ICML 2023 Papers with AI Summaries
International Conference on Machine Learning · 1828 papers
→ ICML 2023 papers with code (421)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 1828 ICML 2023 papers by keyword, author or institution —
free trial on arXivSub.
Mu$^2$SLAM: Multitask, Multilingual Speech and Language Models
Yong Cheng (Google Research), Ankur Bapna (Google Research)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningSimultaneous Localization and MappingTextMultimodalityAudio
🎯 What it does: A multi-task, multi-language unified encoder-decoder pre-training model Mu SLAM has been constructed, capable of simultaneously handling tasks such as automatic speech recognition (ASR), speech translation (AST), and machine translation (MT), and pre-trained on over 100 languages.
"Why did the Model Fail?": Attributing Model Performance Changes to Distribution Shifts
Haoran Zhang (Massachusetts Institute of Technology), Shalmali Joshi (Columbia University)
Domain AdaptationExplainability and InterpretabilityBiomedical Data
🎯 What it does: This paper models the problem of attributing model performance degradation as a cooperative game, using causal graphs to decompose distribution shifts into actionable mechanisms, and allocates overall performance changes to each mechanism through Shapley values.
$\pi$-Tuning: Transferring Multimodal Foundation Models with Optimal Multi-task Interpolation
Chengyue Wu (University of Hong Kong), Ping Luo (University of Hong Kong)
Domain AdaptationKnowledge DistillationTransformerPrompt EngineeringMixture of ExpertsImageTextMultimodality
🎯 What it does: This paper proposes π-Tuning, a parameter-efficient transfer learning method that utilizes task similarity for expert parameter interpolation.
$H$-Consistency Bounds for Pairwise Misranking Loss Surrogates
Anqi Mao (Courant Institute of Mathematical Sciences), Yutao Zhong
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This study investigates the H-consistency upper bounds when using alternative loss functions in score-based ranking. It proves that under commonly used homogeneity assumption sets (such as linear models and single hidden layer ReLU networks), the traditional pairwise misranking loss cannot achieve meaningful H-consistency guarantees. It proposes introducing a 'dropout' mechanism (pairwise or bipartite dropout loss) in ranking to obtain feasible theoretical bounds.
2D-Shapley: A Framework for Fragmented Data Valuation
Liu Zhihong, Ruoxi Jia (Virginia Tech)
Anomaly DetectionData-Centric LearningTabular
🎯 What it does: A 2D-Shapley framework is proposed to quantify the contributions of fragmented data (neither sharing features nor samples), along with corresponding fairness axioms and analytical expressions.
A Category-theoretical Meta-analysis of Definitions of Disentanglement
Yivan Zhang (University of Tokyo), Masashi Sugiyama (RIKEN AIP)
🎯 What it does: A meta-analysis of the definitions of disentanglement is conducted, unifying and comparing various definitions proposed in the academic community using category theory.
A Closer Look at Few-shot Classification Again
Xu Luo (University of Electronic Science and Technology of China), Jingkuan Song (University of Electronic Science and Technology of China)
ClassificationMeta LearningSupervised Fine-TuningImage
🎯 What it does: The system evaluates and demonstrates that training algorithms and adaptation algorithms can be completely decoupled in few-shot classification, with a detailed analysis of the number of classes, data scale, differences between supervised and self-supervised learning during the training phase, and the performance of algorithms such as linear, matching, and fine-tuning during the adaptation phase.
A Closer Look at Self-Supervised Lightweight Vision Transformers
Shaoru Wang (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationKnowledge DistillationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper studies the use of self-supervised pre-training methods on lightweight Vision Transformers and evaluates their impact on downstream tasks.
A Closer Look at the Intervention Procedure of Concept Bottleneck Models
Sungbin Shin (POSTECH), Namhoon Lee (POSTECH)
ClassificationData SynthesisImage
🎯 What it does: The study systematically evaluates the intervention process of the Concept Bottleneck Model (CBM), proposing various criteria for concept selection and analyzing their performance under different intervention levels, training schemes, concept expression methods, and data attributes.
A Complete Expressiveness Hierarchy for Subgraph GNNs via Subgraph Weisfeiler-Lehman Tests
Bohang Zhang (Peking University), Liwei Wang (Peking University)
Graph Neural NetworkGraph
🎯 What it does: This paper systematically studies the expressiveness of subgraph Graph Neural Networks (subgraph GNNs), proposes and fully constructs the Subgraph Weisfeiler–Lehman (SWL) hierarchy, reveals the expressive power of different aggregation and pooling designs, and proves that the strongest form is SSWL.
A Conditional Normalizing Flow for Accelerated Multi-Coil MR Imaging
Jeffrey Wen (Ohio State University), Philip Schniter (Ohio State University)
RestorationGenerationFlow-based ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A conditional normalization flow (CNF) for multi-channel accelerated MRI has been designed and implemented to quickly sample from the posterior distribution to generate complete images.
