NeurIPS 2024 Papers — Page 36
Conference on Neural Information Processing Systems · 4035 papers
Take A Shortcut Back: Mitigating the Gradient Vanishing for Training Spiking Neural Networks
Yufei Guo (Peking University), Zhe Ma (Peking University)
ClassificationSpiking Neural NetworkImage
🎯 What it does: A method is proposed to alleviate the gradient vanishing problem in training spiking neural networks (SNNs) by adding multiple shortcut branches in the network to directly pass gradients from the output to the shallow layers.
Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models
Jack Merullo (Brown University), Ellie Pavlick (Brown University)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By performing singular value decomposition (SVD) on the weights of GPT-2 and Pythia, low-rank communication channels are identified, and interventions are conducted to verify their causal impact on model behavior, further explaining prompt sensitivity and improving performance on list recall tasks.
TALoS: Enhancing Semantic Scene Completion via Test-time Adaptation on the Line of Sight
Hyun-Kurl Jang (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
SegmentationDomain AdaptationAutonomous DrivingContrastive LearningPoint Cloud
🎯 What it does: This paper proposes TALoS, which achieves test-time adaptation for SSC by utilizing line-of-sight information from multi-temporal LiDAR observations.
Taming "data-hungry" reinforcement learning? Stability in continuous state-action spaces
Yaqi Duan (New York University), Martin J Wainwright
Reinforcement LearningSequential
🎯 What it does: This paper proposes an analytical framework for continuous state-action space reinforcement learning and demonstrates faster convergence rates in both offline and online scenarios, specifically 1/n for sample error and log T for cumulative regret.
Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains
Lei Wang (University of Florida), Jie Xu (University of Florida)
ClassificationFederated LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: In the context of federated learning, the FedPLVM method is proposed, utilizing dual-layer prototype clustering (local and global) and α-sparse prototype loss to address the learning imbalance caused by feature variance between different domains.
Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs
Qinpeng Cui (Advanced Micro Devices), Emad Barsoum (Advanced Micro Devices)
RestorationSuper ResolutionDomain AdaptationDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes DoSSR, a domain transfer super-resolution framework based on a pre-trained diffusion model, which directly initiates the diffusion process from low-resolution images and achieves a smooth transition from low resolution to high resolution through a domain shift equation.
Taming Generative Diffusion Prior for Universal Blind Image Restoration
Siwei Tu (Fudan University), Ben Fei (Chinese University of Hong Kong)
RestorationDiffusion modelImage
🎯 What it does: This paper proposes a general framework for blind image restoration using diffusion models under unknown degradation conditions, called BIR-D.
Taming Heavy-Tailed Losses in Adversarial Bandits and the Best-of-Both-Worlds Setting
Duo Cheng (Virginia Tech), Bo Ji (Virginia Tech)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the multi-armed bandit problem under heavy-tailed losses, particularly in the Best of Both Worlds (BOBW) setting, proposing an algorithm and proving its pseudo-regret bounds in adversarial and stochastic environments.
Taming the Long Tail in Human Mobility Prediction
Xiaohang Xu (University of Tokyo), Kaoru Sezaki (University of Tokyo)
Recommendation SystemSafty and PrivacyGraph Neural NetworkTransformerGraph
🎯 What it does: The LoTNext framework is proposed to address the long-tail distribution problem in predicting the next location (POI) for humans, improving the prediction accuracy for low-visit-frequency places.
Tangent Space Causal Inference: Leveraging Vector Fields for Causal Discovery in Dynamical Systems
Kurt Butler (Stony Brook University), Petar Djuric
Time SeriesOrdinary Differential Equation
🎯 What it does: A new causal inference method is proposed - Tangent Space Causal Inference (TSCI), which detects causal relationships by comparing the tangent vector fields of dynamic system embedding spaces.
TAPTRv2: Attention-based Position Update Improves Tracking Any Point
Hongyang Li (South China University of Technology), Lei Zhang (South China University of Technology)
Object TrackingTransformerVideo
🎯 What it does: This paper proposes a Transformer-based TAPTRv2 framework for precise tracking of arbitrary points in videos, improving the cost-volume dependency issue of the previous TAPTR model, and achieving adaptive updates of point positions through Attention-Based Position Update (APU).
Target-Guided Adversarial Point Cloud Transformer Towards Recognition Against Real-world Corruptions
Jie Wang (Beijing Institute of Technology), Jianan Li (Chinese University of Hong Kong)
RecognitionAdversarial AttackTransformerPrompt EngineeringPoint Cloud
🎯 What it does: This paper proposes an adversarial Transformer model for 3D point cloud recognition, named APCT, aimed at enhancing robustness against real-world noise and corruptions.
Targeted Sequential Indirect Experiment Design
Elisabeth Ailer (Technical University of Munich), Niki Kilbertus (Technical University of Munich)
OptimizationTabular
🎯 What it does: This paper proposes an adaptive strategy for designing indirect experiments to optimally obtain information about the target queries regarding the true mechanisms, particularly in the presence of confounding factors and multidimensional nonlinearity.
