ICML 2023 Papers — Page 13
International Conference on Machine Learning · 1828 papers
Optimal Arms Identification with Knapsacks
Shaoang Li (University of Science and Technology of China), Xiangyang Li
OptimizationReinforcement LearningBiomedical Data
🎯 What it does: This paper proposes the Optimal Arm Identification with Knapsack Constraints (OAK) problem, extending the Best Arm Identification (BAI) framework to consider resource consumption. The authors present a new OAK algorithm and prove the upper bound of the algorithm by exploring the relationship between selecting the optimal action and the structure of the feasible region.
Optimal Convergence Rates for Agnostic Nyström Kernel Learning
Jian Li (Institute of Information Engineering, Chinese Academy of Sciences), Weiping Wang (Institute of Information Engineering, Chinese Academy of Sciences)
🎯 What it does: This paper presents the optimal convergence rate of Nystrom Kernel Ridge Regression (KRR-Nystrom) in an agnostic setting where the fitting function does not need to lie within the kernel space, and provides capacity-related optimal theoretical bounds for both uniform sampling and data-dependent sampling.
Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning
Tongzhou Wang (Massachusetts Institute of Technology), Amy Zhang (University of Texas at Austin)
OptimizationRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: A novel reinforcement learning algorithm named QRL is proposed, which models the value function for achieving goals as a differentiable quasimetric model, directly learning the optimal goal-conditioned value function.
Optimal Horizon-Free Reward-Free Exploration for Linear Mixture MDPs
Junkai Zhang (University of California), Quanquan Gu (University of California)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes an algorithm for achieving reward-agnostic exploration in linear mixed MDPs, making the sample complexity independent of the planning horizon.
Optimal LP Rounding and Linear-Time Approximation Algorithms for Clustering Edge-Colored Hypergraphs
Nate Veldt (Texas A&M University)
OptimizationGraphBenchmark
🎯 What it does: This paper studies the approximability of the edge coloring hypergraph clustering (MINECC) problem, proposes an optimal rounding scheme based on linear programming, and designs a linear-time 2-approximation combinatorial algorithm; theoretical analysis and experimental evaluation are conducted on instances.
Optimal No-Regret Learning for One-Sided Lipschitz Functions
Paul Duetting, Joshua Ruizhi Wang
Optimization
🎯 What it does: This paper proposes and analyzes an online learning algorithm that can approximate the maximum value of a one-dimensional one-sided Lipschitz function with a total regret of O(log log T) under noise-free feedback.
Optimal Online Generalized Linear Regression with Stochastic Noise and Its Application to Heteroscedastic Bandits
Heyang Zhao (University of California), Quanquan Gu (University of California)
OptimizationReinforcement Learning
🎯 What it does: This paper studies online generalized linear regression and heteroscedastic multi-armed bandit problems with random noise, proposing an FTRL-based algorithm and providing near-optimal risk bounds.
Optimal randomized multilevel Monte Carlo for repeatedly nested expectations
Yasa Syed (Rutgers University), Guanyang Wang (Rutgers University)
OptimizationComputational EfficiencyFinance Related
🎯 What it does: A recursive unbiased random multi-level Monte Carlo estimator named READ has been designed and implemented to evaluate nested expectations of arbitrary depth, theoretically achieving optimal or near-optimal computational complexity under the conditions of LBS or LBL.
Optimal Rates and Efficient Algorithms for Online Bayesian Persuasion
Martino Bernasconi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the problem of online Bayesian persuasion, proposing new no-regret algorithms for both single and multiple receivers under partial feedback, and achieving polynomial time algorithms for the multi-receiver case by introducing a type-reporting mechanism.
Optimal Sets and Solution Paths of ReLU Networks
Aaron Mishkin (Stanford University), Mert Pilanci (Stanford University)
OptimizationTabular
🎯 What it does: This study establishes a theoretical framework for transforming the training problem of two-layer ReLU networks into a convex optimization program, analytically deriving the geometric structure of the global optimal solution set and all stationary points, proposing an optimal pruning algorithm for minimal width networks, and analyzing the continuity and sensitivity of the regularization path.
Optimal Shrinkage for Distributed Second-Order Optimization
Fangzhao Zhang (Stanford University), Mert Pilanci (Stanford University)
OptimizationTabular
🎯 What it does: This paper addresses the bias problem in the inversion of Hessian matrices in distributed second-order optimization and proposes a shrinkage-based unbiased covariance analytical inverse estimation method, which is applied to distributed Newton and stochastic sketch algorithms.
Optimal Stochastic Non-smooth Non-convex Optimization through Online-to-Non-convex Conversion
Ashok Cutkosky (Boston University), Francesco Orabona (Boston University)
Optimization
🎯 What it does: A novel algorithmic framework is proposed for non-smooth non-convex stochastic optimization problems by transforming the shifted regret from online learning into optimization update directions, along with corresponding theoretical analysis.
