NeurIPS 2024 Papers — Page 37
Conference on Neural Information Processing Systems · 4035 papers
The surprising efficiency of temporal difference learning for rare event prediction
Xiaoou Cheng (New York University), Jonathan Weare (New York University)
Computational EfficiencyReinforcement LearningMultimodalityTime Series
🎯 What it does: In the prediction of rare events in finite state Markov chains, a systematic comparison is made between Least-Squares Temporal Difference Learning (LSTD) and traditional Monte Carlo (MC) estimators, and the central limit theorem and relative variance upper bound for LSTD are provided.
The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning
Moritz Schneider (Bosch Center for Artificial Intelligence), Joschka Boedecker (University of Freiburg)
TransformerReinforcement LearningContrastive LearningImage
🎯 What it does: A systematic evaluation of the effectiveness of pre-trained visual representations (PVR) in model-based reinforcement learning.
The tree autoencoder model, with application to hierarchical data visualization
Miguel Á. Carreira-Perpiñán (University of California), Kuat Gazizov (University of California)
OptimizationRepresentation LearningAuto EncoderImageText
🎯 What it does: A tree-structured autoencoder called PCA Tree is proposed for hierarchical dimensionality reduction in two-dimensional visualization.
The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better
Scott Geng (University of Washington), Ranjay Krishna (Allen Institute for AI)
ClassificationRetrievalTransformerSupervised Fine-TuningDiffusion modelImageRetrieval-Augmented Generation
🎯 What it does: This paper studies the comparison of the effects of fine-tuning models using synthetic images generated by generative models versus real images retrieved directly from upstream datasets in visual classification tasks, demonstrating that retrieval methods often enhance model performance more effectively.
The Value of Reward Lookahead in Reinforcement Learning
Nadav Merlis (ENSAE Paris), Vianney Perchet (Criteo AI Lab)
Reinforcement Learning
🎯 What it does: This study investigates the value of observing future rewards in reinforcement learning, quantifying the competitive ratio between non-foresight agents and multi-step foresight agents.
Theoretical Analysis of Weak-to-Strong Generalization
Hunter Lang (Massachusetts Institute of Technology), Aravindan Vijayaraghavan (Northwestern University)
ClassificationData-Centric LearningText
🎯 What it does: This study investigates the generalization mechanism from weak teachers to strong student models in weakly supervised learning, proving the theoretical foundations of pseudo-label correction and coverage expansion effects.
Theoretical and Empirical Insights into the Origins of Degree Bias in Graph Neural Networks
Arjun Subramonian (University of California), Yizhou Sun (University of California)
ClassificationGraph Neural NetworkGraph
🎯 What it does: Theoretical and empirical analysis of degree bias in node classification tasks using graph neural networks, proving that high-degree nodes have a lower probability of misclassification and exploring the reasons behind this.
Theoretical Characterisation of the Gauss Newton Conditioning in Neural Networks
Jim Zhao (University of Basel), Aurelien Lucchi (University of Basel)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper provides a theoretical analysis of the condition number of the Gauss-Newton matrix in neural networks, offering tight upper bounds for linear networks, residual networks, and convolutional networks of arbitrary depth and width, and extends to nonlinear activations in a single LeakyReLU network.
Theoretical Foundations of Deep Selective State-Space Models
Nicola Muca Cirone (Imperial College London), Terry Lyons (University of Oxford)
Recurrent Neural NetworkTransformerSequentialBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies the Linear Control Differential Equation (CDE) model and provides a theoretical foundation for its expressiveness in sequential tasks, explaining the advantages of modern selectable state space models (SSMs) such as Mamba.
Theoretical guarantees in KL for Diffusion Flow Matching
Marta Gentiloni Silveri (Ecole Polytechnique), Giovanni Conforti (Universita degli Studi di Padova)
Diffusion modelStochastic Differential Equation
🎯 What it does: This paper studies the bridging of target distribution and benchmark distribution through Brownian bridges in finite time, namely Diffusion Flow Matching (DFM), and provides non-asymptotic theoretical guarantees for KL divergence.
Theoretical Investigations and Practical Enhancements on Tail Task Risk Minimization in Meta Learning
Yiqin Lv (National University of Defense Technology), Zheng Xie (National University of Defense Technology)
OptimizationMeta LearningTabular
🎯 What it does: This study investigates the theory and practice of tail task risk minimization in two-stage distributionally robust meta-learning, providing Stackelberg equilibrium, convergence rates, and generalization bounds, and improving the quantification of VaR/CVaR through kernel density estimation.
Thinking Forward: Memory-Efficient Federated Finetuning of Language Models
Kunjal Panchal (University of Massachusetts), Hui Guan (University of Massachusetts)
Federated LearningComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method for low-memory fine-tuning of LLMs in a federated learning environment by splitting trainable weights across clients and using forward-mode automatic differentiation.