A Connection between One-Step RL and Critic Regularization in Reinforcement Learning
Benjamin Eysenbach (Carnegie Mellon University), Ruslan Salakhutdinov (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: This paper proves through theoretical derivation and experimental validation that when the regularization strength λ=1, first-order RL and Critic regularization yield the same policy, and explores their performance in offline RL.
A Coupled Flow Approach to Imitation Learning
Gideon Joseph Freund (Bar Ilan University), Sarit Kraus (Bar Ilan University)
Reinforcement LearningFlow-based ModelSequential
🎯 What it does: Proposes a Coupled Normalizing Flow-based imitation learning method (CFIL), utilizing the Donsker-Varadhan form of KL for distribution matching;
A Critical Revisit of Adversarial Robustness in 3D Point Cloud Recognition with Diffusion-Driven Purification
Jiachen Sun (University of Michigan), Chaowei Xiao (NVIDIA)
RecognitionAdversarial AttackDiffusion modelContrastive LearningPoint Cloud
🎯 What it does: A preprocessing method based on diffusion models, PointDP, is proposed to eliminate adversarial perturbations in 3D point clouds.
A Critical View of Vision-Based Long-Term Dynamics Prediction Under Environment Misalignment
Hanchen Xie (University of Southern California), Wael AbdAlmageed (University of Southern California)
SegmentationDomain AdaptationImageBenchmark
🎯 What it does: In visual long-term dynamic prediction, this paper systematically explores the impact of environmental mismatch (Cross-Domain and Cross-Context) on model performance, using RPCIN as a probe for evaluation.
A Deep Conjugate Direction Method for Iteratively Solving Linear Systems
Ayano Kaneda (Waseda University), Joseph Teran (University of California)
OptimizationConvolutional Neural NetworkTabular
🎯 What it does: A deep learning-based conjugate direction method (DCDM) is proposed to iteratively solve large-scale sparse symmetric positive definite linear systems, focusing on the solution of the discrete Poisson equation.
A Distribution Optimization Framework for Confidence Bounds of Risk Measures
Hao Liang (Chinese University of Hong Kong), Zhi-Quan Luo
OptimizationTabularFinance Related
🎯 What it does: This paper proposes a confidence interval framework based on distribution optimization to solve the confidence upper and lower bounds for multi-class risk measures (such as CVaR, SRM, DRM, ERM, CE, RDEU);
A Fast Optimistic Method for Monotone Variational Inequalities
Michael Sedlmayer (University of Vienna), Radu Ioan Bot (University of Vienna)
OptimizationGenerative Adversarial NetworkImage
🎯 What it does: A new fast optimistic method fOGDA-VI is proposed for solving monotone variational inequalities, using only one operator evaluation and one projection at each step.
A Fast, Well-Founded Approximation to the Empirical Neural Tangent Kernel
Mohamad Amin Mohamadi (University of British Columbia), Danica J. Sutherland (Alberta Machine Intelligence Institute)
OptimizationComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: A fast and theoretically grounded empirical neural tangent kernel (eNTK) approximation method, called 'sum of logits', is proposed, which converges to the true eNTK at initialization.
A Flexible Diffusion Model
weitao Du, Yuanqi Du (Cornell University)
GenerationData SynthesisDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes a general method for parameterizing the forward process of diffusion models using the frameworks of Riemannian geometry and symplectic geometry, called FP-Diffusion;
A Framework for Adapting Offline Algorithms to Solve Combinatorial Multi-Armed Bandit Problems with Bandit Feedback
Guanyu Nie (Iowa State University), Christopher John Quinn (Iowa State University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: A general framework is proposed that adapts offline approximation algorithms to the combinatorial multi-armed bandit (CMAB) problem relying solely on feedback, addressing the issue of stochastic combinatorial multi-armed bandits.
A Fully First-Order Method for Stochastic Bilevel Optimization
Jeongyeol Kwon (University of Wisconsin Madison), Robert D Nowak
OptimizationTabular
🎯 What it does: A completely first-order gradient-based stochastic bilevel optimization algorithm (FSA2 and its momentum version FSA3) is proposed, and its non-asymptotic convergence properties are provided.
A Game-Theoretic Framework for Managing Risk in Multi-Agent Systems
Oliver Slumbers (University College London), Jun Wang (Peking University)
Autonomous DrivingOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes a new Risk-Averse Equilibrium (RAE) for multi-agent systems that considers the risks posed by the behaviors of other agents.
A General Representation Learning Framework with Generalization Performance Guarantees
Junbiao Cui (Shanxi University), Jiye Liang (Shanxi University)
OptimizationRepresentation LearningTabular
🎯 What it does: A representation learning framework based on VC dimension is proposed, which achieves kernel function selection and DNN enhancement through this framework.
A Generalization of ViT/MLP-Mixer to Graphs
Xiaoxin He (National University of Singapore), Xavier Bresson (National University of Singapore)
Representation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes a Graph ViT/MLP-Mixer network that migrates the Vision Transformer / MLP-Mixer architecture to graph structures for graph representation learning.
