AAAI 2023 Papers — Page 11
AAAI Conference on Artificial Intelligence · 1578 papers
OctFormer: Efficient Octree-Based Transformer for Point Cloud Compression with Local Enhancement
Mingyue Cui (Sun Yat-sen University), Huang Kai
CompressionTransformerPoint Cloud
🎯 What it does: Using the octree-based Transformer OctFormer, combined with non-overlapping context windows and shared multi-head self-attention, to achieve efficient point cloud compression;
ODE-RSSM: Learning Stochastic Recurrent State Space Model from Irregularly Sampled Data
Zhaolin Yuan (University of Science and Technology Beijing), Hong-Ning Dai (Hong Kong Baptist University)
Recurrent Neural NetworkAuto EncoderTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: This study investigates how to achieve system identification and prediction in input-output systems with irregular sampling using a continuous-time random state space model (ODE-RSSM).
Off-Policy Proximal Policy Optimization
Wenjia Meng (Shandong University), Yilong Yin (Zhejiang University)
OptimizationReinforcement LearningSequential
🎯 What it does: A variant of PPO based on offline data (Off-Policy PPO) is proposed, which significantly improves sample efficiency by designing a new clipped approximate objective for the safe utilization of offline experiences.
Offline Imitation Learning with Suboptimal Demonstrations via Relaxed Distribution Matching
Lantao Yu (Stanford University), Stefano Ermon (Stanford University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabularBenchmark
🎯 What it does: A framework for offline imitation learning called RelaxDICE is proposed, which can learn high-performance policies in situations with only a few expert demonstrations and a large number of sub-optimal demonstrations.
Offline Quantum Reinforcement Learning in a Conservative Manner
Zhihao Cheng (University of Sydney), Dacheng Tao (JD Explore Academy)
Reinforcement LearningTabular
🎯 What it does: The first offline quantum reinforcement learning algorithm, CQ2L, is proposed, which learns the Q-value function using variational quantum circuits without any interaction with the environment.
OMPQ: Orthogonal Mixed Precision Quantization
Yuexiao Ma (Xiamen University), Rongrong Ji
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A mixed precision quantization method based on network orthogonality, OMPQ, is proposed, which utilizes the orthogonality metric ORM to quickly generate optimal bit-width configurations.
On Error and Compression Rates for Prototype Rules
Omer Kerem (Ben-Gurion University of the Negev), Roi Weiss (Ariel University)
ClassificationCompressionImage
🎯 What it does: This study investigates the error and compression rate of prototype learning rules in non-parametric multi-class classification, providing upper bounds for the error and compression rate of OptiNet in Euclidean space. A novel lossless compression scheme, ProtoComp, is proposed, which can further compress the prototype set without changing the error rate. Additionally, it is proven that the compression rate can be improved under the condition of geometric margin.
On Generalized Degree Fairness in Graph Neural Networks
Zemin Liu (National University of Singapore), Yuan Fang (Singapore Management University)
Graph Neural NetworkGraph
🎯 What it does: A framework called DegFairGNN is proposed to address the fairness issue based on node degree in graph neural networks, defining a general degree bias.
On Grounded Planning for Embodied Tasks with Language Models
Bill Yuchen Lin (University of Southern California), Xiang Ren (Sea AI Lab)
Robotic IntelligenceTransformerLarge Language ModelTabular
🎯 What it does: Proposes the G-PlanET task, allowing language models to generate executable step-by-step plans based on given high-level goals and environmental tables.
On Instance-Dependent Bounds for Offline Reinforcement Learning with Linear Function Approximation
Thanh Nguyen-Tang (Johns Hopkins University), Raman Arora (UC Santa Barbara)
Reinforcement Learning
🎯 What it does: The Bootstrapped and Constrained Pessimistic Value Iteration (BCP-VI) algorithm is proposed for leveraging instance-dependent learning rates using linear feature approximation in offline reinforcement learning.
On Manipulating Weight Predictions in Signed Weighted Networks
Tomasz Lizurej (University of Warsaw), Stefan Dziembowski (University of Warsaw)
Graph
🎯 What it does: This paper studies the feasibility and difficulty of malicious manipulation of the FGA-based trust prediction model in weighted directed signed networks.
On Solution Functions of Optimization: Universal Approximation and Covering Number Bounds
Ming Jin (Virginia Tech), Ruoxi Jia (Virginia Tech)
Optimization
🎯 What it does: This study investigates the expressive power and learnability of convex optimization solution functions, proving that linear programming/quadratic programming solution functions can achieve global approximation in Sobolev spaces. It constructs deep optimization networks to achieve error-free polynomial approximation and provides upper bounds on empirical covering numbers.
On the Calibration and Uncertainty with Pólya-Gamma Augmentation for Dialog Retrieval Models
Tong Ye (University of Science and Technology of China), Jing Xiao (Ping An Technology)
RetrievalTransformerLarge Language ModelText
🎯 What it does: This paper proposes an efficient dialogue retrieval model uncertainty and calibration framework PG-DRR, which uses a Pólya-Gamma enhanced Gaussian process layer to replace traditional dense layers, achieving more reliable confidence and calibration.
