ICLR 2023 Papers — Page 4
International Conference on Learning Representations · 1573 papers
CUTS: Neural Causal Discovery from Irregular Time-Series Data
Yuxiao Cheng (Tsinghua University), Qionghai Dai (Tsinghua University)
Graph Neural NetworkSupervised Fine-TuningTime Series
🎯 What it does: The CUTS framework is proposed, which alternates between data imputation and causal graph learning to discover nonlinear Granger causality from irregular time series data.
Cycle to Clique (Cy2C) Graph Neural Network: A Sight to See beyond Neighborhood Aggregation
Yun Young Choi (National Institute for Mathematical Sciences), U Jin Choi (KAIST)
ClassificationGraph Neural NetworkGraph
🎯 What it does: A Cycle-to-Clique Graph Neural Network (Cy2C-GNN) is proposed, which enhances the expressive power of traditional GNNs by transforming cycles in the graph into cliques and introducing a clique adjacency matrix.
D4AM: A General Denoising Framework for Downstream Acoustic Models
Chi-Chang Lee (National Taiwan University), Chu-Song Chen (Academia Sinica)
RecognitionOptimizationConvolutional Neural NetworkSupervised Fine-TuningAudio
🎯 What it does: Proposes the D4AM framework, which jointly trains the speech enhancement model with the ASR task, achieving a universal noise reduction preprocessor through gradient calibration and regression target weighting.
D4FT: A Deep Learning Approach to Kohn-Sham Density Functional Theory
Tianbo Li (SEA AI Lab), Shuicheng YAN
OptimizationComputational EfficiencyTabularPhysics Related
🎯 What it does: A deep learning-based Kohn-Sham DFT solving method called D4FT is proposed, which directly performs gradient descent on the energy function, eliminates orthogonality constraints through reparameterization, and transfers numerical integration to SGD sampling, significantly reducing computational complexity and achieving differentiable solutions.
DAG Learning on the Permutahedron
Valentina Zantedeschi (ServiceNow Research), Vlad Niculae (Informatics Institute University of Amsterdam)
OptimizationGraph
🎯 What it does: A continuous optimization framework based on Permutahedron is proposed, using vector parameterization of node ordering to achieve DAG learning.
DAG Matters! GFlowNets Enhanced Explainer for Graph Neural Networks
Wenqian Li (National University of Singapore), Yan Pang (National University of Singapore)
Explainability and InterpretabilityGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Proposes an instance-level graph neural network explainer GFlowExplainer based on GFlowNets, improving the subgraph search process.
DamoFD: Digging into Backbone Design on Face Detection
Yang Liu (Alibaba Group), Baigui Sun (Alibaba Group)
Object DetectionNeural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: A NAS-based framework for searching dedicated backbone networks for face detection is designed, utilizing DDSAR-Score to evaluate backbone performance.
DASHA: Distributed Nonconvex Optimization with Communication Compression and Optimal Oracle Complexity
Alexander Tyurin (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated LearningTabular
🎯 What it does: A series of new distributed non-convex optimization algorithms, DASHA (including DASHA-PAGE, DASHA-MVR, DASHA-SYNC-MVR), have been proposed, which can achieve lower gradient operator complexity and communication complexity under the premise of using unbiased compressors.
Data augmentation alone can improve adversarial training
Lin Li (King's College London), Michael W. Spratling (King's College London)
Adversarial AttackData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This paper demonstrates and implements that significant improvements in model robustness and accuracy can be achieved in adversarial training by using only data augmentation, and proposes a new cropping transformation called Cropshift along with a multi-layer enhancement strategy named IDBH.
Data Continuity Matters: Improving Sequence Modeling with Lipschitz Regularizer
Eric Qu (Duke Kunshan University), Dongsheng Li (Microsoft Research Asia)
TransformerTime SeriesSequential
🎯 What it does: A regularizer based on Lipschitz continuity is proposed, which dynamically adjusts the smoothness of sequential data to match the preferences of different models, thereby enhancing the performance of sequence modeling.
Data Valuation Without Training of a Model
Ki Nohyun, Hye Won Chung (Korea Advanced Institute of Science and Technology)
Data-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: A training-independent data value scoring method called Complexity Gap Score (CG-score) is proposed, which can be calculated directly from the data;
Data-Free One-Shot Federated Learning Under Very High Statistical Heterogeneity
Clare Elizabeth Heinbaugh, Huajie Shao (William and Mary)
Federated LearningKnowledge DistillationAuto EncoderImage
🎯 What it does: Two data-free one-shot federated learning methods, FEDCVAE-ENS and FEDCVAE-KD, are designed and proposed to address extremely high statistical heterogeneity environments. They utilize conditional VAE to reconstruct local learning tasks and aggregate decoders through knowledge distillation or ensemble methods.
