NeurIPS 2023 Papers — Page 14
Conference on Neural Information Processing Systems · 3218 papers
Gradient Informed Proximal Policy Optimization
Sanghyun Son (University of Maryland), Ming Lin (University of Maryland)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a reinforcement learning method that combines the analytical gradient of differentiable environments with Proximal Policy Optimization (PPO) — GI-PPO. The core idea is to bridge the RP gradient and PPO updates through the α-policy and adaptively adjust α to control the gradient variance and bias.
Gradient-Based Feature Learning under Structured Data
Alireza Mousavi-Hosseini (University of Toronto and Vector Institute), Murat A Erdogdu
Tabular
🎯 What it does: This paper studies the sample complexity and convergence behavior of training a two-layer neural network to learn a single-index model using gradient flow when the input data covariance has a spike structure.
Gradient-Free Kernel Stein Discrepancy
Matthew A Fisher, Chris J. Oates
Time Series
🎯 What it does: A gradient-free kernel Stein divergence (GF-KSD) is proposed for posterior approximation.
GradOrth: A Simple yet Efficient Out-of-Distribution Detection with Orthogonal Projection of Gradients
Sima Behpour (Bosch Research North America), Liu Ren (Bosch Research North America)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A method for OOD detection based on pre-trained network gradient orthogonal projection, called GradOrth, is proposed, which uses the low-rank subspace projection gradient of ID data to determine whether a sample is OOD.
Grammar Prompting for Domain-Specific Language Generation with Large Language Models
Bailin Wang (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes a grammar prompting method that allows LLMs to first predict a minimized BNF grammar subset with a small number of examples, and then generate corresponding DSL programs according to that grammar.
GRAND-SLAMIN’ Interpretable Additive Modeling with Structural Constraints
Shibal Ibrahim (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: The GRAND-SLAMIN framework is proposed, which can learn interpretable additive models (GAM) with sparse interactions and hierarchical constraints in an end-to-end manner.
Granger Components Analysis: Unsupervised learning of latent temporal dependencies
Jacek Dmochowski
Time SeriesSequentialBiomedical DataMagnetic Resonance ImagingElectrocardiogram
🎯 What it does: An unsupervised separation method based on Granger causality (Granger Components Analysis, GCA) is proposed, which learns latent temporal dependency component pairs by maximizing the Granger causality strength between component pairs.
Graph Contrastive Learning with Stable and Scalable Spectral Encoding
Deyu Bo (Beijing University of Posts and Telecommunications), Chuan Shi (Singapore Management University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A graph contrastive learning framework Sp GCL that integrates spatial views and spectral views is proposed, along with a new spectral encoder EigenMLP;
Graph Convolutional Kernel Machine versus Graph Convolutional Networks
Zhihao Wu (Shenzhen Research Institute of Big Data), Jicong Fan (Chinese University of Hong Kong)
ClassificationOptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This paper proposes Graph Convolutional Kernel Machines (GCKM), which combines graph convolution with kernel methods for graph learning tasks.
Graph Denoising Diffusion for Inverse Protein Folding
Kai Yi (University of New South Wales), Yu Guang Wang (Shanghai Jiao Tong University)
GenerationProtein Structure PredictionGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: A graph denoising diffusion model (GRADE-IF) is proposed, which can generate diverse and foldable amino acid sequences under the condition of a given protein backbone.
Graph Mixture of Experts: Learning on Large-Scale Graphs with Explicit Diversity Modeling
Haotao Wang (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
Graph Neural NetworkMixture of ExpertsGraph
🎯 What it does: This paper proposes the Graph Mixture of Experts (GMoE) model, which allows each node to dynamically select different aggregation experts to better handle diverse graph structures.
Graph of Circuits with GNN for Exploring the Optimal Design Space
Aditya Hemant Shahane (Indian Institute of Technology Delhi), Sandeep Kumar (Indian Institute of Technology Delhi)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A semi-supervised graph model GCX based on graph neural networks is designed, integrating two optimization algorithms EASCO and ASTROG, to achieve simulation circuit parameter prediction and optimization with limited labeled data.
Graph-Structured Gaussian Processes for Transferable Graph Learning
Jun Wu (University of Illinois at Urbana-Champaign), Jingrui He (University of Illinois at Urbana-Champaign)
Graph Neural NetworkGraphAgriculture Related
🎯 What it does: A graph-structured Gaussian process framework called GraphGP is proposed for cross-graph transfer learning.
GraphAdapter: Tuning Vision-Language Models With Dual Knowledge Graph
Xin Li (University of Science and Technology of China), Xinchao Wang (National University of Singapore)
ClassificationDomain AdaptationGraph Neural NetworkTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a new adapter-based fine-tuning method called GraphAdapter, which integrates visual and textual structural knowledge through dual knowledge graphs to improve the transfer performance of few-shot vision-language models.
