ICLR 2024 Papers — Page 5
International Conference on Learning Representations · 2260 papers
Contrastive Preference Learning: Learning from Human Feedback without Reinforcement Learning
Joey Hejna (Stanford University), Dorsa Sadigh (Stanford University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningImage
🎯 What it does: A reinforcement learning method based on contrastive learning is proposed—Contrastive Preference Learning (CPL), which learns optimal policies directly from human preference data, eliminating the need for a reward model and RL optimization steps.
Controlled Text Generation via Language Model Arithmetic
Jasper Dekoninck (ETH Zurich), Martin Vechev (ETH Zurich)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A framework called model arithmetic is proposed, which utilizes weighted KL divergence to combine multiple LLMs and attribute models, supporting fine-grained text control through formulas, and incorporates a novel joint operator and extended speculative sampling for efficient inference.
Controlling Vision-Language Models for Multi-Task Image Restoration
Ziwei Luo (Uppsala University), Thomas B. Schön
RestorationTransformerVision Language ModelDiffusion modelContrastive LearningImage
🎯 What it does: Train a degradation-aware CLIP (DA-CLIP) controller to enable the pre-trained CLIP image encoder to output high-quality features from degraded images and predict the type of degradation, then integrate these embeddings with an image restoration network to enhance multi-task image restoration performance.
ControlVideo: Training-free Controllable Text-to-video Generation
Yabo Zhang (Harbin Institute of Technology), Qi Tian (Huawei Cloud)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: A training-free ControlVideo framework is proposed, utilizing text prompts and motion sequences (such as depth maps and edge maps) to achieve controllable video generation.
Convergence of Bayesian Bilevel Optimization
Shi Fu (University of Science and Technology of China), Dacheng Tao (Nanyang Technological University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: The theoretical convergence guarantee of Bayesian Bi-level Optimization (BBO) is proposed, along with a sublinear tuning loss upper bound for the simultaneous convergence of model parameters and hyperparameters.
Conversational Drug Editing Using Retrieval and Domain Feedback
Shengchao Liu (University of California), Chaowei Xiao (University of Wisconsin-Madison)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: The ChatDrug framework is proposed, utilizing conversational LLMs for drug editing, supporting three types of drugs: small molecules, peptides, and proteins.
Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model
Zihan Zhong (Tsinghua University), Chun Yuan (Tsinghua University)
SegmentationTransformerMixture of ExpertsImageAgriculture Related
🎯 What it does: This paper proposes Conv-LoRA, a PEFT method that fine-tunes only a few parameters on the SAM pre-trained ViT encoder to improve multi-domain semantic segmentation.
Convolutional Deep Kernel Machines
Edward Milsom (University of Bristol), Laurence Aitchison (University of Bristol)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the Convolutional Deep Kernel Machine (DKM) and designs an efficient cross-domain inducing point approximation method, achieving Bayesian representation learning for infinitely wide convolutional networks.
Coordinate-Aware Modulation for Neural Fields
Joo Chan Lee (Sungkyunkwan University), Eunbyung Park (Sungkyunkwan University)
GenerationCompressionNeural Radiance FieldImageVideo
🎯 What it does: Proposes Coordinate-Aware Modulation (CAM), which parallelly injects scaling and translation parameters generated by a single-channel grid into the intermediate features of the MLP, thereby alleviating spectral bias.
Copilot4D: Learning Unsupervised World Models for Autonomous Driving via Discrete Diffusion
Lunjun Zhang (Waabi), Raquel Urtasun (Waabi)
Autonomous DrivingTransformerDiffusion modelAuto EncoderWorld ModelPoint Cloud
🎯 What it does: This paper proposes Copilot4D, which utilizes VQVAE to discretize LiDAR point clouds and then employs an improved discrete diffusion model (a variant of MaskGIT) to predict future point clouds, achieving an unsupervised world model.
COPlanner: Plan to Roll Out Conservatively but to Explore Optimistically for Model-Based RL
Xiyao Wang (University of Maryland), Furong Huang (University of Maryland)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: Proposes the COPlanner framework, which combines conservative model rollouts and optimistic environment exploration to mitigate model errors.
Copula Conformal prediction for multi-step time series prediction
Sophia Huiwen Sun (University of California San Diego), Rose Yu (University of California San Diego)
Time Series
🎯 What it does: The CopulaCPTS algorithm is proposed, which uses copula and induced consistency prediction to generate confidence intervals for multi-step time series.
CoRe-GD: A Hierarchical Framework for Scalable Graph Visualization with GNNs
Florian Grötschla (ETH Zurich), Roger Wattenhofer (ETH Zurich)
Graph Neural NetworkGraphBenchmark
🎯 What it does: This paper presents CoRe-GD, a scalable graph visualization framework that utilizes hierarchical graph refinement and position reconnection techniques to learn and optimize the pressure function of graphs and output node layouts.