TARP-VP: Towards Evaluation of Transferred Adversarial Robustness and Privacy on Label Mapping Visual Prompting Models
Zhen Chen (University of Liverpool), Wenjie Ruan (University of Science and Technology of China)
ClassificationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerPrompt EngineeringImage
🎯 What it does: This study investigates the security of the Label Mapping Visual Prompt (LM-VP) model, systematically assessing its robustness against transfer adversarial attacks and the privacy leakage risk from Membership Inference Attacks (MIA).
TARSS-Net: Temporal-Aware Radar Semantic Segmentation Network
Youcheng Zhang (Intelligent Science and Technology Academy of CASIC), Zhe Ma (Shenzhen International Graduate School Tsinghua University)
SegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: A new framework for radar semantic segmentation, TARSS-Net, is proposed, which utilizes the TRAM module for target historical association learning and aggregation of temporal information to enhance the performance of radar multi-view semantic segmentation.
Task Confusion and Catastrophic Forgetting in Class-Incremental Learning: A Mathematical Framework for Discriminative and Generative Modelings
Milad Khademi Nori (Toronto Metropolitan University), IL MIN KIM
ClassificationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A mathematical framework is proposed to formally define Task Confusion (TC) and Catastrophic Forgetting (CF) in class-incremental learning (class-IL), and through theorems proving that discriminative models inevitably suffer from TC while generative models can overcome TC, further evaluates and compares the theoretical advantages and disadvantages of existing IL strategies.
Task-Agnostic Machine-Learning-Assisted Inference
Jiacheng Miao (University of Wisconsin Madison), Qiongshi Lu (University of Wisconsin Madison)
TabularBiomedical Data
🎯 What it does: A task-agnostic machine learning-assisted inference framework, PSPS, is proposed, allowing any existing statistical inference method based on labeled data to combine unlabeled data and pre-trained ML models to achieve effective and more accurate inference results.
Task-oriented Time Series Imputation Evaluation via Generalized Representers
Zhixian Wang (University of Hong Kong), Yi Wang (Alibaba Group)
Anomaly DetectionOptimizationTime Series
🎯 What it does: This paper addresses the issue of imputing missing values in time series by proposing a task-oriented (using prediction tasks as an example) evaluation framework. This framework can estimate the impact of each imputed value on downstream tasks without retraining the prediction model, and it combines the advantages of different imputation methods based on this impact to achieve better imputation results.
Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning
Grzegorz Rypeść (Warsaw University of Technology), Bartłomiej Twardowski (Autonomous University of Barcelona)
ClassificationKnowledge DistillationGaussian SplattingImage
🎯 What it does: The AdaGauss method is proposed for category incremental learning without sample storage, dynamically adapting the Gaussian distribution for each task and alleviating dimensional collapse and task proximity bias through regularization.
Teach Better or Show Smarter? On Instructions and Exemplars in Automatic Prompt Optimization
Xingchen Wan (Google Cloud AI Research), Sercan O Arik
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper compares and systematically evaluates two types of automatic prompt optimization methods: Instruction Optimization (IO) and Example Optimization (EO), exploring their individual and synergistic effects in a multi-task environment.
Team-Fictitious Play for Reaching Team-Nash Equilibrium in Multi-team Games
Ahmed Said Dönmez, Muhammed O. Sayin (Bilkent University)
OptimizationReinforcement LearningGraphStochastic Differential Equation
🎯 What it does: A variant called Team-Fictitious Play (Team-FP) is proposed for learning team coordination in multi-team games and approaching Team Nash Equilibrium (TNE).
Tell What You Hear From What You See - Video to Audio Generation Through Text
Xiulong Liu (University of Washington), Eli Shlizerman (University of Washington)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodalityAudio
🎯 What it does: A multimodal generation framework VATT has been constructed, capable of text-guided video-to-audio generation and automatic subtitle generation from video to audio.
Template-free Articulated Gaussian Splatting for Real-time Reposable Dynamic View Synthesis
Diwen Wan (Peking University), Gang Zeng (Peking University)
GenerationData SynthesisPose EstimationGaussian SplattingVideo
🎯 What it does: Without using any templates or pose annotations, we automatically reconstruct re-poseable dynamic 3D objects from multi-view videos using 3D Gaussian splatting and superpoint techniques, achieving high-quality new view synthesis within a real-time rendering framework.
Temporal Graph Neural Tangent Kernel with Graphon-Guaranteed
Katherine Tieu (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: A graph neural tangent kernel (Temp-G NTK) is proposed that can measure the similarity of temporal evolution graphs without the need to train neural networks, and an upper bound on its generalization error for graph classification and node prediction tasks is provided.
Temporal Sentence Grounding with Relevance Feedback in Videos
Jianfeng Dong (Zhejiang Gongshang University), Meng Wang (Hefei University of Technology)
RetrievalRecurrent Neural NetworkVideoText
🎯 What it does: The Temporal Sentence Grounding with Relevance Feedback (TSG-RF) task is proposed, which can determine whether there are segments in a video related to a query sentence, and if so, provide precise time boundaries; otherwise, it gives feedback of 'no relevant segments'.
