-lingual voice conversion and controllable prosody synthesis.
SelMatch: Effectively Scaling Up Dataset Distillation via Selection-Based Initialization and Partial Updates by Trajectory Matching
Yongmin Lee (Korea Advanced Institute of Science and Technology), Hye Won Chung (Korea Advanced Institute of Science and Technology)
Knowledge DistillationImage
🎯 What it does: A dataset distillation method named SelMatch is proposed, which effectively expands the distillation scale through selective initialization and partial updates.
Semantic-Aware Human Object Interaction Image Generation
Zhu Xu (Peking University), Yang Liu (Peking University)
GenerationData SynthesisPose EstimationDiffusion modelImageTextBenchmark
🎯 What it does: By guiding the diffusion model with pose quality and interaction boundary information, the SA-HOI framework is proposed, which can generate higher quality person-object interaction images from pure text prompts and achieve multi-step refinement based on iterative inversion.
Semantically-correlated memories in a dense associative model
Thomas F Burns (Brown University)
Graph Neural NetworkVideoGraph
🎯 What it does: The Correlated Dense Associative Memory (CDAM) model is proposed, which unifies automatic association and disassociation into a dense associative memory network, and achieves semantic association through a graph structure. The four dynamic patterns are then analyzed theoretically and numerically, demonstrating applications in tasks such as community detection, sequence recall, and finite automaton simulation.
SeMOPO: Learning High-quality Model and Policy from Low-quality Offline Visual Datasets
Shenghua Wan (Nanjing University), De-Chuan Zhan (Nanjing University)
Reinforcement LearningImageBenchmark
🎯 What it does: This paper proposes an offline visual reinforcement learning method called SeMOPO, which can learn high-quality policies from datasets collected by low-quality policies that contain moving background noise.
Sequence Compression Speeds Up Credit Assignment in Reinforcement Learning
Aditya Ramesh, Jürgen Schmidhuber (Swiss AI Lab IDSIA)
OptimizationReinforcement LearningSequentialFinance Related
🎯 What it does: The Chunked-TD algorithm is proposed, which utilizes the learned forward model to predict the probability of the next state, dynamically determines the λ-return, and accelerates credit assignment in reinforcement learning through sequence compression.
Sequential Asynchronous Action Coordination in Multi-Agent Systems: A Stackelberg Decision Transformer Approach
Bin Zhang (Institute of Automation Chinese Academy of Sciences), Guoliang Fan (Institute of Automation Chinese Academy of Sciences)
OptimizationKnowledge DistillationTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes an asynchronous action coordination method for multi-agent systems based on Stackelberg games—Stackelberg Decision Transformer (STEER), which achieves the learning of Stackelberg equilibrium strategies.
Sequential Disentanglement by Extracting Static Information From A Single Sequence Element
Nimrod Berman (Ben Gurion University of the Negev), Omri Azencot (Ben Gurion University of the Negev)
GenerationData SynthesisAnomaly DetectionRecurrent Neural NetworkAuto EncoderVideoMultimodalityTime SeriesSequentialAudio
🎯 What it does: A variational autoencoder is designed to extract static information from a single frame and subtract static content in a dynamic path, achieving unsupervised decoupling of sequential data.
Sequential Kernel Goodness-of-fit Testing
Zhengyu Zhou (Wuhan University), Weiwei Liu (Wuhan University)
Time SeriesSequential
🎯 What it does: This paper proposes a Sequential Kernel Goodness-of-Fit Test (SKGT) based on betting theory, which can monitor in real-time and decide whether to reject the null hypothesis in streaming data, avoiding the uncontrolled false positive rate of traditional batch tests under continuous monitoring.
Sequential Neural Score Estimation: Likelihood-Free Inference with Conditional Score Based Diffusion Models
Louis Sharrock (Lancaster University), Mark Beaumont (University of Bristol)
GenerationData SynthesisOptimizationDiffusion modelScore-based ModelFlow-based ModelTabularTime SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes a sequence neural posterior score estimation (SNPSE) method based on a conditional score diffusion model for Bayesian inference in simulation models, and presents its non-sequential version (NPSE).
SF-DQN: Provable Knowledge Transfer using Successor Feature for Deep Reinforcement Learning
Shuai Zhang (New Jersey Institute of Technology), Meng Wang (Rensselaer Polytechnic Institute)
Reinforcement LearningSequential
🎯 What it does: This paper studies a reinforcement learning algorithm called SF-DQN, which is based on Successor Features (SF) and Generalization Policy Improvement (GPI). It proves its convergence and generalization performance and achieves knowledge transfer between tasks.