This Too Shall Pass: Removing Stale Observations in Dynamic Bayesian Optimization
Anthony Bardou (Ecole polytechnique fédérale de Lausanne), Giovanni Ranieri (Ecole polytechnique fédérale de Lausanne)
OptimizationTabularTime SeriesBenchmark
🎯 What it does: A dynamic Bayesian optimization (DBO) algorithm W-DBO is proposed, which can online identify and remove observation points that are irrelevant to future predictions.
Thompson Sampling For Combinatorial Bandits: Polynomial Regret and Mismatched Sampling Paradox
Raymond Zhang (Universite Paris Saclay), Richard Combes (Universite Paris Saclay)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes an improved Thompson Sampling (BG-CTS) algorithm for the linear combination of semi-bandit problems and provides its polynomial upper bound non-asymptotic convergence rate.
Thought of Search: Planning with Language Models Through The Lens of Efficiency
Michael Katz (IBM T. J. Watson Research Center), Shirin Sohrabi (IBM T. J. Watson Research Center)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the use of large language models (LLMs) to generate key components of search algorithms (successor functions and goal tests), thereby achieving complete, correct, and efficient planning searches.
Tight Bounds for Learning RUMs from Small Slates
Flavio Chierichetti (Sapienza University of Rome), Andrew Tomkins (Google)
🎯 What it does: This paper studies how to learn Random Utility Models (RUM) using only small subsets (slates) of maximum size k, and proves that k=Θ(√n) is both necessary and sufficient for achieving a uniform approximation of the winning distribution for all subsets;
Tight Rates for Bandit Control Beyond Quadratics
Y. Jennifer Sun (Princeton University), Zhou Lu (Princeton University)
Autonomous DrivingOptimizationReinforcement Learning
🎯 What it does: An algorithm is proposed that can achieve an optimal regret bound of ˜ ( O √ T ) for bandit non-stochastic control problems with strongly convex smooth cost functions in the presence of adversarial perturbations.
Tighter Convergence Bounds for Shuffled SGD via Primal-Dual Perspective
Xufeng Cai (University of Wisconsin Madison), Jelena Diakonikolas (University of Wisconsin Madison)
OptimizationTabular
🎯 What it does: This paper studies the convergence speed of shuffled stochastic gradient descent (shuffled SGD) and provides tighter convergence bounds from a primal-dual perspective.
Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences
Zhaoze Wang (University of Pennsylvania), Vijay Balasubramanian (Santa Fe Institute)
Representation LearningRecurrent Neural NetworkAuto EncoderSequential
🎯 What it does: Train an autoencoder to reconstruct complete perceptions from partially noisy perception sequences in a simulated room, thereby spontaneously generating radial fields similar to hippocampal place cells in the encoding layer and demonstrating phenomena such as remapping, orthogonality, and drift.
Time-Constrained Robust MDPs
Adil Zouitine (IRT Saint-Exupéry), Emmanuel Rachelson
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: Proposes a Time-Constrained Robust MDP (TC-RMDP) framework and designs three algorithms (Oracle-TC, Stacked-TC, vanilla-TC) to address reinforcement learning problems with non-rectangular and time-varying uncertainties.
Time-FFM: Towards LM-Empowered Federated Foundation Model for Time Series Forecasting
Qingxiang Liu (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)
Federated LearningSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: A federated foundation model TIME-FFM based on a pre-trained language model has been constructed for cross-domain time series prediction, addressing data privacy and multi-domain heterogeneity issues.
Time-Reversal Provides Unsupervised Feedback to LLMs
Yerram Varun, Prateek Jain (Google DeepMind)
GenerationRetrievalSafty and PrivacyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a Time-Reversed Language Model (TRLM), which utilizes a reverse (response → question) direction for scoring and generation, providing unsupervised feedback, and further improving generation re-ranking, citation attribution, retrieval, and safety filtering.
Time-Varying LoRA: Towards Effective Cross-Domain Fine-Tuning of Diffusion Models
Zhan Zhuang (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
GenerationDomain AdaptationDiffusion modelContrastive LearningImage
🎯 What it does: A time-varying low-rank adapter called Terra is proposed for cross-domain generation and fine-tuning of diffusion models, capable of achieving continuous domain flow generation and interpolation on a low-rank module.
TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables
Yuxuan Wang (Tsinghua University), Mingsheng Long (Tsinghua University)
TransformerTime SeriesFinance Related
🎯 What it does: The TimeXer model is proposed, utilizing the self-attention and cross-attention of the Transformer to encode endogenous time series in patches and exogenous sequences in variates, and connects the two through globally learnable tokens, achieving precise integration of exogenous variables in time series forecasting.
Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series
Vijay Ekambaram (IBM Research), Jayant Kalagnanam (IBM Research)
Time Series
🎯 What it does: This paper proposes Tiny Time Mixers (TTM), a very small pre-trained model specifically designed for zero/few-shot multivariate time series forecasting.
TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge
Huanan LI, Zhangming Zhu (Xidian University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: Proposes the TinyLUT framework, which utilizes a separation mapping strategy and a dynamic discretization mechanism to map convolutional networks into LUTs with minimal storage, achieving efficient image restoration.
TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices
Hong Jia (University of Cambridge), Cecilia Mascolo (University of Cambridge)
Domain AdaptationComputational EfficiencyImage
🎯 What it does: Proposed the TinyTTA framework, achieving efficient test-time adaptation (TTA) on resource-constrained MCUs, and developed the TinyTTA Engine library;
To Believe or Not to Believe Your LLM: Iterative Prompting for Estimating Epistemic Uncertainty
Yasin Abbasi-Yadkori, Csaba Szepesvari
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the knowledge uncertainty of large language models and proposes a method to detect model hallucinations by constructing a pseudo-joint distribution through iterative prompting and calculating information-theoretic metrics.
To Err Like Human: Affective Bias-Inspired Measures for Visual Emotion Recognition Evaluation
Chenxi Zhao (Nankai University), Jufeng Yang (Nankai University)
RecognitionConvolutional Neural NetworkImage
🎯 What it does: Proposes an emotional distance based on the Mikel emotion wheel to define misclassification costs, and introduces two new visual emotion recognition evaluation metrics: ECC (Overall Evaluation) and EMC (Misclassification Evaluation Only), while validating their effectiveness in semi-supervised learning.
To Learn or Not to Learn, That is the Question — A Feature-Task Dual Learning Model of Perceptual Learning
Xiao Liu (Peking University), Si Wu (Peking University)
Convolutional Neural NetworkImage
🎯 What it does: A dual learning model (task-based learning + feature-based learning) was constructed to explain the specificity and transfer phenomena in perceptual learning.
Token Merging for Training-Free Semantic Binding in Text-to-Image Synthesis
taihang Hu, Yaxing Wang (Nankai University)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: Proposes a training-free Token Merging (ToMe) method to address the semantic binding issue in text-to-image generation;
Tolerant Algorithms for Learning with Arbitrary Covariate Shift
Surbhi Goel (University of Pennsylvania), Arsen Vasilyan (University of California Berkeley)
ClassificationAnomaly DetectionTabular
🎯 What it does: The study addresses learning problems under arbitrary covariate shift (distribution shift) and proposes a fault-tolerant algorithm compatible with PQ learning and TDS learning frameworks. It achieves dimension-efficient and error-controllable selective classification and tolerance distribution detection for basic concept classes such as half-spaces, intersections of half-spaces, and decision trees.
TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment
Wei Li (Zhejiang University), Yi Yang (Zhejiang University)
TransformerLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: A pre-alignment framework called TOPA is constructed, which enables large language models to possess video understanding capabilities using only text.
TopoFR: A Closer Look at Topology Alignment on Face Recognition
Jun Dan (Zhejiang University), Shan Luo (King's College London)
RecognitionImageBenchmarkStochastic Differential Equation
🎯 What it does: This paper proposes TopoFR, which enhances face recognition performance through topological structure alignment and hard sample mining.
TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes
Yanping Fu (Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences), Feng Dai (Institute of Computing Technology, Chinese Academy of Sciences)
Autonomous DrivingExplainability and InterpretabilityGraph Neural NetworkTransformerImage
🎯 What it does: An interpretable lane topology reasoning method called TopoLogic is designed, which combines geometric distance and semantic similarity to determine the connectivity between lanes.
Topological Generalization Bounds for Discrete-Time Stochastic Optimization Algorithms
Rayna Andreeva (University of Edinburgh), Umut Simsekli (Ecole Normale Superieure PSL Research University)
OptimizationGraph Neural NetworkTransformerImage
🎯 What it does: A new class of topology-based complexity measures (α-weighted lifespan and forward quantization) is proposed, and its upper bound on the generalization error of discrete-time stochastic optimization algorithms is provided.
Topological obstruction to the training of shallow ReLU neural networks
Marco Nurisso (Politecnico di Torino), Francesco Vaccarino (Politecnico di Torino)
OptimizationBiomedical Data
🎯 What it does: This study investigates the topological limitations of shallow ReLU networks during the gradient flow optimization process, revealing issues of conservativeness and disconnection in the parameter space due to the homogeneity of the activation function.
Toward a Stable, Fair, and Comprehensive Evaluation of Object Hallucination in Large Vision-Language Models
Hongliang Wei (Harbin Institute of Technology), Debin Zhao (Harbin Institute of Technology)
Object DetectionGenerationTransformerVision Language ModelImageText
🎯 What it does: The study investigates the impact of instruction on object hallucinations produced by large visual language models (LVLM) and proposes the LeHaCE framework to evaluate the degree of hallucination based on length, enhancing the stability and fairness of the assessment.