CORN: Contact-based Object Representation for Nonprehensile Manipulation of General Unseen Objects
Yoonyoung Cho (Korea Advanced Institute of Science and Technology), Beomjoon Kim (Korea Advanced Institute of Science and Technology)
Knowledge DistillationRobotic IntelligenceTransformerReinforcement LearningPoint Cloud
🎯 What it does: A contact-aware object representation (CORN) and a lightweight patch-based transformer are proposed for efficiently training non-grasping manipulation policies (push, flip, roll) in simulation, achieving zero-shot transfer to real objects.
Correlated Noise Provably Beats Independent Noise for Differentially Private Learning
Christopher A. Choquette-Choo (Google), Abhradeep Guha Thakurta (Google)
OptimizationSafty and PrivacyImageText
🎯 What it does: This paper proposes an algorithm DP-FTRL that uses correlated Gaussian noise in differential privacy learning and provides a theoretical proof that it outperforms traditional DP-SGD.
COSA: Concatenated Sample Pretrained Vision-Language Foundation Model
Sihan Chen (University of Chinese Academy of Sciences), Jing Liu
RetrievalTransformerVision Language ModelContrastive LearningImageVideoText
🎯 What it does: During the pre-training phase, pseudo video-paragraph corpora are constructed by online concatenation of multiple images and corresponding texts, thereby jointly learning static and event-level temporal representations of vision and language.
CoT3DRef: Chain-of-Thoughts Data-Efficient 3D Visual Grounding
Eslam Mohamed BAKR, Mohamed Elhoseiny (King Abdullah University of Science and Technology)
Object DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPoint CloudChain-of-Thought
🎯 What it does: An interpretable 3D visual localization framework called CoT3DRef is proposed, which transforms the 3D visual localization task into a sequence-to-sequence (Seq2Seq) problem, first predicting anchor point chains and then localizing the target object.
Counterfactual Density Estimation using Kernel Stein Discrepancies
Diego Martinez-Taboada (Carnegie Mellon University), Edward Kennedy
OptimizationTabular
🎯 What it does: This paper studies a doubly robust minimum kernel Stein discrepancy estimator for density estimation of causal counterfactual distributions under energy models without the need for a regularization constant.
Counting Graph Substructures with Graph Neural Networks
Charilaos Kanatsoulis, Alejandro Ribeiro (University of Pennsylvania)
Graph Neural NetworkGraph
🎯 What it does: This paper analyzes and proves that message-passing graph neural networks (GNNs) can generate equivariant representations through appropriate activation and normalization layers using random uninformative node inputs, thereby counting and enumerating substructures such as cycles, cliques, quasi-cliques, and connected components in graphs, and can generalize to unknown graphs.
Course Correcting Koopman Representations
Mahan Fathi (Google DeepMind), Ross Goroshin (Google DeepMind)
Representation LearningReinforcement LearningAuto EncoderTime SeriesSequential
🎯 What it does: In this paper, the authors investigate a model that combines Koopman theory with autoencoders to learn linear latent representations of nonlinear dynamical systems, and propose an inference mechanism called Periodic Reencoding to enhance the accuracy and stability of long-term trajectory predictions.
CoVLM: Composing Visual Entities and Relationships in Large Language Models Via Communicative Decoding
Junyan Li (University of Massachusetts Amherst), Chuang Gan (University of Massachusetts Amherst)
Object DetectionGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: We propose CoVLM, a multimodal large language model that achieves visual decoding of visual entities and relationships through communication tokens.
CPPO: Continual Learning for Reinforcement Learning with Human Feedback
Han Zhang (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A continuous learning RLHF method called CPPO is proposed, which utilizes sample-level weight balancing strategies for learning and knowledge retention, enhancing the model's adaptability to the ever-changing human preferences.
CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets
Lifan Yuan (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
OptimizationAI Code AssistantTransformerLarge Language ModelTextTabularRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: The CRAFT framework is proposed to automatically generate and retrieve a set of tools for specific tasks to enhance the reasoning and execution capabilities of LLMs.
CrIBo: Self-Supervised Learning via Cross-Image Object-Level Bootstrapping
Tim Lebailly (KU Leuven), Tinne Tuytelaars (KU Leuven)
Object DetectionSegmentationRetrievalKnowledge DistillationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A cross-image object-level self-supervised learning method called CrIBo is proposed, which enhances dense visual representation learning by using object-level nearest neighbor guidance during the training process.
CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing
Zhibin Gou (Tsinghua University), Weizhu Chen (Microsoft Research Asia)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposes the CRITIC framework, allowing frozen LLMs to self-validate and correct outputs through interactive tools;
Critical Learning Periods Emerge Even in Deep Linear Networks
Michael Kleinman (Stanford University), Stefano Soatto (University of California Los Angeles)
TabularOrdinary Differential Equation
🎯 What it does: This study investigates the phenomenon of critical learning periods in deep linear networks and verifies its dependence on depth and data structure through analytical differential equations and numerical simulations.
