AAAI Conference on Artificial Intelligence Β· 696 papers
Policy-Adaptive Estimator Selection for Off-Policy Evaluation
Takuma Udagawa (Sony Group Corporation), Kei Tateno (Sony Group Corporation)
CodeRecommendation SystemOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
π― 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).
Positional Label for Self-Supervised Vision Transformer
Zhemin Zhang (Southwest Jiaotong University), Xun Gong (Southwest Jiaotong University)
CodeRepresentation LearningTransformerImage
π― 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.
π― What it does: This paper proposes a cross-system attack framework called PC-Attack, which learns graph topology knowledge using self-supervised graph contrastive learning on publicly available recommendation system data, and fine-tunes the model with only a small amount of target system data (β€10%) to generate pseudo-user poisoning attacks on the target recommendation system.
π― What it does: A concise architecture is designed that uses only a Simplified Fully Connected Network (SFCN) to predict the next elements in a user's historical behavior sequence.
PrimeNet: Pre-training for Irregular Multivariate Time Series
Ranak Roy Chowdhury (University of California San Diego), Jingbo Shang (University of California San Diego)
CodeClassificationAnomaly DetectionRepresentation LearningTransformerContrastive LearningTime SeriesBiomedical DataElectronic Health Records
π― What it does: This paper proposes PrimeNet, a self-supervised pre-training model that utilizes time-sensitive contrastive learning and time-sensitive reconstruction tasks to learn representations of irregular multivariate time series, and fine-tunes it for downstream tasks.
Principled and Efficient Motif Finding for Structure Learning of Lifted Graphical Models
Jonathan Feldstein (Bennu.AI), Efthymia Tsamoura (Samsung AI)
CodeGraph Neural NetworkGraph
π― What it does: This paper addresses the motif discovery problem in structure learning and proposes a new PRISM algorithm that can automatically identify structural patterns from data and accelerate model building.
Principled Data-Driven Decision Support for Cyber-Forensic Investigations
Soodeh Atefi (University of Houston), Aron Laszka (Pennsylvania State University)
CodeReinforcement Learning
π― What it does: A principled data-driven decision support framework based on Markov Decision Processes (MDP) is proposed to guide the selection of technical priorities in network forensic investigations.
Stephan A. Fahrenkrog-Petersen (Humboldt-Universitat zu Berlin), Matthias Weidlich (Humboldt-Universitat zu Berlin)
CodeOptimizationSafty and PrivacyTabular
π― What it does: This paper presents a study on privacy attacks against publicly available scheduling data. It first constructs a threat model for public scheduling, then defines a distance-based privacy loss metric, and analyzes the theoretical characteristics of both information-free and information-rich attacks. Subsequently, it applies the inverse scheduling problem to single-machine total weighted completion time (TWCT) scheduling, providing a constraint satisfaction solving method, complexity analysis, and validating the effectiveness of information-rich attacks through large-scale synthetic scheduling experiments.
Probabilistic Generalization of Backdoor Trees with Application to SAT
Alexander Semenov (ITMO University), Ibragim Dzhiblavi (ITMO University)
CodeOptimizationTabular
π― What it does: This paper extends the concept of strong backdoor sets (SBS) and introduces the probability backdoor tree (Ο-backdoor tree), providing its theoretical properties. It then implements an efficient search method based on evolutionary algorithms to quickly construct Ο-backdoor trees in SAT instances.
π― What it does: This paper systematically evaluates the robustness of 38 mainstream deep networks in recognizing objects under unusual poses by constructing a synthetic dataset called ObjectPose, and explores the impact of dataset size, network size, training strategies, and multiple transformations on performance.
π― What it does: A Progressive Deep Multi-View Fusion (PDMF) framework is proposed, which learns to assist complete representation during the pre-training phase and captures the consistency and complementarity between different views. In the fine-tuning phase, view-specific encoders are learned, and Multi-View Sparse Batch Normalization (MSBN) is used to achieve alignment and fusion of view-specific representations, ultimately resulting in a comprehensive multi-view representation.