Optimality of Thompson Sampling with Noninformative Priors for Pareto Bandits
Jongyeong Lee (University of Tokyo), Masashi Sugiyama (RIKEN)
OptimizationReinforcement LearningTabular
🎯 What it does: The study investigates the asymptotic optimality of the Thompson Sampling (TS) algorithm with non-informative priors under a two-parameter Pareto distribution and proposes a truncation method to enhance the performance of certain priors.
Optimally-weighted Estimators of the Maximum Mean Discrepancy for Likelihood-Free Inference
Ayush Bharti (Aalto University), Francois-Xavier Briol
🎯 What it does: A maximum mean discrepancy (MMD) estimator based on optimal weighting is proposed to improve the sample complexity of likelihood-free inference;
Optimistic Online Mirror Descent for Bridging Stochastic and Adversarial Online Convex Optimization
Sijia Chen (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the Stochastic Extended Adversarial (SEA) model proposed by Sachs et al. (2022) and provides theoretical guarantees through the Optimistic Online Mirror Descent (OMD) framework.
Optimistic Planning by Regularized Dynamic Programming
Antoine Moulin (Universitat Pompeu Fabra), Gergely Neu (Universitat Pompeu Fabra)
OptimizationReinforcement LearningTabular
🎯 What it does: A new method based on regularized dynamic programming is proposed for optimistic planning in infinite-horizon discounted Markov decision processes.
Optimization for Amortized Inverse Problems
Tianci Liu (Purdue University), Qi Lei (New York University)
RestorationOptimizationFlow-based ModelImage
🎯 What it does: The AIPO (Amortized Inverse Problem Optimization) algorithm is proposed for MAP estimation in inverse problems with a reversible generative model (normalizing flow) prior.
Optimizing DDPM Sampling with Shortcut Fine-Tuning
Ying Fan (University of Wisconsin Madison), Kangwook Lee (University of Wisconsin Madison)
GenerationOptimizationReinforcement LearningDiffusion modelImage
🎯 What it does: Shortcut Fine-Tuning (SFT) is applied to the pre-trained DDPM sampler, achieving fast sampling optimization by directly minimizing the Integral Probability Metric (IPM) and utilizing policy gradient equivalence.
Optimizing Hyperparameters with Conformal Quantile Regression
David Salinas (Amazon Web Services), Cedric Archambeau (Amazon Web Services)
OptimizationHyperparameter SearchTabular
🎯 What it does: A hyperparameter optimization proxy based on Conformal Quantile Regression (CQR) is proposed, which can obtain reliable confidence intervals without assuming Gaussian noise, and implements a unified search framework for single-fidelity and multi-fidelity through Thompson sampling and ASHA.
Optimizing Mode Connectivity for Class Incremental Learning
Haitao Wen (University of Electronic Science and Technology of China), Hongliang Li (University of Electronic Science and Technology of China)
OptimizationImage
🎯 What it does: This paper studies the pattern connectivity between optimal solutions obtained from two consecutive training steps in Class Incremental Learning (CIL) and proposes methods for optimizing the connectivity path (OPC) using Fourier series and gradient projection, as well as EOPC, which is obtained by sampling and averaging in flat regions, serving as a post-processing plugin compatible with various CIL frameworks.
Optimizing NOTEARS Objectives via Topological Swaps
Chang Deng (University of Chicago), Pradeep Kumar Ravikumar
OptimizationGraph
🎯 What it does: This paper studies a class of non-convex optimization problems and proposes a bi-level optimization algorithm that optimizes the objective function by iteratively swapping pairs of nodes in the topological order of a directed acyclic graph (DAG).
Optimizing the Collaboration Structure in Cross-Silo Federated Learning
Wenxuan Bao (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
OptimizationFederated LearningImage
🎯 What it does: This paper proposes the FEDCOLLAB framework, which optimizes the collaborative structure in cross-data center federated learning through clustering based on distribution distance and data volume, thereby reducing negative transfer and improving the performance of models on each client.
Oracles & Followers: Stackelberg Equilibria in Deep Multi-Agent Reinforcement Learning
Matthias Gerstgrasser (Harvard University), David C. Parkes
Meta LearningReinforcement LearningTabularSequential
🎯 What it does: This paper proposes a general framework that utilizes multi-agent reinforcement learning to solve the Stackelberg equilibrium in Markov games, unifying various previous methods under this framework.
Orthogonality-Enforced Latent Space in Autoencoders: An Approach to Learning Disentangled Representations
Jaehoon Cha (Rutherford Appleton Laboratory), Jeyan Thiyagalingam (Rutherford Appleton Laboratory)
Representation LearningAuto EncoderImage
🎯 What it does: A framework for unsupervised factor separation based on deterministic autoencoders (DAE) is proposed, achieving linear decoupling through Euler encoding and orthogonal latent space.