On the Complexity of PAC Learning in Hilbert Spaces
Sergei Chubanov (Bosch Center for Artificial Intelligence)
ClassificationOptimization
🎯 What it does: An algorithm for PAC learning convex polyhedra in Hilbert space is proposed, which can learn using polyhedra as classifiers under given error and confidence levels.
On the Connection between Invariant Learning and Adversarial Training for Out-of-Distribution Generalization
Shiji Xin (Peking University), Yisen Wang (Peking University)
ClassificationDomain AdaptationAdversarial AttackHyperparameter SearchGenerative Adversarial NetworkImage
🎯 What it does: A Domain-wise Adversarial Training (DAT) method is proposed, which suppresses domain-related noise features by introducing domain-specific perturbations in each training domain, thereby enhancing the model's generalization ability on out-of-distribution (OOD) data.
On the Effectiveness of Parameter-Efficient Fine-Tuning
Zihao Fu (University of Cambridge), Nigel Collier (Alibaba Group)
OptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This paper conducts a systematic analysis of Parameter-Efficient Fine-Tuning (PEFT) methods, proposes a unified sparse fine-tuning model framework, and provides theoretical upper bounds on the stability and generalization error of sparsity. It also identifies the issue of projection discontinuity in projection-based PEFT methods and proposes the Second-Order Approximation Method (SAM) to select adjustable parameters. Finally, large-scale experiments were conducted on the GLUE and SuperGLUE tasks using RoBERTa-base.
On the Expressive Flexibility of Self-Attention Matrices
Valerii Likhosherstov (University of Cambridge), Adrian Weller (University of Cambridge)
Transformer
🎯 What it does: This paper analyzes the expressive power of the self-attention matrix from a theoretical perspective, proving that the approximation of any right random matrix is NP-hard, and provides a construction method for approximation in the case of sparse matrices, where the query dimension only needs to be set to a logarithmic level.
On the Sample Complexity of Representation Learning in Multi-Task Bandits with Global and Local Structure
Alessio Russo (KTH Royal Institute of Technology), Alexandre Proutiere (KTH Royal Institute of Technology)
OptimizationRepresentation Learning
🎯 What it does: This paper studies the sample complexity of optimal arm identification in multi-task weighted multi-armed bandits (MAB) when all tasks share the same optimal representation while the predictor is task-specific.
On the Sample Complexity of Vanilla Model-Based Offline Reinforcement Learning with Dependent Samples
Mustafa O. Karabag (University of Texas at Austin), Ufuk Topcu (University of Texas at Austin)
Reinforcement Learning
🎯 What it does: This paper studies the sample complexity of the 'conventional' method for estimating models in model-free offline reinforcement learning with dependent samples, providing a polynomial upper bound.
On the Stability and Generalization of Triplet Learning
Jun Chen (Huazhong Agricultural University), Feng Zheng (Southern University of Science and Technology)
RecognitionRetrievalImageStochastic Differential Equation
🎯 What it does: This paper studies the stability and generalization ability of triplet learning, proposing generalization guarantees based on stability analysis and establishing high-probability generalization bounds for triplet learning algorithms that satisfy uniform stability.
On the Vulnerability of Backdoor Defenses for Federated Learning
Pei Fang (Tongji University), Jinghui Chen (Pennsylvania State University)
Federated LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A new federated learning backdoor attack method (F3BA) is proposed, which successfully implants a backdoor by flipping a small number of low-importance weights in the local model and jointly optimizing the trigger, without significantly affecting normal accuracy.
On Total-Order HTN Plan Verification with Method Preconditions – An Extension of the CYK Parsing Algorithm
Songtuan Lin (Australian National University), Pascal Bercher (Australian National University)
Benchmark
🎯 What it does: This paper proposes a TO HTN planning verification method based on the extended CYK algorithm, which can handle method preconditions.
On Undisputed Sets in Abstract Argumentation
Matthias Thimm (University of Hagen)
🎯 What it does: Two new semantics are proposed - undisputed set and strongly undisputed set - to weaken acceptability judgments in abstract argumentation frameworks, addressing the issue of unacceptability caused by irrelevant attacks such as self-attacks.
One Is All: Bridging the Gap between Neural Radiance Fields Architectures with Progressive Volume Distillation
Shuangkang Fang (Beihang University), Shuchang Zhou (Megvii Research)
Knowledge DistillationRepresentation LearningNeural Radiance FieldImage
🎯 What it does: A Progressive Volume Distillation (PVD) method is proposed, enabling arbitrary one-to-one conversions between different NeRF architectures (such as MLP, sparse tensors, low-rank tensors, hash tables, and their combinations), allowing for quick adaptation to different task requirements in later stages.
One-for-All: Proposal Masked Cross-Class Anomaly Detection
Xincheng Yao (Shanghai Jiao Tong University), Zhenyu Liu (Shanghai Jiao Tong University)
Anomaly DetectionTransformerAuto EncoderImage
🎯 What it does: A cross-category and multi-category anomaly detection method PMAD based on patch-level reconstruction and prototype-guided proposal masking is proposed.