Progressive Neighborhood Aggregation for Semantic Segmentation Refinement
Ting Liu (Northwestern Polytechnical University), Yanning Zhang (Beijing Jiaotong University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: This paper proposes an Advanced Neighborhood Aggregation (PNA) framework, which aims to refine the coarse predictions of semantic segmentation step by step while maintaining the overall structural integrity of the network, utilizing the spatial structural information and detail information of multi-scale features from the backbone network.
Prototypical Partial Optimal Transport for Universal Domain Adaptation
Yucheng Yang (Xi'an Jiaotong University), Jian Sun (Xi'an Jiaotong University)
CodeDomain AdaptationContrastive LearningImage
π― What it does: A unified domain adaptation method based on minimum batch prototype partial optimal transport (m-PPOT) is proposed, which re-weights the source prototypes and target samples using transport plans to distinguish between common categories and unknown categories.
Provable Detection of Propagating Sampling Bias in Prediction Models
Pavan Ravishankar (New York University), Daniel B. Neill (New York University)
CodeTabular
π― What it does: The mechanism of differential sampling bias propagation from the data stage to the prediction stage is studied, and a testable threshold is provided. The theoretical results are then validated on the COMPAS and New York City Police Department stop-and-frisk data.
π― What it does: A single-cycle decentralized proximal stochastic recursive momentum algorithm named DEEPSTORM is proposed for solving non-convex stochastic composite optimization problems.
π― What it does: A pseudo-label based conditional GAN model inverse attack method (PLG-MI) is proposed, which first generates pseudo-labels using a top-n strategy on public data to train a cGAN, and then searches for private images of specified categories in the latent space.
PUnifiedNER: A Prompting-Based Unified NER System for Diverse Datasets
Jinghui Lu (SenseTime Research), Fei Tan (SenseTime Research)
CodeRecognitionTransformerPrompt EngineeringText
π― What it does: This paper proposes a prompt-based unified named entity recognition system (PUnifiedNER) that can handle up to 37 entity types from different domains at once, enabling on-demand entity recognition.
π― What it does: A PUPS framework is proposed, which uses point-level classifiers to directly predict semantic categories and instance segmentation in point clouds, resulting in a unified panoramic segmentation outcome.
π― What it does: Proposed the Q-functional architecture, which uses state-based coefficients and action basis functions to quickly and parallelly evaluate the value of multiple actions, thereby replacing traditional policy networks;
π― What it does: A Quality-Aware Self-Training (QAST) framework is proposed to generate high-quality synthetic samples through GAN on scarce relational datasets and automatically label them with pseudo-labels to alleviate the class imbalance problem.
Query Your Model with Definitions in FrameNet: An Effective Method for Frame Semantic Role Labeling
Ce Zheng (Peking University), Baobao Chang (Peking University)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: A query-based framework AGED is proposed, using the FrameNet framework and frame element definitions as natural language queries to accomplish frame semantic role labeling.
π― What it does: A radar-image joint 3D object detection network called RADIANT is proposed, which predicts the 3D offset from radar return points to the target center through training the network, achieving the association and fusion of radar depth and monocular detection results, significantly improving depth estimation accuracy and overall detection performance.
π― What it does: A non-parametric single-view 3D face reconstruction method is proposed, using hierarchical implicit SDF and trained with pseudo 2D & 3D datasets.
π― What it does: A framework combining random walk encoding with Transformer and convolution, called Random Walk Conformer, is proposed to simultaneously capture global relationships and local subgraph patterns in graphs.
Qianyu Guo (Fudan University), Weifeng Ge (Fudan University)
CodeClassificationMeta LearningImage
π― What it does: A new Few-Shot learning framework called RankDNN is proposed, which transforms multi-class tasks into binary relevance ranking problems and utilizes vector-Kronecker product encoding of image triplets and an MLP classifier for training.
Rawlsian Fairness in Online Bipartite Matching: Two-Sided, Group, and Individual
Seyed Esmaeili (University of Maryland), John P. Dickerson (University of Maryland)
CodeOptimizationTabular
π― What it does: This paper designs an online bipartite matching algorithm that achieves Rawlsian fairness for drivers (offline party) and passengers (online party) while maintaining platform revenue.
