ICML 2024 Papers — Page 13
International Conference on Machine Learning · 2610 papers
Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-constraint
Wei Xiong (University of Illinois Urbana-Champaign), Tong Zhang
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Formally model the RLHF problem as a context-free bandit problem with reverse KL regularization, and provide efficient algorithms with finite sample theoretical guarantees in offline, online, and hybrid settings.
Iterative Regularized Policy Optimization with Imperfect Demonstrations
Gong Xudong, Huaimin Wang (Flight Automatic Control Research Institute AVIC)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: Proposes an Iterative Regularization Policy Optimization (IRPO) framework that alternates between offline imitation learning and online reinforcement learning, using the policies obtained from online learning to generate new demonstrations, thereby continuously improving policy quality under imperfect demonstrations.
Iterative Search Attribution for Deep Neural Networks
Zhiyu Zhu (University of Sydney), Jun Shen (University of Wollongong)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and implements an Iterative Search Attribution method (ISA) based on gradient ascent and descent, which improves attribution accuracy by gradually pruning unimportant features and introducing scale factors.
IW-GAE: Importance weighted group accuracy estimation for improved calibration and model selection in unsupervised domain adaptation
Taejong Joo (Northwestern University), Diego Klabjan (Northwestern University)
Domain AdaptationImage
🎯 What it does: This paper proposes an approach to simultaneously improve model calibration and model selection in an Unsupervised Domain Adaptation (UDA) environment through Importance Weighted Group Accuracy Estimation (IW-GAE).
Jacobian Regularizer-based Neural Granger Causality
Wanqi Zhou (Xi'an Jiaotong University), Badong Chen (Xi'an Jiaotong University)
Recurrent Neural NetworkTime SeriesMagnetic Resonance Imaging
🎯 What it does: This paper proposes a neural Granger causality learning method based on input-output Jacobian matrix regularization, which can simultaneously estimate summary and full-time causal graphs within a single model.
Jetfire: Efficient and Accurate Transformer Pretraining with INT8 Data Flow and Per-Block Quantization
Haocheng Xi (Tsinghua University), Jun Zhu (Tsinghua University)
Computational EfficiencyTransformerImageText
🎯 What it does: Jetfire is proposed, a full quantization training method based on INT8 data streams and block-level quantization for Transformer pre-training.
Joint Composite Latent Space Bayesian Optimization
Natalie Maus (University of Pennsylvania), Eytan Bakshy (Meta)
OptimizationAdversarial Attack
🎯 What it does: The JoCo algorithm is proposed, which jointly trains a neural network encoder and a Gaussian process model in high-dimensional input/output composite Bayesian optimization, compressing high-dimensional space and achieving more efficient sampling.
Junk DNA Hypothesis: Pruning Small Pre-Trained Weights $\textit{Irreversibly}$ and $\textit{Monotonically}$ Impairs ``Difficult" Downstream Tasks in LLMs
Lu Yin (Eindhoven University of Technology), Zhangyang Wang (University of Texas at Austin)
TransformerLarge Language ModelText
🎯 What it does: This paper explores the impact of small weight magnitudes on the performance of downstream tasks of varying difficulty by applying one-time magnitude pruning, SparseGPT, Wanda, and N:M sparse methods to the pre-trained weights of large language models (LLMs), and comparing them with GPTQ quantization.
Just Cluster It: An Approach for Exploration in High-Dimensions using Clustering and Pre-Trained Representations
Stefan Sylvius Wagner (Heinrich Heine University), Stefan Harmeling (Technical University Dortmund)
Robotic IntelligenceReinforcement LearningPoint Cloud
🎯 What it does: A method called 'Just Cluster It' is proposed, which is based on episodic and global clustering. It estimates pseudo counts and generates intrinsic rewards through Gaussian mixture clustering of pre-trained or random features.
Keep the Momentum: Conservation Laws beyond Euclidean Gradient Flows
Sibylle Marcotte (ENS PSL University), Gabriel Peyré (CNRS)
OptimizationRecurrent Neural NetworkPhysics RelatedOrdinary Differential Equation
🎯 What it does: This study investigates the conservation laws in the training process of deep neural networks under momentum and non-Euclidean metrics, providing a complete theoretical representation and solution method.
Kepler codebook
Junrong Lian (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
RestorationGenerationSuper ResolutionTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This study proposes the Kepler Codebook, treating codebook training as a Kepler sphere packing problem. It designs Irwin-Hall distribution regularization and codebook partition strategies to address the issues of codebook collapse and low utilization in traditional VQ training, further achieving high-quality image reconstruction and generation.
