IEEE transactions on artificial intelligence最新文献

{"title":"Generative Models for Time Series Anomaly Detection: A Survey","authors":"Jie Cao;Jiawei Miao;Haicheng Tao;Youquan Wang;Jia Wu;Zidong Wang;Xindong Wu","doi":"10.1109/TAI.2025.3614213","DOIUrl":"https://doi.org/10.1109/TAI.2025.3614213","url":null,"abstract":"Time series anomaly detection (TSAD) is a fundamental practice in information management, aimed at identifying unusual patterns in temporal datasets. This process is critical to maintaining the integrity and reliability of systems. Recently, generative models have significantly advanced the capabilities of artificial general intelligence, presenting novel methodologies to understand and interpret complex data structures. In this review, we examine the latest advancements in applying generative models to TSAD and highlight how these models present a paradigm shift in detecting and analyzing anomalies within sequential data. In particular, we first present the background information, including definitions of key concepts, a taxonomy of anomaly types, and the distinction between generative and discriminative models in time series data. Then, we investigate a range of generative models, offering mathematical summaries of the predominant techniques in TSAD. Furthermore, we provide a summary of the datasets and propose recommendations for appropriate generative methods tailored to various application domains. Finally, we address the significant challenges in current research and propose potential directions for future study.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2253-2275"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147579040","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"EEGcMamba: EEG Clustering via State Space Model","authors":"Junfu Chen;Dechang Pi;Xiaoyi Jiang;Feng Gao;Qiao Yang;Yang Chen","doi":"10.1109/TAI.2025.3614230","DOIUrl":"https://doi.org/10.1109/TAI.2025.3614230","url":null,"abstract":"Electroencephalography (EEG) is extensively employed for supervised analysis across various scenarios, including disease diagnosis, neurocognition, and brain–computer interfaces. It has heightened the demand for effective EEG labeling. However, the manual labeling process is labor-intensive and time-consuming. Furthermore, there are a few studies that attempt to decode information from EEG in an unsupervised manner using deep learning. In this article, we present a novel end-to-end EEG clustering approach via a state space model (SSM), termed EEGcMamba. EEGcMamba introduces a universal backbone, BrainMamba, for EEG feature learning, and incorporates both a weighted instance-level contrast head and a dual-branch cluster-level contrast head for contrastive learning. Specifically, BrainMamba utilizes a slice-aware scanning mechanism to input segmented EEG slices through multiple sequences, capturing fine-grained contextual connections between slices. To mitigate risks of pushing similar EEG samples apart further in the embedding space, we introduce weight terms from the data space when calculating the instance-level contrastive losses. Furthermore, in the cluster-level contrast head, the assignment-discrimination branch accounts for the clustering distribution consistency, while the semantic-aware branch employs pseudolabeling semantics to establish group-instance discrimination. Extensive experiments on 11 EEG benchmark datasets demonstrate the superiority of EEGcMamba over existing advanced methods. The code will be available at GitHub.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2292-2306"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147579022","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Online Multilabel Streaming Feature Selection Based on Agglomeration Degree and Local-Global Correlation","authors":"Qiang Li;Wenhai Yang;Jianhua Dai","doi":"10.1109/TAI.2025.3612314","DOIUrl":"https://doi.org/10.1109/TAI.2025.3612314","url":null,"abstract":"Currently, multilabel streaming feature selection is receiving increasing attention in various applications. Considering that label correlation is a common approach in the field of multilabel streaming feature selection to improve algorithm performance. Researchers commonly explore label relationship through label clustering, a method widely adopted in theoretical and practical domains. However, existing studies only focus on the global semantic structure of the entire label set, overlooking the local semantic structure within label cluster. To address this issue, this article proposes a novel feature evaluation function that simultaneously considers agglomeration degree and local-global correlation. First, we introduce agglomeration degree to explore the local semantic structure of intra-luster labels and accordingly weight different