A Gromov--Wasserstein Geometric View of Spectrum-Preserving Graph Coarsening
Yifan Chen (Hong Kong Baptist University), Jie Chen (IBM Research)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the coarsening problem of graphs from the geometric perspective of Gromov-Wasserstein (GW) distance. An upper bound is provided for the distance difference between the original graph and the coarsened graph in the GW space, and it is proven that this upper bound depends only on the spectral difference, thereby clarifying the effectiveness of spectral preservation in multi-graph tasks. Based on this theory, the coarsening objective is transformed into a weighted kernel k-means clustering problem, and a new coarsening method called KGC is proposed along with its algorithm implementation.
A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining
Shengchao Liu (Mila Quebec Artificial Intelligence Institute), Jian Tang (HEC Montreal)
GenerationDrug DiscoveryGraph Neural NetworkContrastive LearningMultimodalityPoint CloudGraphStochastic Differential Equation
🎯 What it does: This paper proposes a stochastic differential equation (SDE) model that utilizes group symmetry (SE(3) equivariance) for multimodal pre-training of molecular two-dimensional topological structures and three-dimensional conformations.
A Hybrid Quantum-Classical Approach based on the Hadamard Transform for the Convolutional Layer
Hongyi Pan (University of Illinois Chicago), Ahmet Cetin
ClassificationRecognitionComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A hybrid quantum-classical convolution layer based on Hadamard transform (HT-Perceptron) is proposed, which can replace traditional Conv2D to achieve frequency domain convolution.
A Kernel Stein Test of Goodness of Fit for Sequential Models
Jerome Baum (University College London), Arthur Gretton (University College London)
TextSequential
🎯 What it does: A goodness-of-fit test method based on Kernel Stein Divergence (KSD) is proposed, specifically designed for variable-dimensional sequences (such as text and Markov chains), along with the corresponding Zanella-Stein operator and KSD statistic.
A Kernel-Based View of Language Model Fine-Tuning
Sadhika Malladi (Princeton University), Sanjeev Arora (Princeton University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper addresses the fine-tuning of pre-trained language models for downstream tasks from the perspective of the Neural Tangent Kernel (NTK). It derives the early training kernel (SignGD kernel) suitable for the Adam optimizer and demonstrates that during the fine-tuning process using natural language prompts, model updates often follow kernel behavior. Additionally, it explains the effectiveness of low-rank subspace fine-tuning methods (such as LoRA) from a kernel perspective.
A Kernelized Stein Discrepancy for Biological Sequences
Alan Nawzad Amin, Debora Susan Marks
GenerationData SynthesisAnomaly DetectionSequentialBiomedical Data
🎯 What it does: This paper proposes and implements the Kernelized Stein Discrepancy for evaluating biological sequence generation models (KSD-B), and validates its effectiveness both theoretically and experimentally.
A Large-Scale Study of Probabilistic Calibration in Neural Network Regression
Victor Dheur (University of Mons), Souhaib Ben Taieb (University of Mons)
Tabular
🎯 What it does: This study systematically evaluates the probability calibration of neural network regression models on 57 tabular datasets and conducts large-scale experiments on various calibration methods (post-hoc calibration, quantile calibration, quantile prediction, regularization methods, etc.);
A Law of Robustness beyond Isoperimetry
Yihan Wu (University of Maryland), Hongyang Zhang (University of Waterloo)
🎯 What it does: This paper studies the robust interpolation problem under arbitrary data distributions, deriving the Lipschitz lower bound of the interpolation function and proposing a second-order robustness law: when the sample size is polynomial in the degree, over-parameterization may enhance robustness; however, when the sample size grows exponentially, no O(1)-Lipschitz robust interpolation function can exist.
A Mathematical Model for Curriculum Learning for Parities
Elisabetta Cornacchia (École Polytechnique Fédérale de Lausanne), Elchanan Mossel (Massachusetts Institute of Technology)
OptimizationConvolutional Neural NetworkTabular
🎯 What it does: The researchers proposed and analyzed a curriculum learning (CL) strategy for efficiently learning k-parity Boolean functions under different product distributions (i.e., different bias parameters p) by generating and sequentially presenting samples.
A Model-Based Method for Minimizing CVaR and Beyond
Si Yi Meng (Cornell University), Robert M. Gower (Flatiron Institute)
OptimizationTabularFinance Related
🎯 What it does: A model-based stochastic prox-linear method SPL+ is proposed for minimizing CVaR, along with closed-form updates and convergence analysis.
A Model-free Closeness-of-influence Test for Features in Supervised Learning
Mohammad Mehrabi (University of Southern California), Ryan A. Rossi (Adobe)
ImageTabular
🎯 What it does: This paper proposes a model-free feature influence proximity test method to determine whether the effects of two features on a response variable are similar.
A Modern Look at the Relationship between Sharpness and Generalization
Maksym Andriushchenko (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (Tübingen AI Center)
ClassificationOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper systematically evaluates the relationship between adaptive sharpness, which is invariant under reparameterization, and generalization performance in modern deep learning settings (such as ImageNet, CIFAR-10, CLIP-BERT transfer learning, and Vision Transformer).
