arXivSub Start free trial

NeurIPS 2023 Papers — Page 22

Conference on Neural Information Processing Systems · 3218 papers

Online Inventory Problems: Beyond the i.i.d. Setting with Online Convex Optimization

Massil HIHAT, Simon Bussy (Califrais' Machine Learning Lab)

OptimizationTabularTime Series

🎯 What it does: A general online inventory optimization framework (OIO) is proposed, along with the MaxCOSD algorithm, which achieves an unbounded O(√T) expected and high-probability fallback rate under non-i.i.d. demand, arbitrary losses, and various inventory dynamics (such as perishability, stockouts, etc.).

Online Label Shift: Optimal Dynamic Regret meets Practical Algorithms

Dheeraj Baby (University of California Santa Barbara), Yu-Xiang Wang (University of California Santa Barbara)

ClassificationOptimizationImageText

🎯 What it does: This paper studies classifier adaptation under online label drift (both unsupervised and supervised), proposing an online regression-based label ratio estimation and reweighting method that achieves optimal dynamic regret.

Online learning of long-range dependencies

Nicolas Zucchet (ETH Zurich), Joao Sacramento

Recurrent Neural NetworkReinforcement LearningSequential

🎯 What it does: An online learning algorithm is proposed that can achieve nearly complete gradient descent in multi-layer independent recurrent module networks, significantly enhancing the learning capability for long-term dependencies.

Online Learning under Adversarial Nonlinear Constraints

Pavel Kolev (Max Planck Institute for Intelligent Systems), Michael Muehlebach (Max Planck Institute for Intelligent Systems)

OptimizationTabular

🎯 What it does: A new online learning algorithm is proposed—Constraint Violation Velocity Projection (CVV-Pro), which is used for online optimization under unknown, nonlinear, and time-varying constraints, achieving O(√T) regret and O(1/√T) feasibility convergence without knowledge of the feasible set.

Online List Labeling with Predictions

Samuel McCauley (Williams College), Shikha Singh (Williams College)

Time Series

🎯 What it does: A learned online list labeling data structure called learnedLLA is designed, and a theoretical analysis of the prediction error is provided.

Online Map Vectorization for Autonomous Driving: A Rasterization Perspective

Gongjie Zhang, Zuoguan Wang

SegmentationAutonomous DrivingSupervised Fine-TuningPoint Cloud

🎯 What it does: A map vectorization method called MapVR based on differentiable rasterization is proposed, which combines rasterization supervision with vectorization output to achieve more refined HD map construction.

Online Nonstochastic Model-Free Reinforcement Learning

Udaya Ghai (Amazon), Elad Hazan (Princeton University)

Reinforcement LearningSequential

🎯 What it does: A model-free reinforcement learning framework based on pseudo-disturbance is proposed, which enhances robustness by combining disturbance action control (DAC).

Online PCA in Converging Self-consistent Field Equations

Xihan Li (University College London), Jun Wang (University College London)

Auto EncoderTime SeriesPhysics Related

🎯 What it does: The self-consistent field (SCF) equations are treated as a PCA autoencoder model for non-stationary time series, proposing an online PCA technique to improve SCF solving and constructing the Adaptive Online SCF algorithm.

Online Performative Gradient Descent for Learning Nash Equilibria in Decision-Dependent Games

Zihan Zhu (Duke University), Zhuoran Yang (Yale University)

Optimization

🎯 What it does: This paper proposes an online algorithm for decision-dependent games called Online Performative Gradient Descent (OPGD), which can approximate and converge to a unique Nash equilibrium using a parameterized distribution model, even when only the loss function is observable and there is a lack of gradient oracle.

Online POMDP Planning with Anytime Deterministic Guarantees

Moran Barenboim (Technion Israel Institute of Technology), Vadim Indelman (Technion Israel Institute of Technology)

OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular

🎯 What it does: This paper proposes a deterministic bounding method based on subset observations and states, providing deterministic guarantees for online POMDP planning at any time.

Online Pricing for Multi-User Multi-Item Markets

Yigit Efe Erginbas (University of California), Soham Rajesh Phade (Wayve Technologies)

OptimizationTabular

🎯 What it does: This paper proposes an online pricing algorithm for multi-user multi-product markets that achieves approximately optimal revenue through dynamic learning and allocation without knowing user values.

Online RL in Linearly $q^\pi$-Realizable MDPs Is as Easy as in Linear MDPs If You Learn What to Ignore

Gellért Weisz (Google DeepMind), Csaba Szepesvari

Reinforcement Learning

🎯 What it does: This paper proposes a new algorithm SKIPPYELEANOR for online reinforcement learning on linear qπ-realizable Markov Decision Processes (MDP), which can achieve an approximately optimal policy through online interaction without relying on state resets.

Online robust non-stationary estimation

Abishek Sankararaman (Amazon Web Services), Murali Balakrishnan (Amazon Web Services)

Anomaly DetectionOptimizationTabularTime Series

🎯 What it does: A robust online non-stationary estimation framework is proposed, and it is proven that with appropriate parameter tuning, clipped SGD can achieve asymptotic convergence in the presence of drift, corrupted samples, and heavy-tailed data.

