NeurIPS 2023 Papers — Page 27
Conference on Neural Information Processing Systems · 3218 papers
Self-Refine: Iterative Refinement with Self-Feedback
Aman Madaan (Carnegie Mellon University), Peter Clark (Google Deepmind)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The SELF-REFINE method is proposed, which utilizes large language models (LLMs) to generate a draft and then allows the same model to provide feedback and iteratively improve it, thereby enhancing the generation quality.
Self-supervised Graph Neural Networks via Low-Rank Decomposition
Liang Yang (Hebei University of Technology), Chuan Wang (Institute of Information Engineering)
Representation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes a self-supervised graph neural network framework—Low-Rank Decomposition-based GNN (LRD-GNN), which learns node representations by performing low-rank matrix or tensor decomposition on the attribute matrix of each ego-network, and constructs two instances based on matrices and tensors.
Self-Supervised Learning of Representations for Space Generates Multi-Modular Grid Cells
Rylan Schaeffer (Stanford University), Ila R Fiete
Representation LearningRecurrent Neural NetworkContrastive LearningSequential
🎯 What it does: A recurrent neural network is trained using a self-supervised learning framework to naturally generate spatial representations of multi-module grid cells based solely on speed input.
Self-Supervised Learning with Lie Symmetries for Partial Differential Equations
Grégoire Mialon (Meta), Bobak Kiani
Representation LearningConvolutional Neural NetworkContrastive LearningTime SeriesPhysics Related
🎯 What it does: Pre-training on unlabeled data using self-supervised learning (SSL) and the Lie symmetry of PDEs to obtain a general representation of PDEs;
Self-Supervised Motion Magnification by Backpropagating Through Optical Flow
Zhaoying Pan (University of Michigan), Andrew Owens (University of Michigan)
RestorationGenerationData SynthesisGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: This paper proposes a self-supervised motion magnification method that utilizes a differentiable optical flow network as a loss function to train a generative network to amplify subtle motions in input videos by a specified magnification factor;
Self-supervised Object-Centric Learning for Videos
Görkay Aydemir (Koç University), Fatma Guney
Object DetectionObject TrackingSegmentationTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: A completely unsupervised self-supervised object center learning framework, SOLV, is proposed for multi-object segmentation in real-world videos, capable of generating instance segmentation results for tracking without using any additional modalities or supervisory information.
Self-Supervised Reinforcement Learning that Transfers using Random Features
Boyuan Chen (Massachusetts Institute of Technology), Abhishek Gupta (University of Washington)
Reinforcement Learning
🎯 What it does: This paper proposes a self-supervised reinforcement learning framework RaMP, which constructs a task-independent Q base by learning the multi-step accumulation of random features, thereby enabling rapid transfer between different reward functions.
Self-supervised video pretraining yields robust and more human-aligned visual representations
Nikhil Parthasarathy (Google DeepMind), Olivier J Henaff
Representation LearningConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: A new method called VITO is proposed, which learns visual representations through self-supervised video pre-training, aiming to improve the model's performance in image and video understanding tasks.
Self-Supervised Visual Acoustic Matching
Arjun Somayazulu (University of Texas at Austin), Kristen Grauman (Meta Platforms)
Generative Adversarial NetworkMultimodalityAudio
🎯 What it does: Proposes a self-supervised visual-acoustic matching method
Self-Weighted Contrastive Learning among Multiple Views for Mitigating Representation Degeneration
Jie Xu (University of Electronic Science and Technology of China), Xiaofeng Zhu (University of Electronic Science and Technology of China)
Representation LearningAuto EncoderContrastive LearningImageVideo
🎯 What it does: This paper proposes a self-weighted multi-view contrastive learning framework SEM, which utilizes reconstruction regularization to alleviate the representation degradation problem in multi-view contrastive learning.
Semantic HELM: A Human-Readable Memory for Reinforcement Learning
Fabian Paischer (Johannes Kepler University), Sepp Hochreiter (Johannes Kepler University)
TransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A reinforcement learning framework (SHELM) has been designed and implemented to map visual inputs to interpretable language memory, utilizing CLIP to retrieve semantically similar tokens and feed them into a pre-trained language model as memory;
Semantic Image Synthesis with Unconditional Generator
JungWoo Chae (LG CNS AI Research), Youngjung Uh (Yonsei University)
SegmentationGenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: Proposes a semantic image synthesis method that allows spatial control of a pre-trained unconditional generator without requiring extensive semantic annotations.
Semantic segmentation of sparse irregular point clouds for leaf/wood discrimination
Yuchen BAI, Florence Forbes (University of Grenoble Alpes)
ClassificationSegmentationPoint Cloud
🎯 What it does: A SOUL network based on PointNet++ is proposed for semantic segmentation of leaf and wood materials from sparse and heterogeneous point clouds generated by drone LiDAR.
