NeurIPS 2023 Papers — Page 26
Conference on Neural Information Processing Systems · 3218 papers
RL-based Stateful Neural Adaptive Sampling and Denoising for Real-Time Path Tracing
Antoine Scardigli (Huawei Zurich Research Center), Lorenz K Muller
OptimizationComputational EfficiencyReinforcement LearningImage
🎯 What it does: An adaptive sampling and state space encoding framework based on reinforcement learning is proposed, capable of achieving real-time path tracking at low sampling rates.
RoboCLIP: One Demonstration is Enough to Learn Robot Policies
Sumedh Anand Sontakke, Laurent Itti (University of Southern California)
Robotic IntelligenceReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Using a pre-trained video-language model (such as S3D), a robot policy is trained in online reinforcement learning with only a single video or text description as a demonstration, generating rewards through similarity.
Robust and Actively Secure Serverless Collaborative Learning
Nicholas Franzese (Northwestern University), Xiao Wang (University of Toronto)
Federated LearningSafty and PrivacyImage
🎯 What it does: A decentralized peer-to-peer learning protocol is proposed, which can securely aggregate model updates and train a shared model under a malicious threat model where neither the server nor the client is trusted.
Robust Bayesian Satisficing
Artun Saday (Bilkent University), Cem Tekin (Bilkent University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: A Robust Bayesian Satisficing (RBS) framework and RoBOS algorithm are proposed, achieving robust satisfaction against unknown distribution shifts using a threshold τ;
Robust Concept Erasure via Kernelized Rate-Distortion Maximization
Somnath Basu Roy Chowdhury (University of North Carolina at Chapel Hill), Snigdha Chaturvedi (University of North Carolina at Chapel Hill)
Representation LearningData-Centric LearningTextTabular
🎯 What it does: A new target KRaM based on distance metric learning is proposed to remove specific concepts from distributed representations while preserving as much other information as possible.
Robust Contrastive Language-Image Pretraining against Data Poisoning and Backdoor Attacks
Wenhan Yang (University of California Los Angeles), Baharan Mirzasoleiman (University of California Los Angeles)
Representation LearningAdversarial AttackData-Centric LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: In response to the vulnerability of multimodal vision-language models to target data poisoning and backdoor attacks during the pre-training phase, the ROCLIP method is proposed to achieve robust pre-training;
Robust covariance estimation with missing values and cell-wise contamination
gregoire pacreau, Karim Lounici (Ecole Polytechnique)
TabularTime SeriesFinance Related
🎯 What it does: The study designs and analyzes an unbiased covariance estimation method in the presence of missing values and cell-level contamination, combining it with various detection filtering and imputation strategies to achieve robust covariance estimation in high dimensions.
Robust Data Pruning under Label Noise via Maximizing Re-labeling Accuracy
Dongmin Park (KAIST), Jae-Gil Lee (KAIST)
OptimizationData-Centric LearningImage
🎯 What it does: This paper proposes a robust data pruning method called Prune4ReL in the context of label noise, aiming to enhance the re-labeling accuracy and final generalization performance of the Re-labeling model by maximizing the total neighborhood confidence of subsets for all training samples.
Robust Data Valuation with Weighted Banzhaf Values
Weida Li, Yaoliang Yu (University of Waterloo)
Data-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningTabular
🎯 What it does: A robust method for assessing the importance of training data is proposed—the weighted Banzhaf value, along with its theoretical and empirical analysis.
Robust Distributed Learning: Tight Error Bounds and Breakdown Point under Data Heterogeneity
Youssef Allouah (Ecole Polytechnique Federale de Lausanne), Geovani Rizk (Ecole Polytechnique Federale de Lausanne)
OptimizationFederated LearningTabular
🎯 What it does: A theoretical analysis of robust distributed learning under data heterogeneity is presented, along with lower and upper bounds.
Robust Knowledge Transfer in Tiered Reinforcement Learning
Jiawei Huang (ETH Zurich), Niao He (ETH Zurich)
Reinforcement LearningTabular
🎯 What it does: A Tiered RL framework is proposed that can learn in parallel and achieve robust knowledge transfer when the source task and target task are not completely identical.
Robust Learning for Smoothed Online Convex Optimization with Feedback Delay
Pengfei Li (University of California Riverside), Shaolei Ren (University of California Riverside)
OptimizationTabular
🎯 What it does: A robust learning algorithm named RCL is proposed, which combines machine learning prediction with expert online algorithms, ensuring a (1+λ)-competitive ratio in the Smoothed Online Convex Optimization (SOCO) environment with multi-step nonlinear switching costs and feedback delays; during the training phase, a robust perception strategy is employed to enhance average performance.
Robust Learning with Progressive Data Expansion Against Spurious Correlation
Yihe Deng (University of California, Los Angeles), Quanquan Gu (University of California, Los Angeles)
ClassificationOptimizationData-Centric LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: The Progressive Data Expansion (PDE) algorithm is proposed, which studies the learning process of deep learning models in the presence of spurious correlated features and enhances the model's robustness to spurious correlations through a two-stage training approach.