Dataless Knowledge Fusion by Merging Weights of Language Models
Xisen Jin (University of Southern California), Pengxiang Cheng (Bloomberg)
Domain AdaptationKnowledge DistillationTransformerLarge Language ModelTextBenchmark
🎯 What it does: The research achieves knowledge fusion by merging multiple model weights without accessing the training data, proposing a new model merging method called Regression Mean (RegMean).
Dataset Pruning: Reducing Training Data by Examining Generalization Influence
Shuo Yang (University of Technology Sydney), Ping Li
OptimizationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: A sample selection method based on influence functions—Dataset Pruning—is proposed to filter redundant samples from large-scale training sets, maximizing the number of retained samples while ensuring that the model's generalization error does not exceed a preset threshold.
DAVA: Disentangling Adversarial Variational Autoencoder
Benjamin Estermann (ETH Zurich), Roger Wattenhofer (ETH Zurich)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A self-supervised variational autoencoder training method called DAVA is proposed, and the PIPE metric is introduced for unsupervised separation measurement.
DaxBench: Benchmarking Deformable Object Manipulation with Differentiable Physics
Siwei Chen (National University of Singapore), David Hsu (Sea AI Lab)
Robotic IntelligenceReinforcement LearningBenchmarkPhysics Related
🎯 What it does: This paper presents DaXBench—a differentiable simulation framework that supports liquids, ropes, fabrics, and elastoplastic materials for evaluating various manipulation methods for deformable objects.
DBQ-SSD: Dynamic Ball Query for Efficient 3D Object Detection
Jinrong Yang (Huazhong University of Science and Technology), Nanning Zheng (Xi'an Jiaotong University)
Object DetectionAutonomous DrivingComputational EfficiencyPoint Cloud
🎯 What it does: This paper proposes a Dynamic Ball Query (DBQ) module, which adaptively selects query points and dynamically allocates different scales of sampling radii in a single-stage point cloud object detector, thereby selectively discarding background points while retaining foreground information, significantly improving inference speed.
DCI-ES: An Extended Disentanglement Framework with Connections to Identifiability
Cian Eastwood (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Representation LearningImage
🎯 What it does: This paper proposes an extended DCI-ES framework in discrete representation learning to quantify the decoupling, completeness, information content, explicitness, and size of representations;
DDM$^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models
Tiange Xiang (Stanford University), Akshay Chaudhari (Stanford University)
RestorationDiffusion modelBiomedical DataMagnetic Resonance ImagingDiffusion Tensor Imaging
🎯 What it does: A self-supervised diffusion model DDM2 is proposed for denoising low signal-to-noise ratio diffusion MRI under unpaired data conditions.
De Novo Molecular Generation via Connection-aware Motif Mining
Zijie Geng (University of Science and Technology of China), Tie-Yan Liu (Microsoft Research AI4Science)
GenerationDrug DiscoveryGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A connection-aware molecular motif mining and generation framework called MiCaM is proposed, which realizes the data-driven mining of connection-aware motifs and uses them to generate new molecules.
DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing
Pengcheng He (Microsoft), Weizhu Chen (Microsoft)
TransformerLarge Language ModelText
🎯 What it does: DeBERTaV3 is proposed by combining the Disentangled Attention of DeBERTa with the Replaced Token Detection (RTD) of ELECTRA, improving the pre-training tasks and introducing a new embedding sharing method.
DecAF: Joint Decoding of Answers and Logical Forms for Question Answering over Knowledge Bases
Donghan Yu (Carnegie Mellon University), Bing Xiang (Amazon Web Services)
TransformerPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the DECAF framework, which jointly generates answers and logical forms and combines execution results to obtain the final answer, using a retrieval-based textual knowledge base instead of entity linking;
DeCap: Decoding CLIP Latents for Zero-Shot Captioning via Text-Only Training
Wei Li (Zhejiang University), Yi Yang (Zhejiang University)
GenerationRetrievalTransformerVision Language ModelContrastive LearningImageVideoText
🎯 What it does: Proposes the DeCap framework, which uses a text decoder to generate zero-shot image/video captions in the multimodal latent space of CLIP, requiring only text data for training, and during inference, maps image embeddings to the text space through a zero-training projection.
Decentralized Optimistic Hyperpolicy Mirror Descent: Provably No-Regret Learning in Markov Games
Wenhao Zhan, Zhuoran Yang
OptimizationReinforcement Learning
🎯 What it does: Under general function approximation, a decentralized 'DORIS' algorithm is proposed to achieve no-regret learning against adversarial opponents in Markov games and to find an approximate coarse correlated equilibrium in self-play.
