ICML 2023 Papers — Page 11
International Conference on Machine Learning · 1828 papers
Modality-Agnostic Variational Compression of Implicit Neural Representations
Jonathan Richard Schwarz (DeepMind), Jinwoo Shin (KAIST)
CompressionMeta LearningAuto EncoderImageVideoMultimodalityAudio
🎯 What it does: A cross-modal variational compression implicit neural representation (VC-INR) algorithm is proposed, which uniformly compresses various data types such as images, audio, video, 3D voxels, and climate data.
Model Ratatouille: Recycling Diverse Models for Out-of-Distribution Generalization
Alexandre Rame, David Lopez-Paz (Meta AI)
Domain AdaptationSupervised Fine-TuningImageBiomedical DataBenchmark
🎯 What it does: A strategy named Model Ratatouille is proposed, which utilizes multiple fine-tunings of the same base model on various auxiliary tasks as different initializations, followed by parallel fine-tuning on the target task and finally averaging the weights to enhance generalization capability on out-of-distribution (OOD) samples.
Model Transferability with Responsive Decision Subjects
Yatong Chen (University of California), Yang Liu (University of California)
Domain AdaptationRecommendation SystemAnomaly DetectionOptimizationTabularFinance Related
🎯 What it does: The study examines the gap between the performance of a trained model on a new distribution (i.e., induced risk) and the source distribution risk when human decision-makers respond after deploying machine learning models, leading to distribution shift.
Model-agnostic Measure of Generalization Difficulty
Akhilan Boopathy (Massachusetts Institute of Technology), Ila R Fiete
ClassificationOptimizationConvolutional Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: A model-independent metric is proposed to quantify the inductive bias complexity of tasks within an information-theoretic framework, in order to assess the generalization difficulty of tasks.
Model-Aware Contrastive Learning: Towards Escaping the Dilemmas
Zizheng Huang (Nanjing University), Chunlin Chen (Nanjing University)
ClassificationRepresentation LearningContrastive LearningImageTextGraph
🎯 What it does: This paper proposes a model-aware adaptive temperature (MACL) strategy based on contrastive learning, which adjusts the temperature by dynamically estimating the alignment of positive samples and introduces gradient reweighting to alleviate the gradient decay problem.
Model-based Offline Reinforcement Learning with Count-based Conservatism
Byeongchan Kim (Seoul National University), Min-hwan Oh (Seoul National University)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes an offline reinforcement learning method called Count-MORL based on a counting-based conservative model, which quantifies model error using the frequency of state-action pairs and implements a conservative policy through reward penalties.
Model-based Reinforcement Learning with Scalable Composite Policy Gradient Estimators
Paavo Parmas (Kyoto University), Yuma Aoki (University of Tokyo)
Recurrent Neural NetworkReinforcement LearningTabularTime Series
🎯 What it does: The paper studies the gradient estimation problem in model-based reinforcement learning and proposes a scalable Total Propagation X (TPX) composite gradient estimator.
Model-Bellman Inconsistency for Model-based Offline Reinforcement Learning
Yihao Sun (Nanjing University), Yang Yu (Nanjing University)
Reinforcement LearningTabularBenchmark
🎯 What it does: A model-based offline reinforcement learning algorithm called MOBILE is proposed, which utilizes model-Bellman inconsistency to achieve lazy value estimation.
Model-Free Robust Average-Reward Reinforcement Learning
Yue Wang (University at Buffalo), Shaofeng Zou (University at Buffalo)
Reinforcement Learning
🎯 What it does: This paper proposes two model-free robust relative value iteration (RVI) TD and RVI Q-learning algorithms for robust Markov decision processes under average rewards, and provides convergence proofs for them.
MODeL: Memory Optimizations for Deep Learning
Benoit Steiner (Anthropic), James Hegarty (Meta)
OptimizationComputational EfficiencyTransformer
🎯 What it does: This paper proposes an algorithm called MODeL based on integer linear programming to automatically optimize the lifecycle and memory placement of tensors during neural network training, significantly reducing peak memory usage.
ModelDiff: A Framework for Comparing Learning Algorithms
Harshay Shah (Massachusetts Institute of Technology), Aleksander Madry (Massachusetts Institute of Technology)
ClassificationExplainability and InterpretabilityImage
🎯 What it does: The MODELDIFF framework is proposed to compare two learning algorithms, uncovering their differences in training data utilization and generating meaningful input transformations through interpretable 'discriminative subgroups';
Modeling Dynamic Environments with Scene Graph Memory
Andrey Kurenkov (Stanford University), Roberto Martín-Martín (University of Texas at Austin)
Robotic IntelligenceGraph Neural NetworkTransformerReinforcement LearningGraphBenchmark
🎯 What it does: A link prediction framework for object search is proposed on dynamic, partially observable scene graphs, designing the Scene Graph Memory (SGM) and Node Edge Predictor (NEP) models.
Modeling Temporal Data as Continuous Functions with Stochastic Process Diffusion
Marin Biloš (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
GenerationData SynthesisTransformerDiffusion modelTime SeriesSequentialBiomedical DataFinance RelatedStochastic Differential Equation
🎯 What it does: A generative model for denoising diffusion in function space is proposed, aimed at modeling, predicting, and interpolating irregular time series, compatible with continuous time.