Semi-Implicit Denoising Diffusion Models (SIDDMs)
yanwu xu, Tingbo Hou (Boston University)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a Semi-Implicit Denoising Diffusion Model (SIDDM), which decomposes the denoising distribution into implicit edge matching and explicit conditional matching (AFD), achieving large step sampling while maintaining high-quality samples.
Semi-Supervised Contrastive Learning for Deep Regression with Ordinal Rankings from Spectral Seriation
Weihang Dai (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
Contrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A new semi-supervised contrastive learning framework CLSS is proposed, which utilizes the feature similarity matrix of unlabeled samples to recover ordinal relationships through spectral sequence algorithms, and applies it to contrastive learning and predictive supervision.
Semi-Supervised Domain Generalization with Known and Unknown Classes
Lei Zhang (Nanjing University), Wei Wang (Nanjing University)
Domain AdaptationImage
🎯 What it does: This paper proposes a semi-supervised domain generalization method called CWAEE, which can identify known and unknown classes during training and utilize them to enhance generalization capabilities for unknown target domains.
Sensitivity in Translation Averaging
Lalit Manam (Indian Institute of Science), Venu Madhav Govindu (Indian Institute of Science)
OptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper addresses the problem of translation averaging in three-dimensional structured light beams given a specific relative direction, analyzes the sensitivity under input uncertainty, and proposes a condition number determination method based on the angle matrix; subsequently, an efficient algorithm is designed to filter out 'oblique triangles' (i.e., triangles prone to failure) to enhance the robustness of the solution; finally, the effectiveness of this filtering method is validated on real SfM data.
Separable Physics-Informed Neural Networks
Junwoo Cho (Sungkyunkwan University), Eunbyung Park (Sungkyunkwan University)
OptimizationComputational EfficiencyTabularPhysics Related
🎯 What it does: This paper proposes a Separable Physics-Informed Neural Network (SPINN) that efficiently trains PINNs under large-scale sampling points (>10⁷) by decomposing multi-dimensional coordinates into independent one-dimensional MLPs and utilizing forward automatic differentiation, addressing the computational and memory bottlenecks of traditional PINNs in high-dimensional PDE solving.
Sequential Memory with Temporal Predictive Coding
Mufeng Tang (MRC Brain Network Dynamics Unit University of Oxford), Rafal Bogacz (MRC Brain Network Dynamics Unit University of Oxford)
Sequential
🎯 What it does: A temporal prediction model based on predictive coding (tPC) is proposed and implemented for sequence memory.
Sequential Predictive Two-Sample and Independence Testing
Aleksandr Podkopaev (Walmart Global Technology), Aaditya Ramdas (Carnegie Mellon University)
Anomaly DetectionOptimizationConvolutional Neural NetworkRecurrent Neural NetworkImageTabular
🎯 What it does: A new framework for sequential nonparametric two-sample tests and independence tests is proposed, enabling online decision-making based on predictive models and betting principles.
Sequential Preference Ranking for Efficient Reinforcement Learning from Human Feedback
Minyoung Hwang (Seoul National University), Songhwai Oh (Seoul National University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: A framework named SeqRank is proposed, which enhances the efficiency of human feedback through sequential preference ranking, reducing the number of comparisons required.
Sequential Subset Matching for Dataset Distillation
Jiawei Du (Agency for Science Technology and Research), Joey Tianyi Zhou (Agency for Science Technology and Research)
Data SynthesisOptimizationKnowledge DistillationMeta LearningImage
🎯 What it does: The SeqMatch method is proposed, which divides synthetic data into multiple subsets and optimizes them sequentially to address the issue of insufficient high-order feature distillation caused by excessive compression of low-level features.
Setting the Trap: Capturing and Defeating Backdoors in Pretrained Language Models through Honeypots
Ruixiang Tang (Rice University), Xia Hu (Rice University)
ClassificationAdversarial AttackTransformerSupervised Fine-TuningText
🎯 What it does: A honeypot module is proposed, which inserts a small classifier at the lower layers of a pre-trained language model (PLM) to specifically absorb and cover backdoor information during the fine-tuning process, allowing the main network to only learn the original task.
SGFormer: Simplifying and Empowering Transformers for Large-Graph Representations
Qitian Wu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Representation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes SGFormer, a simplified Transformer model that utilizes a single-layer global linear attention combined with GCN to achieve efficient learning of large-scale graph node representations.
SHAP-IQ: Unified Approximation of any-order Shapley Interactions
Fabian Fumagalli (Bielefeld University), Barbara Hammer (Bielefeld University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: We propose SHAP-IQ, a unified sampling approximation method for approximating the Cardinal Interaction Index (CII) of any order, and evaluate it on language models, image classification, and high-dimensional synthetic models.
Shape Non-rigid Kinematics (SNK): A Zero-Shot Method for Non-Rigid Shape Matching via Unsupervised Functional Map Regularized Reconstruction
Souhaib Attaiki (École Polytechnique), Maks Ovsjanikov (École Polytechnique)
Object DetectionSegmentationGraph Neural NetworkDiffusion modelPoint CloudMesh
🎯 What it does: A zero-shot non-rigid shape matching method SNK is proposed, achieving shape matching through unsupervised feature mapping and decoder reconstruction directly on a single pair of shapes.