Open Compound Domain Adaptation with Object Style Compensation for Semantic Segmentation

Tingliang Feng (Tianjin University), Di Lin (Tianjin University)

SegmentationDomain AdaptationAutonomous DrivingConvolutional Neural NetworkImage

🎯 What it does: A mechanism for object style compensation is introduced in open composite domain adaptation for semantic segmentation, achieving compensation for style differences across different categories/instances.

Open Visual Knowledge Extraction via Relation-Oriented Multimodality Model Prompting

Hejie Cui (Emory University), Carl Yang (Emory University)

Object DetectionGenerationRetrievalTransformerPrompt EngineeringVision Language ModelImageTextMultimodality

🎯 What it does: Proposes the OpenVik framework, completing the detection of open-source visual knowledge and the generation of unformatted knowledge, achieving knowledge extraction from images to free text.

Opening the Vocabulary of Egocentric Actions

Dibyadip Chatterjee (National University of Singapore), Angela Yao (National University of Singapore)

RecognitionObject DetectionTransformerContrastive LearningVideoBenchmark

🎯 What it does: Designed and implemented an open vocabulary action recognition framework that decouples verb and object prediction, using an object-agnostic pre-trained verb encoder and a CLIP-based main object prompt encoder.

OpenMask3D: Open-Vocabulary 3D Instance Segmentation

Ayça Takmaz (ETH Zürich), Francis Engelmann (Google)

Object DetectionSegmentationContrastive LearningPoint Cloud

🎯 What it does: A zero-shot open vocabulary 3D instance segmentation method called OpenMask3D is proposed, which can segment instances in 3D scenes based on free-text queries.

OpenShape: Scaling Up 3D Shape Representation Towards Open-World Understanding

Minghua Liu (University of California San Diego), Hao Su (University of California San Diego)

ClassificationRecognitionRetrievalTransformerContrastive LearningMultimodalityPoint Cloud

🎯 What it does: Learning multimodal (text, image, point cloud) joint representation for open-world 3D shape understanding

Operation-Level Early Stopping for Robustifying Differentiable NAS

Shen Jiang (Nanjing University), Yihua Huang (Nanjing University)

Neural Architecture SearchConvolutional Neural NetworkImage

🎯 What it does: This paper proposes an improved method based on Operation-Level Early Stopping (OLES) to enhance the robustness of traditional DARTS, alleviating performance collapse caused by the dominance of skip connections.

Operator Learning with Neural Fields: Tackling PDEs on General Geometries

Louis Serrano (Sorbonne Universite), patrick gallinari

Time SeriesPhysics RelatedOrdinary Differential Equation

🎯 What it does: A meshless coordinate network named CORAL is proposed for operator learning and solving partial differential equations on arbitrary geometries.

Optimal Algorithms for the Inhomogeneous Spiked Wigner Model

Alexander Pak, Florent Krzakala (École Polytechnique Fédérale de Lausanne)

OptimizationPhysics Related

🎯 What it does: This paper studies the abrupt Wigner model under inhomogeneous noise and designs a set of AMP iterations along with a corresponding rigorous state evolution theory; it also proposes an improved spectral method that can achieve signal retrieval at the information-theoretic limit.

Optimal and Fair Encouragement Policy Evaluation and Learning

Angela Zhou (University of Southern California)

Recommendation SystemOptimizationReinforcement LearningTabular

🎯 What it does: This paper proposes an optimal policy learning framework under encouragement settings, incorporating fairness constraints (such as resource/burden equality), and provides corresponding identification, estimation, and online learning algorithms.

Optimal approximation using complex-valued neural networks

Paul Geuchen (Katholic University Eichstätt-Ingolstadt), Felix Voigtlaender (Katholic University Eichstätt-Ingolstadt)

Optimization

🎯 What it does: This paper theoretically studies the error upper bound of Complex-Valued Neural Networks (CVNN) in approximating smooth complex-valued functions, providing a quantized approximation rate of m^{-k/(2n)} for a wide range of activation functions (such as modReLU and cardioid), and proving its optimality under continuous weight selection, while revealing the curse of dimensionality.

Optimal Block-wise Asymmetric Graph Construction for Graph-based Semi-supervised Learning

Zixing Song (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)

ClassificationOptimizationGraph Neural NetworkGraph

🎯 What it does: This paper proposes an optimal asymmetric graph structure for graph semi-supervised learning (GSSL) and designs an efficient block-level graph learning algorithm called BAGL.

Optimal Convergence Rate for Exact Policy Mirror Descent in Discounted Markov Decision Processes

Emmeran Johnson (Imperial College London), Patrick Rebeschini (University of Oxford)

OptimizationReinforcement Learning

🎯 What it does: This study investigates the convergence properties of the Policy Mirror Descent (PMD) algorithm with exact policy evaluation in discounted Markov Decision Processes (MDPs), proving that it can achieve the same linear γ-rate as classical Policy Iteration (PI) and Value Iteration (VI), and provides matching lower bounds and the necessity of adaptive step sizes.

