ICML 2024 Papers — Page 6
International Conference on Machine Learning · 2610 papers
Critical windows: non-asymptotic theory for feature emergence in diffusion models
Marvin Li (Harvard University), Sitan Chen (Harvard University)
GenerationData SynthesisSafty and PrivacyAdversarial AttackDiffusion modelImageStochastic Differential Equation
🎯 What it does: A theory of the 'critical window' occurring during the reverse sampling process of diffusion models is proposed and proven, providing non-asymptotic upper and lower bounds for the appearance of this window, which is explained as a mechanism of hierarchical sampling.
CRoFT: Robust Fine-Tuning with Concurrent Optimization for OOD Generalization and Open-Set OOD Detection
Lin Zhu (Shanghai Jiao Tong University), Nanyang Ye (Shanghai Jiao Tong University)
ClassificationDomain AdaptationAnomaly DetectionTransformerSupervised Fine-TuningGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A lightweight joint fine-tuning framework CRoFT is proposed, which can maintain the pre-trained knowledge of CLIP while addressing OOD generalization and open-set OOD detection.
Cross-domain Open-world Discovery
Shuo Wen (Ecole Polytechnique Federale de Lausanne), Maria Brbic (Ecole Polytechnique Federale de Lausanne)
Domain AdaptationSupervised Fine-TuningContrastive LearningImage
🎯 What it does: A method named CROW is proposed for cross-domain open-world discovery tasks, which involves labeling samples with known categories and discovering unknown categories in the presence of domain shift and category shift.
Cross-Domain Policy Adaptation by Capturing Representation Mismatch
Jiafei Lyu (Tsinghua University), Xiu Li (Tsinghua University)
Domain AdaptationReinforcement LearningContrastive LearningTabular
🎯 What it does: A cross-domain strategy adaptation method based on representation mismatch is proposed, which quantifies the dynamic differences between the source domain and the target domain through state and state-action encoders trained in the target domain, using representation mismatch as a reward penalty to guide SAC learning.
Cross-view Masked Diffusion Transformers for Person Image Synthesis
Trung X. Pham (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisPose EstimationTransformerDiffusion modelImage
🎯 What it does: This paper proposes X-MDPT, a cross-view mask diffusion model based on Transformer for pose-guided portrait synthesis.
CrossGET: Cross-Guided Ensemble of Tokens for Accelerating Vision-Language Transformers
Dachuan Shi (Shanghai AI Laboratory), Jiaqi Wang (Shanghai AI Laboratory)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes a cross-modal guided Token aggregation framework called CrossGET, which can dynamically merge redundant Tokens in a visual-language Transformer, significantly reducing computational load and improving inference speed.
CRUXEval: A Benchmark for Code Reasoning, Understanding and Execution
Alex Gu (Massachusetts Institute of Technology), Sida Wang (Meta)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: The CRUXEval benchmark is proposed to evaluate the capabilities of code language models in code reasoning, understanding, and execution, providing 800 short Python function samples that include two main tasks: input prediction and output prediction.
Curated LLM: Synergy of LLMs and Data Curation for tabular augmentation in low-data regimes
Nabeel Seedat (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: Under low sample conditions (<100), high-quality data augmentation is achieved through the generation of tabular data using LLMs (such as GPT-4) combined with a learning dynamic-driven filtering mechanism.
CurBench: Curriculum Learning Benchmark
Yuwei Zhou (Tsinghua University), Wenwu Zhu (Tsinghua University)
ImageTextGraphBenchmark
🎯 What it does: This paper proposes and implements CurBench, a cross-domain curriculum learning benchmark covering 15 datasets, 9 benchmark models, and 3 training settings (standard, noisy, imbalanced) across computer vision, natural language processing, and graph structure learning;
CuTS: Customizable Tabular Synthetic Data Generation
Mark Vero (ETH Zurich), Martin Vechev (ETH Zurich)
Data SynthesisGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes CuTS, a framework for generating synthetic tabular data that supports various custom constraints (differential privacy, logical constraints, statistical constraints, and downstream model constraints);
CW Complex Hypothesis for Image Data
Yi Wang (Johns Hopkins University), Zhiren Wang (Pennsylvania State University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper studies the distribution structure of image data, challenging the traditional manifold assumption, and proposes that image data is better described using CW complexes (structures formed by gluing submanifolds of different dimensions along their boundaries). This hypothesis is supported through theoretical derivation, metric design, and experimental validation.
D-Flow: Differentiating through Flows for Controlled Generation
Heli Ben-Hamu (Weizmann Institute of Science), Yaron Lipman (Weizmann Institute of Science)
GenerationData SynthesisOptimizationDiffusion modelFlow-based ModelImageOrdinary Differential EquationAudio
🎯 What it does: A D-Flow framework is proposed, which optimizes source noise points by gradient descent on the ODE sampling process of a pre-trained diffusion/flow matching model, thereby achieving controlled generation for tasks such as inverse problems, conditional generation, and free-form inpainting.