Cross-Modal Contextualized Diffusion Models for Text-Guided Visual Generation and Editing
Ling Yang (Peking University), Bin CUI
GenerationData SynthesisTransformerDiffusion modelImageVideoTextMultimodality
🎯 What it does: A cross-modal contextual diffusion model called CONTEXTDIFF is proposed, which enhances text-guided visual generation and editing by incorporating cross-modal context into the forward and backward diffusion processes.
CrossLoco: Human Motion Driven Control of Legged Robots via Guided Unsupervised Reinforcement Learning
Tianyu Li (Georgia Institute of Technology), Sehoon Ha (Georgia Institute of Technology)
Robotic IntelligenceReinforcement LearningTime Series
🎯 What it does: The CrossLoco framework is proposed, which achieves human motion-driven control of quadruped robots through unsupervised reinforcement learning, capable of mapping diverse human actions (such as walking, running, dancing) to robot movements.
CrossQ: Batch Normalization in Deep Reinforcement Learning for Greater Sample Efficiency and Simplicity
Aditya Bhatt (German Research Center for Artificial Intelligence), Jan Peters (German Research Center for Artificial Intelligence)
Reinforcement LearningSequential
🎯 What it does: The Cross Q algorithm is proposed, improving sample efficiency in deep reinforcement learning under continuous control by removing the target network and incorporating batch normalization and a wider critic network, achieving faster convergence while maintaining UTD=1.
Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding
Tatsunori Taniai (OMRON SINIC X Corporation), Kanta Ono (Osaka University)
TransformerGraphPhysics Related
🎯 What it does: A Transformer-based crystal structure encoder called Crystalformer is proposed, which achieves deep feature space aggregation of infinitely repeating atoms in crystals through a physical interpretation of infinite connected attention for crystal periodicity, completing regression predictions for various crystal properties.
Curiosity-driven Red-teaming for Large Language Models
Zhang-Wei Hong (Massachusetts Institute of Technology), Pulkit Agrawal
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: By combining reinforcement learning with curiosity-driven exploration, this paper automates the generation of diverse and effective red team test cases to evaluate the robustness of large language models (LLMs) in producing toxic, erroneous, or harmful content.
Curriculum reinforcement learning for quantum architecture search under hardware errors
Yash J. Patel (Leiden University), Onur Danaci (Leiden University)
OptimizationReinforcement LearningSequentialPhysics Related
🎯 What it does: Developed and evaluated a curriculum learning-based reinforcement learning quantum architecture search algorithm CRLQAS, aimed at automatically generating quantum circuits for VQE in noisy quantum chemistry problems.
Customizable Combination of Parameter-Efficient Modules for Multi-Task Learning
Haowen Wang (Ant Group), Cong Fan (Ant Group)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: The C-Poly model is proposed, which achieves parameter-efficient fine-tuning for multi-task learning by separating task-general and task-specific low-rank adapters.
Cycle Consistency Driven Object Discovery
Aniket Rajiv Didolkar (Mila University of Montreal), Yoshua Bengio (Mila University of Montreal)
Object DetectionSegmentationReinforcement LearningImageVideo
🎯 What it does: Two types of cyclic consistency losses (SLOT-FEATURE-SLOT and FEATURE-SLOT-FEATURE) are proposed to enhance slot-based object discovery models, ensuring that each object corresponds to a unique slot and different slots correspond to different objects.
DAFA: Distance-Aware Fair Adversarial Training
Hyungyu Lee (Seoul National University), Sungroh Yoon (Seoul National University)
ClassificationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a distance-aware fair adversarial training (DAFA) based on inter-class similarity, which enhances the robustness of difficult-to-classify categories by assigning them greater loss weights and adversarial margins.
DAM: Towards a Foundation Model for Forecasting
Luke Nicholas Darlow (Huawei), Amos Storkey (University of Edinburgh)
TransformerTime SeriesFinance Related
🎯 What it does: This paper proposes DAM (Deep Data-Dependent Approximate Analytical Model), a model capable of unified forecasting for time series across different domains, sampling frequencies, and prediction requirements.
Data Debugging with Shapley Importance over Machine Learning Pipelines
Bojan Karlaš (Harvard University), Ce Zhang (University of Chicago)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningImageTextTabular
🎯 What it does: Proposes the Datascope framework for efficiently computing Shapley values across a complete machine learning pipeline (including data preprocessing, feature extraction, etc.) to achieve data error localization and correction.
Data Distillation Can Be Like Vodka: Distilling More Times For Better Quality
Xuxi Chen (University of Texas), Baharan Mirzasoleiman (University of California)
Data SynthesisKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A multi-stage dataset distillation (PDD) framework is proposed, which recursively generates small synthetic subsets at different training stages, capturing the complete training dynamics step by step based on the previous stage's subset, thereby enhancing the generalization performance of the distilled data.