Kernel Debiased Plug-in Estimation: Simultaneous, Automated Debiasing without Influence Functions for Many Target Parameters
Brian M Cho, Ivana Malenica (Harvard University)
🎯 What it does: A method called Kernel Debiased Plug-in Estimation (KDPE) is proposed to eliminate plug-in bias in nonparametric models and achieve efficient estimation without relying on influence functions.
Kernel Semi-Implicit Variational Inference
Ziheng Cheng (Peking University), Cheng Zhang (Peking University)
OptimizationTabularStochastic Differential Equation
🎯 What it does: Proposes Kernel Semi-Implicit Variational Inference (KSIVI), which explicitly solves the lower-level minimization problem of SIVI-SM in RKHS, obtaining the KSD objective and achieving semi-implicit variational inference without lower-level optimization.
Kernel-Based Evaluation of Conditional Biological Sequence Models
Pierre Glaser (Gatsby Computational Neuroscience Unit), Alan Nawzad Amin
Protein Structure PredictionSequentialBiomedical Data
🎯 What it does: A set of absolute evaluation metrics based on kernel functions (ACMMD and ACMMD-Rel) and corresponding hypothesis tests is proposed to measure the fit and reliability of conditional sequence models, thereby guiding hyperparameter tuning and model diagnostics.
KernelSHAP-IQ: Weighted Least Square Optimization for Shapley Interactions
Fabian Fumagalli (Bielefeld University), Barbara Hammer (Bielefeld University)
OptimizationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImageTextTabular
🎯 What it does: The KernelSHAP-IQ method is proposed for approximating high-order Shapley interactions (SII), treating SII as a solution to a weighted least squares (WLS) optimization problem.
KernelWarehouse: Rethinking the Design of Dynamic Convolution
Chao Li (Intel Labs), Anbang Yao (Intel Labs)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: KernelWarehouse is proposed, achieving large-scale dynamic convolution through kernel segmentation, shared warehouses, and contrast-driven attention, enhancing model expressiveness while maintaining parameter efficiency.
Keypoint-based Progressive Chain-of-Thought Distillation for LLMs
Kaituo Feng (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
Knowledge DistillationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: By using key point weighting and progressive chain reasoning distillation, the reasoning ability of large models is transferred to smaller models.
KISA: A Unified Keyframe Identifier and Skill Annotator for Long-Horizon Robotics Demonstrations
Longxin Kou (Tianjin University), Jianye HAO
Robotic IntelligenceTransformerVision Language ModelContrastive LearningVideo
🎯 What it does: Proposes the KISA framework, which can automatically, scalably, and semantically identify key frames from unlabeled demonstrations and assign corresponding skills to each key frame.
KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache
Zirui Liu (Rice University), Xia Hu (Rice University)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A two-bit KV cache quantization method called KIVI is proposed, which uses channel-wise quantization for the key cache and token-wise quantization for the value cache.
KnowFormer: Revisiting Transformers for Knowledge Graph Reasoning
Junnan Liu (Zhongguancun Laboratory), Jianxin Li (Zhongguancun Laboratory)
TransformerGraph
🎯 What it does: A Transformer structure named KNOWFORMER is proposed for knowledge graph reasoning.
Knowledge Distillation with Auxiliary Variable
Bo Peng (University of Technology Sydney), Jie Lu (University of Technology Sydney)
Object DetectionKnowledge DistillationContrastive LearningImage
🎯 What it does: A new knowledge distillation method called AuxKD is proposed, which enhances the student's ability to model the prediction distribution by rephrasing the student's prediction distribution through the introduction of label-related auxiliary variables (instance members);
Knowledge Graphs Can be Learned with Just Intersection Features
Duy Le (Case Western Reserve University), Zhaozhuo Xu (Stevens Institute of Technology)
Recommendation SystemOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes using the intersection features of k-hop neighborhoods of triples (head, relation, tail) in knowledge graphs for link prediction, and provides an efficient randomized estimation method.
Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models
Raviteja Vemulapalli (Apple), Oncel Tuzel (Apple)
Computational EfficiencyKnowledge DistillationSupervised Fine-TuningImage
🎯 What it does: A task-oriented knowledge transfer framework is proposed, which efficiently trains small task-specific models using the task knowledge from large-scale visual foundation models.
Knowledge-aware Reinforced Language Models for Protein Directed Evolution
Yuhao Wang (Zhejiang University), Huajun Chen (Zhejiang University)
Drug DiscoveryReinforcement LearningBiomedical Data
🎯 What it does: A protein directed evolution model KnowRLM based on amino acid knowledge graph and reinforcement learning is proposed, which can efficiently search and screen high-fitness mutants.