GraphMP: Graph Neural Network-based Motion Planning with Efficient Graph Search
Xiao Zang (Rutgers University), Bo Yuan (Rutgers University)
OptimizationRobotic IntelligenceGraph Neural NetworkGraph
🎯 What it does: A neural motion planner named GraphMP is proposed and implemented, which combines Graph Neural Networks (GNN) for graph pattern extraction and uses differentiable A* for graph search, achieving fast and high-quality path planning in low to high-dimensional (up to 14 dimensions) planning tasks.
GraphPatcher: Mitigating Degree Bias for Graph Neural Networks via Test-time Augmentation
Mingxuan Ju (University of Notre Dame), Yanfang Ye (University of Notre Dame)
Graph Neural NetworkGraph
🎯 What it does: A testing-time data augmentation framework called GRAPHPATCHER is proposed, which alleviates the degree bias of graph neural networks by dynamically generating virtual neighbors around low-degree nodes to fill in their sparse neighborhoods.
Grassmann Manifold Flows for Stable Shape Generation
Ryoma Yataka (Mitsubishi Electric Corporation), Masashi Shiraishi (Mitsubishi Electric Corporation)
GenerationData SynthesisFlow-based ModelPoint CloudOrdinary Differential Equation
🎯 What it does: This paper proposes the construction of Continuous Normalizing Flows (GrCNF) on the Grassmann manifold for generating stable shape data that is invariant to rotation and reflection.
Greatness in Simplicity: Unified Self-Cycle Consistency for Parser-Free Virtual Try-On
Chenghu Du (Wuhan University of Technology), Shengwu Xiong (Wuhan University of Technology)
Image TranslationGenerationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a parser-free virtual try-on network (USC-PFN) that utilizes self-consistency and a single generator to achieve high-quality clothing fitting and changing effects.
Greedy Poisson Rejection Sampling
Gergely Flamich (University of Cambridge)
OptimizationComputational EfficiencyTabular
🎯 What it does: Proposes the Greedy Poisson Rejection Sampling (GPRS) algorithm to solve the single-sample channel simulation problem of one-dimensional unmodeled density ratios.
Greedy Pruning with Group Lasso Provably Generalizes for Matrix Sensing
Nived Rajaraman (University of California), Kannan Ramchandran (University of California)
Supervised Fine-Tuning
🎯 What it does: This paper studies the effectiveness of greedy pruning and fine-tuning frameworks in matrix sensing problems, proposing a training method through group Lasso regularization to obtain models suitable for greedy pruning.
Grounded Decoding: Guiding Text Generation with Grounded Models for Embodied Agents
Wenlong Huang (Stanford University), brian ichter
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: A Grounded Decoding method is proposed and implemented, which generates executable robot instruction sequences by combining environmental grounding functions during the autoregressive decoding process of LLMs, supporting long-term spatial-temporal tasks.
Grounding Neural Inference with Satisfiability Modulo Theories
Zifan Wang (Center for AI Safety), Matt Fredrikson (Carnegie Mellon University)
OptimizationTransformerImage
🎯 What it does: This paper proposes an SMTLayer, which embeds an SMT solver into a deep network as a pluggable layer, enabling forward solving of symbolic constraints and utilizing unsatisfied core or MaxSMT information for gradient updates during backpropagation, allowing direct use of symbolic knowledge during both training and inference phases.
Group Fairness in Peer Review
Haris Aziz (University of New South Wales Sydney), Nisarg Shah (University of Toronto)
TabularReview/Survey Paper
🎯 What it does: Proposes the introduction of the concept of 'core' in peer review to construct a review assignment scheme that satisfies fairness for all subgroups.
Group Robust Classification Without Any Group Information
Christos Tsirigotis (University of Montreal), Aaron Courville (University of Montreal)
ClassificationContrastive LearningImage
🎯 What it does: A complete method for achieving group robust classification without any group labels is proposed, utilizing a self-supervised pre-training network as a bias proxy, and requiring no explicit bias annotations during both training and validation phases.
Guarantees for Self-Play in Multiplayer Games via Polymatrix Decomposability
Revan MacQueen (University of Alberta), James R. Wright (University of Alberta)
Reinforcement LearningTabular
🎯 What it does: This paper studies the reliability of self-play learning strategies in multi-player games, proposing two structural features: 'polygon decomposability' and 'subgame stability' to provide performance guarantees for self-play strategies when facing new opponents.
Guide Your Agent with Adaptive Multimodal Rewards
Changyeon Kim (KAIST), Kimin Lee (KAIST)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningVision Language ModelTextMultimodality
🎯 What it does: Proposes the Adaptive Return-conditioned Policy (ARP), which calculates the visual and textual similarity as immediate rewards through a pre-trained multimodal encoder, and uses this reward to train a return-conditioned behavior cloning strategy to enhance generalization in unseen environments and unseen instructions.