Robust Lipschitz Bandits to Adversarial Corruptions
Yue Kang (University of California), Thomas Chun Man Lee
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: The study proposes a robust learning algorithm for the Lipschitz bandit problem with reward corruption in the presence of adaptive opponents under a continuous action set;
Robust low-rank training via approximate orthonormal constraints
Dayana Savostianova (Gran Sasso Science Institute), Francesco Tudisco (Gran Sasso Science Institute)
CompressionOptimizationImage
🎯 What it does: A robust low-rank training algorithm is proposed that maintains near-orthogonal constraints through singular value constraints during low-rank network training.
Robust Matrix Sensing in the Semi-Random Model
Xing Gao (University of Illinois at Chicago), Yu Cheng (Brown University)
Optimization
🎯 What it does: This paper studies the matrix sensing problem under a semi-random model, proposes the weighted RIP (wRIP) assumption, and designs an iterative reweighting gradient descent algorithm (including a weight oracle and error reduction subroutine) to recover low-rank matrices in this model.
Robust Mean Estimation Without Moments for Symmetric Distributions
Gleb Novikov (ETH Zurich), Stefan Tiegel (ETH Zurich)
🎯 What it does: An algorithm for robust mean estimation for symmetric product distributions and elliptical distributions is proposed without the need for any moments of the distribution.
Robust Model Reasoning and Fitting via Dual Sparsity Pursuit
Xingyu Jiang (Huazhong University of Science and Technology), Jiayi Ma (Wuhan University)
Anomaly DetectionOptimizationImage
🎯 What it does: A unified Dual Sparse Tracking (DSP) method is proposed for robust inference and fitting of geometric models in the presence of a large number of outliers and unknown model types.
Robust Multi-Agent Reinforcement Learning via Adversarial Regularization: Theoretical Foundation and Stable Algorithms
Alexander Bukharin (Georgia Institute of Technology), Tuo Zhao (Ford Motor Company)
Autonomous DrivingOptimizationReinforcement LearningGenerative Adversarial Network
🎯 What it does: The ERNIE framework is proposed, utilizing adversarial regularization to control the Lipschitz constant of the policy, thereby enhancing robustness in multi-agent reinforcement learning, with its effectiveness proven both theoretically and experimentally.
Robust Second-Order Nonconvex Optimization and Its Application to Low Rank Matrix Sensing
Shuyao Li (University of Wisconsin Madison), Stephen Wright (University of Wisconsin Madison)
OptimizationTabular
🎯 What it does: This paper studies the problem of finding approximate second-order stationary points (SOSP) under strong pollution models and proposes a general framework that can efficiently find approximate SOSP with dimension-independent accuracy guarantees, applying it to low-rank matrix sensing problems.
Robustifying Generalizable Implicit Shape Networks with a Tunable Non-Parametric Model
Amine Ouasfi (Inria), Adnane Boukhayma (Inria)
SegmentationOptimizationConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud
🎯 What it does: This paper proposes a point cloud reconstruction framework that combines a pre-trained deep occupancy network with adjustable Nystrom kernel ridge regression (NKRR), optimizing the shape function adaptively during testing.
Robustness Guarantees for Adversarially Trained Neural Networks
Poorya Mianjy (Johns Hopkins University), Raman Arora (Johns Hopkins University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new PGD attack method by performing 'symmetrical reflection about the origin' on the loss function within the internal loop of adversarial training, and conducts a theoretical analysis of its convergence; based on this, it provides a global convergence and robust generalization error upper bound for the overall adversarial training of a two-layer Leaky ReLU network, applicable to any width and initialization, and only requires the assumption of a linearly separable margin.
Rotating Features for Object Discovery
Sindy Löwe (University of Amsterdam), Max Welling (University of Amsterdam)
Object DetectionRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderImage
🎯 What it does: A continuous distributed object-centric representation method called Rotating Features is proposed, which extends complex features into high-dimensional rotational features and can unsupervisedly and automatically separate objects from pre-trained vision transformer features, addressing the binding problem.
RRHF: Rank Responses to Align Language Models with Human Feedback
Hongyi Yuan (Tsinghua University), Fei Huang (Alibaba DAMO Academy)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A new learning paradigm called RRHF is proposed, which aligns the output of language models with human preferences by learning to rank responses generated from multiple sources.
RS-Del: Edit Distance Robustness Certificates for Sequence Classifiers via Randomized Deletion
Zhuoqun Huang (University of Melbourne), Benjamin I. P. Rubinstein (University of Melbourne)
Anomaly DetectionSequential
🎯 What it does: A random deletion (RS-Del) random smoothing method for discrete sequence classifiers is proposed, with a proof of robustness in terms of edit distance.
Rubik's Cube: High-Order Channel Interactions with a Hierarchical Receptive Field
Naishan Zheng (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A high-order channel interaction convolution operator, Rubik's Cube Convolution, is proposed as a lightweight alternative to standard convolution in image restoration tasks.