KV-Runahead: Scalable Causal LLM Inference by Parallel Key-Value Cache Generation
Minsik Cho (Apple), Devang Naik (Apple)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes KV-Runahead, a reasoning scheme that accelerates the prompt phase of LLM inference by generating key-value caches in parallel.
LAGMA: LAtent Goal-guided Multi-Agent Reinforcement Learning
Hyungho Na (Korea Advanced Institute of Science and Technology), Il-chul Moon
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: In collaborative multi-agent reinforcement learning, a method is proposed to accelerate learning by generating goal-achieving trajectories in the latent space and providing intrinsic rewards guided by potential goals.
LaMAGIC: Language-Model-based Topology Generation for Analog Integrated Circuits
Chen-Chia Chang (Duke University), Xin Zhang (IBM T. J. Watson Research Center)
GenerationData SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningGraph
🎯 What it does: Based on large language models (LM) through supervised fine-tuning, this method generates analog circuit topologies and duty cycles that meet specified voltage conversion ratios and efficiency targets in a single pass.
LangCell: Language-Cell Pre-training for Cell Identity Understanding
Suyuan Zhao (Tsinghua University), Zaiqing Nie (Tsinghua University)
ClassificationRecognitionData-Centric LearningTransformerContrastive LearningTextBiomedical Data
🎯 What it does: The LangCell framework is proposed, which combines single-cell transcriptome data with natural language descriptions for joint pre-training, achieving cross-modal understanding of cell identities.
Langevin Policy for Safe Reinforcement Learning
Fenghao Lei (Zhejiang University), Chaoyi Pang (NingboTech University)
Safty and PrivacyReinforcement LearningSequentialStochastic Differential Equation
🎯 What it does: Proposes Langevin Policy and Langevin Actor-Critic (LAC) for safe reinforcement learning, directly approximating the safe optimal policy through sampling and achieving efficient training.
Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models
Andy Zhou (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes the Language Agent Tree Search (LATS), which integrates Monte Carlo Tree Search (MCTS) into language models to achieve a unified framework for reasoning, execution, and planning, forming an interactive and self-reflective autonomous agent.
Language Agents with Reinforcement Learning for Strategic Play in the Werewolf Game
Zelai Xu (Tsinghua University), Yi Wu (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A strategic language agent that combines large language models with reinforcement learning for the social reasoning game Werewolf is proposed.
Language Generation with Strictly Proper Scoring Rules
Chenze Shao (Tencent Inc), Jie Zhou (Tencent Inc)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes using non-local strictly proper scoring rules (Brier score, Spherical score) to train language generation models, and provides implementation schemes for distributing scoring rules to the token level, smoothing, and masked log scores.
Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch
Le Yu (Alibaba Group), Yongbin Li (Alibaba Group)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: A DARE method is proposed, which achieves the removal of 90%–99% of parameters without training and without a GPU by randomly dropping and scaling the differential parameters, and can merge multi-task SFT models into a single multifunctional model.
Language Models as Science Tutors
Alexis Chevalier (Princeton University), Danqi Chen
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Designed and implemented the long-context scientific question answering evaluation benchmark TUTOREVAL and the dialogue dataset TUTORCHAT, and based on this, trained a specialized scientific and mathematical tutor-type language model;
Language Models as Semantic Indexers
Bowen Jin (University of Illinois), Xianfeng Tang (Amazon)
RetrievalRecommendation SystemTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Using self-supervised generative language models to learn semantic IDs of documents (semantic indexer)
Language Models Represent Beliefs of Self and Others
Wentao Zhu (Peking University), Yizhou Wang (Peking University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This study investigates how large language models internally represent self and other beliefs, and enhances their theory of mind (ToM) reasoning abilities through activation interventions.
Language Models with Conformal Factuality Guarantees
Christopher Mohri (Stanford University), Tatsunori Hashimoto (Stanford University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the 'Conformal Factuality' framework, which combines conformal prediction with correctness guarantees of language models, providing high-probability guarantees of factuality and correctness.
Language-Driven Cross-Modal Classifier for Zero-Shot Multi-Label Image Recognition
Yicheng Liu (Harbin Institute of Technology), Zheng Zhang (Harbin Institute of Technology)
ClassificationRecognitionTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A cross-modal classifier CoMC is proposed, which is trained only using language data for zero-shot multi-label image recognition.
Language-guided Skill Learning with Temporal Variational Inference
Haotian Fu (Brown University), Xingdi Yuan (Microsoft Research)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMultimodalitySequential
🎯 What it does: The LAST method is proposed to automatically discover reusable skills from expert demonstrations. It first uses a large language model to perform initial segmentation and language labeling of trajectories, and then integrates these segments through temporal variational inference and an MDL-assisted objective to learn skills and low-level policies.
