ICML 2024 Papers — Page 9
International Conference on Machine Learning · 2610 papers
Exploiting Human-AI Dependence for Learning to Defer
Zixi Wei (Chongqing University), Lei Feng (Singapore University of Technology and Design)
ClassificationReinforcement LearningImage
🎯 What it does: A framework for delayed decision-making based on human-machine dependency relationships is proposed, along with a new definition of dependency Bayes optimality; based on this, a consistent DCE loss that can produce bounded confidence is designed.
Exploiting Negative Samples: A Catalyst for Cohort Discovery in Healthcare Analytics
Kaiping Zheng (National University of Singapore), James Wei Luen Yip (National University Heart Centre)
Auto EncoderBiomedical DataElectronic Health Records
🎯 What it does: This study investigates a method for medical cohort discovery using negative sample Shapley values, constructing a negative sample Shapley field and automatically discovering clinically significant groups through manifold learning and density clustering.
Exploration and Anti-Exploration with Distributional Random Network Distillation
Kai Yang (Tsinghua University), Xiu Li (Tsinghua University)
Reinforcement Learning
🎯 What it does: This paper studies and addresses the issue of inconsistent rewards in Random Network Distillation (RND), proposing Distributional Random Network Distillation (DRND). It distills multiple random target networks and utilizes pseudo-counts to achieve more accurate exploration rewards, applying this method to online PPO and offline SAC in reinforcement learning.
Exploration by Optimization with Hybrid Regularizers: Logarithmic Regret with Adversarial Robustness in Partial Monitoring
Taira Tsuchiya (University of Tokyo), Junya Honda (Kyoto University)
OptimizationAdversarial AttackReinforcement Learning
🎯 What it does: In the framework of partial monitoring with limited feedback, a Best of Both Worlds (BOBW) algorithm based on Hybrid Regularization is proposed, and asymptotic optimal or approximately optimal cumulative regret bounds are provided for both stochastic and adversarial environments.
Exploration-Driven Policy Optimization in RLHF: Theoretical Insights on Efficient Data Utilization
Yihan Du (University of Illinois Urbana-Champaign), R. Srikant (University of Illinois Urbana-Champaign)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: Two RLHF algorithms based on policy gradient (PG RLHF- and NN PG RLHF--) are proposed, targeting linear and neural network function approximation, respectively. They can actively collect human preference feedback and learn a reward model for subsequent policy optimization.
Explorations of Self-Repair in Language Models
Cody Rushing (University of Texas at Austin), Neel Nanda (University of Texas at Austin)
TransformerLarge Language ModelText
🎯 What it does: This paper studies the phenomenon of self-repair in large language models, systematically evaluating how the model compensates through subsequent components after the ablation of a single attention head under a complete pre-training distribution.
Exploring Correlations of Self-Supervised Tasks for Graphs
Taoran Fang (Zhejiang University), Yang Yang (Zhejiang University)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: This study investigates the interrelationships between graph self-supervised tasks and constructs the GraphTCM model based on task relevance to enhance representation learning.
Exploring Intrinsic Dimension for Vision-Language Model Pruning
Hanzhang Wang (Shanghai University), Qingyuan Ma (Shanghai University)
RetrievalCompressionVision Language ModelImageMultimodality
🎯 What it does: This paper proposes an intrinsic dimension (ID) estimation for multimodal pre-trained models and introduces an ID-based layer importance metric to guide the pruning of visual-language models.
Exploring the Benefit of Activation Sparsity in Pre-training
Zhengyan Zhang (Tsinghua University), Jie Zhou (Tencent)
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper studies the activation sparsity of pre-trained Transformers during the training process and proposes the Switchable Sparse-Dense Learning (SSD) framework, which dynamically switches between sparse (SMoE) and dense training during the pre-training phase to leverage the sparse activation characteristics to enhance training and inference efficiency.
Exploring the Complexity of Deep Neural Networks through Functional Equivalence
Guohao Shen (Hong Kong Polytechnic University)
Optimization
🎯 What it does: This study explores the complexity of deep neural networks from the perspective of functional equivalence, proposing a new upper bound on the covering number, revealing that the complexity of neural networks can be reduced, and that over-parameterized networks are easier to train.
Exploring the Enigma of Neural Dynamics Through A Scattering-Transform Mixer Landscape for Riemannian Manifold
Tingting Dan (University of North Carolina at Chapel Hill), Guorong Wu (University of North Carolina at Chapel Hill)
ClassificationRecognitionAnomaly DetectionTransformerSupervised Fine-TuningBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: This paper proposes a Riemannian manifold mixed model called DeepHoloBrain based on scattering transform, aimed at integrating the coupling of brain structure and functional networks for early disease diagnosis.
Exploring the LLM Journey from Cognition to Expression with Linear Representations
Yuzi Yan (Baichuan AI), Dong Yan (Tsinghua University)
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This study quantifies the cognitive and expressive abilities of large language models, revealing their different developmental paths during the pre-training, SFT, and RLHF stages.