One-Shot Replay: Boosting Incremental Object Detection via Retrospecting One Object
Dongbao Yang (Institute of Information Engineering, Chinese Academy of Sciences), Weiping Wang (Institute of Information Engineering, Chinese Academy of Sciences)
Object DetectionData SynthesisKnowledge DistillationImage
🎯 What it does: A One-Shot Replay method is proposed, which synthesizes samples in new data by storing only one cropped object for each old category and using a copy-paste approach to achieve incremental object detection.
Online Hyperparameter Optimization for Class-Incremental Learning
Yaoyao Liu (Max Planck Institute for Informatics), Qianru Sun (Singapore Management University)
OptimizationKnowledge DistillationHyperparameter SearchReinforcement LearningImage
🎯 What it does: An online MDP framework is designed to automatically adjust key hyperparameters in class-incremental learning through Exp3 online learning, achieving adaptive control of the balance between stability and plasticity.
Online Noisy Continual Relation Learning
Guozheng Li (Southeast University), Wenjun Ke (Beijing Institute of Computer Technology and Application)
TransformerContrastive LearningText
🎯 What it does: This paper proposes a continuous relation extraction framework S6 aimed at online task-free scenarios with noisy annotations, capable of continuous learning from streaming data while avoiding catastrophic forgetting.
Online Platforms and the Fair Exposure Problem under Homophily
Jakob Schoeffer (Karlsruhe Institute of Technology), Marc Juarez (University of Edinburgh)
Recommendation SystemOptimizationText
🎯 What it does: This study proposes the 'fair exposure' problem, constructs a temporal model considering group homogeneity in dissemination, and analyzes how platforms can maximize user clicks and likes under limited intervention while satisfying fair exposure constraints. It also provides theoretical solutions for fair-unrelated and fair-constrained optimization and evaluates the 'fairness cost' caused by fair constraints.
Online Random Feature Forests for Learning in Varying Feature Spaces
Christian Schreckenberger (University of Mannheim), Heiner Stuckenschmidt (Old Dominion University)
ClassificationExplainability and InterpretabilityComputational EfficiencyTextTabular
🎯 What it does: An Online Random Feature Forest (ORF V) is proposed to handle data streams with Variable Feature Space (VFS), maintaining a shallow decision stump forest for each feature that appears, and performing feature statistical updates, forest and tree-level pruning, online weight updates, and forest reconstruction with each incoming instance.
Online Reinforcement Learning with Uncertain Episode Lengths
Debmalya Mandal (Max Planck Institute for Software Systems), Rupak Majumdar (Max Planck Institute for Software Systems)
Reinforcement Learning
🎯 What it does: This paper studies the reinforcement learning problem with uncertain episode lengths, equating it to the infinite expectation solution of general discounted rewards, and proposes a universal discount algorithm based on UCB-VI;
Online Semi-supervised Learning with Mix-Typed Streaming Features
Di Wu (Southwest University), Yi He (Old Dominion University)
ClassificationAnomaly DetectionGaussian SplattingTabular
🎯 What it does: A framework for online semi-supervised learning with mixed-type streaming features is proposed, and an online algorithm OSLMF based on Gaussian Copula and density peak clustering is designed.
Online Symbolic Regression with Informative Query
Pengwei Jin (Institute of Computing Technology, Chinese Academy of Sciences), Yunji Chen (Institute of Computing Technology, Chinese Academy of Sciences)
TransformerContrastive LearningTabular
🎯 What it does: Proposes the QUOSR query framework, which assists online symbolic regression by actively generating data points with high information content.
Online Tuning for Offline Decentralized Multi-Agent Reinforcement Learning
Jiechuan Jiang (Peking University), Zongqing Lu (Peking University)
Reinforcement LearningAuto EncoderTabular
🎯 What it does: An Online Transfer Correction (OTC) method is proposed, which utilizes limited online experience to correct the differences in transfer dynamics in offline multi-agent reinforcement learning, thereby enhancing the performance of the learned distributed policy during the online fine-tuning phase.
Only a Few Classes Confusing: Pixel-Wise Candidate Labels Disambiguation for Foggy Scene Understanding
Liang Liao (Nanyang Technological University), Shin'ichi Satoh (National Tsing Hua University)
SegmentationDomain AdaptationAutonomous DrivingContrastive LearningImage
🎯 What it does: This paper proposes to improve the quality of pseudo-labels in self-supervised domain adaptation by using a candidate label set (CLS) instead of traditional single hard labels.
Open-Ended Diverse Solution Discovery with Regulated Behavior Patterns for Cross-Domain Adaptation
Kang Xu (Fudan University), Wei Li (Fudan University)
Domain AdaptationReinforcement LearningSequential
🎯 What it does: A set of diverse and constrained policies was trained to achieve efficient transfer in environments with dynamic differences.