Moderately Distributional Exploration for Domain Generalization
Rui Dai (University of Science and Technology of China), Xinmei Tian (University of Science and Technology of China)
Domain AdaptationOptimizationAdversarial AttackImageBenchmark
🎯 What it does: In the domain generalization task, a new domain is constructed to enhance the model's generalization ability to unknown target domains by performing distributionally robust optimization in a subspace where the semantics are the same but non-semantic factors are variable.
MolDiff: Addressing the Atom-Bond Inconsistency Problem in 3D Molecule Diffusion Generation
Xingang Peng (Peking University), Jianzhu Ma (Tsinghua University)
GenerationDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A new 3D molecular generation model, MolDiff, has been designed to simultaneously generate atom types, positions, and chemical bonds during the diffusion process, guided by specific noise scheduling and a bond predictor to maintain atom-bond consistency.
Momentum Ensures Convergence of SIGNSGD under Weaker Assumptions
Tao Sun (National University of Defense Technology), Bao Wang (University of Utah)
OptimizationFederated LearningConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This paper studies the use of momentum in SIGNSGD under weak assumptions and proves its convergence, while also proposing corresponding distributed and federated learning variants.
Monge, Bregman and Occam: Interpretable Optimal Transport in High-Dimensions with Feature-Sparse Maps
marco cuturi, Pierre Ablin (Apple)
OptimizationExplainability and InterpretabilityDrug DiscoveryBiomedical Data
🎯 What it does: The study proposes an interpretable high-dimensional optimal transport (OT) mapping method centered on separable sparse regularization, utilizing the combination of Bregman centers and sparse elastic costs to generate adaptive feature-sparse transport vectors.
MonoFlow: Rethinking Divergence GANs via the Perspective of Wasserstein Gradient Flows
Mingxuan Yi (University of Bristol), Song Liu (University of Bristol)
GenerationData SynthesisGenerative Adversarial NetworkImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the MonoFlow framework, viewing the adversarial training of GANs as vector field learning and sampling of Wasserstein gradient flows.
MonoNeRF: Learning Generalizable NeRFs from Monocular Videos without Camera Poses
Yang Fu (University of California), Xiaolong Wang (University of California)
GenerationPose EstimationDepth EstimationNeural Radiance FieldAuto EncoderOptical FlowVideo
🎯 What it does: Train an autoencoder architecture called MonoNeRF, which can learn a generalizable neural radiance field (NeRF) from monocular videos without camera pose and depth annotations, and achieve single-image depth estimation, camera pose estimation, and novel view synthesis within the same model.
Monotonic Location Attention for Length Generalization
Jishnu Ray Chowdhury (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
Recurrent Neural NetworkSequential
🎯 What it does: A new positioning attention mechanism (Bi-Relative, OneStep, and Monotonic Attention) and a set of new diagnostic tasks (such as ReCopy, Reverse ReCopy, etc.) are proposed to enhance the performance of sequence-to-sequence models in length generalization.
Monotonicity and Double Descent in Uncertainty Estimation with Gaussian Processes
Liam Hodgkinson (University of Melbourne), Michael W. Mahoney (University of California)
TabularComputed Tomography
🎯 What it does: This study investigates the performance of Gaussian processes (GP) in high dimensions for uncertainty quantification (UQ), particularly focusing on the monotonicity of marginal likelihood and cross-validation metrics, as well as the double descent phenomenon.
Motion Question Answering via Modular Motion Programs
Mark Endo (Stanford University), Jiajun Wu (Stanford University)
RecognitionPose EstimationConvolutional Neural NetworkGraph Neural NetworkVideoText
🎯 What it does: This paper proposes the HumanMotionQA task and the BABEL-QA dataset, and introduces the NSPose neural-symbolic method for multi-step question answering on long sequences of human actions.
mPLUG-2: A Modularized Multi-modal Foundation Model Across Text, Image and Video
Haiyang Xu (DAMO Academy, Alibaba Group), Jingren Zhou (DAMO Academy, Alibaba Group)
TransformerLarge Language ModelImageVideoTextMultimodality
🎯 What it does: Proposes the mPLUG-2 multimodal foundation model, which employs a modular design to achieve unified pre-training for text, images, and videos.
Multi-Agent Best Arm Identification with Private Communications
Alexandre Rio (Huawei Noah's Ark Lab), Marta Soare (Université d'Orléans)
OptimizationSafty and PrivacyReinforcement LearningTabular
🎯 What it does: In the multi-agent best arm identification (BAI) problem, this paper proposes two algorithms that ensure privacy under communication constraints—DP-MASE (based on differential privacy) and CORRUPTED ELIMINATION (based on (ε,η)-privacy).
Multi-Agent Learning from Learners
Mine Melodi Caliskan (University of Tuebingen), Setareh Maghsudi (University of Tuebingen)
Reinforcement LearningTabular
🎯 What it does: A multi-agent learner learning (MA-LfL) algorithm is proposed to recover the reward function of learning agents from their trajectories in general-sum Markov games.
Multi-agent Online Scheduling: MMS Allocations for Indivisible Items
Shengwei Zhou (University of Macau), Xiaowei Wu (University of Macau)
OptimizationAgentic AI
🎯 What it does: This paper studies fair allocation of indivisible items (including goods and tasks) in an online environment, focusing on the maximum-minimum share (MMS) fairness, and provides competitive ratio analysis and algorithms for different numbers of agents and types of items.