Shared Adversarial Unlearning: Backdoor Mitigation by Unlearning Shared Adversarial Examples
Shaokui Wei (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
OptimizationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A Shared Adversarial Unlearning (SAU) method is proposed to eliminate backdoor attacks on deep neural networks by generating and removing shared adversarial samples.
Sharp Bounds for Generalized Causal Sensitivity Analysis
Dennis Frauen (LMU Munich), Stefan Feuerriegel (LMU Munich)
Flow-based ModelTabularTime Series
🎯 What it does: This paper proposes a generalized marginal sensitivity model (GMSM) and derives sharp boundaries under various causal effects (average treatment effect, mediation effect, path effect, and distribution effect), providing scalable algorithms to estimate these boundaries from observational data.
Sharp Calibrated Gaussian Processes
Alexandre Capone (Technical University of Munich), Geoff Pleiss (University of British Columbia)
Tabular
🎯 What it does: A new Gaussian Process (GP) regression calibration method is proposed, which trains a GP specifically for predicting means and another GP for predicting quantiles, achieving a balance between calibration and sharpness through adjustable hyperparameters.
Sharp Recovery Thresholds of Tensor PCA Spectral Algorithms
Michael Jacob Feldman, David Donoho
🎯 What it does: This paper studies the matrixization methods of tensor PCA (tensor unfolding, partial tracking, power iteration, recursive unfolding), providing their recovery thresholds and limiting cosine similarities, and proving that exact recovery can be achieved above the unfolding threshold.
Sharp Spectral Rates for Koopman Operator Learning
Vladimir R Kostic, Massimiliano Pontil (Istituto Italiano di Tecnologia)
Time SeriesSequentialPhysics Related
🎯 What it does: This paper presents a non-asymptotic learning theory for Koopman operator learning, providing upper bounds on the spectral estimation errors for two mainstream estimators: EDMD (PCR) and RRR.
Sharpness Minimization Algorithms Do Not Only Minimize Sharpness To Achieve Better Generalization
Kaiyue Wen (Tsinghua University), Tengyu Ma (Stanford University)
OptimizationTabular
🎯 What it does: This study investigates the relationship between 'flatness' in over-parameterized neural networks and generalization performance, proposing three different scenarios and providing theoretical and experimental results.
Sharpness-Aware Minimization Leads to Low-Rank Features
Maksym Andriushchenko (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
ClassificationOptimizationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper studies the impact of Sharpness-Aware Minimization (SAM) on the feature rank of deep networks, finding that SAM can significantly reduce the feature rank across different models (ResNet, ViT, MLP-Mixer, etc.) and different tasks (classification, contrastive learning);
Sheaf Hypergraph Networks
Iulia Duta (University of Cambridge), Pietro Lio
Graph Neural NetworkGraphBenchmark
🎯 What it does: This paper proposes an extension of the cellular layer sheaf structure to hypergraphs and constructs both linear and nonlinear Sheaf Hypergraph Laplacians based on this, leading to the design of two new hypergraph network models: SheafHyperGNN and SheafHyperGCN.
SheetCopilot: Bringing Software Productivity to the Next Level through Large Language Models
Hongxin Li (Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
OptimizationAI Code AssistantTransformerLarge Language ModelTabular
🎯 What it does: This paper proposes a SheetCopilot agent based on large language models that can decompose natural language instructions into atomic actions to automatically control spreadsheets to complete complex tasks.
ShiftAddViT: Mixture of Multiplication Primitives Towards Efficient Vision Transformer
Haoran You (Georgia Institute of Technology), Yingyan Celine Lin (Georgia Institute of Technology)
ClassificationComputational EfficiencyTransformerMixture of ExpertsImage
🎯 What it does: Reparameterize the pre-trained Vision Transformer by replacing multiplication operations with shifts and additions to construct the ShiftAddViT model.
SHOT: Suppressing the Hessian along the Optimization Trajectory for Gradient-Based Meta-Learning
JunHoo Lee (Seoul National University), Nojun Kwak (Seoul National University)
OptimizationMeta LearningImage
🎯 What it does: This paper proposes an explicit suppression of the Hessian matrix on the optimization trajectory in gradient-based meta-learning, achieved by minimizing the distance between the target model and the reference model.
Should I Stop or Should I Go: Early Stopping with Heterogeneous Populations
Hammaad Adam (Massachusetts Institute of Technology), Allison Koenecke (Cornell University)
TabularTime Series
🎯 What it does: This paper proposes a method called CLASH, designed to prevent harmful effects from premature stopping of randomized experiments in heterogeneous populations.
Should Under-parameterized Student Networks Copy or Average Teacher Weights?