Optimal cross-learning for contextual bandits with unknown context distributions

Jon Schneider (Google Research), Julian Zimmert (Google Research)

OptimizationReinforcement Learning

🎯 What it does: Proposed an almost optimal cross-learning contextual bandwidth algorithm that achieves an expected regret of nearly ˜O(√KT) when the context distribution is unknown;

Optimal Excess Risk Bounds for Empirical Risk Minimization on $p$-Norm Linear Regression

Ayoub El Hanchi (University of Toronto), Murat A Erdogdu (University of Toronto)

Optimization

🎯 What it does: This paper studies the non-asymptotic excess risk bounds of empirical risk minimization (ERM) in the p-norm linear regression problem and provides high-probability upper bounds for p ∈ (1, ∞).

Optimal Exploration for Model-Based RL in Nonlinear Systems

Andrew Wagenmaker (University of Washington), Kevin Jamieson (University of Washington)

OptimizationRobotic IntelligenceReinforcement Learning

🎯 What it does: This paper studies optimal exploration in model-based reinforcement learning for nonlinear systems, aiming to quantify which system parameters are most important for learning good controllers, and proposes an efficient exploration algorithm to reduce the uncertainty of these parameters.

Optimal Extragradient-Based Algorithms for Stochastic Variational Inequalities with Separable Structure

Angela Yuan (University of California), Simon Shaolei Du (University of Washington)

Optimization

🎯 What it does: A random extragradient-based algorithm combined with Nesterov acceleration, called AG-EG, is proposed to solve the strongly monotone stochastic variational inequality (VI) problem with a separable structure, achieving optimal convergence rates in bilinear coupled strong convex-strong concave saddle point problems and bilinear games.

Optimal Guarantees for Algorithmic Reproducibility and Gradient Complexity in Convex Optimization

Liang Zhang (ETH Zurich and Max Planck Institute), Niao He (ETH Zurich)

Optimization

🎯 What it does: A regularization framework is proposed in convex optimization and convex-concave min-max problems, achieving optimal algorithm reproducibility and near-optimal gradient complexity.

Optimal Learners for Realizable Regression: PAC Learning and Online Learning

Idan Attias (Ben-Gurion University of the Negev), Grigoris Velegkas (Yale University)

🎯 What it does: This paper constructs an optimal learning algorithm under the framework of PAC learning and online learning for realizable regression, proposes a new combinatorial dimension, and provides necessary and sufficient conditions for the feasibility of learning in relation to the dimension.

Optimal Parameter and Neuron Pruning for Out-of-Distribution Detection

Chao Chen (Alibaba Cloud), Jieping Ye (Alibaba Cloud)

Anomaly DetectionOptimizationConvolutional Neural NetworkImage

🎯 What it does: A post-training parameter and neuron pruning method called OPNP is proposed, which identifies and removes weights and neurons that are detrimental to OOD detection based on gradient sensitivity.

Optimal Preconditioning and Fisher Adaptive Langevin Sampling

Michalis Titsias

OptimizationComputational EfficiencyTabularStochastic Differential Equation

🎯 What it does: This paper proposes a Laplace diffusion preconditioning method optimized based on expected jump distance, and applies it to construct an adaptive MALA (FisherMALA), which learns the global preconditioning matrix (i.e., the inverse of the Fisher information matrix) online, significantly improving high-dimensional sampling efficiency.

Optimal privacy guarantees for a relaxed threat model: Addressing sub-optimal adversaries in differentially private machine learning

Georgios Kaissis (Technical University of Munich), Daniel Rueckert (Technical University of Munich)

OptimizationSafty and PrivacyImageTabular

🎯 What it does: This paper introduces a more realistic threat model (RTM) in differential privacy machine learning and formally characterizes and validates the optimal error rate of membership inference attacks under it.

Optimal Rates for Bandit Nonstochastic Control

Y. Jennifer Sun (Princeton University), Elad Hazan (Princeton University)

OptimizationReinforcement Learning

🎯 What it does: Under semi-adversarial noise, a BBO-with-memory based EBPC algorithm is designed for linear quadratic regulation (LQR/LQG) control problems with time-varying, observable, and unobservable states, achieving an optimal cumulative regret rate of O(√T);

Optimal Regret Is Achievable with Bounded Approximate Inference Error: An Enhanced Bayesian Upper Confidence Bound Framework

Ziyi Huang (Columbia University), Haofeng Zhang (Columbia University)

OptimizationReinforcement LearningTabular

🎯 What it does: A new reinforcement learning algorithm based on Bayesian upper confidence bounds (EBUCB) is designed and analyzed, capable of handling the multi-armed bandit problem under the constraint of bounded approximate Bayesian inference error.

Optimal testing using combined test statistics across independent studies

Lasse Vuursteen (Delft University of Technology), Harry van Zanten (Vrije Universiteit Amsterdam)

🎯 What it does: This paper studies how to combine test statistics from multiple independent experiments to improve statistical power, especially in the context of high-dimensional models and composite hypothesis testing.

Optimal Time Complexities of Parallel Stochastic Optimization Methods Under a Fixed Computation Model

Alexander Tyurin (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)

OptimizationTabular

🎯 What it does: A new time oracle protocol and corresponding lower bound on time complexity are proposed, along with the design of the optimal parallel stochastic gradient method, Rennala SGD, to achieve this lower bound.