Multi-channel Autobidding with Budget and ROI Constraints
Yuan Deng (Google Research), Vahab Mirrokni (Google Research)
OptimizationReinforcement LearningFinance Related
🎯 What it does: This study investigates how to simultaneously run advertisements across multiple advertising platforms to maximize ad conversions (such as click-through rates), while satisfying overall return on investment (ROI) and budget constraints.
Multi-class Graph Clustering via Approximated Effective $p$-Resistance
Shota Saito (University College London), Mark Herbster (University College London)
OptimizationGraph Neural NetworkGraphTabular
🎯 What it does: This paper proposes a multi-class graph clustering method based on approximate effective p-resistance, utilizing the p-pseudoinverse matrix to quickly compute distances between all pairs of points and perform k-medoids clustering.
Multi-Environment Pretraining Enables Transfer to Action Limited Datasets
David Venuto (McGill University), Ofir Nachum (Google DeepMind)
Robotic IntelligenceTransformerReinforcement LearningVideo
🎯 What it does: Achieve high-performance reinforcement learning with only a small amount of action-labeled data in the target environment by pre-training a multi-environment inverse dynamics model.
Multi-Epoch Matrix Factorization Mechanisms for Private Machine Learning
Christopher A. Choquette-Choo (Google Research), Abhradeep Guha Thakurta
OptimizationSafty and PrivacyImageText
🎯 What it does: This paper proposes a differential privacy mechanism for multi-round (multi-epoch) machine learning training, utilizing a matrix decomposition framework to achieve a better privacy-utility-computation trade-off under multiple participation.
Multi-Fidelity Covariance Estimation in the Log-Euclidean Geometry
Aimee Maurais (Massachusetts Institute of Technology), Youssef Marzouk (Massachusetts Institute of Technology)
🎯 What it does: A multi-fidelity covariance estimator is proposed, utilizing logarithmic Euclidean geometry to ensure that the estimated matrix is always positive definite;
Multi-Layer Neural Networks as Trainable Ladders of Hilbert Spaces
Zhengdao Chen (Google Research)
Tabular
🎯 What it does: This paper proposes the Neural Hilbert Ladders (NHL) framework to describe the function space that multi-layer neural networks can approximate at infinite width, and provides a corresponding complexity measure; through this framework, the relationship between approximation error, generalization error, and complexity of multi-layer networks is proven, and the dynamics of NHL evolving under gradient descent is derived in the mean field limit; at the same time, experiments validate the predictive capability of NHL dynamics for the actual gradient descent process.
Multi-Modal Classifiers for Open-Vocabulary Object Detection
Prannay Kaul (Visual Geometry Group University of Oxford), Andrew Zisserman (Visual Geometry Group University of Oxford)
Object DetectionTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A multi-modal open vocabulary object detection framework is proposed, which can dynamically generate category classifiers based on text descriptions, image examples, or a combination of both during inference, enabling the detection of unseen categories during training.
Multi-Objective GFlowNets
Moksh Jain (University of Montreal), Emmanuel Bengio (Recursion)
OptimizationDrug DiscoveryReinforcement LearningTabularSequential
🎯 What it does: This paper proposes Multi-Objective GFlowNets (MOGFNs), which learn multi-objective reward distributions under weighted preferences through conditional GFlowNet, capable of generating diverse candidates while maintaining Pareto optimality.
Multi-Objective Population Based Training
Arkadiy Dushatskiy (Centrum Wiskunde en Informatica), Peter Bosman
OptimizationHyperparameter SearchTabular
🎯 What it does: This paper proposes Multi-Objective Population Based Training (MO-PBT), a parallel hyperparameter search method that can simultaneously optimize multiple conflicting objectives during model training.
Multi-Symmetry Ensembles: Improving Diversity and Generalization via Opposing Symmetries
Charlotte Loh (Massachusetts Institute of Technology), Akash Srivastava (Amazon)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes a Multi-Symmetry Ensemble method (MSE), which trains invariant and equivariant networks for transformations such as rotation and color inversion separately in self-supervised contrastive learning, and then greedily combines models under different symmetries (i.e., different hypotheses) into an ensemble.
Multi-Task Differential Privacy Under Distribution Skew
Walid Krichene (Google Research), Li Zhang (Google Research)
Recommendation SystemSafty and PrivacyTabular
🎯 What it does: This study investigates the multi-task learning problem under user-level differential privacy, analyzes the impact of data distribution skew between tasks on model quality, and proposes a general algorithm based on adaptive re-weighted empirical loss.
Multi-task Hierarchical Adversarial Inverse Reinforcement Learning
Jiayu Chen (Purdue University), Vaneet Aggarwal (King Abdullah University of Science and Technology)
Robotic IntelligenceMeta LearningReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: This paper proposes MH-AIRL, a multi-task hierarchical adversarial inverse reinforcement learning framework based on unlabeled expert demonstrations, capable of learning hierarchical policies suitable for a class of tasks.