Berfin Simsek (New York University), Johanni Brea (École Polytechnique Fédérale de Lausanne)
OptimizationKnowledge DistillationTabular
🎯 What it does: This paper studies the under-parameterized case where the input follows a standard Gaussian distribution, the teacher network has orthogonal weights and is unit normalized, and the width of the student network n is less than the width of the teacher k. It analyzes and solves the critical points in the loss landscape, proposes the 'copy-average' structure, and proves that this structure can achieve global optimality under the erf activation.
Should We Learn Most Likely Functions or Parameters?
Shikai Qiu (New York University), Andrew Gordon Wilson (New York University)
OptimizationImageTabular
🎯 What it does: This paper studies the MAP estimation directly on the function space rather than the parameter space in supervised learning, and derives the corresponding objective function; by analyzing its theoretical properties and experimental results, it proposes a scalable L-MAP approximation method for large neural networks.
Siamese Masked Autoencoders
Agrim Gupta (Stanford University), Li Fei-Fei (Stanford University)
SegmentationPose EstimationTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: This paper proposes Siamese Masked Autoencoders (SiamMAE), which learns visual correspondences by masking a high proportion of future frames in video frame pairs and predicting the missing patches.
SimFBO: Towards Simple, Flexible and Communication-efficient Federated Bilevel Learning
Yifan Yang (University at Buffalo), Kaiyi Ji (University at Buffalo)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes SimFBO and its system heterogeneous robust version ShroFBO, providing a single-loop, no sub-loop, and client sampling without replacement federated bi-level optimization framework;
Similarity-based cooperative equilibrium
Caspar Oesterheld (Carnegie Mellon University), Jakob Nicolaus Foerster
Reinforcement Learning from Human FeedbackTabular
🎯 What it does: This paper proposes a similarity-based cooperative equilibrium model aimed at addressing the lack of cooperation among machine learning agents in social dilemmas (such as the prisoner's dilemma). By introducing a novel concept of a differential meta game, agents only need to observe a single value representing mutual similarity to achieve cooperation.
Similarity, Compression and Local Steps: Three Pillars of Efficient Communications for Distributed Variational Inequalities
Aleksandr Beznosikov (Innopolis University), Alexander Gasnikov (Moscow Institute of Physics and Technology)
OptimizationFederated LearningComputational EfficiencyTabular
🎯 What it does: A novel algorithm (Algorithm 1 and Algorithm 2) is proposed that simultaneously combines compression, data similarity, and local steps in distributed variational inequalities (VI) and saddle point problems (SPP), along with its theoretical convergence analysis and communication complexity.
SimMMDG: A Simple and Effective Framework for Multi-modal Domain Generalization
Hao Dong (ETH Zurich), Olga Fink (EPFL)
Domain AdaptationContrastive LearningMultimodality
🎯 What it does: A framework for multimodal domain generalization, SimMMDG, is proposed by splitting each modality feature into shared and exclusive parts, and enhancing the model's generalization ability to unknown domains through supervised contrastive learning, distance constraints, and cross-modal translation.
SimMTM: A Simple Pre-Training Framework for Masked Time-Series Modeling
Jiaxiang Dong (Tsinghua University), Mingsheng Long (Tsinghua University)
ClassificationTransformerContrastive LearningTime SeriesSequential
🎯 What it does: This paper proposes SimMTM, a self-supervised pre-training framework for time series based on multiple masked sequence aggregation, aimed at improving the performance of prediction and classification tasks.
Simple and Asymmetric Graph Contrastive Learning without Augmentations
Teng Xiao (Pennsylvania State University), Suhang Wang (Zhejiang University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes GraphACL, a framework for asymmetric contrastive learning that does not rely on graph data augmentation. It learns node representations by utilizing the one-hop neighbor context and two-hop homophily, making it applicable to both homophilic and heterophilic graphs.
Simple and Controllable Music Generation
Jade Copet (Meta AI), Alexandre Défossez (Meta AI)
GenerationTransformerLarge Language ModelAudio
🎯 What it does: This paper presents a single-stage Transformer language model called MUSICGEN, designed for music generation based on text or melody, and reduces autoregressive steps through a codebook interleaving pattern, achieving high-quality monophonic and stereo audio.
Simple, Scalable and Effective Clustering via One-Dimensional Projections
Moses Charikar (Stanford University), Erik Waingarten (University of Pennsylvania)
OptimizationComputational EfficiencyTabular
🎯 What it does: We propose a simple random projection clustering algorithm PRONE, which first projects data points to one dimension and then performs k-means++ stacking in that one-dimensional space, with an overall running time of O(nnz(X) + n log n), independent of k.
Simplicity Bias in 1-Hidden Layer Neural Networks
Depen Morwani (Harvard University), Praneeth Netrapalli (Google Research)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper rigorously defines the 'simplicity bias' of a single hidden layer neural network and theoretically proves that such networks primarily rely on low-dimensional input subspaces under the infinite width limit; it also empirically validates this phenomenon on several real datasets and proposes a multi-model training method based on orthogonal projection, OrthoP, to enhance model diversity and robustness.