S-CLIP: Semi-supervised Vision-Language Learning using Few Specialist Captions
Sangwoo Mo (Korean Advanced Institute of Science and Technology), Jinwoo Shin (Korean Advanced Institute of Science and Technology)
ClassificationRetrievalTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: Proposes S-CLIP, a semi-supervised pre-training method for CLIP that utilizes unpaired images;
SA-Solver: Stochastic Adams Solver for Fast Sampling of Diffusion Models
Shuchen Xue (University of Chinese Academy of Sciences), Zhi-Ming Ma (Huawei Noah's Ark Lab)
GenerationData SynthesisComputational EfficiencyDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper presents SA-Solver, a diffusion SDE solver based on a stochastic Adams multi-step method, achieving efficient sampling.
Saddle-to-Saddle Dynamics in Diagonal Linear Networks
Scott Pesme (Ecole Polytechnique Fédérale de Lausanne), Nicolas Flammarion (Ecole Polytechnique Fédérale de Lausanne)
OptimizationTabular
🎯 What it does: The study uses gradient flow in a 2-layer diagonal linear network and derives its complete trajectory in the limit where initialization approaches zero, proving that the trajectory exhibits saddle-to-saddle behavior and ultimately converges to the minimum ℓ1 norm solution.
Safe Exploration in Reinforcement Learning: A Generalized Formulation and Algorithms
Akifumi Wachi (LINE Corporation), Kazumune Hashimoto (Osaka University)
Safty and PrivacyReinforcement LearningSequential
🎯 What it does: This paper proposes the Generalized Safety Exploration (GSE) problem and provides a unified formalization, followed by the design of a safety-constrained meta-algorithm MASE, along with two implementable variants based on GLM and GP;
SafeDICE: Offline Safe Imitation Learning with Non-Preferred Demonstrations
Youngsoo Jang (LG AI Research), Moontae Lee
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: The SafeDICE algorithm is proposed to address the offline safe imitation learning problem, utilizing labeled non-preferred (violating) demonstrations for distribution correction on the stationary distribution in the state-action space, resulting in a policy that satisfies safety constraints while maintaining high returns.
Safety Verification of Decision-Tree Policies in Continuous Time
Christian Schilling (Aalborg University), Kim Guldstrand Larsen (Aalborg University)
Safty and PrivacyReinforcement LearningSequential
🎯 What it does: This paper studies the safety of decision tree control in continuous and discrete time systems, proposing the first algorithm that can directly verify decision tree policies in continuous time, along with corresponding implementable details.
SALSA VERDE: a machine learning attack on LWE with sparse small secrets
Cathy Yuanchen Li (Meta), Kristin E. Lauter
OptimizationAdversarial AttackTransformerTabular
🎯 What it does: A machine learning-based LWE attack method called VERDE has been developed, capable of recovering sparse binary, ternary, and narrow Gaussian secrets at larger dimensions and smaller moduli.
SAME: Uncovering GNN Black Box with Structure-aware Shapley-based Multipiece Explanations
Ziyuan Ye (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: An interpretable method for graph neural networks called SAME is proposed, which can generate multiple structured explanations.
SAMoSSA: Multivariate Singular Spectrum Analysis with Stochastic Autoregressive Noise
Abdullah Omar Alomar, Devavrat Shah (Massachusetts Institute of Technology)
Time Series
🎯 What it does: A two-stage algorithm SAMoSSA is proposed for simultaneously estimating time-varying deterministic components and AR correlated noise in multivariate time series data.
Sample based Explanations via Generalized Representers
Che-Ping Tsai (Carnegie Mellon University), Pradeep Kumar Ravikumar
Explainability and InterpretabilityImageText
🎯 What it does: A general sample-based explanation framework called Generalized Representers is proposed, which quantifies the influence of training samples on model predictions.
Sample Complexity Bounds for Score-Matching: Causal Discovery and Generative Modeling
Zhenyu Zhu (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationScore-based ModelTabular
🎯 What it does: This paper provides statistical sample complexity bounds for score matching and its application in causal discovery. By using stochastic gradient descent to train standard deep ReLU neural networks, it demonstrates that accurate estimation of the score function is achievable and establishes error rate bounds for recovering causal relationships using the score matching method.
Sample Complexity for Quadratic Bandits: Hessian Dependent Bounds and Optimal Algorithms
Qian Yu (Princeton University), Jason D. Lee (Princeton University)
OptimizationTabular
🎯 What it does: This paper studies the quadratic bandit problem in stochastic zeroth-order optimization, providing the first tight characterization of sample complexity dependent on the Hessian, and proposes the corresponding optimal algorithm.
Sample Complexity of Forecast Aggregation
Tao Lin (Harvard University), Yiling Chen (Harvard University)
🎯 What it does: This paper studies the Bayesian predictive aggregation framework, exploring how many samples are needed to approximate the optimal aggregator given that the distribution of experts' private signals is unknown but samples are available.