Multi-Task Off-Policy Learning from Bandit Feedback
Joey Hong (University of California), Mohammad Ghavamzadeh (Google Research)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: A multi-task offline (off-policy) optimization algorithm HierOPO based on a hierarchical Bayesian model is proposed, which estimates the parameters of each task and constructs lower confidence bounds by leveraging the similarity between tasks through hierarchical inference.
Multi-task Representation Learning for Pure Exploration in Linear Bandits
Yihan Du (Tsinghua University), Wen Sun (Cornell University)
Representation LearningReinforcement LearningTabular
🎯 What it does: This paper studies the multi-task pure exploration linear bandit problem, which aims to quickly identify the optimal arm (or strategy) by learning shared low-dimensional feature representations across multiple tasks. Two algorithms are proposed: DouExpDes (for RepBAI-LB) and C DouExpDes- (for RepBPI-CLB).
Multi-Task Structural Learning using Local Task Similarity induced Neuron Creation and Removal
Naresh Kumar Gurulingan (NavInfo Europe), Elahe Arani (NavInfo Europe)
SegmentationAutonomous DrivingConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A multi-task structure learning framework MTSL is proposed, which utilizes task local similarity to drive the creation and removal of neurons, dynamically learning the multi-task network structure and parameters.
Multi-User Reinforcement Learning with Low Rank Rewards
Dheeraj Mysore Nagaraj, Prateek Jain (Google Research)
Reinforcement Learning
🎯 What it does: An algorithm for multi-user collaborative reinforcement learning is designed, significantly reducing sample complexity under the low-rank assumption of the reward matrix.
Multi-View Masked World Models for Visual Robotic Manipulation
Younggyo Seo (KAIST), Pieter Abbeel (University of California Berkeley)
Robotic IntelligenceTransformerReinforcement LearningAuto EncoderWorld ModelVideo
🎯 What it does: This paper proposes a Multi-View Masking Autoencoder (MV-MAE) that jointly trains multi-view representations through view masking and video autoencoding, and then utilizes the frozen representations to learn a world model and perform behavior learning, addressing visual robotic manipulation tasks.
MultiAdam: Parameter-wise Scale-invariant Optimizer for Multiscale Training of Physics-informed Neural Networks
Jiachen Yao (Tsinghua University), Jun Zhu (Tsinghua University)
OptimizationTabularPhysics Related
🎯 What it does: This paper proposes a parameter-level scale-invariant optimizer, MultiAdam, for Physics-Informed Neural Networks (PINN), addressing the loss imbalance issue caused by domain scaling, and proves its effectiveness through theoretical error analysis.
Multicalibration as Boosting for Regression
Ira Globus-Harris (University of Pennsylvania), Jessica Sorrell (University of Pennsylvania)
OptimizationTabular
🎯 What it does: This study investigates the relationship between multi-calibration and boosting, proposing a definition of multi-calibration based on squared error and a 'swap-regret' style condition, and presents a simple boosting/multi-calibration algorithm that only requires a squared error regression oracle.
MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation
Omer Bar-Tal (Weizmann Institute of Science), Tali Dekel (Weizmann Institute of Science)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: Proposes the MultiDiffusion framework, which achieves various controllable generation, such as aspect ratio and region text guidance, by integrating multiple diffusion paths of a pre-trained diffusion model without the need for additional training.
Multiple Thinking Achieving Meta-Ability Decoupling for Object Navigation
Ronghao Dang (Tongji University), Qijun Chen (Tongji University)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes the Meta Ability Decoupling (MAD) paradigm and constructs a Multi-Thought (MT) model based on this paradigm to decouple and collaboratively work on meta abilities such as attention, search, navigation, exploration, and obstacle avoidance in object navigation tasks.
Multiplier Bootstrap-based Exploration
Runzhe Wan (Amazon), Rui Song (North Carolina State University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: A general exploration strategy MBE based on Multiplier Bootstrap is proposed to solve the multi-armed and contextual bandit problems with complex reward models, and near-optimal cumulative regret upper bounds are provided for both instance-related and instance-unrelated cases under sub-Gaussian distributions.
Multiply Robust Off-policy Evaluation and Learning under Truncation by Death
Jianing Chu (North Carolina State University), Wenbin Lu (North Carolina State University)
Biomedical DataElectronic Health Records
🎯 What it does: This paper proposes an offline policy evaluation (OPE) and offline policy learning (OPL) framework in the context of 'death truncation', and based on this, provides non-parametric identification of the survivor value function, efficiency limits, and multiple robust estimators, thereby achieving optimal decisions for the always-surviving population.
MultiRobustBench: Benchmarking Robustness Against Multiple Attacks
Sihui Dai (Princeton University), Prateek Mittal (Princeton University)
OptimizationAdversarial AttackConvolutional Neural NetworkImageBenchmark
🎯 What it does: A unified framework for multi-attack robustness assessment is proposed, and based on this, the MultiRobustBench benchmark is constructed, providing an evaluation system for 9 types of attacks and 20 levels of attack intensity (a total of 180 types of attacks).
Multisample Flow Matching: Straightening Flows with Minibatch Couplings
Aram-Alexandre Pooladian (Meta AI), Ricky T. Q. Chen (Meta AI)
GenerationData SynthesisOptimizationFlow-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes Multisample Flow Matching, a training method that utilizes the joint distribution of multiple samples to construct probability paths in continuous-time generative models, improving the path design of traditional Flow Matching.