Guiding Large Language Models via Directional Stimulus Prompting
Zekun Li (University of California Santa Barbara), Xifeng Yan (University of California Santa Barbara)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: A framework named Directional Stimulus Prompting (DSP) is proposed, which utilizes a small adjustable strategy model to generate instantiated 'directional stimulus' prompts, achieving fine-grained, instance-specific output guidance in black-box large language models (LLMs) without parameter tuning.
Guiding The Last Layer in Federated Learning with Pre-Trained Models
Gwen Legate (Concordia University), Eugene Belilovsky (Concordia University)
Federated LearningImageText
🎯 What it does: In federated learning, an efficient recent class mean head initialization for FedNCM is proposed by utilizing a pre-trained model to fine-tune only the head layer, and further constructing a two-stage FedNCM+FT process.
GUST: Combinatorial Generalization by Unsupervised Grouping with Neuronal Coherence
Hao Zheng (Tsinghua University), Rong Zhao (Tsinghua University)
Object DetectionRepresentation LearningSpiking Neural NetworkAuto EncoderImage
🎯 What it does: A GUST model is designed to achieve unsupervised visual object grouping using neuronal coherence, and it can achieve compositional generalization in scenes with different numbers of objects.
H-InDex: Visual Reinforcement Learning with Hand-Informed Representations for Dexterous Manipulation
Yanjie Ze (Shanghai Qi Zhi Institute), Huazhe Xu (Shanghai AI Lab)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A three-stage visual reinforcement learning framework H-InDex is proposed, utilizing human hand pose priors to enhance the learning efficiency of robotic dexterous manipulation.
H-nobs: Achieving Certified Fairness and Robustness in Distributed Learning on Heterogeneous Datasets
Guanqiang Zhou (George Mason University), Zhi Tian (George Mason University)
OptimizationFederated LearningTabular
🎯 What it does: This paper proposes the H-nobs framework, which combines the fairness-promoting objective function with robust aggregation based on norm-based screening, achieving a joint satisfaction of fairness and robustness in distributed learning.
H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models
Zhenyu Zhang (University of Texas at Austin), Beidi Chen (Carnegie Mellon University)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a KV cache eviction strategy based on Heavy-Hitters (H O), which significantly compresses the KV cache of LLMs by dynamically retaining a small number of words that contribute the most to attention and the most recent words.
H2RBox-v2: Incorporating Symmetry for Boosting Horizontal Box Supervised Oriented Object Detection
Yi Yu (Southeast University), Junchi Yan (Shanghai Jiao Tong University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposes H2RBox-v2, a weakly supervised detection framework that utilizes horizontal box annotations to learn rotation boxes through symmetry self-supervised learning.
H3T: Efficient Integration of Memory Optimization and Parallelism for Large-scale Transformer Training
Yuzhong Wang (Tsinghua University), Maosong Sun (Tsinghua University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes the H3T framework, which automatically integrates memory optimization and parallelization for high-throughput training of large Transformers.
Handling Data Heterogeneity via Architectural Design for Federated Visual Recognition
Sara Pieri (Mohamed Bin Zayed University of Artificial Intelligence), Hisham Cholakkal (Mohamed Bin Zayed University of Artificial Intelligence)
RecognitionFederated LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates the model heterogeneity in visual recognition within federated learning, systematically evaluating the performance of 19 state-of-the-art visual architectures across 4 federated datasets.
HAP: Structure-Aware Masked Image Modeling for Human-Centric Perception
Junkun Yuan (Zhejiang University), Jingdong Wang (Baidu VIS)
RecognitionPose EstimationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Introduce Masked Image Modeling (MIM) and combine it with human structural priors for human-centered pre-training.
Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discovery
Yuxin Wen (University of Maryland), Tom Goldstein (University of Maryland)
GenerationOptimizationTransformerPrompt EngineeringDiffusion modelText
🎯 What it does: This paper studies a discrete optimization algorithm using gradient projection (PEZ) to achieve automatic learning of hard prompts (readable text tokens) to control generative models.
Hardness of Low Rank Approximation of Entrywise Transformed Matrix Products
Tamas Sarlos (Google Research), Qiuyi Zhang
Transformer
🎯 What it does: This paper studies the computational difficulty of low-rank approximation on entry-level transformation matrices (such as kernel matrices and Transformer attention matrices), providing a lower bound based on the Strong Exponential Time Hypothesis and an upper bound for polynomial transformations.
Hardware Resilience Properties of Text-Guided Image Classifiers
Syed Talal Wasim (Mohamed bin Zayed University of AI), Gu-Yeon Wei (Harvard University)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: By initializing the classification layer with rich text descriptions generated by GPT-3 during training and the output of the CLIP text encoder, the robustness of the image classification model to instantaneous bit flips in hardware during deployment is significantly improved.