Open-Vocabulary Multi-Label Classification via Multi-Modal Knowledge Transfer
Sunan He (Shenzhen University), Shu-Tao Xia (Tsinghua University)
ClassificationKnowledge DistillationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This study investigates open vocabulary multi-label classification, utilizing multimodal knowledge transfer to recognize unseen labels.
Opinion Optimization in Directed Social Networks
Haoxin Sun (Fudan University), Zhongzhi Zhang (Fudan University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies how to minimize the average equilibrium opinion under the Friedkin-Johnsen model in unweighted directed social networks by changing the intrinsic opinions of k nodes to 0.
Opposite Online Learning via Sequentially Integrated Stochastic Gradient Descent Estimators
Wenhai Cui (Zhongtai Securities Institute for Financial Studies Shandong University), Xiaodong Yan (Shandong National Center for Applied Mathematics)
OptimizationSequentialStochastic Differential Equation
🎯 What it does: This paper proposes a framework for 'opposite online learning' based on Stochastic Gradient Descent (SGD) and the Two-Armed Bandit (TAB) process for one-sided and two-sided hypothesis testing on streaming data.
OPT-GAN: A Broad-Spectrum Global Optimizer for Black-Box Problems by Learning Distribution
Minfang Lu (University of Jinan), Lin Wang (University of Jinan)
OptimizationGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes a global black-box optimizer OPT-GAN based on GAN, which can find the global optimal solution in multi-dimensional multi-peak functions by learning and continuously updating the target distribution.
Optimal Decision Diagrams for Classification
Alexandre M. Florio (Polytechnique Montreal), Thibaut Vidal (Polytechnique Montreal)
ClassificationOptimizationTabular
🎯 What it does: An optimal decision diagram (ODD) training method based on mixed-integer linear programming (MILP) is proposed, along with a two-stage training process (heuristic construction + MILP optimization), which is also extended to constraints such as fairness, simplicity, and stability.
Optimal Pathfinding on Weighted Grid Maps
Mark Carlson (Monash University), Morteza Ebrahimi (University of Tehran)
Optimization
🎯 What it does: This paper proposes the Weighted Jump Point Search (JPSW), an optimal online path planning algorithm implemented on weighted grid maps.
Optimal Pricing Schemes for Identical Items with Time-Sensitive Buyers
Zhengyang Liu (Beijing Institute of Technology), Zihe Wang (Renmin University of China)
Optimization
🎯 What it does: This study investigates the pricing problem of homogeneous goods in the presence of time-sensitive buyers, providing a theoretical analysis and algorithm implementation for the optimal pricing scheme.
Optimal Sparse Recovery with Decision Stumps
Kiarash Banihashem (University of Maryland), Max Springer (University of Maryland)
OptimizationTabular
🎯 What it does: This paper studies the optimal sparse recovery problem for feature selection using single-layer decision trees (decision stumps) and provides tight finite sample bounds for feature selection in linear regression.
Optimal Sparse Regression Trees
Rui Zhang (Duke University), Cynthia Rudin (University of British Columbia)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: A provably optimal sparse regression tree (OSRT) algorithm is proposed, which can globally optimally construct interpretable regression trees without depth limitations.
Optimism in Face of a Context:Regret Guarantees for Stochastic Contextual MDP
Orin Levy (Tel Aviv University), Yishay Mansour (Google Research)
OptimizationReinforcement Learning
🎯 What it does: A regret minimization algorithm for the Contextual Markov Decision Process (CMDP) is proposed, analyzing three different settings: known dynamics, unknown but context-independent dynamics, and unknown dynamics that depend on context.
Optimistic Whittle Index Policy: Online Learning for Restless Bandits
Kai Wang (Harvard University), Milind Tambe (Google Research)
OptimizationReinforcement LearningBiomedical Data
🎯 What it does: An online learning algorithm UCWhittle based on the Whittle index is proposed, which can simultaneously learn unknown transition probabilities and execute approximately optimal scheduling strategies in RMAB (Random Multi-Armed Bandit).
Optimizing Multiple Simultaneous Objectives for Voting and Facility Location
Yue Han (Rensselaer Polytechnic Institute), Elliot Anshelevich (Rensselaer Polytechnic Institute)
Optimization
🎯 What it does: The research addresses the problem of approximating multiple l-centroid objectives in facility location and spatial social choice optimization.
Orders Are Unwanted: Dynamic Deep Graph Convolutional Network for Personality Detection
Tao Yang (Sun Yat-sen University), Qifan Wang (Meta AI)
ClassificationRecommendation SystemGraph Neural NetworkSupervised Fine-TuningTextGraph
🎯 What it does: A dynamic depth graph convolutional network (D-DGCN) is proposed, which dynamically integrates information from multiple unordered posts on social media to generate user personality profiles and perform multi-label personality prediction.
Out-of-Distribution Generalization by Neural-Symbolic Joint Training
Anji Liu (Peking University), Yitao Liang (Peking University)
Domain AdaptationExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkImageSequential
🎯 What it does: A neural-symbolic joint training framework NTOC is proposed, which can learn generalizable neural features and symbolic rules simultaneously without prior symbolic knowledge, addressing the OOD generalization problem.