Exploring the Low-Pass Filtering Behavior in Image Super-Resolution
Haoyu Deng (University of Electronic Science and Technology of China), Liang-Jian Deng (University of Electronic Science and Technology of China)
RestorationSuper ResolutionImage
🎯 What it does: This paper studies the low-pass filtering behavior of single-image super-resolution networks through signal processing theory, discovering that the impulse response approximates a sinc function, and proposes the HyRA analysis framework and the FSDS spectral similarity evaluation metric.
Exploring Training on Heterogeneous Data with Mixture of Low-rank Adapters
Yuhang Zhou (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
Domain AdaptationMixture of ExpertsImageBiomedical Data
🎯 What it does: This paper proposes the use of Mixture of Low-Rank Adapters (MoLA) to alleviate gradient conflicts in heterogeneous data training, and presents two variants: MoLA-Grad and MoLA-Router.
Exponential Spectral Pursuit: An Effective Initialization Method for Sparse Phase Retrieval
Mengchu Xu (Fudan University), Jian Wang (Fudan University)
OptimizationTabular
🎯 What it does: A new initialization method called Exponential Spectrum Tracking (ESP) is proposed for sparse phase recovery to improve the sampling complexity during the initialization phase.
Expressivity and Generalization: Fragment-Biases for Molecular GNNs
Tom Wollschläger (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: Proposes the Fragment-WL test to measure the expressive power of fragment-biased GNNs, and based on this, designs the FragNet model and a new infinite vocabulary fragmentation scheme.
Extending Test-Time Augmentation with Metamorphic Relations for Combinatorial Problems
Siwei Wei (Chinese Academy of Sciences), Yan Cai (Chinese Academy of Sciences)
OptimizationGraph Neural NetworkGraphTabular
🎯 What it does: This paper proposes the MAGG method, which enhances incremental reasoning of pre-trained machine learning models during the inference phase by utilizing the metamorphic relations of combinatorial problems;
Extracting Training Data From Document-Based VQA Models
Francesco Pinto (University of Oxford), Federico Tombari (Google)
Safty and PrivacyTransformerSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: Investigate the behavior of document visual question answering models in memorizing and extracting information from training data, particularly the risk of sensitive information leakage.
Extreme Compression of Large Language Models via Additive Quantization
Vage Egiazarian (HSE University), Dan Alistarh (IST Austria)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: The AQLM algorithm is proposed, which compresses LLM weights to extremely low bits (2-3 bits) based on Additive Quantization, balancing model accuracy and storage efficiency.
Factored-Reward Bandits with Intermediate Observations
Marco Mussi (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a new factored-reward bandit model (Factored-Reward Bandits), which decomposes each action into several sub-actions, observes the intermediate results of each sub-action, and takes their product as the final reward.
FADAS: Towards Federated Adaptive Asynchronous Optimization
Yujia Wang (Pennsylvania State University), Jinghui Chen (Pennsylvania State University)
OptimizationFederated LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: An asynchronous adaptive federated learning algorithm FADAS is proposed to address the slow client bottleneck caused by traditional synchronous aggregation, incorporating a delay-adaptive learning rate based on a maximum delay threshold.
FAFE: Immune Complex Modeling with Geodesic Distance Loss on Noisy Group Frames
Ruidong Wu (Helixon), Jian Peng (Tsinghua)
Protein Structure PredictionSupervised Fine-TuningBiomedical Data
🎯 What it does: A new geometric distance loss F2E is proposed to address the gradient vanishing problem caused by the FAPE loss in AlphaFold2-Multimer during immune complex modeling, further improved to a group-level F2E (G-F2E) to simultaneously consider rotational and translational errors.
Failures Are Fated, But Can Be Faded: Characterizing and Mitigating Unwanted Behaviors in Large-Scale Vision and Language Models
Som Sagar (Arizona State University), Ransalu Senanayake (Arizona State University)
OptimizationTransformerReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: A post-processing method based on deep reinforcement learning is proposed to explore and correct failure modes of pre-trained visual, language, and cross-modal models, achieving fine-grained improvements with a small amount of human feedback.
Fair Classification with Partial Feedback: An Exploration-Based Data Collection Approach
Vijay Keswani (Duke University), L. Elisa Celis (Yale University)
ClassificationOptimizationTabular
🎯 What it does: In response to the fair classification problem in partial feedback environments (where only the labels of positively predicted samples can be observed), an iterative exploration-exploitation framework is proposed. This framework can actively collect the true labels of unobserved samples while ensuring that each decision step has high utility and controlling the false positive rate, thereby improving overall prediction performance and fairness.
