Conference on Neural Information Processing Systems · 1874 papers
DeformableTST: Transformer for Time Series Forecasting without Over-reliance on Patching
Donghao Luo (Tsinghua University), Xue Wang (Tsinghua University)
CodeTransformerTime Series
🎯 What it does: DeformableTST is proposed, a Transformer structure with optional patching, achieving time series prediction through deformable attention and hierarchical structure.
🎯 What it does: A unified conditional diffusion framework based on Doob h-transformation is proposed, and within this framework, the DEFT method is designed, utilizing fine-tuned minimal networks for efficient conditional generation.
Hikaru Shindo (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
CodeObject DetectionSegmentationTransformerLarge Language ModelPrompt EngineeringImage
🎯 What it does: This paper proposes DeiSAM, which combines large-scale pre-trained models with differentiable logical reasoning to achieve image segmentation based on complex instructive natural language prompts.
Delta-CoMe: Training-Free Delta-Compression with Mixed-Precision for Large Language Models
Bowen Ping (Peking University), Maosong Sun (Tsinghua University)
CodeCompressionTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper studies an untrained mixed-precision Delta compression method called Delta-CoMe, which is used to compress the incremental weights of fine-tuned LLMs, supporting multi-model multi-tenant deployment.
DeltaDEQ: Exploiting Heterogeneous Convergence for Accelerating Deep Equilibrium Iterations
Zuowen Wang (Institute of Neuroinformatics University of Zurich and ETH Zurich), Shih-Chii Liu (Institute of Neuroinformatics University of Zurich and ETH Zurich)
🎯 What it does: The DeltaDEQ method is proposed, which significantly reduces inference computation by using the delta update rule in the fixed-point iteration of the Deep Equilibrium Model (DEQ) to skip converged dimensions.
🎯 What it does: The DeltaDock two-stage framework is proposed, which first selects the best pocket through pocket-ligand alignment, and then uses a dual-layer coarse-to-fine iteration to complete blind docking and pose docking.
Delving into the Reversal Curse: How Far Can Large Language Models Generalize?
Zhengkai Lin (Zhejiang University), Jieping Ye (Alibaba Cloud)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper conducts experiments on large language models across various tasks (open-ended question answering and multiple choice) to study and reveal the so-called 'reversal curse' and the model's 'thinking bias', and explores the impact of training data structure on knowledge transfer and generalization ability.
DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States
Bozhou Zhang (Fudan University), Li Zhang (Fudan University)
CodeAutonomous DrivingTransformerPoint Cloud
🎯 What it does: The DeMo framework is proposed, which splits trajectory prediction queries into pattern queries (capturing directional intent) and state queries (tracking dynamic states), and achieves more accurate multimodal trajectory prediction through a hybrid attention mechanism and the Mamba module.
🎯 What it does: The core mechanisms of Mamba and linear attention Transformers are theoretically compared and experimentally validated, and based on this, the Mamba-Inspired Linear Attention (MILA) model is proposed.
🎯 What it does: This paper proposes a framework called DenoiseRep that integrates the denoising process of diffusion models into feature extraction to enhance the feature representation capability for discriminative tasks.
Huanjin Yao (Baidu Inc.), Jingdong Wang (Chinese University of Hong Kong)
CodeRecognitionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the Dense Connector plugin, which enhances the visual perception capabilities of multimodal large language models by utilizing multi-layer features from visual encoders.
🎯 What it does: A density-based user representation method GPR4DUR based on Gaussian process regression is proposed for multi-interest personalized retrieval.
Yuan Qiu (Georgia Institute of Technology), Peng Chen (Georgia Institute of Technology)
CodeAuto EncoderPhysics Related
🎯 What it does: This paper proposes the Derivative-enhanced Deep Operator Network (DE-DeepONet), which improves the accuracy of predicting parameter functions and their derivatives by incorporating derivative information and dimensionality reduction into DeepONet.