SFC: Achieve Accurate Fast Convolution under Low-precision Arithmetic
Liulu He (Nanjing University), Li Du (Nanjing University)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A fast Fourier convolution algorithm based on symbolic computation (SFC) is proposed, which enhances the efficiency of linear convolution by incorporating correction terms.
Sharp Rates in Dependent Learning Theory: Avoiding Sample Size Deflation for the Square Loss
Ingvar Ziemann (University of Pennsylvania), Nikolai Matni (University of Pennsylvania)
Tabular
🎯 What it does: This study explores statistical learning under dependent (β-mixing) data and squared loss, aiming to find an accurate proxy for noise interaction terms in learning. The research shows that when the topological structure of the hypothesis class F is comparable to L2 and Ψp, the convergence rate of the empirical risk minimizer depends solely on the complexity of the class and second-order statistics.
Sharpness-Aware Data Generation for Zero-shot Quantization
Hoang Anh Dung (Monash University), Thanh-Toan Do (Monash University)
Data SynthesisOptimizationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A data generation method called SADAG is proposed for zero-shot quantization, which combines model sharpness through gradient matching to calibrate the quantization model and reduce sharpness, thereby enhancing generalization.
Shifted Interpolation for Differential Privacy
Jinho Bok (University of Pennsylvania), Jason Altschuler (University of Pennsylvania)
OptimizationSafty and PrivacyImage
🎯 What it does: This paper proposes a new analytical method that quantifies the privacy leakage of differentially private machine learning algorithms by establishing and improving the 'iterative privacy enhancement' phenomenon, particularly in the context of strong convex optimization.
SHINE: Shielding Backdoors in Deep Reinforcement Learning
Zhuowen Yuan (University of Illinois Urbana-Champaign), Dawn Song (University of California Berkeley)
Explainability and InterpretabilityReinforcement Learning
🎯 What it does: Proposes the SHINE method, which uses interpretability techniques to identify backdoor triggers in DRL agents and eliminate the impact of these triggers through retraining, achieving backdoor masking.
Short-Long Convolutions Help Hardware-Efficient Linear Attention to Focus on Long Sequences
Zicheng Liu (Westlake University), Stan Z. Li (Westlake University)
Convolutional Neural NetworkRecurrent Neural NetworkTextSequentialAudio
🎯 What it does: The CHELA model is proposed, which integrates linear attention and long convolution through short-long convolution to achieve hardware-efficient linear attention on long sequences.
Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs
Andries Petrus Smit (InstaDeep), Arnu Pretorius (InstaDeep)
OptimizationHyperparameter SearchTransformerLarge Language ModelPrompt EngineeringTextBiomedical Data
🎯 What it does: Evaluate and compare the effectiveness of Multi-Agent Debate (MAD) with other prompting strategies in multi-domain question answering, exploring its impact on cost, time, and accuracy; and propose enhancing MAD performance by adjusting agent agreement intensity; provide open-source implementation and experimental scripts.
SiBBlInGS: Similarity-driven Building-Block Inference using Graphs across States
Noga Mudrik (Johns Hopkins University), Adam Shabti Charles
Anomaly DetectionOptimizationGraph Neural NetworkTime Series
🎯 What it does: A graph-driven dictionary learning framework named SiBBlInGS is proposed to infer interpretable building blocks (BB) and their temporal trajectories in high-dimensional time series with multiple states and variable durations.
Sign Gradient Descent-based Neuronal Dynamics: ANN-to-SNN Conversion Beyond ReLU Network
Hyunseok Oh (Seoul National University), Youngki Lee (Seoul National University)
OptimizationConvolutional Neural NetworkSpiking Neural NetworkReinforcement LearningImage
🎯 What it does: A novel neuron dynamics model based on sign gradient descent is proposed, which is used for the conversion from ANN to SNN, enabling SNN to approximate various nonlinear activation functions other than ReLU.
Sign is Not a Remedy: Multiset-to-Multiset Message Passing for Learning on Heterophilic Graphs
Langzhang Liang (Harbin Institute of Technology), Yuan Qi (Fudan University)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper addresses the Signed Message Passing (SMP) method used on heterophilic graphs, revealing its two major flaws from both theoretical and experimental perspectives: (1) even if the first-order propagation weights are 'desirable', the multi-hop propagation matrix may become undesirable, leading to incorrect updates of node embeddings; (2) even with negative weights, SMP is still prone to the oversmoothing problem in multi-class graphs. To address these issues, the authors propose a new Multiset-to-Multiset (m-2-m) message passing framework and design the M2M-GNN model based on it. M2M-GNN partitions neighbor embeddings into several 'chunks' using attention soft labels, aggregates and concatenates each chunk separately, thereby maintaining the separation of different class embeddings and enhancing expressive power. Extensive experiments on 11 benchmark datasets demonstrate that M2M-GNN ranks first overall in node classification tasks, achieving at least a top-three performance on each dataset and maintaining robustness in deep models.
