NeurIPS 2024 Papers — Page 20
Conference on Neural Information Processing Systems · 4035 papers
Kraken: Inherently Parallel Transformers For Efficient Multi-Device Inference
Rohan Baskar Prabhakar (Princeton University), David Wentzlaff (Princeton University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: A Transformer variant named Kraken is designed and implemented, which reduces communication bottlenecks in multi-device inference by introducing fixed parallelism at each layer and performing AllReduce only once at the end of the layer.
Kronecker-Factored Approximate Curvature for Physics-Informed Neural Networks
Felix Dangel (Vector Institute), Marius Zeinhofer
OptimizationComputational EfficiencyPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes the application of Kronecker-factored approximate curvature (KFAC) to the Gauss-Newton matrix in the loss of Physics-Informed Neural Networks (PINN), significantly reducing training costs and achieving scalability for large-scale networks.
KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization
Tianyi Zhang (Rice University), Anshumali Shrivastava (ThirdAI Corporation)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A Coupled Quantization (CQ) method is proposed to compress the KV cache of large language models (LLMs) to reduce memory usage and improve inference throughput.
KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization
Coleman Richard Charles Hooper (University of California), Amir Gholami (University of California)
RetrievalCompressionTransformerLarge Language ModelText
🎯 What it does: The KVQuant method is proposed, which performs low-precision (≤3bit) quantization on the KV cache in LLM inference to support million-level context length inference.
L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer
Jin Shin (Seoul National University of Science and Technology), Hyun Kim (Seoul National University of Science and Technology)
SegmentationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A lightweight test-time adaptation method (L-TTA) is proposed, which achieves rapid adaptation to the target domain by only reconstructing the stem layer of the model.
L4GM: Large 4D Gaussian Reconstruction Model
Jiawei Ren (NVIDIA), Huan Ling (University of Toronto)
GenerationData SynthesisGaussian SplattingVideo
🎯 What it does: Developed L4GM, a one-time forward inference method that can transform single-view videos into high-quality 4D (dynamic) object reconstruction models;
Label Delay in Online Continual Learning
Botos Csaba (University of Oxford), Adel Bibi (University of Oxford)
Contrastive LearningTime SeriesSequential
🎯 What it does: The paper addresses the issue of label delay in online continual learning and presents a new experimental framework and evaluation metrics.
Label Noise: Ignorance Is Bliss
Yilun Zhu (University of Michigan), Clayton Scott (University of Michigan)
ClassificationData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A theoretical framework based on Relative Signal Strength (RSS) is proposed to analyze the limits under instance-dependent label noise, and the near-optimality of Noise Ignoring Empirical Risk Minimization (NI-ERM) is validated in both theory and practice.
LACIE: Listener-Aware Finetuning for Calibration in Large Language Models
Elias Stengel-Eskin (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes and implements the Listener-Aware Calibration for Implicit and Explicit confidence (LACIE) method, which fine-tunes large language models through a multi-agent speaker-listener game and preference optimization, enabling them to express confidence more accurately when answering questions.
LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations
Yuhang Li (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
Autonomous DrivingKnowledge DistillationTransformerTime SeriesSequential
🎯 What it does: This paper proposes a length-independent knowledge distillation framework LaKD for trajectory prediction of observation trajectories of arbitrary lengths.
LAM3D: Large Image-Point Clouds Alignment Model for 3D Reconstruction from Single Image
Ruikai Cui (Australian National University), Pan Ji (Tencent XR Vision Labs)
GenerationDepth EstimationTransformerDiffusion modelAuto EncoderImagePoint CloudMesh
🎯 What it does: A 3D mesh generation framework LAM3D is proposed, which can generate high-precision 3D meshes from a single image.
Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases
Hang Yin (Shanghai Jiao Tong University), Chenghu Zhou (Chinese Academy of Sciences)
Graph Neural NetworkContrastive LearningGraph
🎯 What it does: A unified framework called Lambda is proposed to address the issue of unlabeled dangling entities in knowledge graph entity alignment. It first detects dangling entities through iterative PU learning, and then utilizes GNN to select and aggregate them, completing entity alignment with spectral contrastive learning.
Langevin Unlearning: A New Perspective of Noisy Gradient Descent for Machine Unlearning
Eli Chien (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
OptimizationSafty and PrivacyConvolutional Neural NetworkImageStochastic Differential Equation
🎯 What it does: Proposes a Langevin Unlearning framework based on noise gradient descent, providing a 'forgotten' guarantee for approximate machine learning models.