Real or Fake Text?: Investigating Human Ability to Detect Boundaries between Human-Written and Machine-Generated Text
Liam Dugan (University of Pennsylvania), Chris Callison-Burch (University of Pennsylvania)
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper studies the boundary detection problem between machine-generated text and human-written text, using the gamified platform RoFT for annotation collection.
π― What it does: The study focuses on recovering the network structure in partially observable linear stochastic network dynamic systems using observed node time series data.
π― What it does: A recursive structure attention-guided deep super-resolution (RSAG) framework is proposed, which uses a deep contrast network (DCN) to adaptively separate high-frequency/low-frequency components, and employs a recursive structure attention (SA) block to fuse the latest depth predictions with image features, enhancing low-frequency edges through HF&LF feature fusion.
π― What it does: This paper proposes an optimization strategy based on the residual membrane potential, significantly reducing the uneven errors in the conversion from ANN to SNN.
π― What it does: An unsupervised domain adaptation framework is designed, utilizing dual image transfer in the frequency domain (NSCT) and spatial domain (histogram matching) along with multi-teacher distillation to enhance the performance of medical image segmentation in unannotated target domains.
Refined Semantic Enhancement towards Frequency Diffusion for Video Captioning
Xian Zhong (Wuhan University of Technology), Mang Ye (University of Central Florida)
CodeGenerationTransformerDiffusion modelVideoText
π― What it does: A new video subtitle generation model RSFD is proposed, which improves the generation of low-frequency words through frequency-aware diffusion and divergent semantic supervision.
Relation-Aware Language-Graph Transformer for Question Answering
Jinyoung Park (Korea University), Hyunwoo Kim
CodeTransformerLarge Language ModelTextMultimodalityGraph
π― What it does: A question answering framework called QAT (Question Answering Transformer) is proposed, which combines a language model with a knowledge graph. It constructs Meta-Path token pairs to perform relation-centric encoding of multi-hop relationships in the KG, and incorporates Relation-Aware Self-Attention (RASA) and Cross-Modal Relative Position Bias into the self-attention of the Transformer, enabling dynamic interaction and reasoning of multimodal information.
Repair Is Nearly Generation: Multilingual Program Repair with LLMs
Harshit Joshi (Microsoft), Ivan RadiΔek (Microsoft)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes a multilingual program error correction engine RING based on a large code language model (Codex), employing a three-stage prompt-based strategy of error localization, code transformation, and candidate ranking.
π― What it does: This paper proposes the introduction of a rephraser module in non-autoregressive machine translation (NAT) training, which rephrases reference sentences using the output of NAT, thereby providing a more suitable training objective for NAT.
π― What it does: This paper studies a self-supervised learning loss called DILEMMA, which enhances the shape discrimination ability of visual representations by detecting positional errors in image patches.
Daniel Halpern (Harvard University), Manuel WΓΌthrich (Harvard University)
CodeOptimizationRepresentation LearningText
π― What it does: This paper studies the problem of handling incomplete voting in committee elections on online citizen participation platforms and proposes an adaptive query algorithm to achieve fair representation.
RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL
Haoyang Li (Renmin University of China), Hong Chen (Renmin University of China)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a Text-to-SQL framework called RESDSQL that separates pattern matching from syntax parsing. It first ranks and filters database schemas using a cross-encoder, and then generates an SQL skeleton followed by filling in specific fields using a seq2seq model, significantly improving the understanding and generation of complex queries.
π― What it does: This paper proposes the Resilient Binary Neural Network (ReBNN), which suppresses weight oscillation in binary networks through parameterized scaling factors and weighted reconstruction loss, thereby enhancing training stability and performance.
π― What it does: This paper proposes a solution to the task confusion problem in category incremental learning under a dynamically scalable architecture, introducing the Task-Related Incremental Learning (TCIL) framework;
π― What it does: This paper proposes a graph reconstruction method based on adaptive spectral clustering, which reconstructs the adjacency matrix using learnable pseudo-feature weights, significantly improving the node classification performance of classical GNNs while maintaining the original structure of graphs with low homogeneity.
