ICML 2023 Papers — Page 4
International Conference on Machine Learning · 1828 papers
Continual Learners are Incremental Model Generalizers
Jaehong Yoon (Korea Advanced Institute of Science and Technology), Yue Cao (Beijing Academy of Artificial Intelligence)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper explores the potential of continual learning models as pre-trainers, proposing an unsupervised continual learning framework based on Masked Image Modeling (MIM) and designing the GLAD module to maintain global attention diversity during fine-tuning, thereby enhancing the model's generalization ability.
Continual Learning in Linear Classification on Separable Data
Itay Evron (Technion), Daniel Soudry (Toyota Technological Institute at Chicago)
ClassificationOptimizationTabular
🎯 What it does: This paper analyzes continual learning in separable linear classification tasks from a theoretical perspective, proving that weak regularization is equivalent to continuous maximum margin projection, and subsequently provides upper bounds for forgetting, distance, and convergence.
Continual Task Allocation in Meta-Policy Network via Sparse Prompting
Yijun Yang (Southern University of Science and Technology), Yuhui Shi (Southern University of Science and Technology)
Meta LearningReinforcement LearningPrompt EngineeringSequential
🎯 What it does: This paper proposes the CoTASP method, which dynamically extracts sub-networks from a meta-policy network using sparse prompts to achieve continual task learning, addressing the plasticity-stability trade-off problem.
Continual Vision-Language Representation Learning with Off-Diagonal Information
Zixuan Ni (Zhejiang University), Qi Tian (Huawei Cloud)
RetrievalRepresentation LearningTransformerContrastive LearningMultimodality
🎯 What it does: This study investigates the continuous training of CLIP on streaming data and proposes the Mod-X framework to alleviate the spatial confusion problem in continuous training.
Continuation Path Learning for Homotopy Optimization
Xi Lin (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
OptimizationTabular
🎯 What it does: A model-based continuous path learning method (CPL) is proposed, which directly learns the entire homotopy optimization path and can generate solutions at any interpolation level.
Continuous Spatiotemporal Transformer
Antonio Henrique de Oliveira Fonseca (Yale University), David van Dijk (Yale University)
RestorationAutonomous DrivingOptimizationTransformerVideoTime SeriesBiomedical DataOrdinary Differential Equation
🎯 What it does: A Transformer architecture named Continuous Spatiotemporal Transformer (CST) is proposed, specifically designed for learning continuous spatiotemporal dynamical systems, ensuring the model outputs are continuous and smooth, achieving high-quality interpolation and prediction across various tasks.
Continuously Parameterized Mixture Models
Christopher M Bender, Junier Oliva
Anomaly DetectionImageOrdinary Differential Equation
🎯 What it does: A continuous parameterized mixed model (CPMM) utilizing neural ODEs is proposed, achieving density estimation and clustering of high-dimensional complex data through a hierarchical structure and curriculum learning.
ContraBAR: Contrastive Bayes-Adaptive Deep RL
Era Choshen (Technion), Aviv Tamar (Technion)
Recurrent Neural NetworkReinforcement LearningContrastive LearningImage
🎯 What it does: This paper proposes a Bayesian Adaptive Deep Reinforcement Learning framework based on Contrastive Learning (ContraBAR), which learns the hidden state of the task (belief) through Contrastive Predictive Coding (CPC) and uses it as the information state input to the policy.
Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretraining
Zekun Qi (Xi'an Jiaotong University), Li Yi (Tsinghua University)
Representation LearningTransformerContrastive LearningPoint Cloud
🎯 What it does: A 3D representation learning framework called RECON is proposed, which integrates contrastive learning and generative reconstruction.
Contrastive Energy Prediction for Exact Energy-Guided Diffusion Sampling in Offline Reinforcement Learning
Cheng Lu (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationData SynthesisReinforcement LearningDiffusion modelContrastive LearningImageTabular
🎯 What it does: A method for energy-guided diffusion sampling is proposed, which can accurately sample from the distribution p(x) ∝ q(x)e^{‑E(x)} on a pre-trained diffusion model;
Contrastive Learning Meets Homophily: Two Birds with One Stone
Dongxiao He (Tianjin University), Weixiong Zhang (Hong Kong Polytechnic University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: The NeCo method is proposed, which eliminates the need for data augmentation in graph contrastive learning by expanding positive samples to a neighbor set, while also addressing the issue of same-class edge discrimination in GNNs.
Controllability-Aware Unsupervised Skill Discovery
Seohong Park (University of California), Pieter Abbeel (University of California)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes an unsupervised skill discovery method called Controllability-Aware Skill Discovery (CSD), which actively seeks more complex skills by learning a distance function that measures control difficulty.
Controllable Neural Symbolic Regression
Tommaso Bendinelli (CSEM SA), Pierre-Alexandre Kamienny (Sorbonne Université)
TransformerTabular
🎯 What it does: A model NSRwH is proposed that can constrain the symbolic regression output through additional conditions during the inference phase, incorporating prior information into the Transformer.