Toward a Well-Calibrated Discrimination via Survival Outcome-Aware Contrastive Learning
Dongjoon Lee (Chung Ang University), Changhee Lee (Korea University)
Contrastive LearningTime SeriesBiomedical Data
🎯 What it does: A deep survival model based on contrastive learning, ConSurv, is proposed, which constructs contrastive loss using weighted negative sampling based on the similarity of patient event times, thereby enhancing the model's discriminative ability while maintaining calibration.
Toward Approaches to Scalability in 3D Human Pose Estimation
Jun-Hee Kim (Korea University), Seong-Whan Lee (Korea University)
Pose EstimationDepth EstimationImage
🎯 What it does: This paper proposes two key technologies: ① A Biomechanical Pose Generator (BPG) based on the Normal Range of Motion (NROM), which automatically generates diverse 3D human poses that conform to anatomical constraints without relying on source data; ② Binary Depth Coordinates (BDC), which simplifies 3D depth estimation into a binary classification of 'front/back', using bone lengths and 2D positional information to solve through quadratic equations and eliminate depth ambiguity with binary depth.
Toward Conditional Distribution Calibration in Survival Prediction
Shi-ang Qi (University of Alberta), Russell Greiner (University of Alberta)
Biomedical Data
🎯 What it does: This paper proposes a post-processing method called CiPOT for calibrating the conditional distribution of survival prediction models, enhancing both marginal and conditional calibration performance without sacrificing discriminative power.
Toward Dynamic Non-Line-of-Sight Imaging with Mamba Enforced Temporal Consistency
Yue Li (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
RestorationVideo
🎯 What it does: A dynamic NLOS reconstruction framework based on ST-Mamba is proposed, utilizing multi-frame transient information to achieve high-resolution hidden object recovery.
Toward Efficient Inference for Mixture of Experts
Haiyang Huang (Duke University), Benjamin Lee (University of Pennsylvania)
Mixture of ExpertsTextBiomedical Data
🎯 What it does: A systematic analysis of the inference process of the Mixture-of-Experts (MoE) model is conducted, proposing three optimization techniques: dynamic gating, expert buffering, and load balancing, to improve inference throughput, reduce memory usage, and enhance system stability.
Toward Global Convergence of Gradient EM for Over-Paramterized Gaussian Mixture Models
Weihang Xu (University of Washington), Simon Shaolei Du
OptimizationTabular
🎯 What it does: Conducted a global convergence analysis of the gradient EM algorithm for over-parameterized Gaussian mixture models (GMM), proving that under a single Gaussian true value, the gradient EM can converge to the maximum likelihood estimate from any initialization.
Toward Real Ultra Image Segmentation: Leveraging Surrounding Context to Cultivate General Segmentation Model
Sai Wang (Wuhan University), Bo Du (Wuhan University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: Proposes the SGNet framework, which utilizes surrounding context to guide local image patches for ultra-high-resolution image segmentation, compatible with any general semantic segmentation model;
Toward Robust Incomplete Multimodal Sentiment Analysis via Hierarchical Representation Learning
Mingcheng Li (Fudan University), Lihua Zhang (Fudan University)
Representation LearningTransformerContrastive LearningMultimodality
🎯 What it does: A hierarchical representation learning framework (HRLF) is proposed to specifically address the issue of uncertain missing modalities in multimodal sentiment analysis.
Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing
Ye Tian (Tencent AI Lab), Dong Yu (Tencent AI Lab)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The ALPHALLM framework is proposed, achieving self-improvement of LLMs;
Toward Semantic Gaze Target Detection
Samy Tafasca (Idiap Research Institute), Jean-marc Odobez
RecognitionObject DetectionTransformerContrastive LearningImageBenchmark
🎯 What it does: An end-to-end semantic gaze target detection model is designed, capable of predicting both gaze heatmaps and target categories simultaneously.
Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution
Yizi Zhang (Columbia University), Cole Lincoln Hurwitz
TransformerPrompt EngineeringBiomedical Data
🎯 What it does: A neural spike data foundation model based on Multi-task Masking (MtM) is proposed and implemented, capable of learning neural dynamics across animals and brain regions at single-cell and single-spike resolution, and trained and evaluated on the IBL Neuropixels dataset.
Towards a Scalable Reference-Free Evaluation of Generative Models
Azim Ospanov (Chinese University of Hong Kong), Farzan Farnia (Chinese University of Hong Kong)
GenerationComputational EfficiencyImageVideoText
🎯 What it does: A scalable reference-agnostic evaluation method FKEA based on random Fourier features has been developed to efficiently estimate the diversity score of generative models.