Muse: Text-To-Image Generation via Masked Generative Transformers
Huiwen Chang (Google Research), Dilip Krishnan (Google Research)
GenerationData SynthesisTransformerLarge Language ModelImageText
🎯 What it does: Designed and trained Muse, a text-to-image generation model based on Masked Generative Transformer.
MyoDex: A Generalizable Prior for Dexterous Manipulation
Vittorio Caggiano (Meta AI), Vikash Kumar
Robotic IntelligenceReinforcement LearningMultimodality
🎯 What it does: Using multi-task reinforcement learning to learn various complex contact-rich hand manipulation behaviors on the simulated musculoskeletal hand model MyoHand, and to construct a transferable behavior prior MyoDex.
N$\text{A}^\text{2}$Q: Neural Attention Additive Model for Interpretable Multi-Agent Q-Learning
Zichuan Liu (Nanjing University), Chunlin Chen (Nanjing University)
Explainability and InterpretabilityReinforcement Learning
🎯 What it does: An interpretable multi-agent value decomposition method is proposed—Neural Attention Additive Q-learning (NAQ), which achieves high-order interaction decomposition of joint value and identity semantic-driven credit allocation.
Naive imputation implicitly regularizes high-dimensional linear models
Alexis Ayme (Sorbonne Université), Erwan Scornet (Ecole Polytechnique)
Tabular
🎯 What it does: This study investigates the impact of naive imputation using zero filling for missing values in high-dimensional linear regression on the model's implicit regularization, and provides an upper bound on the generalization error using average SGD training; it also validates the effectiveness of this method in the MCAR missing scenario through theoretical and simulation experiments.
Near-Minimax-Optimal Risk-Sensitive Reinforcement Learning with CVaR
Kaiwen Wang (Cornell University), Wen Sun (Cornell University)
Reinforcement Learning
🎯 What it does: This paper proposes and analyzes CVaR (Conditional Value at Risk) reinforcement learning in risk-sensitive contexts, providing optimal lower bounds under multi-armed bandit (MAB) and finite state Markov decision processes (MDP), and designs two corresponding optimal or near-optimal algorithms: BERNSTEIN-UCB (suitable for MAB) and CVaR-UCBVI (suitable for MDP).
Near-Optimal $\Phi$-Regret Learning in Extensive-Form Games
Ioannis Anagnostides (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: Designed and analyzed a decoupled learning dynamic, such that in multi-player incomplete information generalized form games, each player's triggered φ-regret grows with the number of rounds T at O(log T), achieving near-optimal convergence to EFCE.
Near-Optimal Algorithms for Private Online Optimization in the Realizable Regime
Hilal Asi (Tel Aviv University), Kunal Talwar (Tel Aviv University)
OptimizationSafty and Privacy
🎯 What it does: Proposed a differential privacy online optimization algorithm in a realizable setting, achieving an approximately optimal loss bound.
Near-optimal Conservative Exploration in Reinforcement Learning under Episode-wise Constraints
Donghao Li (Pennsylvania State University), Jing Yang (Pennsylvania State University)
Reinforcement Learning
🎯 What it does: In the finite state-action turn-based Markov decision process, an exploration strategy that satisfies the 'conservative' constraint for each turn is proposed, ensuring that the expected return during the learning process is always no less than the threshold of the baseline policy, and achieving the same order of learning return as in the unconstrained case.
Near-Optimal Cryptographic Hardness of Agnostically Learning Halfspaces and ReLU Regression under Gaussian Marginals
Ilias Diakonikolas (University of Wisconsin Madison), Lisheng Ren (University of Wisconsin Madison)
Gaussian Splatting
🎯 What it does: This paper constructs distinguishable instances from the sub-exponential difficulty of the Learning With Errors (LWE) problem and proves that under Gaussian distribution, agnostic learning for half-space and ReLU regression is nearly optimal in terms of computational complexity.
Near-Optimal Quantum Coreset Construction Algorithms for Clustering
Yecheng Xue (Peking University), Shaofeng H.-C. Jiang (Peking University)
OptimizationComputational Efficiency
🎯 What it does: The first quantum algorithm is proposed for constructing ε-coresets for k-median/k-means, reducing the query complexity to ilde{O}(√{nk} d^{3/2}/ε), achieving nearly quadratic speedup.
Nearly Minimax Optimal Regret for Learning Linear Mixture Stochastic Shortest Path
Qiwei Di (University of California), Quanquan Gu (University of California)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper proposes a learning algorithm (LEVIS++) for the stochastic shortest path (SSP) problem in linear mixed models, and provides an upper bound on regret that almost matches the lower bound O(d B* √K);
Nearly Minimax Optimal Reinforcement Learning for Linear Markov Decision Processes
Jiafan He (University of California), Quanquan Gu (University of California)
Computational EfficiencyReinforcement Learning
🎯 What it does: This paper studies reinforcement learning in linear Markov decision processes and proposes a computationally efficient algorithm that achieves an almost minimax optimal regret bound.
Nearly Optimal Algorithms with Sublinear Computational Complexity for Online Kernel Regression
Junfan Li (Tianjin University), Shizhong Liao (Tianjin University)
Optimization
🎯 What it does: This paper proposes two online kernel regression algorithms, AOGD-ALD and NONS-ALD, which achieve near-optimal regret bounds with sublinear computational complexity, and provides theoretical analysis and implementation details of the algorithms.