Fair Federated Learning via the Proportional Veto Core
Bhaskar Ray Chaudhury (University of Illinois Urbana-Champaign), Ariel D. Procaccia (Harvard University)
Federated LearningImage
🎯 What it does: A fair federated learning framework based on the Proportional Veto Core (PVC) is proposed, and a distributed algorithm called Rank-Core-Fed is designed, which can generate PVC stable models without requiring the convexity of the utility function.
Fair Off-Policy Learning from Observational Data
Dennis Frauen (Ludwig Maximilian University Munich), Stefan Feuerriegel (Ludwig Maximilian University Munich)
Reinforcement LearningTabularFinance Related
🎯 What it does: This paper presents FairPol, a fair offline policy learning framework implemented using neural networks, which learns optimal decision rules from observational data under different fairness constraints (action fairness and value fairness).
Fair Resource Allocation in Multi-Task Learning
Hao Ban (University at Buffalo), Kaiyi Ji (University at Buffalo)
OptimizationReinforcement LearningImage
🎯 What it does: A multi-task learning optimization framework called FairGrad based on α-fairness is proposed, applying the concept of α-fairness to existing multi-task learning methods to address gradient conflicts and loss scale imbalance issues; convergence proofs are also provided.
Fair Risk Control: A Generalized Framework for Calibrating Multi-group Fairness Risks
Lujing Zhang (Peking University), Linjun Zhang (Rutgers University)
Text
🎯 What it does: A general multi-dimensional multi-group fair risk control framework (s, G, α)-GMC is proposed, along with an algorithm based on projection updates.
FairProof : Confidential and Certifiable Fairness for Neural Networks
Chhavi Yadav (University of California), Kamalika Chaudhuri (University of California)
Safty and PrivacyTabular
🎯 What it does: Designed and implemented FairProof, a zero-knowledge proof-based mechanism that allows model owners to publicly verify the local individual fairness of neural networks at each input point while ensuring model confidentiality and uniformity.
Faithfulness Measurable Masked Language Models
Andreas Madsen (Mila), Sarath Chandar (Mila)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method for mask fine-tuning on pre-trained masked language models, enabling them to measure the credibility of importance metrics, thereby achieving verifiability of model explanations.
Fast Adversarial Attacks on Language Models In One GPU Minute
Vinu Sankar Sadasivan (University of Maryland), Soheil Feizi (University of Maryland)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper presents BEAST—a gradient-independent fast adversarial attack based on beam search, capable of generating adversarial suffixes in one minute on a single GPU, achieving jailbreaks, inducing hallucinations, and privacy leakage attacks on language models.
Fast Algorithms for Hypergraph PageRank with Applications to Semi-Supervised Learning
Konstantinos Ameranis (University of Chicago), Erasmo Tani (University of Chicago)
ClassificationOptimizationGraph Neural NetworkPoint CloudTabular
🎯 What it does: A fast PageRank and Laplacian system solving algorithm for hypergraphs is proposed, and it is applied to semi-supervised learning and hypergraph clustering.
Fast and Sample Efficient Multi-Task Representation Learning in Stochastic Contextual Bandits
Jiabin Lin (Iowa State University), Namrata Vaswani (Iowa State University)
OptimizationRepresentation LearningImage
🎯 What it does: A multi-task representation learning method based on Alternating Projection Gradient Descent (AltGDmin) and minimization is proposed and validated to address the stochastic contextual linear Bandit problem.
Fast Co-Training under Weak Dependence via Stream-Based Active Learning
Ilias Diakonikolas (University of Wisconsin Madison), Christos Tzamos (University of Athens and Archimedes AI)
Computational Efficiency
🎯 What it does: A black-box method is proposed to transform the problem of collaborative training under weak dependence into online learning, utilizing an active learning framework to achieve a collaborative training algorithm that balances label and computational efficiency.
Fast Decision Boundary based Out-of-Distribution Detector
Litian Liu (Massachusetts Institute of Technology), Yao Qin (University of California Santa Barbara)
Anomaly DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a fast OOD detection method based on the distance to the decision boundary in feature space, called fDBD.
Fast Peer Adaptation with Context-aware Exploration
Long Ma (Peking University), Yizhou Wang (Peking University)
Domain AdaptationReinforcement LearningMultimodality
🎯 What it does: A framework named PACE is proposed, which utilizes context-aware exploration to achieve rapid adaptation to unknown peers.
Fast Sampling-Based Sketches for Tensors
William Joseph Swartworth (Carnegie Mellon University), David Woodruff (Carnegie Mellon University)
🎯 What it does: This paper proposes a new sampling-based sketch method that can quickly perform ℓ0 sampling and ℓ1 embedding on two-dimensional and three-dimensional tensors (especially rank-one tensors) in nearly linear time, providing the corresponding p-sample construction and offering complete theoretical analysis and experimental validation.
Fast Text-to-3D-Aware Face Generation and Manipulation via Direct Cross-modal Mapping and Geometric Regularization
Jinlu Zhang (Xiamen University), Rongrong Ji (Xiamen University)
GenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImageText
🎯 What it does: Proposes E3-FaceNet for rapid generation and editing of 3D faces from text.