Language Generation in the Limit
Jon Kleinberg (Cornell University), Sendhil Mullainathan (University of Chicago)
Generation
🎯 What it does: After providing a finite sample of an unknown language, it is proven that under the Gold-Angluin adversarial model, there exists an algorithm for any countable language set that can generate new unseen strings in the limit, and for finite sets, it can immediately generate infinitely after a finite sample;
Language Grounded Multi-agent Reinforcement Learning with Human-interpretable Communication
Huao Li (University of Pittsburgh), Katia P. Sycara
Explainability and InterpretabilityLarge Language ModelReinforcement LearningText
🎯 What it does: A pipeline (LangGround) is proposed to align the communication space of MARL agents with human natural language, achieving interpretable team communication.
Language Model as Visual Explainer
Xingyi Yang (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerLarge Language ModelDiffusion modelContrastive LearningImageText
🎯 What it does: The LVX method is proposed, which generates a hierarchical attribute tree using large language models, combines text-image generation and visual model embedding to construct an interpretable tree, dynamically refines the tree structure, and generates tree-shaped explanations for given samples. Finally, the explanation tree is used to calibrate the model with contrastive loss.
Language Models as Hierarchy Encoders
Yuan He (University of Oxford), Ian Horrocks (University of Oxford)
TransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: Retrain a Transformer-based language model to become a Hierarchical Transformer Encoder (HIT) that explicitly encodes hierarchical structures.
Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models
Hui-Po Wang (CISPA Helmholtz Center for Information Security), Mario Fritz (CISPA Helmholtz Center for Information Security)
CompressionConvolutional Neural NetworkTransformerLarge Language ModelImage
🎯 What it does: Proposes LM-GC, which combines a pre-trained large language model as a zero-shot gradient prior with arithmetic coding to achieve lossless gradient compression.
Language-Driven Interactive Traffic Trajectory Generation
Junkai XIA, Siheng Chen (Shanghai Jiao Tong University)
GenerationAutonomous DrivingTransformerLarge Language ModelPrompt EngineeringPoint Cloud
🎯 What it does: An interactive traffic trajectory generation model called InteractTraj has been developed, which can generate realistic trajectories containing interactions between vehicles based on user descriptions.
Large Language Model Unlearning
Yuanshun Yao (Meta), Yang Liu (University of California Santa Cruz)
Data-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: This paper proposes a method for rapid unlearning in large language models using only negative samples to eliminate harmful responses, copyright leaks, and hallucinations.
Large Language Model Unlearning via Embedding-Corrupted Prompts
Chris Yuhao Liu (University of California), Yang Liu (University of California)
OptimizationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A lightweight LLM unlearning framework called Embedding-Corrupted (ECO) Prompts is proposed, which learns perturbations of the embeddings of prompts that meet the forgetting conditions during the inference phase, allowing the model to 'forget' specified knowledge without updating weights.
Large language model validity via enhanced conformal prediction methods
John Cherian, Emmanuel Candes
TransformerLarge Language ModelTextBiomedical Data
🎯 What it does: This paper develops an improved conditional conformal prediction framework that provides credibility guarantees and filters out erroneous statements in the outputs of large language models.
Large Language Models as Urban Residents: An LLM Agent Framework for Personal Mobility Generation
Jiawei Wang (University of Tokyo), Chuan Xiao (Osaka University)
GenerationRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelAgentic AITime SeriesSequentialRetrieval-Augmented Generation
🎯 What it does: Using large language models (LLM) as agents, combined with real urban personal mobility trajectory data, a framework based on activity patterns and motivations is proposed to generate interpretable and semantically rich personal daily activity trajectories.
Large Language Models Must Be Taught to Know What They Don’t Know
Sanyam Kapoor (New York University), Andrew Gordon Wilson (New York University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the uncertainty estimation of large language models (LLMs) and proposes to achieve high-quality calibrated uncertainty by fine-tuning the model with LoRA+Prompt on a small amount of labeled data.
Large Language Models Play StarCraft II:Benchmarks and A Chain of Summarization Approach
Weiyu Ma (Institute of Automation, Chinese Academy of Sciences), Haifeng Zhang (Nanjing Artificial Intelligence Research of IA)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Developed a textual StarCraft II environment (TextStarCraft II) and a summary-based decision-making process (Chain of Summarization, CoS), and conducted real-time strategy decision-making and evaluation using large language models (LLM) in this environment.
Large Language Models-guided Dynamic Adaptation for Temporal Knowledge Graph Reasoning
Jiapu Wang (Beijing University of Technology), Baocai Yin (Beijing University of Technology)
Graph Neural NetworkTransformerLarge Language ModelGraphTime Series
🎯 What it does: A dynamic adaptation framework LLM-DA based on large language models is proposed for temporal knowledge graph reasoning.
Large Pre-trained time series models for cross-domain Time series analysis tasks
Harshavardhan Kamarthi (Georgia Institute of Technology), B. Aditya Prakash (Georgia Institute of Technology)
ClassificationAnomaly DetectionOptimizationRecurrent Neural NetworkTransformerTime SeriesFinance Related
🎯 What it does: A cross-domain time series pre-training model LPTM is proposed, utilizing adaptive segmentation and self-supervised masking tasks, supporting zero-shot or few-shot fine-tuning to complete various prediction and classification tasks.