π― What it does: To address the problem of Data-Free Quantization (DFQ), a generative method called AdaSG based on sample adaptability is proposed. It utilizes a zero-sum game composed of a generator and a quantization network to produce synthetic samples beneficial to the quantization network, thereby restoring the performance of the quantized model without accessing real data.
Revisiting Classifier: Transferring Vision-Language Models for Video Recognition
Wenhao Wu (University of Sydney), Wanli Ouyang (University of Sydney)
CodeClassificationRecognitionTransformerVision Language ModelVideo
π― What it does: This paper proposes replacing the linear classifier in video classification with text embeddings from a pre-trained vision-language model, freezing the classifier weights, and only fine-tuning the visual encoder to enhance the model's transfer performance.
Revisiting Denoising Diffusion Probabilistic Models for Speech Enhancement: Condition Collapse, Efficiency and Refinement
Wenxin Tai (University of Electronic Science and Technology of China), Ting Zhong (University of Electronic Science and Technology of China)
CodeRestorationGenerationDiffusion modelAudio
π― What it does: The DR-DiffuSE framework is proposed, which achieves high-quality speech enhancement through a conditional diffusion model, addressing issues such as conditional collapse, low efficiency, and degradation of audio quality after enhancement.
Riemannian Local Mechanism for SPD Neural Networks
Ziheng Chen (Jiangnan University), Josef Kittler (University of Surrey)
CodeRecognitionVideo
π― What it does: A multi-scale submanifold network (MSNet) is designed and implemented, utilizing submatrix selection and logarithmic mapping to extract local geometric information in SPD networks.
RLogist: Fast Observation Strategy on Whole-Slide Images with Deep Reinforcement Learning
Boxuan Zhao (Shanghai Jiao Tong University), Wei Yang (Tencent AI Lab)
CodeClassificationOptimizationConvolutional Neural NetworkTransformerReinforcement LearningImageBiomedical Data
π― What it does: A fast observation strategy RLogist based on deep reinforcement learning is proposed for quickly locating and classifying diagnosis-related areas on Whole-slide Images (WSI).
Robust and Fast Measure of Information via Low-Rank Representation
Yuxin Dong (Xi'an Jiaotong University), Chen Li (Xi'an Jiaotong University)
CodeComputational EfficiencyTabular
π― What it does: A new information metric based on low-rank representation of R' enyi entropy is proposed, aimed at improving robustness to noise and computational efficiency.
π― What it does: Proposes RCGRL, which utilizes actively generated instrumental variables to achieve graph representation learning while removing confounding effects;
Robust Image Denoising of No-Flash Images Guided by Consistent Flash Images
Geunwoo Oh (Gwangju Institute of Science and Technology), Bochang Moon (Gwangju Institute of Science and Technology)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: This paper proposes a robust image denoising framework that utilizes pairs of flash and no-flash images to achieve high-quality denoising under inconsistent lighting conditions.
π― What it does: This paper proposes a robust cooperative multi-agent reinforcement learning framework called ROMANCE, specifically designed to address the policy disturbance problem under limited strategic opponents (limited attack attempts).
π― What it does: A single-sample brain tissue segmentation framework combining Image Alignment Style Transfer (IST) and Feature-aware Content Consistency (FCC) is proposed to address the issue of label and image space mismatch in traditional dual-model iterations.
π― What it does: This paper proposes a clustering method based on a dual-track metric, CBM, for learning robust representations against dispersed distractions such as background, color, and camera perspective in visual reinforcement learning.
Menghui Zhou (Yunnan University), Po Yang (Sheffield University)
CodeAnomaly DetectionOptimizationTabularTime SeriesBiomedical DataAlzheimer's DiseaseAgriculture Related
π― What it does: The study introduces a Robust Time Smoothing framework (RoTS) in multi-task learning, which decomposes each task into smooth and discrete parts, enabling the identification and handling of anomalous tasks while maintaining temporal similarity.
π― What it does: This paper proposes a robust camera pose regression model called RobustLoc, based on multi-view graph neural diffusion and neural differential equations, aimed at achieving high-precision localization in challenging driving environments affected by seasonal changes, weather, lighting, and dynamic object disturbances.
CodeOptimizationExplainability and InterpretabilityGraphBiomedical Data
π― What it does: A rule induction method based on linear programming is proposed to select and weight a compact and interpretable set of rules from knowledge graphs for knowledge graph completion.