Temporal-Difference Learning Using Distributed Error Signals
Jonas Guan (University of Toronto), William A Cunningham
Reinforcement LearningSequential
🎯 What it does: A new deep Q-learning algorithm called ARTIFICIAL DOPAMINE (AD) is designed, which learns using only distributed hierarchical TD errors without the need for backpropagation.
Temporally Consistent Atmospheric Turbulence Mitigation with Neural Representations
Haoming Cai (University of Maryland), Christopher Metzler
RestorationOptical FlowVideo
🎯 What it does: The ConVRT framework is proposed, achieving temporal consistency recovery in video atmospheric turbulence suppression through neural video representation.
Tensor-Based Synchronization and the Low-Rankness of the Block Trifocal Tensor
Daniel Miao (University of Minnesota), Joe Kileel (University of Texas at Austin)
Pose EstimationSimultaneous Localization and MappingImage
🎯 What it does: This study investigates and utilizes the block trifocal tensor for camera pose synchronization through three-view geometry. By constructing its Tucker decomposition, it proves that its multilinear rank is (6,4,4), allowing for the unique recovery of camera position and pose in the absence of noise.
Test Where Decisions Matter: Importance-driven Testing for Deep Reinforcement Learning
Stefan Pranger (Graz University of Technology), Bettina Könighofer (Graz University of Technology)
Safty and PrivacyReinforcement LearningSequential
🎯 What it does: This paper proposes a model-based deep reinforcement learning safety testing framework that uses state importance ranking to guide test point selection and achieves formal safety guarantees over the complete state space through iterative bounded MDP.
Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment
Zhen-Yu Zhang (Center for Advanced Intelligence Project RIKEN), Masashi Sugiyama (Center for Advanced Intelligence Project RIKEN)
Domain AdaptationMeta LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a testing-time adaptive method called Ada-ReAlign in non-stationary environments, which utilizes source domain sketches to achieve dynamic representation alignment for unlabeled data streams, thereby enabling continuous model adaptation.
Test-Time Adaptation Induces Stronger Accuracy and Agreement-on-the-Line
Eungyeup Kim (Carnegie Mellon University), J Zico Kolter
Domain AdaptationImage
🎯 What it does: This study investigates the impact of Test-Time Adaptation (TTA) on model accuracy and Agreement-on-the-Line (ACL/AGL) under different distribution shifts.
Test-Time Dynamic Image Fusion
Bing Cao (Tianjin University), Qinghua Hu (Tianjin University)
RestorationConvolutional Neural NetworkImageMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a Test-Time Dynamic Image Fusion (TTD) method, which significantly improves fusion quality by dynamically fusing multi-source images based on pixel-level relative dominance (RD) weights derived from reconstruction loss.
Testably Learning Polynomial Threshold Functions
Lucas Slot (ETH Zurich), Manuel Wiedmer (ETH Zurich)
🎯 What it does: A testable learning algorithm for polynomial threshold functions (PTFs) is proposed over the standard Gaussian distribution, and it is proven that learning can be achieved in time n poly(1/ε) for any fixed order;
Testing Calibration in Nearly-Linear Time
Lunjia Hu (Harvard University), Chutong Yang (University of Texas at Austin)
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper studies the calibration detection problem of binary classification prediction models and proposes a near-linear time algorithm based on LDTC and smoothed calibration error.
Testing Semantic Importance via Betting
Jacopo Teneggi (Johns Hopkins University), Jeremias Sulam (Johns Hopkins University)
Explainability and InterpretabilityVision Language ModelImage
🎯 What it does: A semantic concept importance testing framework for black-box models (C-SKIT and X-SKIT) was constructed using the theory of conditional independence and sequential betting methods, achieving global and local importance ranking along with control of false discovery rates.
Tetrahedron Splatting for 3D Generation
Chun Gu (Fudan University), Li Zhang (Fudan University)
GenerationData SynthesisOptimizationDiffusion modelGaussian SplattingMesh
🎯 What it does: A new 3D representation method called Tet-Splatting is proposed, which combines voxel rendering with tetrahedral meshes to achieve precise mesh extraction and real-time rendering.
Text-Aware Diffusion for Policy Learning
Calvin Luo (Brown University), Chen Sun (Brown University)
Robotic IntelligenceReinforcement LearningDiffusion modelVideoText
🎯 What it does: Using a pre-trained text-conditioned diffusion model to provide zero-shot text-conditioned rewards for reinforcement learning policies, learning the mapping from natural language descriptions to policies.
Text-DiFuse: An Interactive Multi-Modal Image Fusion Framework based on Text-modulated Diffusion Model
Hao Zhang (Wuhan University), Jiayi Ma
Image TranslationRestorationDiffusion modelImageMultimodality
🎯 What it does: This paper proposes an interactive multimodal image fusion framework called Text-DiFuse based on a text-controlled diffusion model, which can perform denoising, color cast removal, and multimodal information fusion in the presence of composite distortions (such as noise, color cast, underexposure, etc.), while also supporting the emphasis of targets of interest through natural language instructions.