A Near-Optimal Algorithm for Safe Reinforcement Learning Under Instantaneous Hard Constraints
Ming Shi (Ohio State University), Ness Shroff (Ohio State University)
OptimizationSafty and PrivacyReinforcement LearningTabular
🎯 What it does: An approximate optimal reinforcement learning algorithm (LSVI-NEW) is proposed that guarantees safety under instantaneous hard constraints, achieving approximately optimal returns in linear mixed MDPs with unsafe states and actions while ensuring that safety constraints are met at every step.
A Nearly-Optimal Bound for Fast Regression with $\ell_\infty$ Guarantee
Zhao Song (Adobe Research), Lichen Zhang (Massachusetts Institute of Technology)
OptimizationTabular
🎯 What it does: This paper studies the fast regression problem with ℓ ∞ guarantees and proposes a new algorithmic framework that utilizes dense random matrices for regression, ensuring nearly optimal row count and performance.
A Neural PDE Solver with Temporal Stencil Modeling
Zhiqing Sun (Carnegie Mellon University), Shinjae Yoo (Brookhaven National Laboratory)
Super ResolutionOptimizationConvolutional Neural NetworkTime SeriesPhysics Related
🎯 What it does: A Temporal Stencil Modeling (TSM) method has been developed to achieve learnable flux approximation in time-varying partial differential equations in conservation form through a convolution volume method.
A new near-linear time algorithm for k-nearest neighbor search using a compressed cover tree
Yury Elkin (University of Liverpool), Vitaliy Kurlin (University of Liverpool)
OptimizationComputational EfficiencyPoint Cloud
🎯 What it does: A new compressed cover tree algorithm is proposed for k-nearest neighbor search in near-linear time.
A New PHO-rmula for Improved Performance of Semi-Structured Networks
David Rügamer (Ludwig Maximilian University of Munich)
Explainability and InterpretabilityComputational EfficiencyTabularTime Series
🎯 What it does: The paper proposes a Post-Hoc Orthogonalization (PHO) method to address the identifiability issue between structured and unstructured predictors in Semi-Structured Neural Networks (SSN), achieving interpretability of structured effects while maintaining predictive performance.
A Picture of the Space of Typical Learnable Tasks
Rahul Ramesh (University of Pennsylvania), Pratik Chaudhari (University of Pennsylvania)
ClassificationMeta LearningContrastive LearningImage
🎯 What it does: The paper uses information geometry methods to unify the modeling, distance measurement, trajectory re-indexing, and visualization of deep network prediction models trained on different tasks and learning algorithms (supervised, meta-learning, semi-supervised, self-supervised, contrastive learning, etc.), thereby revealing the structure and commonalities of the learnable task space.
A Reinforcement Learning Framework for Dynamic Mediation Analysis
Lin Ge (North Carolina State University), Rui Song (North Carolina State University)
OptimizationReinforcement LearningTime SeriesBiomedical Data
🎯 What it does: A dynamic mediation analysis framework based on reinforcement learning is proposed, constructing a Mediated Markov Decision Process (MMDP), and decomposing the Average Treatment Effect (ATE) into Immediate Direct Effect (IDE), Immediate Mediation Effect (IME), Delayed Direct Effect (DDE), and Delayed Mediation Effect (DME).
A Robust Optimisation Perspective on Counterexample-Guided Repair of Neural Networks
David Boetius (University of Konstanz), Tobias Sutter (University of Konstanz)
OptimizationImage
🎯 What it does: This paper views counterexample-driven neural network repair as a robust optimization problem, exploring its theoretical properties and practical implementation.
A Robust Test for the Stationarity Assumption in Sequential Decision Making
Jitao Wang (University of Michigan), Zhenke Wu (University of Michigan)
Reinforcement LearningTabularTime SeriesSequentialBiomedical DataElectronic Health Records
🎯 What it does: A dual robust testing method in offline reinforcement learning is proposed to detect whether the stationarity assumption of the Markov Decision Process (MDP) holds.
A Scalable Frank-Wolfe-Based Algorithm for the Max-Cut SDP
Chi Bach Pham (Monash University), James Saunderson (Monash University)
OptimizationGaussian SplattingGraph
🎯 What it does: This paper proposes a scalable Frank-Wolfe-based algorithm for solving large-scale Max-Cut semidefinite programming (SDP), achieving low memory implementation through linear optimization of the symmetric maximum eigenvalue.
A Simple Zero-shot Prompt Weighting Technique to Improve Prompt Ensembling in Text-Image Models
James Urquhart Allingham (University of Cambridge), Balaji Lakshminarayanan (Google Research)
ClassificationAnomaly DetectionTransformerPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: A zero-shot prompt weighted combination method (ZPE) is proposed to automatically generate the optimal prompt combination for text-image models without using any labeled validation set.