Large Language Models are Geographically Biased
Rohin Manvi (Stanford University), Stefano Ermon (Stanford University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Using zero-shot prompts to enable large language models (LLMs) to predict numerical/evaluative data on various topics across different global locations, thereby assessing their geographic bias.
Large Language Models Can Automatically Engineer Features for Few-Shot Tabular Learning
Sungwon Han (Korea Advanced Institute of Science and Technology), Tomas Pfister (Google Cloud AI)
ClassificationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: Proposes the FeatLLM framework, which utilizes large language models to generate rule-based features, and then employs a simple linear model for few-shot table learning.
Large Scale Dataset Distillation with Domain Shift
Noel Loo (Massachusetts Institute of Technology), Daniela Rus
Domain AdaptationKnowledge DistillationImage
🎯 What it does: A scalable large-scale dataset distillation algorithm D3S is proposed, reframing the dataset distillation problem as a domain shift problem, and optimizing the distribution and labels of the distilled data by approximating KL divergence;
Larimar: Large Language Models with Episodic Memory Control
Payel Das (IBM AI Research), Pin-Yu Chen
GenerationTransformerLarge Language ModelText
🎯 What it does: A memory module called Larimar has been designed, which allows for external writing and reading, enabling large language models to update and forget knowledge during the testing phase through one-time writing without the need for retraining.
LASER: Linear Compression in Wireless Distributed Optimization
Ashok Vardhan Makkuva (École Polytechnique Fédérale de Lausanne), Michael Gastpar (École Polytechnique Fédérale de Lausanne)
OptimizationConvolutional Neural NetworkTransformerLarge Language ModelImageText
🎯 What it does: The LASER algorithm is proposed, which performs low-rank linear compression and transmission of gradients over wireless noisy channels, enhancing the communication efficiency of distributed training.
Latent Logic Tree Extraction for Event Sequence Explanation from LLMs
Zitao Song (Nanyang Technological University), Shuang Li (Chinese University of Hong Kong)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningSequentialBiomedical DataElectronic Health Records
🎯 What it does: A framework for extracting interpretable logical trees based on LLM priors, called LaTee, is proposed for explaining and predicting event sequences.
Latent Noise Segmentation: How Neural Noise Leads to the Emergence of Segmentation and Grouping
Ben Lonnqvist (Ecole Polytechnique Federale de Lausanne), Michael Herzog
SegmentationAuto EncoderImage
🎯 What it does: A novel method for unsupervised image segmentation and perceptual grouping is proposed using the internal random noise of neural networks.
Latent Optimal Paths by Gumbel Propagation for Variational Bayesian Dynamic Programming
Xinlei Niu (Australian National University), Charles Patrick Martin (Australian National University)
GenerationOptimizationReinforcement LearningFlow-based ModelAuto EncoderTextAudio
🎯 What it does: This paper proposes a framework for probabilizing the classic optimal path problem (Gumbel propagation + Bayesian dynamic programming) and embeds the sparse optimal path as a latent variable into a variational autoencoder, forming BDP-VAE, achieving end-to-end structured latent learning.
Latent Space Symmetry Discovery
Jianke Yang (University of California San Diego), Rose Yu (University of California San Diego)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkTime Series
🎯 What it does: A generative model named LaLiGAN is proposed, which can automatically discover nonlinear symmetries from high-dimensional data and learn the corresponding linear latent space transformations.
Latent variable model for high-dimensional point process with structured missingness
Maksim Sinelnikov (Aalto University), Harri Lähdesmäki
Auto EncoderTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: A deep latent variable model capable of handling high-dimensional longitudinal data with structural missingness is proposed, combining Gaussian process (GP) priors and temporal point processes, introducing independent latent variables for observed values and missing masks, forming the LLSM and LLPPSM models.
Layer-Aware Analysis of Catastrophic Overfitting: Revealing the Pseudo-Robust Shortcut Dependency
Runqi Lin (University of Sydney), Tongliang Liu (University of Sydney)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper studies the phenomenon of catastrophic overfitting in single-step adversarial training, revealing the deformation of different network layers during the overfitting process, and proposes a Layer-Aware Adversarial Perturbation (LAP) method to prevent the formation of pseudo-robust shortcuts, thereby avoiding catastrophic overfitting.
LayerMerge: Neural Network Depth Compression through Layer Pruning and Merging
Jinuk Kim (Seoul National University), Hyun Oh Song (Seoul National University)
GenerationCompressionOptimizationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes a novel deep compression method called LayerMerge, which combines joint pruning and fusion of convolutional layers and activation layers to achieve minimal performance loss under target inference latency.