Data Filtering Networks
Alex Fang (Apple), Vaishaal Shankar (Apple)
RetrievalComputational EfficiencyData-Centric LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: Construct and train a Data Filtering Network (DFN) to automatically filter high-quality samples from a vast amount of unlabeled image-text pairs, thereby generating large pre-training datasets (DFN-2B, DFN-5B) and training CLIP variants on these datasets.
Data-independent Module-aware Pruning for Hierarchical Vision Transformers
Yang He (Agency for Science Technology and Research), Joey Tianyi Zhou (Agency for Science Technology and Research)
ClassificationCompressionTransformerImage
🎯 What it does: A data-independent module-aware pruning method called DIMAP is proposed for hierarchical vision Transformers (Swin Transformer), aimed at compressing model parameters and computational load without relying on input data.
DataInf: Efficiently Estimating Data Influence in LoRA-tuned LLMs and Diffusion Models
Yongchan Kwon (Columbia University), James Zou (Stanford University)
GenerationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A method for efficiently calculating the data influence (influence function) in large generative models (LLMs and diffusion models) is proposed—DataInf.
DATS: Difficulty-Aware Task Sampler for Meta-Learning Physics-Informed Neural Networks
Maryam Toloubidokhti (Rochester Institute of Technology), Linwei Wang (Rochester Institute of Technology)
OptimizationMeta LearningTabularPhysics Related
🎯 What it does: DATS is proposed, a difficulty-aware task sampler that prioritizes training higher difficulty PDE tasks during meta-learning of PINN and dynamically allocates residual points.
Davidsonian Scene Graph: Improving Reliability in Fine-grained Evaluation for Text-to-Image Generation
Jaemin Cho (University of North Carolina), Su Wang (Google Research)
GenerationTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: A text-to-image evaluation framework based on Davidsonian semantic graphs (DSG) is proposed, which examines the fit between generated images and text by generating atomic, unique, and clearly dependent questions.
DDMI: Domain-agnostic Latent Diffusion Models for Synthesizing High-Quality Implicit Neural Representations
Dogyun Park (Korea University), Hyunwoo J. Kim (Korea University)
GenerationData SynthesisDiffusion modelAuto EncoderImageVideo
🎯 What it does: A domain-agnostic latent diffusion model DDMI is designed for synthesizing high-quality implicit neural representations (INR), capable of generating continuous functions across various signal domains such as images, 3D shapes, videos, and NeRF.
De novo Protein Design Using Geometric Vector Field Networks
Weian Mao (Zhejiang University), Chunhua Shen (Zhejiang University)
Protein Structure PredictionDiffusion modelGraph
🎯 What it does: This paper proposes the Vector Field Network (VFN) to achieve stronger geometric feature extraction under protein frame encoding without atomic information, thereby improving the quality and diversity of de novo protein design.
Debiased Collaborative Filtering with Kernel-Based Causal Balancing
Haoxuan Li (Peking University), Peng Cui (Tsinghua University)
Recommendation SystemTabular
🎯 What it does: A kernel-based causal balancing method is proposed, which estimates the propensity score by learning a kernel function that can approximate all balancing functions in collaborative filtering, thereby eliminating selection bias in the observed data.
Debiasing Algorithm through Model Adaptation
Tomasz Limisiewicz (Charles University), Tomáš Musil (Charles University)
TransformerLarge Language ModelText
🎯 What it does: Proposes and implements an algorithm for debiasing through model adaptation (DAMA), which modifies the LLaMA language model without architectural changes to reduce gender bias.
Debiasing Attention Mechanism in Transformer without Demographics
Shenyu Lu (Purdue University), Xiaoqian Wang (Purdue University)
TransformerContrastive LearningImageText
🎯 What it does: A Transformer debiasing method that does not rely on sensitive labels is proposed. After normalizing the Query and Key in the last layer of the Encoder and taking the absolute value, the attention weights are recalculated, and supervised contrastive learning is applied to locally align the Values corresponding to high attention weights, thereby reducing the model's dependence on sensitive attributes and enhancing fairness.
Decentralized Riemannian Conjugate Gradient Method on the Stiefel Manifold
Jun Chen (Zhejiang University), Yong Liu (Zhejiang University)
OptimizationImageTabular
🎯 What it does: A decentralized Riemannian conjugate gradient descent (DRCGD) algorithm is proposed to solve distributed optimization problems on the Stiefel manifold.
Deceptive Fairness Attacks on Graphs via Meta Learning
Jian Kang (University of Rochester), Hanghang Tong (University of Illinois Urbana-Champaign)
Adversarial AttackMeta LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a deceptive fairness attack framework for graph learning (FATE), which maximizes the deviation of specified fairness metrics (such as statistical parity or individual fairness) through controllable modifications to the graph structure or features, while maintaining the performance of downstream tasks (node classification).