Text-Guided Attention is All You Need for Zero-Shot Robustness in Vision-Language Models
Lu Yu (Tianjin University of Technology), Changsheng Xu (Institute of Automation, University of Chinese Academy of Sciences)
Explainability and InterpretabilityAdversarial AttackTransformerVision Language ModelImageMultimodality
🎯 What it does: A new framework TGA-ZSR is proposed, which enhances the zero-shot robustness of visual-language models through a text-guided attention mechanism while maintaining performance on clean samples.
Text-Infused Attention and Foreground-Aware Modeling for Zero-Shot Temporal Action Detection
Yearang Lee (Korea University), Seong-Whan Lee (Korea University)
RecognitionObject DetectionTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: A zero-shot temporal action detection method that integrates text and video information is proposed to address the common sub-action bias problem.
Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts
Mohammad Sadil Khan (DFKI), Muhammad Zeshan Afzal (DFKI)
GenerationTransformerVision Language ModelTextSequential
🎯 What it does: Proposes the Text2CAD framework, which automatically generates parametric CAD models using text prompts, supporting different levels of description from beginners to experts;
Text2NKG: Fine-Grained N-ary Relation Extraction for N-ary relational Knowledge Graph Construction
Haoran Luo (Beijing University of Posts and Telecommunications), Anh Tuan Luu (Nanyang Technological University)
ClassificationComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: Proposes the Text2NKG framework for fine-grained n-ary relation extraction from natural language text, and uses the extraction results to construct four common types of NKGs (hyper-relational, event-based, role-based, hypergraph-based).
TextCtrl: Diffusion-based Scene Text Editing with Prior Guidance Control
Weichao Zeng (Institute of Information Engineering, Chinese Academy of Sciences), Yu Zhou (Nankai University)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes a scene text editing method called TextCtrl based on diffusion models, achieving high-fidelity editing through prior style decoupling and character structure representation.
Textual Training for the Hassle-Free Removal of Unwanted Visual Data: Case Studies on OOD and Hateful Image Detection
Saehyung Lee (Seoul National University), Sungroh Yoon (Seoul National University)
ClassificationAnomaly DetectionTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: A framework named HFTT (Hassle‑Free Textual Training) is proposed, which trains a classifier to detect undesirable visual content (such as OOD samples and hate images) using only text data and a pre-trained vision-language model.
TFG: Unified Training-Free Guidance for Diffusion Models
Haotian Ye (Stanford University), Stefano Ermon (Peking University)
GenerationData SynthesisOptimizationHyperparameter SearchDiffusion modelImageBenchmarkAudio
🎯 What it does: This paper proposes a unified training-free guidance framework (TFG) for achieving conditional generation on diffusion models.
TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering
Jun Dan (Zhejiang University), Yanchao Tan (Fuzhou University)
Domain AdaptationGraph Neural NetworkGraph
🎯 What it does: This paper proposes the TFGDA framework for semi-supervised graph domain adaptation tasks;
TFS-NeRF: Template-Free NeRF for Semantic 3D Reconstruction of Dynamic Scene
Sandika Biswas (Monash University), Hamid Rezatofighi (Monash University)
Object DetectionSegmentationPose EstimationNeural Radiance FieldVideo
🎯 What it does: This paper proposes a template-free dynamic scene neural radiance field TFS-NeRF for achieving semantically separated 3D reconstruction from sparse single-view RGB videos, capable of handling two interacting rigid, non-rigid, or deformable entities.
The ALCHEmist: Automated Labeling 500x CHEaper than LLM Data Annotators
Tzu-Heng Huang (University of Wisconsin Madison), Frederic Sala (University of Wisconsin Madison)
Data-Centric LearningTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: The Alchemist system is proposed, which utilizes large language models to generate executable annotation programs, replacing the implementation of data annotation through individual API calls.
The Bayesian sampling in a canonical recurrent circuit with a diversity of inhibitory interneurons
Eryn Sale (UT Southwestern Medical Center), Wenhao Zhang
Stochastic Differential Equation
🎯 What it does: A nonlinear loop model containing two types of inhibitory neurons, PV and SOM, was constructed, theoretically deriving its Bayesian sampling in the stimulus feature subspace and verifying its functionality through numerical simulations.
The Benefits of Balance: From Information Projections to Variance Reduction
Lang Liu (University of Washington), Zaid Harchaoui (University of Washington)
ClassificationRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper explains and utilizes multi-source data balancing in self-supervised learning through balancing multimodal data (e.g., Sinkhorn iteration), demonstrating that this process can significantly reduce the variance of target estimation and providing a non-asymptotic error upper bound.
The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection
Qingyang Zhang (Tianjin University), Changqing Zhang (Tianjin University)
ClassificationAnomaly DetectionImage
🎯 What it does: A new Decoupled Uncertainty Learning (DUL) framework is proposed to address the contradiction in existing OOD detection methods, which improve detection performance while causing a decline in OOD generalization performance.
