NeurIPS 2023 Papers with AI Summaries
Conference on Neural Information Processing Systems · 3218 papers
→ NeurIPS 2023 papers with code (1376)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 3218 NeurIPS 2023 papers by keyword, author or institution —
free trial on arXivSub.
“Why Not Looking backward?” A Robust Two-Step Method to Automatically Terminate Bayesian Optimization
Shuang Li (Harbin Institute of Technology), Wei Li (Harbin Institute of Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: A two-step automatic termination Bayesian optimization method is proposed, which is based on detecting the search entering a local convex region and using a local reward threshold as a criterion.
(Almost) Provable Error Bounds Under Distribution Shift via Disagreement Discrepancy
Elan Rosenfeld (Carnegie Mellon University), Saurabh Garg (Carnegie Mellon University)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A method is proposed based on Disagreement Discrepancy to provide an error upper bound, utilizing unlabeled test samples to (almost) provably estimate the error of deep networks under distribution shifts.
(Amplified) Banded Matrix Factorization: A unified approach to private training
Christopher A. Choquette-Choo (Google DeepMind), Zheng Xu (Google Research)
Federated LearningSafty and PrivacyImageText
🎯 What it does: A differential privacy mechanism based on Band Matrix Factorization (BANDMF) is proposed, which can efficiently implement privacy-preserving training in both centralized and cross-device federated learning environments.
(S)GD over Diagonal Linear Networks: Implicit bias, Large Stepsizes and Edge of Stability
Mathieu Even (Inria), Nicolas Flammarion (EPFL)
OptimizationTabularStochastic Differential Equation
🎯 What it does: The study investigates the implicit regularization and convergence behavior of stochastic gradient descent (SGD) and ordinary gradient descent (GD) on two-dimensional diagonal linear networks under large step sizes and the influence of random noise.
$\textbf{A}^2\textbf{CiD}^2$: Accelerating Asynchronous Communication in Decentralized Deep Learning
Adel Nabli (Concordia University), Edouard Oyallon (Sorbonne University)
Convolutional Neural NetworkImageStochastic Differential Equation
🎯 What it does: A continuous momentum-based asynchronous decentralized deep learning training algorithm A CiD 2 2 was researched and implemented, significantly accelerating communication between nodes;
$\texttt{TACO}$: Temporal Latent Action-Driven Contrastive Loss for Visual Reinforcement Learning
Ruijie Zheng (University of Maryland), Furong Huang (University of Maryland)
Robotic IntelligenceReinforcement LearningContrastive LearningImage
🎯 What it does: A temporal contrastive learning framework named TACO is proposed, which can simultaneously learn latent representations of states and actions and apply them to visual continuous control in reinforcement learning.
$\varepsilon$-fractional core stability in Hedonic Games.
Simone Fioravanti (Gran Sasso Science Institute), Giovanna Varricchio (University of Calabria)
🎯 What it does: This paper studies the concept of ε-fractional core stability in Hedonic Games and provides results on existence and polynomial-time construction in Simple Fractional and Anonymous HGs.
$H$-Consistency Bounds: Characterization and Extensions
Anqi Mao (Courant Institute of Mathematical Sciences), Yutao Zhong (Courant Institute of Mathematical Sciences)
ClassificationOptimization
🎯 What it does: A general error transformation function is proposed to derive the H-consistency upper bounds for any hypothesis set in multi-class classification, covering two common surrogate loss types: comp-sum and constrained, and extending previous limitations on complete hypothesis sets to bounded hypothesis sets.
$k$-Means Clustering with Distance-Based Privacy
Alessandro Epasto (Google Research), Peilin Zhong (Google Research)
OptimizationSafty and PrivacyTabular
🎯 What it does: This work proposes distance privacy-based k-means and k-median clustering algorithms, and presents a polynomial time (ε, δ, ρ)-dist-DP algorithm;
$L_2$-Uniform Stability of Randomized Learning Algorithms: Sharper Generalization Bounds and Confidence Boosting
Xiaotong Yuan, Ping Li (VecML Inc.)
🎯 What it does: This paper proposes the concept of L₂-uniform stability for stochastic learning algorithms and, based on this, proves the consistency error upper bound for the first moment; subsequently, it presents a bagging-based confidence boosting method that ensures this upper bound holds with high probability regarding both sample and algorithm randomness; the theory is applied to SGD with a decaying learning rate, resulting in tighter high-probability generalization error and risk upper bounds for convex, non-convex, and smooth losses.
$p$-Poisson surface reconstruction in curl-free flow from point clouds
Yesom Park (Seoul National University), Myungjoo Kang (Seoul National University)
Point CloudBenchmark
🎯 What it does: Reconstructing closed surfaces from unordered point clouds through implicit neural representation, resulting in smooth and detail-rich surface geometry.
$p$-value Adjustment for Monotonous, Unbiased, and Fast Clustering Comparison
Kai Klede (Friedrich-Alexander Universität Erlangen-Nürnberg), Bjoern Eskofier (Helmholtz Zentrum München)
ImageTabular
🎯 What it does: Proposed and implemented the p-value adjusted Rand index (PMI₂), addressing type II bias in clustering comparison and ensuring monotonicity; also provided two efficient computation schemes: standardized approximation and Monte Carlo approximation, and validated its superiority on various datasets.