Decision ConvFormer: Local Filtering in MetaFormer is Sufficient for Decision Making
Jeonghye Kim (KAIST), Youngchul Sung (KAIST)
Convolutional Neural NetworkTransformerReinforcement LearningTabularTime Series
🎯 What it does: A Decision ConvFormer (DC) based on the MetaFormer framework is proposed, replacing the original attention module with a 1D convolution token mixer to predict actions in offline reinforcement learning.
Decodable and Sample Invariant Continuous Object Encoder
Dehao Yuan (University of Maryland), Yiannis Aloimonos (University of Maryland)
Point Cloud
🎯 What it does: A training-free high-dimensional function encoding method HDFE is proposed, which maps continuous objects to fixed-dimensional vectors, supporting invariant sample distribution/density, decodability, and isometry.
Decoding Natural Images from EEG for Object Recognition
Yonghao Song (Tsinghua University), Xiaorong Gao (Chinese Academy of Sciences)
ClassificationRecognitionContrastive LearningMultimodalityTime Series
🎯 What it does: A self-supervised EEG-image contrastive learning framework (NICE) has been constructed to decode natural images from EEG signals and achieve zero-shot object recognition.
DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization
Xiangxin Zhou (University of Chinese Academy of Sciences), Quanquan Gu (ByteDance Research)
OptimizationDrug DiscoveryGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: A controllable and decomposable diffusion model (DECOMPOPT) has been developed for structure-based molecular optimization, capable of optimizing multi-objective properties such as binding affinity and synthetic feasibility while maintaining molecular syntax through fine-grained control of molecular substructures.
Decomposed Diffusion Sampler for Accelerating Large-Scale Inverse Problems
Hyungjin Chung (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
RestorationOptimizationComputational EfficiencyDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A sampling strategy called Decomposed Diffusion Sampling (DDS) is proposed, which combines the DDIM sampling of diffusion models with the Krylov subspace method (conjugate gradient CG) to iteratively solve inverse problems quickly and with high quality by using CG in the tangent space of the denoised points.
Decongestion by Representation: Learning to Improve Economic Welfare in Marketplaces
Omer Nahum (Technion), Nir Rosenfeld (Technion)
Recommendation SystemOptimizationRepresentation LearningReinforcement LearningTabular
🎯 What it does: This study investigates how to alleviate congestion in online markets by controlling a subset of product display features on the platform, thereby enhancing social welfare.
Decoupled Marked Temporal Point Process using Neural Ordinary Differential Equations
Yujee Song (Pohang University of Science and Technology), Won Hwa Kim (Pohang University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyTime SeriesSequentialBiomedical DataOrdinary Differential Equation
🎯 What it does: A Dec-ODE framework is proposed to decouple the impact of each event during marked time points and continuously model it using Neural ODE, balancing interpretability and efficient inference.
Decoupling regularization from the action space
Sobhan Mohammadpour, Pierre-Luc Bacon
Drug DiscoveryReinforcement LearningAgentic AITabular
🎯 What it does: This paper proposes the idea of decoupling regularization from the action space, designing a 'decoupled regularizer' that makes the regularization strength independent of the number of actions, and provides both static and dynamic temperature selection schemes; it achieves decoupling of standard regularizers (such as entropy) through algorithm improvements.
Decoupling Weighing and Selecting for Integrating Multiple Graph Pre-training Tasks
Tianyu Fan (Westlake University), Stan Z. Li (Westlake University)
Knowledge DistillationRepresentation LearningDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes a framework called WAS for instance-level fusion of task selection and weighting in graph pre-training.
Deep Confident Steps to New Pockets: Strategies for Docking Generalization
Gabriele Corso (Massachusetts Institute of Technology), Tommi S. Jaakkola
Drug DiscoverySupervised Fine-TuningDiffusion modelBiomedical DataBenchmark
🎯 What it does: A new DOCKGEN blind ligand docking benchmark is proposed, analyzing and enhancing the generalization performance of docking methods based on diffusion models;
Deep Generative Clustering with Multimodal Diffusion Variational Autoencoders
Emanuele Palumbo (ETH Zurich), Julia E Vogt
GenerationData SynthesisDiffusion modelAuto EncoderImageMultimodality
🎯 What it does: The CMVAE model is proposed, which combines multimodal VAE with clustering to achieve clustering structures in a shared latent space.
Deep Geodesic Canonical Correlation Analysis for Covariance-Based Neuroimaging Data
Ce Ju (Nanyang Technological University), Motoaki Kawanabe (Advanced Telecommunications Research Institute International)
Representation LearningMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A self-supervised deep learning framework called DeepGeoCCA based on SPD manifolds is proposed for learning shared low-dimensional representations of multimodal brain imaging data.