The Challenges of the Nonlinear Regime for Physics-Informed Neural Networks
Andrea Bonfanti (Basque Center for Applied Mathematics University of the Basque Country), Cristina Cipriani (Technical University of Munich)
OptimizationComputational EfficiencyTabularPhysics Related
🎯 What it does: This study investigates the training dynamics of Physics-Informed Neural Networks (PINNs) under nonlinear partial differential equations (PDEs), revealing the randomness and dynamics of the Neural Tangent Kernel (NTK) in nonlinear scenarios, and proving that second-order optimization methods can alleviate spectral bias and accelerate convergence.
The Closeness of In-Context Learning and Weight Shifting for Softmax Regression
Shuai Li (Shanghai Jiao Tong University), Tianyi Zhou (University of Southern California)
OptimizationTransformerText
🎯 What it does: This paper analyzes the role of the self-attention layer in the Transformer in context learning from the perspective of softmax regression, providing an upper bound on the data transformation of single-layer self-attention and gradient descent updates, proving that the behavior of both is highly similar when training softmax regression tasks.
The Collusion of Memory and Nonlinearity in Stochastic Approximation With Constant Stepsize
Dongyan Lucy Huo (Cornell University), Qiaomin Xie (University of Wisconsin Madison)
OptimizationTabularStochastic Differential Equation
🎯 What it does: This study explores the stochastic approximation (SA) method with Markov data and nonlinear updates under a constant step size, analyzing the interaction between the Markov dependence of the data and the nonlinear update rules.
The Dormant Neuron Phenomenon in Multi-Agent Reinforcement Learning Value Factorization
Haoyuan Qin (Xiamen University), Siqi Shen (Xiamen University)
Reinforcement Learning
🎯 What it does: This paper studies the phenomenon of 'dormant neurons' in value decomposition methods within multi-agent reinforcement learning and proposes a new parameter perturbation method called ReBorn.
The Edge-of-Reach Problem in Offline Model-Based Reinforcement Learning
Anya Sims (University of Oxford), Yee Whye Teh (University of Oxford)
Reinforcement Learning
🎯 What it does: This study investigates the 'edge-of-reach' problem caused by model rollouts truncation in offline model-based reinforcement learning and proposes a RAVL method that directly applies value pessimistic penalties to these edge states.
The Empirical Impact of Neural Parameter Symmetries, or Lack Thereof
Derek Lim (Massachusetts Institute of Technology), Stefanie Jegelka (Technical University of Munich)
OptimizationMeta LearningGraph Neural NetworkImageGraph
🎯 What it does: This study proposes two structures to eliminate the symmetry in the parameter space of deep networks and experimentally evaluates their impact on phenomena such as linear mode connectivity, Bayesian inference, meta-networks, and monotonic linear interpolation.
The Evolution of Statistical Induction Heads: In-Context Learning Markov Chains
Ezra Edelman (University of Pennsylvania), Surbhi Goel (University of Pennsylvania)
TransformerSequential
🎯 What it does: This study investigates the learning dynamics of the Transformer in in-context learning tasks based on random Markov chains, revealing the process by which the model gradually achieves statistical inference of pairs through a 'statistical induction head'.
The Expressive Capacity of State Space Models: A Formal Language Perspective
Yash Sarrof (Saarland University), Michael Hahn (Saarland University)
Recurrent Neural NetworkTransformerSequential
🎯 What it does: The theoretical analysis of the expressive power of linear state space models (SSM) is conducted, comparing it with Transformers and RNNs, and studying its representability in regular languages, counting languages, and bounded hierarchical structures.
The Factorization Curse: Which Tokens You Predict Underlie the Reversal Curse and More
Ouail Kitouni (Massachusetts Institute of Technology), Mark Ibrahim (Meta)
RetrievalTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: Proposes the 'Factorization Curse' framework to study the knowledge retrieval failure of language models under different factorizations, and experimentally verifies that the factorization-agnostic objective can alleviate the reversal curse.
The Fairness-Quality Tradeoff in Clustering
Rashida Hakim (Columbia University), Mihalis Yannakakis (Columbia University)
OptimizationTabular
🎯 What it does: This paper proposes a general algorithm for calculating the Pareto front between quality (cost) and fairness in clustering problems.
The Feature Speed Formula: a flexible approach to scale hyper-parameters of deep neural networks
Lénaïc Chizat (École Polytechnique Fédérale de Lausanne), Praneeth Netrapalli (Google DeepMind)
OptimizationHyperparameter Search
🎯 What it does: A feature speed formula is proposed, and based on this, the feature learning rate of deep networks is analyzed, leading to the design of an automated hyperparameter scaling method.
The Fine-Grained Complexity of Gradient Computation for Training Large Language Models
Josh Alman (Columbia University), Zhao Song (University of California)
OptimizationComputational EfficiencyLarge Language ModelText
🎯 What it does: Analyzed the fine-grained complexity of forward and backward computations in the training of large language models, and provided a feasibility boundary with B≈√log n as the threshold.