Decepticons: Corrupted Transformers Breach Privacy in Federated Learning for Language Models
Liam H Fowl, Tom Goldstein (University of Maryland)
Federated LearningSafty and PrivacyAdversarial AttackTransformerText
🎯 What it does: Proposes an attack scheme where a malicious server exploits the parameters of a deformed Transformer to leak user text in federated learning;
Decision S4: Efficient Sequence-Based RL via State Spaces Layers
Shmuel Bar David, Lior Wolf (Tel Aviv University)
TransformerReinforcement LearningSequential
🎯 What it does: An efficient serialized reinforcement learning method based on the S4 state space layer is proposed, covering two stages: offline sequence training and online fine-tuning.
Decision Transformer under Random Frame Dropping
Kaizhe Hu (Tsinghua University), Huazhe Xu (Tsinghua University)
Robotic IntelligenceTransformerReinforcement LearningVideo
🎯 What it does: Developed an offline reinforcement learning algorithm DeFog based on Decision Transformer, capable of achieving robust decision-making in remote control environments with frame loss.
Decomposed Prompting: A Modular Approach for Solving Complex Tasks
Tushar Khot (Allen Institute for AI), Ashish Sabharwal (Allen Institute for AI)
Large Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the Decomposed Prompting (DECOMP) framework, which utilizes a small number of examples to decompose complex tasks into several sub-tasks, each completed with dedicated LLM prompts or symbolic modules;
Decompositional Generation Process for Instance-Dependent Partial Label Learning
Congyu Qiao (Southeast University), Xin Geng (Southeast University)
ClassificationOptimizationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an instance-dependent partial label learning method based on a decomposition generative process, IDGP, which explicitly models the candidate label generation using MAP optimization.
Deconstructing Distributions: A Pointwise Framework of Learning
Gal Kaplun (Harvard University), Preetum Nakkiran (University of California San Diego)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes pointwise learning profiles (accuracy profile and softmax profile) to systematically study the performance of single input points across models of different resource levels, revealing the structural characteristics of data points and models, constructing a negatively correlated OOD dataset CIFAR-10-NEG, and quantitatively analyzing the impact of pre-training.
Decoupled Training for Long-Tailed Classification With Stochastic Representations
Giung Nam (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
ClassificationKnowledge DistillationImage
🎯 What it does: This study adopts a separable learning framework for long-tail classification tasks, first utilizing Stochastic Weight Averaging (SWA) to enhance the generalization ability of the feature extractor, and then generating random representations through SWAG for self-distillation, further improving the classifier's decision boundary and uncertainty estimation.
Deep Declarative Dynamic Time Warping for End-to-End Learning of Alignment Paths
Ming Xu (Queensland University of Technology), Stephen Gould (Australian National University)
Autonomous DrivingOptimizationMultimodalityTime SeriesAudio
🎯 What it does: This paper proposes a differentiable Dynamic Time Warping (DTW) layer—DecDTW—that enables end-to-end learning of alignment paths for time series in deep networks.
Deep Ensembles for Graphs with Higher-order Dependencies
Steven Krieg, Nitesh Chawla
Representation LearningGraph Neural NetworkGraph
🎯 What it does: A deep graph ensemble method (DGE) is proposed for graph data with high-order dependencies, training multiple GNNs in different high-order subspaces of the same node to capture neighborhood variance and enhance representation learning.
Deep Generative Modeling on Limited Data with Regularization by Nontransferable Pre-trained Models
Yong Zhong (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: The Reg-DGM framework is proposed, using a pre-trained non-transferable model as a regularization term to train deep generative models, in order to reduce variance on limited data and improve generation quality.
Deep Generative Symbolic Regression
Samuel Holt (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
GenerationOptimizationTransformerReinforcement LearningTabularBenchmarkPhysics Related
🎯 What it does: A deep generative symbolic regression (DGSR) framework is proposed, which utilizes a pre-trained conditional generative model to learn the invariance of equations and data, and enhances the performance of symbolic regression in high-dimensional variables through gradient refinement during inference.
Deep Learning From Crowdsourced Labels: Coupled Cross-Entropy Minimization, Identifiability, and Regularization
Shahana Ibrahim (Oregon State University), Xiao Fu (Oregon State University)
ClassificationData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: An end-to-end learning framework based on Coupled Cross-Entropy Minimization (CCEM) is proposed for noisy labels in crowdsourcing, along with theoretical identifiability under limited samples, missing labels, and weak annotator correlation, as well as two geometric regularization improvement methods.
Deep Learning meets Nonparametric Regression: Are Weight-Decayed DNNs Locally Adaptive?
Kaiqi Zhang (University of California Santa Barbara), Yu-Xiang Wang (University of California Santa Barbara)
Supervised Fine-TuningTabular
🎯 What it does: This paper studies the performance of neural networks in nonparametric regression problems, particularly their ability to estimate functions with heterogeneous smoothness.