Layerwise Change of Knowledge in Neural Networks
Xu Cheng (Nanjing University of Science and Technology), Quanshi Zhang (Shanghai Jiao Tong University)
ClassificationConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper analyzes each layer of deep networks by introducing interaction primitives (AND/OR interactions), quantifying and tracking new interactions that emerge and old interactions that are forgotten during forward propagation, thereby revealing the hierarchical changes in network learning behavior, generalization ability, and representation stability.
Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning
Jinsoo Yoo, Geoff Pleiss (University of British Columbia)
OptimizationImage
🎯 What it does: Proposes Layerwise Proximal Replay (LPR), which combines experience replay with proximal point optimization in online continual learning to stabilize network updates.
LCA-on-the-Line: Benchmarking Out of Distribution Generalization with Class Taxonomies
Jia Shi (Carnegie Mellon University), Shu Kong (University of Macau)
Domain AdaptationPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes and validates the use of Lowest Common Ancestor (LCA) distance as a unified metric to predict model performance on various Out-of-Distribution (OOD) datasets from In-Distribution (ID) data, demonstrating its effectiveness for both Vision Models (VM) and Vision-Language Models (VLM).
LeaPformer: Enabling Linear Transformers for Autoregressive and Simultaneous Tasks via Learned Proportions
Victor Agostinelli III, Lizhong Chen (Oregon State University)
TransformerTextBenchmarkAudio
🎯 What it does: The Learned Proportions (LeaP) module and the linear Transformer variant LeaPformer based on this module are proposed, improving the position reweighting mechanism to be independent of specific sequence lengths, supporting autoregressive and parallel tasks.
Learning 1-Bit Tiny Object Detector with Discriminative Feature Refinement
Sheng Xu (Beihang University), Xiao Sun (Shanghai Artificial Intelligence Laboratory)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A foreground discriminative feature refinement method (DFR-Det) for 1-bit tiny object detection is proposed, which does not add extra burden during the inference phase.
Learning a Diffusion Model Policy from Rewards via Q-Score Matching
Michael Psenka (University of California), Yi Ma (University of California)
OptimizationReinforcement LearningDiffusion modelMultimodalityStochastic Differential Equation
🎯 What it does: An algorithm named Q-Score Matching (QSM) is proposed, which utilizes the scores of diffusion models and matches them with the action gradients of the Q-function to learn policies in continuous control tasks.
Learning Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking
Yongxin Li (Guilin University of Technology), Shuiwang Li (Guilin University of Technology)
Object TrackingTransformerContrastive LearningVideo
🎯 What it does: We propose AVTrack, an adaptive computation Vision Transformer framework for real-time tracking of drones, which learns viewpoint-invariant features through mutual information maximization.
Learning and Forgetting Unsafe Examples in Large Language Models
Jiachen Zhao (University of Massachusetts Amherst), Mengye Ren (New York University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates how publicly released large language models learn unsafe examples during custom downstream fine-tuning and how they forget these examples during subsequent safety fine-tuning. It proposes a ForgetFilter filtering method based on the forgetting rate to remove unsafe data and evaluates its effects on safety and downstream performance.
Learning Associative Memories with Gradient Descent
Vivien Cabannes (Meta), Alberto Bietti (Flatiron)
OptimizationTransformerSequential
🎯 What it does: This paper studies the dynamics of a single-layer associative memory model (storing input-output external products) under cross-entropy loss during gradient descent training, and transforms this process into an interacting particle system.
Learning Causal Domain-Invariant Temporal Dynamics for Few-Shot Action Recognition
Yuke Li (Boston College), Donglai Wei (Boston College)
ClassificationRecognitionRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: This paper proposes the CDTD (Causal Domain-Invariant Temporal Dynamics) framework, which separates time-invariant and variable information through unsupervised causal representation learning, thereby achieving few-shot action recognition.
Learning Causal Dynamics Models in Object-Oriented Environments
Zhongwei Yu (Institute of Automation Chinese Academy of Sciences), Dengpeng Xing (Institute of Automation Chinese Academy of Sciences)
Tabular
🎯 What it does: Proposes the Object-Oriented Causal Dynamics Model (OOCDM), which can learn causal dynamics models in large-scale objectified environments.
Learning Causal Relations from Subsampled Time Series with Two Time-Slices
Anpeng Wu (Zhejiang University), Fei Wu (Shanghai AI Laboratory)
Time Series
🎯 What it does: This study investigates how to construct a Descendant Hierarchical Topology (DHT) to recover a summary causal graph from an under-sampled time series with only two time slices through conditional independence testing.
Learning Cognitive Maps from Transformer Representations for Efficient Planning in Partially Observed Environments
Antoine Dedieu (Google DeepMind), Miguel Lazaro-Gredilla
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerReinforcement LearningText
🎯 What it does: Train a Transformer with a Discrete Bottleneck (TDB) in partially observable environments (POE) to construct interpretable cognitive maps from its quantized output indices for efficient path planning; simultaneously validated its interpretability and reasoning ability on text data.