Fast Timing-Conditioned Latent Audio Diffusion
Zach Evans (Stability AI), Jordi Pons (Stability AI)
GenerationData SynthesisComputational EfficiencyDiffusion modelAuto EncoderAudio
🎯 What it does: A time-conditioned latent audio diffusion model (Stable Audio) is proposed, capable of generating up to 95 seconds of 44.1kHz stereo music and sound effects from text and duration prompts, supporting variable length control.
Fast White-Box Adversarial Streaming Without a Random Oracle
Ying Feng (Carnegie Mellon University), David Woodruff
Computational EfficiencyAdversarial AttackLarge Language Model
🎯 What it does: A sparse recovery algorithm is designed under a white-box adversarial flow model that does not require a random oracle and can detect sparsity errors, suitable for stream, distributed, and matrix/tensor recovery tasks.
Fast-Slow Test-Time Adaptation for Online Vision-and-Language Navigation
Junyu Gao (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
Domain AdaptationOptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: A Fast-Slow Test-Time Adaptation (FSTTA) framework for online visual and language navigation (VLN) is proposed, which can update model parameters in real-time on unannotated test samples.
Fast, Scalable, Warm-Start Semidefinite Programming with Spectral Bundling and Sketching
Rico Angell (University of Massachusetts Amherst), Andrew McCallum (University of Massachusetts Amherst)
OptimizationTabular
🎯 What it does: A unified spectral bundling and matrix sampling method, USBS, has been designed and implemented to solve large-scale semidefinite programming (SDP) problems with equality and inequality constraints, supporting warm-start initialization.
Faster Adaptive Decentralized Learning Algorithms
Feihu Huang (Nanjing University of Aeronautics and Astronautics), Jianyu Zhao (Nanjing University of Aeronautics and Astronautics)
OptimizationConvolutional Neural NetworkImageTabular
🎯 What it does: This paper proposes two decentralized optimization algorithms based on adaptive matrices, momentum, and gradient tracking: AdaMDOS (for stochastic optimization) and AdaMDOF (for finite sums), and provides convergence analysis.
Faster Maximum Inner Product Search in High Dimensions
Mo Tiwari (Stanford University), Martin Jinye Zhang (Carnegie Mellon University)
Recommendation SystemOptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes BanditMIPS and its improved version BanditMIPSα, which addresses the high-dimensional maximum inner product search problem, achieving O(1) dimensional complexity through adaptive sampling using multi-armed bandits.
Faster Sampling via Stochastic Gradient Proximal Sampler
Xunpeng Huang (Hong Kong University of Science and Technology), Tong Zhang (University of Illinois Urbana-Champaign)
OptimizationTabularStochastic Differential Equation
🎯 What it does: A Stochastic Proximal Sampler (SPS) framework is proposed, which can achieve unbiased sampling using stochastic gradients under non-log-convex target distributions.
Faster Streaming and Scalable Algorithms for Finding Directed Dense Subgraphs in Large Graphs
Slobodan Mitrović (University of California), Theodore Pan (University of California)
Graph Neural NetworkGraph
🎯 What it does: A directed dense subgraph algorithm has been designed that can achieve a (2+ε) approximation in a semi-streaming and MPC environment with a single pass through random streams or large-scale parallel computation, using only O(n polylog n) memory, and implementing O(1) or O(√log n) rounds of parallel algorithms under superlinear/near-linear memory.
Fault Tolerant ML: Efficient Meta-Aggregation and Synchronous Training
Tehila Dahan (Technion), Kfir Yehuda Levy
OptimizationFederated LearningAdversarial AttackMeta LearningImage
🎯 What it does: This paper proposes an efficient Centralized Trimming Meta-Aggregator (CTMA) and a dual momentum-based µ²-SGD algorithm to address the Byzantine robust training problem in distributed machine learning.
Feasibility Consistent Representation Learning for Safe Reinforcement Learning
Zhepeng Cen (Carnegie Mellon University), Ding Zhao (Carnegie Mellon University)
Safty and PrivacyRepresentation LearningReinforcement LearningContrastive LearningImageTabular
🎯 What it does: This paper proposes a feasibility consistency-based representation learning framework (FCSRL) that extracts safety-related information from the raw state through self-supervised learning, further enhancing the performance of safe reinforcement learning.
Feasible Reachable Policy Iteration
Shentao Qin (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Robotic IntelligenceReinforcement LearningTabular
🎯 What it does: A feasible reachable function (FR function) and a feasible reachable policy iteration (FRPI) algorithm based on this function are proposed to solve reinforcement learning problems with safety constraints and goal-reaching tasks; the algorithm improves sample efficiency through three steps: region identification, region expansion, and policy improvement.
Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation
Xuexin Chen (Guangdong University of Technology), José Miguel Hernández-Lobato (University of Cambridge)
Explainability and InterpretabilityGraph Neural NetworkImageGraph
🎯 What it does: The FANS method is proposed, which measures feature importance through probabilistic necessity and sufficiency (PNS) and achieves causal explanation through a two-stage perturbation test.
Feature Contamination: Neural Networks Learn Uncorrelated Features and Fail to Generalize
Tianren Zhang (Tsinghua University), Feng Chen (Tsinghua University)
ClassificationDomain AdaptationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper investigates the fundamental reasons for the failure of deep neural networks in out-of-distribution (OOD) generalization and proposes the mechanism of 'feature contamination';
Feature Distribution on Graph Topology Mediates the Effect of Graph Convolution: Homophily Perspective
Soo Yong Lee (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)
ClassificationRecommendation SystemGraph Neural NetworkGraph
🎯 What it does: This study investigates the impact of randomly shuffling feature vectors (Feature Shuffle) among nodes of the same category on the performance of Graph Neural Networks (GNNs). It proposes a category-controlled feature homogeneity metric (CFH) and a new random graph model CSBM-X, and theoretically and experimentally demonstrates the regulatory effect of CFH on graph convolution performance.
Feature Importance Disparities for Data Bias Investigations
Peter W Chang (Harvard Business School), Seth Neel (Harvard Business School)
OptimizationData-Centric LearningTabular
🎯 What it does: A data bias investigation method based on Feature Importance Difference (FID) is proposed, which can efficiently identify the differences in feature importance across different subgroups in a large-scale rich subgroup space, thereby locating potential sources of training data bias.
Feature Reuse and Scaling: Understanding Transfer Learning with Protein Language Models
Francesca-Zhoufan Li (California Institute of Technology), Alex Xijie Lu
Protein Structure PredictionConvolutional Neural NetworkTransformerLarge Language ModelBiomedical Data
🎯 What it does: This paper systematically evaluates the transfer learning effects of large-scale pre-trained protein language models (PLMs) on various protein function and structure prediction tasks, and explores the impact of different factors such as model size, depth, pre-training stage, and random initialization through 370 experiments.
FedBAT: Communication-Efficient Federated Learning via Learnable Binarization
Shiwei Li (Huazhong University of Science and Technology), Ruixuan Li (FiT Tencent)
Federated LearningComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The FedBAT framework is proposed in federated learning, which learns binary model updates directly during local training and adapts the step size, thus achieving efficient communication compression.
FedBPT: Efficient Federated Black-box Prompt Tuning for Large Language Models
Jingwei Sun (Duke University), Holger R Roth
OptimizationFederated LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The FedBPT framework is proposed to achieve black-box prompt tuning for large language models in federated learning, allowing clients to optimize prompts through forward inference without accessing model parameters.
FedCal: Achieving Local and Global Calibration in Federated Learning via Aggregated Parameterized Scaler
Hongyi Peng (Nanyang Technological University), Xiaoxiao Li (University of British Columbia)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: A federated learning framework named FedCal is proposed, which achieves local and global probability calibration by training local scalers on each client and aggregating them on the server to obtain a global scaler.
Federated Combinatorial Multi-Agent Multi-Armed Bandits
Fares Fourati (King Abdullah University of Science and Technology), Vaneet Aggarwal (Purdue University)
OptimizationFederated LearningImage
🎯 What it does: A general federated learning framework C-MA-MAB is proposed for the bandwidth-constrained multi-agent combinatorial multi-armed bandit problem (bandit feedback), which transforms any single-agent (α-ε) approximation algorithm into an online multi-agent algorithm, achieving sublinear α-regret and linear speedup.
Federated Continual Learning via Prompt-based Dual Knowledge Transfer
Hongming Piao (City University of Hong Kong), Ying Wei (Nanyang Technological University)
Federated LearningKnowledge DistillationTransformerPrompt EngineeringImage
🎯 What it does: Proposes Powder, a federated continual learning method that utilizes prompts for dual knowledge transfer.
Federated Full-Parameter Tuning of Billion-Sized Language Models with Communication Cost under 18 Kilobytes
Zhen Qin (Zhejiang University), Shuiguang Deng (Zhejiang University)
OptimizationFederated LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A federated full-parameter fine-tuning method named FedKSeed is proposed, utilizing zero-order optimization and limited seed reuse to achieve kilobyte-level communication costs, supporting full-parameter tuning of trillion-parameter language models on edge devices.
Federated Neuro-Symbolic Learning
Pengwei Xing (Nanyang Technological University), Han Yu (Nanyang Technological University)
Federated LearningTransformerText
🎯 What it does: A neural symbolic learning framework for federated learning, FedNSL, is proposed, utilizing rule distribution as a communication medium to address cross-domain rule heterogeneity issues.