DEEP NEURAL NETWORK INITIALIZATION WITH SPARSITY INDUCING ACTIVATIONS
Ilan Price (Mathematical Institute University of Oxford), Jared Tanner (Mathematical Institute University of Oxford)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: The study investigates the use of sparse activation functions in deep networks and proposes a modified, prunable version to address training instability.
Deep Neural Networks Tend To Extrapolate Predictably
Katie Kang (University of California Berkeley), Sergey Levine (University of California Berkeley)
ClassificationDomain AdaptationConvolutional Neural NetworkContrastive LearningImageText
🎯 What it does: This study investigates the predictive behavior of neural networks on out-of-distribution (OOD) inputs under high-dimensional input conditions, finding that they often tend towards the 'Optimal Constant Solution (OCS)', which is a constant prediction that minimizes training loss while ignoring the input.
Deep Orthogonal Hypersphere Compression for Anomaly Detection
Yunhe Zhang (Chinese University of Hong Kong), Jicong Fan (Chinese University of Hong Kong)
Anomaly DetectionGraph Neural NetworkAuto EncoderImageGraphTabular
🎯 What it does: Two anomaly detection models based on deep orthogonal hypersphere compression (DOHSC and DO2HSC) are proposed, which achieve decision boundaries that better conform to the hypersphere assumption through an orthogonal projection layer, and introduce dual hypersphere compression to alleviate high-dimensional bubble effects and sparsity issues.
Deep Reinforcement Learning for Modelling Protein Complexes
Ziqi Gao (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
Protein Structure PredictionGraph Neural NetworkReinforcement LearningGenerative Adversarial NetworkGraphBiomedical Data
🎯 What it does: This study proposes a deep reinforcement learning framework based on Generative Adversarial Policy Networks, capable of automatically predicting the assembly pathways of multi-chain protein complexes and generating three-dimensional structures.
Deep Reinforcement Learning Guided Improvement Heuristic for Job Shop Scheduling
Cong Zhang (Nanyang Technological University), Jie Zhang (Nanyang Technological University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: The research proposes an improved heuristic method based on deep reinforcement learning, using graph neural networks to represent complete job shop scheduling solutions and directly selecting neighborhood operations through learned policies to enhance solution quality.
Deep SE(3)-Equivariant Geometric Reasoning for Precise Placement Tasks
Ben Eisner (Carnegie Mellon University), David Held (Carnegie Mellon University)
Pose EstimationRobotic IntelligenceGraph Neural NetworkReinforcement LearningPoint Cloud
🎯 What it does: Learning SE(3) equivariant geometric reasoning for relative placement tasks of robots under high-dimensional perception, predicting the precise relative poses between objects end-to-end.
Deep Temporal Graph Clustering
Meng Liu (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
Graph Neural NetworkGraphTime Series
🎯 What it does: A general deep temporal graph clustering framework TGC is proposed, which combines deep clustering techniques with temporal graph learning using an interactive sequence batch processing mode, compatible with existing temporal graph models (HTNE, TGN, TREND, etc.)
DeepSPF: Spherical SO(3)-Equivariant Patches for Scan-to-CAD Estimation
Driton Salihu (Technical University of Munich), Eckehard Steinbach (Technical University of Munich)
RecognitionRetrievalConvolutional Neural NetworkGaussian SplattingPoint Cloud
🎯 What it does: A DeepSPF encoder is proposed, constructed based on SO(3) rotation equivariant spherical patch fields (SPF) and Patch Gaussian layers (PG-Layer), for the Scan-to-CAD task of point clouds.
DeepZero: Scaling Up Zeroth-Order Optimization for Deep Model Training
Aochuan Chen (Michigan State University), Sijia Liu (Michigan State University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the DeepZero framework, aimed at achieving zero-order (gradient-free) optimization training of deep networks from scratch.
Defining and extracting generalizable interaction primitives from DNNs
Lu Chen (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelImageText
🎯 What it does: A method is proposed to extract generalizable interaction primitives between multiple models to explain the reasoning logic of DNNs.
Defining Expertise: Applications to Treatment Effect Estimation
Alihan Hüyük, Mihaela van der Schaar (University of Cambridge)
Tabular
🎯 What it does: This paper studies the role of decision-makers' expert knowledge in causal inference, proposing two types of experts: predictive and prognostic, and treating them as prior biases in treatment effect estimation.
Delphic Offline Reinforcement Learning under Nonidentifiable Hidden Confounding
Alizée Pace (ETH Zurich), Guy Tennenholtz (Google Research)
Reinforcement LearningBiomedical DataElectronic Health Records
🎯 What it does: The study focuses on offline reinforcement learning in the presence of unidentifiable hidden confounders, proposing a delphic uncertainty measure for confounding bias and constructing a penalized offline RL algorithm.