Designing Cell-Type-Specific Promoter Sequences Using Conservative Model-Based Optimization
Aniketh Janardhan Reddy (University of California), Nilah M Ioannidis
CodeOptimizationTransformerSupervised Fine-TuningBiomedical Data
🎯 What it does: This paper proposes an offline model-driven optimization (MBO) framework based on Conservative Objective Models (COMs) for the efficient design of cell type-specific promoters targeting similar leukemia cell lines (Jurkat, K562, THP1) under data-limited conditions.
🎯 What it does: For visual Transformers using the token sparsification (TS) mechanism, a DeSparsify attack is proposed, which utilizes a custom loss function to disable the TS mechanism, resulting in a decrease in model usability.
Detecting and Measuring Confounding Using Causal Mechanism Shifts
Abbavaram Gowtham Reddy (Indian Institute of Technology Hyderabad), Vineeth N. Balasubramanian
CodeGraph
🎯 What it does: Three metrics for measuring confounding (CNF-1, CNF-2, CNF-3) in multiple environments have been proposed, which can distinguish between observed and unobserved confounding and assess the strength of confounding. These metrics are applicable in scenarios where different environmental information and causal paths are known or unknown.
🎯 What it does: The study introduces the concept of 'margin consistency' in robustly trained deep classifiers, utilizing logit margins as an approximation of input space margins to detect the vulnerability of individual samples, and enhances the detection performance of models with insufficient margin consistency by learning pseudo-margins.
🎯 What it does: Proposes the DeTeCtive framework, which utilizes multi-task assisted multi-level contrastive learning and KNN retrieval to achieve AI-generated text detection.
Deterministic Uncertainty Propagation for Improved Model-Based Offline Reinforcement Learning
Abdullah Akgül (University of Southern Denmark), Melih Kandemir (University of Southern Denmark)
CodeReinforcement LearningTabular
🎯 What it does: The MOMBO algorithm is proposed, which utilizes a deterministic moment matching method to propagate the uncertainty of the environment model through the Q-network, thereby enabling more accurate Bellman targets for conservative value iteration in model-based offline RL.
DeTikZify: Synthesizing Graphics Programs for Scientific Figures and Sketches with TikZ
Jonas Belouadi (University of Mannheim), Steffen Eger (University of Technology Nuremberg)
CodeGenerationData SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageTextMultimodality
🎯 What it does: This paper presents DeTi k Zify, a multimodal language model capable of automatically generating TikZ graphic programs based on sketches or existing images, achieving high-quality reproduction of scientific graphics and sketch transformation.
🎯 What it does: On small AI accelerators, the spatial information of the original image is expanded into additional channels through patch-wise even sampling and channel-wise stacking, rewriting the input of the first layer CNN, thereby making full use of unused processors and memory, improving inference accuracy without increasing latency.
DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization
Yueming Xu (Fudan University), Li Zhang (Fudan University)
CodePose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper presents DG-SLAM, a real-time SLAM system for dynamic environments based on 3D Gaussian splatting, achieving high-precision camera pose estimation and high-quality map reconstruction.
CodeTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper proposes an intrinsic evaluation metric based on effective rank difference (Diff-eRank) to quantify the denoising capability of LLMs in training.
🎯 What it does: A single-step diffusion-denoising data augmentation method based on diffusion models (DiffAug) is proposed and applied to train image classifiers to enhance robustness.
🎯 What it does: This paper proposes a method to extend sparse topological gradients into a globally smooth vector field through differentiable homeomorphic interpolation. Based on this, a gradient descent algorithm that can be combined with subsampling is designed for topological optimization of large-scale point clouds and topological regularization of the latent space in black-box autoencoders.
🎯 What it does: This paper directly optimizes the edge weights of the task graph through maximum likelihood estimation for differentiable optimization, proposing the TGML loss function, which can learn interpretable task graphs in neural networks through gradient descent.
Differential Privacy in Scalable General Kernel Learning via $K$-means Nystr{\"o}m Random Features
Bonwoo Lee (KAIST), Cheolwoo Park (KAIST)
CodeOptimizationSafty and PrivacyGaussian SplattingTabular
🎯 What it does: A scalable differential privacy kernel learning framework based on K-means Nyström approximation is proposed, compatible with any kernel and achieving DP ERM, KME, and general data publishing.
Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement
Jeremiah Birrell (Texas State University), Jason Pacheco (University of Arizona)
CodeOptimizationSafty and PrivacyImage
🎯 What it does: A new differential privacy stochastic gradient descent (DP-SGD) accounting method is proposed, with a comprehensive R' enyi differential privacy (RDP) analysis for fixed-size subsampling (FSwoR and FSwR).
🎯 What it does: This paper proposes DiffHammer, a selective attack method for diffusion-based adversarial purification that effectively circumvents the gradient dilemma and enhances evaluation efficiency.
DiffLight: A Partial Rewards Conditioned Diffusion Model for Traffic Signal Control with Missing Data
Hanyang Chen (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CodeAutonomous DrivingTransformerReinforcement LearningDiffusion modelTime Series
🎯 What it does: In the context of missing data, DiffLight is proposed, which jointly utilizes conditional diffusion models to achieve traffic signal control and data missing compensation.
DiffNorm: Self-Supervised Normalization for Non-autoregressive Speech-to-speech Translation
Weiting Tan (Johns Hopkins University), Philipp Koehn (Johns Hopkins University)
CodeTransformerDiffusion modelAudio
🎯 What it does: Proposed DIFFNORM (self-supervised normalization based on diffusion models) and Classifier-Free Guidance to improve non-autoregressive speech-to-speech translation systems.
DiffPhyCon: A Generative Approach to Control Complex Physical Systems
Long Wei (Westlake University), Tailin Wu (Westlake University)
CodeOptimizationRobotic IntelligenceReinforcement LearningDiffusion modelTime SeriesPhysics Related
🎯 What it does: This paper proposes DiffPhyCon, a generative control framework based on diffusion models, which simultaneously generates control sequences and physical system trajectories while optimizing the overall control objectives.
DiffPO: A causal diffusion model for learning distributions of potential outcomes
Yuchen Ma (LMU Munich), Stefan Feuerriegel (LMU Munich)
CodeDiffusion model
🎯 What it does: A causal diffusion model named DiffPO is proposed to learn the complete distribution of potential outcomes (PO) from observational data, providing both point estimates and uncertainty quantification.
🎯 What it does: This paper proposes a point diffusion-based box correction method called DiffuBox, aimed at improving the localization and size of 3D object detection boxes across different domains.
🎯 What it does: Designed and implemented an online reinforcement learning algorithm DACER based on the reverse process of diffusion models, aimed at learning multimodal policies and controlling exploration through an entropy regulator.
🎯 What it does: Proposes the EncDiff framework, utilizing diffusion models and cross-attention as prior biases to achieve unsupervised separable representation learning.
🎯 What it does: This paper theoretically proves that diffusion classifiers have an O(1) Lipschitz constant and derives their verifiable robustness; it further extends diffusion classifiers to data with arbitrary noise levels, constructing Noised Diffusion Classifiers (NDC) and proposing two implementations: Exact Posterior and Approximate Posterior, utilizing random smoothing to enhance the robustness radius; at the same time, it designs two techniques, variance reduction and Sift-and-Refine, to significantly reduce computational complexity.
🎯 What it does: A learnable multivariate adaptive noise (MuLAN) diffusion process is proposed to improve the log-likelihood estimation of image generation.
Diffusion of Thought: Chain-of-Thought Reasoning in Diffusion Language Models
Jiacheng Ye (University of Hong Kong), Lingpeng Kong
CodeDiffusion modelTextChain-of-Thought
🎯 What it does: This paper proposes Diffusion-of-Thought (DoT), which combines diffusion models with chain-of-thought (CoT) to achieve parallel reasoning steps, and conducts experiments in areas such as multi-step reasoning, error self-correction, and reasoning efficiency adjustment.