Federated Offline Reinforcement Learning: Collaborative Single-Policy Coverage Suffices
Jiin Woo (Carnegie Mellon University), Yuejie Chi (Carnegie Mellon University)
Federated LearningReinforcement LearningTabular
🎯 What it does: A federated Q-learning algorithm for offline reinforcement learning, FedLCB-Q, is proposed, which enables multiple agents to collaboratively learn an approximately optimal policy without sharing offline datasets.
Federated Optimization with Doubly Regularized Drift Correction
Xiaowen Jiang (CISPA Helmholtz Center for Information Security), Sebastian U Stich (CISPA Helmholtz Center for Information Security)
OptimizationFederated LearningComputational EfficiencyTabular
🎯 What it does: Proposes two federated learning frameworks, DANE+ and FedRed, which achieve a significant reduction in communication volume through dual regularization drift correction while maintaining or improving local computational efficiency.
Federated Representation Learning in the Under-Parameterized Regime
Renpu Liu (Pennsylvania State University), Jing Yang (Pennsylvania State University)
Federated LearningRepresentation LearningImage
🎯 What it does: This paper proposes the FLUTE algorithm, specifically designed for the scenario of under-parameterized federated representation learning (FRL) on resource-constrained devices, addressing the issue of traditional FRL's inability to effectively aggregate local representations in under-parameterized environments.
Federated Self-Explaining GNNs with Anti-shortcut Augmentations
Linan Yue (University of Science and Technology of China), Fangzhou Yao (University of Science and Technology of China)
Federated LearningExplainability and InterpretabilityGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A federated graph reasoning method called FedGR is proposed, which combines two types of anti-spurious augmentations (complement-aware augmentation and difference-aware augmentation) to achieve self-explanatory graph neural networks.
FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees
Jiahao Liu (Sun Yat-sen University), Quan Z. Sheng (Macquarie University)
Federated LearningComputational EfficiencyKnowledge DistillationImage
🎯 What it does: Proposes two low-rank model training frameworks, FedLMT and pFedLMT, to address the issue of system heterogeneity in federated learning;
FedMBridge: Bridgeable Multimodal Federated Learning
Jiayi Chen (University of Virginia), Aidong Zhang (University of Virginia)
Federated LearningGraph Neural NetworkMultimodality
🎯 What it does: A multi-modal federated learning framework named FedMBridge is proposed to achieve implicit parameter sharing in scenarios where different clients have completely different network structures and fusion strategies (AMFL).
FedRC: Tackling Diverse Distribution Shifts Challenge in Federated Learning by Robust Clustering
Yongxin Guo (Chinese University of Hong Kong), Tao Lin (Westlake University)
OptimizationFederated LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes the FedRC scheme, which addresses the simultaneous issues of concept, label, and feature distribution shifts in federated learning using robust clustering principles.
FedREDefense: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction Error
Yueqi XIE, Neil Zhenqiang Gong (Duke University)
Federated LearningKnowledge DistillationAdversarial AttackImage
🎯 What it does: Proposes FedREDefense, which detects and filters malicious clients in federated learning by reconstructing model update errors to defend against model poisoning attacks.
FedSC: Provable Federated Self-supervised Learning with Spectral Contrastive Objective over Non-i.i.d. Data
Shusen Jing (University of California), Songyang Zhang (University of Louisiana)
Federated LearningRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes FedSC, a provably federated self-supervised learning framework based on spectral contrastive loss, which shares correlation matrices among clients and incorporates differential privacy protection to achieve cross-client comparison and enhance representation quality.
Feedback Efficient Online Fine-Tuning of Diffusion Models
Masatoshi Uehara (Genentech), Tommaso Biancalani (Genentech)
GenerationDrug DiscoveryReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageBiomedical Data
🎯 What it does: This paper proposes an online RL fine-tuning method named SEIKO, which utilizes a reward model and an uncertainty model to gradually guide the pre-trained diffusion model to efficiently explore the feasible sample space and obtain high-reward samples while maintaining KL regularization.
Feedback Loops With Language Models Drive In-Context Reward Hacking
Alexander Pan (University of California), Jacob Steinhardt (University of California)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This study investigates the feedback loops formed by large language models (LLMs) when interacting with the external world, demonstrating how these loops can lead to contextually induced reward harm (ICRH) — that is, negative side effects (such as toxicity or violations) arising while enhancing the agent's objectives (like Twitter engagement rates or bill payments).
Feel-Good Thompson Sampling for Contextual Dueling Bandits
Xuheng Li (University of California), Quanquan Gu (University of California)
Reinforcement LearningTabular
🎯 What it does: An algorithm named FGTS.CDB is proposed for the linear contextual bandit problem, which is based on Feel-Good Thompson sampling.
FESSNC: Fast Exponentially Stable and Safe Neural Controller
Jingdong Zhang (Fudan University), Wei Lin (Fudan University)
OptimizationSafty and PrivacyReinforcement LearningTime SeriesStochastic Differential Equation
🎯 What it does: A neural controller framework called FESSNC is designed to learn quickly and ensure exponential stability and safety guarantees when controlling stochastic differential equations.