clusters. Furthermore, we design an intra-cluster label weight calculation method that leverages the high correlation characteristics of intra-cluster labels. Additionally, this algorithm also takes into account the local relevance among labels and the global relevance among features. Finally, we integrate this feature evaluation function into the multilabel streaming feature selection framework. Through online significance analysis, online correlation analysis, and online redundancy analysis, we effectively filter irrelevant and redundant features, retaining those critical for label identification. Experimental results indicate that the proposed method outperforms other compared representative online and offline multilabel feature selection methods.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2127-2141"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147579019","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A Simple and Effective Architecture Selection Method for Differentiable Architecture Search","authors":"Boxu Chen;Le Yang;Ziwei Zheng;Fan Li;Shiji Song;Gao Huang;Chao Shen","doi":"10.1109/TAI.2025.3610384","DOIUrl":"https://doi.org/10.1109/TAI.2025.3610384","url":null,"abstract":"Although differentiable architecture search (DARTS) improves the searching efficiency of neural architecture search (NAS), the widely applied magnitude-based selection method of DARTS can frequently lead to deteriorating architectures with degenerated performance. Most existing works propose to address this issue by improving the supernet’s optimization to guarantee the applicability of the magnitude-based method, while little attention has been paid to the selection criterion to obtain the final architecture. In this brief, we introduce a novel, simple, and effective architecture selection method, Manda (Magnitudes and activations), which estimates the contribution of an operation in an optimized supernet by both its architecture parameter’s magnitude and corresponding generated activation. Notably, Manda can effectively address the notorious degeneration issue in DARTS without any modification of the supernet’s optimization procedure, indicating the instability in DARTS can be attributed to the widely applied magnitude-based selection method. The experimental results on both NAS-Bench-201 and DARTS search spaces show the effectiveness of our method.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2412-2422"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147578990","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Modality-Specific Knowledge Distillation With Wasserstein Distance Minimization for Vision-Language Pretrained Models","authors":"Lung-Hao Lee;Hsiu-Min Shih;Jin Wang","doi":"10.1109/TAI.2025.3613686","DOIUrl":"https://doi.org/10.1109/TAI.2025.3613686","url":null,"abstract":"The vast costs of storage and computational resources raise obstacles to deploying vision-language pretrained models (VL-PTMs) on resource-constrained devices for services computing. To compress these models, recent studies have suggested using knowledge distillation (KD) to train a compact student model under the supervision of a complicated teacher model. A critical issue of the KD methods is one-to-one layer mapping, where each student layer can be distilled only by one specific teacher layer. In addition, different modality features contain different amounts of knowledge, which may lead to an imbalanced distribution of different modalities such that the dominant modality will overwhelm the minor modalities. This study proposes a modality-specific knowledge distillation (MKD) by minimizing the Wasserstein distance for a vision-language pretrained model. Instead of empirically finding the one-to-one layer mapping, the proposed MKD performs a many-to-many layer mapping strategy, where each layer of the student model has a chance to learn from different intermediate layers of the teacher model. To balance the distribution between modalities, we used two extra inputs (text-only and image-only), and two auxiliary loss objectives to encourage more effective distillation. Experiments conducted on four multimodal tasks demonstrate the effectiveness of the proposed MKD with Wasserstein distance minimization for compressing vision-language pretrained models.