$S^3$: Increasing GPU Utilization during Generative Inference for Higher Throughput
Yunho Jin (Harvard University), Gu-Yeon Wei (Harvard University)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The S3 system is proposed, which utilizes predicted output sequence lengths and performs dynamic scheduling based on predictions to maximize GPU resource utilization and improve inference throughput.
$SE(3)$ Equivariant Convolution and Transformer in Ray Space
Yinshuang Xu (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
RestorationGenerationConvolutional Neural NetworkTransformerPoint Cloud
🎯 What it does: Learning geometry priors based on multi-view perspectives, proposing SE(3)-equivariant convolution and Transformer in ray space (light field) for 3D reconstruction and neural rendering.
2Direction: Theoretically Faster Distributed Training with Bidirectional Communication Compression
Alexander Tyurin (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated Learning
🎯 What it does: A new acceleration method called 2Direction is proposed to address the communication bottleneck in distributed convex optimization problems, particularly in the bidirectional communication between servers and worker nodes.
3D Copy-Paste: Physically Plausible Object Insertion for Monocular 3D Detection
Yunhao Ge (Stanford University), Jiajun Wu (Stanford University)
Object DetectionData SynthesisPoint Cloud
🎯 What it does: This paper proposes a 3D Copy-Paste technique that can insert virtual 3D objects into real indoor scenes in a physically feasible manner, generating annotated data with real positions, sizes, poses, and lighting;
3D Indoor Instance Segmentation in an Open-World
Mohamed El Amine Boudjoghra (Mohamed Bin Zayed University of Artificial Intelligence), Fahad Khan
Object DetectionSegmentationTransformerContrastive LearningPoint Cloud
🎯 What it does: The first open-world indoor 3D instance segmentation framework, 3D-OWIS, is proposed, capable of recognizing known categories and unknown objects during inference and achieving incremental learning upon obtaining labels.
3D molecule generation by denoising voxel grids
Pedro O. Pinheiro (Prescient Design), Saeed Saremi (Prescient Design)
GenerationData SynthesisDrug DiscoveryDiffusion modelScore-based ModelPoint Cloud
🎯 What it does: A molecular generation method based on 3D voxel grids, VoxMol, has been developed, using score-based single-step denoising and Langevin walk-jump sampling to generate 3D molecules.
3D-Aware Visual Question Answering about Parts, Poses and Occlusions
Xingrui Wang (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
Object DetectionPose EstimationNeural Radiance FieldImageMesh
🎯 What it does: Proposes the 3D-aware VQA task, constructs the Super-CLEVR-3D dataset, and designs the PO3D-VQA model for 3D structure reasoning and answer generation.
3D-IntPhys: Towards More Generalized 3D-grounded Visual Intuitive Physics under Challenging Scenes
Haotian Xue (Georgia Institute of Technology), Hsiao-Yu Tung (Massachusetts Institute of Technology)
Object DetectionSegmentationData SynthesisGraph Neural NetworkNeural Radiance FieldVideoPoint CloudPhysics Related
🎯 What it does: The 3D-IntPhys framework is proposed to unsupervisedly reconstruct 3D point clouds from multi-view RGB images and instance masks, and to predict the physical dynamics of point clouds using graph neural networks, achieving long-term predictions for complex materials such as fluids, particles, and rigid bodies.
3D-LLM: Injecting the 3D World into Large Language Models
Yining Hong (University of California), Chuang Gan (Massachusetts Institute of Technology)
Object DetectionGenerationRetrievalTransformerLarge Language ModelVision Language ModelTextPoint Cloud
🎯 What it does: A novel 3D-LLM is proposed, capable of inputting 3D point clouds and their features along with language prompts to perform various 3D-related tasks;
4D Panoptic Scene Graph Generation
Jingkang Yang (Nanyang Technological University), Ziwei Liu (Hong Kong Baptist University)
SegmentationGenerationRobotic IntelligenceTransformerLarge Language ModelVideoPoint Cloud
🎯 What it does: This paper proposes the 4D Panoramic Scene Graph (PSG-4D) task and constructs a panoramic segmentation and dynamic relationship joint generation framework called PSG4DFormer, which includes RGB-D videos and point cloud sequences, and demonstrates collaborative decision-making with large language models in real service robots.
4M: Massively Multimodal Masked Modeling
David Mizrahi (Swiss Federal Institute of Technology Lausanne), Amir Zamir (Swiss Federal Institute of Technology Lausanne)
Object DetectionSegmentationGenerationDepth EstimationTransformerImageTextMultimodality
🎯 What it does: A unified Transformer framework called 4M is proposed, which utilizes multi-modal masking modeling for pre-training on different input/output modalities (text, images, geometry, semantics, feature maps, etc.) to obtain a foundational model for multi-task and multi-modal applications.
A Batch-to-Online Transformation under Random-Order Model
Jing Dong (Chinese University of Hong Kong), Yuichi Yoshida (National Institute of Informatics)
OptimizationTabular
🎯 What it does: A batch-to-online conversion framework is proposed, utilizing average sensitivity and core sample construction to transform offline approximation algorithms into online algorithms under the random order model, achieving polynomial logarithmic ε-approximate regret in problems such as k-means clustering, low-rank matrix approximation, and linear regression.