Towards a Theoretical Understanding of the 'Reversal Curse' via Training Dynamics
Hanlin Zhu (University of California Berkeley), Stuart Russell (University of California Berkeley)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper investigates the flaws of autoregressive large language models in reverse logical reasoning (the 'reversal curse') through theoretical analysis and experiments, and further explores the impact of chain-of-thought (COT) reasoning.
Towards a theory of how the structure of language is acquired by deep neural networks
Francesco Cagnetta (École Polytechnique Fédérale de Lausanne), Matthieu Wyart (École Polytechnique Fédérale de Lausanne)
Convolutional Neural NetworkTransformerText
🎯 What it does: The study investigated the amount of data required to learn language structure through next-token prediction, using a synthetic dataset generated by probabilistic context-free grammar (PCFG).
Towards Accurate and Fair Cognitive Diagnosis via Monotonic Data Augmentation
Zheng Zhang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Tabular
🎯 What it does: To address the issue of data sparsity in cognitive diagnosis, a data augmentation framework based on monotonicity constraints, CMCD, is proposed to enhance the accuracy and fairness of diagnosis.
Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learning
Lanqing Li (Zhejiang Lab), Pheng-Ann Heng (Chinese University of Hong Kong)
Robotic IntelligenceMeta LearningReinforcement LearningContrastive LearningSequential
🎯 What it does: A unified information-theoretic framework for context-based offline meta reinforcement learning (COMRL) is proposed, integrating existing COMRL algorithms and revealing their relationships through theoretical analysis.
Towards Calibrated Robust Fine-Tuning of Vision-Language Models
Changdae Oh (University of Wisconsin Madison), Kyungwoo Song (Yonsei University)
ClassificationDomain AdaptationOptimizationTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: A method called CaRot is proposed for robust fine-tuning of Vision-Language models (such as CLIP) in distribution shift environments, which simultaneously improves OOD accuracy and confidence calibration.
Towards Combating Frequency Simplicity-biased Learning for Domain Generalization
Xilin He (Shenzhen University), Linlin Shen (University of Exeter)
RecognitionDomain AdaptationAdversarial AttackImage
🎯 What it does: This study investigates the impact of frequency shortcut learning on single-source domain generalization and proposes two adversarial frequency augmentation modules, AAUA and AAD, to dynamically adjust the frequency characteristics of the dataset to suppress frequency shortcut learning.
Towards Croppable Implicit Neural Representations
Maor Ashkenazi (Ben Gurion University of the Negev), Eran Treister (Ben Gurion University of the Negev)
Image TranslationRestorationSuper ResolutionImageVideoAudio
🎯 What it does: Proposes the Local-Global SIREN architecture, which can directly crop or extend the already encoded implicit neural representation without retraining or fine-tuning, and can be further used for various downstream tasks.
Towards Diverse Device Heterogeneous Federated Learning via Task Arithmetic Knowledge Integration
Mahdi Morafah (University of California San Diego), Bill Lin (Czech Technical University in Prague)
Federated LearningKnowledge DistillationConvolutional Neural NetworkTransformerImageText
🎯 What it does: In the context of federated learning with heterogeneous devices, the TAKFL framework is proposed, which utilizes knowledge distillation to treat the model set of each device prototype as an independent task for knowledge transfer, and achieves multi-source knowledge fusion through self-supervised regularization and adaptive task arithmetic.
Towards Dynamic Message Passing on Graphs
Junshu Sun (Institute of Computing Technology, Chinese Academy of Sciences), Shuhui Wang (Institute of Computing Technology, Chinese Academy of Sciences)
ClassificationRecommendation SystemGraph Neural NetworkGraph
🎯 What it does: This paper proposes a novel graph neural network model N2 based on dynamic information transmission, which constructs variable transmission paths using measurable similarity between pseudo-nodes and graph nodes in a shared state space, achieving low-complexity global message passing.
Towards Editing Time Series
Baoyu Jing (University of Illinois at Urbana-Champaign), Kan Ren (ShanghaiTech University)
TransformerDiffusion modelTime Series
🎯 What it does: Proposed and implemented the Time Series Editing (TSE) task, allowing for editing based on specified attributes while maintaining the original sequence details.
Towards Effective Planning Strategies for Dynamic Opinion Networks
Bharath Chandra Muppasani (University of South Carolina), Michael Huhns (University of South Carolina)
OptimizationGraph Neural NetworkSupervised Fine-TuningReinforcement LearningGraph
🎯 What it does: This paper studies the suppression of rumor spreading in dynamic opinion networks by identifying key nodes and disseminating accurate information to them, proposing a ranking-based supervised learning and centralized dynamic programming based on reinforcement learning.
Towards Efficient and Optimal Covariance-Adaptive Algorithms for Combinatorial Semi-Bandits
Julien Zhou (Criteo AI Lab), Julyan Arbel (Univ. Grenoble Alpes)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: Two adaptive algorithms for combinatorial semi-bandits are proposed: OLS-UCB-C (UCB based on covariance estimation) and COS-V (sampling-based TS heuristic).