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2225-2237"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147578998","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Ent-BAM: A Generative Framework for Biomedical Argument Mining Enriched With Entity Information","authors":"Nilmadhab Das;Yash Sunil Choudhary;V. Vijaya Saradhi;Ashish Anand","doi":"10.1109/TAI.2025.3616076","DOIUrl":"https://doi.org/10.1109/TAI.2025.3616076","url":null,"abstract":"Biomedical argument mining (BAM) aims to identify and extract argumentative structures, specifically, <italic>argumentative components</i> (<italic>ACs</i>) and <italic>argumentative relations</i> (<italic>ARs</i>), within biomedical texts. Primary challenges in BAM arise from the rich presence of specialized terms and a diverse range of entities. These entities often <italic>occur within ACs</i> and <italic>co-occur in ARs</i>, providing valuable cues for identifying ACs and ARs. Most of the existing approaches do not leverage these cues. This article presents <italic>ent-BAM</i>, a BAM framework that incorporates entity (co-)occurrence information in a text-to-text generation paradigm. Plain text serves as the input, and the output is formatted in augmented natural language (ANL). First, we employ a prompt-based strategy using a large language model (LLM) to extract (co-)occurred entities from the standard BAM dataset. Then, we construct the ANL output by embedding AC and AR labels, along with those extracted entities, within a single sequence. Ent-BAM significantly outperforms existing state-of-the-art (SoTA) methods by an average task margin of up to 4.86% micro-F1 Score, highlighting the strong potential of our approach.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2346-2356"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147579028","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Toward Quantum Image Generation on Single Qubit Using Quantum Information Bottleneck","authors":"Yijie Zhu;Vaneet Aggarwal;Debanjan Konar;Yuri Pashkin;Plamen Angelov;Richard Jiang","doi":"10.1109/TAI.2025.3615187","DOIUrl":"https://doi.org/10.1109/TAI.2025.3615187","url":null,"abstract":"Amidst the rapidly evolving landscape of information technology, the convergence of quantum computing and machine learning—referred to as quantum machine learning—offers promising potential to enhance classical algorithms. However, significant challenges remain in both hardware and software implementation during the noisy intermediate-scale quantum (NISQ) era, including imperfect qubits, architectural constraints, and high noise levels. In response to these obstacles, this research introduces a novel solution: quantum convolutional variational autoencoders (QCVAEs), designed to operate with only a single qubit. This innovative approach efficiently utilizes a single qubit to manage large-scale data, making it particularly well suited for quantum computers with limited resources. Simulation results demonstrate the robustness of QCVAE in handling image data, and its deployment on a real quantum computer showcases the model’s practical viability. Additionally, the proposed approach leverages the information bottleneck principle to optimize quantum embeddings, effectively mitigating the impact of prevalent quantum noise. By addressing these core challenges, QCVAE presents a compelling solution for advancing quantum computing applications within the constraints of current NISQ technology.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2321-2331"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147579036","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"HEAL: Brain-Inspired Hyperdimensional Efficient Active Learning","authors":"Yang Ni;Zhuowen Zou;Wenjun Huang;Hanning Chen;William Youngwoo Chung;Samuel Cho;Ranganath Krishnan;Pietro Mercati;Mohsen Imani","doi":"10.1109/TAI.2025.3617929","DOIUrl":"https://doi.org/10.1109/TAI.2025.3617929","url":null,"abstract":"Drawing inspiration from the outstanding learning capability of our human brains, hyperdimensional computing (HDC) emerges as a novel computing paradigm, and it leverages high-dimensional vector representation and operations for brain-like lightweight machine learning (ML). Practical deployments of HDC have significantly enhanced the learning efficiency compared with current deep ML methods on a broad spectrum of applications. However, boosting the data efficiency of HDC classifiers in supervised learning remains an open question. In this article, we introduce hyperdimensional efficient active learning (<inline-formula><tex-math>$mathsf{HEAL}$</tex-math></inline-formula>), a novel active learning (AL) framework tailored for HDC classification. <inline-formula><tex-math>$mathsf{HEAL}$</tex-math></inline-formula> proactively annotates unlabeled data points via uncertainty and diversity-guided acquisition, leading to a more efficient dataset annotation and lowering labor costs. Unlike conventional AL methods that only support classifiers built upon deep neural networks (DNN), <inline-formula><tex-math>$mathsf{HEAL}$</tex-math></inline-formula> operates without the need for gradient or probabilistic computations. This allows it to be effortlessly integrated with any existing HDC classifier architecture. The key design of <inline-formula><tex-math>$mathsf{HEAL}$</tex-math></inline-formula> is