A Statistical Perspective on Retrieval-Based Models
Soumya Basu (Google), Manzil Zaheer (Google DeepMind)
ClassificationRetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper systematically studies the theoretical properties of retrieval models in classification tasks from a statistical perspective, and provides generalization error upper bounds for two types of retrieval learning frameworks: Local Empirical Risk Minimization (Local-ERM) and kernel methods used in an extended feature space.
A Study of Global and Episodic Bonuses for Exploration in Contextual MDPs
Mikael Henaff (Meta AI Research), Roberta Raileanu (Meta AI Research)
Reinforcement Learning
🎯 What it does: The study investigates the effects of global and episodic exploration rewards in contextual MDPs and proposes a combination strategy.
A Study on Transformer Configuration and Training Objective
Fuzhao Xue (National University of Singapore), Yang You
TransformerAuto EncoderImageTextBenchmark
🎯 What it does: This study investigates the relationship between Transformer configurations and training objectives, demonstrating that the Masked Autoencoder can alleviate the over-smoothing problem, and based on this, proposes a deeper and narrower configuration (Bamboo).
A Theoretical Analysis of the Learning Dynamics under Class Imbalance
Emanuele Francazi (Eidgenössische Technische Hochschule Zürich), Aurelien Lucchi (University of Basel)
OptimizationTabular
🎯 What it does: This paper theoretically analyzes the learning dynamics of Gradient Descent (GD) and Stochastic Gradient Descent (SGD) under class imbalance, revealing the phenomenon of minority class learning inertia (MID), and proposes a method to alleviate this issue through per-class normalized gradient descent (PCNGD/PCNSGD).
A theory of continuous generative flow networks
Salem Lahlou (Mila), Nikolay Malkin (Mila)
GenerationData SynthesisOptimizationFlow-based ModelImageTabular
🎯 What it does: A general theory of Generative Flow Networks (GFlowNets) for continuous and mixed spaces is proposed, providing three training objectives: flow matching, detailed balance, and trajectory balance, and designing corresponding gradient-computable loss functions within this framework.
A theory of representation learning gives a deep generalisation of kernel methods
Adam X. Yang (University of Bristol), Laurence Aitchison (University of Bristol)
Representation LearningTabular
🎯 What it does: A new infinite width limit - Bayesian representation learning limit is proposed, which can maintain representation learning in deep Bayesian networks (including Deep Gaussian Processes, DGP); under this limit, the posterior of DGP is multivariate Gaussian, and the corresponding DKM (Deep Kernel Machine) objective is provided; sparse DKM is achieved through inducing point techniques, resulting in linear time complexity;
A Three-regime Model of Network Pruning
Yefan Zhou (University of California), Michael W. Mahoney (University of California)
CompressionOptimizationHyperparameter SearchConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a three-stage model based on two parameters: 'temperature' and 'load', to predict how hyperparameters during the training phase (such as the number of training epochs and batch size) affect the subsequent network pruning results. The model categorizes the gradient landscapes of the network before and after pruning into three regimes (Regime I, II-A, II-B).
A Toy Model of Universality: Reverse Engineering how Networks Learn Group Operations
Bilal Chughtai, Neel Nanda
TransformerGraph
🎯 What it does: This study investigates the internal mechanisms of small neural networks when performing finite group composition tasks, discovering that the network implements a general algorithm based on representation theory (GCR), which is validated through reverse engineering and ablation studies.
A Two-Stage Active Learning Algorithm for k-Nearest Neighbors
Nicholas Rittler, Kamalika Chaudhuri (University of California)
Classification
🎯 What it does: A two-stage active learning algorithm is proposed for training k-nearest neighbor classifiers.
A Unified Audio-Visual Learning Framework for Localization, Separation, and Recognition
Shentong Mo (Carnegie Mellon University), Pedro Morgado (University of Wisconsin Madison)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningMultimodalityAudio
🎯 What it does: A unified audio-visual learning framework, OneAVM, is proposed, achieving single-model inference for audio source localization, separation, and recognition tasks.
A Unified Optimization Framework of ANN-SNN Conversion: Towards Optimal Mapping from Activation Values to Firing Rates
Haiyan Jiang (Mohamed bin Zayed University of Artificial Intelligence), Bin Gu (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationSpiking Neural NetworkImage
🎯 What it does: This study investigates a unified conversion framework from ANN to SNN and proposes the SlipReLU activation function to balance conversion error and model performance loss.
A Unifying Framework to the Analysis of Interaction Methods using Synergy Functions
Daniel Lundstrom (University of Southern California), Meisam Razaviyayn (University of Southern California)
🎯 What it does: This paper proposes a unified framework that utilizes the synergy function to analyze and construct attribution and interaction methods for continuous input models, providing a corresponding axiomatic foundation.
A Universal Unbiased Method for Classification from Aggregate Observations
Zixi Wei (Chongqing University), Heng Tao Shen (University of Electronic Science and Technology of China)
ClassificationConvolutional Neural NetworkContrastive LearningImageTabular
🎯 What it does: This paper proposes a general unbiased risk estimator for classification learning under the condition of only providing aggregated observations of instance groups (CFAO).