A Bayesian Approach To Analysing Training Data Attribution In Deep Learning
Elisa Nguyen (Tübingen AI Center University of Tübingen), Seong Joon Oh (Tübingen AI Center University of Tübingen)
Convolutional Neural NetworkTransformerImage
🎯 What it does: A reliability assessment of Training Data Attribution (TDA) is conducted from a Bayesian perspective, quantifying its noise and variance, and comparing various approximation methods.
A Bayesian Take on Gaussian Process Networks
Enrico Giudice (University of Basel), Giusi Moffa (University College London)
Gaussian SplattingTabular
🎯 What it does: This paper proposes a method for complete Bayesian structure learning on Gaussian Process Networks, utilizing MCMC sampling, Laplace approximation, and importance sampling to achieve computable posterior distributions.
A Bounded Ability Estimation for Computerized Adaptive Testing
Yan Zhuang (University of Science and Technology of China), Xin Li (State Key Laboratory of Cognitive Intelligence)
Reinforcement LearningTabular
🎯 What it does: A computer adaptive testing framework based on definable ability estimation (BECAT) is proposed, which approximates students' true abilities as estimated values under a complete item bank, and implements explicit ability estimation through a greedy selection algorithm based on expected gradient difference approximation.
A case for reframing automated medical image classification as segmentation
Sarah Hooper, Christopher Re
ClassificationSegmentationConvolutional Neural NetworkImageBiomedical DataComputed Tomography
🎯 What it does: This paper explores reframing the classification task of medical imaging as a segmentation task and proposes a method for classification through segmentation networks (segmentation-for-classification).
A Causal Framework for Decomposing Spurious Variations
Drago Plecko (Columbia University), Elias Bareinboim (Columbia University)
Tabular
🎯 What it does: This paper proposes a new tool based on the 'Partially Abducted Submodel' for non-parametrically decomposing spurious variation in causal models, and provides theoretical decomposition formulas and identifiability conditions under Markovian and Semi-Markovian structures.
A Closer Look at the Robustness of Contrastive Language-Image Pre-Training (CLIP)
Weijie Tu (Australian National University), Tom Gedeon (Curtin University)
Anomaly DetectionTransformerContrastive LearningImageMultimodality
🎯 What it does: This study systematically evaluates the performance of 83 CLIP models and 127 traditional ImageNet models on safety attributes such as robustness to visual factors, anomaly detection, and prediction uncertainty.
A Combinatorial Algorithm for Approximating the Optimal Transport in the Parallel and MPC Settings
Nathaniel Lahn (Radford University), Kaiyi Zhang (Virginia Tech)
OptimizationImageTextBenchmark
🎯 What it does: A parallel combinatorial algorithm has been designed and implemented for ε-approximate solutions to the optimal transport (OT) problem and assignment problem in parallel and MPC environments, providing a GPU-friendly matrix implementation.
A Competitive Algorithm for Agnostic Active Learning
Yihan Zhou (University of Texas at Austin), Eric Price (University of Texas at Austin)
Reinforcement Learning
🎯 What it does: A competitive algorithm is proposed to solve the unbiased active learning problem, which can compete with the optimal algorithm under any binary hypothesis class and distribution.
A Computation and Communication Efficient Method for Distributed Nonconvex Problems in the Partial Participation Setting
Alexander Tyurin (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated LearningComputational Efficiency
🎯 What it does: A novel distributed non-convex optimization algorithm DASHA-PP is proposed, which can simultaneously achieve variance reduction, compressed communication, and partial participation.
A Computationally Efficient Sparsified Online Newton Method
Fnu Devvrit, Inderjit S Dhillon
OptimizationComputational EfficiencyTransformerAuto EncoderImageGraph
🎯 What it does: A sparse online Newton method (SONew) is proposed, achieving linear time/space second-order preconditioners through LogDet divergence and structured sparse graphs;
A Cross-Moment Approach for Causal Effect Estimation
Yaroslav Kivva (Ecole Polytechnique Federale de Lausanne), Negar Kiyavash (Ecole Polytechnique Federale de Lausanne)
Tabular
🎯 What it does: In a linear structural causal model (SCM) with only a single proxy variable and potential confounding, a method based on three-variable cross-moments is proposed to estimate the causal effect of the treatment variable on the outcome variable.
A Data-Free Approach to Mitigate Catastrophic Forgetting in Federated Class Incremental Learning for Vision Tasks
Sara Babakniya (University of Southern California), Salman Avestimehr (University of Southern California)
Federated LearningKnowledge DistillationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: The MFCL framework is proposed to address the catastrophic forgetting problem in class-incremental learning under federated learning. The framework generates models through data-free knowledge distillation training on the server side, while clients train using the generated synthetic samples along with local real data after each task, thus avoiding the need to store old data on the client side.
A Deep Instance Generative Framework for MILP Solvers Under Limited Data Availability
Zijie Geng (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationGraph Neural NetworkAuto EncoderTabular
🎯 What it does: The G2MILP framework is proposed, which uses deep generative models to generate new mixed-integer linear programming instances under limited data conditions.