Sample Complexity of Goal-Conditioned Hierarchical Reinforcement Learning
Arnaud Robert (Imperial College London), Aldo A. Faisal
Reinforcement LearningTabular
🎯 What it does: This paper studies the sample complexity lower bound of goal-conditioned hierarchical reinforcement learning and designs a hierarchical Q-learning algorithm named Stationary Hierarchical Q-Learning (SHQL) based on this lower bound.
Sample Efficient Reinforcement Learning in Mixed Systems through Augmented Samples and Its Applications to Queueing Networks
Honghao Wei (Washington State University), Lei Ying (University of Michigan)
OptimizationReinforcement LearningTabular
🎯 What it does: A sample-efficient reinforcement learning framework for hybrid systems (including stochastic and pseudo-stochastic states) is proposed, which compensates for the scarcity of samples in pseudo-stochastic states by generating virtual samples (Augmented Sample Generator), thereby significantly reducing the demand for real samples.
Sample-Conditioned Hypothesis Stability Sharpens Information-Theoretic Generalization Bounds
Ziqiao Wang (University of Ottawa), Yongyi Mao (University of Ottawa)
🎯 What it does: This paper proposes sample condition hypothesis stability (SCH stability) and a sample-related hypothesis matrix, using them to construct tighter upper bounds on input-output mutual information (IOMI) and conditional mutual information (CMI) generalization errors.
Sample-Efficient and Safe Deep Reinforcement Learning via Reset Deep Ensemble Agents
Woojun Kim (Carnegie Mellon University), Youngchul Sung (KAIST)
Reinforcement LearningTabular
🎯 What it does: A reset-based deep ensemble learning framework (RDE) is proposed, which sequentially resets N deep networks and adaptively integrates them to address the primacy bias in deep reinforcement learning and prevent performance collapse after resets.
Sample-efficient Multi-objective Molecular Optimization with GFlowNets
Yiheng Zhu (Zhejiang University), Jian Wu (Zhejiang University)
OptimizationDrug DiscoveryFlow-based ModelGraph
🎯 What it does: A sample-efficient multi-objective molecular optimization framework is proposed, utilizing a hypernetwork-based GFlowNet (HN-GFN) as a sampler in Bayesian optimization (MOBO) to sample diversified batches of molecules from the approximate Pareto front.
Sampling from Gaussian Process Posteriors using Stochastic Gradient Descent
Jihao Andreas Lin (University of Cambridge), Alexander Terenin (Cornell University)
OptimizationTabular
🎯 What it does: This paper proposes a method for approximately solving the posterior mean and sampling of Gaussian processes using stochastic gradient descent (SGD), and extends it to the setting of inducing points.
Sampling from Structured Log-Concave Distributions via a Soft-Threshold Dikin Walk
Oren Mangoubi (Worcester Polytechnic Institute), Nisheeth K Vishnoi
OptimizationComputational Efficiency
🎯 What it does: A Dikin Walk algorithm with a soft threshold is proposed for sampling L-Lipschitz or β-smooth log-concave distributions under polyhedral constraints.
Sampling weights of deep neural networks
Erik Lien Bolager (Technical University of Munich), Felix Dietrich (Technical University of Munich)
ClassificationOptimizationExplainability and InterpretabilityTabular
🎯 What it does: A weight sampling method based on data point pairs (SWIM) is proposed, which can construct fully connected deep neural networks without the need for gradient optimization.
SANFlow: Semantic-Aware Normalizing Flow for Anomaly Detection
Daehyun Kim (Hanyang University), Tae Hyun Kim (Hanyang University)
Anomaly DetectionFlow-based ModelImage
🎯 What it does: A framework for anomaly detection and localization based on regularized flows (SANFlow) has been designed and implemented, achieving finer density estimation by mapping input image features to different base distributions at different spatial locations and semantics.
SatLM: Satisfiability-Aided Language Models Using Declarative Prompting
Xi Ye (University of Texas at Austin), Greg Durrett (University of Texas at Austin)
Large Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the SATLM framework, which transforms natural language inference tasks into SAT problems and utilizes SAT solvers to complete the reasoning.
SaVeNet: A Scalable Vector Network for Enhanced Molecular Representation Learning
Sarp Aykent (Auburn University), Tian Xia (Auburn University)
Representation LearningDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A scalable vector network (SAVENET) has been designed and implemented for efficient learning of geometric representations of 3D molecular graphs, providing excellent predictions across various tasks.
Saving 100x Storage: Prototype Replay for Reconstructing Training Sample Distribution in Class-Incremental Semantic Segmentation
Jinpeng Chen (City University of Hong Kong), Sam Kwong (Lingnan University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: The STAR method is proposed, which utilizes the compact prototypes of each old foreground class and the repetition of background pixels to reconstruct the complete class distribution of the single-step training set, thereby addressing the issues of class distribution bias and background shift, significantly improving the performance of class incremental semantic segmentation.