RWEN-TTS: Relation-Aware Word Encoding Network for Natural Text-to-Speech Synthesis
Shinhyeok Oh (Netmarble AI Center), Insoo Oh (Netmarble AI Center)
CodeGenerationData SynthesisRecurrent Neural NetworkGraph Neural NetworkLarge Language ModelTextAudio
π― What it does: This paper proposes a Relation-Aware Word Encoding Network (RWEN) to improve the naturalness and expressiveness in text-to-speech (TTS) synthesis. The network consists of two parts: Semantic Relation Encoding (SRE) and Adjacent Word Relation Encoding (AWRE), which effectively utilize syntactic tree and contextual information at the word level.
Safe Interval Path Planning with Kinodynamic Constraints
Zain Alabedeen Ali (Moscow Institute of Physics and Technology), Konstantin Yakovlev (Federal Research Center for Computer Science and Control of Russian Academy of Sciences)
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingTabularBenchmark
π― What it does: A complete and optimal variant of SIPP (SIPP-IP) is proposed, capable of planning single-robot paths in dynamic obstacle environments while considering kinodynamic constraints.
π― What it does: A sub-quadratic time algorithm SAH (Shifting-Aware Hashing) is proposed for the reverse k maximum inner product search (RkMIPS) problem in high-dimensional data.
Scalable and Globally Optimal Generalized Lβ K-center Clustering via Constraint Generation in Mixed Integer Linear Programming
Aravinth Chembu (University of Toronto), Akshat Kumar (Singapore Management University)
CodeAnomaly DetectionOptimizationTabular
π― What it does: This paper proposes a mixed-integer linear programming (MILP) model for the generalized L1 k-center clustering problem, achieving global optimal solutions for millions of samples through constraint generation techniques; it also incorporates outlier handling in the model.
π― What it does: This paper proposes a scalable attribute graph subspace clustering algorithm (SAGSC), which learns initial representations through graph convolutional networks, then constructs a non-negative similarity matrix using a factorized self-representation matrix and kernel mapping, and finally performs spectral clustering implicitly on this matrix.
Scalable Bayesian Meta-Learning through Generalized Implicit Gradients
Yilang Zhang (University of Minnesota), Georgios B. Giannakis (University of Minnesota)
CodeOptimizationMeta LearningImage
π― What it does: This paper proposes a general implicit Bayesian meta-learning (iBaML) framework that efficiently computes the gradients of Bayesian meta-learning using implicit differentiation and the conjugate gradient method, avoiding the complexity bottleneck caused by explicit higher-order derivatives.
Andrea Cini (UniversitΓ della Svizzera italiana), Cesare Alippi (Politecnico di Milano)
CodeGraph Neural NetworkGraphTime SeriesBenchmark
π― What it does: This paper proposes a scalable spatiotemporal graph neural network (SGP) that utilizes a Deep Echo State Network (random recurrent network) to perform multi-scale encoding of each sensor's time series. Information is then propagated and aggregated in the spatial dimension using a graph shift operator, ultimately resulting in spatiotemporal embeddings for the nodes. This embedding can be computed during the preprocessing stage, after which only a lightweight multi-layer perceptron decoder needs to be trained to complete the predictions.
π― What it does: A scalable neural symbolic regularization framework NGPU is proposed, which constrains the scene graph generation model during training using negative consistency constraints, enhancing the model's consistency and generalization ability regarding knowledge.
CodeClassificationAnomaly DetectionTransformerBiomedical Data
π― What it does: A hierarchical Transformer ScatterFormer based on invariant scattering transform is proposed for patient-independent epilepsy seizure detection.
CodeGraph Neural NetworkScore-based ModelGraphBiomedical Data
π― What it does: This study investigates a scheme to enhance the learning of graphical event models (GEM) in low data environments through background knowledge.
Script, Language, and Labels: Overcoming Three Discrepancies for Low-Resource Language Specialization
Jaeseong Lee (Seoul National University), Seung-won Hwang (Seoul National University)
CodeTransformerSupervised Fine-TuningText
π― What it does: This paper proposes a solution to three types of mismatches (script, language, label) encountered when specializing multilingual pre-trained models for low-resource languages, along with corresponding solutions.