🎯 What it does: This paper proposes Diffusion Trusted Q-Learning (DTQL), which achieves efficient training and inference in offline reinforcement learning through a dual policy (diffusion policy and first-order policy) and a diffusion trust region loss.
Diffusion Priors for Variational Likelihood Estimation and Image Denoising
Jun Cheng (Huazhong University of Science and Technology), Shan Tan (Huazhong University of Science and Technology)
CodeRestorationDiffusion modelImage
🎯 What it does: A self-supervised image denoising method based on diffusion priors and variational Bayes is proposed: during the reverse diffusion process, the noise precision posterior is adaptively estimated, the independent and identically distributed (i.i.d.) Gaussian likelihood is dynamically updated, and local Gaussian convolution is used to correct the noise variance, ultimately obtaining the denoised image through MAP inference; at the same time, the local prior of the low-resolution diffusion model is used to directly process high-resolution noisy images, and the method is extended to non-Gaussian noise and demosaicing tasks.
🎯 What it does: Transfer learning is applied to pre-trained diffusion models, proposing the Diff-Tuning method, which utilizes the chain forgetting trend of the reverse process for knowledge retention and reconstruction.
🎯 What it does: This paper studies a fraction-based diffusion model called Twigs for conditional graph generation, particularly for tasks such as molecular design and network graph optimization.
🎯 What it does: A hierarchical semantic reconstruction framework based on diffusion models is proposed, using multi-layer feature extraction and diffusion reconstruction to identify unsupervised anomaly distributions.
🎯 What it does: A course reinforcement learning method based on conditional diffusion models, DiCuRL, is proposed, which utilizes diffusion models to generate intermediate goals with progressively increasing difficulty, helping agents learn efficiently in environments without domain knowledge.
🎯 What it does: This paper proposes a Diffusion-based Truncated Sampler (DITS) to improve the modality alignment issue in text-video retrieval systems.
🎯 What it does: Proposes two methods, DiffusionBlend and DiffusionBlend++, which utilize 3D-patch diffusion score fusion to learn 3D CT image priors and achieve 3D reconstruction under sparse view/limited angle conditions.
🎯 What it does: Proposes the DiffusionFake framework, which utilizes the pre-trained Stable Diffusion to perform inverse reconstruction of source and target images, guiding the detector to learn source/target features, thereby enhancing the cross-domain generalization ability of deepfake detection.
DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning
Hao Bai (University of Illinois at Urbana-Champaign), Aviral Kumar (Carnegie Mellon University)
CodeRobotic IntelligenceReinforcement LearningVision Language ModelSequentialBenchmark
🎯 What it does: A self-reinforcement learning framework called DigiRL was constructed and trained to perform device control tasks on Android devices using a pre-trained VLM.
🎯 What it does: This paper proposes a diffusion model architecture that integrates spatial and frequency information—DiMSUM, which combines Mamba state space networks with discrete wavelet transforms, and further fuses spatial features and frequency sub-bands through cross-attention to generate high-quality images.
DiPEx: Dispersing Prompt Expansion for Class-Agnostic Object Detection
Jia Syuen Lim (University of Queensland), Yadan Luo (University of Queensland)
CodeObject DetectionPrompt EngineeringVision Language ModelImage
🎯 What it does: A self-supervised Dispersing Prompt Expansion (DiPEx) method is proposed for class-agnostic object detection, enhancing recall through the diffusion of non-overlapping prompts.
🎯 What it does: The Director3D framework is proposed to generate real-world 3D scenes from text and automatically generate suitable camera trajectories.
🎯 What it does: A distribution-based model editing method called DISCEDIT is proposed, which utilizes discriminative filters to identify and prune key network components for structured pruning and category forgetting.
Discovering Preference Optimization Algorithms with and for Large Language Models
Chris Lu (Sakana AI), Robert Tjarko Lange (Sakana AI)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Utilizing large language models to automatically generate and evaluate preference optimization loss functions, thereby discovering new offline preference optimization algorithms; ultimately proposing DiscoPOP (Logarithmic Ratio Modulation Loss) and validating its effectiveness across various dialogue, summarization, and sentiment generation tasks.