The GAN is dead; long live the GAN! A Modern GAN Baseline
Nick Huang, James Tompkin (Brown University)
GenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper studies a more stable GAN training method and builds a simplified baseline R3GAN that does not require traditional techniques.
The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations
Tyler LaBonte (Georgia Institute of Technology), Abhishek Kumar (Georgia Institute of Technology)
ClassificationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageTextBenchmark
🎯 What it does: Conducted detailed experiments on the fine-tuning process of deep models in the presence of spurious correlations, studying the effects of class balance, model size, and spectral imbalance on the worst-group accuracy, and proposed a mixed balance method.
The High Line: Exact Risk and Learning Rate Curves of Stochastic Adaptive Learning Rate Algorithms
Elizabeth Collins-Woodfin (McGill University), Courtney Paquette (Google DeepMind)
OptimizationTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper develops a theoretical framework for high-dimensional optimization problems, providing precise analytical expressions for the risk and learning rate curves using a single pass of Stochastic Gradient Descent (SGD) with adaptive learning rates, and conducts an in-depth analysis of idealized Exact Line Search and AdaGrad-Norm algorithms on least squares problems.
The Impact of Geometric Complexity on Neural Collapse in Transfer Learning
Michael Munn (Google Research), Javier Gonzalvo (Google Research)
ClassificationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper theoretically and experimentally studies the impact of geometric complexity (GC) on neural collapse (NC) during the pre-training phase of neural networks, and further explores the decisive role of GC in transfer learning, especially in few-shot transfer performance.
The Impact of Initialization on LoRA Finetuning Dynamics
Soufiane Hayou (University of California Berkeley), Bin Yu (University of California Berkeley)
TransformerSupervised Fine-TuningText
🎯 What it does: Study the impact of two random initialization schemes in LoRA fine-tuning on training dynamics and performance.
The Implicit Bias of Adam on Separable Data
Chenyang Zhang (University of Hong Kong), Yuan Cao (University of Hong Kong)
OptimizationTabular
🎯 What it does: This study investigates and proves the implicit bias of Adam on linearly separable data, showing that its iterations tend towards the direction of the maximum ℓ∞-margin, and that the convergence speed under general learning rates is polynomial-level.
The Implicit Bias of Gradient Descent on Separable Multiclass Data
Hrithik Ravi (University of Michigan), Yutong Wang (Illinois Institute of Technology)
OptimizationTabular
🎯 What it does: This paper studies the implicit bias when using gradient descent on linearly separable multi-class data, proving that losses satisfying the exponential tail property under the PERM loss framework lead the model to converge towards the hard-margin SVM direction.
The Implicit Bias of Gradient Descent toward Collaboration between Layers: A Dynamic Analysis of Multilayer Perceptions
Zheng Wang (University of Exeter), Wenjie Ruan (University of Science and Technology of China)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study investigates the implicit bias of gradient descent on inter-layer cooperation regarding adversarial robustness, proposing a correlation measure to assess inter-layer cooperation.
The Implicit Bias of Heterogeneity towards Invariance: A Study of Multi-Environment Matrix Sensing
Yang Xu (Peking University), Cong Fang (Peking University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the implicit bias of standard stochastic gradient descent (SGD) in over-parameterized models under heterogeneous data across multiple environments, proving that it can automatically learn invariant signals across environments and suppress environment-related spurious signals.
The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains
Eric Qu (University of California Berkeley), Aditi S. Krishnapriyan (University of California Berkeley)
Graph Neural NetworkTransformerTabular
🎯 What it does: This study investigates the scalability of neural network interatomic potentials (NNIP) and proposes an efficient scalable attention interatomic potential (EScAIP) model, which is evaluated on multiple chemical datasets.
The Importance of Online Data: Understanding Preference Fine-tuning via Coverage
Yuda Song (Carnegie Mellon University), Wen Sun (Cornell University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: By comparing online RLHF and offline comparative preference fine-tuning methods, a hybrid optimization algorithm HyPO is proposed, and a theoretical analysis of the convergence conditions of both methods is conducted from the perspective of coverage, followed by experimental validation on multiple benchmarks.
The Intelligible and Effective Graph Neural Additive Network
Maya Bechler-Speicher (Tel Aviv University), Ran Gilad-Bachrach (Tel Aviv University)
ClassificationExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This paper designs and evaluates an interpretable graph neural network GNAN, which achieves global visual explanations and predictions for graph data based on an additivity model;
The Iterative Optimal Brain Surgeon: Faster Sparse Recovery by Leveraging Second-Order Information
Diyuan Wu (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
CompressionOptimizationTransformerImageText
🎯 What it does: The Iterative Optimal Brain Surgeon (I-OBS) algorithm is proposed, which introduces second-order information into Iterative Hard Thresholding (IHT) to achieve theoretical analysis and practical implementation of sparse recovery and model compression.