Sign Rank Limitations for Inner Product Graph Decoders
Su Hyeong Lee (University of Chicago), Risi Kondor (University of Chicago)
OptimizationRepresentation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: This paper theoretically analyzes and experimentally demonstrates that the inner product decoder has limitations in representational capacity for graph reconstruction tasks, and proposes improvements such as the addition of a cutoff and complex embeddings to enhance expressiveness.
SignSGD with Federated Defense: Harnessing Adversarial Attacks through Gradient Sign Decoding
Chanho Park (POSTECH), Namyoon Lee (Korea University)
OptimizationFederated LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A new distributed learning algorithm signSGD-FD is proposed, which utilizes weighted majority voting combined with a federated defense mechanism, maintaining convergence speed unaffected in the face of malicious attacks.
SILVER: Single-loop variance reduction and application to federated learning
Kazusato Oko (University of Tokyo), Taiji Suzuki (University of Tokyo)
OptimizationFederated LearningTabular
🎯 What it does: A single-cycle variance reduction algorithm SILVER is proposed, which can solve non-convex finite-sum optimization problems without the need for periodic full gradient calculations and provides first-order, second-order optimality, and exponential convergence under the PL condition.
Simple Ingredients for Offline Reinforcement Learning
Edoardo Cetin (Sakana AI), Ahmed Touati (Meta)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper addresses the performance degradation caused by the mixing of multi-source data in offline reinforcement learning, proposing the MOOD testing platform and systematically evaluating various algorithms and design assumptions.
Simple linear attention language models balance the recall-throughput tradeoff
Simran Arora (Stanford University), Christopher Re (Stanford University)
GenerationRetrievalTransformerLarge Language ModelText
🎯 What it does: A hybrid language model architecture named BASED is proposed, which combines global linear attention (Taylor approximation) with local sliding window attention to achieve a trade-off between high recall rate and high throughput.
Simplicity Bias of Two-Layer Networks beyond Linearly Separable Data
Nikita Tsoy (INSAIT), Nikola Konstantinov (INSAIT)
Image
🎯 What it does: This paper studies the simplicity bias of two-layer neural networks on non-linearly separable datasets, proving that during the early training phase, the network features converge in a few directions, and this bias intensifies in the later training phase.
Simplicity Bias via Global Convergence of Sharpness Minimization
Khashayar Gatmiry (Massachusetts Institute of Technology), Stefanie Jegelka (Massachusetts Institute of Technology)
OptimizationTabular
🎯 What it does: This study investigates the simplicity bias achieved by sharpness minimization in two-layer networks using label noise SGD, proving that it can converge to a low-rank feature matrix under high-dimensional data.
SimPro: A Simple Probabilistic Framework Towards Realistic Long-Tailed Semi-Supervised Learning
Chaoqun Du (Tsinghua University), Gao Huang (Tsinghua University)
ClassificationOptimizationImage
🎯 What it does: Proposes the SimPro framework for training in real long-tail semi-supervised learning scenarios where the distribution of labeled data is imbalanced and the distribution of unlabeled data is unknown.
Simulation of Graph Algorithms with Looped Transformers
Artur Back de Luca (University of Waterloo), Kimon Fountoulakis (University of Waterloo)
TransformerGraph
🎯 What it does: A cyclic Transformer model is constructed, and it is proven that it can simulate various graph algorithms (such as Dijkstra, BFS, DFS, Kosaraju, etc.) and multi-task combinations with constant width.
Simulation-Based Inference with Quantile Regression
He Jia (Princeton University)
Recurrent Neural NetworkTime SeriesBenchmark
🎯 What it does: This paper proposes Neural Quantile Estimation (NQE), a method that utilizes conditional quantile regression for autoregressive learning of univariate quantiles for each posterior dimension, and obtains high-quality posterior samples through quantile interpolation and post-processing calibration.
Simultaneous identification of models and parameters of scientific simulators
Cornelius Schröder (University of Tübingen), Jakob H. Macke (Max Planck Institute for Intelligent Systems)
Time SeriesSequentialPhysics Related
🎯 What it does: This paper proposes a simulation-based Bayesian inference method (SBMI) that can simultaneously identify the model structure and parameters of scientific simulators, learning the complete posterior distribution directly from observational data.