Optimal Transport for Treatment Effect Estimation

Hao Wang (Zhejiang University), Ruiming Tang (Huawei Noah's Ark Lab)

OptimizationTabularBenchmark

🎯 What it does: An optimal transport framework based on relaxed mass preservation (ESCFR) is proposed to simultaneously address the mini-batch sampling effect and treatment selection bias caused by unobserved confounders in observational data.

Optimal Transport Model Distributional Robustness

Van-Anh Nguyen (Monash University), Dinh Phung (VinAI)

ClassificationOptimizationConvolutional Neural NetworkImageStochastic Differential Equation

🎯 What it does: Introduce an optimal transport-based distributional robustness framework (OT-MDR) in the model space and extend it to three scenarios: single model, ensemble model, and Bayesian neural networks;

Optimal Transport-Guided Conditional Score-Based Diffusion Model

Xiang Gu (Xi'an Jiaotong University), Zongben Xu (Xi'an Jiaotong University)

Image TranslationGenerationDiffusion modelScore-based ModelImageStochastic Differential Equation

🎯 What it does: This paper proposes a conditionally score-based diffusion model guided by optimal transport (OTCS) for image translation using unpaired or partially paired training sets.

Optimal Treatment Allocation for Efficient Policy Evaluation in Sequential Decision Making

Ting Li (Shanghai University of Finance and Economics), Hongtu Zhu (University of North Carolina at Chapel Hill)

OptimizationReinforcement LearningSequential

🎯 What it does: The study investigates how to design experimental allocation schemes in sequential decision-making to maximize the amount of information obtained from online experiments and accurately estimate treatment effects.

Optimal Treatment Regimes for Proximal Causal Learning

Tao Shen (National University of Singapore), Yifan Cui (Zhejiang University)

TabularBiomedical DataElectronic Health Records

🎯 What it does: In observational studies that do not meet the unbiased assumption, a new individualized treatment regimen (ITR) based on the proximal causal inference framework is proposed, which identifies and learns the optimal treatment decision by simultaneously utilizing treatment and outcome confounding bridges.

Optimal Unbiased Randomizers for Regression with Label Differential Privacy

Ashwinkumar Badanidiyuru (Google), Chiyuan Zhang (Google Research)

OptimizationSafty and PrivacyTabular

🎯 What it does: A new unbiased label randomization method is proposed for training regression models under label differential privacy.

Optimality in Mean Estimation: Beyond Worst-Case, Beyond Sub-Gaussian, and Beyond $1+\alpha$ Moments

Trung Dang (University of Texas at Austin), Paul Valiant (Purdue University)

Optimization

🎯 What it does: This paper studies one-dimensional mean estimation, proposes a new 'neighborhood optimality' framework, and provides a general indistinguishable construction, proving that mean estimation cannot surpass the optimal sub-Gaussian error upper bound for almost all distributions.

Optimality of Message-Passing Architectures for Sparse Graphs

Aseem Baranwal (University of Waterloo), Aukosh Jagannath (University of Waterloo)

ClassificationOptimizationGraph Neural NetworkGraph

🎯 What it does: This study investigates the node classification problem on very sparse feature maps, providing a locally Bayesian optimal message passing architecture and analyzing its generalization error.

Optimistic Active Exploration of Dynamical Systems

Bhavya Sukhija (ETH Zürich), Andreas Krause (ETH Zürich)

OptimizationReinforcement LearningTime Series

🎯 What it does: This paper proposes the OPAX algorithm for actively exploring unknown dynamical systems and learning globally consistent dynamical models, thereby achieving multi-task zero-shot planning.

Optimistic Exploration in Reinforcement Learning Using Symbolic Model Estimates

Sarath Sreedharan (Colorado State University), Michael Katz (IBM)

OptimizationRobotic IntelligenceReinforcement Learning

🎯 What it does: This paper proposes an optimistic exploration framework based on symbolic models, utilizing an optimistic symbolic approximation model with online learning in conjunction with a diversified planner to guide RL agents in goal-oriented exploration within sparse reward environments, continuously updating the model through environmental feedback.

Optimistic Meta-Gradients

Sebastian Flennerhag (Google DeepMind), Satinder Singh (Google DeepMind)

OptimizationMeta LearningImage

🎯 What it does: This study explores the connection between gradient-based meta-learning and convex optimization, proving the convergence rate of meta-learning in a single-task setting and demonstrating the importance of optimism in accelerating convergence.

Optimistic Natural Policy Gradient: a Simple Efficient Policy Optimization Framework for Online RL

Qinghua Liu (Princeton University), Csaba Szepesvari

OptimizationReinforcement Learning

🎯 What it does: A simple and efficient policy optimization framework for online reinforcement learning, OPTIMISTIC NPG, is proposed, along with a theoretical convergence analysis under linear MDPs and general function approximation.

Optimistic Rates for Multi-Task Representation Learning

Austin Watkins (Johns Hopkins University), Raman Arora (Johns Hopkins University)

OptimizationRepresentation Learning

🎯 What it does: This paper presents an adaptive (optimistic) risk upper bound for transfer learning and multi-task learning within the framework of Multi-Task Representation Learning (MTRL), and proves a fast convergence rate under smooth non-negative loss functions using local Rademacher complexity; it also proposes a local Rademacher chain rule for composite predictor classes.

