Artificial Intelligence Review最新文献

Intelligent Tutoring Systems (ITS) and allied digital platforms now constitute core infrastructure in many classrooms, where they automate formative assessment, personalise pacing and supply fine-grained analytics that would otherwise exceed human capacity. Against this backdrop, Large Language Models (LLMs) have emerged as a disruptive layer of functionality, expanding educational AI from rule-based tutoring to open-ended dialogue, generative content and real-time adaptation. Early classroom prototypes already leverage multi-agent LLM frameworks to orchestrate teacher-student and peer interactions, demonstrating richer discourse patterns and enhanced engagement when benchmarked with established observation rubrics. Most consequential, however, is the accelerating shift towards full simulation of teacher work. Emerging evidence suggests that prompting an LLM to rehearse lessons, generate reflective commentary, and iteratively revise materials can raise the quality of teaching plans to a level comparable to those crafted by expert educators. While the narrative highlights practical applications and pedagogical implications, this review is grounded in a systematic methodology combined with narrative analysis, ensuring analytical depth and thematic cohesion.

Percutaneous coronary intervention with stent implantation has become a widely used strategy to treat coronary artery disease. Stent malapposition (SM) may increase the risk of late stent thrombosis due to stent tissue coverage reduction, attracting much attention clinically. Recently, optical coherence tomography (OCT) images have been utilized to visually assess the stent apposition/malapposition. However, automated OCT-based SM recognition has been under-explored previously. Therefore, this paper proposes a novel enhanced pixel-wise style fusion network (EPSF-Net) to recognize SM automatically from OCT images. In the EPSF-Net, considering SM information is subtle, we design a novel enhanced pixel-wise style fusion (EPSF) block, which first applies the pixel-wise style pooling to aggregate pixel-wise style context, then enhances pixel-wise style context with multi-scale learning, and finally fuses enhanced pixel-wise style context via a pixel-wise fusion operator. Moreover, the re-parameterizing technique is utilized to reduce the parameters and computational cost of EPSF at the inference stage. Additionally, considering there is no publicly available OCT dataset for SM recognition, we construct an OCT image dataset of SM, named SM-OCT, to validate the effectiveness of our method, which will be available. The extensive experiments on the SM-OCT dataset show that our proposed EPSF-Net achieves better SM recognition performance than state-of-the-art methods. Additionally, two publicly available OCT datasets are employed to verify the generalization of our method.

Concept drift—changes in the underlying data distribution over time—poses a significant challenge to machine learning systems deployed in dynamic environments. While numerous drift detection methods have been developed for structured data such as tabular and time-series streams, concept drift in image data remains an underexplored area due to the unstructured and high-dimensional nature of visual information. This survey presents the first comprehensive review of concept drift detection methods tailored for image data streams. We propose a novel taxonomy that categorizes existing approaches based on key properties such as image feature handling, detection strategy, detection level, concept drift cause, and evaluation considerations. Through the lens of this taxonomy, we analyze 14 representative concept drift detection methods designed for image data, highlighting current approaches to the field, their strengths and limitations. Based on this analysis, we outline promising future research directions to advance the field of concept drift detection in image-based systems.

The increasing size and complexity of pre-trained language models have demonstrated superior performance in many applications, but they usually require large training datasets to be adequately trained. Insufficient training sets could unexpectedly lead the model to overfit and fail to cope with complex tasks. Large language models (LLMs) trained on extensive corpora have prominent text generation capabilities, which improve the quality and quantity of data and play a crucial role in data augmentation. Specifically, distinctive prompt templates are given in personalised tasks to guide LLMs in generating the required content. Recently, promising retrieval-based techniques have further enhanced the expressive performance of LLMs in data augmentation by introducing external knowledge, enabling them to produce more grounded data. This survey provides an in-depth analysis of data augmentation in LLMs, classifying the techniques into Simple Augmentation, Prompt-based Augmentation, Retrieval-based Augmentation, and Hybrid Augmentation. Additionally, we conduct extensive experiments across four techniques, systematically compare and analyse their performance, and provide key insights. Following this, we connect data augmentation with three critical optimisation techniques. Finally, we introduce existing challenges and future opportunities that could further improve data augmentation. This survey provides researchers and practitioners of the text modality with avenues to address data scarcity and improve data quality, helping scholars understand the evolution of text data augmentation from traditional methods to the application of human-like generation and agent search in the era of LLMs.