Deep Learning on Implicit Neural Representations of Shapes
Luca De Luigi (University of Bologna), Luigi di Stefano
ClassificationSegmentationGenerationRetrievalGenerative Adversarial NetworkPoint CloudMesh
🎯 What it does: In this work, the authors propose the inr2vec framework, which can directly obtain compact latent vectors from the weights of a single implicit neural representation (INR) and utilize these vectors to perform downstream 3D tasks such as retrieval, classification, segmentation, and generation.
Deep Ranking Ensembles for Hyperparameter Optimization
Abdus Salam Khazi (University of Freiburg), Josif Grabocka (University of Freiburg)
OptimizationHyperparameter SearchMeta LearningBenchmark
🎯 What it does: A deep ranking ensemble model (Deep Ranking Ensembles) based on learning to rank is proposed for hyperparameter search in Bayesian optimization.
Deep Reinforcement Learning for Cost-Effective Medical Diagnosis
Zheng Yu (Princeton University), Mengdi Wang (Princeton University)
Anomaly DetectionOptimizationReinforcement LearningTabularBiomedical DataElectronic Health Records
🎯 What it does: A dynamic medical diagnosis framework based on reinforcement learning is proposed, which can select experimental test panels and make diagnostic decisions at a limited cost;
Deep Transformers without Shortcuts: Modifying Self-attention for Faithful Signal Propagation
Bobby He (University of Oxford), Yee Whye Teh (DeepMind)
TransformerLarge Language ModelText
🎯 What it does: This paper studies how to train deep Transformer networks without using residual skip connections and normalization layers, and proposes three methods (E-SPA, U-SPA, Value-SkipInit) to maintain signal propagation and avoid rank collapse.
Deep Variational Implicit Processes
Luis A. Ortega (Universidad Autónoma de Madrid), Daniel Hernández-Lobato (Universidad Autónoma de Madrid)
TabularTime Series
🎯 What it does: This paper proposes a Deep Variational Implicit Process (DVIP) model, which stacks implicit processes (IP) in a multi-layer form to construct a flexible function space prior and achieve scalable variational inference.
Defending against Adversarial Audio via Diffusion Model
Shutong Wu (Arizona State University), Chaowei Xiao (Arizona State University)
GenerationAdversarial AttackDiffusion modelAudio
🎯 What it does: This paper proposes AudioPure, a defense method for audio purification using diffusion models against adversarial attacks.
Deja Vu: Continual Model Generalization for Unseen Domains
Chenxi Liu (Northwestern University), Qi Zhu (Northwestern University)
Domain AdaptationKnowledge DistillationContrastive LearningImage
🎯 What it does: A framework named RaTP is proposed to address the performance degradation of models during the 'strange period' in continuous domain drift environments, capable of providing good performance immediately when a new domain appears while retaining memory of the old domain.
DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTATION
Bowen Zhao (Tsinghua University), Shu-Tao Xia (Tsinghua University)
Domain AdaptationSupervised Fine-TuningImageVideo
🎯 What it does: An improved scheme is proposed for Fully Test-Time Adaptation, which can adjust the pre-trained model in real-time and in a one-time manner during the testing stream.
Delving into Semantic Scale Imbalance
Yanbiao Ma (Xidian University), Xu Liu (Xidian University)
ClassificationRecognitionImageBiomedical Data
🎯 What it does: This paper proposes the use of subspace volume quantization to define semantic scales, and based on this, it studies the imbalance of semantic scales and its impact on model bias, proposing a dynamic semantic scale balancing learning method.
Denoising Diffusion Error Correction Codes
Yoni Choukroun (Tel Aviv University), Lior Wolf (Tel Aviv University)
Diffusion model
🎯 What it does: Using a denoising diffusion model to implement soft decoding of linear codes, treating channel distortion as a forward diffusion process, and gradually recovering codewords during the reverse diffusion process.
Denoising Diffusion Samplers
Francisco Vargas (University of Cambridge), Arnaud Doucet (DeepMind)
GenerationData SynthesisOptimizationDiffusion modelTabularTime SeriesBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A sampling framework based on Denoising Diffusion Sampler (DDS) is proposed, which samples from unnormalized target distributions and estimates the normalization constant by learning the time reversal of the diffusion process.
Denoising Masked Autoencoders Help Robust Classification
QuanLin Wu, Di He (Peking University)
ClassificationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: A self-supervised method called Denoising Masked AutoEncoders (DMAE) is proposed, which trains a Vision Transformer encoder by adding Gaussian noise to images and masking certain blocks to reconstruct the original image, serving directly as a base classifier for random smoothing robust classification.
DENSE RGB SLAM WITH NEURAL IMPLICIT MAPS
Heng Li (Hong Kong University of Science and Technology), Ping Tan (Hong Kong University of Science and Technology)
Pose EstimationOptimizationRobotic IntelligenceSimultaneous Localization and MappingImage
🎯 What it does: A dense RGB SLAM method utilizing neural implicit mapping has been developed, capable of achieving camera pose tracking and panoramic 3D scene reconstruction using only a standard RGB camera.