Simplifying Neural Network Training Under Class Imbalance
Ravid Shwartz-Ziv (New York University), Andrew Gordon Wilson (New York University)
ClassificationOptimizationData-Centric LearningConvolutional Neural NetworkContrastive LearningImageTabular
🎯 What it does: By adjusting standard training components such as batch size, data augmentation, optimizers, and label smoothing, state-of-the-art performance has been achieved on class-imbalanced data.
Simultaneous embedding of multiple attractor manifolds in a recurrent neural network using constrained gradient optimization
Haggai Agmon (Hebrew University of Jerusalem), Yoram Burak (Hebrew University of Jerusalem)
OptimizationRecurrent Neural NetworkSequential
🎯 What it does: Multiple continuous attractors are embedded simultaneously in a recurrent neural network, and subtle adjustments to synaptic weights are made using gradient optimization, flattening the energy landscape, significantly reducing memory drift, and enhancing network stability.
Single-Call Stochastic Extragradient Methods for Structured Non-monotone Variational Inequalities: Improved Analysis under Weaker Conditions
Sayantan Choudhury (Johns Hopkins University), Nicolas Loizou (Johns Hopkins University)
Optimization
🎯 What it does: This paper proposes a new convergence analysis framework for the Single-Call Stochastic Extragradient methods (such as SPEG and SOG), providing convergence proofs for two types of structured non-monotone variational inequalities (quasi-strongly monotone and weak Minty) under weak assumptions (without the need for bounded variance or growth conditions), and unifying the consideration of arbitrary sampling (including uniform, importance sampling, and arbitrary mini-batch).
Single-Pass Pivot Algorithm for Correlation Clustering. Keep it simple!
Konstantin Makarychev, Sayak Chakrabarty
Optimization
🎯 What it does: A simple single-pass semi-streaming algorithm is proposed for correlated clustering, achieving a (3 + ε)-approximation with O(n/ε) memory usage.
Single-Stage Visual Query Localization in Egocentric Videos
Hanwen Jiang (University of Texas at Austin), Kristen Grauman (Meta)
RecognitionObject DetectionTransformerVideo
🎯 What it does: A single-stage, end-to-end trainable visual query localization framework VQLoC is proposed, specifically designed for the visual query localization task in first-person long-duration videos.
Sketching Algorithms for Sparse Dictionary Learning: PTAS and Turnstile Streaming
Gregory Dexter (Purdue University), Taisuke Yasuda (Carnegie Mellon University)
Optimization
🎯 What it does: A PTAS based on sketching is proposed for sparse dictionary learning and Euclidean k-means clustering, along with corresponding low-space streaming algorithms.
Sketchy: Memory-efficient Adaptive Regularization with Frequent Directions
Vladimir Feinberg (Google DeepMind), Elad Hazan (Princeton University)
OptimizationGraph Neural NetworkImageAudio
🎯 What it does: This paper proposes a Sketchy optimizer based on Frequent Directions, which uses low-rank sparse approximations to replace the full matrix AdaGrad and Shampoo, significantly reducing memory requirements while maintaining approximate performance.
Skill-it! A data-driven skills framework for understanding and training language models
Mayee F Chen, Christopher Re
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A skill framework oriented towards language models is proposed, defining skills and ordered skill sets, and based on this, an online data sampling algorithm SKILL-IT is designed to improve model performance on downstream tasks under a fixed data budget.
SLaM: Student-Label Mixing for Distillation with Unlabeled Examples
Vasilis Kontonis (University of Texas at Austin), Erik Vee (Google Research)
ClassificationKnowledge DistillationImageText
🎯 What it does: A Student-Label Mixing (SLaM) method based on a teacher noise model is proposed for knowledge distillation on unlabeled samples, correcting noise by mixing student predictions with teacher predictions in the student loss.
SLIBO-Net: Floorplan Reconstruction via Slicing Box Representation with Local Geometry Regularization
Jheng-Wei Su (National Tsing Hua University), Hung-Kuo Chu (National Tsing Hua University)
GenerationTransformerPoint Cloud
🎯 What it does: To automatically reconstruct 2D floor plans from unstructured 3D point clouds, the SLIBO-Net method is proposed for end-to-end prediction.
Slimmed Asymmetrical Contrastive Learning and Cross Distillation for Lightweight Model Training
Jian Meng (Cornell Tech), Jae-sun Seo (Cornell Tech)
Computational EfficiencyKnowledge DistillationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper designs a lightweight contrastive learning framework SACL-XD, which can train small networks from scratch and achieve high accuracy.
SLM: A Smoothed First-Order Lagrangian Method for Structured Constrained Nonconvex Optimization
Songtao Lu (IBM Research)
OptimizationTabular
🎯 What it does: This paper studies structured non-convex function constrained optimization problems and proposes a Smooth First-order Lagrangian Method (SLM).