DAG-Based Column Generation for Adversarial Team Games
Youzhi Zhang (Hong Kong Institute of Science and Innovation), Daniel Dajun Zeng (Institute of Automation, Chinese Academy of Sciences)
OptimizationTabular
🎯 What it does: This paper proposes a directed acyclic graph-based team belief column generation framework (DCG) for efficiently solving the team minimax equilibrium (TMECor) in adversarial team games.
Data Engineering for Scaling Language Models to 128K Context
Yao Fu (University of Edinburgh), Hao Peng (University of Illinois at Urbana-Champaign)
TransformerLarge Language ModelText
🎯 What it does: By performing full attention continuous pre-training on 7B/13B LLaMA-2 with 1-5B tokens, the model's context window is expanded to 128K.
Data Poisoning Attacks against Conformal Prediction
Yangyi Li (Iowa State University), Mengdi Huai (Iowa State University)
Adversarial AttackData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This paper is the first to study data poisoning attacks on Conformal Prediction, demonstrating that attackers can manipulate the model's uncertainty estimates while keeping the labels correct.
Data-efficient Large Vision Models through Sequential Autoregression
Zhiwei Hao (Beijing Institute of Technology), Chang Xu (University of Sydney)
RestorationSegmentationPose EstimationKnowledge DistillationTransformerImage
🎯 What it does: This paper proposes a data-efficient autoregressive visual model DeLVM, which achieves multi-task learning on limited data through data augmentation and knowledge distillation.
Data-Efficient Learning via Clustering-Based Sensitivity Sampling: Foundation Models and Beyond
Kyriakos Axiotis (Google Research), Michael Wunder (Google Research)
Data-Centric LearningImageTabular
🎯 What it does: This paper studies the data selection problem and proposes a new data selection method based on k-means clustering and sensitivity sampling to efficiently train machine learning models.
Data-Efficient Molecular Generation with Hierarchical Textual Inversion
Seojin Kim (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
GenerationDrug DiscoveryTransformerLarge Language ModelGraph
🎯 What it does: A data-efficient molecular generation method called HI-Mol is proposed, which utilizes hierarchical text inversion to learn the distribution of a small amount of molecular data and generates new molecules through multi-layer token interpolation.
Data-free Distillation of Diffusion Models with Bootstrapping
Jiatao Gu (Apple), Joshua M. Susskind
GenerationKnowledge DistillationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A data-independent single-step diffusion model distillation method called BOOT has been developed, which achieves rapid generation through bootstrapping target learning of the time-conditioned student model.
Data-free Neural Representation Compression with Riemannian Neural Dynamics
Zhengqi Pei (Institute of Computing Technology Chinese Academy of Sciences), Qingming Huang (School of Computer Science and Technology University of Chinese Academy of Sciences)
CompressionComputational EfficiencyImage
🎯 What it does: Reconstruct the weight matrix of the pre-trained neural network under Riemannian metrics to achieve data-free compression and inference acceleration.
DataFreeShield: Defending Adversarial Attacks without Training Data
Hyeyoon Lee (Seoul National University), Jinho Lee (Seoul National University)
Data SynthesisAdversarial AttackImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: In the absence of access to the original training data, a method is proposed to enhance adversarial robustness by generating diverse synthetic samples and refining the model through gradient refinement.
DE-COP: Detecting Copyrighted Content in Language Models Training Data
André Vicente Duarte (Instituto Superior Técnico), Lei Li (Carnegie Mellon University)
Large Language ModelTextBenchmark
🎯 What it does: This study proposes a method to detect whether copyrighted content is included in the training data of large language models through multiple-choice question answering.
Dealing With Unbounded Gradients in Stochastic Saddle-point Optimization
Gergely Neu (Universitat Pompeu Fabra), Nneka Okolo (Universitat Pompeu Fabra)
OptimizationReinforcement LearningSequential
🎯 What it does: This study investigates the performance of stochastic first-order methods in finding saddle points of convex-concave functions and proposes a simple and effective regularization technique to stabilize iterations and provide meaningful performance guarantees, even in cases where the iteration size and gradient noise grow linearly.
Debating with More Persuasive LLMs Leads to More Truthful Answers
Akbir Khan (University College London), Ethan Perez (Anthropic)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The study proposes using LLM debates in reading comprehension tasks, allowing weak models (without text access) to determine the correct answer by evaluating the debates of expert models (with text access).
Debiased Distribution Compression
Lingxiao Li (Massachusetts Institute of Technology), Lester Mackey (Microsoft Research)
CompressionOptimizationTabular
🎯 What it does: A series of distribution compression methods for biased sampling sequences are proposed, capable of generating high-quality, sparse representative point sets (coresets) to approximate the target distribution P.