Discovering Sparsity Allocation for Layer-wise Pruning of Large Language Models
Lujun Li (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
CodeOptimizationTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper proposes the DSA framework, which automatically searches for hierarchical sparse allocation functions for hierarchical pruning of large language models, significantly reducing the model parameter count while maintaining or improving performance.
🎯 What it does: A discrete dictionary decomposition layer (D3) is proposed to enhance the decomposition capability of tensor product representation (TPR) models in system generalization tasks.
🎯 What it does: A diffusion model based on discrete boundary conditions is proposed, which first estimates the boundaries of discrete samples and scales the trajectories in both the forward and backward processes, making the continuous diffusion model more suitable for generating discrete data.
CodeData SynthesisSafty and PrivacyAuto EncoderContrastive LearningImageText
🎯 What it does: This paper proposes an implicit Zero-Watermarking scheme based on decoupled style domains for detecting unauthorized data usage in text-to-image models.
Disentangling Interpretable Factors with Supervised Independent Subspace Principal Component Analysis
Jiayu Su (Columbia University), Raul Rabadan (Columbia University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningBiomedical Data
🎯 What it does: A linear multi-subspace PCA extension named sisPCA is proposed to decouple and explain multiple independent subspaces in high-dimensional biological data.
Disentangling Linear Quadratic Control with Untrusted ML Predictions
Tongxin Li (Chinese University of Hong Kong), Yisong Yue (California Institute of Technology)
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTime Series
🎯 What it does: An online control strategy named DISC is proposed, which can adaptively learn the confidence parameters of machine learning predictions in the presence of unobservable disturbances in linear quadratic control problems, and achieve control by combining decoupled predictive information.
Dissecting Query-Key Interaction in Vision Transformers
Xu Pan (Harvard University), Odelia Schwartz (University of Miami)
CodeTransformerImage
🎯 What it does: This paper analyzes the attention preferences of self-attention at different levels for similar and dissimilar tokens by performing singular value decomposition on the query-key interaction matrix in visual Transformers, and visualizes semantically interpretable interaction patterns.
🎯 What it does: This paper derives the predictive mean and kernel function expressions of deep multi-head self-attention networks (Transformer approximation models) under the thermodynamic limit of finite width based on the framework of Bayesian statistical mechanics, revealing the task-related kernel combination mechanism among attention paths.
🎯 What it does: A continuous graph neural network framework called DRAGON is proposed, which incorporates learnable distributed fractional-order operators to replace traditional single fractional or integer-order models, achieving more flexible feature updates.
Distribution-Aware Data Expansion with Diffusion Models
haoweiz, Bin Wang (Tsinghua University)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes DistDiff, a training-free data augmentation framework that generates new samples consistent with the real distribution and diverse by using energy guidance and residual multiplicative transformations at intermediate steps of the diffusion model.
Distributional Reinforcement Learning with Regularized Wasserstein Loss
Ke Sun (University of Alberta), Linglong Kong (University of Alberta)
CodeReinforcement LearningSequential
🎯 What it does: A distributed reinforcement learning algorithm called SinkhornDRL based on Sinkhorn divergence is proposed, which uses regularized Wasserstein loss to measure the difference between the current and target return distributions.
DistrictNet: Decision-aware learning for geographical districting
Cheikh Ahmed (Polytechnique Montreal), Thibaut Vidal (Polytechnique Montreal)
CodeOptimizationGraph Neural NetworkGraph
🎯 What it does: The DISTRICTNET framework is proposed, which utilizes graph neural networks to generate edge weights and employs a differentiable capacity minimum spanning tree (CMST) solver to predict and solve the urban districting problem.
CodeTransformerLarge Language ModelSupervised Fine-TuningTextMagnetic Resonance Imaging
🎯 What it does: This paper systematically compares the differences in processing narrative texts between language models (GPT-2 XL) and the human brain, using MEG data to evaluate model prediction errors, automatically generating and validating two major hypotheses regarding social/emotional and physical common sense differences, and enhancing the consistency with brain responses through fine-tuning the model on corresponding datasets.