Slot-guided Volumetric Object Radiance Fields
DI QI, Xiangyu Zhang (MEGVII Technology Inc)
SegmentationRepresentation LearningTransformerNeural Radiance FieldPoint Cloud
🎯 What it does: A slot-guided volume radiance field (sVORF) framework is proposed to achieve single-view unsupervised 3D scene decomposition and object-level representation learning.
SlotDiffusion: Object-Centric Generative Modeling with Diffusion Models
Ziyi Wu (University of Toronto), Animesh Garg (University of Toronto)
SegmentationGenerationDiffusion modelImageVideo
🎯 What it does: We propose SlotDiffusion, an unsupervised object center learning framework based on latent diffusion models, which achieves high-quality scene segmentation and generation in images and videos.
Slow and Weak Attractor Computation Embedded in Fast and Strong E-I Balanced Neural Dynamics
Xiaohan Lin (Peking University), Si Wu (Peking University)
Spiking Neural Network
🎯 What it does: By simultaneously incorporating strong fast random E-I synapses and weak slow structured CANN synapses within the same neural circuit, the coexistence and collaborative operation of continuous attractor networks and balance networks were achieved.
Small batch deep reinforcement learning
Johan Samir Obando Ceron (Mila Université de Montréal), Pablo Samuel Castro (Google DeepMind)
Computational EfficiencyReinforcement LearningSequential
🎯 What it does: In value-based deep reinforcement learning, the impact of batch size on learning performance and computational efficiency has been systematically studied, revealing that smaller batch sizes can significantly enhance performance.
Small Total-Cost Constraints in Contextual Bandits with Knapsacks, with Application to Fairness
Evgenii E Chzhen (Universite Paris-Saclay), Gilles Stoltz (Universite Paris-Saclay)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes a direct and adaptive dual strategy that utilizes projected gradient descent to solve the contextual multi-armed bandit problem under constraints of contextual robustness and affordable costs, achieving controllable total costs and fast convergence in scenarios with fairness constraints.
SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation
Mengcheng Lan (Nanyang Technological University), Wayne Zhang (SenseTime Research)
SegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Designed the SmooSeg method, which utilizes smoothness priors and energy minimization to achieve unsupervised semantic segmentation.
Smooth Flipping Probability for Differential Private Sign Random Projection Methods
Ping Li (LinkedIn Ads), Xiaoyun Li (LinkedIn Ads)
ClassificationRetrievalSafty and PrivacyGaussian SplattingImageText
🎯 What it does: A differential privacy algorithm based on Random Projection (RP) and Sign Random Projection (SignRP) is proposed, significantly enhancing privacy protection and practicality through the optimal Gaussian mechanism and smooth flipping probability.
Smooth, exact rotational symmetrization for deep learning on point clouds
Sergey Pozdnyakov, Michele Ceriotti (Ecole Polytechnique Federale de Lausanne)
Computational EfficiencyDrug DiscoveryTransformerPoint CloudBenchmarkPhysics Related
🎯 What it does: The ECSE protocol is proposed, which adds rotation and other transformations to any point cloud model in the later stage, and designs a non-invariant Point Edge Transformer (PET). By implementing ECSE, state-of-the-art performance is achieved in chemical and materials science tasks after achieving invariance.
Smoothed Analysis of Sequential Probability Assignment
Alankrita Bhatt (California Institute of Technology), Abhishek Shetty (University of California)
Sequential
🎯 What it does: This paper studies the smooth analysis of the sequence probability distribution problem with context, exploring the information-theoretically optimal minimax rate and an algorithmic simplification framework involving the maximum likelihood estimation (MLE) oracle.
Smoothed Online Learning for Prediction in Piecewise Affine Systems
Adam Block (Massachusetts Institute of Technology), Russ Tedrake (Massachusetts Institute of Technology)
Optimization
🎯 What it does: This paper studies the online learning and prediction problem in piecewise affine (PWA) systems and proposes a stochastic smoothing-based algorithm that achieves low regret performance.
SmoothHess: ReLU Network Feature Interactions via Stein's Lemma
Max Torop (Northeastern University), Jennifer Dy (Northeastern University)
OptimizationAdversarial AttackConvolutional Neural NetworkGaussian SplattingImageTime SeriesBiomedical Data
🎯 What it does: Using Gaussian convolution on the output of ReLU networks to obtain a smooth function, we estimate its Hessian to capture the second-order interactions of features.
Smoothing the Landscape Boosts the Signal for SGD: Optimal Sample Complexity for Learning Single Index Models
Alex Damian (Princeton University), Jason D. Lee (Princeton University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the task of learning single-index models under isotropic Gaussian distributions and proposes a method to improve the learning efficiency of online stochastic gradient descent (SGD) by smoothing the loss function.
SNAP: Self-Supervised Neural Maps for Visual Positioning and Semantic Understanding
Paul-Edouard Sarlin (Google), Simon Lynen (Google)
RecognitionPose EstimationRetrievalContrastive LearningImageMultimodality
🎯 What it does: Utilizing multi-modal images (ground view and aerial images) for self-supervised learning of a neural 2D map, achieving visual localization and semantic understanding.
SnapFusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds
Yanyu Li (Snap Inc.), Jian Ren (Snap Inc.)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationDiffusion modelImage
🎯 What it does: A text-to-image diffusion model has been implemented on mobile devices, completing in under 2 seconds using 8 denoising steps, with image quality comparable to Stable Diffusion v1.5.
SNEkhorn: Dimension Reduction with Symmetric Entropic Affinities
Hugues Van Assel (École normale supérieure de Lyon), Nicolas Courty (Université Bretagne Sud)
OptimizationImageBiomedical Data
🎯 What it does: This paper proposes Symmetric Entropic Affinities and a dimensionality reduction algorithm based on this matrix called SNEkhorn.
SOAR: Improved Indexing for Approximate Nearest Neighbor Search
Philip Sun (Google Research), Sanjiv Kumar (Google Research)
RetrievalCompressionOptimizationTabular
🎯 What it does: Proposes the SOAR index, which enhances the quality of nearest neighbor retrieval through orthogonal amplification of residuals for multi-allocation.
SOC: Semantic-Assisted Object Cluster for Referring Video Object Segmentation
Zhuoyan Luo (Tsinghua University), Yujiu Yang (Tsinghua University)
Object DetectionSegmentationTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: Proposes the SOC framework to achieve video-level multimodal understanding, enhancing RVOS through semantic-assisted object clustering.
Social Motion Prediction with Cognitive Hierarchies
Wentao Zhu (Peking University), Yizhou Wang (Peking University)
TransformerReinforcement LearningGenerative Adversarial NetworkVideo
🎯 What it does: This study focuses on social action prediction, constructing a multi-person 3D action dataset based on team basketball called Wusi, and proposes modeling multi-person motion prediction as a multi-agent reinforcement learning (MARL) problem, incorporating cognitive hierarchy theory to achieve recursive strategic reasoning.
SODA: Robust Training of Test-Time Data Adaptors
Zige Wang (Peking University), Bo Han (Hong Kong Baptist University)
Domain AdaptationOptimizationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: In scenarios where model parameters are inaccessible, a test-time data adaptation method called SODA based on zero-order optimization is proposed.
Soft-Unification in Deep Probabilistic Logic
Jaron Maene (KU Leuven), Luc De Raedt (KU Leuven)
GraphTabular
🎯 What it does: A soft-unification mechanism based on deep learning is proposed, and the DeepSoftLog framework is constructed, integrating probabilistic logic programming with learnable embeddings to achieve end-to-end differentiable neural symbolic reasoning.
Softmax Output Approximation for Activation Memory-Efficient Training of Attention-based Networks
Changhyeon Lee (Ulsan National Institute of Science and Technology), Seulki Lee (Ulsan National Institute of Science and Technology)
TransformerText
🎯 What it does: This paper proposes a method to approximate the softmax output during the training process of attention networks such as Transformer, reducing the activation memory usage.
SOL: Sampling-based Optimal Linear bounding of arbitrary scalar functions
Yuriy Biktairov (University of Southern California), Jyotirmoy Deshmukh (University of Southern California)
OptimizationImage
🎯 What it does: An optimization linear upper bound (SOL) method based on sampling is proposed to generate approximately optimal linear upper and lower bounds for any Lipschitz continuous scalar function.
Solving a Class of Non-Convex Minimax Optimization in Federated Learning
Xidong Wu (University of Pittsburgh), Heng Huang (University of Maryland)
OptimizationFederated LearningImage
🎯 What it does: This paper studies non-convex-convex, non-convex-strongly convex, and non-convex-PL (Polyak-Lojasiewicz) minimax optimization problems in a federated learning environment, and proposes two improved algorithms, FedSGDA+ and FedSGDA-M.
Solving Inverse Physics Problems with Score Matching
Benjamin Holzschuh (Technical University of Munich), Nils Thuerey (Technical University of Munich)
Score-based ModelContrastive LearningTime SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a method for inverse physical solving that combines score matching with differentiable physical simulation, capable of inferring the initial state from the final state of the system and sampling the posterior distribution.
Solving Linear Inverse Problems Provably via Posterior Sampling with Latent Diffusion Models
Litu Rout (University of Texas at Austin), Sanjay Shakkottai (University of Texas at Austin)
RestorationSuper ResolutionDiffusion modelImageStochastic Differential Equation
🎯 What it does: Using pre-trained Latent Diffusion Models to solve linear inverse problems (such as image inpainting, denoising, super-resolution, etc.)
Sorting with Predictions
Xingjian Bai (University of Oxford), Christian Coester (University of Oxford)
OptimizationComputational EfficiencyReinforcement LearningTabular
🎯 What it does: This paper explores solving sorting problems through learning-enhanced algorithms, designing two new algorithms that utilize position prediction and dirty comparison to improve sorting efficiency.