π― What it does: This paper proposes a Structure-Embedded Transformer (SEFormer) and constructs a multi-scale SEFormer network for LiDAR 3D object detection, which can retain and encode local structural information in sparse point clouds.
SegFormer: A Topic Segmentation Model with Controllable Range of Attention
Haitao Bai (Xi'an Jiaotong University), Zhou Su (Xi'an Jiaotong University)
CodeClassificationSegmentationTransformerText
π― What it does: This paper presents SegFormer, an end-to-end neural topic segmentation model that utilizes unidirectional attention blocks, a context aggregator, and topic classification loss.
π― What it does: This paper proposes an end-to-end weakly supervised semantic segmentation framework called TSCD, which is centered around Self Correspondence Distillation (SCD) and a Variation-Aware Refine Module (VARM) based on pixel variation, aimed at refining CAM-based pseudo labels.
π― What it does: A self-symmetric invertible network (SAIN) is proposed for compressed sensing image rescaling, capable of performing downsampling and simulating compression in a single forward pass, while recovering details in the reverse process.
π― What it does: A Self-Contrastive (SelfCon) learning framework is proposed, which generates multiple features through a multi-exit network under a single view and compares them, achieving supervised contrastive learning without data augmentation.
π― What it does: A self-emphasizing network (SEN) is proposed, which highlights hand and facial information through lightweight spatial and temporal self-emphasizing modules to enhance continuous sign language recognition performance.
π― What it does: This paper proposes a self-supervised skeletal sequence representation learning framework based on a three-stream structure (PSTL), which generates local masked samples through Central Space Masking (CSM) and Motion Attention Temporal Masking (MATM), and uses a cross-correlation matrix to learn the spatial-temporal relationships of the skeleton.
π― What it does: A self-supervised bidirectional learning method IA-SSGM is proposed, which utilizes the Hungarian solver to generate pseudo-labels and enhances graph matching performance through cyclic consistency constraints.
Self-Supervised Graph Learning for Long-Tailed Cognitive Diagnosis
Shanshan Wang (Anhui University), Xingyi Zhang (University of Science and Technology of China)
CodeGraph Neural NetworkContrastive LearningGraph
π― What it does: In the cognitive diagnosis task, a self-supervised graph learning framework SCD is introduced, which generates sparse views through importance-driven edge dropout on the original student-exercise interaction graph, and utilizes graph contrastive learning to assist the main task, enhancing the diagnostic effectiveness for long-tail students.
π― What it does: A self-supervised interest transfer network (SITN) is proposed for cross-domain recommendation tasks, achieving the transfer of interest knowledge across different domains through instance and cluster-level contrastive learning.
Self-Supervised Logic Induction for Explainable Fuzzy Temporal Commonsense Reasoning
Bibo Cai (Harbin Institute of Technology), Lifeng Shang (Huawei Noah's Ark Lab)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextSequential
π― What it does: A self-supervised logical induction framework called LECTER is proposed for fuzzy temporal commonsense reasoning, which can explicitly derive temporal dependencies and perform interpretability validation.
Self-Supervised Primal-Dual Learning for Constrained Optimization
Seonho Park (Georgia Institute of Technology), Pascal Van Hentenryck (Georgia Institute of Technology)
CodeOptimization
π― What it does: A self-supervised dual learning (Primal-Dual Learning, PDL) framework is proposed, which directly approximates the optimal solution of constrained optimization problems using neural networks, without the need to pre-generate or solve instances.
π― What it does: A method for semantic-aware superpixel discovery based on self-supervised visual Transformer (DINO) features is proposed, and this superpixel is used for seed expansion-based weakly supervised semantic segmentation training.
π― What it does: An advanced semi-supervised homography estimation method is proposed, capable of achieving accurate homography estimation in large baseline scenarios.