Scalable Fair Influence Maximization
Xiaobin Rui (China University of Mining and Technology), Wei Chen (Microsoft Research Asia)
OptimizationGraph
🎯 What it does: A Fair Impact Maximization Algorithm (FIMM) based on Reverse Influence Sampling (RIS) is proposed, which can maximize the welfare fairness objective function under a given community structure and budget.
Scalable Membership Inference Attacks via Quantile Regression
Martin Andres Bertran (Amazon Web Services Artificial Intelligence Machine Learning), Steven Wu
Computational EfficiencyAdversarial AttackHyperparameter SearchImageTabular
🎯 What it does: Proposes a scalable membership inference attack based on quantile regression, which can complete black-box attacks with a single trained model.
Scalable Primal-Dual Actor-Critic Method for Safe Multi-Agent RL with General Utilities
Donghao Ying (University of California Berkeley), Javad Lavaei (University of California Berkeley)
Safty and PrivacyReinforcement Learning
🎯 What it does: A scalable Primal-Dual Actor-Critic method is proposed to address the Safe Multi-Agent Reinforcement Learning (Safe MARL) problem under the condition of no global observation, utilizing shadow rewards and κ-hop policies, and supporting general utility forms of objectives and constraints.
Scalable Transformer for PDE Surrogate Modeling
Zijie Li (Carnegie Mellon University), Amir Barati Farimani (Carnegie Mellon University)
TransformerMeshPhysics Related
🎯 What it does: A scalable Transformer structure (FactFormer) is proposed for proxy modeling of PDEs, which maps high-dimensional functions to one-dimensional sub-functions through learned projections, and employs axial factorization kernel integrals to achieve efficient attention computation.
Scalarization for Multi-Task and Multi-Domain Learning at Scale
Amelie Royer, Babak Ehteshami Bejnordi (Qualcomm AI Research)
Domain AdaptationOptimizationImage
🎯 What it does: This paper evaluates and optimizes the scalarization strategy in multi-task/multi-domain learning through large-scale experiments, exploring the impact of model capacity, weight selection, and gradient conflicts on performance, and proposes using population-based training (PBT) for efficient search of optimal weights;
Scale Alone Does not Improve Mechanistic Interpretability in Vision Models
Roland S. Zimmermann (Max Planck Institute for Intelligent Systems), Wieland Brendel (Max Planck Institute for Intelligent Systems)
Explainability and InterpretabilityImage
🎯 What it does: Conducted large-scale psychophysical experiments on 9 different visual models of varying sizes and architectures to quantify their unit-level mechanistic interpretability;
Scale-Space Hypernetworks for Efficient Biomedical Image Analysis
Jose Javier Gonzalez Ortiz (Massachusetts Institute of Technology), Adrian V Dalca (Massachusetts Institute of Technology)
SegmentationComputational EfficiencyConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingUltrasound
🎯 What it does: A learnable Scale-Space Hypernetworks (SSHN) model is proposed, which can generate CNN weights under various resampling coefficients at different scales within a single model, thus achieving an accurate accuracy-efficiency trade-off.
Scale-teaching: Robust Multi-scale Training for Time Series Classification with Noisy Labels
Zhen Liu (South China University of Technology), Qianli Ma (South China University of Technology)
ClassificationConvolutional Neural NetworkTime Series
🎯 What it does: Proposes the Scale-Teaching framework, which utilizes multi-scale temporal data and a cross-teacher mechanism to handle time series classification tasks with noisy labels.
ScaleLong: Towards More Stable Training of Diffusion Model via Scaling Network Long Skip Connection
Zhongzhan Huang (Sun Yat-Sen University), Liang Lin (Sun Yat-Sen University)
GenerationDiffusion modelImage
🎯 What it does: This paper analyzes the training instability encountered when using UNet in diffusion models from a theoretical perspective, attributing it to the influence of long skip connection coefficients, and proposes two coefficient scaling-based improvement methods (Constant Scaling CS and Learnable Scaling LS) to enhance training stability.
Scaling Data-Constrained Language Models
Niklas Muennighoff (Hugging Face), Colin Raffel (Hugging Face)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: In scenarios with limited unique data, we systematically explore the scale of language models and computational allocation, proposing and validating an extended Chinchilla scaling law for data duplication, and conducting over 400 experiments to quantify the benefits and diminishing returns of multi-round training.
Scaling Laws for Hyperparameter Optimization
Arlind Kadra (University of Freiburg), Josif Grabocka (University of Freiburg)
OptimizationHyperparameter SearchImageTextTabularBenchmark
🎯 What it does: A Bayesian optimization method for multi-order precision hyperparameter optimization based on a set of deep power law functions (Deep Power Law, DPL) is proposed.
Scaling laws for language encoding models in fMRI
Richard Antonello, Alexander Huth
TransformerLarge Language ModelMultimodalityMagnetic Resonance ImagingAudio
🎯 What it does: This study investigates the scaling laws of large language models and audio models in fMRI brain encoding, assessing the impact of model parameters and training data volume on encoding performance.