SoTTA: Robust Test-Time Adaptation on Noisy Data Streams
Taesik Gong (Nokia Bell Labs), Sung-Ju Lee (KAIST)
Domain AdaptationImage
🎯 What it does: This paper proposes a testing-time adaptive method for noisy samples (SoTTA), aimed at enhancing the model's robustness to noisy samples in real-world testing streams.
Sounding Bodies: Modeling 3D Spatial Sound of Humans Using Body Pose and Audio
Xudong XU, Alexander Richard (Meta Reality Labs Research)
GenerationData SynthesisPose EstimationMultimodalityAudio
🎯 What it does: Using audio captured by a head-mounted microphone and human joint posture information, a multimodal network was constructed that can generate a three-dimensional sound field covering the human body, enabling spatial audio rendering at any spatial location.
SPA: A Graph Spectral Alignment Perspective for Domain Adaptation
Zhiqing Xiao (Zhejiang University), Junbo Zhao (Zhejiang University)
Domain AdaptationGraph Neural NetworkImage
🎯 What it does: This paper addresses the problem of unsupervised domain adaptation and proposes a framework based on spectral alignment, which captures cross-domain transferability while enhancing intra-domain discriminability.
SPACE: Single-round Participant Amalgamation for Contribution Evaluation in Federated Learning
Yi-Chung Chen (National Taiwan University), Ming-Syan Chen (National Taiwan University)
Federated LearningImage
🎯 What it does: A single-round participant fusion evaluation method called SPACE is proposed for efficiently assessing the contributions of each client in federated learning.
SPAE: Semantic Pyramid AutoEncoder for Multimodal Generation with Frozen LLMs
Lijun Yu (Google), Lu Jiang (Google)
GenerationData SynthesisLarge Language ModelAuto EncoderImageMultimodality
🎯 What it does: Proposes the Semantic Pyramid AutoEncoder (SPAE), which maps images to a sequence of semantic vocabulary that can be understood by a frozen LLM, thereby achieving cross-modal reasoning for image understanding and generation.
Sparse Deep Learning for Time Series Data: Theory and Applications
Mingxuan Zhang (Purdue University), Faming Liang (Purdue University)
Recurrent Neural NetworkTime SeriesFinance Related
🎯 What it does: This paper establishes the Bayesian theory and empirical evidence of sparse deep learning (sparse RNN) on time series, proving posterior consistency, structural selection consistency, and asymptotic normality of predictions, and utilizes this theory to achieve efficient prediction uncertainty quantification and model compression.
Sparse Modular Activation for Efficient Sequence Modeling
Liliang Ren (University of Illinois at Urbana-Champaign), ChengXiang Zhai
OptimizationComputational EfficiencyTransformerTextSequentialAudio
🎯 What it does: Designed and implemented a differentiable Sparse Modular Activation (SMA) mechanism, constructing the SeqBoat network, which utilizes SMA for dynamic sparse activation state space models (SSM) and gated attention units (GAU) to achieve efficient sequence modeling;
Sparse Parameterization for Epitomic Dataset Distillation
Xing Wei (Xi'an Jiaotong University), Zhiheng Ma (Shenzhen Institute of Advanced Technology)
Data SynthesisKnowledge DistillationRecurrent Neural NetworkTransformerImage
🎯 What it does: A sparse parameterization-based Epitomic dataset distillation framework SPEED is proposed, capable of synthesizing information-rich synthetic data in extremely low storage space.
SparseProp: Efficient Event-Based Simulation and Training of Sparse Recurrent Spiking Neural Networks
Rainer Engelken (Columbia University)
Spiking Neural Network
🎯 What it does: An algorithm named SparseProp is proposed for efficient and accurate event-driven simulation and training of large-scale sparse spiking neural networks (SNNs).
Sparsity-Preserving Differentially Private Training of Large Embedding Models
Badih Ghazi (Google Research), Chiyuan Zhang (Google Research)
Safty and PrivacyComputational EfficiencyTransformerSupervised Fine-TuningTabularTime Series
🎯 What it does: Two algorithms, DP-FEST and DP-AdaFEST, are proposed, which utilize frequency filtering and an adaptive selection mechanism to inject noise only into the high-contribution rows of sparse gradients, thereby preserving the sparsity of the embedding layer gradients while ensuring differential privacy and significantly improving training efficiency.
Spatial-frequency channels, shape bias, and adversarial robustness
Ajay Subramanian (New York University), Denis G. Pelli (New York University)
ClassificationRecognitionAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and utilizes critical band masking experiments to compare the object recognition performance of humans and various deep networks under the influence of frequency domain noise, assessing their spatial frequency channel characteristics.
Spatially Resolved Gene Expression Prediction from Histology Images via Bi-modal Contrastive Learning
Ronald Xie (University of Toronto), Gary Bader
Representation LearningData-Centric LearningConvolutional Neural NetworkContrastive LearningImageBiomedical Data
🎯 What it does: A dual-modal contrastive learning-based embedding framework called BLEEP is proposed, which can predict the spatial resolution of gene expression from H&E tissue slice images.