Controlled Differential Equations on Long Sequences via Non-standard Wavelets
Sourav Pal (University of Wisconsin Madison), Vikas Singh (University of Wisconsin Madison)
ClassificationOptimizationComputational EfficiencyRecurrent Neural NetworkAuto EncoderTime SeriesSequentialBiomedical DataElectrocardiogramOrdinary Differential Equation
🎯 What it does: A non-standard wavelet decomposition (BCR-DE) framework based on the Beylkin–Coifman–Rokhlin (BCR) algorithm is proposed, simplifying long-time series neural controlled differential equations (NCDE) into a single integral transform and constraining the operator to Calderon-Zygmund (CZ) class, achieving efficient training and inference.
Controlled Text Generation with Natural Language Instructions
Wangchunshu Zhou (ETH Zurich), Mrinmaya Sachan (ETH Zurich)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a framework for controlled text generation through natural language instructions, called INSTRUCTCTG, which utilizes instruction tuning rather than modifying the decoding process.
Controlling Posterior Collapse by an Inverse Lipschitz Constraint on the Decoder Network
Yuri Kinoshita (University of Tokyo), Shin-ichi Maeda (Preferred Networks)
GenerationData SynthesisAuto EncoderImageText
🎯 What it does: This paper proposes the incorporation of an inverse Lipschitz neural network into the VAE decoder to theoretically control the degree of posterior collapse, and provides corresponding implementation methods.
Controlling Type Confounding in Ad Hoc Teamwork with Instance-wise Teammate Feedback Rectification
Dong Xing (Zhejiang University), Gang Pan (Zhejiang University)
Robotic IntelligenceReinforcement LearningAgentic AITabularSequentialBiomedical Data
🎯 What it does: This paper first theoretically analyzes the phenomenon of 'type confusion' that occurs in ad hoc team cooperation and proposes a method called Instance-level Teammate Feedback Correction (CTCAT) to eliminate this confusion, thereby enhancing the agents' cooperative performance in front of unknown teammates.
Convergence of First-Order Methods for Constrained Nonconvex Optimization with Dependent Data
Ahmet Alacaoglu (University of Wisconsin Madison), Hanbaek Lyu (University of Wisconsin Madison)
OptimizationGraph
🎯 What it does: This study investigates the convergence and complexity of Projection Stochastic Gradient (PSGD), AdaGrad, and momentum-based stochastic gradient methods under constrained non-convex smooth optimization problems in the presence of data correlation.
Convergence of Proximal Point and Extragradient-Based Methods Beyond Monotonicity: the Case of Negative Comonotonicity
Eduard Gorbunov (Mohamed bin Zayed University of Artificial Intelligence), Gauthier Gidel (Universite de Montreal)
Optimization
🎯 What it does: Under the assumption of negative comonotonicity, the convergence properties of the Proximal Point (PP), Extragradient (EG), and Optimistic Gradient (OG) methods are studied, and the optimal non-asymptotic convergence rates along with corresponding worst-case examples are provided.
Convex Geometry of ReLU-layers, Injectivity on the Ball and Local Reconstruction
Daniel Haider (Austrian Academy of Sciences), Peter Balazs (University of Vienna)
OptimizationConvolutional Neural NetworkTabular
🎯 What it does: This paper studies the injectivity of ReLU layers on a finite radius sphere (and its non-negative part) within the framework of frame theory, and provides a computable upper bound bias estimation method (PBE) along with the corresponding reconstruction formula.
Cooperation in the Latent Space: The Benefits of Adding Mixture Components in Variational Autoencoders
Oskar Kviman (KTH Royal Institute of Technology), Jens Lagergren (KTH Royal Institute of Technology)
GenerationAuto EncoderImageBiomedical Data
🎯 What it does: This paper proposes the use of a Mixture Variational Autoencoder (Mixture VAE) trained through MISELBO, allowing the mixture components to collaboratively cover the target distribution, thereby improving the quality of the posterior approximation of the VAE.
Cooperative Multi-Agent Reinforcement Learning: Asynchronous Communication and Linear Function Approximation
Yifei Min (Yale University), Quanquan Gu (University of California Los Angeles)
Reinforcement Learning
🎯 What it does: An asynchronous communication multi-agent reinforcement learning algorithm (Async-Coop-LSVI-UCB) is proposed, achieving low returns and low communication complexity collaborative learning under the linear MDP framework.
Cooperative Open-ended Learning Framework for Zero-Shot Coordination
Yang Li (University of Manchester), Wei Pan (University of Manchester)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: The COLE framework is proposed, which evaluates and overcomes cooperative incompatibility through Graphical Form Games (GFG) and Preference Graphs (P-GFG), achieving zero-shot coordination;
Coordinate Descent Methods for Fractional Minimization
Ganzhao Yuan (Peng Cheng Laboratory)
OptimizationText
🎯 What it does: This paper proposes two coordinate descent (CD) algorithms (FCD and PCD) specifically designed to solve structured fractional minimization problems, and provides theoretical convergence and optimality analysis.