Optimization and Bayes: A Trade-off for Overparameterized Neural Networks

Zhengmian Hu (University of Maryland), Heng Huang (University of Maryland)

Optimization

🎯 What it does: An algorithm named Transformative Bayesian Learning (TransBL) is proposed, which combines gradient optimization and Bayesian learning through importance sampling to achieve a trade-off between generalization and sampling efficiency for over-parameterized neural networks.

Optimization of Inter-group criteria for clustering with minimum size constraints

Eduardo Sany Laber, Lucas Murtinho (Pontifical Catholic University of Rio de Janeiro)

OptimizationTabular

🎯 What it does: An optimization method for clustering under the minimum spacing (Min-Sp) and minimum spanning tree spacing (MST-Sp) is proposed, addressing the minimum size constraint that each cluster must contain at least L samples.

Optimization or Architecture: How to Hack Kalman Filtering

Ido Greenberg (Technion), Shie Mannor (Technion)

Object TrackingAutonomous DrivingOptimizationVideoTime SeriesStochastic Differential Equation

🎯 What it does: An optimized noise parameter learning framework based on MSE gradient optimization (Optimized KF, OKF) is introduced into the traditional Kalman filter (KF), demonstrating significant superiority over traditional KF and neural network baselines across various tasks.

Optimize Planning Heuristics to Rank, not to Estimate Cost-to-Goal

Leah Chrestien (Czech Technical University in Prague), Tomáš Pevný (Czech Technical University in Prague)

OptimizationConvolutional Neural NetworkTabular

🎯 What it does: A learning ranking heuristic is proposed and validated based on a ranking loss function, which is used to learn heuristics that can be directly applied to A* and GBFS from solved instances;

Optimized Covariance Design for AB Test on Social Network under Interference

Qianyi Chen (Tsinghua University), Yong Wang (Tencent)

OptimizationGraph

🎯 What it does: In A/B testing considering network interference in social networks, a covariance-based experimental randomization scheme is proposed, aiming to balance the bias and variance of the estimator and improve the estimation accuracy of the Global Average Treatment Effect (GATE).

Optimizing over trained GNNs via symmetry breaking

Shiqiang Zhang (Imperial College London), Ruth Misener (Imperial College London)

OptimizationDrug DiscoveryGraph Neural NetworkGraph

🎯 What it does: This paper proposes a framework for optimization on trained Graph Neural Networks (GNNs), providing two types of symmetry-breaking constraints to address the symmetry issues caused by graph isomorphism, and designs a feasible graph indexing algorithm.

Optimizing Prompts for Text-to-Image Generation

Yaru Hao (Microsoft Research), Furu Wei (Microsoft Research)

GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringImageText

🎯 What it does: This paper proposes an automated prompt optimization framework (Promptist) that first performs supervised fine-tuning on manually designed prompts and then further improves them using reinforcement learning to generate prompts that better align with the preferences of text-to-image models, thereby enhancing the aesthetic quality of generated images and their relevance to the original user intent.

Optimizing Solution-Samplers for Combinatorial Problems: The Landscape of Policy-Gradient Method

Constantine Caramanis (University of Texas at Austin), Christos Tzamos (University of Wisconsin)

OptimizationReinforcement LearningGraph

🎯 What it does: A theoretical framework is proposed for analyzing and designing compressible generative models that can be optimized through gradient descent, aimed at generating approximate optimal solutions to combinatorial optimization problems.

Oracle Complexity of Single-Loop Switching Subgradient Methods for Non-Smooth Weakly Convex Functional Constrained Optimization

Yankun Huang (University of Iowa), Qihang Lin (University of Iowa)

OptimizationTabular

🎯 What it does: This study investigates the theoretical complexity of the Switching Subgradient (SSG) method for non-smooth weakly convex constrained optimization problems, providing an operator complexity of O(1/ε⁴); experiments are conducted on binary classification tasks with fairness constraints.

Order Matters in the Presence of Dataset Imbalance for Multilingual Learning

Dami Choi (University of Toronto), Behrooz Ghorbani (OpenAI)

TransformerText

🎯 What it does: The study proposes a two-stage training scheme for multi-task learning under data imbalance, where pre-training is first conducted on high-resource tasks, followed by joint fine-tuning of both high and low-resource tasks.

Ordering-based Conditions for Global Convergence of Policy Gradient Methods

Jincheng Mei (Google DeepMind), Dale Schuurmans (Google DeepMind)

OptimizationReinforcement Learning

🎯 What it does: This paper studies the global convergence properties of Policy Gradient methods (Softmax PG and Natural PG) in single-state multi-armed bandits under linear function approximation.

Orthogonal Non-negative Tensor Factorization based Multi-view Clustering

Jing Li (Xidian University), Wei Xia (Xidian University)

OptimizationImage

🎯 What it does: This paper proposes an Orthogonal Non-negative Tensor Factorization (Orth-NTF) method based on one-sided orthogonal constraints, using tensor Schatten p-norm regularization for joint decomposition of multi-view anchor graphs, thereby fully mining complementary information between views while maintaining spatial structure, achieving intuitive clustering without post-processing.