A Definition of Continual Reinforcement Learning
David Abel (Google DeepMind), Satinder Singh (Google DeepMind)
Reinforcement LearningSequential
🎯 What it does: A formal mathematical definition of Continuous Reinforcement Learning (CRL) is proposed, introducing the operators of generates and reaches to characterize the implicit search behavior of agents based on a foundation.
A Diffusion-Model of Joint Interactive Navigation
Matthew Niedoba (University of British Columbia), Frank Wood (University of British Columbia)
GenerationAutonomous DrivingTransformerDiffusion modelTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A diffusion model named DJINN is proposed for generating complete joint traffic scene trajectories, capable of producing realistic traffic scenes under any conditions (such as targets, behavior classes, or manual editing).
A Dual-Stream Neural Network Explains the Functional Segregation of Dorsal and Ventral Visual Pathways in Human Brains
Minkyu Choi (University of Michigan), Zhongming Liu (University of Michigan)
RecognitionObject DetectionConvolutional Neural NetworkRecurrent Neural NetworkImageVideoMagnetic Resonance Imaging
🎯 What it does: A dual-stream neural network inspired by the dual pathways (dorsal/ventral) of the human visual system is proposed and implemented. It utilizes retinal transformations to generate different field inputs, with WhereCNN learning spatial attention and eye movement control, and WhatCNN learning object recognition, recursively constructing scene representations through multiple fixations.
A Dynamical System View of Langevin-Based Non-Convex Sampling
Mohammad Reza Karimi Jaghargh, Andreas Krause (ETH Zurich)
OptimizationStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a unified framework for dynamical systems, proving that under non-convex objectives, a wide range of Langevin-Robbins-Monro sampling algorithms (including stochastic midpoint, mirror Langevin, Runge-Kutta, etc.) converge to the target distribution in Wasserstein distance, and the convergence is for the last iteration;
A Fast and Accurate Estimator for Large Scale Linear Model via Data Averaging
Rui Wang (Renmin University of China), Wangli Xu (Renmin University of China)
OptimizationComputational EfficiencyTabular
🎯 What it does: A sketching method based on data averaging is proposed for large-scale linear regression problems (New), achieving faster and more accurate parameter estimates with a computational complexity of O(Np + p³).
A fast heuristic to optimize time-space tradeoff for large models
Akifumi Imanishi (Preferred Networks), Emilio Castillo (Preferred Networks)
OptimizationComputational EfficiencyTransformerLarge Language ModelImage
🎯 What it does: A fast heuristic algorithm based on simulated annealing (FastSA) is proposed to generate activation recomputation plans on any computation graph, significantly reducing peak memory usage while keeping computational overhead acceptable.
A Finite-Particle Convergence Rate for Stein Variational Gradient Descent
Jiaxin Shi (Stanford University), Lester Mackey (Microsoft Research)
Optimization
🎯 What it does: Under the conditions of sub-Gaussian and gradient smoothness for the target distribution, a non-asymptotic convergence rate analysis of finite particle Stein Variational Gradient Descent (SVGD) is provided, proving that its kernel Stein discrepancy converges to 0 with the sequence of iteration steps, with a rate of 1/√log log n.
A Finite-Sample Analysis of Payoff-Based Independent Learning in Zero-Sum Stochastic Games
Zaiwei Chen (California Institute of Technology), Adam Wierman (Massachusetts Institute of Technology)
Reinforcement Learning
🎯 What it does: This study explores two-person zero-sum stochastic games and develops a variant of smooth best response learning dynamics that combines independent learning dynamics of matrix games with minimax iteration of stochastic games.
A Fractional Graph Laplacian Approach to Oversmoothing
Sohir Maskey (Ludwig Maximilian University of Munich), Gitta Kutyniok (Ludwig Maximilian University of Munich)
Graph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: A graph neural ODE framework based on fractional graph Laplacians is proposed to address the over-smoothing problem in directed graphs.
A Framework for Fast and Stable Representations of Multiparameter Persistent Homology Decompositions
David Loiseaux (Inria d'Université Côte d'Azur), Andrew Blumberg
ClassificationComputational EfficiencyRepresentation LearningPoint CloudGraphTime Series
🎯 What it does: This paper proposes a general framework T-CDR for representing multi-parameter persistent homology decompositions as vectors, and based on this, provides a provably stable S-CDR representation.
A General Framework for Equivariant Neural Networks on Reductive Lie Groups
Ilyes Batatia (University of Cambridge), Christoph Ortner (University of British Columbia)
RecognitionPoint Cloud
🎯 What it does: A general G-Equivariant Cluster Expansion and G-MACE architecture is proposed, which can construct equivariant neural networks for finite-dimensional representations of any simple Lie group, and the lie-nn library implementing this framework has been released.