Coordinated Dynamic Bidding in Repeated Second-Price Auctions with Budgets
Yurong Chen (Peking University), Xiaotie Deng (Peking University)
OptimizationTabular
🎯 What it does: This paper studies the design of a collaborative online bidding algorithm in repeated binary auctions under budget constraints, ensuring that the utility of coalition members is at least not lower than the optimal level of their independent bidding.
Correcting discount-factor mismatch in on-policy policy gradient methods
Fengdi Che (University of Alberta), A. Rupam Mahmood (University of Alberta)
Reinforcement LearningSequential
🎯 What it does: This paper addresses the issue of discount factor mismatch caused by using undiscounted state distributions in on-policy policy gradient methods, proposing a distribution correction based on the average discount factor.
Corruption-Robust Algorithms with Uncertainty Weighting for Nonlinear Contextual Bandits and Markov Decision Processes
Chenlu Ye (Hong Kong University of Science and Technology), Tong Zhang
Reinforcement Learning
🎯 What it does: This paper proposes robust algorithms, namely CR-Eluder-UCB and CR-LSVI-UCB, that utilize uncertainty-weighted regression to address nonlinear contextual bandit and Markov decision process (MDP) problems in the presence of adversarial contamination.
Counterfactual Analysis in Dynamic Latent State Models
Martin B Haugh, Raghav Singal (Dartmouth College)
OptimizationBiomedical Data
🎯 What it does: In dynamic latent state models (such as hidden Markov models), an optimization-based framework is proposed to compute the upper and lower bounds of counterfactual queries (such as the necessity of death) given observed data.
Counterfactual Identifiability of Bijective Causal Models
Arash Nasr-Esfahany (Massachusetts Institute of Technology), Devavrat Shah (Massachusetts Institute of Technology)
GenerationData SynthesisFlow-based ModelVideo
🎯 What it does: This paper proposes and studies a unified class of models called Bijective Generation Mechanism (BGMs), proving that counterfactual identification can be achieved under three common structures: Markovian, Instrumental Variable (IV), and Backdoor Criterion (BC). It also presents a learnable method based on structured generative models such as Conditional Normalizing Flow.
Coupled Variational Autoencoder
Xiaoran Hao (Rutgers University), Patrick Shafto (Institute for Advanced Study)
GenerationData SynthesisAuto EncoderImage
🎯 What it does: This paper proposes the Coupled Variational Autoencoder (C-VAE), which transforms the traditional VAE problem into an optimal transport (OT) problem with entropy regularization. It jointly learns the encoder and decoder through the dual/semi-dual form of OT, achieving flexible matching to any prior distribution and naturally addressing the Prior Hole problem.
Covariate balancing using the integral probability metric for causal inference
Insung Kong (Seoul National University), Yongdai Kim (Seoul National University)
Adversarial AttackTabular
🎯 What it does: Two weighting methods (parametric and non-parametric CBIPM) were designed using Integral Probability Metrics (IPM) to achieve covariate balance and estimate average treatment effects.
Crafting Training Degradation Distribution for the Accuracy-Generalization Trade-off in Real-World Super-Resolution
Ruofan Zhang (Tsinghua University), Wenming Yang (Tsinghua University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A method is designed to customize the training degradation distribution using a small number of reference images, enhancing the accuracy and generalization ability of super-resolution in real scenes.
Cramming: Training a Language Model on a single GPU in one day.
Jonas Geiping (University of Maryland), Tom Goldstein (University of Maryland)
TransformerLarge Language ModelText
🎯 What it does: Train a Transformer language model from scratch in just one day on a single consumer-grade GPU, exploring how to achieve performance close to BERT on downstream GLUE tasks with limited computational resources.
CRISP: Curriculum based Sequential neural decoders for Polar code family
S Ashwin Hebbar, Pramod Viswanath (Princeton University)
Convolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningSequential
🎯 What it does: A learning-based polar code and PAC code decoder called CRISP is proposed, achieving near-MAP reliability through curriculum learning training.
Critical Points and Convergence Analysis of Generative Deep Linear Networks Trained with Bures-Wasserstein Loss
Pierre Bréchet (Max Planck Institute for Mathematics in the Sciences), Guido Montufar
GenerationOptimizationTabular
🎯 What it does: Analyzed and proved the critical point structure and convergence of deep linear networks when trained with the Bures-Wasserstein loss.
Cross-Entropy Loss Functions: Theoretical Analysis and Applications
Anqi Mao (New York University), Yutao Zhong (New York University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper conducts a theoretical analysis of cross-entropy (logistic) and its generalized form—the comp-sum loss function. It provides H-consistency bounds for any complete and symmetric hypothesis set and introduces a smooth adversarial comp-sum loss to achieve adversarially robust training.