DensePure: Understanding Diffusion Models for Adversarial Robustness
Chaowei Xiao (Arizona State University), Dawn Song (University of California Berkeley)
OptimizationAdversarial AttackTransformerDiffusion modelScore-based ModelImage
🎯 What it does: The DensePure method is proposed, which utilizes the reverse process of the diffusion model to denoise the input multiple times and obtains the final prediction through voting, thereby achieving provable improvements in adversarial robustness.
DEP-RL: Embodied Exploration for Reinforcement Learning in Overactuated and Musculoskeletal Systems
Pierre Schumacher (Max Planck Institute for Intelligent Systems), Georg Martius (Max Planck Institute for Intelligent Systems)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes the use of the self-organizing learning rule DEP for exploration in reinforcement learning, enabling efficient motion control in large-scale over-actuated muscular systems.
Depth Separation with Multilayer Mean-Field Networks
Yunwei Ren (Carnegie Mellon University), Rong Ge (Duke University)
🎯 What it does: Proves that a three-layer neural network can learn a target function (ReLU(1-||x||)) that can only be approximated by a three-layer network and not by a two-layer network in polynomial time, thus achieving algorithmic depth separation.
DepthFL : Depthwise Federated Learning for Heterogeneous Clients
Minjae Kim (Seoul National University), Soo-Mook Moon (Seoul National University)
Federated LearningKnowledge DistillationImageText
🎯 What it does: This paper proposes DepthFL, a federated learning framework that constructs local models of different scales through depth scaling rather than width scaling, and enhances the performance of the global model by utilizing mutual distillation among multi-layer classifiers.
Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling
Keyu Tian (Peking University), Zehuan Yuan (Bytedance Inc)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The SparK framework is proposed, utilizing sparse convolution and a hierarchical decoder to achieve BERT-style masked image pre-training for convolutional networks.
Deterministic training of generative autoencoders using invertible layers
Gianluigi Silvestri (OnePlanet Research Center), Luca Ambrogioni (Donders Institute for Brain Cognition and Behaviour)
GenerationData SynthesisFlow-based ModelAuto EncoderImage
🎯 What it does: A deterministic generative autoencoder (AEF) constructed using reversible layers is proposed, achieving maximum likelihood training for traditional variational autoencoders (VAE);
DexDeform: Dexterous Deformable Object Manipulation with Human Demonstrations and Differentiable Physics
Sizhe Li (Massachusetts Institute of Technology), Chuang Gan (University of Massachusetts Amherst)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningPoint CloudPhysics Related
🎯 What it does: This paper proposes a framework called DexDeform, based on human demonstrations and differentiable physics, for learning dexterous manipulation of flexible objects by multi-fingered robotic hands.
DFlow: Learning to Synthesize Better Optical Flow Datasets via a Differentiable Pipeline
Kwon Byung-Ki (POSTECH), Tae-Hyun Oh (Yonsei University)
Data SynthesisDomain AdaptationOptimizationContrastive LearningOptical FlowImageVideo
🎯 What it does: A differentiable optical flow data generation pipeline DFlow is proposed, which automatically synthesizes a better synthetic optical flow dataset for the target domain by optimizing the generation parameters through gradient optimization.
DFPC: Data flow driven pruning of coupled channels without data.
Tanay Narshana (Observe.AI), Chiranjib Bhattacharyya (Indian Institute of Science)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A data-free structured pruning method (DFPC) is proposed, specifically targeting the coupling channels in multi-branch convolutional networks for pruning.
Diagnosing and Rectifying Vision Models using Language
Yuhui Zhang (Stanford University), Serena Yeung (Stanford University)
ClassificationDomain AdaptationExplainability and InterpretabilityPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the DrML framework, which uses text embeddings generated from natural language as proxies for image embeddings, enabling the diagnosis and correction of visual classifier errors without visual samples.
Dichotomy of Control: Separating What You Can Control from What You Cannot
Sherry Yang, Ofir Nachum (Google Research)
TransformerReinforcement LearningContrastive LearningTabularSequential
🎯 What it does: A Dichotomy of Control (DoC) framework is proposed, which separates controllable actions of the policy from uncontrollable environmental randomness using future conditional supervised learning and mutual information constraints, ensuring conditional consistency for given rewards in offline reinforcement learning.
DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking
Gabriele Corso (Massachusetts Institute of Technology), Tommi S. Jaakkola
GenerationOptimizationDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A molecular docking method called DIFFDOCK based on diffusion generative models is proposed, which generates 3D structures for protein binding in the ligand pose space through the diffusion process of displacement, rotation, and torsion angles.