Debiased Offline Representation Learning for Fast Online Adaptation in Non-stationary Dynamics
Xinyu Zhang (Nanjing University), Yang Yu (Nanjing University)
Representation LearningRecurrent Neural NetworkReinforcement LearningContrastive LearningSequential
🎯 What it does: This paper proposes a context encoder DORA that learns from offline data, capable of quickly identifying and adapting to non-stationary dynamic environments, and trains transferable meta-policies within an offline reinforcement learning framework.
Decentralized Convex Finite-Sum Optimization with Better Dependence on Condition Numbers
Yuxing Liu (Fudan University), Luo Luo (Fudan University)
OptimizationTabular
🎯 What it does: The study focuses on distributed convex finite optimization and proposes a random variance reduction algorithm CESAR that can adaptively select the mini-batch size at each node, significantly reducing dependence on the condition number.
Deciphering RNA Secondary Structure Prediction: A Probabilistic K-Rook Matching Perspective
Cheng Tan (Zhejiang University), Stan Z. Li (Westlake University)
Protein Structure PredictionConvolutional Neural NetworkSupervised Fine-TuningBiomedical Data
🎯 What it does: A framework is proposed that transforms RNA secondary structure prediction into a K-Rook matching problem, and the RFold method based on this idea is implemented;
DecisionNCE: Embodied Multimodal Representations via Implicit Preference Learning
Jianxiong Li (Tsinghua University), Xianyuan Zhan (Artificial Intelligence Research)
Representation LearningRobotic IntelligenceVision-Language-Action ModelContrastive LearningVideoMultimodality
🎯 What it does: A unified multimodal representation learning framework called DecisionNCE is proposed, which transforms the Bradley-Terry model into visual-language contrastive learning through implicit preference learning, thereby capturing both local and global information of task progress, achieving temporal consistency, and completing trajectory-level alignment of language instructions.
Decoding Compressed Trust: Scrutinizing the Trustworthiness of Efficient LLMs Under Compression
Junyuan Hong (University of Texas at Austin), Bo Li (University of Chicago)
CompressionTransformerLarge Language ModelText
🎯 What it does: A systematic evaluation of the performance of large language models in a compressed state (quantization and pruning) across eight dimensions of trustworthiness (such as fairness, ethics, robustness, etc.).
Decoding-time Realignment of Language Models
Tianlin Liu (University of Basel), Mathieu Blondel (Google DeepMind)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a real-time alignment method during decoding (DeRa), which utilizes a mixture of the reference model and the aligned model's probability distributions to adjust the KL regularization strength, allowing exploration of different alignment levels without retraining;
Decomposable Submodular Maximization in Federated Setting
Akbar Rafiey (University of California San Diego)
OptimizationFederated LearningSafty and Privacy
🎯 What it does: This paper proposes an algorithm for maximizing decomposable submodular functions in a federated learning environment, with the core methods being Federated Continuous Greedy (FEDCG) and its improved version FEDCG+. The paper also provides theoretical convergence guarantees; additionally, a discrete Federated Discrete Greedy algorithm is introduced for discrete submodular optimization problems such as maximum coverage and facility location.
Decomposing and Editing Predictions by Modeling Model Computation
Harshay Shah (Massachusetts Institute of Technology), Aleksander Madry (Massachusetts Institute of Technology)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkTransformerVision Language ModelImageText
🎯 What it does: This paper studies the impact of internal components of deep models (such as convolutional kernels, attention heads, etc.) on prediction results, proposing a 'component modeling' framework. In its special case, 'component attribution,' the COAR algorithm is designed to estimate the adversarial impact of individual components on predictions. Based on the attribution results, COAR-EDIT is implemented to achieve model editing through component ablation.
Decomposing Uncertainty for Large Language Models through Input Clarification Ensembling
Bairu Hou (University of California Santa Barbara), Yang Zhang (MIT IBM Watson Artificial Intelligence Laboratory IBM Research)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: A framework based on input clarification ensembling is proposed to decompose the uncertainty of large language model (LLM) outputs into data uncertainty (aleatoric) and model uncertainty (epistemic);
Deconstructing the Goldilocks Zone of Neural Network Initialization
Artem M Vysogorets, Julia Kempe (New York University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Analyze and redefine the Goldilocks zone during neural network initialization, derive the essential conditions that lead to excessive positive curvature, and clarify its relationship with model confidence, initial loss, and gradient vanishing through theory and experiments.
DeCoOp: Robust Prompt Tuning with Out-of-Distribution Detection
Zhi Zhou (Nanjing University), Yu-Feng Li (Nanjing University)
ClassificationRecognitionAnomaly DetectionTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: Proposes an open-world prompt tuning method DECOOP that can handle both base classes and new classes after base class training.
Decouple then Classify: A Dynamic Multi-view Labeling Strategy with Shared and Specific Information
Xinhang Wan (National University of Defense Technology), En Zhu (National University of Defense Technology)
ClassificationAuto EncoderMultimodality
🎯 What it does: This paper proposes a dynamic multi-view labeling strategy that first decouples shared information from specific information using an autoencoder and a view discriminator, and then trains two independent classifiers separately, actively selecting low-confidence samples for labeling based on uncertainty.