Harnessing Hard Mixed Samples with Decoupled Regularizer
Zicheng Liu (Zhejiang University), Stan Z. Li (Westlake University)
ClassificationData-Centric LearningSupervised Fine-TuningImage
🎯 What it does: A new Mixup training objective called Decoupled Mixup is proposed, which can mine discriminative features from hard mixed samples while maintaining smoothness.
Harnessing the power of choices in decision tree learning
Guy Blanc (Stanford University), Mo Tiwari (Stanford University)
ClassificationExplainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: This paper proposes a method to extend traditional greedy decision tree algorithms (such as ID3, C4.5, CART) to Topk, where at each node, instead of considering only the single best feature, it considers the k best features, thereby improving learning effectiveness while maintaining interpretability.
HASSOD: Hierarchical Adaptive Self-Supervised Object Detection
Shengcao Cao (University of Illinois at Urbana-Champaign), Yu-Xiong Wang (University of Illinois at Urbana-Champaign)
Object DetectionSegmentationContrastive LearningImage
🎯 What it does: A completely unsupervised object detection and segmentation method called HASSOD is proposed, which can automatically discover objects in images and understand their hierarchical structure.
Have it your way: Individualized Privacy Assignment for DP-SGD
Franziska Boenisch (CISPA Helmholtz Center for Information Security), Nicolas Papernot (University of Toronto)
OptimizationSafty and PrivacyConvolutional Neural NetworkSupervised Fine-TuningImageText
🎯 What it does: This study investigates the allocation of individualized privacy budgets for each data point in differential privacy stochastic gradient descent (DP-SGD), enabling the model to provide corresponding protection based on different users' privacy preferences.
HeadSculpt: Crafting 3D Head Avatars with Text
Xiao Han (University of Surrey), Kwan-Yee K. Wong (The University of Hong Kong)
GenerationDiffusion modelImage
🎯 What it does: Using a pre-trained text-to-image diffusion model (Stable Diffusion) and ControlNet, high-resolution 3D head avatars are generated from text descriptions, supporting fine-grained editing.
HEDNet: A Hierarchical Encoder-Decoder Network for 3D Object Detection in Point Clouds
Gang Zhang (Tsinghua University), Xiaolin Hu (Tsinghua University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkPoint Cloud
🎯 What it does: HEDNet is proposed, a hierarchical encoder-decoder network for 3D object detection in point clouds.
HiBug: On Human-Interpretable Model Debug
Muxi Chen (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelVision Language ModelDiffusion modelImage
🎯 What it does: An automated and interpretable model debugging framework called HiBug has been developed, which utilizes large pre-trained models to automatically generate task-related visual attributes, label images, and identify low-performance data slices to discover systematic errors in the model and conduct root cause analysis.
Hidden Poison: Machine Unlearning Enables Camouflaged Poisoning Attacks
Jimmy Z. Di (University of Waterloo), Ayush Sekhari (Massachusetts Institute of Technology)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A 'camouflaged poisoning attack' that can be triggered in a machine unlearning scenario is proposed, which causes the model to misclassify specific test points by adding a small number of poisoned samples and camouflaged samples to the training set without learning.
Hierarchical Adaptive Value Estimation for Multi-modal Visual Reinforcement Learning
Yangru Huang (Peking University), Yonghong Tian (Peking University)
Autonomous DrivingReinforcement LearningImageMultimodality
🎯 What it does: This paper proposes a Hierarchical Adaptive Value Estimation framework (HAVE) for multimodal visual reinforcement learning, which can dynamically estimate and allocate the contributions of different sensors such as RGB, event cameras, and depth, achieving collaborative decision-making with multimodal information.
Hierarchical clustering with dot products recovers hidden tree structure
Annie Gray (University of Bristol), Nick Whiteley (University of Bristol)
TabularFinance Related
🎯 What it does: This paper proposes a hierarchical clustering algorithm based on dot product similarity and proves within a general graph model framework that it can consistently recover the hidden tree structure of the data.
Hierarchical Decomposition of Prompt-Based Continual Learning: Rethinking Obscured Sub-optimality
Liyuan Wang (Tsinghua University), Jun Zhu (Tsinghua University)
ClassificationRepresentation LearningTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes HiDe-Prompt, which explicitly optimizes prompt parameters using hierarchical decomposition (within-task prediction, task-identity inference, task-adaptive prediction) to enhance continuous learning performance under self-supervised pre-training.
Hierarchical Gaussian Mixture based Task Generative Model for Robust Meta-Learning
Yizhou Zhang (University of Southern California), Yan Liu (University of Southern California)
Meta LearningImage
🎯 What it does: A Hierarchical Gaussian Mixture Task Generation Model (HTGM) is proposed to characterize multi-source task distributions and achieve new task detection in meta-learning.