A General Framework for Robust G-Invariance in G-Equivariant Networks
Sophia Sanborn (University of California Santa Barbara), Nina Miolane (University of California Santa Barbara)
ClassificationConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes a convolutional neural network based on Group Triple Correlation (G-TC) layers to achieve group invariance (G-invariance);
A General Theory of Correct, Incorrect, and Extrinsic Equivariance
Dian Wang (Northeastern University), Robin Walters (Northeastern University)
ClassificationRobotic IntelligenceConvolutional Neural NetworkTabular
🎯 What it does: This paper presents a general theory for analyzing the error limits when there is an imperfect match between real functions and equivariant network symmetry, and introduces the concepts of point-to-point correctness, incorrectness, and external equivariance; it then provides lower bounds on errors for classification, balanced regression, and equivariant regression tasks, and validates the theory through experiments.
A generative model of the hippocampal formation trained with theta driven local learning rules
Tom George, Tomoki Fukai (Okinawa Institute of Science and Technology)
GenerationData SynthesisAuto EncoderTime SeriesSequential
🎯 What it does: A biologically interpretable hippocampal generative model based on the Helmholtz machine has been established, utilizing theta rhythm to switch between forward (wake) and backward (sleep) information flow, and employing local synaptic rules for learning to infer and generate networks. This ultimately achieves autoencoding of random high-dimensional perceptual inputs, path integration of 1D trajectories, and structural transfer across environments.
A Graph-Theoretic Framework for Understanding Open-World Semi-Supervised Learning
Yiyou Sun (University of Wisconsin), Yixuan Li (University of Wisconsin)
Representation LearningGraph Neural NetworkContrastive LearningImage
🎯 What it does: A graph theory-based open-world semi-supervised learning framework is proposed, and the Spectral Open-world Representation Learning (SORL) method is designed;
A graphon-signal analysis of graph neural networks
Ron Levie (Mathematics Technion Israel Institute of Technology)
Graph Neural NetworkGraph
🎯 What it does: This paper extends spectral graph theory to the graph signal level, proposing the spectral-signal cut distance and proving the weak regularity, compactness, and sampling lemma of spectral signals. It further demonstrates the Lipschitz continuity of message-passing-based graph neural networks under this metric, providing generalization error bounds and subsampling stability bounds without distribution assumptions.
A Guide Through the Zoo of Biased SGD
Yury Demidovich (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationTabular
🎯 What it does: The convergence of Biased Stochastic Gradient Descent (BiasedSGD) with a biased gradient estimator is studied, and a unified hypothesis framework is constructed.
A Heat Diffusion Perspective on Geodesic Preserving Dimensionality Reduction
Guillaume Huguet (Université de Montréal), Smita Krishnaswamy (Yale University)
Diffusion modelTabularBiomedical Data
🎯 What it does: This paper proposes a direct association between thermal diffusion and geodesic distance on manifolds through the Varadhan formula, defining thermal geodesic similarity, and based on this, constructs a Heat Geo dimensionality reduction method;
A Heavy-Tailed Algebra for Probabilistic Programming
Feynman T. Liang (University of California), Michael W. Mahoney (University of California)
Flow-based Model
🎯 What it does: This paper proposes a heavy-tailed static analysis method for probabilistic programming languages, utilizing a heavy-tailed algebra based on the generalized gamma distribution (GGA) to perform closed operations on the tail behavior of random variables and infer tail parameters at compile time.
A Hierarchical Spatial Transformer for Massive Point Samples in Continuous Space
Wenchong He (University of Florida), Christine Angelini (University of Florida)
TransformerPoint Cloud
🎯 What it does: A Hierarchical Spatial Transformer (HST) is proposed for modeling and predicting up to millions of point samples in continuous space.
A Hierarchical Training Paradigm for Antibody Structure-sequence Co-design
Fang Wu (Tsinghua University), Stan Z. Li (Westlake University)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: A hierarchical training framework (HTP) is proposed for the joint design of antibody sequences and structures.
A Holistic Approach to Unifying Automatic Concept Extraction and Concept Importance Estimation
Thomas FEL, Thomas Serre (Brown University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a unified theoretical framework that relates concept extraction and concept importance assessment to dictionary learning and attribution methods. It further defines new evaluation metrics based on this framework and constructs a visualization tool (strategic cluster graph) using ImageNet data to reveal the model's classification strategies.
A Logic for Expressing Log-Precision Transformers
William Merrill (New York University), Ashish Sabharwal (Allen Institute for AI)
Transformer
🎯 What it does: This paper studies Transformers with log-precision and proves that they can be expressed as first-order logic with multiple quantifiers (FO M) sentences, providing an upper bound for this type of Transformer;
A Long $N$-step Surrogate Stage Reward for Deep Reinforcement Learning
Junmin Zhong (Arizona State University), Jennie Si (Arizona State University)
Reinforcement Learning
🎯 What it does: This paper proposes a Long N-Step Pseudo Stage Reward (LNSS) estimator for deep reinforcement learning to alleviate high variance issues and improve the convergence speed and final performance of continuous control tasks.