Few-shot Adaptation to Distribution Shifts By Mixing Source and Target Embeddings
Yihao Xue (University of California), Baharan Mirzasoleiman (University of California)
Domain AdaptationTransformerImage
🎯 What it does: This paper proposes a few-shot distribution transfer method called MixPro, which achieves model adaptation in the target domain by linearly mixing the feature embeddings from the source and target domains before training a linear classifier.
Few-Shot Character Understanding in Movies as an Assessment to Meta-Learning of Theory-of-Mind
Mo Yu (Pattern Recognition Center WeChat AI), Jie Zhou
Meta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A few-shot character understanding dataset based on movie scripts, TOM-IN-AMC, was constructed to evaluate the machine's ability to quickly understand characters' theory of mind (ToM) in new movies.
Few-Shot Unsupervised Implicit Neural Shape Representation Learning with Spatial Adversaries
Amine Ouasfi (Inria), Adnane Boukhayma (Inria)
Representation LearningPoint Cloud
🎯 What it does: This paper proposes a method for unsupervised few-shot implicit neural shape representation learning on sparse unstructured point clouds, utilizing local adversarial query point pairs to regularize SDF learning.
Fewer Truncations Improve Language Modeling
Hantian Ding (Amazon Web Services AI Labs), Stefano Soatto (Amazon Web Services AI Labs)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: The Best-fit Packing method is proposed to reduce document truncation in the pre-training of large language models, improving data integrity.
FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning
Wenzhe Li (Princeton University), Chi Jin (Princeton University)
Reinforcement LearningVideoBenchmark
🎯 What it does: This paper proposes the FightLadder platform, providing a real-time fighting game benchmark for competitive multi-agent reinforcement learning, and implements various algorithms and evaluation metrics.
Finding NEM-U: Explaining unsupervised representation learning through neural network generated explanation masks
Bjørn Leth Møller (University of Copenhagen), Bulat Ibragimov (Norwegian Computing Center)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Proposed the NEM framework and implemented NEM-U, enabling the unsupervised representation learning model to generate occlusion explanation masks in parallel during inference by training a U-Net mask network.
Fine-Grained Causal Dynamics Learning with Quantization for Improving Robustness in Reinforcement Learning
Inwoo Hwang (Seoul National University), Sanghack Lee (Seoul National University)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: This paper proposes a fine-grained causal dynamics learning framework that utilizes vector quantization to partition the state-action space into subgroups and learns the corresponding local causal graphs for each subgroup, thereby achieving modeling and prediction of fine-grained causal relationships.
Fine-grained Classes and How to Find Them
Matej Grcic, Maria Brbic
ClassificationContrastive LearningImageBiomedical Data
🎯 What it does: A FALCON method is proposed, which automatically discovers fine-grained categories and learns fine-grained classifiers under unsupervised conditions using only coarse-grained labels.
Fine-grained Local Sensitivity Analysis of Standard Dot-Product Self-Attention
Aaron J Havens, Bin Hu (University of Illinois Urbana-Champaign)
OptimizationAdversarial AttackTransformerImage
🎯 What it does: This paper conducts a fine-grained theoretical analysis of the local sensitivity of the standard dot-product self-attention mechanism, proposing a new local sensitivity measure and providing a computable closed-form upper bound.
Fine-tuning Reinforcement Learning Models is Secretly a Forgetting Mitigation Problem
Maciej Wolczyk, Piotr Miłoś (deepsense.ai)
Reinforcement Learning
🎯 What it does: This paper discusses the forgetting problem (FPC) that arises when fine-tuning pre-trained models in reinforcement learning and proposes the use of knowledge retention techniques to alleviate this issue.
Finite Smoothing Algorithm for High-Dimensional Support Vector Machines and Quantile Regression
Qian Tang (University of Iowa), Boxiang Wang (University of Iowa)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a Finite Smoothing Algorithm (FSA) aimed at addressing the computational challenges of applying Support Vector Machines (SVM) and Quantile Regression in high-dimensional data.
Finite Time Logarithmic Regret Bounds for Self-Tuning Regulation
Rahul Singh (Indian Institute of Science), Panganamala Kumar
OptimizationReinforcement LearningTime Series
🎯 What it does: The PIECE adaptive minimum variance control algorithm is proposed, and it is proven to achieve a logarithmic order cumulative penalty upper bound for systems with bounded or sub-Gaussian noise within a finite time;
Finite Volume Features, Global Geometry Representations, and Residual Training for Deep Learning-based CFD Simulation
Loh Sher En Jessica (Nanyang Technological University), Adams Wai-Kin Kong
Graph Neural NetworkMeshGraph
🎯 What it does: This paper proposes a method based on finite volume features (FVF), global geometric representations (SV, DID), and residual training to enhance the accuracy of CFD simulations based on graph neural networks.