Scaling MLPs: A Tale of Inductive Bias
Gregor Bachmann (ETH Zurich), Thomas Hofmann (ETH Zurich)
ClassificationRecognitionImage
🎯 What it does: This study systematically evaluates the performance of Multi-Layer Perceptrons (MLP) in computer vision tasks, including training from scratch, transfer learning, large-scale pre-training, and scale law analysis.
Scaling Open-Vocabulary Object Detection
Matthias Minderer (Google DeepMind), Neil Houlsby (Google DeepMind)
Object DetectionTransformerContrastive LearningImageText
🎯 What it does: This paper improves the performance of open vocabulary object detection through large-scale self-supervised training, proposing the OWLv2 structure and the OWL-ST training process.
Scaling Riemannian Diffusion Models
Aaron Lou (Stanford University), Stefano Ermon (Stanford University)
GenerationData SynthesisDiffusion modelPoint CloudOrdinary Differential Equation
🎯 What it does: A numerical improvement to the Riemannian diffusion model is proposed, enabling efficient computation of the heat kernel on symmetric spaces, thereby achieving diffusion generation on high-dimensional manifolds.
Scaling Up Differentially Private LASSO Regularized Logistic Regression via Faster Frank-Wolfe Iterations
Edward Raff (Booz Allen Hamilton), Fred Lu (Booz Allen Hamilton)
ClassificationOptimizationSafty and PrivacyComputational EfficiencyText
🎯 What it does: A sparse data-friendly Frank-Wolfe algorithm is proposed for training LASSO regularized logistic regression models with differential privacy constraints on high-dimensional sparse data.
Scan and Snap: Understanding Training Dynamics and Token Composition in 1-layer Transformer
Yuandong Tian (Meta AI), Simon Shaolei Du
TransformerContrastive LearningText
🎯 What it does: This paper conducts a rigorous theoretical analysis of the SGD training dynamics of a single-layer Transformer in the next word prediction task, revealing the process by which the self-attention layer gradually focuses on the co-occurrence of the query word and a unique keyword, forming a 'scan-and-snap' mechanism.
Scattering Vision Transformer: Spectral Mixing Matters
Badri Narayana Patro (Microsoft), Vijay Srinivas Agneeswaran (Microsoft)
ClassificationObject DetectionSegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes the Scattering Vision Transformer (SVT), which decomposes images into low-frequency and high-frequency components through an invertible scattering network, and then uses a spectral gating network to achieve efficient feature mixing.
Scenario Diffusion: Controllable Driving Scenario Generation With Diffusion
Ethan Pronovost (Zoox), Nicholas Roy (Zoox)
GenerationData SynthesisAutonomous DrivingDiffusion modelAuto Encoder
🎯 What it does: A controllable Scenario Diffusion model is proposed for autonomous driving scene synthesis.
SceneScape: Text-Driven Consistent Scene Generation
Rafail Fridman (Weizmann Institute of Science), Tali Dekel (Weizmann Institute of Science)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: This paper proposes an unsupervised long video generation method based on text prompts and camera trajectories, utilizing a pre-trained text-to-image diffusion model and a monocular depth prediction model. By continuously optimizing during testing, a unified 3D mesh is constructed, enabling text-driven 3D consistent roaming video generation.
Schema-learning and rebinding as mechanisms of in-context learning and emergence
Sivaramakrishnan Swaminathan (Google DeepMind), Dileep George (Google DeepMind)
TransformerTextRetrieval-Augmented Generation
🎯 What it does: By introducing clone-structured causal graphs (CSCG), it is demonstrated that task transfer in in-context learning (ICL) can be achieved through template learning, context retrieval, and re-binding mechanisms.
Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Time
Zichang Liu (Rice University), Anshumali Shrivastava (Rice University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The SCISSORHANDS method is proposed, which dynamically compresses the KV cache during inference by leveraging the importance persistence hypothesis of LLMs, significantly reducing memory consumption.
Score-based Data Assimilation
François Rozet (University of Liège), Gilles Louppe (University of Liège)
Data SynthesisOptimizationScore-based ModelTime SeriesSequentialPhysics RelatedStochastic Differential Equation
🎯 What it does: A score-based data assimilation method (SDA) is proposed, which approximates the score of an entire long trajectory by learning the scores of short segments, enabling one-time parallel generation/inference of state trajectories of arbitrary lengths; at the same time, the observation model is decoupled from the score network, using observation information only during inference, thus supporting zero-shot observation scenarios.
Score-based Generative Modeling through Stochastic Evolution Equations in Hilbert Spaces
Sungbin Lim (Korea University), Sungjoon Choi (Korea University)
GenerationData SynthesisSuper ResolutionDiffusion modelScore-based ModelImageTime SeriesStochastic Differential Equation
🎯 What it does: A general framework for continuous-time fractional generative models in Hilbert space is proposed, deriving a time-reversal formula applicable to stochastic evolution equations with time-varying coefficients, and based on this, Hilbert Diffusion Models (HDM) are introduced, including HDM-SDE (infinite-dimensional SDE) and HDM-SPDE (stochastic partial differential equations).