Hierarchical Integration Diffusion Model for Realistic Image Deblurring
Zheng Chen (Shanghai Jiao Tong University), Xin Yuan (Shanghai Jiao Tong University)
RestorationTransformerDiffusion modelImage
🎯 What it does: Hierarchically fuse the prior features generated by the latent diffusion model with the Transformer regression model for image deblurring.
Hierarchical Multi-Agent Skill Discovery
Mingyu Yang (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Robotic IntelligenceTransformerReinforcement LearningAgentic AISequentialBenchmark
🎯 What it does: A hierarchical multi-agent skill discovery method (HMASD) is proposed, which simultaneously learns team skills and individual skills, and achieves skill coordination through a skill allocator.
Hierarchical Open-vocabulary Universal Image Segmentation
Xudong Wang (Berkeley AI Research), Trevor Darrell (Berkeley AI Research)
Object DetectionSegmentationTransformerVision Language ModelImage
🎯 What it does: A unified framework named HIPIE is proposed, capable of hierarchical, instance, semantic, panoptic, part, and referential segmentation in an open vocabulary manner, and achieving object detection.
Hierarchical Randomized Smoothing
Yan Scholten (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
ClassificationAdversarial AttackGraph Neural NetworkImageGraph
🎯 What it does: A Hierarchical Randomized Smoothing framework is proposed, which injects local noise into complex data that can be decomposed into multiple entities (such as image pixels or graph nodes) and provides corresponding robustness proofs.
Hierarchical Semi-Implicit Variational Inference with Application to Diffusion Model Acceleration
Longlin Yu (Peking University), Cheng Zhang (Peking University)
GenerationOptimizationComputational EfficiencyDiffusion modelImageStochastic Differential Equation
🎯 What it does: A hierarchical semi-implicit variational inference (HSIVI) method is proposed to construct stronger semi-implicit distributions in multi-layer structures and apply it to high-dimensional Bayesian inference and diffusion model acceleration.
Hierarchical VAEs provide a normative account of motion processing in the primate brain
Hadi Vafaii (University of Maryland), Daniel A. Butts (University of Maryland)
Data SynthesisRepresentation LearningAuto EncoderOptical FlowImageVideo
🎯 What it does: This study investigates the representation of hierarchical VAE in motion perception and its alignment with the nervous system, proposing a new synthetic data framework called ROFL.
Hierarchical Vector Quantized Transformer for Multi-class Unsupervised Anomaly Detection
Ruiying Lu (Xidian University), Ruimin Hu (Xidian University)
Anomaly DetectionTransformerMixture of ExpertsImage
🎯 What it does: A Hierarchical Vector Quantization Transformer (HVQ-Trans) for multi-class unsupervised anomaly detection is proposed, achieving efficient localization and detection of anomalies through discrete prototype reconstruction.
Hierarchically Gated Recurrent Neural Network for Sequence Modeling
Zhen Qin (Shanghai Artificial Intelligence Laboratory), Yiran Zhong (Shanghai Artificial Intelligence Laboratory)
Recurrent Neural NetworkTextSequential
🎯 What it does: A Hierarchical Gated Recurrent Network (HGRN) is proposed, which introduces a learnable lower bound for the forget gate in multi-layer RNNs, allowing lower layers to focus on short-term information while higher layers capture long-term dependencies.
High dimensional, tabular deep learning with an auxiliary knowledge graph
Camilo Ruiz (Stanford University), Jure Leskovec (Stanford University)
Tabular
🎯 What it does: The PLATO method is proposed, which infers the weights of the first layer of a multilayer perceptron (MLP) using an auxiliary knowledge graph that describes features, thereby achieving strong predictive performance in tabular data where the feature dimension is much larger than the sample size (d ≫ n).
High Precision Causal Model Evaluation with Conditional Randomization
Chao Ma (Microsoft Research), Cheng Zhang (Microsoft Research)
Tabular
🎯 What it does: A low-variance estimator (pairs estimator) for evaluating causal model errors in conditional random experiments is proposed, which offsets variance by using the same inverse probability weighting (IPW) estimator for model predictions and true effects.
High-dimensional Asymptotics of Denoising Autoencoders
Hugo Cui (École Polytechnique Fédérale de Lausanne), Lenka Zdeborova
RestorationAuto EncoderImage
🎯 What it does: This paper studies denoising autoencoders (DAE) for high-dimensional Gaussian mixture data and provides a closed-form expression for its testing mean squared error;
High-dimensional Contextual Bandit Problem without Sparsity
Junpei Komiyama (New York University), Masaaki Imaizumi (University of Tokyo)
Reinforcement LearningTabular
🎯 What it does: This study explores the high-dimensional linear contextual bandit problem, proposing a minimum norm interpolation estimator that does not rely on sparsity, and introduces the Explore-then-Commit (EtC) algorithm and the Adaptive Explore-then-Commit (AEtC) algorithm.