Learning Constraints from Offline Demonstrations via Superior Distribution Correction Estimation
Guorui Quan (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes an offline inverse constraint reinforcement learning algorithm called ICSDICE, which can learn safety constraints and obtain constraint-satisfying policies solely from expert demonstration data.
Learning Coverage Paths in Unknown Environments with Deep Reinforcement Learning
Arvi Jonnarth (Linkoping University), Michael Felsberg (Linkoping University)
Robotic IntelligenceConvolutional Neural NetworkReinforcement Learning
🎯 What it does: An end-to-end deep reinforcement learning method is designed to achieve coverage path planning in unknown environments using continuous control signals.
Learning Decision Policies with Instrumental Variables through Double Machine Learning
Daqian Shao (University of Oxford), Marta Kwiatkowska (University of Oxford)
OptimizationReinforcement LearningTabular
🎯 What it does: Using instrumental variables (IV) and a double machine learning (DML) framework in offline data, a nonlinear IV regression method DML-IV is proposed for learning causal effects and deriving decision strategies.
Learning Decision Trees and Forests with Algorithmic Recourse
Kentaro Kanamori (Fujitsu Limited), Yuichi Ike (Kyushu University)
ClassificationOptimizationExplainability and InterpretabilityAdversarial AttackTabularFinance Related
🎯 What it does: This study investigates a method for learning accurate tree models and forest models while ensuring executable algorithmic regression.
Learning Divergence Fields for Shift-Robust Graph Representations
Qitian Wu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Domain AdaptationRepresentation LearningGraph Neural NetworkTransformerDiffusion modelGraphStochastic Differential Equation
🎯 What it does: A geometric diffusion model with a learnable divergence field is proposed to learn data with interdependencies, enhancing the robustness of graph representations under distribution shifts.
Learning Exceptional Subgroups by End-to-End Maximizing KL-Divergence
Sascha Xu (CISPA Helmholtz Center for Information Security), Jilles Vreeken (CISPA Helmholtz Center for Information Security)
Anomaly DetectionOptimizationExplainability and InterpretabilityFlow-based ModelTabular
🎯 What it does: Learning to identify anomalous subgroups and providing interpretable rule descriptions
Learning from Integral Losses in Physics Informed Neural Networks
Ehsan Saleh (University of Illinois Urbana-Champaign), Matthew West (University of Illinois Urbana-Champaign)
Reinforcement LearningTabularPhysics Related
🎯 What it does: For partial integral equations that include integral terms, this paper proposes three methods for training Physics-Informed Neural Networks (PINN): deterministic sampling, double-sampling technique, and delayed target method, with a regularization improvement on the latter;
Learning from Memory: Non-Parametric Memory Augmented Self-Supervised Learning of Visual Features
Thalles Silva (University of Campinas), Adín Ramírez Rivera (University of Oslo)
RetrievalRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A self-supervised visual feature learning framework called MaSSL is proposed, which utilizes representations of past experiences to compare with current views in order to learn robust visual features.
Learning from Streaming Data when Users Choose
Jinyan Su (Cornell University), Sarah Dean (Cornell University)
Recommendation SystemOptimizationTabular
🎯 What it does: A streaming data learning model for user selection based on preferences and bounded rationality on multi-service platforms has been established, and a decentralized algorithm MSGD relying solely on single-point gradient updates from selected users has been proposed, aiming to minimize overall user loss.
Learning from Students: Applying t-Distributions to Explore Accurate and Efficient Formats for LLMs
Jordan Dotzel (Cornell University), Zhiru Zhang (Cornell University)
Convolutional Neural NetworkTransformerLarge Language ModelText
🎯 What it does: This paper first conducts a distribution analysis of the weights and activations of 30 large language models and convolutional networks, finding that most conform to the Student's t-distribution. Based on this distribution, a new four-bit floating-point quantization format, Student Float (SF4), is derived, which improves accuracy by about 0.7% compared to the existing NF4 across various LLMs. Furthermore, 'super-normal' support (SR, SP) is added to non-integer formats such as E2M1 and APoT, further enhancing precision and reducing hardware area consumption. Finally, a precision-area Pareto curve is drawn to illustrate the trade-offs of different data types.