A Watermark for Large Language Models
John Kirchenbauer (University of Maryland), Tom Goldstein (University of Maryland)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A method for implicit watermarking that can be embedded in the output of large language models is proposed, along with a publicly available detection algorithm, allowing for the determination of whether a short text fragment is machine-generated.
A/B Testing in Network Data with Covariate-Adaptive Randomization
Jialu Wang (George Washington University), Feifang Hu (George Washington University)
GraphTabularAgriculture Related
🎯 What it does: An adaptive randomization algorithm that simultaneously balances covariates and network structure is proposed for estimating average treatment effects in network A/B testing.
AbODE: Ab initio antibody design using conjoined ODEs
Yogesh Verma (Aalto University), Vikas Garg (YaiYai Ltd)
GenerationData SynthesisDrug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerGraphBiomedical DataOrdinary Differential Equation
🎯 What it does: A generative model named AbODE is proposed, which jointly designs the CDR sequences and three-dimensional structures of antibodies, and performs conditional generation based on antigen information.
Abstract-to-Executable Trajectory Translation for One-Shot Task Generalization
Stone Tao (University of California San Diego), Hao Su (University of California San Diego)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: Construct a simplified abstract environment, generate abstract trajectories, and then translate them into executable trajectories using a sequence-to-sequence model, thereby achieving one-shot task generalization.
Abstracting Imperfect Information Away from Two-Player Zero-Sum Games
Samuel Sokota (Carnegie Mellon University), Noam Brown (Meta AI)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a method to transform imperfect information in two-player zero-sum games into a complete information game through public strategy announcements, and solves the regularized minimization (UG) equilibrium of PuB-AMG (Public Belief Alternating Markov Game) to avoid the non-correspondence issues present in traditional methods.
ACAT: Adversarial Counterfactual Attention for Classification and Detection in Medical Imaging
Alessandro Fontanella (University of Edinburgh), Amos Storkey (University of Edinburgh)
ClassificationObject DetectionConvolutional Neural NetworkAuto EncoderImageBiomedical DataComputed Tomography
🎯 What it does: The ACAT framework is proposed, which utilizes significant maps generated by adversarial counterfactuals to modulate features with soft attention at multiple scales, thereby improving the accuracy of medical image classification and localization.
Accelerated Cyclic Coordinate Dual Averaging with Extrapolation for Composite Convex Optimization
Cheuk Yin Lin (University of Wisconsin Madison), Jelena Diakonikolas (University of Wisconsin Madison)
OptimizationTabular
🎯 What it does: An accelerated cyclic coordinate dual averaging algorithm (A-CODER) and its variant quantization version (VR-ACODER) are proposed for solving combinatorial optimization problems that can be expressed as smooth convex plus non-smooth convex.
Accelerated Infeasibility Detection of Constrained Optimization and Fixed-Point Iterations
Jisun Park (Seoul National University), Ernest K. Ryu (Seoul National University)
Optimization
🎯 What it does: This study investigates the convergence rates of fixed point iterations (Picard, Krasnosel’skiĭ-Mann, Halpern, Mann) in the case of non-convergence (infeasibility) when approximating infinitesimal displacement vectors (i.e., evidence of infeasibility), providing matching upper and lower bounds;
Accelerated Primal-Dual Methods for Convex-Strongly-Concave Saddle Point Problems
Mohammad Khalafi (Southern Methodist University), Digvijay Boob (Southern Methodist University)
Optimization
🎯 What it does: An accelerated linearized primal-dual (ALPD) method for convex-strongly concave saddle point problems is proposed, improving the performance of traditional linearized PD methods under strong duality concavity conditions.
Accelerated Stochastic Optimization Methods under Quasar-convexity
Qiang Fu (Sun Yat-sen University), Ashia Camage Wilson
OptimizationTabularTime Series
🎯 What it does: Two accelerated stochastic optimization methods for star-convex functions are proposed—QASGD and QASVRG, which can achieve fast convergence under L-smooth and (strong) star-convex conditions;
Accounting For Informative Sampling When Learning to Forecast Treatment Outcomes Over Time
Toon Vanderschueren (KU Leuven), Mihaela van der Schaar (University of Cambridge)
Drug DiscoveryReinforcement LearningTime SeriesSequentialBiomedical Data
🎯 What it does: The study investigates how to learn to predict treatment outcomes over time in the presence of informative sampling in observational data.
Accuracy on the Curve: On the Nonlinear Correlation of ML Performance Between Data Subpopulations
Weixin Liang (Stanford University), James Zou (Stanford University)
Domain AdaptationAuto EncoderImage
🎯 What it does: The study investigates the correlation between the performance of machine learning models across different subpopulations under conditions of subpopulation distribution shift, and systematically explores the nonlinear ('moon-shaped') characteristics of this correlation.