Large Scale Transfer Learning for Tabular Data via Language Modeling
Joshua P Gardner, Ludwig Schmidt (Stanford University)
TransformerLarge Language ModelTabular
🎯 What it does: Developed TABULA-8B, a language model based on Llama 3-8B for unsupervised table prediction, and constructed a T4 training dataset with a scale of 4.2M tables and 2.1B rows.
Large Spatial Model: End-to-end Unposed Images to Semantic 3D
Zhiwen Fan (University of Texas at Austin), Yue Wang (University of Southern California)
SegmentationGenerationDepth EstimationTransformerNeural Radiance FieldImagePoint Cloud
🎯 What it does: A unified Transformer framework called LSM is proposed, capable of reconstructing dense geometry, semantics, and novel view rendering from uncalibrated RGB images in a single forward pass.
Large Stepsize Gradient Descent for Non-Homogeneous Two-Layer Networks: Margin Improvement and Fast Optimization
Yuhang Cai (University of California Berkeley), Peter Bartlett
OptimizationImage
🎯 What it does: This study investigates the convergence, margin stability, and regularization effects of large step gradient descent in two-layer networks.
LaSCal: Label-Shift Calibration without target labels
Teodora Popordanoska (KU Leuven), Matthew B. Blaschko (KU Leuven)
Domain AdaptationOptimizationConvolutional Neural NetworkTransformerImageTextMultimodality
🎯 What it does: A consistent Calibration Error (CE) estimator under the label shift assumption is proposed, and an unsupervised post-processing calibration method called LaSCal is implemented based on this estimator.
LaSe-E2V: Towards Language-guided Semantic-aware Event-to-Video Reconstruction
Kanghao Chen (Hong Kong University of Science and Technology), Lin Wang (Hong Kong University of Science and Technology)
RestorationGenerationData SynthesisDiffusion modelVideoText
🎯 What it does: This paper proposes a language-guided event camera video reconstruction framework, LaSe-E2V, which combines event streams with textual descriptions to achieve semantically aware high-quality video reconstruction.
Last-Iterate Convergence for Generalized Frank-Wolfe in Monotone Variational Inequalities
Zaiwei Chen (Purdue University), Eric Mazumdar (California Institute of Technology)
Optimization
🎯 What it does: A generalized Frank-Wolfe (FW) algorithm is proposed and analyzed to solve the constrained monotone variational inequality (MVI) problem, providing last-iterate convergence rates of O(T⁻¹/2) and O(T⁻¹/6) for deterministic and stochastic cases, respectively.
Last-Iterate Global Convergence of Policy Gradients for Constrained Reinforcement Learning
Alessandro Montenegro (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: A general primal-dual algorithm based on policy gradient, C-PG, and its action-based and parameter-based versions, C-PGAE and C-PGPE, are proposed to solve constrained reinforcement learning problems in continuous control, supporting constraints based on risk metrics.
Latent Diffusion for Neural Spiking Data
Jaivardhan Kapoor (University of Tübingen), Jakob H. Macke (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisSpiking Neural NetworkDiffusion modelAuto EncoderTime SeriesSequentialAudio
🎯 What it does: LDNS combines autoencoders with diffusion models to generate realistic, variable-length, and conditionally neural spike data in latent space, achieving low-dimensional latent variable inference.
Latent Functional Maps: a spectral framework for representation alignment
Marco Fumero (IST Austria), Emanuele Rodolà (Sapienza University of Rome)
RetrievalDomain AdaptationRepresentation LearningImageMultimodality
🎯 What it does: Proposes the Latent Functional Maps (LFM) framework, which aligns, evaluates similarity, and transfers information across different models' latent spaces using spectral geometry and functional mapping methods.
Latent Intrinsics Emerge from Training to Relight
Xiao Zhang (University of Chicago), Anand Bhattad (Toyota Technological Institute at Chicago)
Image TranslationRestorationGenerationAuto EncoderImage
🎯 What it does: A completely data-driven image relighting model is proposed, which maps the intrinsic properties of the scene and lighting to latent variables through an autoencoder, and synthesizes images under new lighting in the decoder;
Latent Learning Progress Drives Autonomous Goal Selection in Human Reinforcement Learning
Gaia Molinaro (University of California), Anne Collins
Reinforcement LearningSequential
🎯 What it does: By designing a hierarchical reinforcement learning task, this study investigates how humans drive learning through self-set goals and verifies whether the latent learning progress (LLP) can explain human goal selection.
Latent Neural Operator for Solving Forward and Inverse PDE Problems
Tian Wang (Chinese Academy of Sciences), Chuang Wang (Chinese Academy of Sciences)
TransformerTime SeriesPhysics Related
🎯 What it does: Designed and implemented the Latent Neural Operator (LNO), which utilizes Physics-Cross-Attention (PhCA) to learn the operations of PDEs in a learnable latent space, capable of solving both forward and inverse problems.