The Metaverse is an emerging virtual reality space that merges digital and physical worlds and provides users with immersive, interactive, and persistent virtual environments. The Metaverse leverages multiple technologies, including digital twins, blockchain, artificial intelligence, extended reality, and edge computing to realize the seamless connectivity and interaction between both worlds: physical and virtual. Artificial Intelligence (AI) empowers intelligent decisions in such complex dynamic environments. More specifically, Reinforcement Learning (RL) is uniquely effective in the context of Metaverse applications due to the natural process of learning through interaction and its modeling of sequential decision making, allowing it to be flexible, dynamic, and able to discover complex strategies and emergent behavior in complicated environments where programming explicit rules is impractical. Although multiple works have explored the research on the Metaverse and AI-based applications, there remains a significant gap in the literature that addresses the contribution of RL algorithms within the Metaverse. Therefore, this review presents a comprehensive overview of RL algorithms for Metaverse applications. We examine the architecture of Metaverse networks, the role of RL in enhancing virtual interactions, and the potential for transferring learned behaviors to real-world applications. Furthermore, we categorize the key challenges, opportunities, and research directions associated with deploying RL in the Metaverse.

The use of Artificial Intelligence (AI) and Unmanned Aerial Vehicles (UAVs), also known as drones, is changing the way future communication and networking systems are designed. UAVs can collect data, support wireless networks, and help deliver services from the sky, which makes them an important part of modern technology. To understand these developments, we reviewed almost 250 research papers published between 2015 and 2024. Our review focuses on UAV network design, communication methods, energy management, AI-based optimization, and future challenges. Unlike previous surveys that mainly summarize individual technical domains, this work introduces a new AI-driven UAV classification framework that connects these aspects under one structure. The framework organizes UAV systems across five dimensions–mission adaptability, autonomy level, communication intelligence, scalability, and deployment context–providing a unified way to compare current and future UAV technologies. This analytical structure highlights how artificial intelligence enables UAVs to move from static, pre-defined operations toward dynamic, real-time decision-making and mission-specific adaptation. We found that deep learning and reinforcement learning are the most common AI methods used to improve routing, flight planning, resource use, and network performance. These techniques help UAV networks adapt to changing conditions and reduce communication delays. However, we also found several open challenges, such as improving real-time energy efficiency, increasing security and privacy, managing large drone groups (swarms), and dealing with regulatory and policy issues. By combining this new framework with an extensive literature review, the paper offers a holistic view that not only summarizes past progress but also maps existing gaps and trends for future research. This paper provides a clear summary of current research, explains key trends, and points out gaps such as the need for lightweight AI models and better swarm coordination. The insights from this review can help researchers and engineers build smarter, safer, and more efficient UAV networks in the future.

Generative AI has emerged as a transformative technology in healthcare, enabling the generation of high-fidelity synthetic data for applications such as medical imaging, electronic health records, biomedical signal processing, and drug discovery. The increasing reliance on machine learning in healthcare necessitates large-scale, high-quality datasets, yet real-world data acquisition is often constrained by privacy regulations, heterogeneity, and limited accessibility. Generative AI models provide a viable solution by generating realistic and diverse synthetic datasets while preserving patient confidentiality. Unlike prior reviews that primarily focus on specific model classes or applications, this study fills a significant research gap by offering a unified, comparative evaluation of diverse generative models, including Generative Adversarial Networks, Variational Autoencoders, Transformers, and Diffusion Models, as well as their adaptations for privacy-preserving Federated Learning environments. Each model class is examined in terms of its variants, underlying methodologies, performance in healthcare applications, strengths, limitations, and computational feasibility. The study also investigates practical considerations for deploying generative AI in clinical settings, including challenges related to training stability, bias mitigation, model interpretability, and regulatory compliance. The insights from this review provide guidance for researchers and healthcare practitioners in selecting and optimizing generative AI models for medical applications, laying the foundation for future advancements in AI-driven healthcare solutions.