Oscillation-free Quantization for Low-bit Vision Transformers
Shih-yang Liu, Kwang-Ting Cheng (Hong Kong University of Science and Technology)
ClassificationTransformerImage
🎯 What it does: This paper studies and addresses the issue of weight oscillation that occurs during the quantization of low-bit Vision Transformers. It proposes three techniques: Statistical weight quantization (StatsQ), confidence-guided annealing (CGA), and query-key reparameterization (QKR) to construct a non-oscillating low-bit ViT quantization model.
Out-of-Distribution Generalization of Federated Learning via Implicit Invariant Relationships
Yaming Guo (Jilin University), Yi Chang (Jilin University)
Domain AdaptationFederated LearningImage
🎯 What it does: This paper proposes the FEDIIR method, which achieves out-of-distribution (OOD) generalization for non-participating clients by implicitly aligning gradients across clients in federated learning to learn invariant relationships.
Out-of-Domain Robustness via Targeted Augmentations
Irena Gao (Stanford University), Percy Liang (Stanford University)
Domain AdaptationSupervised Fine-TuningAudio
🎯 What it does: Proposes and validates targeted augmentation by randomizing domain-dependent pseudo-features while preserving robust features to enhance the model's generalization ability on unseen domains.
Outline, Then Details: Syntactically Guided Coarse-To-Fine Code Generation
Wenqing Zheng (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
GenerationAI Code AssistantTransformerSupervised Fine-TuningText
🎯 What it does: A multi-step coarse-to-fine code generation model called ChainCoder is designed, which first generates a code outline and then details in a hierarchical manner, using an AST-decomposed layout framework and accessory sequences.
Over-parametrization via Lifting for Low-rank Matrix Sensing: Conversion of Spurious Solutions to Strict Saddle Points
Ziye Ma (University of California Berkeley), Somayeh Sojoudi (University of California Berkeley)
Recommendation SystemOptimization
🎯 What it does: This paper studies the role of over-parameterization in solving non-convex optimization problems, particularly in low-rank matrix recovery problems, proposing a method that forms an infinite hierarchy of non-convex problems through lifting techniques and the Burer-Monteiro decomposition.
Overcoming Simplicity Bias in Deep Networks using a Feature Sieve
Rishabh Tiwari (Google Research India), Pradeep Shenoy (Google Research India)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A feature selection method called SIFER is proposed, which uses an auxiliary network to alternately identify and suppress easily predictable simple features in the network, thereby encouraging the model to learn more complex and robust features.
PAC Generalization via Invariant Representations
Advait U Parulekar, Sanjay Shakkottai (University of Texas at Austin)
Tabular
🎯 What it does: This paper proposes and analyzes the use of approximate invariant representations to achieve distribution-free generalization of linear structural equation models (SEM) under finite sample conditions, and provides an upper bound on the PAC sample complexity in relation to the number of training environments and sample size.
PAC Prediction Sets for Large Language Models of Code
Adam Khakhar (University of Pennsylvania), Osbert Bastani (University of Pennsylvania)
GenerationExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper proposes a PAC prediction set method for code generation using large language models, representing the uncertain parts as 'partial programs' with holes, thereby achieving interpretable predictions with theoretical confidence.
PAC-Bayesian Generalization Bounds for Adversarial Generative Models
Sokhna Diarra Mbacke (Universite Laval), Pascal Germain (Universite Laval)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper extends PAC-Bayesian theory to adversarial generative models, providing generalization bounds based on Wasserstein distance and total variation, and validating its feasibility through experiments.
PAC-Bayesian Offline Contextual Bandits With Guarantees
Otmane Sakhi (Criteo AI Lab), Nicolas Chopin (CREST)
OptimizationReinforcement LearningTabular
🎯 What it does: A framework for offline contextual multi-armed bandit learning based on PAC-Bayes theory is proposed, which directly optimizes the generalization bound on offline data using prior information from the logging policy, ensuring that the new policy is guaranteed to outperform the original policy.
Paging with Succinct Predictions
Antonios Antoniadis (University of Twente), Bertrand Simon (IN2P3 Computing Center and CNRS)
🎯 What it does: This paper studies the learning-enhanced paging problem where only one bit of prediction can be provided for each request, and proposes deterministic and randomized algorithms for two simplified prediction settings: discard and phase.
Pairwise Ranking Losses of Click-Through Rates Prediction for Welfare Maximization in Ad Auctions
Boxiang Lyu (University of Chicago Booth School of Business), Oluwasanmi O Koyejo
Recommendation SystemOptimizationKnowledge DistillationTabular
🎯 What it does: A loss function for CTR prediction in advertising bidding is designed to enhance social welfare.