Diversity-Driven Synthesis: Enhancing Dataset Distillation through Directed Weight Adjustment
Jiawei Du (Agency for Science Technology and Research), Joey Tianyi Zhou (Agency for Science Technology and Research)
CodeData SynthesisKnowledge DistillationImage
🎯 What it does: This paper proposes a new dataset distillation method called Directed Weight Adjustment (DWA), which enhances the diversity and representativeness of the synthetic dataset by dynamically adjusting the weights during the synthesis process.
Divide-and-Conquer Meets Consensus: Unleashing the Power of Functions in Code Generation
Jingchang Chen (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
CodeAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: A code generation framework called FUNCODER is proposed, which combines divide-and-conquer with functional consensus to recursively decompose functions and filter through functional consensus to achieve code generation for complex requirements.
🎯 What it does: In Bayesian inverse problems, a denoising diffusion model (DDM) is used as a prior, and a block decomposition posterior sampling framework (DCPS) is proposed, which gradually approximates the target posterior by constructing a series of intermediate posteriors;
🎯 What it does: This paper proposes a new plugin method DMPlug, which uses a pre-trained diffusion model to solve inverse problems, addressing issues such as insufficient manifold feasibility and measurement feasibility in traditional methods, as well as lack of robustness to unknown noise.
🎯 What it does: Evaluate the performance of models trained with causal features versus all features on 16 tabular datasets in cross-domain generalization.
Do LLMs Build World Representations? Probing Through the Lens of State Abstraction
Zichao Li (Mila), Jackie CK Cheung
CodeRepresentation LearningTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies how large language models construct world representations in text planning tasks and proposes a detection framework based on reinforcement learning state abstraction.
🎯 What it does: A zero-shot online action detection framework, OV-OAD, is proposed, which utilizes large-scale video-text pairs for pre-training and achieves real-time, open-vocabulary action recognition solely based on text supervision.
DoFIT: Domain-aware Federated Instruction Tuning with Alleviated Catastrophic Forgetting
Binqian Xu (Nanjing University of Science and Technology), Jinhui Tang (National University of Singapore)
CodeFederated LearningSupervised Fine-TuningTextFinance Related
🎯 What it does: Proposes the DoFIT framework to address the catastrophic forgetting problem caused by cross-domain data heterogeneity in federated instruction tuning;
🎯 What it does: A bidirectional synthetic planning algorithm DESP is proposed to address the computer-aided synthetic planning problem with starting material constraints.
Doubly Mild Generalization for Offline Reinforcement Learning
Yixiu Mao (Tsinghua University), Xiangyang Ji (Tsinghua University)
CodeReinforcement LearningTabular
🎯 What it does: This paper studies how to moderately utilize generalization capabilities in offline reinforcement learning, proposing the Doubly Mild Generalization (DMG) method, which balances mild action generalization and mild generalization propagation to enhance offline RL performance and achieve seamless transfer to online learning.
Drago: Primal-Dual Coupled Variance Reduction for Faster Distributionally Robust Optimization
Ronak Mehta (University of Washington), Zaid Harchaoui (University of Washington)
CodeOptimizationTabularBenchmark
🎯 What it does: A random primal-dual algorithm named DRAGO is proposed to solve the penalized distributionally robust optimization (DRO) problem with closed, convex uncertainty sets.
DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation
Yuang Ai (Institute of Automation, Chinese Academy of Sciences), Hongxia Yang (ByteDance, Inc)
CodeRestorationData SynthesisTransformerLarge Language ModelMixture of ExpertsDiffusion modelImage
🎯 What it does: This paper proposes the GenIR unified dataset construction pipeline and two key components, DreamClear, achieving a large-scale, privacy-preserving real image restoration system.
🎯 What it does: This paper presents DreamSteerer, a pluggable method that significantly enhances the editability of source images under personalized concept conditions by fine-tuning existing text-to-image personalization diffusion models.