With the rapid development of high-power density instruments, spray cooling has drawn increasing interest in industry as a high-efficiency thermal management technology. Despite extensive research, typical spray cooling systems only function effectively within constrained conditions. Therefore, creating a general model for spray cooling is essential for accurately predicting its functionalities. In this work, we employed six machine learning (ML) algorithms to analyze the thermal performance and hydraulic properties of spray cooling. Leveraging data from 25 previous studies encompassing different working fluids, spray atomization types, Reynolds numbers ((:Re)), Nusselt numbers (:left(Nuright)), and Weber numbers ((:We)), our ML models significantly enhance the prediction of spray cooling functionalities compared to traditional correlations. The effectiveness of these ML algorithms was experimentally validated, yielding mean absolute percentage errors (MAPEs) of 6(% -)20(% -) for (:Nu) and 4(% -)16(% -) for mean droplet diameter (:{d}_{d}), respectively. Then, we proposed a general correlation for the thermal performances of various working fluids, atomization methods, and operational conditions, achieving a 38% reduction in MAPE compared to the most accurate existing correlation. Subsequently, this general correlation was integrated into the ML models, resulting in MAPEs ranging from 0.48% to 2.3%. Furthermore, we optimized the key factors of spray cooling with the (:Nu) number reaching 220. Finally, we employed SHapley Additive exPlanations (SHAP) approach to interpret the ML models and to identify an optimal strategy towards greatly enhanced thermal performance. This study demonstrates that ML significantly outperforms the empirical correlations for evaluating spray cooling performance and functionalities, paving a new avenue for thermoregulation of modern power systems.

Pose estimation and tracking in sports has gained significant attention due to its potential to revolutionize performance analysis, injury prevention, and strategic decision-making. This survey presents a comprehensive overview of the methodologies, datasets, challenges, and future directions in this rapidly evolving field. We explore traditional approaches, including geometric and statistical models, and highlight the transformative impact of deep learning techniques, such as convolutional neural networks, transformers, and hybrid architectures, which have enabled highly accurate and robust pose estimation. The paper also discusses dataset creation and ground-truthing techniques tailored to sports contexts, emphasizing the importance of multimodal data, scalability, and representativeness. Applications across diverse sports, from individual to team-based activities, demonstrate the versatility of pose estimation systems in both real-time and offline settings. However, challenges such as occlusions, dynamic backgrounds, and computational efficiency persist, necessitating further innovation. We identify future research directions, including the integration of multimodal data, edge computing, and ethical considerations, to enhance accuracy, interpretability, and generalizability. This survey aims to provide a foundational reference for researchers and practitioners, fostering advancements in pose estimation and tracking technologies that meet the unique demands of sports analytics.

Over the past few decades, Human Skeleton Modeling (HSM) has gained considerable attention in computer vision, exploring various practical applications such as the video surveillance, the human-computer interaction, the medical assistance analysis, and the autonomous driving through images and videos. The performance of HSM and its applications on challenging datasets has been significantly improved due to recent advancements of deep learning methods. These advancements have been extended to non-Euclidean or graph data with multiple nodes and edges. Because human joints and skeleton combinations are represented as graph structures, graph networks are appropriate for the non-Euclidean HSM. In recent years, graph networks have become essential tools for the HSM and behavioral analyses. However, prior surveys are often siloed, focusing either on a narrow class of models such as GCNs or on a single application like action recognition. A unified framework that systematically analyzes diverse graph network learning paradigms across the entire HSM pipeline has been notably absent. We conduct a survey of graph network methods for HSM and their application domains. This comprehensive overview includes a taxonomy of graph network techniques, a detailed study of benchmark datasets for HSM, extensive descriptions of the performance of graph networks in three major application domains, and a collection of related resources and open-source codes. Finally, we provided insightful recommendations for future research directions and trends of graph networks for HSM. This survey serves as the introductory material for beginners in graph network-based HSM and as the reference materials for advanced researchers.