Cross-Modal Fine-Tuning: Align then Refine
Junhong Shen (Carnegie Mellon University), Ameet Talwalkar (Hewlett Packard Enterprise)
Domain AdaptationOptimizationNeural Architecture SearchTransformerSupervised Fine-TuningMultimodalityTabularTime SeriesBenchmark
🎯 What it does: A cross-modal fine-tuning framework named ORCA is proposed, which first aligns the distribution of target data and then fine-tunes the pre-trained model to achieve efficient transfer for various different modal tasks.
CrossSplit: Mitigating Label Noise Memorization through Data Splitting
Jihye Kim (Samsung Advanced Institute of Technology), Simon Lacoste-Julien (Mila)
ClassificationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A robust training framework named CrossSplit is proposed, which randomly divides the training set with noisy labels into two parts, trains two networks on the two parts of the data respectively, and uses each other's predictions to soften the labels, combined with semi-supervised learning to reduce the model's memory of incorrect labels.
CSP: Self-Supervised Contrastive Spatial Pre-Training for Geospatial-Visual Representations
Gengchen Mai (University of Georgia), Stefano Ermon (Stanford University)
ClassificationRecognitionRepresentation LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: A geospatial pre-training framework CSP based on self-supervised contrastive learning is proposed, which encodes images and geographic coordinates through dual encoders and performs contrastive learning during the pre-training phase to learn the correspondence between location and visual features.
Curiosity in Hindsight: Intrinsic Exploration in Stochastic Environments
Daniel Jarrett (DeepMind), Michal Valko (DeepMind)
Reinforcement LearningContrastive LearningVideo
🎯 What it does: This study investigates the exploration problem in sparse or no-reward environments and proposes an intrinsic exploration method based on structural causal models, called 'Hindsight Curiosity'.
Curious Replay for Model-based Adaptation
Isaac Kauvar (Stanford University), Nick Haber (Stanford University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes a priority mechanism for experience replay called Curious Replay, which utilizes curiosity signals to prioritize training experiences, thereby enhancing the model's prediction and adaptability.
Curriculum Co-disentangled Representation Learning across Multiple Environments for Social Recommendation
Xin Wang (Tsinghua University), Wenwu Zhu (Tsinghua University)
Recommendation SystemRepresentation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A cross-consumption environment and social environment co-decoupled representation learning framework, CurCoDis, is proposed, achieving user representation decoupling and interpretability through curriculum subgraph training and prototype routing mechanisms.
Cut your Losses with Squentropy
Like Hui (University of California), Stephen Wright (University of Wisconsin)
ClassificationOptimizationImageTextTabularAudio
🎯 What it does: A new multi-class loss function called 'squentropy' is proposed, which adds cross-entropy to the average squared loss of incorrect classes for training neural networks.
Cyclic Block Coordinate Descent With Variance Reduction for Composite Nonconvex Optimization
Xufeng Cai (University of Wisconsin Madison), Jelena Diakonikolas (University of Wisconsin Madison)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Cyclic Block Coordinate Descent (CBCD) method and provides non-asymptotic convergence guarantees for deterministic and stochastic non-convex optimization problems.
D2Match: Leveraging Deep Learning and Degeneracy for Subgraph Matching
Xuanzhou Liu (Tsinghua University), Haiqin Yang (International Digital Economy Academy)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes the D2Match model for subgraph matching, which first reduces subgraph matching to subtree matching, and then equates it to perfect matching in bipartite graphs, achieving linear-time matching through tree structure aggregation using GNN.
DADAO: Decoupled Accelerated Decentralized Asynchronous Optimization
Adel Nabli (Sorbonne Université), Edouard Oyallon (Sorbonne Université)
OptimizationTabular
🎯 What it does: A DADAO algorithm is proposed for asynchronously, decoupled, and accelerated minimization of the sum of strongly convex smooth functions in distributed networks.
Data Efficient Neural Scaling Law via Model Reusing
Peihao Wang (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
Knowledge DistillationData-Centric LearningTransformerImageText
🎯 What it does: This paper studies the neural scaling law of large Transformers under extremely low data scarcity conditions and recovers the original power law through model reuse techniques.
Data Feedback Loops: Model-driven Amplification of Dataset Biases
Rohan Taori (Stanford University), Tatsunori Hashimoto
GenerationData-Centric LearningImageText
🎯 What it does: Discusses the feedback loop generated when model outputs are used as future training data and studies its impact on bias amplification.
Data Poisoning Attacks Against Multimodal Encoders
Ziqing Yang (CISPA Helmholtz Center for Information Security), Yang Zhang (CISPA Helmholtz Center for Information Security)
RetrievalAdversarial AttackContrastive LearningImageTextMultimodality
🎯 What it does: This study investigates data poisoning attacks on multimodal encoders, exploring the vulnerability of both visual and textual modalities.