A Measure-Theoretic Axiomatisation of Causality
Junhyung Park (Max Planck Institute for Intelligent Systems), Krikamol Muandet (CISPA Helmholtz Center for Information Security)
Supervised Fine-Tuning
🎯 What it does: This paper constructs a causal space framework based on the combination of measure-theoretic probability spaces and transition probability kernels, proposing to directly encode causal information as 'causal kernels', and provides definitions, properties, and intervention operations;
A Metadata-Driven Approach to Understand Graph Neural Networks
Ting Wei Li (University of Michigan), Jiaqi Ma (University of Illinois Urbana-Champaign)
Graph Neural NetworkGraph
🎯 What it does: A metadata-based multivariate sparse regression method is proposed to identify key graph data attributes that affect GNN performance, and the impact of degree distribution on GNN is verified through theoretical analysis and controlled experiments.
A Neural Collapse Perspective on Feature Evolution in Graph Neural Networks
Vignesh Kothapalli (New York University), Joan Bruna (New York University)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the feature evolution of Graph Neural Networks (GNN) in node-level classification tasks and quantifies it from the perspective of Neural Collapse (NC).
A new perspective on building efficient and expressive 3D equivariant graph neural networks
weitao Du, Zhi-Ming Ma (Chinese Academy of Sciences)
Computational EfficiencyRepresentation LearningDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes a new local-global hierarchical 3D graph isomorphism framework to evaluate and enhance the expressive power of 3D equivariant graph neural networks, and based on this, designs two main modules: LSE (Local Substructure Encoding) and FTE (Frame Transition Encoding), ultimately realizing the LEFTNet model.
A normative theory of social conflict
Sergey A. Shuvaev (Institute of Cytology and Genetics Siberian Branch of Russian Academy of Sciences), Alexei A. Koulakov (Cold Spring Harbor Laboratory)
Reinforcement LearningTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: By combining game theory with partially observable Markov decision processes, a first-order theory of mind (1-ToM) model based on Bayesian inference was constructed to simulate the attack and defense decisions of mice in chronic social conflict experiments, and neural circuits related to the model's hidden variables were detected using whole-brain c-Fos imaging data.
A Novel Approach for Effective Multi-View Clustering with Information-Theoretic Perspective
Chenhang Cui (University of Electronic Science and Technology of China), Lifang He (Lehigh University)
Multimodality
🎯 What it does: This paper proposes a new method called SUMVC based on information bottleneck and variational inference for multi-view clustering.
A Novel Framework for Policy Mirror Descent with General Parameterization and Linear Convergence
Carlo Alfano (University of Oxford), Patrick Rebeschini (University of Oxford)
OptimizationReinforcement LearningTabular
🎯 What it does: A general strategy optimization framework AMPO based on mirror descent is proposed, which can handle arbitrary parameterization and mirror mappings, and is theoretically proven to have linear convergence.
A One-Size-Fits-All Approach to Improving Randomness in Paper Assignment
Yixuan Even Xu (Tsinghua University), Fei Fang (Carnegie Mellon University)
OptimizationText
🎯 What it does: A general random paper allocation method (Perturbed Maximization, PM) is proposed, which significantly enhances randomness while maintaining high matching quality through a concave perturbation function.
A Partially-Supervised Reinforcement Learning Framework for Visual Active Search
Anindya Sarkar (Washington University in St Louis), Yevgeniy Vorobeychik (Washington University in St Louis)
Object DetectionMeta LearningConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A partially supervised reinforcement learning framework PSVAS is proposed, which divides visual active search into a prediction module and a search module, and combines meta-learning to achieve task adaptation.
A Path to Simpler Models Starts With Noise
Lesia Semenova (Duke University), Cynthia Rudin (Duke University)
ClassificationOptimizationTabular
🎯 What it does: This paper proposes and proves how noise leads to an increase in the Rashomon set, thereby explaining why simple models often achieve similar or close accuracy to complex models on high-noise datasets.
A polar prediction model for learning to represent visual transformations
Pierre-Etienne H Fiquet, Eero P Simoncelli
OptimizationRepresentation LearningConvolutional Neural NetworkVideo
🎯 What it does: A polarization prediction model is constructed through self-supervised learning, achieving frame-level accurate predictions by utilizing transformation rules such as translation and rotation in natural videos.
A Privacy-Friendly Approach to Data Valuation
Jiachen T. Wang (Princeton University), Prateek Mittal (Princeton University)
Safty and PrivacyComputational EfficiencyTextTabularBenchmark
🎯 What it does: This paper proposes TKNN-Shapley, an improved and more computationally efficient and privacy-friendly data value assessment method based on KNN-Shapley, and presents its differentially private version DP-TKNN-Shapley.
A Pseudo-Semantic Loss for Autoregressive Models with Logical Constraints
Kareem Ahmed (University of California), Guy Van den Broeck (University of California)
GenerationOptimizationRecurrent Neural NetworkTransformerLarge Language ModelTextSequential
🎯 What it does: A pseudo-semantic loss is proposed, which approximates the likelihood of constraints using pseudolikelihood in the local neighborhood sampled by the model, thereby incorporating logical constraints into autoregressive generative models.
A Randomized Approach to Tight Privacy Accounting
Jiachen T. Wang (Princeton University), Prateek Mittal (Princeton University)
Safty and PrivacyTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a new privacy guarantee paradigm—Estimate-Verify-Release (EVR), which achieves a more compact upper bound on privacy leakage by converting privacy parameter estimation into formal guarantees; it also introduces a Monte Carlo-based privacy verifier and privacy accountant.