Latent Paraphrasing: Perturbation on Layers Improves Knowledge Injection in Language Models
Minki Kang (KRAFTON), Jaewoong Cho (KRAFTON)
TransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Enhancing knowledge injection capabilities in large language models by injecting learned noise (i.e., 'latent rewriting') in the early layers, avoiding the repeated use of external generative models for data-level rewriting.
Latent Plan Transformer for Trajectory Abstraction: Planning as Latent Space Inference
Deqian Kong (University of California, Los Angeles), Ying Nian Wu (University of California, Los Angeles)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes a Latent Plan Transformer based on an implicit variational autoregressive model for planning tasks, utilizing trajectory-reward pairs as training data to directly infer trajectories from global rewards.
Latent Representation Matters: Human-like Sketches in One-shot Drawing Tasks
Victor Boutin (Centre de Recherche Cerveau & Cognition CNRS), Rufin VanRullen (Artificial and Natural Intelligence Toulouse Institute)
GenerationRepresentation LearningDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: The study introduces various representational inductive biases in a one-shot drawing task and evaluates their impact on the ability of Latent Diffusion Models to generate human-like sketches.
Layer-Adaptive State Pruning for Deep State Space Models
Minseon Gwak (POSTECH), PooGyeon Park (POSTECH)
CompressionComputational EfficiencyRecurrent Neural NetworkTime SeriesSequentialAudio
🎯 What it does: This paper studies a layer adaptive state pruning method called LAST for deep state space models, aimed at reducing the dimensionality of the model states and lowering computational and memory costs while maintaining performance.
LCGen: Mining in Low-Certainty Generation for View-consistent Text-to-3D
Zeng Tao (Fudan University), Wenqiang Zhang (Fudan University)
GenerationData SynthesisDiffusion modelNeural Radiance FieldText
🎯 What it does: This paper proposes a low-confidence generation method named LCGen, aimed at addressing the Janus problem in text-to-3D generation based on SDS, allowing the model to produce different confidence distributions from different perspectives, thereby enhancing viewpoint consistency.
LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling
Yaohua Zha (Tsinghua University), Shu-Tao Xia (Tsinghua University)
ClassificationObject DetectionSegmentationTransformerPoint Cloud
🎯 What it does: A local constraint compact point cloud model named LCM is proposed for the Masked Point Modeling (MPM) task, which includes a locally aggregated encoder and a Mamba-based decoder.
Learn more, but bother less: parameter efficient continual learning
Fuli Qiao (Pennsylvania State University), Mehrdad Mahdavi (Pennsylvania State University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In large-scale language models, a parameter-efficient continuous learning framework LB-CL is proposed. This framework utilizes low-rank SVD adapters combined with sensitivity analysis to extract and initialize knowledge from prior tasks, and maintains orthogonality between the gradient subspace of new tasks and old tasks through orthogonal gradient projection, thus achieving a balance between forward transfer and preventing catastrophic forgetting.
Learn To be Efficient: Build Structured Sparsity in Large Language Models
Haizhong Zheng (University of Michigan), Atul Prakash (University of Michigan)
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Train an algorithm called LTE to enable large language models to activate fewer neurons during inference, constructing structured sparsity to improve inference efficiency.
Learnability Matters: Active Learning for Video Captioning
Yiqian Zhang (National University of Singapore), Jun Yu (National University of Singapore)
GenerationTransformerVision Language ModelVideo
🎯 What it does: A proactive learning framework for video subtitle generation tasks is proposed, focusing on learnability, diversity, and uncertainty, and introducing a caption-wise sampling protocol to utilize human annotations more efficiently.
Learnability of high-dimensional targets by two-parameter models and gradient flow
Dmitry Yarotsky (Skolkovo Institute of Science and Technology)
🎯 What it does: This paper proves that when the target dimension d is greater than the parameter dimension W (especially when W=2), it is still possible to learn high-dimensional targets almost surely under a given probability distribution through gradient flow (GF); it also provides a general conclusion on unlearnability.
Learning 1D Causal Visual Representation with De-focus Attention Networks
Chenxin Tao (Tsinghua University), Jifeng Dai (Tsinghua University)
ClassificationObject DetectionRetrievalTransformerImage
🎯 What it does: This paper proposes De-focus Attention Networks, which significantly enhance the performance of 1D causal visual models through techniques such as learnable bandpass filters, larger Drop Path linear scheduling, and global average feature auxiliary loss, making them competitive with 2D non-causal models in tasks like classification, detection, and retrieval.
Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation
Xin Hu (Xi'an Jiaotong University), Jian Sun (Xi'an Jiaotong University)
Pose EstimationRepresentation LearningTransformerPoint CloudMesh
🎯 What it does: This paper proposes a Pose-Invariant Implicit Function based on Local 3D Patches (PEIF), which normalizes poses and learns pose-invariant patch representations through a multi-head memory bank, directly predicting the displacement vectors of query points to achieve continuous reconstruction of 3D surfaces.
Learning 3D Garment Animation from Trajectories of A Piece of Cloth
Yidi Shao (Nanyang Technological University), Bo Dai (University of Hong Kong)
OptimizationGraph Neural NetworkContrastive LearningMesh
🎯 What it does: Using the motion trajectory of a single fabric to learn the constitutive relationships of clothing, and then animating various garments through energy optimization.
Learning a Single Neuron Robustly to Distributional Shifts and Adversarial Label Noise
Shuyao Li (University of Wisconsin Madison), Jelena Diakonikolas (University of Wisconsin Madison)
OptimizationComputational EfficiencyAdversarial AttackTabular
🎯 What it does: This study investigates the problem of learning a single neuron under adversarial distribution changes and label noise, with the goal of finding the optimal fitting function.
Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training
Haoran He (Hong Kong University of Science and Technology), Xuelong Li (Institute of Artificial Intelligence China Telecom)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerDiffusion modelVideo
🎯 What it does: This paper proposes a robot strategy learning framework based on a discrete diffusion model called VPDD, which utilizes large-scale unlabeled human videos for pre-training and fine-tunes with a small number of robot demonstrations, ultimately achieving multi-task visual control.
Learning and Transferring Sparse Contextual Bigrams with Linear Transformers
Yunwei Ren (Princeton University), Jason D. Lee (Princeton University)
TransformerSequentialOrdinary Differential Equation
🎯 What it does: Defines the Sparse Contextual Bigram (SCB) model and analyzes the training dynamics and sample complexity of learning this model using a single-layer linear Transformer through gradient methods;
Learning Better Representations From Less Data For Propositional Satisfiability
Mohamed Ghanem (CISPA Helmholtz Center for Information Security), Bernd Finkbeiner (CISPA Helmholtz Center for Information Security)
OptimizationRepresentation LearningGraph Neural NetworkTabular
🎯 What it does: A neural symbolic model called NeuRes has been developed for propositional satisfiability (SAT) problems, capable of parallelly generating verifiable unsatisfiability proofs or satisfying assignments, ensuring that the decision results are verifiable.
Learning Bregman Divergences with Application to Robustness
Mohamed-Hicham LEGHETTAS, Markus Püschel (ETH Zurich)
OptimizationAdversarial AttackImage
🎯 What it does: This paper proposes a method for learning Bregman divergence from raw high-dimensional pixel data, and utilizes the learned divergence for mirror descent adversarial training, significantly enhancing the model's robustness against common image corruptions (such as contrast, fog, blur, etc.).
Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification
Jiangming Shi (Xiamen University), Yanyun Qu (Xiamen University)
RecognitionRetrievalContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an advanced contrastive learning framework PCLHD for unsupervised visible-infrared person re-identification, combining hard prototypes and dynamic prototypes to explore the differences between identities and sample diversity.
Learning Complete Protein Representation by Dynamically Coupling of Sequence and Structure
Bozhen Hu (Zhejiang University), Stan Z. Li (Westlake University)
Representation LearningProtein Structure PredictionGraph Neural NetworkGraphBiomedical Data
🎯 What it does: We propose CoupleNet, a complete representation learning framework that constructs a dynamic dual graph between protein sequences and three-dimensional structures, performing convolution simultaneously on both nodes and edges.
Learning Cooperative Trajectory Representations for Motion Forecasting
Hongzhi Ruan (Tsinghua University), Zaiqing Nie (Tsinghua University)
Autonomous DrivingGraph Neural NetworkTransformerSupervised Fine-TuningGraphTime Series
🎯 What it does: The V2X-Graph framework is proposed to achieve graph-based collaborative trajectory feature fusion, enhancing motion prediction performance.
Learning Cortico-Muscular Dependence through Orthonormal Decomposition of Density Ratios
Shihan Ma (Imperial College London), Jose C Principe
ClassificationRecognitionMultimodalityBiomedical Data
🎯 What it does: Utilizing orthogonal normalized density ratio decomposition (FMCA) to learn multidimensional statistical dependencies from parallel EEG-EMG recordings, extracting feature projection spaces that capture movement types and subject identities;
Learning Cut Generating Functions for Integer Programming
Hongyu Cheng (Johns Hopkins University), Amitabh Basu (Johns Hopkins University)
OptimizationTabular
🎯 What it does: This paper proposes a data-driven method for selecting cutting planes, providing an upper bound on the sample complexity for choosing the best cutting plane generator function (CGF) family under a given instance distribution, and experimentally verifies its superiority over the traditional GMI cuts.