Decoupling Feature Extraction and Classification Layers for Calibrated Neural Networks
Mikkel Jordahn (Technical University of Denmark), Pablo M. Olmos (Universidad Carlos III de Madrid)
ClassificationOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes to split the training process of the feature extraction layer and the classification layer into two stages (TST) and to incorporate Gaussian prior variational training (V-TST) in the second stage to enhance the calibration performance of deep networks.
Decoupling Learning and Decision-Making: Breaking the $\mathcal{O}(\sqrt{T})$ Barrier in Online Resource Allocation with First-Order Methods
Wenzhi Gao (Stanford University), Yinyu Ye (Stanford University)
OptimizationReinforcement Learning
🎯 What it does: Proposes an online linear programming framework that decouples learning from decision-making, breaking the O(√T) regret upper bound for the first time and achieving O(T^{1/3}) regret and constraint violation.
Deep Demonstration Tracing: Learning Generalizable Imitator Policy for Runtime Imitation from a Single Demonstration
Xiong-Hui Chen (Nanjing University), Zongzhang Zhang (Nanjing University)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: An algorithm named Deep Demonstration Tracing (DDT) is proposed for achieving generalizable imitation learning in the face of sudden environmental changes after a single demonstration.
Deep Equilibrium Models are Almost Equivalent to Not-so-deep Explicit Models for High-dimensional Gaussian Mixtures
Zenan Ling (Huazhong University of Science and Technology), Zhenyu Liao (Huazhong University of Science and Technology)
ClassificationOptimizationComputational EfficiencyImage
🎯 What it does: Using random matrix theory (RMT) to conduct high-dimensional approximate analysis of the conjugate kernel (CK) and neural tangent kernel (NTK) of implicit deep equilibrium networks (DEQ) under high-dimensional Gaussian mixture models (GMM), it is proven that they can be completely described by four nonlinear equations, and based on this, an equivalent shallow explicit neural network (ENN) is constructed.
Deep Functional Factor Models: Forecasting High-Dimensional Functional Time Series via Bayesian Nonparametric Factorization
Yirui Liu (J.P. Morgan), Liying Wang (University of Liverpool)
Recurrent Neural NetworkTime SeriesFinance Related
🎯 What it does: A deep Bayesian nonparametric factor model DF2M is proposed for high-dimensional function time series prediction, combining the Indian buffet process, sparse factor structure, Gaussian process dynamics, and deep kernels;
Deep Fusion: Efficient Network Training via Pre-trained Initializations
Hanna Mazzawi (Google Research), Benoit Dherin (Google)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The Deep Fusion method is proposed, which first trains a small model and then merges it into a large model using a fusion operator, thereby achieving network growth during the intermediate training phase, significantly accelerating training and reducing computational costs.
Deep Networks Always Grok and Here is Why
Ahmed Imtiaz Humayun (Rice University), Richard Baraniuk (Rice University)
Adversarial AttackImageText
🎯 What it does: The study and popularization of the phenomenon of delayed generalization and delayed robustness (grokking) of deep networks under various practical settings;
Deep Neural Room Acoustics Primitive
Yuhang He (University of Oxford), Andrew Markham (University of Oxford)
Neural Radiance FieldAudio
🎯 What it does: Using two synchronized sound sources and microphone detectors, we learn implicit neural room acoustic primitives (RIR) in a self-supervised manner, enabling the prediction of corresponding indoor sound propagation impulse responses for any source-receiver pair without the need for real RIR data.
Deep Regression Representation Learning with Topology
Shihao Zhang (National University of Singapore), Angela Yao (National University of Singapore)
Depth EstimationSuper ResolutionRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes a regression representation learning regularizer called PH-Reg, based on information bottleneck and topological analysis, which ensures that the feature space and target space are consistent in intrinsic dimension and topological structure to enhance the generalization ability of regression models.
Deep Stochastic Mechanics
Elena Orlova (University of Chicago), Rebecca Willett (University of Chicago)
OptimizationComputational EfficiencyDiffusion modelTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: A deep stochastic mechanics (DSM) method based on stochastic mechanics and diffusion models is proposed, using neural networks to learn the drift term directly from quantum density sampling to solve the time-evolution Schrödinger equation, breaking through the curse of dimensionality.
Deeper or Wider: A Perspective from Optimal Generalization Error with Sobolev Loss
Yahong Yang (Pennsylvania State University), Juncai He (King Abdullah University of Science and Technology)
Physics Related
🎯 What it does: The study investigates the optimal generalization error of deep (DeNN) and wide (WeNN) neural networks under Sobolev loss, providing theoretical proofs and experimental validations, and further applies it to deep Ritz and PINN for PDE solving.
DeepPolar: Inventing Nonlinear Large-Kernel Polar Codes via Deep Learning
S Ashwin Hebbar (Princeton University), Pramod Viswanath (Princeton University)
Supervised Fine-Tuning
🎯 What it does: A Polar code based on a large kernel nonlinear neural network (DEEPPOLAR) has been designed and implemented, achieving better performance for short to medium-length codes by replacing the traditional linear kernel.
Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization
Xingyi Zhao (Utah State University), Shuhan Yuan (Utah State University)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In the context of backdoor attacks on pre-trained language models, this paper proposes a backdoor elimination method named PURE that is trigger-free and does not require a secure pre-trained model.
Defense against Model Extraction Attack by Bayesian Active Watermarking
Zhenyi Wang (University of Maryland), Heng Huang (University of Maryland)
Computational EfficiencyAdversarial AttackImage
🎯 What it does: A Bayesian active watermarking method is designed to defend against model extraction attacks: a minimal amount of fine-tuning is performed on the pre-trained model, and the posterior distribution of the watermark is learned. During deployment, the watermark is randomly sampled and added to each query, actively disrupting the attacker's query process.
Defining Neural Network Architecture through Polytope Structures of Datasets
Sangmin Lee (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper analyzes the polyhedral structure of the dataset, providing upper and lower bounds on the width of neural networks, and proposes an algorithm to reverse-engineer the polyhedral coverage from a trained network.
Degeneration-free Policy Optimization: RL Fine-Tuning for Language Models without Degeneration
Youngsoo Jang (LG AI Research), Moontae Lee (University of Illinois Chicago)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A reinforcement learning framework named Degeneration-free Policy Optimization (DfPO) is proposed to fine-tune language models to improve downstream metrics in generation tasks while avoiding text degeneration.
DéjàVu: KV-cache Streaming for Fast, Fault-tolerant Generative LLM Serving
Foteini Strati (ETH Zurich), Ana Klimovic
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A distributed LLM service system named D'ej'Vu has been constructed, utilizing the KV cache streaming library (D'ej'VuLib) to achieve the separation of prompting and generation phases, micro-batch exchange, and KV cache replication, thereby improving throughput, reducing GPU memory usage, and enabling fault tolerance.
Delaunay Graph: Addressing Over-Squashing and Over-Smoothing Using Delaunay Triangulation
Hugo Attali (Universite Sorbonne Nord), Nathalie Pernelle (Universite Sorbonne Nord)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a graph rewiring method based on Delaunay triangulation (Delaunay Rewiring, abbreviated as DR), which alleviates the issues of over-squashing and over-smoothing in GNNs by reconstructing the graph structure using only node features, achieving higher node classification accuracy.
Deletion-Anticipative Data Selection with a Limited Budget
Rachael Hwee Ling Sim (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationData-Centric LearningImageTabular
🎯 What it does: Under a limited budget, we propose Deletion-Aware Data Subset Selection (DADS), which anticipates the probability of data owners deleting data in the future to pre-select a training set that maximizes the utility of the model after deletions.
Delving into Differentially Private Transformer
Youlong Ding (Shenzhen University), Weike Pan (Shenzhen University)
Recommendation SystemSafty and PrivacyTransformerText
🎯 What it does: This study investigates how to train Transformer models under differential privacy and proposes a modular approach to achieve dimensionality reduction and optimization during the training process.
Delving into the Convergence of Generalized Smooth Minimax Optimization
Wenhan Xian (University of Maryland), Heng Huang (University of Maryland)
OptimizationTabular
🎯 What it does: This paper studies non-convex-strongly concave minimax optimization problems under non-Lipschitz smoothness conditions, proving that classical GDA, SGDA, GDmax, and SGDmax can still converge after introducing appropriate step size strategies, and provides gradient or SFO complexity that is the same as in traditional cases.
Demystifying SGD with Doubly Stochastic Gradients
Kyurae Kim (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
Optimization
🎯 What it does: This paper studies the convergence properties of doubly stochastic gradient descent (doubly SGD) in optimization problems where each component has an unmeasurable expectation (i.e., 'finite sum with infinite data'), particularly considering the case where component gradient estimators may be correlated, and provides a general variance upper bound.
Denoising Autoregressive Representation Learning
Yazhe Li (Google DeepMind), Ting Chen (xAI)
Object DetectionSegmentationGenerationRepresentation LearningTransformerDiffusion modelImage
🎯 What it does: A self-regressive visual pre-training method DARL based on a decoder Transformer is proposed, which can learn visual representations and achieve image generation.
Dense Reward for Free in Reinforcement Learning from Human Feedback
Alex James Chan, Mihaela van der Schaar (University of Cambridge)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes the Attention Based Credit (ABC) method, which uses the attention weights from the reward model to allocate the final sparse rewards to each token, generating dense rewards and enhancing the stability and convergence speed of RLHF training.
Density Ratio Estimation with Doubly Strong Robustness
Ryosuke Nagumo (Graduate University for Advanced Studies), Hironori Fujisawa (Institute of Statistical Mathematics)
Anomaly DetectionOptimizationTime Series
🎯 What it does: Two density ratio estimation methods for dual contamination of reference and target distributions, Weighted DRE and γ-DRE, are proposed, utilizing a weight function to suppress the influence of outliers, achieving dual strong robustness.