A Recurrent Neural Circuit Mechanism of Temporal-scaling Equivariant Representation
Junfeng Zuo (Peking University), Wenhao Zhang
GenerationRepresentation LearningRecurrent Neural NetworkSequential
🎯 What it does: A controllable recurrent neural circuit model based on Continuous Attractor Networks (CAN) is proposed, which can achieve time-scale equivariant representation and generate motion sequences at different speeds by adjusting the amplitude of control inputs.
A Reduction-based Framework for Sequential Decision Making with Delayed Feedback
Yunchang Yang (Peking University), Simon Shaolei Du
OptimizationReinforcement LearningSequential
🎯 What it does: This paper studies a framework for handling random delayed feedback in general sequential decision-making problems, including multi-armed bandits, single-agent Markov decision processes (MDP), and Markov games (MG). A new reduction-based framework is proposed that transforms any multi-batch algorithm with immediate feedback into a sample-efficient algorithm capable of handling random delays.
A Regularized Conditional GAN for Posterior Sampling in Image Recovery Problems
Matthew C Bendel, Philip Schniter (Ohio State University)
RestorationGenerationGenerative Adversarial NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A regularized conditional GAN (rcGAN) is proposed for quickly sampling the posterior distribution of image recovery problems from constrained measurements with high quality;
A Riemannian Exponential Augmented Lagrangian Method for Computing the Projection Robust Wasserstein Distance
Bo Jiang (Nanjing Normal University), Ya-Feng Liu (Chinese Academy of Sciences)
Optimization
🎯 What it does: This paper proposes a method for calculating the projection robust Wasserstein (PRW) distance optimization on the Stiefel manifold and Euclidean space, which mainly includes the Riemannian Exponential Augmented Lagrangian Method (REALM) and an approximate Riemannian Barzilai-Borwein and Sinkhorn iteration (iRBBS) algorithm used in the subproblems.
A Rigorous Link between Deep Ensembles and (Variational) Bayesian Methods
Veit David Wild, Jeremias Knoblauch (University College London)
OptimizationTabularStochastic Differential Equation
🎯 What it does: This paper elevates the common non-convex optimization problems in deep learning to the probability measure space, constructing a strict convex optimization framework, and derives a new class of infinite-dimensional variational inference (ID-GVI) algorithms using Wasserstein gradient flows, unifying various uncertainty quantification methods such as Bayesian, variational Bayesian, and deep ensemble.
A Robust and Opponent-Aware League Training Method for StarCraft II
Ruozi Huang (Tencent AI Lab), Yang Wei (Tencent AI Lab)
Supervised Fine-TuningReinforcement LearningSequential
🎯 What it does: A new alliance training framework called ROA-Star is proposed in StarCraft II to build more robust superhuman AI.
A Robust Exact Algorithm for the Euclidean Bipartite Matching Problem
Akshaykumar G Gattani, Pouyan Shirzadian (Virginia Tech)
OptimizationPoint Cloud
🎯 What it does: A divide-and-conquer Hungarian algorithm based on random translation quadtrees is proposed, which can solve the Euclidean bipartite matching (i.e., 1-Wasserstein distance) on a two-dimensional random point set in expected time ˜(n^{7/4} log Δ), and is generalized to arbitrary dimensions.
A Scalable Neural Network for DSIC Affine Maximizer Auction Design
Zhijian Duan (Peking University), Xiaotie Deng (Peking University)
Transformer
🎯 What it does: A scalable neural network, AMenuNet, is proposed for designing multi-item auction mechanisms that satisfy dominant strategy incentive compatibility (DSIC) and individual rationality (IR).
A Scale-Invariant Sorting Criterion to Find a Causal Order in Additive Noise Models
Alexander Gilbert Reisach, Sebastian Weichwald (University of Copenhagen)
TabularTime Series
🎯 What it does: The concept of R² rankability is proposed, and based on this, the R²‑SortnRegress algorithm is developed to infer the causal order of linear additive noise models (ANM) from observational data.
A Simple Solution for Offline Imitation from Observations and Examples with Possibly Incomplete Trajectories
Kai Yan (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: A simple and effective offline imitation learning framework called TAILO is proposed, which utilizes a discriminator to identify expert states and thresholds the discounted sum of their scores along future trajectories, directly using it as the weights for weighted behavior cloning.
A Simple Yet Effective Strategy to Robustify the Meta Learning Paradigm
Cheems Wang, Jincai Huang (National University of Defense Technology)
Meta LearningImage
🎯 What it does: This paper proposes a meta-learning framework based on distributed robustness, which enhances the robustness of rapid adaptation by minimizing the task risk's CVaR, and presents a two-stage approximate optimization algorithm.
A Single 2D Pose with Context is Worth Hundreds for 3D Human Pose Estimation
Qitao Zhao (Robotics Institute Carnegie Mellon University), Chen Chen (Center for Research in Computer Vision University of Central Florida)
Pose EstimationTransformerSupervised Fine-TuningImageVideo
🎯 What it does: This paper proposes a method to enhance single-frame 3D human pose estimation using intermediate visual features generated by a 2D pose detector, constructing the Context-Aware PoseFormer model.