Out-of-distribution Detection Learning with Unreliable Out-of-distribution Sources

Haotian Zheng (Xidian University), Bo Han (Hong Kong Baptist University)

GenerationAnomaly DetectionGenerative Adversarial NetworkContrastive LearningImage

🎯 What it does: This paper proposes an OOD detection method based on auxiliary tasks called ATOL, which utilizes fake OOD data generated by a generator to improve the OOD detection performance of the predictor and reduce misjudgments.

Outlier-Robust Gromov-Wasserstein for Graph Data

Lemin Kong (Chinese University of Hong Kong), Anthony Man-Cho So (Chinese University of Hong Kong)

Anomaly DetectionOptimizationGraph Neural NetworkGraph

🎯 What it does: This paper proposes a robust RGW model for Gromov-Wasserstein distance, which can still approximate the original GW distance in the presence of outliers.

Outlier-Robust Wasserstein DRO

Sloan Nietert (Cornell University), Soroosh Shafiee (Cornell University)

Anomaly DetectionOptimizationTabular

🎯 What it does: This paper proposes a distributionally robust optimization framework that simultaneously defends against geometric (Wasserstein) perturbations and non-geometric (total variation TV) contamination (Outlier-Robust Wasserstein DRO), providing theoretical guarantees and computable forms.

Overcoming Recency Bias of Normalization Statistics in Continual Learning: Balance and Adaptation

Yilin Lyu (Beijing Jiaotong University), Liping Jing (Beijing Jiaotong University)

Convolutional Neural NetworkImage

🎯 What it does: The study addresses the recursive bias problem of batch normalization (BN) in continual learning and proposes the AdaB2N method to achieve adaptive balancing and adaptation of BN statistics, enhancing continual learning performance.

P-Flow: A Fast and Data-Efficient Zero-Shot TTS through Speech Prompting

Sungwon Kim (Seoul National University), Bryan Catanzaro (NVIDIA)

GenerationTransformerPrompt EngineeringFlow-based ModelAudio

🎯 What it does: A zero-shot text-to-speech (TTS) model named P-Flow is proposed, which utilizes short audio prompts for speaker adaptation and employs a flow-matching generator to achieve non-autoregressive, real-time speech synthesis.

PAC Learning Linear Thresholds from Label Proportions

Anand Paresh Brahmbhatt (Google Research India), Aravindan Raghuveer (Google Research India)

ClassificationOptimizationStochastic Differential Equation

🎯 What it does: Proposes a PAC-LLP algorithm for efficiently learning linear threshold functions (LTF) from bags with known label proportions under Gaussian distribution.

PAC-Bayes Generalization Certificates for Learned Inductive Conformal Prediction

Apoorva Sharma (NVIDIA Research), Anirudha Majumdar (Princeton University)

ClassificationOptimizationComputational EfficiencyGaussian SplattingTabularSequential

🎯 What it does: This paper proposes a learning framework that utilizes PAC-Bayes theory to enhance the efficiency and generalization guarantees of Inductive Conformal Prediction (ICP), allowing for direct learning of model and score function parameters on calibration data without the need for an additional validation set, thus achieving dual guarantees of expected coverage and efficiency.

PAC-Bayesian Spectrally-Normalized Bounds for Adversarially Robust Generalization

Jiancong Xiao (University of Pennsylvania), Zhi-Quan Luo (Chinese University of Hong Kong)

OptimizationAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: A PAC-Bayes based spectral norm robust generalization bound is proposed, which provides a theoretical upper bound on the error of adversarial training models without additional assumptions.

PackQViT: Faster Sub-8-bit Vision Transformers via Full and Packed Quantization on the Mobile

Peiyan Dong (Northeastern University), Yanzhi Wang (Northeastern University)

ClassificationObject DetectionComputational EfficiencyTransformerImage

🎯 What it does: The PackQViT framework is proposed to achieve full sub-8-bit quantization for accelerating mobile inference of Vision Transformers.

PaintSeg: Painting Pixels for Training-free Segmentation

Xiang Li (Carnegie Mellon University), Bhiksha Raj (Microsoft)

SegmentationTransformerDiffusion modelImage

🎯 What it does: A completely unsupervised image segmentation method called PaintSeg is proposed, which utilizes contrastive painting (AMCP) to create a contrast between the original image and the generated painting image, thereby gradually improving the segmentation mask.

Pairwise Causality Guided Transformers for Event Sequences

Xiao Shou (Rensselaer Polytechnic Institute), Kristin Bennett (Rensselaer Polytechnic Institute)

TransformerTime SeriesSequential

🎯 What it does: This study proposes injecting paired causal relationships into the Transformer model to improve the predictive performance of multivariate event sequences.

PanoGen: Text-Conditioned Panoramic Environment Generation for Vision-and-Language Navigation

Jialu Li (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)

GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageTextMultimodality

🎯 What it does: This paper generates infinitely diverse panoramic indoor environments through text descriptions and uses these generated environments for pre-training and fine-tuning in visual-language navigation (VLN), enhancing the agent's generalization ability in unseen environments.