Density-Softmax: Efficient Test-time Model for Uncertainty Estimation and Robustness under Distribution Shifts
Ha Manh Bui (Johns Hopkins University), Anqi Liu (Johns Hopkins University)
ClassificationDomain AdaptationComputational EfficiencyConvolutional Neural NetworkFlow-based ModelImage
🎯 What it does: This paper proposes Density-Softmax, a model that requires only one forward pass during inference, is sampling-free, and lightweight, aimed at improving uncertainty estimation and robustness under distribution shifts.
Designing Decision Support Systems using Counterfactual Prediction Sets
Eleni Straitouri (Max Planck Institute for Software Systems), Manuel Gomez Rodriguez (Max Planck Institute for Software Systems)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper designs a decision support system based on a prediction set, which forces experts to make judgments within a given prediction set to improve the average prediction accuracy of the experts.
Detecting and Identifying Selection Structure in Sequential Data
Yujia Zheng (Carnegie Mellon University), Kun Zhang (MBZUAI)
SequentialAudio
🎯 What it does: This paper proposes a framework for identifying selection structures, direct causal relationships, and potential confounding relationships in sequential data. It proves the identifiability under non-parametric, non-interventional conditions for the first time and provides a constraint-based algorithm.
Detecting Any instruction-to-answer interaction relationship:Universal Instruction-to-Answer Navigator for Med-VQA
Zhongze Wu (Central South University), Chang Xu (University of Sydney)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The Uni-Med framework is proposed, utilizing the interactive relationship between instructions and answers, as well as visual explanations, to enhance the interpretability and accuracy of medical visual question answering.
Detecting Influence Structures in Multi-Agent Reinforcement Learning
Fabian Raoul Pieroth, Lenz Belzner (Technische Hochschule Ingolstadt)
Reinforcement Learning
🎯 What it does: A unified impact measurement framework is proposed, defining Total Impact Measurement (TIM) and State Impact Measurement (SIM) in average reward multi-agent systems, along with decoupled approximate algorithms and convergence proofs.
DetKDS: Knowledge Distillation Search for Object Detectors
Lujun Li (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
Object DetectionKnowledge DistillationImage
🎯 What it does: Proposes DetKDS, a framework for automated searching of object detection knowledge distillation strategies;
DFA-RAG: Conversational Semantic Router for Large Language Model with Definite Finite Automaton
Yiyou Sun (NEC Laboratories America), Haifeng Chen (NEC Laboratories America)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the DFA-RAG framework, which embeds the Deterministic Finite Automaton (DFA) learned from training dialogues into large language models, achieving semantic routing of dialogues through retrieval-augmented generation with labeled data.
DFD: Distilling the Feature Disparity Differently for Detectors
Kang Liu (Tsinghua University), Rizen Guo (Tencent)
Object DetectionSegmentationPose EstimationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a knowledge distillation method called Disparity Feature Distillation (DFD), which separates spatial regions based on the feature response differences between the teacher model and the student model, applying different distillation constraints to regions with varying degrees of disparity.
DFlow: A Generative Model Combining Denoising AutoEncoder and Normalizing Flow for High Fidelity Waveform Generation
Chenfeng Miao (Ping An Technology), Jing Xiao (Ping An Technology)
GenerationData SynthesisFlow-based ModelAuto EncoderAudio
🎯 What it does: Designed and implemented DFlow, a generative framework that combines denoising autoencoders and normalizing flows for high-fidelity speech waveform synthesis.
Diagnosing the Compositional Knowledge of Vision Language Models from a Game-Theoretic View
Jin Wang (University of Hong Kong), Ping Luo (Shanghai AI Laboratory)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelMultimodality
🎯 What it does: Systematically diagnose the combined knowledge of visual language models (VLM), evaluating the image encoder, text encoder, and the multimodal understanding ability of both.
DIDI: Diffusion-Guided Diversity for Offline Behavioral Generation
Jinxin Liu (Westlake University), Donglin Wang (Westlake University)
GenerationRobotic IntelligenceReinforcement LearningDiffusion modelMultimodality
🎯 What it does: This paper proposes a context strategy DIDI based on diffusion models, which can learn diverse and controllable skills from unlabeled multimodal offline data.
diff History for Neural Language Agents
Ulyana Piterbarg (New York University), Rob Fergus (New York University)
CompressionReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A technique called diff history is proposed, which generates differential summaries using the Unix diff command by observing continuous text, thereby compressing redundant information and providing a denser historical context for training neural language models to make decisions.
DiffAug: Enhance Unsupervised Contrastive Learning with Domain-Knowledge-Free Diffusion-based Data Augmentation
Zelin Zang (Westlake University), Yang You (National University of Singapore)
Representation LearningData-Centric LearningDiffusion modelContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes DiffAug, an augmentation framework for unsupervised contrastive learning that generates conditional vectors through a semantic encoder and uses a conditional diffusion model to generate positive samples, enhancing representation learning effectiveness.