Learning Graph Representation via Graph Entropy Maximization
Ziheng Sun (Chinese University of Hong Kong), Jicong Fan (Chinese University of Hong Kong)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Proposes the GeMax method, which learns representations of graphs and nodes by maximizing the entropy of the Korner graph;
Learning High-Frequency Functions Made Easy with Sinusoidal Positional Encoding
Chuanhao Sun (University of Edinburgh), Mahesh K. Marina (University of Edinburgh)
GenerationData SynthesisNeural Radiance FieldAudio
🎯 What it does: A new pose encoding method called SPE is proposed to efficiently learn high-frequency functions and replace traditional fixed Fourier features.
Learning High-Order Relationships of Brain Regions
Weikang Qiu (Yale University), Rex Ying (Yale University)
Biomedical DataMagnetic Resonance Imaging
🎯 What it does: Utilize fMRI signals to learn the higher-order relationships of brain regions and predict cognitive performance.
Learning in Deep Factor Graphs with Gaussian Belief Propagation
Seth Nabarro (Imperial College London), Andrew Davison
ClassificationRestorationGraph Neural NetworkGaussian SplattingImageVideo
🎯 What it does: This paper proposes the use of Gaussian Belief Propagation (GBP) for learning and inference on deep factor graphs, treating parameters, inputs, outputs, and other elements as random variables, with training and prediction completed through the same inference process.
Learning in Feature Spaces via Coupled Covariances: Asymmetric Kernel SVD and Nyström method
Qinghua Tao (KU Leuven), Johan Suykens
ClassificationOptimizationComputational EfficiencyGraph Neural NetworkGraphTabular
🎯 What it does: The paper proposes a method for learning asymmetric kernel heterogeneous SVD (KSVD) in the feature space using Coupled Covariance Feature (CCE) problems, and accelerates large-scale computations through the Nyström method.
Learning Iterative Reasoning through Energy Diffusion
Yilun Du (Massachusetts Institute of Technology), Joshua B. Tenenbaum
OptimizationDiffusion modelScore-based ModelGraphTabular
🎯 What it does: By learning the energy function, reasoning and planning tasks are formulated as energy minimization problems, and solved using an iterative method of energy diffusion;
Learning Label Shift Correction for Test-Agnostic Long-Tailed Recognition
Tong Wei (Southeast University), Min-Ling Zhang (Southeast University)
ClassificationRecognitionImage
🎯 What it does: A simple Label Shift Correction (LSC) method is proposed to enhance model generalization performance in long-tail recognition tasks where the label distribution is unknown during testing.
Learning Latent Dynamic Robust Representations for World Models
Ruixiang Sun (Beijing Institute of Technology), Riashat Islam (DreamFold AI)
Representation LearningRobotic IntelligenceReinforcement LearningWorld ModelVideo
🎯 What it does: Developed Hybrid-RSSM and introduced masked latent reconstruction and dual simulation principles to learn robust world models in visual environments with external noise.
Learning Latent Space Hierarchical EBM Diffusion Models
Jiali Cui (Stevens Institute of Technology), Tian Han (Stevens Institute of Technology)
GenerationData SynthesisDiffusion modelContrastive LearningImage
🎯 What it does: Proposes using a diffusion probability framework to learn energy-based models (EBM) on multi-layer generative models to address the prior hole problem.
Learning Latent Structures in Network Games via Data-Dependent Gated-Prior Graph Variational Autoencoders
Xue Yu (Renmin University of China), Renjie Liao (University of British Columbia)
Graph Neural NetworkTransformerAuto EncoderGraph
🎯 What it does: This paper proposes an unsupervised learning framework called GPGVAE, which infers the underlying network structure and interaction types (complementary or mutually exclusive strategies) from observed actions in online games.
Learning Linear Block Error Correction Codes
Yoni Choukroun (Tel Aviv University), Lior Wolf (Tel Aviv University)
Transformer
🎯 What it does: An end-to-end differentiable coding-decoding joint training framework is proposed, capable of simultaneously learning the generator matrix of binary linear block codes and the corresponding Transformer neural decoder.
Learning Low-dimensional Latent Dynamics from High-dimensional Observations: Non-asymptotics and Lower Bounds
Yuyang Zhang (Harvard University), Na Li (Harvard University)
OptimizationMeta LearningTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies how to learn low-dimensional latent dynamics from high-dimensional observations, proposing an algorithm to recover high-dimensional features, embed data into low dimensions, and learn low-dimensional model parameters.
Learning Mixtures of Gaussian Processes through Random Projection
Emmanuel Akeweje (Trinity College Dublin), Mimi Zhang (I-Form Advanced Manufacturing Research Centre)
Time Series
🎯 What it does: This paper proposes an ensemble clustering framework that first randomly projects functional data to one dimension and then fits a unidimensional Gaussian mixture model, aiming to efficiently learn high-dimensional Gaussian process mixture models.