a novel approach for uncertainty estimation in HDC classifiers through a lightweight HDC ensemble with prior hypervectors. Additionally, by exploiting hypervectors as prototypes (i.e., compact representations), we develop a sample acquisition strategy for <inline-formula><tex-math>$mathsf{HEAL}$</tex-math></inline-formula> to select diverse samples within each batch for annotation. Our evaluation shows that <inline-formula><tex-math>$mathsf{HEAL}$</tex-math></inline-formula> surpasses a diverse set of baselines in AL quality and achieves notably faster acquisition than many existing state-of-the-art AL methods, recording 12<inline-formula><tex-math>$times$</tex-math></inline-formula> to 45,700<inline-formula><tex-math>$times$</tex-math></inline-formula> speedup in acquisition runtime per batch.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2371-2386"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147578992","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Instance-Wise Joint Feature and Expert Decision Acquisition for Classification","authors":"Sachini Piyoni Ekanayake;Daphney-Stavroula Zois;Nipun Dulantha Wickramasinghe","doi":"10.1109/TAI.2025.3614214","DOIUrl":"https://doi.org/10.1109/TAI.2025.3614214","url":null,"abstract":"In many real-world applications, such as medical diagnosis, financial decision-making, and AI-driven content moderation, obtaining features and expert inputs is both time-consuming and expensive. To address these challenges, we present a supervised learning framework for sequential feature and expert decision acquisition. The objective of this framework, which operates in two phases, is to assign labels to each instance while reducing the cost of acquiring features and requesting experts’ decisions. The process begins with an initial belief about the label of a data instance, and at each step, there is a choice to either stop or continue acquiring features. Once feature acquisition terminates, the framework either assigns a label, or switches to expert decision acquisition, both based on the acquired features. The framework’s performance is evaluated using 13 publicly available datasets, showing improvements in accuracy while acquiring fewer features and expert decisions on average than baselines. Its generalizability across expert types is demonstrated through experiments using standard supervised learning models, widely-used ensemble methods, and neural networks as experts, highlighting its broad applicability across domains.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2276-2291"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147579031","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Physical Reservoir Computing for Optical Stethoscope-Based Heart Sound Biometric Identification","authors":"Yuqi Ding;Haobo Li;Xiangpeng Liang;Marija Vaskeviciute;Daniele Faccio;Hadi Heidari","doi":"10.1109/TAI.2025.3617382","DOIUrl":"https://doi.org/10.1109/TAI.2025.3617382","url":null,"abstract":"Heart sound signal has emerged as a promising solution to biometric identification. In this article, we use an optical flow algorithm to retrieve optical stethoscope-based heart sound signals from a laser-camera system. We apply physical reservoir computing (RC) for the classification algorithm. As a bio-inspired algorithm, physical RC has attracted growing research interests in recent years. We aim to create an efficient identification system by applying a recently proposed physical RC model called rotating neuron reservoir (RNR) as the processing core. Unlike conventional machine learning classifiers, RNR is a hardware-based neuromorphic model that preserves the majority of computing in the analogue domain, holding the promise of a next-generation machine learning accelerator. At the same time, the RNR, as a recurrent neural network (RNN), is suitable for time series data processing. The proposed system is verified by an experimentally collected heart sound dataset by laser-camera system achieving an overall accuracy of 89.03% in identifying twelve testing subjects. In addition, the elevated heart sound from eight subjects have been blended with their normal heart sounds to assess the robustness of the proposed system. The classification accuracy reaches over 83% in this mixed test. Furthermore, the identification system was assessed under individuals with different types of heart murmurs and abnormal heart sounds, achieving an overall accuracy of around 90%. The successful demonstration promises a novel application of physical RC for future biometric identification.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 4","pages":"2357-2370"},"PeriodicalIF":0.0,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147579041","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}