DiffEdit: Diffusion-based semantic image editing with mask guidance
Guillaume Couairon (Meta AI), Matthieu Cord (Sorbonne Université)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: The paper presents DIFFEDIT, a method for semantic image editing that utilizes diffusion models without the need for manual masks.
Differentiable Gaussianization Layers for Inverse Problems Regularized by Deep Generative Models
Dongzhuo Li (ExxonMobil Technology and Engineering Company)
RestorationOptimizationGenerative Adversarial NetworkImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes a differentiable Gaussianization layer to enforce the latent tensors of deep generative models to maintain a standard Gaussian distribution in inverse problems, thereby improving the quality of the solutions.
Differentiable Mathematical Programming for Object-Centric Representation Learning
Adeel Pervez (Informatics Institute University of Amsterdam), Efstratios Gavves (Informatics Institute University of Amsterdam)
Object TrackingOptimizationRepresentation LearningConvolutional Neural NetworkImageVideoGraph
🎯 What it does: A differentiable graph cut and matching framework is proposed, which approximates the minimum s-t cut problem using equality-constrained quadratic programming and embeds it into convolutional networks for object-centric representation learning.
Differentially Private $L_2$-Heavy Hitters in the Sliding Window Model
Jeremiah Blocki (Purdue University), Samson Zhou (University of California Berkeley and Rice University)
Safty and Privacy
🎯 What it does: This paper proposes an algorithm for achieving differential privacy for the L2 centroid subproblem under the sliding window model, capable of outputting all frequency approximations that meet the threshold within sub-logarithmic space.
Differentially Private Adaptive Optimization with Delayed Preconditioners
Tian Li (Carnegie Mellon University), Virginia Smith (Carnegie Mellon University)
Recommendation SystemOptimizationSafty and PrivacyText
🎯 What it does: The DP2 method is proposed, which uses a delayed preconditioner in differential privacy training to reduce noise and improve convergence speed.
DiffMimic: Efficient Motion Mimicking with Differentiable Physics
Jiawei Ren (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
Robotic IntelligenceReinforcement LearningVideoPhysics Related
🎯 What it does: Using a differentiable physics simulator (DPS) to directly optimize control strategies through state matching, achieving motion imitation of physical characters.
DIFFormer: Scalable (Graph) Transformers Induced by Energy Constrained Diffusion
Qitian Wu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
ClassificationRepresentation LearningGraph Neural NetworkTransformerDiffusion modelGraphTabularTime SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes an energy-constrained diffusion-based Transformer (DIFFORMER) that propagates information across the entire dataset using a learnable full-instance diffusion coefficient, generating representations that balance global consistency and local features.
DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models
Shansan Gong (Shanghai AI Laboratory), Lingpeng Kong (University of Hong Kong)
GenerationTransformerDiffusion modelText
🎯 What it does: A sequence-to-sequence text generation framework called DIFFUSEQ based on diffusion models is proposed, which can simultaneously achieve conditional generation and diversity control within a single model.
DiffusER: Diffusion via Edit-based Reconstruction
Machel Reid (Google Research), Graham Neubig (Carnegie Mellon University)
GenerationTransformerDiffusion modelText
🎯 What it does: A text generation framework based on diffusion models, DIFFUSER, is proposed, which implements stepwise revision from any sequence to the target text using Levenshtein edit operations (insertion, deletion, substitution, retention).
Diffusion Adversarial Representation Learning for Self-supervised Vessel Segmentation
Boah Kim (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
SegmentationRepresentation LearningAdversarial AttackDiffusion modelGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A self-supervised vascular segmentation framework based on Diffusion Adversarial Representation Learning (DARL) is proposed, capable of generating vascular segmentation masks in one go and achieving label-free segmentation of X-ray coronary and retinal images.
Diffusion Models Already Have A Semantic Latent Space
Mingi Kwon (Yonsei University), Youngjung Uh (Yonsei University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: In the frozen diffusion model, an asymmetric reverse process (Asyrp) is designed to discover and utilize U-Net bottleneck features in h-space for semantic image attribute editing.
Diffusion Models for Causal Discovery via Topological Ordering
Pedro Sanchez (University of Edinburgh), Sotirios A. Tsaftaris
Diffusion modelScore-based ModelGraphBiomedical Data
🎯 What it does: This paper proposes a causal discovery method called DiffAN based on the Diffusion Probability Model (DPM), which achieves topological sorting by learning the score of the data distribution and approximating its Hessian, thereby inferring a directed acyclic graph.
Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning
Zhendong Wang (University of Texas), Mingyuan Zhou (University of Texas)
Reinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: Proposes using diffusion models as high-expression strategies for offline reinforcement learning;
Diffusion Posterior Sampling for General Noisy Inverse Problems
Hyungjin Chung (KAIST), Jong Chul Ye (KAIST)
RestorationSuper ResolutionDiffusion modelImageStochastic Differential Equation
🎯 What it does: A posterior sampling method based on diffusion models (Diffusion Posterior Sampling, DPS) is proposed, which can achieve high-quality reconstruction for linear and nonlinear inverse problems under noise (Gaussian or Poisson).