DiffDA: a Diffusion model for weather-scale Data Assimilation
Langwen Huang (ETH Zurich), Torsten Hoefler (ETH Zurich)
Graph Neural NetworkDiffusion modelTime Series
🎯 What it does: This paper proposes DiffDA, a weather-scale data assimilation method based on diffusion models, which can integrate sparse observations and 48-hour forecasts into a high-quality initial field at a resolution of 0.25°.
Differentiability and Optimization of Multiparameter Persistent Homology
Luis Scoccola (Mathematical Institute, University of Oxford), Steve Oudot (GeomeriX)
OptimizationGraph Neural NetworkPoint CloudGraph
🎯 What it does: A general differentiable and gradient descent convergence framework is proposed, suitable for various descriptors of multi-parameter persistent homology, and end-to-end optimization of multi-parameter persistent homology in machine learning is achieved within this framework.
Differentiable Annealed Importance Sampling Minimizes The Jensen-Shannon Divergence Between Initial and Target Distribution
Johannes Zenn (University of Tübingen), Robert Bamler (University of Tübingen)
OptimizationTabular
🎯 What it does: This study proves the importance of differentiable entropy harmonic importance sampling (DAIS) in minimizing the Jensen-Shannon (JS) divergence between its initial distribution q₀ and the target distribution f/Z under multi-step transition limits; it also proposes using q₀ directly as a variational approximation (DAIS₀) and evaluates its inference performance on synthetic and real data.
Differentiable Combinatorial Scheduling at Scale
Mingju Liu (University of Maryland), CUNXI YU
OptimizationReinforcement LearningGraph
🎯 What it does: A differentiable combinatorial scheduling framework is proposed, which utilizes Gumbel-Softmax and the constrained Gumbel Trick to transform the scheduling problem (SDC-based linear programming) into a continuous optimization problem solvable by gradient descent;
Differentiable Distributionally Robust Optimization Layers
Xutao Ma (Shanghai Jiao Tong University), WenLi Du
OptimizationReinforcement LearningTabularTime SeriesFinance Related
🎯 What it does: This paper proposes a differentiable distributionally robust optimization (DRO) layer and embeds it into a decision-focused learning pipeline, supporting mixed-integer decisions and parameterized second-order cone (SOC) uncertainty sets.
Differentiable Mapper for Topological Optimization of Data Representation
Ziyad Oulhaj (Nantes Universite), Bertrand Michel (Nantes Universite)
OptimizationRepresentation LearningPoint CloudBiomedical Data
🎯 What it does: This paper introduces Soft Mapper, a differentiable probabilistic modification of the Mapper graph, which automatically optimizes the filter function using topological loss (such as total persistence) to obtain Mapper representations with richer topological information.
Differentiable Model Scaling using Differentiable Topk
Kai Liu (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)
OptimizationComputational EfficiencyNeural Architecture SearchConvolutional Neural NetworkTransformerImage
🎯 What it does: A model scaling method based on differentiable Top-k operations (Differentiable Model Scaling, DMS) is proposed, which can directly and differentiably search for network width and depth.
Differentiable Weightless Neural Networks
Alan Tendler Leibel Bacellar, Felipe M.G. França
ClassificationOptimizationComputational EfficiencyTabular
🎯 What it does: Proposes the Differentiable Weightless Neural Network (DWN), a multi-layer weightless neural network based on lookup tables (LUT), and achieves training through gradient descent;
Differentially Private Bias-Term Fine-tuning of Foundation Models
Zhiqi Bu (Amazon), George Karypis (Amazon)
Safty and PrivacyComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a differentially private bias-term fine-tuning method (DP-BiTFiT) that only trains the bias terms of pre-trained models, achieving fine-tuning on private data while maintaining high accuracy.
Differentially Private Decentralized Learning with Random Walks
Edwige Cyffers (Universite de Lille), Jalaj Upadhyay (Rutgers University)
OptimizationFederated LearningSafty and PrivacyGraph Neural NetworkGraphTabular
🎯 What it does: This paper proposes a private random walk-based stochastic gradient descent (RW-DP-SGD) algorithm for decentralized distributed learning, and provides convergence analysis under strongly convex functions and a closed-form estimate of privacy leakage for arbitrary graph structures.
Differentially Private Domain Adaptation with Theoretical Guarantees
Raef Bassily (Ohio State University), Mehryar Mohri (Courant Institute of Mathematical Sciences)
Domain AdaptationOptimizationSafty and PrivacyImageTabular
🎯 What it does: Two (ε, δ)-differential privacy supervised domain adaptation algorithms are proposed, aimed at utilizing data from a public source domain to improve the prediction accuracy of a private target domain.