The Ladder in Chaos: Improving Policy Learning by Harnessing the Parameter Evolving Path in A Low-dimensional Space
Hongyao Tang (Tianjin University), Jianye HAO
Reinforcement LearningTime Series
🎯 What it does: This study investigates the learning paths of the policy networks of typical deep reinforcement learning agents, discovering that they evolve in low-dimensional subspaces. The Policy Path Trimming and Boosting (PPTB) method is proposed to trim minor directions and reinforce the main direction, significantly improving the learning performance of TD3, RAD, and DoubleDQN.
The Limits of Differential Privacy in Online Learning
Bo Li (Hong Kong University of Science and Technology), Peng Ye (Hong Kong University of Science and Technology)
OptimizationSafty and Privacy
🎯 What it does: This paper studies the fundamental limitations of differential privacy in online learning, proving that pure DP is not learnable for point functions under adaptive adversaries, while approximate DP is learnable; it provides a lower bound on the error of any private learner of Ω(log T), which is further improved to Ω(LD(H) log T); it also presents an online learning algorithm under pure DP against memoryless adversaries, revealing the differences in learnability between pure DP and approximate DP.
The Limits of Transfer Reinforcement Learning with Latent Low-rank Structure
Tyler Sam (Cornell University), Christina Yu
Reinforcement Learning
🎯 What it does: This paper proposes a method to enhance the learning efficiency of the target MDP by utilizing the low-rank representation of the source MDP in transfer reinforcement learning.
The Mamba in the Llama: Distilling and Accelerating Hybrid Models
Junxiong Wang (Cornell University), Tri Dao (Princeton University)
OptimizationComputational EfficiencyKnowledge DistillationRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Distill large Transformer models (such as Llama-3 8B Instruct) into a linear RNN (Mamba) mixture model through weight mapping and a stepwise mixing structure, and based on this, propose hardware-aware multi-step speculative decoding to accelerate inference.
The Many Faces of Optimal Weak-to-Strong Learning
Mikael Møller Høgsgaard (Aarhus University), Markus Engelund Mathiasen (Aarhus University)
ClassificationOptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes and theoretically analyzes a weak-to-strong learning algorithm called MAJORITY-OF-5, which divides the training set into 5 parts and trains each part using AdaBoost, then takes a majority vote to obtain the final classifier. The sample complexity of this method is also proven; subsequently, experiments are conducted comparing it with existing optimal weak-to-strong learners on five datasets.
The Map Equation Goes Neural: Mapping Network Flows with Graph Neural Networks
Christopher Blöcker (University of Zurich), Ingo Scholtes (Julius-Maximilians-Universität Würzburg)
OptimizationGraph Neural NetworkGraph
🎯 What it does: Rewrite the objective function of the Map Equation in information theory into a differentiable tensor form, making it a loss function that can be directly used by Graph Neural Networks (GNNs), thus achieving end-to-end unsupervised community detection.
The Minimax Rate of HSIC Estimation for Translation-Invariant Kernels
Florian Kalinke (Karlsruhe Institute of Technology), Zoltán Szabó (London School of Economics)
🎯 What it does: This study investigates the lower bound of the estimation of Hilbert-Schmidt Independence Criterion (HSIC) under continuous bounded translation-invariant kernels and proves that its optimal convergence rate is $O(n^{-1/2})$, thereby confirming the optimality of existing U-statistics, V-statistics, and Nyström estimators in this context.
The motion planning neural circuit in goal-directed navigation as Lie group operator search
Junfeng Zuo (Peking University), Wenhao Zhang
OptimizationRobotic Intelligence
🎯 What it does: This paper views motion planning as a Lie group operator search problem and derives a two-layer feedforward neural circuit suitable for one-dimensional rotation groups. This circuit is similar to the target navigation circuit of fruit flies and further constructs a complete perception-movement loop model.
The Poisson Midpoint Method for Langevin Dynamics: Provably Efficient Discretization for Diffusion Models
Saravanan Kandasamy (Cornell University), Dheeraj Mysore Nagaraj
GenerationData SynthesisComputational EfficiencyDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The Poisson Midpoint Method (PMM) is proposed to more efficiently discretize Langevin dynamics and is applied to the inference of diffusion models for image generation.
The Power of Extrapolation in Federated Learning
Hanmin Li (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated Learning
🎯 What it does: This paper proposes and studies a server-side extrapolation strategy to enhance the theoretical and empirical convergence performance of the FedProx algorithm in Federated Learning, introducing Extrapolated FedProx (FedExProx) and two adaptive extrapolation rules.
The Power of Hard Attention Transformers on Data Sequences: A formal language theoretic perspective
Pascal Bergsträßer (RPTU Kaiserslautern-Landau), Georg Zetzsche (MPI-SWS)
TransformerSequential
🎯 What it does: This paper studies the expressive power of the Unique Hard Attention Transformer (UHAT) on numerical sequences (data sequences);
The Power of Resets in Online Reinforcement Learning
Zakaria Mhammedi (Google Research), Alexander Rakhlin (Massachusetts Institute of Technology)
Reinforcement Learning
🎯 What it does: This paper studies the impact of using local simulators (resettable) on learning efficiency in online reinforcement learning, proposing two types of algorithms—SimGolf (a combination of global optimism and local simulators) and RVFS (Recursive Value Function Search). It demonstrates that they can achieve sample efficiency and computational efficiency in high-dimensional nonlinear function approximation environments such as low coverability MDPs and Exogenous Block MDPs.