Delta-AI: Local objectives for amortized inference in sparse graphical models
Jean-Pierre René Falet (Mila - Quebec AI Institute), Yoshua Bengio (Mila - Quebec AI Institute)
OptimizationGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A local constraint-based Δ-AI algorithm is proposed for cost-effective approximate inference in sparse graphical models.
Democratizing Fine-grained Visual Recognition with Large Language Models
Mingxuan Liu (Hefei University of Technology), Elisa Ricci (Fondazione Bruno Kessler)
ClassificationRecognitionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a fine-grained visual recognition method called FineR, which extracts image attributes using a visual question answering model, infers fine-grained category names with a large language model (LLM), and then performs zero-shot classification using a vision-language model (VLM).
Demonstration-Regularized RL
Daniil Tiapkin (Ecole Polytechnique), Pierre Menard (ENS Lyon)
Reinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes a demonstration regularization method for reinforcement learning, quantifying the impact of expert demonstrations on the sample complexity of reinforcement learning, particularly in the context of optimal policy identification in finite and linear Markov decision processes and human feedback reinforcement learning scenarios.
Demystifying CLIP Data
Hu Xu (FAIR, Meta AI), Christoph Feichtenhofer (FAIR, Meta AI)
ClassificationRetrievalData-Centric LearningTransformerContrastive LearningImageText
🎯 What it does: By reconstructing the metadata of CLIP and combining substring matching with balanced sampling, the MetaCLIP algorithm is proposed, which trims the massive pool of image-text pairs from the web into a high-quality, distribution-balanced dataset, and validates its superiority on multi-scale ViT models.
Demystifying Embedding Spaces using Large Language Models
Guy Tennenholtz (Google Research), Craig Boutilier (Google Research)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Utilizing large language models to explain and explore domain embedding spaces, we construct the Embedding Language Model (ELM), allowing embedding vectors to be directly queried and described in natural language.
Demystifying Linear MDPs and Novel Dynamics Aggregation Framework
Joongkyu Lee (Seoul National University), Min-hwan Oh (Seoul National University)
Reinforcement Learning
🎯 What it does: The paper proposes a new dynamic aggregation framework, proving that there is a lower bound between the feature dimension of linear MDPs and the size of the state space. Based on this framework, the UC-HRL algorithm is designed, achieving hierarchical reinforcement learning with asymptotic performance guarantees.
Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition
Faisal Hamman (University of Maryland), Sanghamitra Dutta (University of Maryland)
OptimizationFederated LearningTabular
🎯 What it does: This paper studies the trade-off between local and global fairness in federated learning, utilizing the PID method from information theory to decompose the sources of unfairness, and proposes the AGLFOP optimization framework to evaluate the theoretical limits of accuracy and fairness.
Demystifying Poisoning Backdoor Attacks from a Statistical Perspective
Ganghua Wang (University of Minnesota), Jie Ding (University of Minnesota)
Adversarial AttackConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: This paper conducts a theoretical analysis of data poisoning backdoor attacks from a statistical perspective, deriving upper and lower bounds for finite sample and asymptotic risks, and identifying the determining factors for successful attacks (backdoor ratio, trigger amplitude, direction, and data distribution).
DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING
Shitong Duan (Fudan University), Ning Gu (Microsoft Research Asia)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a dynamic generative evaluation framework called DeNEVIL, which is used to automatically detect the ethical value tendencies of large language models (LLMs). Based on this, a MoralPrompt dataset containing over 500 value principles was constructed, and a benchmark test for value bias was conducted on 27 types of LLMs.
Denoising Diffusion Bridge Models
Linqi Zhou (Stanford University), Stefano Ermon (Stanford University)
Image TranslationGenerationDiffusion modelScore-based ModelRectified FlowImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes Denoising Diffusion Bridge Models (DDBMs), a universal framework for generation and image translation between any two distributions.
Denoising Diffusion Step-aware Models
Shuai Yang (Hong Kong University of Science and Technology), Ying-Cong Chen
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: Proposes DDSM (Denoising Diffusion Step-aware Models), which uses networks of different sizes at different steps of the diffusion process to reduce redundant computations;
Denoising Diffusion via Image-Based Rendering
Titas Anciukevičius (University of Edinburgh), Paul Henderson (University of Glasgow)
RestorationGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes GIBR, an unsupervised 3D scene generation and reconstruction method implemented using diffusion models and image-based rendering.
Denoising Task Routing for Diffusion Models
Byeongjun Park (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
RestorationGenerationDiffusion modelImage
🎯 What it does: A method called Denoising Task Routing (DTR) is proposed, which establishes dedicated information channels for different denoising tasks (corresponding to different time steps) in diffusion models through channel masking, explicitly addressing the negative transfer problem in multi-task learning.