Overcoming Concept Shift in Domain-Aware Settings through Consolidated Internal Distributions
Mohammad Rostami (Information Sciences Institute, University of Southern California), Aram Galstyan (Information Sciences Institute, University of Southern California)
Domain AdaptationAuto EncoderImageMultimodality
🎯 What it does: A sequential model adaptation method based on internal distribution, SMAUI, is proposed to address the problem of concept drift in the absence of available source data.
Overcoming Forgetting in Fine-Grained Urban Flow Inference via Adaptive Knowledge Replay
Haoyang Yu (University of Electronic Science and Technology of China), Fan Zhou (University of Electronic Science and Technology of China)
Convolutional Neural NetworkTime Series
🎯 What it does: The CUFAR framework is proposed to achieve continuous learning for fine-grained urban traffic inference.
PAC Learning and Stabilizing Hedonic Games: Towards a Unifying Approach.
Simone Fioravanti (Gran Sasso Science Institute), Giovanna Varricchio (Goethe-Universitat)
🎯 What it does: This paper studies the learnability and stability issues of Hedonic Games (HG) within the PAC learning framework, extending the known classes of learnable/stable HGs and attempting to provide unified structural conditions for learnability/stability.
Painterly Image Harmonization in Dual Domains
Junyan Cao (Shanghai Jiao Tong University), Li Niu (Shanghai Jiao Tong University)
Image TranslationImage HarmonizationGenerative Adversarial NetworkImage
🎯 What it does: This study investigates the harmonization problem of painterly images and photographic foregrounds, proposing a dual-domain network PHDNet to achieve bidirectional harmonization in both spatial and frequency domains.
Panoramic Video Salient Object Detection with Ambisonic Audio Guidance
Xiang Li (Carnegie Mellon University), Bhiksha Raj (Bytedance Inc.)
Object DetectionKnowledge DistillationTransformerVideoMultimodalityAudio
🎯 What it does: This paper proposes a framework for salient object detection using panoramic video and omnidirectional audio;
ParaFormer: Parallel Attention Transformer for Efficient Feature Matching
Xiaoyong Lu (Southeast University), Songlin Du (Southeast University)
Computational EfficiencyGraph Neural NetworkTransformerImage
🎯 What it does: This paper proposes ParaFormer and its variant ParaFormer-U, which are designed for sparse feature matching tasks with a parallel attention mechanism and waveform position encoding, and incorporates attention pooling into the U-Net architecture to achieve efficient and accurate feature matching.
Parameter-Efficient Model Adaptation for Vision Transformers
Xuehai He (University of California Santa Cruz), Xin Eric Wang (Microsoft Research)
ClassificationTransformerImageBenchmark
🎯 What it does: This paper studies how to efficiently adapt large pre-trained Vision Transformer (ViT) models for image classification tasks with very few trainable parameters.
Parameterized Algorithms for Colored Clustering
Leon Kellerhals (Technische Universität Berlin), Rolf Niedermeier (Technische Universität Berlin)
🎯 What it does: This paper studies the clustering problem of colored graphs (Colored Clustering), providing parameterized complexity results under several parameters, along with corresponding algorithms and lower bounds.
Parametric Surface Constrained Upsampler Network for Point Cloud
Pingping Cai (University of South Carolina), Song Wang (University of South Carolina)
RestorationGenerationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: A point cloud upsampling network based on parametric surface constraints is proposed, capable of generating high-density, smooth point clouds from sparse point clouds, and can also be transferred to point cloud completion tasks.
PaRot: Patch-Wise Rotation-Invariant Network via Feature Disentanglement and Pose Restoration
Dingxin Zhang (University of Sydney), Weidong Cai (University of Sydney)
ClassificationSegmentationPose EstimationPoint Cloud
🎯 What it does: A Patch-wise Rotation-Invariant network called PaRot is proposed, which achieves robust point cloud classification and segmentation against arbitrary rotations through feature decoupling and pose recovery.
Partial-Label Regression
Xin Cheng (Chongqing University), Bo An (Nanyang Technological University)
TabularBenchmark
🎯 What it does: The Partial-Label Regression (PLR) problem is proposed, and three solutions are provided: average loss, minimum loss identification method, and progressive identification method;
Participatory Budgeting Designs for the Real World
Roy Fairstein (Ben-Gurion University of the Negev), Kobi Gal (Ben-Gurion University of the Negev)
Tabular
🎯 What it does: Experimental evaluation of voting formats and aggregation rules in participatory budgeting (PB), analyzing the impact of different voting methods on user experience, outcome stability, and social welfare.
Partitioning Friends Fairly
Lily Li (University of Toronto), Nisarg Shah (University of Toronto)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the fair allocation problem of dividing n agents into k approximately equal-sized groups in social networks, with the goal of simultaneously satisfying both core and envy-free fairness guarantees.