Semi-Supervised Deep Regression with Uncertainty Consistency and Variational Model Ensembling via Bayesian Neural Networks
Weihang Dai (Hong Kong University of Science and Technology), Kwang-Ting Cheng (Hong Kong University of Science and Technology)
CodeImageUltrasound
π― What it does: This paper proposes a semi-supervised deep regression framework UCVME based on Bayesian neural networks, which utilizes uncertainty consistency loss and variational model ensemble to generate high-quality pseudo-labels, enhancing the performance of regression tasks.
SEnsor Alignment for Multivariate Time-Series Unsupervised Domain Adaptation
Yucheng Wang (Nanyang Technological University), Lihua Xie (Nanyang Technological University)
CodeDomain AdaptationRecurrent Neural NetworkGraph Neural NetworkContrastive LearningTime Series
π― What it does: This paper studies unsupervised domain adaptation for multivariate time series data and proposes the SEnsor Alignment (SEA) framework, which utilizes local (sensor feature and correlation) and global (global feature) alignment to achieve cross-domain knowledge transfer.
Sequence Generation with Label Augmentation for Relation Extraction
Bo Li (Peking University), Shikun Zhang (Peking University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study explores the application of sequence generation (Seq2Seq) models to the relation extraction (RE) task, finding that directly generating relation names can leverage the semantic and associative information within those names. Subsequently, the RELA method is proposed, which enhances relation names through three automatic label augmentation strategies: paraphrasing, inquiry, and synonym retrieval using GPT-2, thereby improving model performance. Finally, an in-depth analysis of BART's attention and hidden states in the RE task is conducted.
π― What it does: This paper proposes a single-stage shadow removal network called ShadowFormer based on Transformer, which utilizes global contextual information to restore shadowed areas.
SharpSSAT: A Witness-Generating Stochastic Boolean Satisfiability Solver
Yu-Wei Fan (National Taiwan University), Jie-Hong R. Jiang (National Taiwan University)
CodeTabularBenchmark
π― What it does: This paper presents a new SSAT solver called SharpSSAT, which includes the capability to generate Skolem function proofs; it also implements a proof generation scheme for the existing ClauSSat solver.
Show, Interpret and Tell: Entity-Aware Contextualised Image Captioning in Wikipedia
Khanh Nguyen (Computer Vision Center Universitat Autonoma de Barcelona), Dimosthenis Karatzas (Computer Vision Center Universitat Autonoma de Barcelona)
CodeGenerationDomain AdaptationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: The study focuses on the task of context-aware image captioning for entities on Wikipedia images, utilizing images, paragraphs, and descriptions to generate captions that are consistent with specific contexts.
π― What it does: This paper proposes SHUNIT, a method for unpaired image-to-image translation that achieves 'style harmonization' to convert source domain images into target domain styles while maintaining semantic consistency.
Stefano Fiorini (University of Milano-Bicocca), Enza Messina (University of Milano-Bicocca)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes SigMaNet, a spectral convolutional neural network capable of handling both directed and undirected graphs with edge weights that can be positive or negative.
π― What it does: Synthesize high-resolution 3D indoor scenes from a small number of indoor images (RGB-D or single RGB), and generate consistent images and videos under large viewpoint changes;
π― What it does: This paper proposes a simple and efficient heterogeneous graph neural network, SeHGNN, which utilizes mean aggregation to precompute neighbor information and integrates the semantic features of various long meta-paths through a single-layer structure using Transformer, to achieve node representation learning.
Simultaneously Updating All Persistence Values in Reinforcement Learning
Luca Sabbioni (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
CodeReinforcement LearningTabular
π― What it does: This paper proposes a new All-Persistence Bellman Operator that can simultaneously update action value functions across all time scales in a single experience transfer, and based on this, extends classical Q-learning and DQN to obtain Persistent Q-learning and Persistent DQN.
SKDBERT: Compressing BERT via Stochastic Knowledge Distillation
Zixiang Ding (Meituan), Wei Lin
CodeCompressionKnowledge DistillationTransformerLarge Language ModelText
π― What it does: This paper proposes a Stochastic Knowledge Distillation (SKD) method for compressing the BERT language model, resulting in a smaller and faster SKDBERT.
π― What it does: A neural symbolic model named SKIER is proposed to recognize emotions in multi-party dialogues, explicitly modeling discourse relations and integrating symbolic knowledge.