PAL: Program-aided Language Models
Luyu Gao (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Large Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Utilize large language models (LLM) to generate procedural reasoning steps for natural language questions, and execute the computation process using a Python interpreter, thereby achieving natural language reasoning.
PaLM-E: An Embodied Multimodal Language Model
Danny Driess (Robotics at Google), Pete Florence (Google Research)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A embodied multimodal language model named PaLM-E is proposed, which can directly embed continuous perceptual information such as vision and state into a large language model to achieve tasks like robot planning, visual question answering, and image captioning.
Parallel $Q$-Learning: Scaling Off-policy Reinforcement Learning under Massively Parallel Simulation
Zechu Li (Massachusetts Institute of Technology), Pulkit Agrawal (Massachusetts Institute of Technology)
Reinforcement Learning
🎯 What it does: Proposed the Parallel Q-Learning (PQL) scheme, which utilizes GPU for large-scale parallel simulation on a single workstation to achieve efficient training of offline policy learning;
Parallel Neurosymbolic Integration with Concordia
Jonathan Feldstein (University of Edinburgh), Efthymia Tsamoura (Samsung AI)
Recommendation SystemMixture of ExpertsVideoText
🎯 What it does: A parallel neural-symbolic integration framework called Concordia is proposed, which combines probabilistic logic theories with deep models, supporting arbitrary probabilistic logics (such as MLN, PSL, etc.).
Parallel Online Clustering of Bandits via Hedonic Game
Xiaotong Cheng (University of Tuebingen), Setareh Maghsudi (University of Tuebingen)
Recommendation SystemOptimizationTabular
🎯 What it does: A parallel online clustering Bandit algorithm called CLUB-HG based on hedonic games is proposed to automatically discover user clusters and share information in large-scale user environments, thereby improving recommendation quality.
Parameter-Level Soft-Masking for Continual Learning
Tatsuya Konishi (KDDI Research), Bing Liu (University of Illinois at Chicago)
ClassificationOptimizationImage
🎯 What it does: A soft mask based on parameter gradient importance (SPG) is proposed to achieve continuous learning, which can prevent catastrophic forgetting, encourage knowledge transfer, and reduce network capacity consumption.
Pareto Manifold Learning: Tackling multiple tasks via ensembles of single-task models
Nikolaos Dimitriadis (Ecole Polytechnique Federale de Lausanne), François Fleuret (University of Geneva)
ClassificationSegmentationOptimizationImage
🎯 What it does: This study investigates Pareto frontier learning in multi-task learning and proposes the Pareto Manifold Learning (PaMaL) method, which constructs a continuously adjustable Pareto frontier in a single training session by utilizing linear combinations of single-task models in the weight space.
Pareto Regret Analyses in Multi-objective Multi-armed Bandit
Mengfan Xu (Northwestern University), Diego Klabjan (Northwestern University)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper studies Pareto regret in multi-objective multi-armed bandits (MO-MAB), proposing a general Pareto regret metric suitable for both stochastic and adversarial environments, and provides corresponding theoretical upper and lower bounds.
Partial Optimality in Cubic Correlation Clustering
David Stein (TU Dresden), Bjoern Andres (TU Dresden)
OptimizationPoint Cloud
🎯 What it does: The study investigated the partial optimality conditions for the cubic clustering problem on complete graphs and implemented the corresponding testing algorithm.
Partially Observable Multi-agent RL with (Quasi-)Efficiency: The Blessing of Information Sharing
Xiangyu Liu (University of Maryland), Kaiqing Zhang (University of Maryland)
Computational EfficiencyReinforcement Learning
🎯 What it does: The research utilizes information sharing to construct computable and sample-efficient approximate equilibrium strategies in partially observable multi-agent reinforcement learning (POSG).
PASTA: Pessimistic Assortment Optimization
Juncheng Dong (Duke University), Vahid Tarokh (Duke University)
Recommendation SystemOptimizationTabular
🎯 What it does: This paper studies the optimization problem of selecting a combination of products in an offline data environment and proposes the PASTA framework based on the pessimistic principle to address the issue of insufficient data coverage.
Patch-level Contrastive Learning via Positional Query for Visual Pre-training
Shaofeng Zhang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Object DetectionSegmentationTransformerContrastive LearningImageVideo
🎯 What it does: This paper proposes a patch-level contrastive learning method based on position queries, called PQCL, which achieves patch-level contrast without the need for patch correspondence and enhances the effectiveness of contrastive learning through a cross-attention mechanism.
Patch-level Routing in Mixture-of-Experts is Provably Sample-efficient for Convolutional Neural Networks
Mohammed Nowaz Rabbani Chowdhury (Rensselaer Polytechnic Institute), Pin-Yu Chen (IBM Research)
ClassificationData SynthesisComputational EfficiencyConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: The paper demonstrates through theoretical analysis and experimental validation that using patch-level Mixture-of-Experts (pMoE) in convolutional neural networks can significantly reduce sample complexity and model complexity while maintaining or improving generalization performance.