Nearly Optimal Competitive Ratio for Online Allocation Problems with Two-sided Resource Constraints and Finite Requests
Qixin Zhang (City University of Hong Kong), Yu Yang (City University of Hong Kong)
Optimization
🎯 What it does: The study investigates the online allocation problem with upper and lower bound constraints and proposes a series of nearly optimal competitive ratio algorithms.
Nearly-Linear Time and Streaming Algorithms for Outlier-Robust PCA
Ilias Diakonikolas (University of Wisconsin Madison), Thanasis Pittas (University of Wisconsin Madison)
Anomaly DetectionOptimization
🎯 What it does: A robust principal component analysis (PCA) algorithm with near-linear time complexity has been designed and implemented, and a single-channel streaming implementation has been proposed, which can recover the main variance direction of the data in the presence of a small number of outliers.
Nearly-Optimal Hierarchical Clustering for Well-Clustered Graphs
Steinar Laenen (University of Edinburgh), He Sun (University of Edinburgh)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes two nearly linear time hierarchical clustering algorithms that can produce hierarchical trees with costs comparable to the optimal Dasgupta cost for graphs with clear clustering structures, theoretically achieving O(1) approximation while maintaining almost linear time complexity.
Nearly-tight Bounds for Deep Kernel Learning
Yifan Zhang, Min-Ling Zhang (Southeast University)
🎯 What it does: This paper proposes almost tight capacity-based generalization bounds for Deep Kernel Learning (DKL) and Deep Multi-Kernel Learning (DMKL), and provides a theoretical analysis of deep kernel machines.
NerfDiff: Single-image View Synthesis with NeRF-guided Distillation from 3D-aware Diffusion
Jiatao Gu (Apple), Ravi Ramamoorthi (University of California, San Diego)
GenerationData SynthesisKnowledge DistillationDiffusion modelNeural Radiance FieldImage
🎯 What it does: Proposes the NerfDiff method, which utilizes a 3D-aware conditional diffusion model for view synthesis from a single input image;
NeRFool: Uncovering the Vulnerability of Generalizable Neural Radiance Fields against Adversarial Perturbations
Yonggan Fu (Georgia Institute of Technology), Celine Lin
Adversarial AttackNeural Radiance FieldImage
🎯 What it does: The NeRFool framework is proposed to systematically analyze the adversarial robustness of GNeRF, providing an attackable NeRFool+ method and designing a defense scheme based on adversarial training.
Nested Elimination: A Simple Algorithm for Best-Item Identification From Choice-Based Feedback
Junwen Yang (National University of Singapore), Yifan Feng (National University of Singapore)
Recommendation SystemOptimizationComputational EfficiencyTabular
🎯 What it does: An algorithm based on Nested Elimination (NE) is proposed for efficiently identifying the best products from choice-based feedback.
Nesterov Meets Optimism: Rate-Optimal Separable Minimax Optimization
Chris Junchi Li (University of California), Michael Jordan
Optimization
🎯 What it does: A new single-call first-order optimization algorithm AG-OG is designed, providing accelerated convergence rates for separable convex-concave minimax problems.
Network Effects in Performative Prediction Games
Xiaolu Wang (Chinese University of Hong Kong), Hoi To Wai
TabularFinance Related
🎯 What it does: This paper studies the existence and uniqueness of stable equilibria and Nash equilibria in multi-agent performative prediction (Multi-PP) games on multilayer networks, providing closed-form solutions, numerical examples, and convergence analysis.
Neural Algorithmic Reasoning with Causal Regularisation
Beatrice Bevilacqua (Purdue University), Petar Veličković (DeepMind)
Graph Neural NetworkContrastive LearningBenchmark
🎯 What it does: A self-supervised method based on causal regularization is proposed, utilizing the property that different inputs can maintain the same computation at a certain step during the algorithm execution process. By achieving indistinguishability of input subset variations through data augmentation and contrastive learning, the generalization ability of neural algorithm inference models under OOD conditions is enhanced.
Neural Collapse in Deep Linear Networks: From Balanced to Imbalanced Data
Hien Dang (FPT Software AI Center), Tan Minh Nguyen
OptimizationImage
🎯 What it does: This paper provides a theoretical analysis of the global optimal solution for deep linear networks under the Unconstrained Features Model (UFM) using Mean Squared Error (MSE) and Cross Entropy (CE) loss. It proves that the global optimal solution of any deep network possesses the property of Neural Collapse (NC) and extends this analysis to the case of imbalanced samples for the first time, resulting in the geometric structure of the General Orthogonal Frame (GOF). Subsequently, the theoretical conclusions are experimentally validated under various settings of network depth, width, and the presence or absence of bias.
Neural Continuous-Discrete State Space Models for Irregularly-Sampled Time Series
Abdul Fatir Ansari (Amazon Web Services), Harold Soh (National University of Singapore)
Data SynthesisAnomaly DetectionRepresentation LearningAuto EncoderVideoTime SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a continuous-discrete state space model (NCDSSM) capable of handling irregularly sampled time series. It utilizes auxiliary variables to decouple representation learning from dynamics, and achieves precise state estimation and uncertainty quantification through classical continuous-discrete Bayesian inference.