PASS: Patch Automatic Skip Scheme for Efficient Real-Time Video Perception on Edge Devices
Qihua Zhou (Hong Kong Polytechnic University), Jingren Zhou (Alibaba Group)
Object TrackingPose EstimationComputational EfficiencyConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: This paper proposes the Patch Automatic Skip Scheme (PASS), a task-agnostic automatic patch skipping strategy that accelerates real-time video perception tasks on edge devices.
PaTeCon: A Pattern-Based Temporal Constraint Mining Method for Conflict Detection on Knowledge Graphs
Jianhao Chen (Nanjing University), Yuzhong Qu (Nanjing University)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes an automatic mining method based on structural patterns, called PaTeCon, for automatically generating temporal constraints and detecting temporal conflicts from knowledge graphs.
PATRON: Perspective-Aware Multitask Model for Referring Expression Grounding Using Embodied Multimodal Cues
Md Mofijul Islam (University of Virginia), Tariq Iqbal (University of Virginia)
Object DetectionRepresentation LearningConvolutional Neural NetworkTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes a perspective-oriented multi-task learning model PATRON, designed for joint localization of relationships and target objects in embodied environments through linguistic and non-linguistic cues (gaze, pointing gestures).
PDFormer: Propagation Delay-Aware Dynamic Long-Range Transformer for Traffic Flow Prediction
Jiawei Jiang (Beihang University), Jingyuan Wang (Renmin University of China)
TransformerTime Series
🎯 What it does: The PDFormer model is proposed, which utilizes spatial self-attention to capture dynamic, long-range spatial dependencies and explicitly models the time delay of traffic information propagation to achieve accurate traffic flow prediction.
PDRF: Progressively Deblurring Radiance Field for Fast Scene Reconstruction from Blurry Images
Cheng Peng (Johns Hopkins University), Rama Chellappa (Johns Hopkins University)
RestorationNeural Radiance FieldImage
🎯 What it does: A method called Progressively Deblurring Radiance Field (PDRF) is designed to quickly reconstruct high-quality radiance fields from blurred images and achieve visual rendering.
PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers
Xiaoyi Dong (University of Science and Technology of China), Baining Guo (Microsoft Research Asia)
Object DetectionSegmentationTransformerAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes a new visual Transformer pre-training objective—Perceptual Codebook (PeCo), which incorporates a perceptual similarity loss based on visual Transformers into VQ-VAE training to obtain discrete visual words that better align with human perception, and uses this as a target for BERT-style Masked Image Modeling (MIM) pre-training.
Peeling the Onion: Hierarchical Reduction of Data Redundancy for Efficient Vision Transformer Training
Zhenglun Kong (Northeastern University), Yanzhi Wang (University of Texas at Austin)
Object DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: A three-layer sparsification framework called Tri-Level E-ViT is proposed, which removes redundant data at the sample, token, and attention connection levels to significantly accelerate the training and inference of Vision Transformers.
PEN: Prediction-Explanation Network to Forecast Stock Price Movement with Better Explainability
Shuqi Li (MADE by DATA), Rui Yan (Renmin University of China)
ClassificationRecommendation SystemAnomaly DetectionOptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningRecurrent Neural NetworkAuto EncoderTextTime SeriesFinance Related
🎯 What it does: A Prediction-Explanation Network (PEN) framework is proposed, which jointly models text streams and price streams and filters text through a Shared Representation Learning module (SRL) to achieve stock price trend prediction and explainability.
Periodic Multi-Agent Path Planning
Kazumi Kasaura (OMRON SINIC X Corporation), Mai Nishimura (OMRON SINIC X Corporation)
OptimizationRobotic IntelligenceSequential
🎯 What it does: Proposed and solved the Periodic Multi-Agent Path Planning (Periodic MAPP) problem, seeking a set of conflict-free trajectories that can be reused within a period to achieve high throughput of agent flows through a two-dimensional environment;
Personalized Dialogue Generation with Persona-Adaptive Attention
Qiushi Huang (University of Surrey), H Tang (ByteDance AI Lab)
GenerationTransformerLarge Language ModelText
🎯 What it does: Proposes the Persona-Adaptive Attention (PAA) framework, which uses dual encoders to encode persona and dialogue context separately, and implements dynamic weighting and masking of persona and context information through self-attention and cross-attention in the GPT-2 decoder, thereby generating responses that are both persona-compliant and contextually relevant.
Persuasion Strategies in Advertisements
Yaman Kumar (Indraprastha Institute of Information Technology Delhi), Changyou Chen (Adobe Media and Data Science Research)
ClassificationObject DetectionTransformerVision Language ModelImageTextMultimodality
🎯 What it does: The first advertising image dataset was constructed and 20 types of persuasion strategies were annotated, proposing a multimodal attention fusion model to predict these strategies.
PGSS: Pitch-Guided Speech Separation
Xiang Li (Peking University), Jing Chen (Peking University)
Generative Adversarial NetworkAudio
🎯 What it does: This work proposes a pitch-guided single-microphone speech separation framework called PGSS.