Towards Estimating Bounds on the Effect of Policies under Unobserved Confounding
Alexis Bellot (Google DeepMind), Silvia Chiappa (Google DeepMind)
Tabular
🎯 What it does: A method for partial identification and estimation of policy effects using causal graphs in the presence of unobserved confounding is proposed.
Towards Exact Gradient-based Training on Analog In-memory Computing
Zhaoxian Wu (Rensselaer Polytechnic Institute), Tianyi Chen (IBM)
OptimizationImagePhysics Related
🎯 What it does: This paper conducts an in-depth study of the theoretical model and convergence behavior of gradient descent training in Analog In-Memory Computing (AIMC), proposing a new analytical framework and verifying that the Tiki-Taka algorithm can eliminate the asymptotic error of traditional Analog SGD.
Towards Flexible 3D Perception: Object-Centric Occupancy Completion Augments 3D Object Detection
Chaoda Zheng (Shenzhen University), Zhen Li (Chinese University of Hong Kong)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes an Object-Centric Occupancy representation and implements an enhanced 3D object detection and shape completion network based on this representation.
Towards Flexible Visual Relationship Segmentation
Fangrui Zhu (Northeastern University), Huaizu Jiang (Northeastern University)
Object DetectionSegmentationTransformerPrompt EngineeringImage
🎯 What it does: This paper presents FleVRS, a unified one-stage visual relation segmentation framework capable of performing standard relation segmentation, promptable relation segmentation, and open-vocabulary relation segmentation.
Towards Global Optimal Visual In-Context Learning Prompt Selection
Chengming Xu (Fudan University), Yanwei Fu (Fudan University)
Image TranslationObject DetectionSegmentationTransformerContrastive LearningImage
🎯 What it does: A Transformer-based list-wise ranking model and consistency-aware aggregator are proposed for Visual Context Learning (VICL) to select the optimal contextual examples from multiple candidate samples, thereby enhancing the model's segmentation, detection, and coloring performance.
Towards Harmless Rawlsian Fairness Regardless of Demographic Prior
Xuanqian Wang (Beihang University), Yew-Soon Ong (Nanyang Technological University)
ClassificationOptimizationTabular
🎯 What it does: A VFair method based on loss variance minimization is proposed, achieving Rawlsian fairness during training without using any sensitive attributes; and harmless fairness is achieved through dynamic gradient adjustment.
Towards Human-AI Complementarity with Prediction Sets
Giovanni De Toni (Fondazione Bruno Kessler and University of Trento), Manuel Gomez Rodriguez
ClassificationOptimizationMixture of ExpertsImageTabular
🎯 What it does: This paper addresses the problem of constructing an optimal prediction set for human experts in multi-class tasks and proves its NP-hardness and approximation difficulty.
Towards Learning Group-Equivariant Features for Domain Adaptive 3D Detection
Sangyun Shin (University of Oxford), Niki Trigoni (University of Oxford)
Object DetectionDomain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: A domain-adaptive 3D object detection framework called GroupExp-DA is proposed, which utilizes a grouping strategy to balance training attention and improves detection accuracy through group equivariant spatial features.
Towards Multi-dimensional Explanation Alignment for Medical Classification
Lijie Hu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
ClassificationExplainability and InterpretabilityLarge Language ModelVision Language ModelImageBiomedical DataComputed Tomography
🎯 What it does: An end-to-end multidimensional interpretable medical image classification framework Med-MICN is proposed, which can automatically generate concept labels and achieve multidimensional alignment by combining neural symbolic reasoning, conceptual semantics, and saliency maps.
Towards Multi-Domain Learning for Generalizable Video Anomaly Detection
MyeongAh Cho (Kyung Hee University), Sangyoun Lee (Yonsei University)
Anomaly DetectionVideo
🎯 What it does: Proposes the Multi-Domain Video Anomaly Detection (MDVAD) task to address the conflict in anomaly definitions across different datasets and to build a general model;
Towards Neuron Attributions in Multi-Modal Large Language Models
Junfeng Fang (University of Science and Technology of China), Tat-Seng Chua (National University of Singapore)
SegmentationGenerationTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A neuron attribution method named NAM for multimodal large language models (MLLM) is proposed, which can identify key neurons during text and image generation processes and enable image editing.
Towards Next-Generation Logic Synthesis: A Scalable Neural Circuit Generation Framework
Zhihai Wang (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationNeural Architecture SearchReinforcement Learning
🎯 What it does: A differentiable neural circuit generation framework T-Net is proposed, based on a triangular network structure, regularized skip connections, and hardness-aware loss, to accurately synthesize large-scale AIG circuits, combined with reinforcement learning and evolutionary algorithms for circuit optimization.