Achieving Hierarchy-Free Approximation for Bilevel Programs with Equilibrium Constraints
Jiayang Li (Northwestern University), Zhaoran Wang (Northwestern University)
Optimization
🎯 What it does: This paper proposes a method to transform the originally difficult-to-solve bi-level optimization problem with equilibrium constraints into two single-level solvable approximate problems by constructing T-step Cournot games and T-step monopoly models, and provides theoretical proof that both can yield arbitrarily convergent upper and lower bounds, thus achieving a hierarchy-free approximate solution.
Achieving High Accuracy with PINNs via Energy Natural Gradient Descent
Johannes Müller, Marius Zeinhofer (Simula Research Laboratory)
OptimizationSequentialPhysics Related
🎯 What it does: An optimization method based on energy natural gradient is proposed for training Physics-Informed Neural Networks (PINNs) and the Deep Ritz method.
Achieving Linear Speedup in Non-IID Federated Bilevel Learning
Minhui Huang (University of California), Kaiyi Ji (University at Buffalo)
OptimizationFederated LearningImage
🎯 What it does: This paper studies the federated bi-level optimization problem under non-i.i.d. data distribution and proposes a dual-loop algorithm called FedMBO, which supports partial client participation and achieves theoretical linear speedup.
Action Matching: Learning Stochastic Dynamics from Samples
Kirill Neklyudov (Vector Institute), Alireza Makhzani (Vector Institute)
GenerationData SynthesisTime SeriesSequentialBiomedical DataPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the Action Matching method, which utilizes cross-temporal sample learning of continuous dynamics and is capable of simulating entire particle trajectories.
Active causal structure learning with advice
Davin Choo (National University of Singapore), Arnab Bhattacharyya (National University of Singapore)
Graph
🎯 What it does: Proposes an adaptive intervention search algorithm that utilizes expert-provided (possibly imperfect) DAG suggestions in causal structure learning to actively determine experimental designs that minimize the number of interventions.
Active Learning based Structural Inference
Aoran Wang (University of Luxembourg), Jun Pang (University of Luxembourg)
Anomaly DetectionOptimizationGraph Neural NetworkReinforcement LearningContrastive LearningTime Series
🎯 What it does: A structure inference framework ALaSI based on deep active learning is proposed for inferring directed connectivity from time series states in large-scale dynamic systems.
Active Policy Improvement from Multiple Black-box Oracles
Xuefeng Liu (University of Chicago), Yuxin Chen (University of Chicago)
Robotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes an active policy improvement framework called MAPS and its extension MAPS-SE, which can actively select the most suitable expert for the current state and perform value function estimation in environments where only multiple sub-optimal black-box experts are available, thereby enhancing the performance of the learning policy.
Active Ranking of Experts Based on their Performances in Many Tasks
El Mehdi Saad (INRAE Mistea Institut Agro Univ Montpellier), Alexandra Carpentier (Institut für Mathematik Universität Potsdam)
🎯 What it does: Proposes a method for expert active ranking and optimal expert identification under the assumption of monotonicity, which can dynamically allocate query amounts by adapting to performance differences between tasks;
Actor-Critic Alignment for Offline-to-Online Reinforcement Learning
Zishun Yu (University of Illinois Chicago), Xinhua Zhang (University of Illinois Chicago)
Reinforcement LearningTabular
🎯 What it does: A new offline-to-online reinforcement learning framework is proposed, which aligns the policy and value function obtained from offline learning through an 'actor-critic alignment' step, followed by online fine-tuning.
AdaBoost is not an Optimal Weak to Strong Learner
Mikael Møller Høgsgaard (Aarhus University), Martin Ritzert (Georg-August-University Göttingen)
Adversarial Attack
🎯 What it does: This paper studies the sample complexity of the AdaBoost algorithm, proving that it is not optimal in the process of transforming weak learners into strong learners, as there exists a suboptimal logarithmic factor in terms of the required accuracy.
AdaNPC: Exploring Non-Parametric Classifier for Test-Time Adaptation
YiFan Zhang, Tieniu Tan (Nanjing University)
Domain AdaptationImage
🎯 What it does: Proposes AdaNPC, a framework that utilizes a KNN non-parametric classifier for online adaptation during testing, storing and gradually updating source domain features and labels in memory.
AdaptDiffuser: Diffusion Models as Adaptive Self-evolving Planners
Zhixuan Liang (University of Hong Kong), Ping Luo (University of Hong Kong)
Robotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: This paper proposes AdaptDiffuser, a self-evolving diffusion model planner that uses reward gradients to guide the generation of synthetic expert data and filters it through a discriminator to enhance the performance and generalization of offline RL.
Adapting to game trees in zero-sum imperfect information games
Côme Fiegel (ENSAE Paris), Michal Valko (DeepMind)
OptimizationReinforcement LearningSequential
🎯 What it does: In zero-sum games with incomplete information, ε-optimal strategies are learned through self-play trajectory feedback, proposing and analyzing two FTRL-based algorithms—Balanced FTRL and Adaptive FTRL.
Adaptive Annealed Importance Sampling with Constant Rate Progress
Shirin Goshtasbpour (ETH Zurich), Fernando Perez-Cruz (ETH Zurich)
TabularOrdinary Differential Equation
🎯 What it does: Research and improve the intermediate distribution path and scheduling strategy of Annealed Importance Sampling (AIS), proposing a constant rate scheduling and implementing the Constant Rate AIS (CR-AIS) algorithm.