High-Fidelity Audio Compression with Improved RVQGAN
Rithesh Kumar (Descript Inc), Kundan Kumar (Descript Inc)
CompressionGenerative Adversarial NetworkAudio
🎯 What it does: An improved high-fidelity universal neural audio compression algorithm, Improved RVQGAN, is proposed, which compresses 44.1 kHz audio to 8 kbps, achieving approximately 90 times compression;
Higher-Order Uncoupled Dynamics Do Not Lead to Nash Equilibrium - Except When They Do
Sarah Asad Toonsi, Jeff S Shamma
Reinforcement Learning
🎯 What it does: This study investigates high-order uncoupled learning dynamics, proving the existence of high-order gradient play that can lead to the convergence of mixed strategy Nash equilibria in any multi-agent game, and demonstrates cases where convergence does not occur for certain games.
HiNeRV: Video Compression with Hierarchical Encoding-based Neural Representation
Ho Man Kwan (University of Bristol), David Bull (University of Bristol)
CompressionConvolutional Neural NetworkVideo
🎯 What it does: Proposes HiNeRV, a hierarchical encoding video compression framework based on implicit neural representations.
HIQL: Offline Goal-Conditioned RL with Latent States as Actions
Seohong Park (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes HIQL, a hierarchical offline goal-conditioned reinforcement learning method that learns high-level subgoal policies, low-level action policies, and subgoal representations by training a single target-conditioned value function.
History Filtering in Imperfect Information Games: Algorithms and Complexity
Christopher Solinas (University of Alberta), Michael Buro (University of Alberta)
Computational EfficiencyReinforcement LearningSequential
🎯 What it does: This paper studies the computational complexity of history filtering during depth-limited search through subgame decomposition in incomplete information games and proposes a feasible generation method.
Holistic Transfer: Towards Non-Disruptive Fine-Tuning with Partial Target Data
Cheng-Hao Tu (Ohio State University), Wei-Lun Chao (Ohio State University)
Domain AdaptationKnowledge DistillationSupervised Fine-TuningImage
🎯 What it does: This paper proposes a new transfer learning paradigm—Holistic Transfer (HT), aimed at maintaining and enhancing the performance of the source model across all categories in the target domain while only utilizing samples from partial categories in the target domain.
Homotopy-based training of NeuralODEs for accurate dynamics discovery
Joon-Hyuk Ko (Seoul National University), Wonho Jhe (Seoul National University)
Time SeriesOrdinary Differential Equation
🎯 What it does: A new training method based on synchronization and homotopy optimization is proposed to improve the training accuracy of Neural Ordinary Differential Equations (NeuralODEs) on long time series data.
Honesty Is the Best Policy: Defining and Mitigating AI Deception
Francis Rhys Ward (Imperial College London), Tom Everitt (DeepMind)
Graph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a formal definition of deception within the framework of Structural Causal Games (SCG) and designs a Path-Specific Objective (PSO) method for anti-deception through graphical standards, validating its effectiveness in reinforcement learning agents and language models.
Horospherical Decision Boundaries for Large Margin Classification in Hyperbolic Space
Xiran Fan (University of Florida), Baba C. Vemuri (University of Florida)
ClassificationOptimizationGraph
🎯 What it does: This paper proposes a hypersphere large margin classifier HoroSVM based on horosphere decision boundaries, and provides a closed-form distance formula and optimization framework in the Poincaré ball model.
HotBEV: Hardware-oriented Transformer-based Multi-View 3D Detector for BEV Perception
Peiyan Dong (Northeastern University), Yanzhi Wang (Northeastern University)
Object DetectionAutonomous DrivingComputational EfficiencyNeural Architecture SearchTransformerImage
🎯 What it does: This paper presents HotBEV, a hardware-oriented Transformer-based multi-view 3D detection framework, focusing on addressing the latency and efficiency issues of real-time BEV perception on onboard devices.
How a Student becomes a Teacher: learning and forgetting through Spectral methods
Lorenzo Giambagli (University of Florence), Duccio Fanelli (University of Florence)
OptimizationKnowledge DistillationTabular
🎯 What it does: A training method for neural networks based on spectral parameterization is proposed, which achieves network learning by directly optimizing the feature vectors and eigenvalues on the weight matrix, and uses spectral regularization to filter the most important nodes.
How do Minimum-Norm Shallow Denoisers Look in Function Space?
Chen Zeno (Technion), Daniel Soudry (Technion)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the minimum representation cost solution of shallow ReLU networks as denoisers under low noise, deriving a closed-form expression in one dimension and proving its superiority over empirical MMSE.
How Does Adaptive Optimization Impact Local Neural Network Geometry?
Kaiqi Jiang (Princeton University), Yuanzhi Li (Carnegie Mellon University)
OptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This study investigates the impact of adaptive optimizers (such as Adam) on the local Hessian diagonal geometry during the training of deep networks and proposes the statistic R_OPT_med to measure the uniformity of the Hessian diagonal; experiments on the Transformer language model demonstrate that Adam's optimization trajectory tends to converge towards regions with a more uniform diagonal; theoretical analysis on a simplified two-layer linear network shows that Adam maintains R ≈ 1, while R for SGD+M increases over time.