Diffusion Probabilistic Fields
Peiye Zhuang (Stanford University), Miguel Ángel Bautista (Apple)
GenerationData SynthesisTransformerDiffusion modelImagePoint Cloud
🎯 What it does: This paper proposes Diffusion Probabilistic Fields (DPF), a diffusion generative model capable of directly learning the distribution of continuous functions (fields), unifying data generation across different domains such as images, 3D geometry, and spheres.
Diffusion Probabilistic Modeling of Protein Backbones in 3D for the motif-scaffolding problem
Brian L. Trippe (Massachusetts Institute of Technology), Tommi S. Jaakkola
GenerationProtein Structure PredictionGraph Neural NetworkDiffusion modelBiomedical Data
🎯 What it does: A diffusion probabilistic model called ProtDiff based on E(3)-equivariant graph neural networks is proposed to generate complete 3D coordinates of protein backbones; and the SMCDiff algorithm is introduced, which achieves conditional sampling (generating scaffolds under the condition of given motifs) through particle filtering, theoretically allowing for precise conditional samples.
Diffusion-based Image Translation using disentangled style and content representation
Gihyun Kwon (KAIST), Jong Chul Ye (KAIST)
Image TranslationTransformerDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes an image translation framework called DiffuseIT based on an unsupervised diffusion model (DDPM), which utilizes a pre-trained Vision Transformer (ViT) to separate content and style, and achieves high-quality style transfer guided by text or a single target image through various losses during the reverse diffusion process.
Diffusion-GAN: Training GANs with Diffusion
Zhendong Wang (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes Diffusion-GAN, which uses the forward diffusion chain as a method for injecting instance noise and employs a temporal dependency mechanism in the discriminator to achieve gradient backpropagation for the generator, thereby stabilizing GAN training.
DiGress: Discrete Denoising diffusion for graph generation
Clement Vignac, Pascal Frossard (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: This paper presents DiGress, a discrete denoising diffusion model for graphs with categorical nodes and edge attributes, designed to generate sparse and structurally reasonable graphs.
Dilated convolution with learnable spacings
Ismail Khalfaoui Hassani, Timothée Masquelier
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a learnable dilation convolution (DCLS) to expand the receptive field of convolutional networks without increasing the number of parameters.
Diminishing Return of Value Expansion Methods in Model-Based Reinforcement Learning
Daniel Palenicek (Technical University of Darmstadt), Jan Peters (Technical University of Darmstadt)
Reinforcement Learning
🎯 What it does: By using a perfect (oracle) dynamics model and learning model, systematic experiments were conducted on the sampling efficiency of model-based value expansion methods (Critic Expansion and Actor Expansion) in continuous control tasks, and the impact of model accuracy and rollout step length on learning effectiveness was explored.
DINO as a von Mises-Fisher mixture model
Hariprasath Govindarajan (Qualcomm Technologies, Inc.), Fredrik Lindsten (Linköping University)
ClassificationRetrievalRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Treating DINO as a von Mises-Fisher mixture model, DINO-vMF is proposed to achieve more flexible prototype learning and improve representation quality by incorporating a normalization constant;
DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection
Hao Zhang, Heung-Yeung Shum
Object DetectionTransformerContrastive LearningImage
🎯 What it does: This paper proposes an end-to-end object detection model based on DETR called DINO, which improves denoising anchor box training, query initialization, and box prediction methods.
Direct Embedding of Temporal Network Edges via Time-Decayed Line Graphs
Sudhanshu Chanpuriya (University of Massachusetts Amherst), Cameron N Musco
Graph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes a temporary network edge embedding method based on Time Decay Line Graph (TDLG), which directly utilizes continuous timestamps to construct a weighted line graph and uses its row vectors as edge features;
Dirichlet-based Uncertainty Calibration for Active Domain Adaptation
Mixue Xie (Beijing Institute of Technology), Chi Harold Liu (Beijing Institute of Technology)
ClassificationSegmentationDomain AdaptationImage
🎯 What it does: A Dirichlet distribution-based evidence deep learning model is proposed for uncertainty calibration and sample selection in active domain adaptation.
Discovering Evolution Strategies via Meta-Black-Box Optimization
Robert Tjarko Lange (Technical University Berlin), Sebastian Flennerhag (DeepMind)
OptimizationMeta LearningTransformerTabularBenchmark
🎯 What it does: This paper proposes an evolutionary strategy architecture based on self-attention and employs MetaBBO for meta-learning its parameters, allowing it to adaptively update the search distribution in black-box optimization.