The Prevalence of Neural Collapse in Neural Multivariate Regression
George Andriopoulos (New York University), Keith W. Ross
Autonomous DrivingOptimizationRobotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkTabular
🎯 What it does: This paper studies and proves that in multi-variable regression tasks, neural networks exhibit a geometric collapse of three features—feature vectors, weight vectors, and covariance matrices (NRC).
The Price of Implicit Bias in Adversarially Robust Generalization
Nikolaos Tsilivis (New York University), Julia Kempe (Meta)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Analyzed the implicit bias in the optimization process of Robust Empirical Risk Minimization (Robust ERM) and studied its impact on robust generalization.
The Reliability of OKRidge Method in Solving Sparse Ridge Regression Problems
Xiyuan Li (Wuhan University), Weiwei Liu (Wuhan University)
OptimizationTabular
🎯 What it does: This paper conducts a high-dimensional theoretical analysis of the error in the OKRidge algorithm for sparse ridge regression problems, utilizing the CGMT framework to convert estimation errors into PO/AO problems and solve them, resulting in the limit expression of NSE;
The Representation Landscape of Few-Shot Learning and Fine-Tuning in Large Language Models
Diego Doimo (Area Science Park), Alberto Cazzaniga (Area Science Park)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Compared the probability distributions of internal representations of large language models under two methods: in-context learning (ICL) with few examples and supervised fine-tuning (SFT), and analyzed their geometric structures through density peak clustering.
The Road Less Scheduled
Aaron Defazio (Meta), Ashok Cutkosky (Boston University)
OptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageTextMagnetic Resonance ImagingBenchmark
🎯 What it does: An optimization method that does not require a preset learning rate schedule is proposed—Schedule-Free learning;
The Sample Complexity of Gradient Descent in Stochastic Convex Optimization
Roi Livni (Tel Aviv University)
Optimization
🎯 What it does: This paper analyzes the sample complexity of full-batch gradient descent (GD) in non-smooth stochastic convex optimization (SCO), proving that both the upper and lower bounds reach Θ(d/m + 1/√m), which is comparable to the worst-case empirical risk minimization (ERM).
The Sample-Communication Complexity Trade-off in Federated Q-Learning
Sudeep Salgia (Carnegie Mellon University), Yuejie Chi (Carnegie Mellon University)
Federated LearningReinforcement LearningTabular
🎯 What it does: This paper studies the trade-off between sample complexity and communication complexity in federated Q-learning and proposes the Fed-DVR-Q algorithm, which can achieve optimal sample and communication complexity simultaneously.
The Secretary Problem with Predicted Additive Gap
Alexander Braun (University of Bonn), Sherry Sarkar (Carnegie Mellon University)
OptimizationTabular
🎯 What it does: This paper proposes an algorithm for decision-making in the secretary problem with randomly permuted sequences, utilizing a single additive gap (the difference between the highest weight and the k-th highest weight). Three variants are introduced: the precise gap algorithm, the robust-consistency balance algorithm, and the error-bounded algorithm.
The Selective $G$-Bispectrum and its Inversion: Applications to $G$-Invariant Networks
Simon Mataigne (UCLouvain), Nina Miolane (UC Santa Barbara)
Computational EfficiencyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: The research implemented the computation and inverse transformation of selective G-Bispectrum, significantly reducing the time/space complexity of traditional G-Bispectrum, and applied it as a complete G-invariant layer in G-CNN.
The Space Complexity of Approximating Logistic Loss
Gregory Dexter (LinkedIn Corporation), Rajiv Khanna (Purdue University)
OptimizationTabular
🎯 What it does: The paper studies the space complexity of data structures when compressing/approximating logistic regression loss, providing a lower bound of Ω(d/ϵ²) under constant complexity, and a lower bound of Ω(d·µ_y(X)) under arbitrary complexity; it also proposes an efficient algorithm for implementing µ_y(X) using linear programming and compares it with existing sketch-based estimation methods.
The Star Geometry of Critic-Based Regularizer Learning
Oscar Leong (University of California), Yong Sheng Soh (National University of Singapore)
RestorationOptimizationImage
🎯 What it does: This paper studies the star geometric structure in reviewer-based regularization learning and proposes a new method to optimize the reviewer loss function to learn specific types of regularizers, particularly the metrics of star-shaped bodies.
The Surprising Effectiveness of SP Voting with Partial Preferences
Hadi Hosseini (Penn State University), Amrit Puhan (Penn State University)
Tabular
🎯 What it does: A scalable voting algorithm based on the idea of 'surprising popularity' is proposed, which recovers the true ranking using local preferences instead of complete rankings;