Towards Next-Level Post-Training Quantization of Hyper-Scale Transformers
Junhan Kim (Samsung Research), Yongkweon Jeon (Samsung Research)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A new post-training quantization algorithm called aespa is proposed, specifically targeting weight quantization for large-scale Transformer models, balancing accuracy and efficiency.
Towards Open-Vocabulary Semantic Segmentation Without Semantic Labels
Heeseong Shin (KAIST), Seungryong Kim (KAIST)
SegmentationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes PixelCLIP, which utilizes unlabeled masks from visual foundation models (such as SAM and DINO) to fine-tune the image encoder of CLIP, achieving open-vocabulary semantic segmentation without semantic labels.
Towards Principled Graph Transformers
Luis Müller (RWTH Aachen University), Christopher Morris (RWTH Aachen University)
Graph Neural NetworkTransformerGraph
🎯 What it does: A graph learning framework based on Edge Transformer is proposed, achieving the expressiveness of 3-WL through triangular attention.
Towards Robust Multimodal Sentiment Analysis with Incomplete Data
Haoyu Zhang (Chinese University of Hong Kong), Tianshu Yu (Wuhan University)
Adversarial AttackTransformerMultimodality
🎯 What it does: A robust model called LNLN is proposed to address data missingness in multimodal emotion analysis.
Towards Safe Concept Transfer of Multi-Modal Diffusion via Causal Representation Editing
Peiran Dong (Hong Kong Polytechnic University), Zicong Hong (Hong Kong Polytechnic University)
GenerationDiffusion modelImageMultimodality
🎯 What it does: A Causal Representation Editing (CRE) framework is proposed to achieve safe concept transfer in multimodal diffusion models.
Towards Scalable and Stable Parallelization of Nonlinear RNNs
Xavier Gonzalez (Stanford University), Scott Linderman
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: This paper proposes a scalable and stable parallelization method for nonlinear RNNs, overcoming the computational complexity and numerical instability issues of the original DEER method.
Towards Stable Representations for Protein Interface Prediction
Ziqi Gao (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
Adversarial AttackProtein Structure PredictionGraph Neural NetworkGraphBiomedical Data
🎯 What it does: The ATProt framework is proposed, which achieves representation stability through adversarial training of protein graph representations, thereby improving protein interface prediction performance.
Towards the Dynamics of a DNN Learning Symbolic Interactions
Qihan Ren (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
ImageTextPoint Cloud
🎯 What it does: This study demonstrates the two-stage dynamics of interaction in deep neural networks (DNN), providing a theoretical mechanism for how DNN transitions from underfitting to overfitting during training.
Towards the Transferability of Rewards Recovered via Regularized Inverse Reinforcement Learning
Andreas Schlaginhaufen (École Polytechnique Fédérale de Lausanne), Maryam Kamgarpour (École Polytechnique Fédérale de Lausanne)
Reinforcement Learning
🎯 What it does: This study investigates the transferability of reward functions learned from limited expert demonstrations within the framework of regularized Inverse Reinforcement Learning (IRL) across different environments (with different transition dynamics), proposing new theoretical conditions and algorithms.
Towards training digitally-tied analog blocks via hybrid gradient computation
Timothy Nest (Montreal Institute of Learning Algorithms), Maxence Ernoult (Rain AI)
OptimizationImage
🎯 What it does: A mixed digital-analog feedforward coupling energy model (ff-EBMs) is proposed, along with an end-to-end BP-EP gradient chain algorithm for training deep networks on hardware mixed architectures.
Towards Understanding Evolving Patterns in Sequential Data
QIUHAO Zeng, Boyu Wang (Western University)
Recurrent Neural NetworkTransformerAuto EncoderVideoTime SeriesSequentialFinance Related
🎯 What it does: A metric based on mutual information (EVORATE) is proposed to quantitatively assess evolutionary patterns in sequence data, and it is extended to scenarios without correspondence (EVORATE W), thereby helping to determine the suitability of using sequence models, estimate the order of sequences, and perform feature selection.
Towards Understanding Extrapolation: a Causal Lens
Lingjing Kong (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Domain AdaptationAuto EncoderImage
🎯 What it does: This study investigates the extrapolation problem with only a single target sample and the target distribution not supported by the training distribution. It proposes a latent variable model centered on the principle of minimal change and provides theoretical identifiability conditions for achievable extrapolation.
Towards Understanding How Transformers Learn In-context Through a Representation Learning Lens
Ruifeng Ren (Renmin University of China), Yong Liu (Renmin University of China)
OptimizationRepresentation LearningTransformerContrastive LearningTabular
🎯 What it does: This paper analyzes the in-context learning (ICL) process of Transformer from the perspective of representation learning and proposes a framework that views the attention layer as a dual model gradient descent.