How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model
Michael Hanna (University of Amsterdam), Alexandre Variengien (Redwood Research)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This study investigates how GPT-2 small achieves greater-than comparisons in the 'year span prediction' task and locates its internal computational circuits.
How many samples are needed to leverage smoothness?
Vivien Cabannes (Meta AI), Stefano Vigogna (University of Rome Tor Vergata)
Tabular
🎯 What it does: This paper clarifies through theoretical and numerical experiments that while smoothness can alleviate the curse of dimensionality, insufficient samples make it difficult to estimate higher-order derivatives, resulting in actual generalization errors that are much higher than the classical convergence rate.
How Re-sampling Helps for Long-Tail Learning?
Jiang-Xin Shi (Nanjing University), Yu-Feng Li (Nanjing University)
ClassificationData-Centric LearningImageBenchmark
🎯 What it does: This study investigates the effectiveness of resampling methods in long-tail learning and proposes a Context-Shift Augmentation (CSA) module to alleviate irrelevant context overfitting caused by resampling, thereby enhancing the generalization performance of single-stage long-tail learning.
How to Fine-tune the Model: Unified Model Shift and Model Bias Policy Optimization
Hai Zhang (Tongji University), Chen Ye (Tongji University)
OptimizationReinforcement Learning
🎯 What it does: A new model-based reinforcement learning framework called USB-PO is proposed, which unifies the handling of model bias and model drift through model refinement during the training process.
How to Scale Your EMA
Dan Busbridge (Apple), Russell Webb
OptimizationConvolutional Neural NetworkTransformerImageStochastic Differential EquationAudio
🎯 What it does: A rule is proposed for exponentially scaling the momentum of the model's EMA (Exponential Moving Average) when changing the batch size, ensuring that the training dynamics remain consistent across different batch sizes.
How to Select Which Active Learning Strategy is Best Suited for Your Specific Problem and Budget
Guy Hacohen (Hebrew University of Jerusalem), Daphna Weinshall (Hebrew University of Jerusalem)
OptimizationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: The SelectAL method is proposed, which can automatically select the most suitable active learning (AL) strategy under a given budget, achieving optimal performance in both low-budget and high-budget scenarios.
How to Turn Your Knowledge Graph Embeddings into Generative Models
Lorenzo Loconte (University of Edinburgh), Antonio Vergari (University of Edinburgh)
GenerationData SynthesisOptimizationGraph Neural NetworkGenerative Adversarial NetworkGraph
🎯 What it does: Reinterpret the energy function of mainstream knowledge graph embedding models as probabilistic circuits (GeKC), and achieve a model that is generative, can be trained with maximum likelihood, allows for precise sampling, and meets logical constraints through activation restrictions or squaring.
How2comm: Communication-Efficient and Collaboration-Pragmatic Multi-Agent Perception
Dingkang Yang (Fudan University), Lihua Zhang (Fudan University)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: A multi-agent collaborative perception framework named How2comm is proposed, aiming to significantly reduce communication bandwidth consumption while maintaining perception performance.
HQA-Attack: Toward High Quality Black-Box Hard-Label Adversarial Attack on Text
Han Liu (Dalian University of Technology), Xianchao Zhang (Dalian University of Technology)
Adversarial AttackTransformerText
🎯 What it does: A high-quality attack framework HQA-Attack is proposed for black-box hard-label text attack scenarios, aiming to generate adversarial texts with high semantic similarity and low perturbation rate under limited query budgets.
HubRouter: Learning Global Routing via Hub Generation and Pin-hub Connection
Xingbo Du (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
GenerationOptimizationReinforcement LearningGenerative Adversarial NetworkTabular
🎯 What it does: A two-stage generative global routing framework called HubRouter is proposed, transforming global routing into a hub-pin connectivity problem.
HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face
Yongliang Shen (Zhejiang University), Yueting Zhuang (Zhejiang University)
TransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: Designed and implemented HuggingGPT, a system that uses a large language model as a controller to automatically plan tasks, select expert models from Hugging Face, execute multimodal tasks, and generate final responses through a language interface;
Human spatiotemporal pattern learning as probabilistic program synthesis
Tracey Mills, Samuel J Cheyette
Time SeriesSequential
🎯 What it does: This paper designs a two-dimensional spatiotemporal sequence learning experiment, recording participants' behavior in predicting point locations and comparing it with various statistical and program synthesis models (Bayesian Ridge, non-combinatorial GP, GP structure learning, and LoT program synthesis).