Neural Diffusion Processes
Vincent Dutordoir (University of Cambridge), Fergus Simpson (Secondmind)
GenerationOptimizationMeta LearningDiffusion modelImageMultimodality
🎯 What it does: Proposed the Neural Diffusion Processes (NDP), a generative model for diffusion on finite margins of functions;
Neural FIM for learning Fisher information metrics from point cloud data
Oluwadamilola Fasina (Yale University), Smita Krishnaswamy (Yale University)
OptimizationExplainability and InterpretabilityDiffusion modelPoint CloudOrdinary Differential Equation
🎯 What it does: This paper proposes a method called Neural FIM, which utilizes neural networks to learn Fisher information metrics from discrete point cloud data, thereby constructing a continuous Riemannian measure space that can be used to compute geometric quantities such as volume and geodesics. It applies this method to PHATE parameter selection, single-cell data visualization, and trajectory inference.
Neural Inverse Operators for Solving PDE Inverse Problems
Roberto Molinaro (ETH Zurich), Siddhartha Mishra (ETH Zurich)
TabularPhysics Related
🎯 What it does: A new neural network architecture called Neural Inverse Operators (NIO) is proposed for directly learning the mapping of PDE inverse problems.
Neural Latent Aligner: Cross-trial Alignment for Learning Representations of Complex, Naturalistic Neural Data
Cheol Jun Cho (University of California), Gopala Anumanchipalli
Representation LearningTransformerAuto EncoderContrastive LearningTime SeriesSequentialAudio
🎯 What it does: An unsupervised Neural Latent Aligner (NLA) framework is proposed, which learns behavior-related representations of natural speech ECoG using cross-trial alignment and a differentiable time-warping model.
Neural Markov Jump Processes
Patrick Seifner (University of Bonn), Ramses J Sanchez
Time SeriesSequentialOrdinary Differential Equation
🎯 What it does: A Markov Jump Process (MJP) model called NeuralMJP is proposed, which is based on neural variational inference and neural ordinary differential equations for posterior inference and parameter estimation of discrete-time observation sequences.
Neural Network Accelerated Implicit Filtering: Integrating Neural Network Surrogates With Provably Convergent Derivative Free Optimization Methods
Brian Irwin (University of British Columbia), Avi Ziv (IBM Research)
OptimizationTabularBenchmark
🎯 What it does: This paper studies a zero-order optimization method called NNAIF that uses neural networks to accelerate implicit filtering, enabling rapid convergence to critical points on noisy black-box functions.
Neural Network Approximations of PDEs Beyond Linearity: A Representational Perspective
Tanya Marwah (Carnegie Mellon University), Andrej Risteski (Carnegie Mellon University)
OptimizationPhysics Related
🎯 What it does: This paper studies the representational capacity of deep neural networks in approximating solutions to nonlinear variational partial differential equations (PDEs). It proves that under certain growth conditions of the Barron norm, the Barron norm of the solution can be controlled, allowing for polynomial-scale approximation using multilayer networks.
Neural networks trained with SGD learn distributions of increasing complexity
Maria Refinetti (Laboratoire de Physique de l'Ecole Normale Supérieure), Sebastian Goldt (International School of Advanced Studies)
ClassificationConvolutional Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes and validates the Distribution Simplicity Bias (DSB): in neural networks trained using stochastic gradient descent, the network first classifies using low-order statistics (mean, covariance) of the input distribution, and then gradually introduces higher-order statistics until optimal classification is achieved.
Neural Prediction Errors enable Analogical Visual Reasoning in Human Standard Intelligence Tests
Lingxiao Yang (Sun Yat-sen University), Ru-Yuan Zhang (Shanghai Jiao Tong University)
ClassificationRecognitionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Predictive Reasoning Block (PRB) integrated into a residual network called PredRNet, aimed at solving abstract visual reasoning problems such as Raven Progressive Matrices (RPM).
Neural signature kernels as infinite-width-depth-limits of controlled ResNets
Nicola Muca Cirone (Imperial College London), Cristopher Salvi
Time SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies controlled residual networks (ResNets) with random initialization, viewing them as the Euler discretization of neural controlled differential equations (Neural CDEs), and demonstrates that these architectures converge to Gaussian processes that satisfy specific partial differential equations (PDEs) in the limits of infinite width and depth.
Neural Status Registers
Lukas Faber (ETH Zurich), Roger Wattenhofer (ETH Zurich)
Convolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkImageGraphTabularSequential
🎯 What it does: A differentiable neural state register (NSR) is proposed to address numerical comparison problems;
Neural Stochastic Differential Games for Time-series Analysis
Sungwoo Park (LG AI Research), Changhee Lee (Chung-Ang University)
Time SeriesStochastic Differential Equation
🎯 What it does: This paper proposes a time series prediction framework based on Multi-Agent Stochastic Differential Equations (MaSDE), utilizing cooperative differential games to achieve information aggregation and collaborative decision-making.