Learning De-Biased Representations for Remote-Sensing Imagery
Zichen Tian (Singapore Management University), Qianru Sun (Singapore Management University)
ClassificationObject DetectionContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised method called deb LoRA, which alleviates the feature bias of long-tail categories in remote sensing data by training low-rank adaptation layers on a pre-trained base model, thereby improving the performance of minority classes in classification and detection tasks.
Learning diffusion at lightspeed
Antonio Terpin (ETH Zurich), Florian Dorfler
OptimizationExplainability and InterpretabilityComputational EfficiencyDiffusion modelBiomedical Data
🎯 What it does: A JKOnet* model based on the first-order optimal condition of JKO is proposed to learn the energy function of diffusion processes from population data.
Learning Diffusion Priors from Observations by Expectation Maximization
François Rozet (University of Liège), Gilles Louppe (University of Liège)
RestorationGenerationDiffusion modelImageBiomedical DataMagnetic Resonance ImagingStochastic Differential Equation
🎯 What it does: Using the Expectation-Maximization (EM) algorithm, a diffusion prior model capable of Bayesian inference is trained solely with incomplete and noisy observations.
Learning Discrete Concepts in Latent Hierarchical Models
Lingjing Kong (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
GenerationData SynthesisExplainability and InterpretabilityDiffusion modelGraph
🎯 What it does: This paper proposes transforming the concept learning framework into an identifiable problem of discrete hierarchical latent causal models, providing theoretical conditions and implementing the corresponding identification algorithm;
Learning Discrete Latent Variable Structures with Tensor Rank Conditions
Zhengming Chen (Guangdong University of Technology), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Tabular
🎯 What it does: A discrete latent variable structure learning method based on tensor rank conditions is proposed, which can identify the causal structure of latent variables without being constrained by linear assumptions.
Learning Disentangled Representations for Perceptual Point Cloud Quality Assessment via Mutual Information Minimization
Ziyu Shan (Shanghai Jiao Tong University), Yiling Xu (Shanghai Jiao Tong University)
Representation LearningTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposes the DisPA framework, which learns point cloud content and distortion features through a dual-branch approach, and achieves explicit decoupling by minimizing mutual information (MI);
Learning Distinguishable Trajectory Representation with Contrastive Loss
Tianxu Li (Nanjing University of Aeronautics and Astronautics), Yang Zhang (Nanjing University of Aeronautics and Astronautics)
Representation LearningReinforcement LearningContrastive LearningSequential
🎯 What it does: This paper proposes a trajectory representation method based on contrastive learning (CTR), which encourages the diversification of multi-agent systems by learning distinguishable trajectory representations.
Learning Distributions on Manifolds with Free-Form Flows
Peter Sorrenson (Heidelberg University), Ullrich Koethe
GenerationData SynthesisOptimizationFlow-based ModelPoint Cloud
🎯 What it does: A single-step generative model based on free-form flows (M-FFF) is proposed, which can directly learn distributions and sample on any Riemannian manifold.
Learning diverse causally emergent representations from time series data
David McSharry (Imperial College London), Pedro A. M. Mediano
Representation LearningConvolutional Neural NetworkRecurrent Neural NetworkContrastive LearningTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: An end-to-end differentiable model is proposed, utilizing the causal collapse measure from information theory to automatically discover macro-appearable variables in time series data.
Learning Elastic Costs to Shape Monge Displacements
Michal Klein (Apple), marco cuturi
OptimizationBiomedical Data
🎯 What it does: This paper proposes a numerical method for solving the Monge problem using elastic cost, and based on this, designs a bi-level loss function for learning the elastic regularization parameters (especially subspace regularization);
Learning Equilibria in Adversarial Team Markov Games: A Nonconvex-Hidden-Concave Min-Max Optimization Problem
Fivos Kalogiannis (University of California Irvine), Ioannis Panageas (University of California Irvine)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes an algorithm that can approximate Nash equilibrium in infinite-horizon adversarial team Markov games (ATMG) through independent learning using only trajectory samples and without communication, with both iteration and sample complexity being polynomial.
Learning Formal Mathematics From Intrinsic Motivation
Gabriel Poesia (Stanford University), Noah Goodman
TransformerLarge Language ModelReinforcement Learning
🎯 What it does: The MINIMO system is proposed, which utilizes a formal mathematical domain given only axioms to simultaneously learn to automatically generate challenging propositions (conjectures) and find their proofs, forming a self-reinforcing training loop.
Learning Frequency-Adapted Vision Foundation Model for Domain Generalized Semantic Segmentation
Qi Bi (University of Amsterdam), Yefeng Zheng (Westlake University)
SegmentationDomain AdaptationAutonomous DrivingImage
🎯 What it does: An improvement of domain generalization semantic segmentation for Vision Foundation Model using frequency adaptive learning methods.