Towards Understanding the Working Mechanism of Text-to-Image Diffusion Model
Mingyang Yi (Renmin University of China), Zhenguo Li (Huawei Noah's Ark Lab)
GenerationData SynthesisComputational EfficiencyDiffusion modelImageText
🎯 What it does: The working mechanism of the text-to-image diffusion model was explored and verified, revealing that the denoising process first restores the overall shape and then fills in details; it was also demonstrated that the special token [EOS] in the text prompt plays a dominant role in the overall shape during the early generation phase.
Towards Unified Multimodal Editing with Enhanced Knowledge Collaboration
Kaihang Pan (Zhejiang University), Qianru Sun (Singapore Management University)
GenerationRetrievalTransformerContrastive LearningMultimodality
🎯 What it does: A unified multimodal editing framework called UniKE is proposed, which balances internal knowledge editing and external knowledge retrieval.
Towards Universal Mesh Movement Networks
Mingrui Zhang (Imperial College London), Matthew D Piggott
Graph Neural NetworkTransformerMesh
🎯 What it does: This paper proposes a grid movement network UM2N that can perform zero-shot transfer across different types of PDEs, boundary geometries, and mesh sizes.
Towards Unsupervised Model Selection for Domain Adaptive Object Detection
Hengfu Yu (University of Electronic Science and Technology of China), Lixin Duan (University of Electronic Science and Technology of China)
Object DetectionDomain AdaptationImage
🎯 What it does: An unsupervised model selection method called Detection Adaptation Score (DAS) is proposed, which can select nearly optimal checkpoints for domain adaptive object detection (DAOD) models without target domain labels.
Toxicity Detection for Free
Zhanhao Hu (University of California), David Wagner (University of California)
ClassificationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A toxicity detection method utilizing the output of large language models (the logits of the first token) for querying without additional models is proposed—MULI;
TPC: Test-time Procrustes Calibration for Diffusion-based Human Image Animation
Sunjae Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationPose EstimationDiffusion modelImageVideo
🎯 What it does: A testing-time Procrustes calibration (TPC) is proposed, which addresses the quality fluctuation problem caused by the mismatch in scale and rotation between the reference image and the target pose through shape alignment and iterative propagation in a diffusion-based portrait animation model.
TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning
Hui Chen (Xi'an Jiaotong University), Xin Yu (University of Queensland)
ClassificationRecognitionVision Language ModelContrastive LearningImageText
🎯 What it does: This paper addresses the scenario of generalized zero-shot learning (GZSL) without base/new category partition information, proposing a framework called Topology-Preserving Reservoir (TPR) based on the pre-trained visual-language model CLIP, which achieves the ability to recognize both base and new classes simultaneously.
Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs
Ching-An Cheng (Microsoft Research), Adith Swaminathan (Netflix)
OptimizationLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a comprehensive automation optimization framework based on execution trajectories called Trace, and presents its mathematical abstraction—Optimization and Trajectory Operator (OPTO) along with a general LLM-driven optimizer called OptoPrime. Through Trace, any traceable Python workflow can be automatically transformed into OPTO instances, achieving end-to-end optimization of heterogeneous parameters (code, prompts, hyperparameters, etc.).
Tracing Hyperparameter Dependencies for Model Parsing via Learnable Graph Pooling Network
Xiao Guo (Michigan State University), Xiaoming Liu (Michigan State University)
GenerationHyperparameter SearchGraph Neural NetworkImage
🎯 What it does: This paper studies the task of model interpretation, which involves predicting the hyperparameters of a generative model given a generated image.
TrackIME: Enhanced Video Point Tracking via Instance Motion Estimation
Seong Hyeon Park (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
Object TrackingSegmentationOptical FlowVideoBenchmark
🎯 What it does: Utilize instance motion estimation to crop the search space for video point tracking, and combine point tracking with instance segmentation to improve tracking accuracy and achieve zero-shot video instance segmentation.
TrAct: Making First-layer Pre-Activations Trainable
Felix Petersen (Stanford University), Stefano Ermon (Stanford University)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: A training strategy called TrAct is proposed, which applies gradient descent to the first layer activations during the training of visual models, directly modifying the update method of the first layer weights.
Trade-Offs of Diagonal Fisher Information Matrix Estimators
Alexander Soen (Australian National University), Ke Sun (CSIRO)
ClassificationOptimizationAuto EncoderImage
🎯 What it does: This paper studies two commonly used estimators of the Fisher information matrix, analyzes their accuracy and sample complexity, and explores their applications in neural network regression and classification.
Trading off Consistency and Dimensionality of Convex Surrogates for Multiclass Classification
Enrique Nueve (University of Colorado Boulder), Jessica Finocchiaro
ClassificationOptimizationComputational Efficiency
🎯 What it does: This paper studies the trade-off between consistency and dimensionality in multi-class classification and proposes two methods for balancing consistency and dimensionality when using convex surrogate loss.