DePT: Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning
Zhengxiang Shi (University College London), Aldo Lipani (University College London)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningPrompt EngineeringTextMultimodality
🎯 What it does: The research proposes Decomposed Prompt Tuning (DEPT) for parameter-efficient fine-tuning, addressing the issue of overly long input sequences caused by soft prompts;
Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit
Blake Bordelon (Harvard University), Cengiz Pehlevan (Harvard University)
OptimizationHyperparameter SearchConvolutional Neural NetworkTransformerImagePhysics Related
🎯 What it does: This paper proposes a residual network that combines μP parameterization with a 1/√L residual branch scale, demonstrating and experimentally validating the transferability of hyperparameters (learning rate, momentum, regularization, etc.) in terms of width and depth.
Designing Skill-Compatible AI: Methodologies and Frameworks in Chess
Karim Hamade (University of Toronto), Ashton Anderson (University of Toronto)
Reinforcement LearningTabular
🎯 What it does: Three methods (Tree, Expector, Attuned) were proposed and evaluated to build AI chess players compatible with low-skilled partners, and two cooperative game frameworks (STT and HB) were designed to verify their effectiveness.
Det-CGD: Compressed Gradient Descent with Matrix Stepsizes for Non-Convex Optimization
Hanmin Li (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated LearningImage
🎯 What it does: This paper proposes two new Compressed Gradient Descent (CGD) algorithms that use matrix step sizes to optimize non-convex objectives with matrix smoothness, providing convergence proofs in both single-machine and distributed settings. By designing a hierarchical structure for the compressor and step size, lossless compression is achieved, and faster convergence rates are validated through experiments.
Detecting Machine-Generated Texts by Multi-Population Aware Optimization for Maximum Mean Discrepancy
Shuhai Zhang (South China University of Technology), Mingkui Tan (South China University of Technology)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A multi-population perception optimization method for maximum mean discrepancy (MMD) called MMD-MP is proposed and implemented, and paragraph-level and sentence-level machine-generated text detectors are constructed based on this method.
Detecting Pretraining Data from Large Language Models
Weijia Shi (University of Washington), Luke Zettlemoyer (University of Washington)
Large Language ModelTextBenchmark
🎯 What it does: This study investigates the detection problem of pre-training data for large language models, specifically determining whether a model has been pre-trained on a given text with black-box access.
Detecting, Explaining, and Mitigating Memorization in Diffusion Models
Yuxin Wen (University of Maryland), Lingjuan Lyu (Sony AI)
GenerationExplainability and InterpretabilityDiffusion modelText
🎯 What it does: A detection method based on text-conditioned noise prediction amplitude is proposed for the rapid identification of memorized outputs in diffusion models, further providing word-level importance assessment and mitigation strategies during inference/training.
DFormer: Rethinking RGBD Representation Learning for Semantic Segmentation
Bowen Yin (Nankai University), Qibin Hou (Nankai University)
Object DetectionSegmentationRepresentation LearningTransformerImageMultimodality
🎯 What it does: A pre-training framework for RGB-D, called DFormer, is proposed, which can learn transferable RGB-D representations and be used directly in downstream tasks;
Diagnosing Transformers: Illuminating Feature Spaces for Clinical Decision-Making
Aliyah R. Hsu (University of California Berkeley), Bin Yu (University of California Berkeley)
Anomaly DetectionExplainability and InterpretabilityTransformerSupervised Fine-TuningTextBiomedical Data
🎯 What it does: A framework called SUFO (Supervised Probing + Unsupervised similarity + Feature Dynamics + Outlier analysis) is proposed to systematically explain the feature space and its evolution process of fine-tuned transformers in clinical tasks.
DIAGNOSIS: Detecting Unauthorized Data Usages in Text-to-image Diffusion Models
Zhenting Wang (Rutgers University), Shiqing Ma (University of Massachusetts Amherst)
GenerationData SynthesisAnomaly DetectionSupervised Fine-TuningDiffusion modelImage
🎯 What it does: A mechanism is proposed to implant invisible memory in text-to-image diffusion models to detect unauthorized use of specific data during training or fine-tuning.
Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking
Kaifeng Lyu (Princeton University), Wei Hu (University of Michigan)
Tabular
🎯 What it does: This paper theoretically proves that under the conditions of large initialization and small weight decay, the training of neural networks experiences an implicit bias separation from kernel mode to rich model in two stages, with a sharp performance transition (grokking) occurring at 1/λ log α.
Dictionary Contrastive Learning for Efficient Local Supervision without Auxiliary Networks
Suhwan Choi (Seoul National University), Myungjoo Kang (Seoul National University)
ClassificationComputational EfficiencyRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes Dictionary Contrastive Learning (DCL), a contrastive learning framework that achieves efficient local supervision without auxiliary networks.
DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation
Roi Benita (Technion Israel Institute of Technology), Joseph Keshet (Technion Israel Institute of Technology)
GenerationData SynthesisDiffusion modelAudio
🎯 What it does: An end-to-end autoregressive diffusion model called DiffAR is proposed for directly generating raw speech waveforms.