Path Neural Networks: Expressive and Accurate Graph Neural Networks
Gaspard Michel (Ecole Polytechnique), Michalis Vazirgiannis (Ecole Polytechnique)
Graph Neural NetworkGraph
🎯 What it does: Proposes Path Neural Networks (PathNNs), which update node representations by aggregating paths originating from each node, and presents three variants: single shortest path, all shortest paths, and all simple paths (length ≤ K).
PCA-based Multi-Task Learning: a Random Matrix Approach
Malik Tiomoko (Huawei Noah's Ark Lab), Frederic Pascal
ClassificationComputational EfficiencyImageText
🎯 What it does: A PCA-based extended multi-task learning method (SPCA-MTL) based on random matrix theory is proposed, along with its theoretical error analysis and optimal label design.
Performative Recommendation: Diversifying Content via Strategic Incentives
Itay Eilat (Technion Israel Institute of Technology), Nir Rosenfeld (Technion Israel Institute of Technology)
Recommendation SystemTabular
🎯 What it does: The study introduces performative regularization in recommendation learning to actively promote content diversity by leveraging the strategic behavior of content creators.
Performative Reinforcement Learning
Debmalya Mandal (Max Planck Institute for Software Systems), Goran Radanovic (Max Planck Institute for Software Systems)
OptimizationReinforcement LearningTabular
🎯 What it does: A performative reinforcement learning framework is proposed, defining executable stable policies, and it is proven that repeatedly optimizing the regularized RL objective can converge to stable policies under smooth environmental changes.
Personalized Federated Learning under Mixture of Distributions
Yue Wu (University of California), Wei Cheng (NEC Laboratories America)
Anomaly DetectionFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes FedGMM—a framework that utilizes Gaussian Mixture Models (GMM) to model the heterogeneity of joint distributions in federated learning, achieve personalized models, and support uncertainty quantification and new sample detection.
Personalized Federated Learning with Inferred Collaboration Graphs
Rui Ye (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai AI Laboratory)
Recommendation SystemFederated LearningText
🎯 What it does: A new personalized federated learning framework called pFedGraph is proposed, which utilizes a learnable collaborative graph to determine the degree of collaboration between clients in each communication round, and optimizes the local model on the client side by aggregating the model.
Personalized Subgraph Federated Learning
Jinheon Baek (KAIST), Sung Ju Hwang (KAIST)
Federated LearningGraph Neural NetworkGraph
🎯 What it does: A personalized framework for subgraph federated learning, FED-PUB, is proposed, which can achieve similarity matching and weighted aggregation between subgraphs solely through model parameters without sharing any data, and localizes subgraph-specific weights through sparse masking.
Perturbation Analysis of Neural Collapse
Tom Tirer (Bar-Ilan University), Jonathan Niles-Weed (New York University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A constraint-free feature model constrained near a preset feature matrix is proposed and analyzed, studying the neural collapse behavior of deep networks at the termination training stage.
PFGM++: Unlocking the Potential of Physics-Inspired Generative Models
Yilun Xu (Massachusetts Institute of Technology), Tommi S. Jaakkola
GenerationData SynthesisDiffusion modelImagePhysics RelatedOrdinary Differential Equation
🎯 What it does: The PFGM++ model is proposed, extending PFGM into a multi-dimensional augmented generative framework, unifying the theory and practice of Diffusion and PFGM.
PFNs4BO: In-Context Learning for Bayesian Optimization
Samuel Müller, Frank Hutter (University of Freiburg)
OptimizationHyperparameter SearchTransformerTabular
🎯 What it does: This paper proposes using Prior-Data Fitted Networks (PFNs) as a general surrogate model for Bayesian Optimization (BO), and implements various priors (simple Gaussian processes, HEBO heuristic priors, Bayesian neural network priors) as well as user priors, input scaling, and non-greedy acquisition functions within this framework.
Phase Transitions in the Detection of Correlated Databases
Dor Elimelech (Ben Gurion University), Wasim Huleihel (Tel Aviv University)
Tabular
🎯 What it does: The study investigates the correlation detection problem between two Gaussian databases, formalizing it as a hypothesis testing problem and determining the phase transition thresholds under different asymptotic conditions.
Phase-aware Adversarial Defense for Improving Adversarial Robustness
Dawei Zhou (Xidian University), Tongliang Liu (University of Sydney)
Adversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: A joint adversarial defense method based on image phase information is proposed, combining phase-level adversarial training and amplitude preprocessing.
PINA: Leveraging Side Information in eXtreme Multi-label Classification via Predicted Instance Neighborhood Aggregation
Eli Chien (University of Illinois), Hsiang-Fu Yu (Amazon)
ClassificationTransformerText
🎯 What it does: The PINA method is proposed, which enhances extreme multi-label classification performance by leveraging label metadata and instance association signals through neighbor prediction aggregation.
Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding
Kenton Lee (Google Research), Kristina Toutanova (Google Research)
RecognitionImage TranslationTransformerVision Language ModelImageText
🎯 What it does: Developed a Pixel-to-Text model called Pix2Struct, which utilizes webpage screenshots and corresponding HTML for self-supervised pre-training to achieve unified visual language understanding.
PixelAsParam: A Gradient View on Diffusion Sampling with Guidance
Anh-Dung Dinh (University of Sydney), Chang Xu (University of Sydney)
GenerationOptimizationDiffusion modelImage
🎯 What it does: This paper proposes viewing the conditional guidance process of diffusion models as a gradient optimization problem for image pixels, and alleviates gradient conflicts through analysis and projection methods, thereby improving image quality and diversity while maintaining conditional information.
PLay: Parametrically Conditioned Layout Generation using Latent Diffusion
Chin-Yi Cheng (Google Research), Yang Li (Google Research)
GenerationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: A potential diffusion model called PLay is designed based on guideline conditions to generate typographic layouts in vector graphic space that meet user design intentions.
Poisoning Generative Replay in Continual Learning to Promote Forgetting
Siteng Kang (University of Illinois Chicago), Xinhua Zhang (University of Illinois Chicago)
GenerationAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: In the context of continual learning, a training set poisoning method utilizing input-aware backdoors has been designed, allowing for high accuracy on the current task while the subsequently generated replay data causes the model to quickly forget previous tasks.
Poisoning Language Models During Instruction Tuning
Alexander Wan (University of California Berkeley), Dan Klein (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study injects toxic samples into large language models during instruction tuning, causing the model to produce erroneous or degraded outputs when trigger words are present in the input.
Polarity Is All You Need to Learn and Transfer Faster
Qingyang Wang (Johns Hopkins University), Joshua T Vogelstein
OptimizationKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study investigates and experiments with the pre-setting and fixing of weight polarity (positive and negative signs) in deep neural networks, exploring its impact on learning speed, data efficiency, and knowledge transfer.
Policy Contrastive Imitation Learning
Jialei Huang (Tsinghua University), Yang Gao (Tsinghua University)
Robotic IntelligenceReinforcement LearningContrastive LearningSequential
🎯 What it does: The PCIL method is proposed, which improves the representation space of adversarial imitation learning through contrastive learning of different strategies and drives policy learning with cosine similarity rewards.
Policy Gradient in Robust MDPs with Global Convergence Guarantee
Qiuhao Wang (City University of Hong Kong), Marek Petrik (University of New Hampshire)
OptimizationReinforcement LearningTabular
🎯 What it does: A double-loop robust policy gradient algorithm (DRPG) is proposed for solving s-rectangular RMDP, and its global convergence guarantee is provided; at the same time, a new parameterized gradient method is designed to solve the internal robust evaluation subproblem.
Policy Mirror Ascent for Efficient and Independent Learning in Mean Field Games
Batuhan Yardim (ETH Zurich), Niao He
OptimizationReinforcement Learning
🎯 What it does: This paper proposes an algorithm based on Policy Mirror Ascent (PMA) for efficiently learning the approximate Nash equilibrium of large-scale uniform anonymous multi-agent games (Mean Field Game, MFG) in the absence of a generative model and with only a single observable trajectory.
Policy Regularization with Dataset Constraint for Offline Reinforcement Learning
Yuhang Ran (National Key Laboratory for Novel Software Technology Nanjing University), Yang Yu (National Key Laboratory for Novel Software Technology Nanjing University)
Reinforcement Learning
🎯 What it does: This paper proposes a strategy regularization method based on dataset constraints, called PRDC, for offline reinforcement learning.
Polyhedral Complex Extraction from ReLU Networks using Edge Subdivision
Arturs Berzins (SINTEF)
OptimizationComputational EfficiencyRecurrent Neural NetworkMesh
🎯 What it does: By performing edge subdivision layer by layer on the folded hyperplanes of ReLU networks instead of traditional region subdivision, a complete extraction of the polyhedral complex generated by the piecewise affine functions of the network is achieved.
Polynomial Preconditioning for Gradient Methods
Nikita Doikov (Ecole Polytechnique Federale de Lausanne), Anton Rodomanov (Universite Catholique de Louvain)
OptimizationTabular
🎯 What it does: This paper proposes a class of preconditioners based on symmetric polynomials to improve the convergence speed of gradient methods in structural nonlinear convex optimization.
Polynomial Time and Private Learning of Unbounded Gaussian Mixture Models
Jamil Arbas (McMaster University), Christopher Liaw (Google)
Safty and PrivacyComputational Efficiency
🎯 What it does: This paper proposes a general reduction from non-private learning to private learning, utilizing this reduction to achieve (ε,δ)-differentially private parameter estimation for unbounded high-dimensional Gaussian mixture models (GMMs), with both sample complexity and running time being polynomial.