Data Representations' Study of Latent Image Manifolds
Ilya Kaufman (Ben Gurion University of the Negev), Omri Azencot (Ben Gurion University of the Negev)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper explores the evolution of the curvature of deep network hidden layer representations with layer depth by measuring the principal curvature of the latent image manifold in CNNs.
Data Structures for Density Estimation
Anders Aamand (Massachusetts Institute of Technology), Sandeep Silwal (Massachusetts Institute of Technology)
🎯 What it does: A data structure is proposed that can identify the known discrete distribution v_i closest to an unknown distribution p under a given sublinear sample (relative to the size of the distribution domain n), and implementation methods are provided in two settings (proper and improper).
Data-Copying in Generative Models: A Formal Framework
Robi Bhattacharjee (University of California San Diego), Kamalika Chaudhuri (University of California San Diego)
GenerationData SynthesisTabular
🎯 What it does: This paper proposes a formal definition and method for detecting training data memorized by generative models.
Data-Driven Subgroup Identification for Linear Regression
Zachary Izzo (Stanford University), James Zou (Stanford University)
OptimizationExplainability and InterpretabilityBiomedical Data
🎯 What it does: This paper proposes a data-driven subgroup identification method called DDGroup, which aims to find interpretable subsets where features and targets have a unified linear relationship in linear regression.
Data-Efficient Contrastive Self-supervised Learning: Most Beneficial Examples for Supervised Learning Contribute the Least
Siddharth Joshi (University of California Los Angeles), Baharan Mirzasoleiman (University of California Los Angeles)
ClassificationData-Centric LearningContrastive LearningImage
🎯 What it does: This paper studies which training samples in contrastive self-supervised learning can most enhance model performance and proposes a subset selection method called SAS based on maximizing expected augmented similarity.
Data-OOB: Out-of-bag Estimate as a Simple and Efficient Data Value
Yongchan Kwon (Columbia University), James Zou (Stanford University)
ClassificationAnomaly DetectionComputational EfficiencyData-Centric LearningTabular
🎯 What it does: A method for estimating data value based on a random sampling bagging model (Data-OOB) is proposed, which can directly calculate the influence of each sample on existing weak learners.
Dataset Distillation with Convexified Implicit Gradients
Noel Loo (Massachusetts Institute of Technology), Daniela Rus (Massachusetts Institute of Technology)
OptimizationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A composite reparameterization and convex data distillation algorithm based on implicit gradients (RCIG) is proposed, which can compress the original dataset into a minimal number of synthetic samples through a single gradient update and train high-quality models on various tasks.
DDGR: Continual Learning with Deep Diffusion-based Generative Replay
Rui Gao (Wuhan University), Weiwei Liu (Wuhan University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A generative replay method based on deep diffusion models (DDGR) is proposed to address catastrophic forgetting in continual learning.
Decentralized SGD and Average-direction SAM are Asymptotically Equivalent
Tongtian Zhu (Zhejiang University), Dacheng Tao (Sydney University)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: Through theoretical analysis and experiments, it is proven that Distributed SGD (D-SGD) is asymptotically equivalent to Average-direction SAM under non-convex and non-β-smooth objectives, revealing the trade-off between regularization and optimization in decentralized learning.
Decentralized Stochastic Bilevel Optimization with Improved per-Iteration Complexity
Xuxing Chen (University of California), Krishna Balasubramanian
OptimizationImageTabular
🎯 What it does: A distributed stochastic two-layer optimization algorithm is proposed, which utilizes only first-order stochastic gradients, Hessian-vector, and Jacobian-vector products, avoiding the computation of the full Hessian and Jacobian;
Decoding Layer Saliency in Language Transformers
Elizabeth Mary Hou, Gregory David Castanon
ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Decode the intermediate layer features of the stacked Transformer model and map them back to the word vector space, proposing a 'Decoded Grad-CAM' method based on Grad-CAM to visualize the words that have the most influence on classification decisions in the text input.
DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design
Jiaqi Guan (University of Illinois Urbana-Champaign), Quanquan Gu (ByteDance Research)
Drug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A 3D structural drug design diffusion model named DecompDiff is proposed, which splits ligands into two parts: branches (arms) and scaffolds, and uses a data-driven decomposition prior to guide molecular generation.
Deep Anomaly Detection under Labeling Budget Constraints
Aodong Li (University of California), Maja Rudolph (Bosch Center for Artificial Intelligence)
Anomaly DetectionContrastive LearningImageVideoTabular
🎯 What it does: Proposes a deep anomaly detection method SOEL under limited labeling budget, combining diversified queries and a semi-supervised appearance exposure learning framework.
Deep Clustering with Incomplete Noisy Pairwise Annotations: A Geometric Regularization Approach
Tri Nguyen (Oregon State University), Xiao Fu (Oregon State University)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A deep constrained clustering method based on geometric regularization is proposed, which learns clustering structures using pairwise similarity annotations in the presence of incomplete data and noise.