Learning from higher-order correlations, efficiently: hypothesis tests, random features, and neural networks
Eszter Szekely, Sebastian Goldt (International School of Advanced Studies)
Tabular
🎯 What it does: This study investigates the role of higher-order covariances (fourth order and above) in neural network learning, proposing a whitened higher-order peak model and analyzing statistical and computational sample complexity through likelihood ratios and lower-order methods.
Learning from Highly Sparse Spatio-temporal Data
Leyan Deng (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
RestorationAnomaly DetectionGraph Neural NetworkTime Series
🎯 What it does: This paper proposes a one-step propagation and confidence refinement (OPCR) method for highly sparse spatiotemporal data, aimed at efficiently recovering missing information and enhancing the performance of subsequent tasks.
Learning from Noisy Labels via Conditional Distributionally Robust Optimization
Hui Guo (University of Western Ontario), Boyu Wang (University of Western Ontario)
ClassificationOptimizationSupervised Fine-TuningImage
🎯 What it does: A framework based on Conditional Distribution Robust Optimization (CDRO) is proposed, which minimizes extreme risk within a distance-constrained fuzzy set using estimated true label posteriors, and generates robust pseudo-labels through analytical solutions for each sample, ultimately resulting in a robust training method that adaptively adjusts Lagrange multipliers in a single update.
Learning from Offline Foundation Features with Tensor Augmentations
Emir Konuk (KTH Royal Institute of Technology), Kevin Smith (KTH Royal Institute of Technology)
ClassificationRecognitionComputational EfficiencyTransformerSupervised Fine-TuningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study proposes a training framework named LOFF-TA, which extracts features from the training set images using a large foundational model in a single pass and caches them, then trains a lightweight classifier on the cached features, utilizing tensor augmentation instead of traditional image augmentation.
Learning from Pattern Completion: Self-supervised Controllable Generation
Zhiqiang Chen (Beijing Academy of Artificial Intelligence), Shan Yu (Institute of Automation, Chinese Academy of Sciences)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelAuto EncoderImage
🎯 What it does: A self-supervised controllable generation framework SCG is proposed, which first trains a modular autoencoder through equivariant constraints to enable different modules to specialize spontaneously; subsequently, these modules are used to complete missing information and achieve controlled generation.
Learning from Snapshots of Discrete and Continuous Data Streams
Pramith Devulapalli (Purdue University), Steve Hanneke (Purdue University)
🎯 What it does: Two novel continuous data stream learning frameworks (Update-and-Deploy and Blind-Prediction) are proposed, and their learnability is analyzed theoretically.
Learning from Teaching Regularization: Generalizable Correlations Should be Easy to Imitate
Can Jin (Rutgers University), Marco Pavone (Stanford University)
Knowledge DistillationConvolutional Neural NetworkRecurrent Neural NetworkTransformerReinforcement LearningImageText
🎯 What it does: By training the teacher model and the student model together, the student's ability to imitate the teacher is used as a regularization signal to enhance the model's generalization performance on unseen data.
Learning from Uncertain Data: From Possible Worlds to Possible Models
Jiongli Zhu (University of California), Babak Salimi (University of Illinois)
OptimizationTabular
🎯 What it does: Using possible world semantics and abstract interpretation, symbolic gradient descent is applied to uncertain training data to obtain a set of all possible linear models, providing model parameters and prediction intervals for inference.
Learning General Parameterized Policies for Infinite Horizon Average Reward Constrained MDPs via Primal-Dual Policy Gradient Algorithm
Qinbo Bai (Purdue University), Vaneet Aggarwal (Purdue University)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a primal-dual policy gradient algorithm for solving the infinite-horizon average reward constrained Markov decision process (CMDP) with a general parameterized policy, and provides a sublinear (˜O(T⁴⁵)) upper bound on expected reward loss and constraint violation.
Learning Generalized Linear Programming Value Functions
Tu Anh-Nguyen (Google Research), Christian Tjandraatmadja (Google Research)
Optimization
🎯 What it does: A theory-driven unsupervised learning method has been developed, using neural networks to approximate the Generalized Value Function (GVF), and applied for fast heuristic solving of two-stage mixed-integer linear programming.
Learning Goal-Conditioned Representations for Language Reward Models
Vaskar Nath (Scale AI), Sean M. Hendryx
Representation LearningReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningText
🎯 What it does: Introduce goal-based contrastive learning in reward model training to learn representations that can approximate the Q-function;
Learning Group Actions on Latent Representations
Yinzhu Jin (University of Virginia), Tom Fletcher
Image TranslationRestorationRepresentation LearningAuto EncoderImageMeshMagnetic Resonance Imaging
🎯 What it does: This paper proposes a method for learning group actions in the latent space of an autoencoder, achieving transformations on latent factors rather than directly in the data space.
Learning Human-like Representations to Enable Learning Human Values
Andrea Wynn, Thomas L. Griffiths (Princeton University)
Representation LearningReinforcement Learning from Human FeedbackText
🎯 What it does: This paper studies the impact of human representational alignment on the rapid, safe learning and generalization of human values by AI systems through the design of multi-armed bandit experiments and the collection of human morality and similarity data.