Adaptive Barrier Smoothing for First-Order Policy Gradient with Contact Dynamics
Shenao Zhang (Northwestern University), Zhaoran Wang (Northwestern University)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: Proposes the Adaptive Barrier Smoothing (ABS) method to reduce the gradient variance of FOPG in hard contact models and control bias, thereby achieving more stable reinforcement learning optimization.
Adaptive Compositional Continual Meta-Learning
Bin Wu (Zhejiang University), Qiang Zhang (Zhejiang University)
Meta LearningImage
🎯 What it does: An adaptive combinatorial continuous meta-learning algorithm ACML is proposed for learning shareable meta-knowledge in non-stationary task sequences.
Adaptive Computation with Elastic Input Sequence
Fuzhao Xue (Google), Yang You (National University of Singapore)
ClassificationRecognitionComputational EfficiencyTransformerImage
🎯 What it does: Proposes AdaTape, which achieves adaptive computation on input sequences through variable-length tape tokens;
Adaptive Coordination in Social Embodied Rearrangement
Andrew Szot (Georgia Institute of Technology), Akshara Rai (Georgia Institute of Technology)
Robotic IntelligenceReinforcement LearningPoint Cloud
🎯 What it does: This paper proposes a vision-based multi-robot collaborative task called Social Rearrangement and designs a zero-shot coordination framework named Behavior Diversity Play (BDP) for this task.
Adaptive Estimation of Graphical Models under Total Positivity
Jiaxi Ying (Hong Kong University of Science and Technology), Daniel P. Palomar (Hong Kong University of Science and Technology)
OptimizationGraph Neural NetworkGraphTime SeriesFinance Related
🎯 What it does: An adaptive multi-stage estimation method is proposed for estimating the precision matrix of M-matrices or diagonally dominant M-matrices under the MTP2 Gaussian graphical model.
Adaptive Identification of Populations with Treatment Benefit in Clinical Trials: Machine Learning Challenges and Solutions
Alicia Curth (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
TabularBiomedical Data
🎯 What it does: This paper studies how to adaptively identify beneficiary populations in confirmatory clinical trials and proposes two meta-algorithms, AdaGGI and AdaGCPI.
Adaptive IMLE for Few-shot Pretraining-free Generative Modelling
Mehran Aghabozorgi (Simon Fraser University), Ke Li (Simon Fraser University)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: Proposes the Adaptive IMLE method, improving IMLE to address the mode collapse problem in few-shot generation, using adaptive thresholds and curriculum learning to dynamically adjust for different training sample difficulties.
Adaptive Smoothing Gradient Learning for Spiking Neural Networks
Ziming Wang (Zhejiang University), Huajin Tang (Zhejiang University)
ClassificationOptimizationSpiking Neural NetworkImageTime SeriesSequentialAudio
🎯 What it does: An Adaptive Smoothing Gradient Learning (ASGL) method is proposed for directly training deep Spiking Neural Networks (SNNs), enabling the network to gradually converge to a true spiking network by mixing simulated activations and random spike noise during training.
Adaptive Whitening in Neural Populations with Gain-modulating Interneurons
Lyndon Duong, Eero P Simoncelli
Recurrent Neural NetworkSequential
🎯 What it does: An adaptive whitening network based on neural gain modulation is designed, which can adjust the whitening of inputs online without changing the synaptic weights.
Adaptively Weighted Data Augmentation Consistency Regularization for Robust Optimization under Concept Shift
Yijun Dong (University of Texas at Austin), Rachel Ward (University of Texas at Austin)
SegmentationOptimizationConvolutional Neural NetworkTransformerImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposes the Adaptive Weighted Data Augmentation Consistency Regularization (AdaWAC) algorithm to address the issues of concept drift and the imbalance of sparse/dense label information in medical image segmentation.
Additive Causal Bandits with Unknown Graph
Alan Malek (DeepMind), Silvia Chiappa (DeepMind)
Graph
🎯 What it does: This paper proposes an algorithm for the causal bandit problem with unknown causal graphs, utilizing the absence of latent confounding and the additive assumption of outcomes, transforming the problem into a combinatorial linear bandit and designing the MODL algorithm.
Addressing Budget Allocation and Revenue Allocation in Data Market Environments Using an Adaptive Sampling Algorithm
Boxin Zhao (University of Chicago), mladen kolar
OptimizationFederated LearningImage
🎯 What it does: A linear-time adaptive sampling algorithm is proposed, which simultaneously addresses the budget allocation and revenue distribution issues in data markets.
Adversarial Cheap Talk
Chris Lu (FLAIR University of Oxford), Jakob Nicolaus Foerster
Adversarial AttackReinforcement LearningTabular
🎯 What it does: A minimization attack framework called Cheap Talk MDP is proposed, which studies attacking reinforcement learning agents by appending 'useless' features to the observations;