Discovering Generalizable Multi-agent Coordination Skills from Multi-task Offline Data
Fuxiang Zhang (Nanjing University), Zongzhang Zhang (Nanjing University)
TransformerReinforcement LearningSequential
🎯 What it does: This paper proposes an algorithm for offline multi-task multi-agent reinforcement learning called ODIS, which can learn general coordination skills from limited sources of multi-task offline data and achieve efficient collaboration without tuning in unseen tasks.
Discovering Informative and Robust Positives for Video Domain Adaptation
Chang Liu (Northeastern University), Yun Fu (Northeastern University)
Domain AdaptationContrastive LearningVideo
🎯 What it does: This paper studies unsupervised video domain adaptation and proposes a non-contrastive learning framework that does not use negative samples. It generates information-rich in-domain positive samples through spatial/temporal augmentation and nearest neighbor generation, and achieves cross-domain alignment by using the reverse source domain as an anchor and sampling from the target distribution to generate robust cross-domain positive samples.
Discovering Latent Knowledge in Language Models Without Supervision
Collin Burns (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
TransformerLarge Language ModelContrastive LearningText
🎯 What it does: Unsupervised discovery of latent knowledge in the internal activations of language models, using logical consistency features to answer yes/no questions.
Discovering Policies with DOMiNO: Diversity Optimization Maintaining Near Optimality
Tom Zahavy (DeepMind), Satinder Singh (DeepMind)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes the DOMiNO algorithm, which utilizes the diversity objective of state-action occupancy measures to generate diverse policies while maintaining near-optimal rewards.
Discrete Contrastive Diffusion for Cross-Modal Music and Image Generation
Ye Zhu (Illinois Institute of Technology), Yan Yan (Illinois Institute of Technology)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningImageMultimodalityAudio
🎯 What it does: A Conditional Discrete Contrastive Diffusion (CDCD) model is proposed, which explicitly maximizes the mutual information between conditional inputs and generated outputs for cross-modal music and image generation.
Discrete Predictor-Corrector Diffusion Models for Image Synthesis
Jose Lezama, Irfan Essa (Georgia Institute of Technology)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper introduces a discrete prediction-correction diffusion model (DPC) that corrects the concurrent decoding errors of non-autoregressive Transformers by learning a discrete MCMC correction kernel, thereby improving the quality and efficiency of image synthesis.
Disentanglement of Correlated Factors via Hausdorff Factorized Support
Karsten Roth (University of Tübingen), Diane Bouchacourt (Meta AI)
GenerationRepresentation LearningAuto EncoderImage
🎯 What it does: This paper proposes a Hausdorff distance-based Factorized Support (HFS) criterion to separate generative factors influenced by correlations under unsupervised conditions, thereby obtaining more interpretable and generalizable representations.
Disentanglement with Biological Constraints: A Theory of Functional Cell Types
James C. R. Whittington, Timothy Behrens
Representation LearningMultimodality
🎯 What it does: This paper proves and experimentally verifies that biological constraints (neuron non-negativity and energy efficiency) can lead to the decoupling of representational solutions in linear networks. It extends this principle to more general nonlinear networks and self-supervised learning, and through simulations, it illustrates how these constraints explain the emergence of functional cell types such as grid cells and object vector cells in the brain, as well as the deformation phenomena of their spatial fields.
Disentangling Learning Representations with Density Estimation
Eric Yeats (Duke University), Hai Li (Duke University)
Representation LearningAuto EncoderImage
🎯 What it does: This paper proposes the Gaussian Channel Autoencoder (GCAE), an unsupervised learning method that achieves reliable disentanglement using Gaussian noise and conditional density estimation.
Disentangling the Mechanisms Behind Implicit Regularization in SGD
Zachary Novack (University of California San Diego), Zachary Chase Lipton
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: Comparatively and empirically evaluate whether various implicit regularizations (gradient norm, Fisher information, Jacobian regularization) can recover the generalization performance of small-batch SGD in large-batch SGD, and study the effects of micro-batch size and dataset on their performance.
Disparate Impact in Differential Privacy from Gradient Misalignment
Maria S. Esipova (Layer 6 AI), Jesse C Cresswell
OptimizationSafty and PrivacyTabular
🎯 What it does: This study investigates the shortcomings of Differentially Private Stochastic Gradient Descent (DPSGD) in terms of fairness, particularly the differential impact caused by gradient mismatch, and proposes a method to eliminate gradient mismatch and reduce unfairness through global scaling.
Distilling Cognitive Backdoor Patterns within an Image
Hanxun Huang (University of Melbourne), James Bailey (University of Melbourne)
Anomaly DetectionKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Cognitive Distillation (CD) method that automatically extracts the minimal patterns (cognitive patterns) from images that can determine model predictions, and uses these patterns for backdoor sample detection and bias identification.