DRIP: Unleashing Diffusion Priors for Joint Foreground and Alpha Prediction in Image Matting
Xiaodi Li (Zhejiang University), Yi Yang (Zhejiang University)
CodeSegmentationGenerationDiffusion modelImage
🎯 What it does: The DRIP method is proposed, utilizing a pre-trained latent diffusion model to jointly predict the foreground color and alpha mask of a single image, achieving high-quality image matting.
🎯 What it does: Proposed the DHD framework for trajectory prediction in multi-drone collaboration, which includes ground-prior-based BEV generation and sliding window sparse interaction.
DropBP: Accelerating Fine-Tuning of Large Language Models by Dropping Backward Propagation
Sunghyeon Woo (Seoul National University), Dongsoo Lee (Seoul National University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: During the fine-tuning process of LLMs, certain layers of backpropagation are randomly dropped to reduce computational and activation storage overhead.
🎯 What it does: Proposes the Signed Graph Augmentation framework (SGA), which addresses the sparsity and imbalance triangle problem in SGNN through structural augmentation, candidate edge selection, and difficulty-based curriculum learning, significantly improving link sign prediction performance.
🎯 What it does: A speech decoding framework called Du-IN based on sEEG has been developed, which achieves speech decoding through mask modeling guided by region-level discrete coding.
Dual Cone Gradient Descent for Training Physics-Informed Neural Networks
Youngsik Hwang (Ulsan National Institute of Science and Technology), Dongyoung Lim
CodeOptimizationPhysics Related
🎯 What it does: This paper proposes the Dual Cone Gradient Descent (DCGD) framework to address the pathologies caused by gradient imbalance and conflicts during the training of Physics-Informed Neural Networks (PINNs), ensuring that each gradient update falls within the dual cone, thereby achieving multi-objective optimization.
Mathieu Tanneau (Georgia Institute of Technology), Pascal Van Hentenryck (Georgia Institute of Technology)
CodeOptimizationTabular
🎯 What it does: A Dual Lagrangian Learning (DLL) method is proposed, utilizing dual cone theory and machine learning to predict the dual feasible solutions of parametric cone optimization problems, thereby providing effective Lagrangian dual bounds.
Dual Prototype Evolving for Test-Time Generalization of Vision-Language Models
Ce Zhang (Carnegie Mellon University), Yaqi Xie (Carnegie Mellon University)
CodeClassificationDomain AdaptationOptimizationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the Dual Prototype Evolving (DPE) method, which achieves adaptive generalization of VLM through dual-modal prototype evolution during testing.
Dual Risk Minimization: Towards Next-Level Robustness in Fine-tuning Zero-Shot Models
Kaican Li (Hong Kong University of Science and Technology), Nevin L. Zhang (Hong Kong University of Science and Technology)
CodeDomain AdaptationOptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImage
🎯 What it does: This paper proposes Dual Risk Minimization (DRM), which balances expected performance and worst-case performance when fine-tuning zero-shot pre-trained models to enhance robustness against distribution shifts.
🎯 What it does: A Dual-Diffusion framework is proposed, utilizing diffusion models to simultaneously denoise binocular 2D keypoints and their corresponding 3D joints, thereby improving the accuracy of 3D human pose estimation.
🎯 What it does: This paper studies a fully self-supervised three-dimensional particle tracking velocimetry (PTV) framework that implements dual-frame flow estimation through optimization during testing.
Dual-Personalizing Adapter for Federated Foundation Models
yiyuan yang, Michael Blumenstein (University of Technology Sydney)
CodeFederated LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes FedDPA—a dual personalized adapter designed to simultaneously address client personalization and testing distribution shift issues within the Federated Foundation Model (FedFM) framework.
🎯 What it does: This paper proposes an end-to-end trainable Dual-Perspective Activation (DPA) mechanism, which identifies and suppresses the activation of irrelevant channels in neural networks through online forward and backward criteria, thereby achieving channel denoising and sparse representation.