Differentially private exact recovery for stochastic block models
Dung Nguyen (University of Virginia), Anil Kumar Vullikanti
OptimizationSafty and PrivacyGraphStochastic Differential Equation
🎯 What it does: This paper studies the exact recovery theoretical thresholds under the edge differential privacy model for three types of random block models: binary asymmetric, masking, and general structure, and proposes a polynomial-time private algorithm.
Differentially Private Post-Processing for Fair Regression
Ruicheng Xian (University of Illinois Urbana-Champaign), Han Zhao (University of Illinois Urbana-Champaign)
OptimizationSafty and PrivacyTabular
🎯 What it does: A post-processing algorithm based on differential privacy is proposed, which can adjust the fairness of the regressor without changing the training phase.
Differentially Private Representation Learning via Image Captioning
Tom Sander (Meta), Chuan Guo (Meta)
Safty and PrivacyRepresentation LearningImageTextMultimodality
🎯 What it does: This study proposes a method for differential privacy representation learning through image caption generation, demonstrating the feasibility of effective differential privacy representation learning on internet-scale multimodal datasets.
Differentially Private Sum-Product Networks
Xenia Heilmann (Johannes Gutenberg University Mainz), Ernst Althaus (Johannes Gutenberg University Mainz)
ClassificationData SynthesisSafty and PrivacyGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes a differential privacy model based on Sum-Product Networks (DPSPN), which can accomplish both classification and synthetic data generation tasks in a single training session.
Differentially Private Synthetic Data via Foundation Model APIs 2: Text
Chulin Xie (University of Illinois Urbana-Champaign), Sergey Yekhanin (Microsoft Research)
Data SynthesisSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an algorithm named AUG-PE, which generates differential privacy (DP) synthetic text by utilizing only the API of large language models (LLMs) without the need for model training.
Differentially Private Worst-group Risk Minimization
Xinyu Zhou (Ohio State University), Raef Bassily (Ohio State University)
OptimizationSafty and Privacy
🎯 What it does: Proposed a worst-case group risk minimization algorithm under (ϵ,δ)-differential privacy constraints, and provided an approximately optimal error upper bound.
DiffFPR: Diffusion Prior for Oversampled Fourier Phase Retrieval
Ji Li (Capital Normal University), Chao Wang (University of Kansas Medical Center)
RestorationDiffusion modelImage
🎯 What it does: A method called DiffFPR is proposed, which combines diffusion models with traditional iterative engines to address the oversampling Fourier phase reconstruction problem for multi-channel color images.
DiffStitch: Boosting Offline Reinforcement Learning with Diffusion-based Trajectory Stitching
Guanghe Li (Jilin University), Weinan Zhang (Shanghai Jiao Tong University)
Reinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: This paper proposes a trajectory stitching method based on diffusion models, called DiffStitch, to enhance offline reinforcement learning datasets.
Diffuse, Sample, Project: Plug-And-Play Controllable Graph Generation
Kartik Sharma (Georgia Institute of Technology), Rakshit Trivedi (Massachusetts Institute of Technology)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A plug-and-play method named PRODIGY is proposed, which achieves hard constraints on graphs (such as the number of edges, triangles, node degrees, atom counts, molecular weights, etc.) using any pre-trained graph diffusion model without the need for retraining or additional labels.
Diffusion Language Models Are Versatile Protein Learners
Xinyou Wang (ByteDance Research), Quanquan Gu (ByteDance Research)
GenerationProtein Structure PredictionTransformerDiffusion modelMultimodalityBiomedical Data
🎯 What it does: This study presents DPLM, a protein language model based on a discrete diffusion probability framework, capable of generating and learning representations of protein sequences.
Diffusion Model-Augmented Behavioral Cloning
Shang-Fu Chen (National Taiwan University), Shao-Hua Sun (National Taiwan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackDiffusion modelSequential
🎯 What it does: A framework for imitation learning is proposed, which combines modeling of conditional probabilities and joint probabilities, called Diffusion Model Enhanced Behavior Cloning (DBC).
Diffusion Models Demand Contrastive Guidance for Adversarial Purification to Advance
Mingyuan Bai (RIKEN), Qibin Zhao (RIKEN)
RestorationAdversarial AttackDiffusion modelContrastive LearningImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A novel adversarial denoising diffusion model is proposed, utilizing contrastive guidance to achieve efficient purification of adversarial samples.
Diffusion Models Encode the Intrinsic Dimension of Data Manifolds
Jan Pawel Stanczuk (University of Cambridge), Carola-Bibiane Schönlieb (University of Cambridge)
GenerationData SynthesisDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: Using score fields of diffusion models to approximate the normal bundle of data manifolds, thereby estimating the intrinsic dimension of the manifold.
Diffusion Posterior Sampling is Computationally Intractable
Shivam Gupta (University of Texas at Austin), Zhiyang Xun (University of Texas at Austin)
Computational EfficiencyDiffusion model
🎯 What it does: It is proven that under the condition of a given approximate smoothness score, posterior sampling of diffusion models is computationally infeasible in the worst case, and any efficient posterior sampling algorithm necessarily requires super-polynomial time.