Learning Identifiable Factorized Causal Representations of Cellular Responses
Haiyi Mao (Genentech), Lin Qiu (Genentech)
Representation LearningAuto EncoderBiomedical Data
🎯 What it does: A recognizable causal decomposition representation (FCR) model is proposed, which can decompose the response of single cells to treatments into independent latent spaces of cellular background, the treatment itself, and their interactions.
Learning Image Priors Through Patch-Based Diffusion Models for Solving Inverse Problems
Jason Hu (University of Michigan), Jeffrey A Fessler
RestorationSuper ResolutionDiffusion modelScore-based ModelImageComputed TomographyStochastic Differential Equation
🎯 What it does: The paper proposes a strategy for training diffusion models using only image patches and positional encoding to construct a global image prior (PaDIS), and applies it to solve various inverse problems (CT reconstruction, deblurring, super-resolution).
Learning in Markov Games with Adaptive Adversaries: Policy Regret, Fundamental Barriers, and Efficient Algorithms
Thanh Nguyen-Tang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
Reinforcement Learning
🎯 What it does: This paper studies the learning problem in two-player Markov games when facing adaptive opponents, focusing on policy regret, clarifying the fundamental difficulties of learning, and proposing two efficient algorithms for different memory lengths.
Learning Infinitesimal Generators of Continuous Symmetries from Data
Gyeonghoon Ko (KAIST), Juho Lee (KAIST)
GenerationImageTime SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes a continuous symmetric generator based on neural ODEs that learns without prior knowledge and can automatically discover affine and non-affine symmetries from data.
Learning Interaction-aware 3D Gaussian Splatting for One-shot Hand Avatars
Xuan Huang (Shenzhen Campus of Sun Yat-Sen University), CHENQIANG GAO
GenerationData SynthesisPose EstimationTransformerGaussian SplattingImagePoint Cloud
🎯 What it does: A two-stage, interaction-aware 3D Gaussian point rendering framework is proposed to quickly construct animatable interactive hand avatars from a single image.
Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity
Jikai Jin (Stanford University), Vasilis Syrgkanis (Stanford University)
Representation LearningTabular
🎯 What it does: This paper studies causal representation learning under multiple environments and without hard interventions (single-node soft interventions or general environments), providing identifiability theory, the core confounding concept (SNA), and the implementable algorithm LiNGCReL.
Learning Low-Rank Feature for Thorax Disease Classification
Yancheng Wang (Arizona State University), Yingzhen Yang (Arizona State University)
ClassificationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: For the classification of diseases in chest X-ray images, a Low-Rank Feature Learning (LRFL) method is proposed.
Learning Macroscopic Dynamics from Partial Microscopic Observations
Mengyi Chen (National University of Singapore), Qianxiao Li (National University of Singapore)
Auto EncoderPhysics Related
🎯 What it does: A method is proposed to learn macroscopic dynamics by utilizing the sparsity assumption and only calculating the forces of a subset of microscopic coordinates.
Learning Mixtures of Unknown Causal Interventions
Abhinav Kumar (Massachusetts Institute of Technology Broad Institute of MIT and Harvard), Caroline Uhler (Massachusetts Institute of Technology Broad Institute of MIT and Harvard)
TabularBiomedical Data
🎯 What it does: This study investigates how to separate each intervention distribution from a mixed distribution containing observed data and unknown interventions (potential confounders) in linear Gaussian structural equation models (Linear-SEM) and subsequently identify the causal graph.
Learning Multimodal Behaviors from Scratch with Diffusion Policy Gradient
Zechu Li (Technical University of Darmstadt), Georgia Chalvatzaki (Technical University of Darmstadt)
Robotic IntelligenceReinforcement LearningDiffusion modelMultimodalityBenchmark
🎯 What it does: A multimodal behavior learning framework based on diffusion models is proposed for continuous control tasks starting from scratch, capable of mastering multiple feasible paths or strategies simultaneously.
Learning Neural Contracting Dynamics: Extended Linearization and Global Guarantees
Sean Jaffe (University of California), Francesco Bullo (University of California)
Flow-based ModelTime Series
🎯 What it does: This paper proposes and implements ELCD (Extended Linearized Contracting Dynamics), a neural network-based dynamic model that provides global convergence and convergence guarantees in arbitrary metric spaces.
Learning Noisy Halfspaces with a Margin: Massart is No Harder than Random
Gautam Chandrasekaran (University of Texas at Austin), Kevin Tian (University of Texas at Austin)
ClassificationOptimization
🎯 What it does: The study addresses the PAC learning problem of halfspaces and Generalized Linear Models (GLM) under the Massart noise model, proposing a new 'Perspectron' algorithm for effective learning.