Learning Modality Knowledge Alignment for Cross-Modality Transfer
Wenxuan Ma (Beijing Institute of Technology), Jingxuan Kang (University of Illinois Urbana-Champaign)
Domain AdaptationMeta LearningContrastive LearningImageVideoMultimodalityAudio
🎯 What it does: This paper studies the problem of source-target knowledge misalignment in cross-modal transfer and proposes a meta-learning-based modality knowledge alignment method (MoNA) that reduces modality knowledge differences by learning the target modality embedder before transfer.
Learning Multiple Secrets in Mastermind
Milind Prabhu (University of Michigan), David Woodruff (Carnegie Mellon University)
Point Cloud
🎯 What it does: A two-round adaptive algorithm is proposed to learn the hidden point set H, requiring at most exp( ˜ O ( √ d log n )) queries.
Learning Optimal Deterministic Policies with Stochastic Policy Gradients
Alessandro Montenegro (Politecnico di Milano), Matteo Papini (Politecnico di Milano)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This study investigates the policy gradient methods for action space and parameter space exploration using white noise in continuous control environments, and proves that they can globally converge to the optimal deterministic policy.
Learning Optimal Projection for Forecast Reconciliation of Hierarchical Time Series
Asterios Tsiourvas (Massachusetts Institute of Technology), Yada Zhu (IBM Research)
OptimizationTransformerTime Series
🎯 What it does: This paper proposes a learnable oblique projection method for collaborative integration in hierarchical time series forecasting, achieving end-to-end prediction and consistency correction.
Learning Pseudo-Contractive Denoisers for Inverse Problems
Deliang Wei (East China Normal University), Fang Li (East China Normal University)
RestorationSuper ResolutionImage
🎯 What it does: A pseudo-contractive deep denoiser is proposed, and a convergent plug-and-play recovery algorithm is designed based on this, achieving global convergence using Ishikawa iteration;
Learning Reward for Robot Skills Using Large Language Models via Self-Alignment
Yuwei Zeng (National University of Singapore), Lin Shao (National University of Singapore)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningMultimodalityChain-of-Thought
🎯 What it does: Utilize large language models (LLM) to learn the reward function for robotic skills through self-alignment;
Learning Scale-Aware Spatio-temporal Implicit Representation for Event-based Motion Deblurring
Wei Yu (Harbin Institute of Technology), Xiangyang Ji (Tsinghua University)
RestorationConvolutional Neural NetworkImageVideo
🎯 What it does: A Scale-Aware Spatio-Temporal Network (SASNet) is proposed to achieve event-driven motion deblurring at arbitrary spatial and temporal scales.
Learning Shadow Variable Representation for Treatment Effect Estimation under Collider Bias
Baohong Li (Zhejiang University), Kun Kuang (Zhejiang University)
Representation LearningAuto EncoderTabular
🎯 What it does: The study investigates how to estimate causal effects in the presence of collider bias in observational data and proposes a method for automatically learning shadow variable representations without prior knowledge.
Learning Solution-Aware Transformers for Efficiently Solving Quadratic Assignment Problem
Zhentao Tan (Peking University), Yadong MU
OptimizationTransformerReinforcement LearningTabular
🎯 What it does: A learning enhancement method based on Transformer, called SAWT, has been developed for efficiently solving the Quadratic Assignment Problem (QAP).
Learning Surrogates for Offline Black-Box Optimization via Gradient Matching
Minh Hoang (Princeton University), Trong Nghia Hoang (Washington State University)
OptimizationTabularBenchmark
🎯 What it does: This paper proposes a gradient matching-based offline black-box optimization surrogate model learning method (MATCH-OPT) and provides a theoretical proof of the upper bound of the surrogate gradient error and optimization performance loss.
Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Making
Vivek Myers (University of California), Benjamin Eysenbach (Princeton University)
Reinforcement LearningContrastive LearningSequential
🎯 What it does: This paper proposes a time distance metric that satisfies the triangle inequality in stochastic Markov decision processes, and implements this metric using successor features obtained through contrastive learning.
Learning the Target Network in Function Space
Kavosh Asadi (Amazon), Rasool Fakoor (Amazon)
Reinforcement LearningVideo
🎯 What it does: A new reinforcement learning value function approximation algorithm called Lookahead-Replicate (LR) is proposed, achieving equivalence between the online network and the target network in function space rather than parameter space.
Learning the Uncertainty Sets of Linear Control Systems via Set Membership: A Non-asymptotic Analysis
Yingying Li (University of Illinois Urbana-Champaign), Adam Wierman (California Institute of Technology)
OptimizationReinforcement LearningTime Series
🎯 What it does: Analyze the non-asymptotic convergence rate of the parameter uncertainty set for unknown linear systems, and propose a UCB algorithm based on SME.