Score-based Generative Models with Lévy Processes
Eunbi Yoon (Korea University), Sungbin Lim (Korea University)
GenerationData SynthesisScore-based ModelImageStochastic Differential Equation
🎯 What it does: This paper proposes a fractional-guided generative model using α-stable Lévy processes (Lévy-Itô model, LIM) and provides an exact time-reversal formula for its backward SDE;
Score-based Source Separation with Applications to Digital Communication Signals
Tejas Jayashankar (Massachusetts Institute of Technology), Gregory Wornell
Diffusion modelScore-based ModelAudio
🎯 What it does: The α-RGS method is proposed, using a pre-trained diffusion model as a statistical prior, performing MAP estimation with randomized Gaussian smoothing at different noise levels to achieve single-channel source separation.
SE(3) Diffusion Model-based Point Cloud Registration for Robust 6D Object Pose Estimation
Haobo Jiang (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Pose EstimationDiffusion modelPoint Cloud
🎯 What it does: A point cloud registration framework based on SE(3) diffusion models is proposed to achieve robust 6D object pose estimation.
SE(3) Equivariant Augmented Coupling Flows
Laurence Illing Midgley, José Miguel Hernández-Lobato (University of Cambridge)
GenerationDrug DiscoveryGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: Developed a SE(3) equivariant augmented coupling flow model for efficiently generating molecular conformations in Cartesian coordinates and learning the complete Boltzmann distribution.
Searching for Optimal Per-Coordinate Step-sizes with Multidimensional Backtracking
Frederik Kunstner (University of British Columbia), Nick Harvey (University of British Columbia)
OptimizationTabular
🎯 What it does: A Multidimensional Backtracking algorithm is proposed to automatically find diagonal preconditioners (coordinate-wise step sizes), which uses the Armijo condition for judgment at each step and constructs separating hyperplanes using supergradients for the search.
Secure Out-of-Distribution Task Generalization with Energy-Based Models
Shengzhuang Chen (City University of Hong Kong), Ying Wei (Nanyang Technological University)
Meta LearningDrug DiscoveryTabularSequential
🎯 What it does: This paper proposes an Energy-Based Meta-Learning (EBML) framework to simultaneously detect and adapt to out-of-distribution (OOD) tasks in meta-learning, while being compatible with various existing meta-learning algorithms.
SEEDS: Exponential SDE Solvers for Fast High-Quality Sampling from Diffusion Models
Martin Gonzalez (IRT SystemX), Nader Masmoudi (New York University)
GenerationComputational EfficiencyDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes SEEDS (Stochastic Explicit Exponential Derivative-free Solvers) — a discrete-time SDE solver that is training-free and designed for fast high-quality sampling, achieving image generation quality comparable to optimal SDE solvers while maintaining a low number of function evaluations.
Seeing is not Believing: Robust Reinforcement Learning against Spurious Correlation
Wenhao Ding (Carnegie Mellon University), Ding Zhao (Carnegie Mellon University)
Autonomous DrivingReinforcement LearningTabularBenchmark
🎯 What it does: A robust reinforcement learning framework for false correlations in state space is proposed—RSC-MDPs, along with corresponding theoretical proofs and an algorithm implementation based on RSC-SAC.
SEENN: Towards Temporal Spiking Early Exit Neural Networks
Yuhang Li (Yale University), Priyadarshini Panda (Yale University)
Computational EfficiencySpiking Neural NetworkReinforcement LearningImage
🎯 What it does: The SEENN method is proposed, which dynamically adjusts the time steps for each input sample during SNN inference to achieve early exit.
SEGA: Instructing Text-to-Image Models using Semantic Guidance
Manuel Brack (German Research Center for Artificial Intelligence), Kristian Kersting (German Research Center for Artificial Intelligence)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A technique called SEGA (Semantic Guidance) is proposed, which directly generates or edits images in the noise estimation space of diffusion models using text prompts, providing fine-grained semantic control.
Segment Any Point Cloud Sequences by Distilling Vision Foundation Models
Youquan Liu (Shanghai AI Laboratory), Ziwei Liu (Nanyang Technological University)
SegmentationAutonomous DrivingKnowledge DistillationRepresentation LearningConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: This paper proposes the Seal framework, which utilizes semantic superpixels generated by 2D vision foundation models (such as SAM, X-Decoder, etc.) to perform unsupervised cross-modal contrastive learning on automotive point cloud sequences, addressing the issues of high annotation costs, cross-modal consistency, and generalization.
Segment Anything in 3D with NeRFs
Jiazhong Cen (Shanghai Jiao Tong University), Qi Tian (Huawei Inc.)