A Single-Loop Accelerated Extra-Gradient Difference Algorithm with Improved Complexity Bounds for Constrained Minimax Optimization
Yuanyuan Liu (Xidian University), Zhouchen Lin (Peking University)
OptimizationAdversarial AttackImage
🎯 What it does: This paper proposes an Extra Gradient Differential Algorithm (EGDA) with a single loop acceleration for solving constrained non-convex-non-concave (NC-NC) and its special cases (NC-C, C-NC) two-stage optimization problems.
A Smooth Binary Mechanism for Efficient Private Continual Observation
Joel Daniel Andersson (Copenhagen University), Rasmus Pagh (Copenhagen University)
Safty and PrivacyComputational EfficiencyGaussian SplattingTime Series
🎯 What it does: A 'smooth binary mechanism' is proposed, which utilizes balanced binary tree leaves to reduce noise variance in the problem of differentially private counting under continuous observation, while maintaining a consistent error distribution at each time step.
A Spectral Algorithm for List-Decodable Covariance Estimation in Relative Frobenius Norm
Ilias Diakonikolas (University of Wisconsin-Madison), Thanasis Pittas (University of Wisconsin-Madison)
🎯 What it does: This paper proposes a list-decodable high-dimensional Gaussian covariance estimation algorithm based solely on spectral techniques, which can output a candidate covariance list of size O(1/α) under the conditions that both the sample size and time complexity are polynomial, and at least one candidate satisfies a relative Frobenius norm error of poly(1/α);
A Spectral Theory of Neural Prediction and Alignment
Abdulkadir Canatar (Flatiron Institute), SueYeon Chung (New York University)
Convolutional Neural NetworkRecurrent Neural NetworkTransformerImage
🎯 What it does: Using spectral theory to decompose and geometrically interpret the errors of deep neural networks (DNN) in visual cortical neural predictions.
A State Representation for Diminishing Rewards
Ted Moskovitz (University College London), Maneesh Sahani (Google DeepMind)
Reinforcement Learning
🎯 What it does: A new state representation λ R is proposed to address the non-Markovian reinforcement learning problem where rewards decrease with the number of visits.
A Sublinear-Time Spectral Clustering Oracle with Improved Preprocessing Time
Ranran Shen (University of Science and Technology of China), Pan Peng (University of Science and Technology of China)
Computational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A sublinear time spectral clustering preprocessor is designed to achieve fast querying and partitioning for graphs with good clustering properties;
A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence
Junyi Zhang (Shanghai Jiao Tong University), Ming-Hsuan Yang (University of California Merced)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningImage
🎯 What it does: This paper utilizes the internal features of Stable Diffusion (a text-to-image diffusion model) and the features of the self-supervised ViT pre-trained model DINO to construct zero-shot semantic correspondence (dense correspondence) and instance exchange tasks.
A Theoretical Analysis of Optimistic Proximal Policy Optimization in Linear Markov Decision Processes
Han Zhong (Peking University), Tong Zhang (Hong Kong University of Science and Technology)
OptimizationReinforcement Learning
🎯 What it does: A new PPO-based optimization algorithm OPPO+ is proposed to achieve approximately optimal policies in linear MDPs (linear transitions with arbitrary rewards) with full information feedback.
A Theoretical Analysis of the Test Error of Finite-Rank Kernel Ridge Regression
Tin Sum Cheng (University of Basel), David Belius (McMaster University and Vector Institute)
Tabular
🎯 What it does: This paper conducts a theoretical analysis of the test error of finite-rank kernel ridge regression (KRR) and derives its non-asymptotic upper and lower bounds to address the shortcomings of existing statistical learning guarantees.
A Theory of Link Prediction via Relational Weisfeiler-Leman on Knowledge Graphs
Xingyue Huang (University of Oxford), Pablo Barcelo
Recommendation SystemGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper proposes a Conditional Message Passing Neural Network (C-MPNN) framework for knowledge graph link prediction and conducts a theoretical analysis of its expressive power. It also provides a unified perspective with traditional models such as R-MPNN and NBFNet, and experimentally validates the effectiveness of the design choices.
A Theory of Multimodal Learning
Zhou Lu (Princeton University)
OptimizationMultimodality
🎯 What it does: This paper proposes a theoretical framework that explains why multimodal learning can outperform unimodal learning with limited data, and provides corresponding upper and lower bounds on generalization error.
A Theory of Transfer-Based Black-Box Attacks: Explanation and Implications
Yanbo Chen (Wuhan University), Weiwei Liu (Wuhan University)
Adversarial AttackRecurrent Neural Network
🎯 What it does: A 'manifold attack model' is proposed within a unified theoretical framework, which systematically explains and quantifies the transferability and low success rate phenomena of transfer attacks by modeling the semantic and geometric information of natural data on low-dimensional manifolds.
A Theory of Unsupervised Translation Motivated by Understanding Animal Communication
Shafi Goldwasser (University of California Berkeley), Orr Paradise (University of California Berkeley)
Large Language ModelGraph
🎯 What it does: A theoretical framework for Unsupervised Machine Translation (UMT) is proposed, and provable sample complexity upper bounds are provided on two simplified language models, exploring its feasibility for decoding animal communication.