Posterior Sampling for Deep Reinforcement Learning
Remo Sasso (Queen Mary University of London), Paulo Rauber (Queen Mary University of London)
Recurrent Neural NetworkReinforcement LearningAuto EncoderImage
🎯 What it does: A scalable posterior sampling deep reinforcement learning algorithm, PSDRL, is proposed, which achieves efficient exploration through latent state space modeling and Bayesian uncertainty quantification based on a model.
POUF: Prompt-Oriented Unsupervised Fine-tuning for Large Pre-trained Models
Korawat Tanwisuth (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
ClassificationDomain AdaptationTransformerPrompt EngineeringImageText
🎯 What it does: An unsupervised fine-tuning framework called POUF is proposed, which directly fine-tunes large-scale pre-trained models on unlabeled target data through prompt or parameter tuning.
PPG Reloaded: An Empirical Study on What Matters in Phasic Policy Gradient
Kaixin Wang (Technion), Shie Mannor (Technion)
Reinforcement LearningSequential
🎯 What it does: Conduct large-scale experiments on the Phasic Policy Gradient (PPG) framework to explore the key factors affecting its performance.
Practical and Matching Gradient Variance Bounds for Black-Box Variational Bayesian Inference
Kyurae Kim (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
OptimizationTabular
🎯 What it does: It is proven that the gradient variance of Black Box Variational Inference (BBVI) satisfies the ABC condition, and corresponding upper and lower bounds are provided.
Pre-computed memory or on-the-fly encoding? A hybrid approach to retrieval augmentation makes the most of your compute
Michiel de Jong (University of Southern California), William W. Cohen (Google Research)
RetrievalComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A hybrid retrieval-augmented language model LUMEN is proposed, combining pre-encoded retrieval memory with real-time on-demand encoding to reduce encoding costs and enhance question-answering performance.
Pre-training for Speech Translation: CTC Meets Optimal Transport
Phuong-Hang Le (University Grenoble Alpes), Didier Schwab (University Grenoble Alpes)
TransformerTextAudio
🎯 What it does: This paper proposes a scheme that combines CTC and OT during the pre-training phase in speech-to-text translation (ST) to reduce the modality gap between speech and text.
Predictable MDP Abstraction for Unsupervised Model-Based RL
Seohong Park (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A novel unsupervised Predictable MDP Abstraction (PMA) method is developed, which first transforms the original MDP into an abstract MDP that only allows predictable transitions through a learned latent action space, and trains a dynamic model on this abstract MDP to achieve zero-shot model-based RL;
Predicting Ordinary Differential Equations with Transformers
Sören Becker (Helmholtz Center Munich), Niki Kilbertus (Helmholtz Center Munich)
TransformerTime SeriesOrdinary Differential Equation
🎯 What it does: Using a transformer-based sequence-to-sequence model to directly recover the symbolic form of scalar first-order autonomous ODEs from a single irregular, noisy ODE trajectory.
Predicting Rare Events by Shrinking Towards Proportional Odds
Gregory Faletto (University of Southern California), Jacob Bien (University of Southern California)
TabularBiomedical Data
🎯 What it does: A new model called PRESTO is proposed to improve the probability estimation of extremely rare events by utilizing previously more common intermediate category data.
Predictive Flows for Faster Ford-Fulkerson
Sami Davies (Northwestern University), Yuyan Wang (Google Research)
SegmentationOptimizationFlow-based ModelImage
🎯 What it does: A method is proposed to warm start the Ford-Fulkerson algorithm using predicted flows, first projecting the predicted flow into a feasible flow, and then continuing to execute Ford-Fulkerson to solve the maximum flow.
Prefer to Classify: Improving Text Classifiers via Auxiliary Preference Learning
Jaehyung Kim (Korea Advanced Institute of Science and Technology), Dongyeop Kang (University of Minnesota)
ClassificationTransformerContrastive LearningText
🎯 What it does: A P2C framework is proposed to enhance text classification performance using sample pair preference learning.
PreNAS: Preferred One-Shot Learning Towards Efficient Neural Architecture Search
Haibin Wang (Alibaba Group), Xiuyu Sun (Alibaba Group)
Neural Architecture SearchTransformerImage
🎯 What it does: This paper proposes PreNAS, a NAS method that first uses zero-cost proxies to filter high-quality sub-networks and then performs one-shot shared training in a limited subspace.
Preprocessors Matter! Realistic Decision-Based Attacks on Machine Learning Systems
Chawin Sitawarin (University of California), Nicholas Carlini (Google DeepMind)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper discusses how preprocessors can significantly reduce the effectiveness of decision-based black-box attacks in practical machine learning systems, and proposes reverse and adaptive attack methods against preprocessors.