Deep Generative Symbolic Regression with Monte-Carlo-Tree-Search
Pierre-Alexandre Kamienny (Meta AI), Marco Virgolin (Centrum Wiskunde en Informatica)
OptimizationExplainability and InterpretabilityTransformerReinforcement LearningTabularPhysics Related
🎯 What it does: This paper proposes an algorithm called DGSR-MCTS that combines deep generative symbolic regression with Monte Carlo Tree Search to quickly generate high-quality, interpretable symbolic expressions on a given dataset.
Deep Graph Representation Learning and Optimization for Influence Maximization
Chen Ling (Emory University), Liang Zhao (Emory University)
OptimizationKnowledge DistillationRepresentation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A framework named DeepIM is proposed, which utilizes autoencoders to learn the latent distribution of seed sets and fits a diffusion model in a continuous latent space through an end-to-end GNN, directly optimizing the seed node set for maximum influence.
Deep Laplacian-based Options for Temporally-Extended Exploration
Martin Klissarov (Mila), Marlos C. Machado (Alberta Machine Intelligence Institute)
Convolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This paper proposes an online deep reinforcement learning algorithm DCEO for learning options based on Laplacian operator features, thereby achieving more effective temporal abstraction exploration.
Deep Latent State Space Models for Time-Series Generation
Linqi Zhou (Stanford University), Stefano Ermon (Stanford University)
GenerationData SynthesisRecurrent Neural NetworkAuto EncoderTime SeriesSequentialBiomedical DataOrdinary Differential Equation
🎯 What it does: A latent generative model LS4 based on a linear state space model is proposed for generating and predicting time series data.
Deep Perturbation Learning: Enhancing the Network Performance via Image Perturbations
Zifan Song (Tongji University), Cai Rong Zhao
ClassificationObject DetectionSegmentationImage
🎯 What it does: This paper proposes a framework called Deep Perturbation Learning (DPL), which improves the training distribution of the network by learning small perturbations on training samples, thereby enhancing model performance.
Deep Regression Unlearning
Ayush Kumar Tarun, Mohan Kankanhalli (National University of Singapore)
ImageTextTabular
🎯 What it does: A machine unlearning method is proposed for deep regression models to achieve the deletion of specified training samples.
Deep Temporal Sets with Evidential Reinforced Attentions for Unique Behavioral Pattern Discovery
Dingrong Wang (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
ClassificationRecognitionAnomaly DetectionRecurrent Neural NetworkReinforcement LearningMultimodalityTime SeriesSequential
🎯 What it does: A model that combines Deep Time Sets (DTS) and Evidence Reinforced Attention (ERA) has been designed and implemented to discover feature behavior patterns from multimodal (gaze and touch) behavior sequences, and to differentiate between children with Autism Spectrum Disorder (ASD) and typically developing (TD) children.
Defects of Convolutional Decoder Networks in Frequency Representation
Ling Tang (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
Convolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper analyzes the defects of convolutional decoder networks in representing different frequency components in the frequency domain, proving that convolution, zero padding, and upsampling operations weaken high-frequency information, repeat strong frequency signals, and make it difficult to learn targets with slight frequency offsets.
Deja Vu: Contextual Sparsity for Efficient LLMs at Inference Time
Zichang Liu (Rice University), Beidi Chen (Carnegie Mellon University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A dynamic sparse structure is proposed for the inference process of large-scale language models (LLMs), which retains only a portion of attention heads and MLP neurons for each input. The DEJAVU system is constructed to achieve real-time sparse prediction and hardware-friendly execution, significantly reducing inference latency.
Delay-Adapted Policy Optimization and Improved Regret for Adversarial MDP with Delayed Bandit Feedback
Tal Lancewicki (Tel Aviv University), Dmitry Sotnikov (Amazon Science)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper proposes the Delay-Adapted Policy Optimization (DAPO) algorithm suitable for adversarial MDPs with delayed bandit feedback, and provides theoretical and experimental results in discrete, linear function approximation, and deep reinforcement learning (PPO) environments.
Delay-agnostic Asynchronous Coordinate Update Algorithm
Xuyang Wu (KTH Royal Institute of Technology), Mikael Johansson (Stockholm University)
OptimizationTabular
🎯 What it does: This paper proposes a new delay-independent asynchronous coordinate update algorithm called DEGAS, aimed at solving operator fixed-point problems, and applies it to asynchronous BCD and ADMM.
Delayed Bandits: When Do Intermediate Observations Help?
Emmanuel Esposito (University of Milan), Yevgeny Seldin (University of Copenhagen)
OptimizationMeta LearningReinforcement Learning from Human FeedbackTabular
🎯 What it does: This study investigates the delayed feedback and intermediate observation model of K-armed multi-armed bandits, analyzing whether intermediate observations can reduce the impact of delays under different state-to-loss mappings, and provides matching upper and lower bounds.