Neural Wasserstein Gradient Flows for Discrepancies with Riesz Kernels
Fabian Altekrüger (Humboldt-Universität zu Berlin), Gabriele Steidl (Technische Universität Berlin)
OptimizationFlow-based ModelImage
🎯 What it does: This paper proposes using neural networks to approximate the Wasserstein gradient flow and the Steepest Descent flow with backward (JKO) and forward Euler discretization schemes, addressing non-absolutely continuous Riesz kernel MMD functions, including singular (Dirac) measures; it also provides an analytical closed form and convergence proof for the interaction energy at the Dirac initial point; the effectiveness of the neural method is validated through numerical experiments.
Neural Wave Machines: Learning Spatiotemporally Structured Representations with Locally Coupled Oscillatory Recurrent Neural Networks
T. Anderson Keller (University of Amsterdam), Max Welling (University of Amsterdam)
ClassificationRepresentation LearningRecurrent Neural NetworkTime SeriesSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper presents the Neural Wave Machine (NWM), a recurrent neural network that implements trainable traveling waves through locally coupled oscillators, and demonstrates its ability to learn spatial and temporal structured representations.
NeuralSlice: Neural 3D Triangle Mesh Reconstruction via Slicing 4D Tetrahedral Meshes
Chenbo Jiang (Chinese Academy of Sciences), Lin Gao (Chinese Academy of Sciences)
GenerationData SynthesisPoint CloudMesh
🎯 What it does: The NEURALSLICE method is proposed, which achieves high-fidelity and topologically flexible 3D mesh reconstruction by representing 3D shapes as the intersection of a 4D tetrahedral mesh and a 4D hyperplane.
NeuralStagger: Accelerating Physics-constrained Neural PDE Solver with Spatial-temporal Decomposition
Xinquan Huang (King Abdullah University of Science and Technology), Tie-Yan Liu (Microsoft Research AI4Science)
OptimizationComputational EfficiencyConvolutional Neural NetworkTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A method called NeuralStagger is proposed, which significantly accelerates the inference of neural PDE solvers by decomposing the physical field into several coarse-resolution sub-tasks in both space and time, and training them in parallel using lightweight networks while maintaining physical constraint training.
Neuro-Symbolic Continual Learning: Knowledge, Reasoning Shortcuts and Concept Rehearsal
Emanuele Marconato (University of Trento), Stefano Teso (University of Trento)
Knowledge DistillationReinforcement LearningSequential
🎯 What it does: A framework for Neural Symbolic Continual Learning (NeSy-CL) is proposed, along with a novel concept-level continual learning strategy called COOL, aimed at maintaining high-quality concepts and avoiding reasoning shortcuts in multi-task settings.
Never mind the metrics---what about the uncertainty? Visualising binary confusion matrix metric distributions to put performance in perspective
David Lovell (Queensland University of Technology), Andrew P. Bradley (University of the Sunshine Coast)
Classification
🎯 What it does: This paper visualizes and quantitatively analyzes the uncertainty of performance metrics (such as MCC, BA, F1, etc.) for binary classifiers by constructing a three-dimensional discrete confusion matrix space (confusion simplex) and its slices on the ROC plane, in conjunction with the Bayesian beta-binomial model.
New metrics and search algorithms for weighted causal DAGs
Davin Choo (National University of Singapore), Kirankumar Shiragur (Broad Institute of MIT and Harvard)
Graph
🎯 What it does: This paper studies the problem of achieving complete identification of causal graphs (DAGs) using adaptive interventions under the condition of different intervention costs at the vertices, and proposes a new evaluation metric and corresponding adaptive search algorithm.
NNSplitter: An Active Defense Solution for DNN Model via Automated Weight Obfuscation
Tong Zhou (Northeastern University), Xiaolin Xu (Northeastern University)
ClassificationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: The NNSplitter scheme is proposed, which splits the pre-trained model into a non-functional ordinary model and a complete model accessible only to authorized users through automated weight obfuscation and recoverable model secrets, thereby achieving proactive model IP protection.
No One Idles: Efficient Heterogeneous Federated Learning with Parallel Edge and Server Computation
Feilong Zhang (Harbin Institute of Technology), Xiangyang Ji (Tsinghua University)
Federated LearningComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: By exchanging the execution order of server-side aggregation and broadcasting, the server and edge devices can work in parallel, eliminating idle time caused by waiting on the server and devices in traditional synchronous federated learning (FL), thus achieving efficient federated learning under heterogeneous devices.
Node Embedding from Neural Hamiltonian Orbits in Graph Neural Networks
Qiyu Kang (Nanyang Technological University), Wee Peng Tay (Nanyang Technological University)
ClassificationRepresentation LearningGraph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: Proposes the use of Hamiltonian orbits to evolve node features in the graph embedding space, achieving adaptive geometric learning.
Non-autoregressive Conditional Diffusion Models for Time Series Prediction
Lifeng Shen (Hong Kong University of Science and Technology), James Kwok
Convolutional Neural NetworkDiffusion modelTime Series
🎯 What it does: A non-autoregressive conditional diffusion model called TimeDiff is proposed for time series forecasting.
Non-stationary Reinforcement Learning under General Function Approximation
Songtao Feng (Ohio State University), Yingbin Liang (Ohio State University)
Reinforcement Learning
🎯 What it does: A theoretical framework for using universal function approximation in non-stationary Markov Decision Processes (MDP) is proposed, along with a corresponding Dynamic Bellman Exploration Dimension (DBE) metric;