Pretraining Language Models with Human Preferences
Tomasz Korbak (University of Sussex), Ethan Perez (Anthropic)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The research directly incorporates human preference feedback during the pre-training phase of language models, aiming to guide the model from the outset to generate text that aligns with human expectations.
Pricing Experimental Design: Causal Effect, Expected Revenue and Tail Risk
David Simchi-Levi (Massachusetts Institute of Technology), Chonghuan Wang (Massachusetts Institute of Technology)
Finance Related
🎯 What it does: In response to the lack of historical sales data for new product pricing experiments, this paper proposes an optimal experimental design framework that balances causal effect estimation, revenue maximization during the experimental period, and tail risk control.
Primal and Dual Analysis of Entropic Fictitious Play for Finite-sum Problems
Atsushi Nitanda (Kyushu Institute of Technology), Taiji Suzuki (University of Tokyo)
Data SynthesisOptimizationImageStochastic Differential Equation
🎯 What it does: The Entropy Virtual Game (EFP) algorithm for finite summation problems is proposed and analyzed, with global convergence guarantees provided for both continuous and discrete time;
Principled Acceleration of Iterative Numerical Methods Using Machine Learning
Sohei Arisaka (National University of Singapore), Qianxiao Li (National University of Singapore)
OptimizationMeta Learning
🎯 What it does: A meta-learning based iterative numerical method acceleration framework is proposed, which theoretically analyzes and addresses the issue of limited iteration count increase caused by using solution error as loss, ultimately presenting a differentiable surrogate loss that directly minimizes the number of iterations.
Principled Offline RL in the Presence of Rich Exogenous Information
Riashat Islam (McGill University), John Langford (Microsoft Research)
Reinforcement LearningContrastive LearningVideo
🎯 What it does: In offline reinforcement learning, a multi-step inverse dynamics model is proposed to learn an agent-centric representation (ACRO) that contains only controllable information, and this representation is used for offline policy optimization.
Principled Reinforcement Learning with Human Feedback from Pairwise or K-wise Comparisons
Banghua Zhu (University of California), Jiantao Jiao (University of California)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: A theoretical framework for RLHF is proposed, analyzing the convergence of MLE based on comparative data under linear rewards, and exploring its impact on the policy.
Privacy-Aware Compression for Federated Learning Through Numerical Mechanism Design
Chuan Guo (Meta AI), Michael Rabbat
CompressionFederated LearningSafty and PrivacyImage
🎯 What it does: A new interpolation mechanism, called Interpolation Minimum Variance Unbiased (I-MVU) mechanism, is proposed for achieving privacy-preserving compression in federated learning.
Private Federated Learning with Autotuned Compression
Enayat Ullah (Johns Hopkins University), Sewoong Oh (Google Research)
Federated LearningSafty and PrivacyText
🎯 What it does: A communication compression scheme is proposed in federated learning that can adaptively adjust the compression rate, is compatible with secure aggregation, and satisfies differential privacy. The core idea is to achieve automatic adjustment of the compression rate through distributed mean estimation.
Private Statistical Estimation of Many Quantiles
Clément Lalanne (University of Lyon), Rémi Gribonval (University of Lyon)
Safty and PrivacyComputational EfficiencyTabular
🎯 What it does: This paper studies how to efficiently estimate multiple quantiles of a distribution under the differential privacy framework, proposing the Recursive Exponential Mechanism (RecExp) and histogram density estimation methods, and providing statistical error upper bounds for both.
Probabilistic Attention-to-Influence Neural Models for Event Sequences
Xiao Shou (Rensselaer Polytechnic Institute), Kristin Bennett
TransformerSequential
🎯 What it does: A probabilistic attention-influence neural model is proposed for timestamp-free multi-label event sequences, which can predict the next event and learn the influence set of each event type.
Probabilistic Categorical Adversarial Attack and Adversarial Training
Han Xu (Michigan State University), Jiliang Tang (Michigan State University)
ClassificationAdversarial AttackConvolutional Neural NetworkRecurrent Neural NetworkText
🎯 What it does: This study investigates adversarial attacks and defenses for discrete (categorical) data, proposing a Probability Classification Adversarial Attack (PCAA) framework, which utilizes gradient attack methods from continuous domains and adapts them to discrete domains; further, based on PCAA, an adversarial training method called PADVT is developed to enhance the robustness of classification models.
Probabilistic Concept Bottleneck Models
Eunji Kim (Seoul National University), Sungroh Yoon (Seoul National University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkGaussian SplattingImage
🎯 What it does: This paper proposes the ProbCBM (Probabilistic Concept Bottleneck Model), which introduces probabilistic embeddings into the concept bottleneck framework to model the uncertainty between concepts and categories, and explains decisions through confidence.