Delayed Feedback in Kernel Bandits
Sattar Vakili (MediaTek Research), Ciara Pike-Burke (Imperial College London)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the kernel bandwidth problem with stochastic delayed feedback and proposes an algorithm called BPE-Delay, aimed at optimizing the learning process of kernel functions.
Delving into Noisy Label Detection with Clean Data
Chenglin Yu (Wuhan University), Weiwei Liu (Wuhan University)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: A Benjamini–Hochberg (BH) multiple hypothesis testing framework (BHN) is proposed, which utilizes a small amount of clean data to identify and eliminate noisy labels in neural networks by treating noisy label detection as a multiple hypothesis testing problem and calibrating empirical p-values.
Demonstration-free Autonomous Reinforcement Learning via Implicit and Bidirectional Curriculum
Jigang Kim (Seoul National University), H. Jin Kim (Seoul National University)
Robotic IntelligenceReinforcement LearningAgentic AITabularBenchmark
🎯 What it does: The IBC algorithm is proposed—a self-directed RL method that requires no demonstrations and no resets, utilizing conditionally activated auxiliary agents and bidirectional target curricula to achieve learning in sparse reward tasks.
Demystifying Disagreement-on-the-Line in High Dimensions
Donghwan Lee (University of Pennsylvania), Hamed Hassani (University of Pennsylvania)
Domain AdaptationGaussian SplattingImage
🎯 What it does: This study investigates the disagreement of random feature models in the high-dimensional limit and its linear relationship in distribution shift (between source and target domains), exploring the effects of regularization and different shared randomness on this relationship.
Demystifying Uneven Vulnerability of Link Stealing Attacks against Graph Neural Networks
He Zhang (Monash University), Xingliang YUAN
Adversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper investigates whether there is an unevenness in the vulnerability of Graph Neural Networks (GNN) to link stealing attacks and proposes a customized threshold attack method based on node label groups, further validating the superiority of this method on various real datasets.
Denoising MCMC for Accelerating Diffusion-Based Generative Models
Beomsu Kim (KAIST), Jong Chul Ye (KAIST)
GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The Denoising MCMC (DMCMC) framework is proposed, which combines MCMC with reverse SDE/ODE solvers to accelerate sampling of score models.
DeSRA: Detect and Delete the Artifacts of GAN-based Real-World Super-Resolution Models
Liangbin Xie (University of Macau), Chao Dong (Shenzhen Institute of Advanced Technology)
RestorationSuper ResolutionGenerative Adversarial NetworkImage
🎯 What it does: The DeSRA method is proposed, which uses MSE-SR results as a reference to first detect artifact areas in GAN-SR inference, and then performs few-shot fine-tuning with pseudo GT to ultimately eliminate artifacts and enhance visual quality.
DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature
Eric Mitchell (Stanford University), Chelsea Finn (Stanford University)
GenerationAnomaly DetectionTransformerLarge Language ModelText
🎯 What it does: A zero-shot machine-generated text detection method based on probabilistic curvature, called DetectGPT, is proposed.
Detecting Adversarial Data by Probing Multiple Perturbations Using Expected Perturbation Score
Shuhai Zhang (South China University of Technology), Mingkui Tan (South China University of Technology)
Anomaly DetectionAdversarial AttackConvolutional Neural NetworkTransformerDiffusion modelImageStochastic Differential Equation
🎯 What it does: Proposes and uses the Expected Perturbation Score (EPS) as a statistic for detecting adversarial samples, and constructs an EPS-based Maximum Mean Discrepancy (MMD) detection method called EPS-AD;
Detecting Adversarial Directions in Deep Reinforcement Learning to Make Robust Decisions
Ezgi Korkmaz, Jonah Brown-Cohen
Adversarial AttackReinforcement LearningVideo
🎯 What it does: A detection method based on deep reinforcement learning strategy loss local second-order approximation (INRD) is proposed to identify adversarial directions in the states seen by the agent, thereby improving decision robustness.
Detecting Out-of-distribution Data through In-distribution Class Prior
Xue Jiang (Southern University of Science and Technology), Bo Han (Hong Kong Baptist University)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: In the case of class-imbalanced training, this study investigates how to improve OOD detection methods during inference by proposing the use of ID class priors instead of the traditional uniform distribution.
Deterministic equivalent and error universality of deep random features learning
Dominik Schröder, Bruno Loureiro (Ecole Normale Supérieure)
Gaussian SplattingTabular
🎯 What it does: This paper studies the problem of learning random Gaussian network functions using fully connected networks, where the intermediate layers are frozen and the output layer is trainable. This is viewed as a natural extension of random feature models to deeper architectures.
DevFormer: A Symmetric Transformer for Context-Aware Device Placement
Haeyeon Kim (Korea Advanced Institute of Science and Technology), Jinkyoo Park (Korea Advanced Institute of Science and Technology)
OptimizationTransformerTabular
🎯 What it does: A Transformer model named DevFormer is proposed for offline context hardware design optimization, particularly for the decap placement problem.