SegmentationNeural Radiance FieldImage
🎯 What it does: Using pre-trained NeRF as a 3D prior, combined with SAM for single-view manual click prompts, we propose the SA3D framework. Through an iterative process of 'mask inverse rendering' and 'cross-view self-prompting', the 2D segmentation results are projected onto a 3D voxel grid to achieve segmentation of 3D objects.
Segment Anything in High Quality
Lei Ke (ETH Zürich), Fisher Yu (ETH Zürich)
SegmentationTransformerImage
🎯 What it does: An improved Segment Anything Model (SAM) is proposed, named HQ-SAM, which retains the original prompt and zero-shot capabilities of SAM while generating higher quality and more refined boundary segmentation masks.
Segment Everything Everywhere All at Once
Xueyan Zou, Yong Jae Lee
SegmentationTransformerImageVideoTextMultimodality
🎯 What it does: We propose SEEM, a unified multimodal segmentation model that supports various prompts such as text, points, boxes, strokes, and reference images, completing all pixel and semantic segmentation in one go.
SegRefiner: Towards Model-Agnostic Segmentation Refinement with Discrete Diffusion Process
Mengyu Wang (Beijing Jiaotong University), Yunchao Wei (Beijing Jiaotong University)
SegmentationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: We propose a model-agnostic segmentation mask refinement method called SegRefiner, which enhances mask quality through a multi-step denoising process using discrete diffusion.
Selective Amnesia: A Continual Learning Approach to Forgetting in Deep Generative Models
Alvin Heng (National University of Singapore), Harold Soh (National University of Singapore)
GenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: A selective forgetting framework based on continual learning, called Selective Amnesia (SA), is proposed, which can delete or remap specified concepts in pre-trained conditional variational generative models (such as VAE, DDPM, Stable Diffusion) without retraining the model or accessing the original data.
Selective Sampling and Imitation Learning via Online Regression
Ayush Sekhari (Massachusetts Institute of Technology), Runzhe Wu (Cornell University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper studies interactive imitation learning using selective sampling and online regression in the presence of noisy expert feedback, and presents the algorithms SAGE and RAVIOLI.
Selectively Sharing Experiences Improves Multi-Agent Reinforcement Learning
Matthias Gerstgrasser (Harvard University), Sarah Keren (Technion Israel Institute of Technology)
Reinforcement Learning
🎯 What it does: This paper studies an algorithm called SUPER that accelerates learning in multi-agent reinforcement learning through limited prioritized experience sharing.
Selectivity Drives Productivity: Efficient Dataset Pruning for Enhanced Transfer Learning
Yihua Zhang (Michigan State University), Sijia Liu (Michigan State University)
Computational EfficiencyData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This study investigates strategies for pruning the source dataset in transfer learning to enhance pre-training efficiency and downstream performance.
Self-Adaptive Motion Tracking against On-body Displacement of Flexible Sensors
Chengxu Zuo (Xiamen University), Yipeng Qin (Cardiff University)
Object TrackingDomain AdaptationRecurrent Neural NetworkSupervised Fine-TuningTime Series
🎯 What it does: This study investigates the data distribution drift caused by displacement of wearable flexible sensors at different wearing positions and proposes an unsupervised adaptive motion tracking network.
Self-Chained Image-Language Model for Video Localization and Question Answering
Shoubin Yu (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
RecognitionRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoText
🎯 What it does: This paper proposes a Self-Chained Video Localization and Answering (SeViLA) framework based on a single image-language model BLIP-2, which first uses a Localizer to locate key frames with language perception, and then employs an Answerer to perform video question answering or event prediction on these frames, refining the Localizer through pseudo-labels generated from the answers.
Self-Consistent Velocity Matching of Probability Flows
Lingxiao Li (Massachusetts Institute of Technology), Justin Solomon (Massachusetts Institute of Technology)
OptimizationFlow-based ModelTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a self-consistent velocity matching method that utilizes the Lagrangian perspective of probability flow to solve large-scale mass conservation partial differential equations, avoiding time and space discretization.
Self-Correcting Bayesian Optimization through Bayesian Active Learning
Carl Hvarfner (Lund University), Luigi Nardi (Stanford University)
OptimizationHyperparameter Search
🎯 What it does: Two sampling strategies are proposed: Statistical Distance-based Active Learning (SAL) and Self-Correcting Bayesian Optimization (SCoreBO), which explicitly utilize the uncertainty of Gaussian process hyperparameters to guide query points.
Self-Evaluation Guided Beam Search for Reasoning
Yuxi Xie (National University of Singapore), Qizhe Xie
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Using the self-evaluation mechanism of LLM for step-by-step guidance in multi-step reasoning to improve the quality of the reasoning chain.
Self-Predictive Universal AI
Elliot Catt (Google DeepMind), Joel Veness (Google DeepMind)
Reinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes Self-AIXI, a general AI agent based on self-prediction (self-modeling) rather than traditional planning, and proves its convergence to optimal AIXI in expectation, inheriting the intelligence measure of Legg-Hutter and self-optimizing properties.