Human-Aligned Calibration for AI-Assisted Decision Making
Nina L. Corvelo Benz (Max Planck Institute for Software Systems), Manuel Gomez Rodriguez (Max Planck Institute for Software Systems)
Text
🎯 What it does: This paper proposes a theoretical framework regarding the impact of confidence values in AI-assisted decision-making on the behavior of decision-makers. It demonstrates that relying solely on traditional probability calibration may lead decision-makers to make suboptimal choices and introduces the concept of 'human-aligned calibration.' Subsequently, it employs a multicalibration method to achieve this alignment and conducts experimental validation using real human-machine interaction data.
Human-Guided Complexity-Controlled Abstractions
Andi Peng (Massachusetts Institute of Technology), Julie Shah
ClassificationRepresentation LearningAuto EncoderImage
🎯 What it does: A framework for human-machine collaboration is proposed, allowing users to generate a series of discrete representations of varying complexity during pre-training based on task requirements, and to perform few-shot fine-tuning by selecting appropriate representations through human choice.
Human-in-the-Loop Optimization for Deep Stimulus Encoding in Visual Prostheses
Jacob Granley (University of California), Michael Beyeler (University of California)
OptimizationAuto EncoderImage
🎯 What it does: A human-computer interaction optimization framework has been developed, utilizing deep encoders and preference Bayesian optimization to achieve personalized stimulation coding for visual prosthetics.
Human-like Few-Shot Learning via Bayesian Reasoning over Natural Language
Kevin Ellis (Cornell University)
Explainability and InterpretabilityMeta LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes a few-shot concept learning model based on Bayesian inference, utilizing language models to generate natural language candidate hypotheses and fine-tuning priors with human data, making machine learning closer to human-like rapid and broad concept acquisition.
Hybrid Policy Optimization from Imperfect Demonstrations
Hanlin Yang (Sun Yat-sen University), Siji Chen (Sun Yat-sen University)
OptimizationKnowledge DistillationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: The HYPO algorithm is proposed to accelerate online learning in sparse reward environments using a small number of low-quality demonstrations.
Hybrid Search for Efficient Planning with Completeness Guarantees
Kalle Kujanpää (Aalto University), Alexander Ilin (Aalto University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a hybrid search framework (HIPSε) that combines high-level subgoal search with low-level action search, achieving completeness by inserting low-level actions into the subgoal search and further enhancing search efficiency.
HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution
Eric Nguyen (Stanford University), Stephen Baccus
TransformerPrompt EngineeringBiomedical DataBenchmark
🎯 What it does: A pre-trained model for genomic sequences at single nucleotide resolution, HyenaDNA, based on the Hyena operation, has been constructed, capable of processing up to 1 million bases of context at once, breaking through the length and resolution limitations of traditional Transformers in DNA sequence modeling.
HyP-NeRF: Learning Improved NeRF Priors using a HyperNetwork
Bipasha Sen (Massachusetts Institute of Technology), Srinath Sridhar (Brown University)
GenerationData SynthesisCompressionNeural Radiance FieldImage
🎯 What it does: This paper proposes HyP-NeRF, a method that utilizes hypernetworks to generate instance-specific multi-resolution hash encodings (MRHE) and MLP weights as a prior for NeRF, supporting high-quality NeRF generation under single-view, text, and cluttered scene conditions;
Hyper-HMM: aligning human brains and semantic features in a common latent event space
Caroline Lee, Christopher Baldassano (Columbia University)
TransformerVideoMultimodalityTime SeriesMagnetic Resonance Imaging
🎯 What it does: Proposes the Hyper-HMM model, which jointly aligns the spatiotemporal signals of fMRI from multiple subjects with the semantic embeddings of stimuli, learning a shared low-dimensional event space, and maps and clusters brain signals with semantic features in this space.
Hyperbolic Graph Neural Networks at Scale: A Meta Learning Approach
Nurendra Choudhary (Virginia Tech), Chandan K. Reddy (Virginia Tech)
Meta LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a scalable hyperbolic space graph neural network meta-learning framework H-GRAM, which learns transferable prior biases from local subgraphs and enables rapid few-shot learning on new subgraphs.
Hyperbolic Space with Hierarchical Margin Boosts Fine-Grained Learning from Coarse Labels
ShuLin Xu, Yi Yang (Zhejiang University)
ClassificationRecognitionRetrievalContrastive LearningImage
🎯 What it does: This paper proposes a method for learning fine-grained embeddings in hyperbolic space and achieves the transfer from coarse-grained labels to fine-grained recognition through hierarchical cosine margin constraints.
Hyperbolic VAE via Latent Gaussian Distributions
Seunghyuk Cho (POSTECH), Dongwoo Kim (POSTECH)
GenerationReinforcement LearningAuto EncoderImage
🎯 What it does: Designed and implemented a Gaussian distribution statistical manifold-based variational autoencoder (GM-VAE), and conducted experiments on two major tasks: image density estimation and model-based reinforcement learning.