PanoGRF: Generalizable Spherical Radiance Fields for Wide-baseline Panoramas

Zheng Chen (Tsinghua University), Song-Hai Zhang (Tsinghua University)

GenerationData SynthesisDepth EstimationConvolutional Neural NetworkNeural Radiance FieldImage

🎯 What it does: This paper proposes PanoGRF, a general spherical radiance field for wide baseline panoramas that directly utilizes panoramic images for view synthesis.

PAPR: Proximity Attention Point Rendering

Yanshu Zhang (Simon Fraser University), Ke Li (Simon Fraser University)

GenerationData SynthesisNeural Radiance FieldGaussian SplattingImagePoint Cloud

🎯 What it does: A sparse point cloud scene representation and differentiable rendering method called PAPR is proposed, which can learn from scratch and generate high-quality views using only RGB images.

ParaFuzz: An Interpretability-Driven Technique for Detecting Poisoned Samples in NLP

Lu Yan (Purdue University), Xiangyu Zhang (Purdue University)

Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: This paper proposes a testing framework for backdoor detection based on model predictive interpretability, called PARAFUZZ. It utilizes ChatGPT to rewrite inputs while maintaining semantics, in order to remove potential triggers and determine whether samples have been poisoned.

Parallel Sampling of Diffusion Models

Andy Shih (Stanford University), Nima Anari (Stanford University)

GenerationRobotic IntelligenceDiffusion modelImageStochastic Differential Equation

🎯 What it does: This paper proposes a method for parallelizing diffusion model sampling using Picard iteration, called ParaDiGMS, which significantly reduces the time delay for a single sample while maintaining sample quality.

Parallel Spiking Neurons with High Efficiency and Ability to Learn Long-term Dependencies

Wei Fang (Peking University), Yonghong Tian (Peking University)

Spiking Neural NetworkImage

🎯 What it does: This paper proposes a reset-free and parallelizable series of Parallel Spiking Neurons (PSN) to significantly enhance the simulation speed and long-term dependency learning capability of spiking neural networks.

Parallel Submodular Function Minimization

Deeparnab Chakrabarty (Dartmouth College), Aaron Sidford (Stanford University)

Optimization

🎯 What it does: Two new algorithms for parallel submodular function minimization (SFM) are proposed, completing O~(nM^2) queries in ˜O(n^{1/3}M^{2/3}) rounds, and a simple algorithm that performs O(n(M+1)) queries in two rounds is provided.

Parallel-mentoring for Offline Model-based Optimization

Can Chen (MILA - Quebec AI Institute), Christopher Pal (MILA - Quebec AI Institute)

OptimizationTabular

🎯 What it does: A tri-mentoring method for offline model-based optimization is proposed and evaluated, aimed at enhancing the quality of optimal designs through mutual guidance among agents in multi-dimensional design tasks.

Parameter and Computation Efficient Transfer Learning for Vision-Language Pre-trained Models

Qiong Wu (Xiamen University), Rongrong Ji (Xiamen University)

Computational EfficiencyTransformerReinforcement LearningVision Language ModelMultimodality

🎯 What it does: This paper proposes a transfer learning method that achieves dual efficiency in parameters and computation on visual-language pre-trained models, called Dynamic Architecture Skipping (DAS).

Parameter-efficient Tuning of Large-scale Multimodal Foundation Model

Haixin Wang (Peking University), Qi Tian (Huawei Cloud and AI)

RetrievalRepresentation LearningTransformerPrompt EngineeringImageVideoTextMultimodality

🎯 What it does: A lightweight cross-modal transfer framework called Aurora is proposed, achieving efficient cross-modal task transfer by training only about 0.1M parameters on the frozen BLIP large model.

Parameterizing Context: Unleashing the Power of Parameter-Efficient Fine-Tuning and In-Context Tuning for Continual Table Semantic Parsing

Yongrui Chen (Southeast University), Xinnan Guo (Southeast University)

OptimizationKnowledge DistillationTransformerPrompt EngineeringTabular

🎯 What it does: This paper proposes a continuous table semantic parsing method (C3) that combines Parameter-Efficient Fine-Tuning (PEFT) and In-Context Tuning (ICT). It captures and compresses contextual information in few-shot scenarios through a teacher-student framework, completely avoiding catastrophic forgetting.

Parameterizing Non-Parametric Meta-Reinforcement Learning Tasks via Subtask Decomposition

Suyoung Lee (KAIST), Youngchul Sung (KAIST)

Meta LearningReinforcement LearningTabularBenchmark

🎯 What it does: A new meta reinforcement learning method SDVT is proposed, which achieves better generalization of task diversity by decomposing non-parametric tasks into shared sub-tasks and parameterizing them.

Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense

Kalpesh Krishna (University of Massachusetts Amherst), Mohit Iyyer (University of Massachusetts Amherst)

GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation

🎯 What it does: This paper proposes a controllable long-text rewriting model called DIPPER, designed to rewrite AI-generated text to evade existing detection methods; it also presents a retrieval-based defense scheme that uses a generated text library to detect whether the text is machine-generated.