Phrase-Level Temporal Relationship Mining for Temporal Sentence Localization
Minghang Zheng (Peking University), Yang Liu (Peking University)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a Temporal Relation Mining (TRM) framework based on phrase-level temporal relations, modeling the temporal localization of video sentences using the temporal relations between sentences and phrases, and enhancing phrase-level prediction performance through consistency and exclusivity constraints.
PiCor: Multi-Task Deep Reinforcement Learning with Policy Correction
Fengshuo Bai (University of Chinese Academy of Sciences), Bo Xu (Chinese Academy of Sciences)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes the PiCor framework, which enhances sample efficiency and task generalization ability in multi-task deep reinforcement learning by separating policy optimization and policy correction into two stages.
PINAT: A Permutation INvariance Augmented Transformer for NAS Predictor
Shun Lu (Chinese Academy of Sciences), Ji Liu
Neural Architecture SearchTransformerGraphBenchmark
🎯 What it does: A NAS performance predictor named PINAT is proposed, which achieves efficient representation and prediction of network structures by incorporating partial permutation invariant embedding layers (PITE) and self-attention layers (PISA) into the Transformer architecture, using the Laplacian matrix as positional encoding.
Pixel Is All You Need: Adversarial Trajectory-Ensemble Active Learning for Salient Object Detection
Zhenyu Wu (Beihang University), Shuo Li (Case Western Reserve University)
Object DetectionGenerative Adversarial NetworkImage
🎯 What it does: A sparse point annotation method based on Adversarial Trajectory-Integrated Active Learning (ATAL) is proposed, allowing for significant object detection performance close to pixel-level training using only 10 points per image.
Pixel-Wise Warping for Deep Image Stitching
Hyeokjun Kweon (Korean Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korean Advanced Institute of Science and Technology)
Image TranslationGenerationDomain AdaptationGenerative Adversarial NetworkOptical FlowImage
🎯 What it does: A pixel-level transformation-based deep image stitching framework is proposed to address the artifact issues caused by the planar homography assumption in large disparity scenes.
PIXEL: Physics-Informed Cell Representations for Fast and Accurate PDE Solvers
Namgyu Kang (Sungkyunkwan University), Eunbyung Park (Sungkyunkwan University)
OptimizationComputational EfficiencyRepresentation LearningDrug DiscoveryPhysics RelatedOrdinary Differential Equation
🎯 What it does: A new PDE solver called PIXEL is proposed, which combines a trainable grid representation with a small MLP, utilizing automatic differentiation and traditional optimization algorithms to solve forward and inverse PDE problems.
Planning and Learning with Adaptive Lookahead
Aviv Rosenberg (Amazon Science), Gal Dalal (Nvidia Research)
Reinforcement LearningTabular
🎯 What it does: This paper proposes a strategy for dynamically selecting the planning lookahead depth based on state and value estimates in reinforcement learning, and presents two adaptive planning algorithms TLPI and QLPI, the latter of which is further extended to the deep Q network QL-DQN.
Planning for Learning Object Properties
Leonardo Lamanna (Fondazione Bruno Kessler), Paolo Traverso (Fondazione Bruno Kessler)
Object DetectionRobotic IntelligenceConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an online learning framework based on PDDL planning, enabling robots to autonomously collect training samples and train neural networks to recognize object attributes in unknown environments.
Planning with Hidden Parameter Polynomial MDPs
Clarissa Costen (Oxford Robotics Institute University of Oxford), Nick Hawes (Oxford Robotics Institute University of Oxford)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a Hidden Parameter Polynomial MDP (HP-MDP), which uses closed-form polynomial representations of transition probabilities within a Bayesian adaptive MDP framework to maintain posterior distributions, achieving more accurate Bayesian adaptive planning.
POEM: Polarization of Embeddings for Domain-Invariant Representations
Sang-Yeong Jo (Ulsan National Institute of Science and Technology), Sung Whan Yoon (Ulsan National Institute of Science and Technology)
Domain AdaptationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes the POEM method, which utilizes category embeddings and domain embeddings within the same model, and achieves orthogonality between them by zeroing out their cosine similarity, thereby obtaining domain-invariant representations.
Point-Teaching: Weakly Semi-supervised Object Detection with Point Annotations
Yongtao Ge (University of Adelaide), Hao Li (Alibaba Group)
Object DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A weakly semi-supervised object detection framework called Point-Teaching is proposed, which effectively utilizes point annotations through point matching, MIL, and copy-paste augmentation.
PointCA: Evaluating the Robustness of 3D Point Cloud Completion Models against Adversarial Examples
Shengshan Hu (Huazhong University of Science and Technology), Lichao Sun
Data SynthesisAdversarial AttackPoint Cloud
🎯 What it does: Attacked a 3D point cloud completion model by designing adversarial point clouds targeting specific geometric shapes.