Finite-Time Convergence and Sample Complexity of Actor-Critic Multi-Objective Reinforcement Learning
Tianchen Zhou (Amazon), Yan Gao (Amazon)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: A multi-objective Actor-Critic algorithm named MOAC is proposed, which uses MGDA-style gradient updates to simultaneously optimize multiple conflicting rewards.
First-Order Manifold Data Augmentation for Regression Learning
Ilya Kaufman (Ben-Gurion University of the Negev), Omri Azencot (Ben-Gurion University of the Negev)
Data-Centric LearningRecurrent Neural NetworkTabularTime Series
🎯 What it does: A domain-independent data augmentation method for regression tasks, FOMA, is proposed, which generates new samples by scaling the small singular values in the singular value decomposition of training samples to simulate the tangent space of the data distribution.
FiT: Flexible Vision Transformer for Diffusion Model
Zeyu Lu (Shanghai Artificial Intelligence Laboratory), LEI BAI
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: A flexible visual Transformer (FiT) capable of image generation at arbitrary resolutions and aspect ratios is proposed and applied to diffusion models;
FlashST: A Simple and Universal Prompt-Tuning Framework for Traffic Prediction
Zhonghang Li (South China University of Technology), Chao Huang (University of Hong Kong)
Knowledge DistillationGraph Neural NetworkPrompt EngineeringTime Series
🎯 What it does: Proposes FlashST, a lightweight spatiotemporal Prompt-Tuning framework for quickly adapting pre-trained models to various urban traffic prediction tasks;
Flexible Residual Binarization for Image Super-Resolution
Yulun Zhang (Shanghai Jiao Tong University), Fisher Yu (ETH Zurich)
Super ResolutionCompressionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a flexible residual binarization method to compress image super-resolution networks to 1 bit while preserving high-frequency details.
Flextron: Many-in-One Flexible Large Language Model
Ruisi Cai (NVIDIA), Pavlo Molchanov (NVIDIA)
TransformerLarge Language ModelText
🎯 What it does: Proposes the FLEXTRON architecture and post-training optimization framework, which converts a trained LLM into a single model that can dynamically select sub-networks under different latency/memory targets without additional fine-tuning.
Floating Anchor Diffusion Model for Multi-motif Scaffolding
Ke Liu (Zhejiang University), Chunhua Shen (Ant Group)
Protein Structure PredictionDiffusion modelBiomedical DataBenchmark
🎯 What it does: The Floating Anchor Diffusion (FADiff) model is proposed for the design of multifunctional motif protein scaffolds, allowing motifs to maintain rigidity and float freely during the diffusion process, automatically determining their positions within the protein.
Flora: Low-Rank Adapters Are Secretly Gradient Compressors
Yongchang Hao (University of Alberta), Lili Mou (University of Alberta)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: The FLORA method is proposed, which achieves sublinear memory usage by compressing gradients and momentum through random projection, overcoming the low-rank limitations of LoRA;
FlowMM: Generating Materials with Riemannian Flow Matching
Benjamin Kurt Miller (University of Amsterdam), Brandon M Wood
GenerationData SynthesisGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: A continuous normalization flow model named FlowMM has been developed, utilizing Riemannian Flow Matching to generate periodic crystal structures and achieve two major tasks: crystal structure prediction (CSP) and novel material generation (DNG).
Fool Your (Vision and) Language Model with Embarrassingly Simple Permutations
Yongshuo Zong (University of Edinburgh), Timothy Hospedales (University of Edinburgh)
Adversarial AttackTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper reveals that both large language models and vision-language models are highly susceptible to attacks on multiple-choice question answers through simple perturbations of answer arrangements, with accuracy potentially dropping below random levels.
Forget Sharpness: Perturbed Forgetting of Model Biases Within SAM Dynamics
Ankit Vani (Mila), Hossein Sharifi-Noghabi (Borealis AI)
ClassificationTransformerImage
🎯 What it does: This paper proposes viewing the perturbation of SAM as a 'perturbed forgetting' mechanism to eliminate the model's useless biases, thereby explaining the generalization advantages of SAM;
Foundation Policies with Hilbert Representations
Seohong Park (University of California), Sergey Levine (University of California)
Reinforcement LearningPrompt Engineering
🎯 What it does: The research proposes a Hilbert space representation and direction-based reward multi-task policy (HILP) in offline unsupervised pre-training, achieving zero-shot adaptation for various downstream tasks.
Foundations of Testing for Finite-Sample Causal Discovery
Tom Yan (Carnegie Mellon University), Zachary Chase Lipton
Graph
🎯 What it does: A 'always effective' testing framework is proposed to implement the update steps of causal discovery under limited samples and soft interventions, and it is combined with a structured multi-constraint Bandit algorithm to complete causal validation.