Pareto Frontiers in Deep Feature Learning: Data, Compute, Width, and Luck

Benjamin L. Edelman (Harvard University), Cyril Zhang (Microsoft Research)

ClassificationOptimizationData-Centric LearningRecurrent Neural NetworkSupervised Fine-TuningTabular

🎯 What it does: This study investigates the multi-resource Pareto front of sparse singular learning, theoretically deriving and experimentally validating the improvement of sample efficiency in MLP due to width and sparse initialization, and transferring the conclusions to tabular data.

Parsel🐍: Algorithmic Reasoning with Language Models by Composing Decompositions

Eric Zelikman (Stanford University), Nick Haber (Stanford University)

Robotic IntelligenceAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: The Parsel framework is proposed, which utilizes large language models to hierarchically decompose, implement, and verify complex algorithmic tasks, thereby achieving the automated generation of complex programs and robotic planning tasks.

Partial Counterfactual Identification of Continuous Outcomes with a Curvature Sensitivity Model

Valentyn Melnychuk (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)

Flow-based ModelTabular

🎯 What it does: This paper studies counterfactual inference for continuous outcomes in Markov structural causal models, proposing a Curvature-Sensitive Model (CSM) and implementing a Pseudo-Invertible Decoder (APID) based on Residual Regularization Flow and Variational Augmentation, achieving interval estimation for partial identification.

Partial Label Learning with Dissimilarity Propagation guided Candidate Label Shrinkage

Yuheng Jia (Southeast University), Yongqiang Dong (Southeast University)

ClassificationTabular

🎯 What it does: A partial label learning method based on adversarial priors, DPCLS, is proposed, which utilizes a label confidence matrix and its transposed second-order similarity matrix, as well as a semantic dissimilarity matrix constructed and propagated through a candidate label set to achieve candidate label contraction and label discrimination.

Partial Matrix Completion

Elad Hazan (Princeton University), Y. Jennifer Sun (Princeton University)

Recommendation SystemOptimizationTabular

🎯 What it does: A Partial Matrix Completion framework is proposed, with offline and online algorithms designed, providing theoretical guarantees on coverage and error, and experimentally validating the relationship between confidence and prediction error.

Partial Multi-Label Learning with Probabilistic Graphical Disambiguation

Jun-Yi Hang (Southeast University), Min-Ling Zhang (Southeast University)

ClassificationGraph Neural NetworkTabular

🎯 What it does: This paper proposes a joint framework PARD that utilizes probabilistic graphical models for denoising and predicting candidate labels in partial multi-label learning.

Participatory Personalization in Classification

Hailey Joren (University of California San Diego), Berk Ustun (University of California San Diego)

ClassificationBiomedical DataElectronic Health Records

🎯 What it does: A classification model (participatory system) is proposed that allows individuals to choose whether to provide personal information during prediction, achieving informed consent and data minimization by allowing users to opt-in/opt-out and informing them of their benefits.

Particle-based Variational Inference with Generalized Wasserstein Gradient Flow

Ziheng Cheng (Peking University), Cheng Zhang (Peking University)

Tabular

🎯 What it does: A particle variational inference framework based on generalized Wasserstein gradient flow (GWG) is designed, and an adaptive version Ada-GWG is proposed to improve sampling efficiency.

Parts of Speech–Grounded Subspaces in Vision-Language Models

James Oldfield (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)

GenerationRepresentation LearningTransformerVision Language ModelContrastive LearningImageText

🎯 What it does: This paper proposes to utilize part of speech (nouns, adjectives, verbs, adverbs) as supervision in the shared visual-language space of CLIP to learn corresponding subspaces/manifolds, in order to split image or text representations into two visual modes: 'content (objects/actions)' and 'appearance (color, style)'.

Passive learning of active causal strategies in agents and language models

Andrew Kyle Lampinen, Jane X Wang

Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringText

🎯 What it does: This paper studies whether agents and language models can learn generalizable active experimental and causal intervention strategies through passive observation data alone (such as expert demonstrations or self-supervised training of pre-trained language models), and utilize these strategies to achieve goals during the interactive testing phase.

Patch Diffusion: Faster and More Data-Efficient Training of Diffusion Models

Zhendong Wang (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)

GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelImage

🎯 What it does: This paper proposes a Patch Diffusion training framework, which significantly accelerates the training of diffusion models and improves data efficiency by performing conditional score matching at the image patch level.

Patch n’ Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution

Mostafa Dehghani (Google DeepMind), Neil Houlsby (Google DeepMind)

Object DetectionSegmentationRetrievalTransformerContrastive LearningImageVideo

🎯 What it does: By introducing the Patch n' Pack technique in Vision Transformer, the model can be trained and inferred at native resolution and arbitrary aspect ratios, thereby improving performance under the same computational budget and achieving more flexible control over inference costs.

Path following algorithms for $\ell_2$-regularized $M$-estimation with approximation guarantee

Yunzhang Zhu (Ohio State University), Renxiong Liu (Nokia Bell Labs)

OptimizationTabular

🎯 What it does: This paper studies the path-following algorithm for ℓ2 regularization M-estimation and proposes a grid point selection and stopping criterion based on interpolation error and optimization error, achieving a theoretical guarantee for global approximation error.