Pointerformer: Deep Reinforced Multi-Pointer Transformer for the Traveling Salesman Problem
Yan Jin (Huazhong University of Science and Technology), Jiang Bian (Microsoft Research Asia)
OptimizationTransformerReinforcement LearningGraph
🎯 What it does: This paper presents Pointerformer, an end-to-end deep reinforcement learning framework that utilizes a multi-pointer Transformer to solve the 2D Euclidean TSP, scalable to over 500 nodes.
Poisoning with Cerberus: Stealthy and Colluded Backdoor Attack against Federated Learning
Xiaoting Lyu (Beijing Jiaotong University), Xiangliang Zhang (University of Notre Dame)
Federated LearningAdversarial AttackImageTabularFinance Related
🎯 What it does: This paper proposes a distributed backdoor attack method called Cerberus Poisoning, which utilizes colluding malicious participants to achieve covert attacks against various defense mechanisms in federated learning through trigger fine-tuning and model bias regularization.
PolarFormer: Multi-Camera 3D Object Detection with Polar Transformer
Yanqin Jiang, Yu-Gang Jiang (Fudan University)
Object DetectionAutonomous DrivingTransformerMultimodality
🎯 What it does: This paper studies a multi-camera 3D object detection framework called PolarFormer, which utilizes polar coordinate transformation to generate Polar representations in BEV and achieves object detection through cross-view attention.
Polarization-Aware Low-Light Image Enhancement
Chu Zhou (Peking University), Boxin Shi (Peking University)
RestorationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A pipeline for low-light polarization image enhancement in the Stokes domain is proposed, and a dual-branch network is designed to separately enhance the unpolarized component and the polarization differential component, significantly restoring the degree of polarization (DoP) and the angle of polarization (AoP) of the image.
Policy-Adaptive Estimator Selection for Off-Policy Evaluation
Takuma Udagawa (Sony Group Corporation), Kei Tateno (Sony Group Corporation)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper addresses the estimator selection problem in offline policy evaluation and presents an adaptive estimator selection method based on importance fitting (PAS-IF).
Policy-Based Primal-Dual Methods for Convex Constrained Markov Decision Processes
Donghao Ying (University of California Berkeley), Zuo-Jun Shen (University of California Berkeley)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper studies the convex-constrained Markov decision process and proposes a primal-dual projection algorithm based on policy gradient, proving global convergence under nonlinear constraints and non-convex objectives.
Policy-Independent Behavioral Metric-Based Representation for Deep Reinforcement Learning
Weijian Liao (Nanjing University), Yang Yu (Nanjing University)
Representation LearningReinforcement LearningVideo
🎯 What it does: A new behavior metric called Conservative State-Action Discrepancy is proposed, which is combined with SAC as a representation learning objective to form Q²-learning;
Popularizing Fairness: Group Fairness and Individual Welfare
Andrew Estornell (Washington University in Saint Louis), Yevgeniy Vorobeychik (Washington University in Saint Louis)
OptimizationTabular
🎯 What it does: This paper proposes the popularity of group fairness machine learning models (i.e., what most people consider better) as a new metric, and presents two post-processing methods (DOS and k-QLS) to achieve both fairness and popularity in models.
Pose-Guided 3D Human Generation in Indoor Scene
Minseok Kim (Ulsan National Institute of Science and Technology), Kyungdon Joo (Ulsan National Institute of Science and Technology)
GenerationData SynthesisPose EstimationGenerative Adversarial NetworkPoint CloudMesh
🎯 What it does: A posture-guided 3D human generation framework is proposed, which can achieve geometric alignment in potential contact areas based on scene information and human posture, thereby generating 3D human models for natural interaction.
Pose-Oriented Transformer with Uncertainty-Guided Refinement for 2D-to-3D Human Pose Estimation
Han Li (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
Pose EstimationTransformerImage
🎯 What it does: This paper proposes a two-stage Transformer-based 3D human pose estimation framework, which includes Pose-Oriented Transformer (POT) and Uncertainty-Guided Refinement Network (UGRN).
Positional Label for Self-Supervised Vision Transformer
Zhemin Zhang (Southwest Jiaotong University), Xun Gong (Southwest Jiaotong University)
Representation LearningTransformerImage
🎯 What it does: This paper proposes the use of absolute and relative position labels as self-supervised tasks in Vision Transformers to enhance the model's ability to model spatial structures.
Positive Distribution Pollution: Rethinking Positive Unlabeled Learning from a Unified Perspective
Qianqiao Liang (Zhejiang University), Xiaolin Zheng (Ant Group)
ClassificationAnomaly DetectionData-Centric LearningFlow-based ModelImageTabularFinance Related
🎯 What it does: A unified perspective on positive distribution contamination is proposed, and the CoVPU model is designed to simultaneously address data imbalance, selection bias, and unknown priors in PU learning.
Post-hoc Uncertainty Learning Using a Dirichlet Meta-Model
Maohao Shen (Massachusetts Institute of Technology), Gregory Wornell (Massachusetts Institute of Technology)
Domain AdaptationAnomaly DetectionImage
🎯 What it does: A post-hoc uncertainty learning framework is proposed, which enhances uncertainty estimation on a trained base model using a Dirichlet-based Bayesian meta-model.