Diagnosis, Feedback, Adaptation: A Human-in-the-Loop Framework for Test-Time Policy Adaptation
Andi Peng (Massachusetts Institute of Technology), Pulkit Agrawal (Massachusetts Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningGenerative Adversarial NetworkImage
🎯 What it does: An interactive framework is proposed that adapts strategies during testing using user feedback;
DIFF2: Differential Private Optimization via Gradient Differences for Nonconvex Distributed Learning
Tomoya Murata (NTT DATA Mathematical Systems Inc.), Taiji Suzuki (RIKEN)
OptimizationFederated LearningSafty and PrivacyTabular
🎯 What it does: This paper proposes a new differential privacy non-convex optimization framework called DIFF2, which constructs a global gradient estimate using gradient differences, significantly reducing the noise magnitude.
Difference of submodular minimization via DC programming
Marwa El Halabi (Samsung), Tim Hoheisel (McGill University)
OptimizationReinforcement LearningTabularAudio
🎯 What it does: This paper transforms the difference submodular (DS) minimization problem into difference convex (DC) optimization, solving it through the difference convex algorithm (DCA) and its complete form (CDCA), and provides theoretical guarantees for convergence and local optimality.
Difference-in-Differences Meets Tree-based Methods: Heterogeneous Treatment Effects Estimation with Unmeasured Confounding
Caizhi Tang (Ant Group), JUN ZHOU
Data SynthesisRecommendation SystemOptimizationTabularTime SeriesFinance Related
🎯 What it does: This paper proposes a framework that combines difference-in-differences and tree models to estimate the conditional average treatment effect (CATT) in the presence of unmeasured confounding in balanced panel data.
Differentiable and Transportable Structure Learning
Jeroen Berrevoets (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Graph Neural NetworkGraph
🎯 What it does: A differentiable and transportable DAG structure learning method called D-Struct is proposed, aimed at addressing the issue of traditional NOTEARS-based methods being unable to reproduce structures across different data distributions.
Differentiable Multi-Target Causal Bayesian Experimental Design
Panagiotis Tigas (University of Oxford), Stefan Bauer (Helmholtz AI)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A gradient-based differentiable Bayesian experimental design method is proposed for batch multi-objective causal model learning.
Differentiable Simulations for Enhanced Sampling of Rare Events
Martin Sipka (Charles University), Rafael Gomez-Bombarelli (Massachusetts Institute of Technology)
OptimizationDrug DiscoveryAuto EncoderGraphPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Using differentiable simulation (DiffSim) without pre-selecting collective variables (CV), it automatically discovers the transition paths of chemical reactions and enhances the sampling efficiency of rare events through end-to-end path integral optimization.
Differentiable Tree Operations Promote Compositional Generalization
Paul Soulos (Johns Hopkins University), Jianfeng Gao (Microsoft Research)
Explainability and InterpretabilityRepresentation LearningTransformerSequential
🎯 What it does: A differentiable tree machine (DTM) is proposed, which compiles discrete tree operations into continuous matrix operations and utilizes external tree storage and neural tree agents to achieve learnable tree structure transformations.
Differential Privacy has Bounded Impact on Fairness in Classification
Paul Mangold (University of Lille), Marc Tommasi (University of Lille)
ClassificationSafty and PrivacyTabular
🎯 What it does: This study investigates the impact of differential privacy on the fairness of classification models and provides a theoretical upper bound on fairness differences.
Differential Privacy, Linguistic Fairness, and Training Data Influence: Impossibility and Possibility Theorems for Multilingual Language Models
Phillip Rust (University of Copenhagen), Anders Søgaard (University of Copenhagen)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study investigates the compatibility and conflicts between multilingual language models in differential privacy, language fairness, and the impact of training data sparsity.
Differentially Private Distributed Bayesian Linear Regression with MCMC
Barış Alparslan (Sabancı University), Ilker Birbil (University of Amsterdam)
Safty and PrivacyComputational EfficiencyTabular
🎯 What it does: A Bayesian linear regression framework suitable for distributed data under differential privacy protection is proposed, along with the corresponding MCMC inference algorithm.
Differentially Private Episodic Reinforcement Learning with Heavy-tailed Rewards
Yulian Wu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
OptimizationSafty and PrivacyReinforcement Learning
🎯 What it does: The study addresses the heavy-tailed reward problem in finite-horizon, tabular Markov Decision Processes (MDP) under the constraints of Joint Differential Privacy (JDP) and Local Differential Privacy (LDP), and proposes a UCB-based value iteration and policy optimization framework (Private-Heavy-UCBVI and Private-Heavy-UCBPO). Privacy protection is achieved through robust mean estimators and tree-based mechanisms/Laplace mechanisms, with corresponding upper and lower bounds provided.